246}
247
248void
249zio_buf_free(void *buf, size_t size)
250{
251 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
252
253 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
254
255 if (zio_use_uma)
256 kmem_cache_free(zio_buf_cache[c], buf);
257 else
258 kmem_free(buf, size);
259}
260
261void
262zio_data_buf_free(void *buf, size_t size)
263{
264 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
265
266 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
267
268 if (zio_use_uma)
269 kmem_cache_free(zio_data_buf_cache[c], buf);
270 else
271 kmem_free(buf, size);
272}
273
274/*
275 * ==========================================================================
276 * Push and pop I/O transform buffers
277 * ==========================================================================
278 */
279static void
280zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
281 zio_transform_func_t *transform)
282{
283 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
284
285 zt->zt_orig_data = zio->io_data;
286 zt->zt_orig_size = zio->io_size;
287 zt->zt_bufsize = bufsize;
288 zt->zt_transform = transform;
289
290 zt->zt_next = zio->io_transform_stack;
291 zio->io_transform_stack = zt;
292
293 zio->io_data = data;
294 zio->io_size = size;
295}
296
297static void
298zio_pop_transforms(zio_t *zio)
299{
300 zio_transform_t *zt;
301
302 while ((zt = zio->io_transform_stack) != NULL) {
303 if (zt->zt_transform != NULL)
304 zt->zt_transform(zio,
305 zt->zt_orig_data, zt->zt_orig_size);
306
307 if (zt->zt_bufsize != 0)
308 zio_buf_free(zio->io_data, zt->zt_bufsize);
309
310 zio->io_data = zt->zt_orig_data;
311 zio->io_size = zt->zt_orig_size;
312 zio->io_transform_stack = zt->zt_next;
313
314 kmem_free(zt, sizeof (zio_transform_t));
315 }
316}
317
318/*
319 * ==========================================================================
320 * I/O transform callbacks for subblocks and decompression
321 * ==========================================================================
322 */
323static void
324zio_subblock(zio_t *zio, void *data, uint64_t size)
325{
326 ASSERT(zio->io_size > size);
327
328 if (zio->io_type == ZIO_TYPE_READ)
329 bcopy(zio->io_data, data, size);
330}
331
332static void
333zio_decompress(zio_t *zio, void *data, uint64_t size)
334{
335 if (zio->io_error == 0 &&
336 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
337 zio->io_data, data, zio->io_size, size) != 0)
338 zio->io_error = EIO;
339}
340
341/*
342 * ==========================================================================
343 * I/O parent/child relationships and pipeline interlocks
344 * ==========================================================================
345 */
346/*
347 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
348 * continue calling these functions until they return NULL.
349 * Otherwise, the next caller will pick up the list walk in
350 * some indeterminate state. (Otherwise every caller would
351 * have to pass in a cookie to keep the state represented by
352 * io_walk_link, which gets annoying.)
353 */
354zio_t *
355zio_walk_parents(zio_t *cio)
356{
357 zio_link_t *zl = cio->io_walk_link;
358 list_t *pl = &cio->io_parent_list;
359
360 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
361 cio->io_walk_link = zl;
362
363 if (zl == NULL)
364 return (NULL);
365
366 ASSERT(zl->zl_child == cio);
367 return (zl->zl_parent);
368}
369
370zio_t *
371zio_walk_children(zio_t *pio)
372{
373 zio_link_t *zl = pio->io_walk_link;
374 list_t *cl = &pio->io_child_list;
375
376 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
377 pio->io_walk_link = zl;
378
379 if (zl == NULL)
380 return (NULL);
381
382 ASSERT(zl->zl_parent == pio);
383 return (zl->zl_child);
384}
385
386zio_t *
387zio_unique_parent(zio_t *cio)
388{
389 zio_t *pio = zio_walk_parents(cio);
390
391 VERIFY(zio_walk_parents(cio) == NULL);
392 return (pio);
393}
394
395void
396zio_add_child(zio_t *pio, zio_t *cio)
397{
398 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
399
400 /*
401 * Logical I/Os can have logical, gang, or vdev children.
402 * Gang I/Os can have gang or vdev children.
403 * Vdev I/Os can only have vdev children.
404 * The following ASSERT captures all of these constraints.
405 */
406 ASSERT(cio->io_child_type <= pio->io_child_type);
407
408 zl->zl_parent = pio;
409 zl->zl_child = cio;
410
411 mutex_enter(&cio->io_lock);
412 mutex_enter(&pio->io_lock);
413
414 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
415
416 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
417 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
418
419 list_insert_head(&pio->io_child_list, zl);
420 list_insert_head(&cio->io_parent_list, zl);
421
422 pio->io_child_count++;
423 cio->io_parent_count++;
424
425 mutex_exit(&pio->io_lock);
426 mutex_exit(&cio->io_lock);
427}
428
429static void
430zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
431{
432 ASSERT(zl->zl_parent == pio);
433 ASSERT(zl->zl_child == cio);
434
435 mutex_enter(&cio->io_lock);
436 mutex_enter(&pio->io_lock);
437
438 list_remove(&pio->io_child_list, zl);
439 list_remove(&cio->io_parent_list, zl);
440
441 pio->io_child_count--;
442 cio->io_parent_count--;
443
444 mutex_exit(&pio->io_lock);
445 mutex_exit(&cio->io_lock);
446
447 kmem_cache_free(zio_link_cache, zl);
448}
449
450static boolean_t
451zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
452{
453 uint64_t *countp = &zio->io_children[child][wait];
454 boolean_t waiting = B_FALSE;
455
456 mutex_enter(&zio->io_lock);
457 ASSERT(zio->io_stall == NULL);
458 if (*countp != 0) {
459 zio->io_stage >>= 1;
460 zio->io_stall = countp;
461 waiting = B_TRUE;
462 }
463 mutex_exit(&zio->io_lock);
464
465 return (waiting);
466}
467
468static void
469zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
470{
471 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
472 int *errorp = &pio->io_child_error[zio->io_child_type];
473
474 mutex_enter(&pio->io_lock);
475 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
476 *errorp = zio_worst_error(*errorp, zio->io_error);
477 pio->io_reexecute |= zio->io_reexecute;
478 ASSERT3U(*countp, >, 0);
479 if (--*countp == 0 && pio->io_stall == countp) {
480 pio->io_stall = NULL;
481 mutex_exit(&pio->io_lock);
482 zio_execute(pio);
483 } else {
484 mutex_exit(&pio->io_lock);
485 }
486}
487
488static void
489zio_inherit_child_errors(zio_t *zio, enum zio_child c)
490{
491 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
492 zio->io_error = zio->io_child_error[c];
493}
494
495/*
496 * ==========================================================================
497 * Create the various types of I/O (read, write, free, etc)
498 * ==========================================================================
499 */
500static zio_t *
501zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
502 void *data, uint64_t size, zio_done_func_t *done, void *private,
503 zio_type_t type, int priority, enum zio_flag flags,
504 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
505 enum zio_stage stage, enum zio_stage pipeline)
506{
507 zio_t *zio;
508
509 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
510 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
511 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
512
513 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
514 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
515 ASSERT(vd || stage == ZIO_STAGE_OPEN);
516
517 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
518 bzero(zio, sizeof (zio_t));
519
520 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
521 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
522
523 list_create(&zio->io_parent_list, sizeof (zio_link_t),
524 offsetof(zio_link_t, zl_parent_node));
525 list_create(&zio->io_child_list, sizeof (zio_link_t),
526 offsetof(zio_link_t, zl_child_node));
527
528 if (vd != NULL)
529 zio->io_child_type = ZIO_CHILD_VDEV;
530 else if (flags & ZIO_FLAG_GANG_CHILD)
531 zio->io_child_type = ZIO_CHILD_GANG;
532 else if (flags & ZIO_FLAG_DDT_CHILD)
533 zio->io_child_type = ZIO_CHILD_DDT;
534 else
535 zio->io_child_type = ZIO_CHILD_LOGICAL;
536
537 if (bp != NULL) {
538 zio->io_bp = (blkptr_t *)bp;
539 zio->io_bp_copy = *bp;
540 zio->io_bp_orig = *bp;
541 if (type != ZIO_TYPE_WRITE ||
542 zio->io_child_type == ZIO_CHILD_DDT)
543 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
544 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
545 zio->io_logical = zio;
546 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
547 pipeline |= ZIO_GANG_STAGES;
548 }
549
550 zio->io_spa = spa;
551 zio->io_txg = txg;
552 zio->io_done = done;
553 zio->io_private = private;
554 zio->io_type = type;
555 zio->io_priority = priority;
556 zio->io_vd = vd;
557 zio->io_offset = offset;
558 zio->io_orig_data = zio->io_data = data;
559 zio->io_orig_size = zio->io_size = size;
560 zio->io_orig_flags = zio->io_flags = flags;
561 zio->io_orig_stage = zio->io_stage = stage;
562 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
563
564 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
565 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
566
567 if (zb != NULL)
568 zio->io_bookmark = *zb;
569
570 if (pio != NULL) {
571 if (zio->io_logical == NULL)
572 zio->io_logical = pio->io_logical;
573 if (zio->io_child_type == ZIO_CHILD_GANG)
574 zio->io_gang_leader = pio->io_gang_leader;
575 zio_add_child(pio, zio);
576 }
577
578 return (zio);
579}
580
581static void
582zio_destroy(zio_t *zio)
583{
584 list_destroy(&zio->io_parent_list);
585 list_destroy(&zio->io_child_list);
586 mutex_destroy(&zio->io_lock);
587 cv_destroy(&zio->io_cv);
588 kmem_cache_free(zio_cache, zio);
589}
590
591zio_t *
592zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
593 void *private, enum zio_flag flags)
594{
595 zio_t *zio;
596
597 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
598 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
599 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
600
601 return (zio);
602}
603
604zio_t *
605zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
606{
607 return (zio_null(NULL, spa, NULL, done, private, flags));
608}
609
610zio_t *
611zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
612 void *data, uint64_t size, zio_done_func_t *done, void *private,
613 int priority, enum zio_flag flags, const zbookmark_t *zb)
614{
615 zio_t *zio;
616
617 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
618 data, size, done, private,
619 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
620 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
621 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
622
623 return (zio);
624}
625
626zio_t *
627zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
628 void *data, uint64_t size, const zio_prop_t *zp,
629 zio_done_func_t *ready, zio_done_func_t *done, void *private,
630 int priority, enum zio_flag flags, const zbookmark_t *zb)
631{
632 zio_t *zio;
633
634 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
635 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
636 zp->zp_compress >= ZIO_COMPRESS_OFF &&
637 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
638 zp->zp_type < DMU_OT_NUMTYPES &&
639 zp->zp_level < 32 &&
640 zp->zp_copies > 0 &&
641 zp->zp_copies <= spa_max_replication(spa) &&
642 zp->zp_dedup <= 1 &&
643 zp->zp_dedup_verify <= 1);
644
645 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
646 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
647 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
648 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
649
650 zio->io_ready = ready;
651 zio->io_prop = *zp;
652
653 return (zio);
654}
655
656zio_t *
657zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
658 uint64_t size, zio_done_func_t *done, void *private, int priority,
659 enum zio_flag flags, zbookmark_t *zb)
660{
661 zio_t *zio;
662
663 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
664 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
665 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
666
667 return (zio);
668}
669
670void
671zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
672{
673 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
674 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
675 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
676 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
677
678 zio->io_prop.zp_copies = copies;
679 zio->io_bp_override = bp;
680}
681
682void
683zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
684{
685 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
686}
687
688zio_t *
689zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
690 enum zio_flag flags)
691{
692 zio_t *zio;
693
694 dprintf_bp(bp, "freeing in txg %llu, pass %u",
695 (longlong_t)txg, spa->spa_sync_pass);
696
697 ASSERT(!BP_IS_HOLE(bp));
698 ASSERT(spa_syncing_txg(spa) == txg);
699 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
700
701 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
702 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
703 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
704
705 return (zio);
706}
707
708zio_t *
709zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
710 zio_done_func_t *done, void *private, enum zio_flag flags)
711{
712 zio_t *zio;
713
714 /*
715 * A claim is an allocation of a specific block. Claims are needed
716 * to support immediate writes in the intent log. The issue is that
717 * immediate writes contain committed data, but in a txg that was
718 * *not* committed. Upon opening the pool after an unclean shutdown,
719 * the intent log claims all blocks that contain immediate write data
720 * so that the SPA knows they're in use.
721 *
722 * All claims *must* be resolved in the first txg -- before the SPA
723 * starts allocating blocks -- so that nothing is allocated twice.
724 * If txg == 0 we just verify that the block is claimable.
725 */
726 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
727 ASSERT(txg == spa_first_txg(spa) || txg == 0);
728 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
729
730 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
731 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
732 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
733
734 return (zio);
735}
736
737zio_t *
738zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
739 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
740{
741 zio_t *zio;
742 int c;
743
744 if (vd->vdev_children == 0) {
745 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
746 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
747 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
748
749 zio->io_cmd = cmd;
750 } else {
751 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
752
753 for (c = 0; c < vd->vdev_children; c++)
754 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
755 done, private, priority, flags));
756 }
757
758 return (zio);
759}
760
761zio_t *
762zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
763 void *data, int checksum, zio_done_func_t *done, void *private,
764 int priority, enum zio_flag flags, boolean_t labels)
765{
766 zio_t *zio;
767
768 ASSERT(vd->vdev_children == 0);
769 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
770 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
771 ASSERT3U(offset + size, <=, vd->vdev_psize);
772
773 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
774 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
775 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
776
777 zio->io_prop.zp_checksum = checksum;
778
779 return (zio);
780}
781
782zio_t *
783zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
784 void *data, int checksum, zio_done_func_t *done, void *private,
785 int priority, enum zio_flag flags, boolean_t labels)
786{
787 zio_t *zio;
788
789 ASSERT(vd->vdev_children == 0);
790 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
791 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
792 ASSERT3U(offset + size, <=, vd->vdev_psize);
793
794 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
795 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
796 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
797
798 zio->io_prop.zp_checksum = checksum;
799
800 if (zio_checksum_table[checksum].ci_eck) {
801 /*
802 * zec checksums are necessarily destructive -- they modify
803 * the end of the write buffer to hold the verifier/checksum.
804 * Therefore, we must make a local copy in case the data is
805 * being written to multiple places in parallel.
806 */
807 void *wbuf = zio_buf_alloc(size);
808 bcopy(data, wbuf, size);
809 zio_push_transform(zio, wbuf, size, size, NULL);
810 }
811
812 return (zio);
813}
814
815/*
816 * Create a child I/O to do some work for us.
817 */
818zio_t *
819zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
820 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
821 zio_done_func_t *done, void *private)
822{
823 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
824 zio_t *zio;
825
826 ASSERT(vd->vdev_parent ==
827 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
828
829 if (type == ZIO_TYPE_READ && bp != NULL) {
830 /*
831 * If we have the bp, then the child should perform the
832 * checksum and the parent need not. This pushes error
833 * detection as close to the leaves as possible and
834 * eliminates redundant checksums in the interior nodes.
835 */
836 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
837 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
838 }
839
840 if (vd->vdev_children == 0)
841 offset += VDEV_LABEL_START_SIZE;
842
843 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
844
845 /*
846 * If we've decided to do a repair, the write is not speculative --
847 * even if the original read was.
848 */
849 if (flags & ZIO_FLAG_IO_REPAIR)
850 flags &= ~ZIO_FLAG_SPECULATIVE;
851
852 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
853 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
854 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
855
856 return (zio);
857}
858
859zio_t *
860zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
861 int type, int priority, enum zio_flag flags,
862 zio_done_func_t *done, void *private)
863{
864 zio_t *zio;
865
866 ASSERT(vd->vdev_ops->vdev_op_leaf);
867
868 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
869 data, size, done, private, type, priority,
870 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
871 vd, offset, NULL,
872 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
873
874 return (zio);
875}
876
877void
878zio_flush(zio_t *zio, vdev_t *vd)
879{
880 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
881 NULL, NULL, ZIO_PRIORITY_NOW,
882 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
883}
884
885void
886zio_shrink(zio_t *zio, uint64_t size)
887{
888 ASSERT(zio->io_executor == NULL);
889 ASSERT(zio->io_orig_size == zio->io_size);
890 ASSERT(size <= zio->io_size);
891
892 /*
893 * We don't shrink for raidz because of problems with the
894 * reconstruction when reading back less than the block size.
895 * Note, BP_IS_RAIDZ() assumes no compression.
896 */
897 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
898 if (!BP_IS_RAIDZ(zio->io_bp))
899 zio->io_orig_size = zio->io_size = size;
900}
901
902/*
903 * ==========================================================================
904 * Prepare to read and write logical blocks
905 * ==========================================================================
906 */
907
908static int
909zio_read_bp_init(zio_t *zio)
910{
911 blkptr_t *bp = zio->io_bp;
912
913 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
914 zio->io_child_type == ZIO_CHILD_LOGICAL &&
915 !(zio->io_flags & ZIO_FLAG_RAW)) {
916 uint64_t psize = BP_GET_PSIZE(bp);
917 void *cbuf = zio_buf_alloc(psize);
918
919 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
920 }
921
922 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
923 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
924
925 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
926 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
927
928 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
929 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
930
931 return (ZIO_PIPELINE_CONTINUE);
932}
933
934static int
935zio_write_bp_init(zio_t *zio)
936{
937 spa_t *spa = zio->io_spa;
938 zio_prop_t *zp = &zio->io_prop;
939 enum zio_compress compress = zp->zp_compress;
940 blkptr_t *bp = zio->io_bp;
941 uint64_t lsize = zio->io_size;
942 uint64_t psize = lsize;
943 int pass = 1;
944
945 /*
946 * If our children haven't all reached the ready stage,
947 * wait for them and then repeat this pipeline stage.
948 */
949 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
950 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
951 return (ZIO_PIPELINE_STOP);
952
953 if (!IO_IS_ALLOCATING(zio))
954 return (ZIO_PIPELINE_CONTINUE);
955
956 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
957
958 if (zio->io_bp_override) {
959 ASSERT(bp->blk_birth != zio->io_txg);
960 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
961
962 *bp = *zio->io_bp_override;
963 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
964
965 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
966 return (ZIO_PIPELINE_CONTINUE);
967
968 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
969 zp->zp_dedup_verify);
970
971 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
972 BP_SET_DEDUP(bp, 1);
973 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
974 return (ZIO_PIPELINE_CONTINUE);
975 }
976 zio->io_bp_override = NULL;
977 BP_ZERO(bp);
978 }
979
980 if (bp->blk_birth == zio->io_txg) {
981 /*
982 * We're rewriting an existing block, which means we're
983 * working on behalf of spa_sync(). For spa_sync() to
984 * converge, it must eventually be the case that we don't
985 * have to allocate new blocks. But compression changes
986 * the blocksize, which forces a reallocate, and makes
987 * convergence take longer. Therefore, after the first
988 * few passes, stop compressing to ensure convergence.
989 */
990 pass = spa_sync_pass(spa);
991
992 ASSERT(zio->io_txg == spa_syncing_txg(spa));
993 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
994 ASSERT(!BP_GET_DEDUP(bp));
995
996 if (pass > SYNC_PASS_DONT_COMPRESS)
997 compress = ZIO_COMPRESS_OFF;
998
999 /* Make sure someone doesn't change their mind on overwrites */
1000 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1001 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1002 }
1003
1004 if (compress != ZIO_COMPRESS_OFF) {
1005 void *cbuf = zio_buf_alloc(lsize);
1006 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1007 if (psize == 0 || psize == lsize) {
1008 compress = ZIO_COMPRESS_OFF;
1009 zio_buf_free(cbuf, lsize);
1010 } else {
1011 ASSERT(psize < lsize);
1012 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1013 }
1014 }
1015
1016 /*
1017 * The final pass of spa_sync() must be all rewrites, but the first
1018 * few passes offer a trade-off: allocating blocks defers convergence,
1019 * but newly allocated blocks are sequential, so they can be written
1020 * to disk faster. Therefore, we allow the first few passes of
1021 * spa_sync() to allocate new blocks, but force rewrites after that.
1022 * There should only be a handful of blocks after pass 1 in any case.
1023 */
1024 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1025 pass > SYNC_PASS_REWRITE) {
1026 ASSERT(psize != 0);
1027 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1028 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1029 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1030 } else {
1031 BP_ZERO(bp);
1032 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1033 }
1034
1035 if (psize == 0) {
1036 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1037 } else {
1038 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1039 BP_SET_LSIZE(bp, lsize);
1040 BP_SET_PSIZE(bp, psize);
1041 BP_SET_COMPRESS(bp, compress);
1042 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1043 BP_SET_TYPE(bp, zp->zp_type);
1044 BP_SET_LEVEL(bp, zp->zp_level);
1045 BP_SET_DEDUP(bp, zp->zp_dedup);
1046 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1047 if (zp->zp_dedup) {
1048 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1049 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1050 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1051 }
1052 }
1053
1054 return (ZIO_PIPELINE_CONTINUE);
1055}
1056
1057static int
1058zio_free_bp_init(zio_t *zio)
1059{
1060 blkptr_t *bp = zio->io_bp;
1061
1062 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1063 if (BP_GET_DEDUP(bp))
1064 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1065 }
1066
1067 return (ZIO_PIPELINE_CONTINUE);
1068}
1069
1070/*
1071 * ==========================================================================
1072 * Execute the I/O pipeline
1073 * ==========================================================================
1074 */
1075
1076static void
1077zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1078{
1079 spa_t *spa = zio->io_spa;
1080 zio_type_t t = zio->io_type;
1081 int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1082
1083 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1084
1085 /*
1086 * If we're a config writer or a probe, the normal issue and
1087 * interrupt threads may all be blocked waiting for the config lock.
1088 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1089 */
1090 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1091 t = ZIO_TYPE_NULL;
1092
1093 /*
1094 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1095 */
1096 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1097 t = ZIO_TYPE_NULL;
1098
1099 /*
1100 * If this is a high priority I/O, then use the high priority taskq.
1101 */
1102 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1103 spa->spa_zio_taskq[t][q + 1] != NULL)
1104 q++;
1105
1106 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1107#ifdef _KERNEL
1108 (void) taskq_dispatch_safe(spa->spa_zio_taskq[t][q],
1109 (task_func_t *)zio_execute, zio, flags, &zio->io_task);
1110#else
1111 (void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1112 (task_func_t *)zio_execute, zio, flags);
1113#endif
1114}
1115
1116static boolean_t
1117zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1118{
1119 kthread_t *executor = zio->io_executor;
1120 spa_t *spa = zio->io_spa;
1121
1122 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1123 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1124 return (B_TRUE);
1125
1126 return (B_FALSE);
1127}
1128
1129static int
1130zio_issue_async(zio_t *zio)
1131{
1132 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1133
1134 return (ZIO_PIPELINE_STOP);
1135}
1136
1137void
1138zio_interrupt(zio_t *zio)
1139{
1140 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1141}
1142
1143/*
1144 * Execute the I/O pipeline until one of the following occurs:
1145 * (1) the I/O completes; (2) the pipeline stalls waiting for
1146 * dependent child I/Os; (3) the I/O issues, so we're waiting
1147 * for an I/O completion interrupt; (4) the I/O is delegated by
1148 * vdev-level caching or aggregation; (5) the I/O is deferred
1149 * due to vdev-level queueing; (6) the I/O is handed off to
1150 * another thread. In all cases, the pipeline stops whenever
1151 * there's no CPU work; it never burns a thread in cv_wait().
1152 *
1153 * There's no locking on io_stage because there's no legitimate way
1154 * for multiple threads to be attempting to process the same I/O.
1155 */
1156static zio_pipe_stage_t *zio_pipeline[];
1157
1158void
1159zio_execute(zio_t *zio)
1160{
1161 zio->io_executor = curthread;
1162
1163 while (zio->io_stage < ZIO_STAGE_DONE) {
1164 enum zio_stage pipeline = zio->io_pipeline;
1165 enum zio_stage stage = zio->io_stage;
1166 int rv;
1167
1168 ASSERT(!MUTEX_HELD(&zio->io_lock));
1169 ASSERT(ISP2(stage));
1170 ASSERT(zio->io_stall == NULL);
1171
1172 do {
1173 stage <<= 1;
1174 } while ((stage & pipeline) == 0);
1175
1176 ASSERT(stage <= ZIO_STAGE_DONE);
1177
1178 /*
1179 * If we are in interrupt context and this pipeline stage
1180 * will grab a config lock that is held across I/O,
1181 * or may wait for an I/O that needs an interrupt thread
1182 * to complete, issue async to avoid deadlock.
1183 *
1184 * For VDEV_IO_START, we cut in line so that the io will
1185 * be sent to disk promptly.
1186 */
1187 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1188 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1189 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1190 zio_requeue_io_start_cut_in_line : B_FALSE;
1191 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1192 return;
1193 }
1194
1195 zio->io_stage = stage;
1196 rv = zio_pipeline[highbit(stage) - 1](zio);
1197
1198 if (rv == ZIO_PIPELINE_STOP)
1199 return;
1200
1201 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1202 }
1203}
1204
1205/*
1206 * ==========================================================================
1207 * Initiate I/O, either sync or async
1208 * ==========================================================================
1209 */
1210int
1211zio_wait(zio_t *zio)
1212{
1213 int error;
1214
1215 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1216 ASSERT(zio->io_executor == NULL);
1217
1218 zio->io_waiter = curthread;
1219
1220 zio_execute(zio);
1221
1222 mutex_enter(&zio->io_lock);
1223 while (zio->io_executor != NULL)
1224 cv_wait(&zio->io_cv, &zio->io_lock);
1225 mutex_exit(&zio->io_lock);
1226
1227 error = zio->io_error;
1228 zio_destroy(zio);
1229
1230 return (error);
1231}
1232
1233void
1234zio_nowait(zio_t *zio)
1235{
1236 ASSERT(zio->io_executor == NULL);
1237
1238 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1239 zio_unique_parent(zio) == NULL) {
1240 /*
1241 * This is a logical async I/O with no parent to wait for it.
1242 * We add it to the spa_async_root_zio "Godfather" I/O which
1243 * will ensure they complete prior to unloading the pool.
1244 */
1245 spa_t *spa = zio->io_spa;
1246
1247 zio_add_child(spa->spa_async_zio_root, zio);
1248 }
1249
1250 zio_execute(zio);
1251}
1252
1253/*
1254 * ==========================================================================
1255 * Reexecute or suspend/resume failed I/O
1256 * ==========================================================================
1257 */
1258
1259static void
1260zio_reexecute(zio_t *pio)
1261{
1262 zio_t *cio, *cio_next;
1263
1264 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1265 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1266 ASSERT(pio->io_gang_leader == NULL);
1267 ASSERT(pio->io_gang_tree == NULL);
1268
1269 pio->io_flags = pio->io_orig_flags;
1270 pio->io_stage = pio->io_orig_stage;
1271 pio->io_pipeline = pio->io_orig_pipeline;
1272 pio->io_reexecute = 0;
1273 pio->io_error = 0;
1274 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1275 pio->io_state[w] = 0;
1276 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1277 pio->io_child_error[c] = 0;
1278
1279 if (IO_IS_ALLOCATING(pio))
1280 BP_ZERO(pio->io_bp);
1281
1282 /*
1283 * As we reexecute pio's children, new children could be created.
1284 * New children go to the head of pio's io_child_list, however,
1285 * so we will (correctly) not reexecute them. The key is that
1286 * the remainder of pio's io_child_list, from 'cio_next' onward,
1287 * cannot be affected by any side effects of reexecuting 'cio'.
1288 */
1289 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1290 cio_next = zio_walk_children(pio);
1291 mutex_enter(&pio->io_lock);
1292 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1293 pio->io_children[cio->io_child_type][w]++;
1294 mutex_exit(&pio->io_lock);
1295 zio_reexecute(cio);
1296 }
1297
1298 /*
1299 * Now that all children have been reexecuted, execute the parent.
1300 * We don't reexecute "The Godfather" I/O here as it's the
1301 * responsibility of the caller to wait on him.
1302 */
1303 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1304 zio_execute(pio);
1305}
1306
1307void
1308zio_suspend(spa_t *spa, zio_t *zio)
1309{
1310 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1311 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1312 "failure and the failure mode property for this pool "
1313 "is set to panic.", spa_name(spa));
1314
1315 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1316
1317 mutex_enter(&spa->spa_suspend_lock);
1318
1319 if (spa->spa_suspend_zio_root == NULL)
1320 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1321 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1322 ZIO_FLAG_GODFATHER);
1323
1324 spa->spa_suspended = B_TRUE;
1325
1326 if (zio != NULL) {
1327 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1328 ASSERT(zio != spa->spa_suspend_zio_root);
1329 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1330 ASSERT(zio_unique_parent(zio) == NULL);
1331 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1332 zio_add_child(spa->spa_suspend_zio_root, zio);
1333 }
1334
1335 mutex_exit(&spa->spa_suspend_lock);
1336}
1337
1338int
1339zio_resume(spa_t *spa)
1340{
1341 zio_t *pio;
1342
1343 /*
1344 * Reexecute all previously suspended i/o.
1345 */
1346 mutex_enter(&spa->spa_suspend_lock);
1347 spa->spa_suspended = B_FALSE;
1348 cv_broadcast(&spa->spa_suspend_cv);
1349 pio = spa->spa_suspend_zio_root;
1350 spa->spa_suspend_zio_root = NULL;
1351 mutex_exit(&spa->spa_suspend_lock);
1352
1353 if (pio == NULL)
1354 return (0);
1355
1356 zio_reexecute(pio);
1357 return (zio_wait(pio));
1358}
1359
1360void
1361zio_resume_wait(spa_t *spa)
1362{
1363 mutex_enter(&spa->spa_suspend_lock);
1364 while (spa_suspended(spa))
1365 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1366 mutex_exit(&spa->spa_suspend_lock);
1367}
1368
1369/*
1370 * ==========================================================================
1371 * Gang blocks.
1372 *
1373 * A gang block is a collection of small blocks that looks to the DMU
1374 * like one large block. When zio_dva_allocate() cannot find a block
1375 * of the requested size, due to either severe fragmentation or the pool
1376 * being nearly full, it calls zio_write_gang_block() to construct the
1377 * block from smaller fragments.
1378 *
1379 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1380 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1381 * an indirect block: it's an array of block pointers. It consumes
1382 * only one sector and hence is allocatable regardless of fragmentation.
1383 * The gang header's bps point to its gang members, which hold the data.
1384 *
1385 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1386 * as the verifier to ensure uniqueness of the SHA256 checksum.
1387 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1388 * not the gang header. This ensures that data block signatures (needed for
1389 * deduplication) are independent of how the block is physically stored.
1390 *
1391 * Gang blocks can be nested: a gang member may itself be a gang block.
1392 * Thus every gang block is a tree in which root and all interior nodes are
1393 * gang headers, and the leaves are normal blocks that contain user data.
1394 * The root of the gang tree is called the gang leader.
1395 *
1396 * To perform any operation (read, rewrite, free, claim) on a gang block,
1397 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1398 * in the io_gang_tree field of the original logical i/o by recursively
1399 * reading the gang leader and all gang headers below it. This yields
1400 * an in-core tree containing the contents of every gang header and the
1401 * bps for every constituent of the gang block.
1402 *
1403 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1404 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1405 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1406 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1407 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1408 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1409 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1410 * of the gang header plus zio_checksum_compute() of the data to update the
1411 * gang header's blk_cksum as described above.
1412 *
1413 * The two-phase assemble/issue model solves the problem of partial failure --
1414 * what if you'd freed part of a gang block but then couldn't read the
1415 * gang header for another part? Assembling the entire gang tree first
1416 * ensures that all the necessary gang header I/O has succeeded before
1417 * starting the actual work of free, claim, or write. Once the gang tree
1418 * is assembled, free and claim are in-memory operations that cannot fail.
1419 *
1420 * In the event that a gang write fails, zio_dva_unallocate() walks the
1421 * gang tree to immediately free (i.e. insert back into the space map)
1422 * everything we've allocated. This ensures that we don't get ENOSPC
1423 * errors during repeated suspend/resume cycles due to a flaky device.
1424 *
1425 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1426 * the gang tree, we won't modify the block, so we can safely defer the free
1427 * (knowing that the block is still intact). If we *can* assemble the gang
1428 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1429 * each constituent bp and we can allocate a new block on the next sync pass.
1430 *
1431 * In all cases, the gang tree allows complete recovery from partial failure.
1432 * ==========================================================================
1433 */
1434
1435static zio_t *
1436zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1437{
1438 if (gn != NULL)
1439 return (pio);
1440
1441 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1442 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1443 &pio->io_bookmark));
1444}
1445
1446zio_t *
1447zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1448{
1449 zio_t *zio;
1450
1451 if (gn != NULL) {
1452 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1453 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1454 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1455 /*
1456 * As we rewrite each gang header, the pipeline will compute
1457 * a new gang block header checksum for it; but no one will
1458 * compute a new data checksum, so we do that here. The one
1459 * exception is the gang leader: the pipeline already computed
1460 * its data checksum because that stage precedes gang assembly.
1461 * (Presently, nothing actually uses interior data checksums;
1462 * this is just good hygiene.)
1463 */
1464 if (gn != pio->io_gang_leader->io_gang_tree) {
1465 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1466 data, BP_GET_PSIZE(bp));
1467 }
1468 /*
1469 * If we are here to damage data for testing purposes,
1470 * leave the GBH alone so that we can detect the damage.
1471 */
1472 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1473 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1474 } else {
1475 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1476 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1477 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1478 }
1479
1480 return (zio);
1481}
1482
1483/* ARGSUSED */
1484zio_t *
1485zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1486{
1487 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1488 ZIO_GANG_CHILD_FLAGS(pio)));
1489}
1490
1491/* ARGSUSED */
1492zio_t *
1493zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1494{
1495 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1496 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1497}
1498
1499static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1500 NULL,
1501 zio_read_gang,
1502 zio_rewrite_gang,
1503 zio_free_gang,
1504 zio_claim_gang,
1505 NULL
1506};
1507
1508static void zio_gang_tree_assemble_done(zio_t *zio);
1509
1510static zio_gang_node_t *
1511zio_gang_node_alloc(zio_gang_node_t **gnpp)
1512{
1513 zio_gang_node_t *gn;
1514
1515 ASSERT(*gnpp == NULL);
1516
1517 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1518 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1519 *gnpp = gn;
1520
1521 return (gn);
1522}
1523
1524static void
1525zio_gang_node_free(zio_gang_node_t **gnpp)
1526{
1527 zio_gang_node_t *gn = *gnpp;
1528
1529 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1530 ASSERT(gn->gn_child[g] == NULL);
1531
1532 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1533 kmem_free(gn, sizeof (*gn));
1534 *gnpp = NULL;
1535}
1536
1537static void
1538zio_gang_tree_free(zio_gang_node_t **gnpp)
1539{
1540 zio_gang_node_t *gn = *gnpp;
1541
1542 if (gn == NULL)
1543 return;
1544
1545 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1546 zio_gang_tree_free(&gn->gn_child[g]);
1547
1548 zio_gang_node_free(gnpp);
1549}
1550
1551static void
1552zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1553{
1554 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1555
1556 ASSERT(gio->io_gang_leader == gio);
1557 ASSERT(BP_IS_GANG(bp));
1558
1559 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1560 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1561 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1562}
1563
1564static void
1565zio_gang_tree_assemble_done(zio_t *zio)
1566{
1567 zio_t *gio = zio->io_gang_leader;
1568 zio_gang_node_t *gn = zio->io_private;
1569 blkptr_t *bp = zio->io_bp;
1570
1571 ASSERT(gio == zio_unique_parent(zio));
1572 ASSERT(zio->io_child_count == 0);
1573
1574 if (zio->io_error)
1575 return;
1576
1577 if (BP_SHOULD_BYTESWAP(bp))
1578 byteswap_uint64_array(zio->io_data, zio->io_size);
1579
1580 ASSERT(zio->io_data == gn->gn_gbh);
1581 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1582 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1583
1584 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1585 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1586 if (!BP_IS_GANG(gbp))
1587 continue;
1588 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1589 }
1590}
1591
1592static void
1593zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1594{
1595 zio_t *gio = pio->io_gang_leader;
1596 zio_t *zio;
1597
1598 ASSERT(BP_IS_GANG(bp) == !!gn);
1599 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1600 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1601
1602 /*
1603 * If you're a gang header, your data is in gn->gn_gbh.
1604 * If you're a gang member, your data is in 'data' and gn == NULL.
1605 */
1606 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1607
1608 if (gn != NULL) {
1609 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1610
1611 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1612 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1613 if (BP_IS_HOLE(gbp))
1614 continue;
1615 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1616 data = (char *)data + BP_GET_PSIZE(gbp);
1617 }
1618 }
1619
1620 if (gn == gio->io_gang_tree)
1621 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1622
1623 if (zio != pio)
1624 zio_nowait(zio);
1625}
1626
1627static int
1628zio_gang_assemble(zio_t *zio)
1629{
1630 blkptr_t *bp = zio->io_bp;
1631
1632 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1633 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1634
1635 zio->io_gang_leader = zio;
1636
1637 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1638
1639 return (ZIO_PIPELINE_CONTINUE);
1640}
1641
1642static int
1643zio_gang_issue(zio_t *zio)
1644{
1645 blkptr_t *bp = zio->io_bp;
1646
1647 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1648 return (ZIO_PIPELINE_STOP);
1649
1650 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1651 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1652
1653 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1654 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1655 else
1656 zio_gang_tree_free(&zio->io_gang_tree);
1657
1658 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1659
1660 return (ZIO_PIPELINE_CONTINUE);
1661}
1662
1663static void
1664zio_write_gang_member_ready(zio_t *zio)
1665{
1666 zio_t *pio = zio_unique_parent(zio);
1667 zio_t *gio = zio->io_gang_leader;
1668 dva_t *cdva = zio->io_bp->blk_dva;
1669 dva_t *pdva = pio->io_bp->blk_dva;
1670 uint64_t asize;
1671
1672 if (BP_IS_HOLE(zio->io_bp))
1673 return;
1674
1675 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1676
1677 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1678 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1679 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1680 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1681 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1682
1683 mutex_enter(&pio->io_lock);
1684 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1685 ASSERT(DVA_GET_GANG(&pdva[d]));
1686 asize = DVA_GET_ASIZE(&pdva[d]);
1687 asize += DVA_GET_ASIZE(&cdva[d]);
1688 DVA_SET_ASIZE(&pdva[d], asize);
1689 }
1690 mutex_exit(&pio->io_lock);
1691}
1692
1693static int
1694zio_write_gang_block(zio_t *pio)
1695{
1696 spa_t *spa = pio->io_spa;
1697 blkptr_t *bp = pio->io_bp;
1698 zio_t *gio = pio->io_gang_leader;
1699 zio_t *zio;
1700 zio_gang_node_t *gn, **gnpp;
1701 zio_gbh_phys_t *gbh;
1702 uint64_t txg = pio->io_txg;
1703 uint64_t resid = pio->io_size;
1704 uint64_t lsize;
1705 int copies = gio->io_prop.zp_copies;
1706 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1707 zio_prop_t zp;
1708 int error;
1709
1710 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1711 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1712 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1713 if (error) {
1714 pio->io_error = error;
1715 return (ZIO_PIPELINE_CONTINUE);
1716 }
1717
1718 if (pio == gio) {
1719 gnpp = &gio->io_gang_tree;
1720 } else {
1721 gnpp = pio->io_private;
1722 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1723 }
1724
1725 gn = zio_gang_node_alloc(gnpp);
1726 gbh = gn->gn_gbh;
1727 bzero(gbh, SPA_GANGBLOCKSIZE);
1728
1729 /*
1730 * Create the gang header.
1731 */
1732 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1733 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1734
1735 /*
1736 * Create and nowait the gang children.
1737 */
1738 for (int g = 0; resid != 0; resid -= lsize, g++) {
1739 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1740 SPA_MINBLOCKSIZE);
1741 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1742
1743 zp.zp_checksum = gio->io_prop.zp_checksum;
1744 zp.zp_compress = ZIO_COMPRESS_OFF;
1745 zp.zp_type = DMU_OT_NONE;
1746 zp.zp_level = 0;
1747 zp.zp_copies = gio->io_prop.zp_copies;
1748 zp.zp_dedup = 0;
1749 zp.zp_dedup_verify = 0;
1750
1751 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1752 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1753 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1754 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1755 &pio->io_bookmark));
1756 }
1757
1758 /*
1759 * Set pio's pipeline to just wait for zio to finish.
1760 */
1761 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1762
1763 zio_nowait(zio);
1764
1765 return (ZIO_PIPELINE_CONTINUE);
1766}
1767
1768/*
1769 * ==========================================================================
1770 * Dedup
1771 * ==========================================================================
1772 */
1773static void
1774zio_ddt_child_read_done(zio_t *zio)
1775{
1776 blkptr_t *bp = zio->io_bp;
1777 ddt_entry_t *dde = zio->io_private;
1778 ddt_phys_t *ddp;
1779 zio_t *pio = zio_unique_parent(zio);
1780
1781 mutex_enter(&pio->io_lock);
1782 ddp = ddt_phys_select(dde, bp);
1783 if (zio->io_error == 0)
1784 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1785 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1786 dde->dde_repair_data = zio->io_data;
1787 else
1788 zio_buf_free(zio->io_data, zio->io_size);
1789 mutex_exit(&pio->io_lock);
1790}
1791
1792static int
1793zio_ddt_read_start(zio_t *zio)
1794{
1795 blkptr_t *bp = zio->io_bp;
1796
1797 ASSERT(BP_GET_DEDUP(bp));
1798 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1799 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1800
1801 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1802 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1803 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1804 ddt_phys_t *ddp = dde->dde_phys;
1805 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1806 blkptr_t blk;
1807
1808 ASSERT(zio->io_vsd == NULL);
1809 zio->io_vsd = dde;
1810
1811 if (ddp_self == NULL)
1812 return (ZIO_PIPELINE_CONTINUE);
1813
1814 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1815 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1816 continue;
1817 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1818 &blk);
1819 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1820 zio_buf_alloc(zio->io_size), zio->io_size,
1821 zio_ddt_child_read_done, dde, zio->io_priority,
1822 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1823 &zio->io_bookmark));
1824 }
1825 return (ZIO_PIPELINE_CONTINUE);
1826 }
1827
1828 zio_nowait(zio_read(zio, zio->io_spa, bp,
1829 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1830 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1831
1832 return (ZIO_PIPELINE_CONTINUE);
1833}
1834
1835static int
1836zio_ddt_read_done(zio_t *zio)
1837{
1838 blkptr_t *bp = zio->io_bp;
1839
1840 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1841 return (ZIO_PIPELINE_STOP);
1842
1843 ASSERT(BP_GET_DEDUP(bp));
1844 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1845 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1846
1847 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1848 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1849 ddt_entry_t *dde = zio->io_vsd;
1850 if (ddt == NULL) {
1851 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1852 return (ZIO_PIPELINE_CONTINUE);
1853 }
1854 if (dde == NULL) {
1855 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1856 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1857 return (ZIO_PIPELINE_STOP);
1858 }
1859 if (dde->dde_repair_data != NULL) {
1860 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1861 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1862 }
1863 ddt_repair_done(ddt, dde);
1864 zio->io_vsd = NULL;
1865 }
1866
1867 ASSERT(zio->io_vsd == NULL);
1868
1869 return (ZIO_PIPELINE_CONTINUE);
1870}
1871
1872static boolean_t
1873zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1874{
1875 spa_t *spa = zio->io_spa;
1876
1877 /*
1878 * Note: we compare the original data, not the transformed data,
1879 * because when zio->io_bp is an override bp, we will not have
1880 * pushed the I/O transforms. That's an important optimization
1881 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1882 */
1883 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1884 zio_t *lio = dde->dde_lead_zio[p];
1885
1886 if (lio != NULL) {
1887 return (lio->io_orig_size != zio->io_orig_size ||
1888 bcmp(zio->io_orig_data, lio->io_orig_data,
1889 zio->io_orig_size) != 0);
1890 }
1891 }
1892
1893 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1894 ddt_phys_t *ddp = &dde->dde_phys[p];
1895
1896 if (ddp->ddp_phys_birth != 0) {
1897 arc_buf_t *abuf = NULL;
1898 uint32_t aflags = ARC_WAIT;
1899 blkptr_t blk = *zio->io_bp;
1900 int error;
1901
1902 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1903
1904 ddt_exit(ddt);
1905
1906 error = arc_read_nolock(NULL, spa, &blk,
1907 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1908 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1909 &aflags, &zio->io_bookmark);
1910
1911 if (error == 0) {
1912 if (arc_buf_size(abuf) != zio->io_orig_size ||
1913 bcmp(abuf->b_data, zio->io_orig_data,
1914 zio->io_orig_size) != 0)
1915 error = EEXIST;
1916 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1917 }
1918
1919 ddt_enter(ddt);
1920 return (error != 0);
1921 }
1922 }
1923
1924 return (B_FALSE);
1925}
1926
1927static void
1928zio_ddt_child_write_ready(zio_t *zio)
1929{
1930 int p = zio->io_prop.zp_copies;
1931 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1932 ddt_entry_t *dde = zio->io_private;
1933 ddt_phys_t *ddp = &dde->dde_phys[p];
1934 zio_t *pio;
1935
1936 if (zio->io_error)
1937 return;
1938
1939 ddt_enter(ddt);
1940
1941 ASSERT(dde->dde_lead_zio[p] == zio);
1942
1943 ddt_phys_fill(ddp, zio->io_bp);
1944
1945 while ((pio = zio_walk_parents(zio)) != NULL)
1946 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1947
1948 ddt_exit(ddt);
1949}
1950
1951static void
1952zio_ddt_child_write_done(zio_t *zio)
1953{
1954 int p = zio->io_prop.zp_copies;
1955 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1956 ddt_entry_t *dde = zio->io_private;
1957 ddt_phys_t *ddp = &dde->dde_phys[p];
1958
1959 ddt_enter(ddt);
1960
1961 ASSERT(ddp->ddp_refcnt == 0);
1962 ASSERT(dde->dde_lead_zio[p] == zio);
1963 dde->dde_lead_zio[p] = NULL;
1964
1965 if (zio->io_error == 0) {
1966 while (zio_walk_parents(zio) != NULL)
1967 ddt_phys_addref(ddp);
1968 } else {
1969 ddt_phys_clear(ddp);
1970 }
1971
1972 ddt_exit(ddt);
1973}
1974
1975static void
1976zio_ddt_ditto_write_done(zio_t *zio)
1977{
1978 int p = DDT_PHYS_DITTO;
1979 zio_prop_t *zp = &zio->io_prop;
1980 blkptr_t *bp = zio->io_bp;
1981 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1982 ddt_entry_t *dde = zio->io_private;
1983 ddt_phys_t *ddp = &dde->dde_phys[p];
1984 ddt_key_t *ddk = &dde->dde_key;
1985
1986 ddt_enter(ddt);
1987
1988 ASSERT(ddp->ddp_refcnt == 0);
1989 ASSERT(dde->dde_lead_zio[p] == zio);
1990 dde->dde_lead_zio[p] = NULL;
1991
1992 if (zio->io_error == 0) {
1993 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1994 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1995 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1996 if (ddp->ddp_phys_birth != 0)
1997 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1998 ddt_phys_fill(ddp, bp);
1999 }
2000
2001 ddt_exit(ddt);
2002}
2003
2004static int
2005zio_ddt_write(zio_t *zio)
2006{
2007 spa_t *spa = zio->io_spa;
2008 blkptr_t *bp = zio->io_bp;
2009 uint64_t txg = zio->io_txg;
2010 zio_prop_t *zp = &zio->io_prop;
2011 int p = zp->zp_copies;
2012 int ditto_copies;
2013 zio_t *cio = NULL;
2014 zio_t *dio = NULL;
2015 ddt_t *ddt = ddt_select(spa, bp);
2016 ddt_entry_t *dde;
2017 ddt_phys_t *ddp;
2018
2019 ASSERT(BP_GET_DEDUP(bp));
2020 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2021 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2022
2023 ddt_enter(ddt);
2024 dde = ddt_lookup(ddt, bp, B_TRUE);
2025 ddp = &dde->dde_phys[p];
2026
2027 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2028 /*
2029 * If we're using a weak checksum, upgrade to a strong checksum
2030 * and try again. If we're already using a strong checksum,
2031 * we can't resolve it, so just convert to an ordinary write.
2032 * (And automatically e-mail a paper to Nature?)
2033 */
2034 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2035 zp->zp_checksum = spa_dedup_checksum(spa);
2036 zio_pop_transforms(zio);
2037 zio->io_stage = ZIO_STAGE_OPEN;
2038 BP_ZERO(bp);
2039 } else {
2040 zp->zp_dedup = 0;
2041 }
2042 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2043 ddt_exit(ddt);
2044 return (ZIO_PIPELINE_CONTINUE);
2045 }
2046
2047 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2048 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2049
2050 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2051 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2052 zio_prop_t czp = *zp;
2053
2054 czp.zp_copies = ditto_copies;
2055
2056 /*
2057 * If we arrived here with an override bp, we won't have run
2058 * the transform stack, so we won't have the data we need to
2059 * generate a child i/o. So, toss the override bp and restart.
2060 * This is safe, because using the override bp is just an
2061 * optimization; and it's rare, so the cost doesn't matter.
2062 */
2063 if (zio->io_bp_override) {
2064 zio_pop_transforms(zio);
2065 zio->io_stage = ZIO_STAGE_OPEN;
2066 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2067 zio->io_bp_override = NULL;
2068 BP_ZERO(bp);
2069 ddt_exit(ddt);
2070 return (ZIO_PIPELINE_CONTINUE);
2071 }
2072
2073 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2074 zio->io_orig_size, &czp, NULL,
2075 zio_ddt_ditto_write_done, dde, zio->io_priority,
2076 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2077
2078 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2079 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2080 }
2081
2082 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2083 if (ddp->ddp_phys_birth != 0)
2084 ddt_bp_fill(ddp, bp, txg);
2085 if (dde->dde_lead_zio[p] != NULL)
2086 zio_add_child(zio, dde->dde_lead_zio[p]);
2087 else
2088 ddt_phys_addref(ddp);
2089 } else if (zio->io_bp_override) {
2090 ASSERT(bp->blk_birth == txg);
2091 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2092 ddt_phys_fill(ddp, bp);
2093 ddt_phys_addref(ddp);
2094 } else {
2095 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2096 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2097 zio_ddt_child_write_done, dde, zio->io_priority,
2098 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2099
2100 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2101 dde->dde_lead_zio[p] = cio;
2102 }
2103
2104 ddt_exit(ddt);
2105
2106 if (cio)
2107 zio_nowait(cio);
2108 if (dio)
2109 zio_nowait(dio);
2110
2111 return (ZIO_PIPELINE_CONTINUE);
2112}
2113
2114ddt_entry_t *freedde; /* for debugging */
2115
2116static int
2117zio_ddt_free(zio_t *zio)
2118{
2119 spa_t *spa = zio->io_spa;
2120 blkptr_t *bp = zio->io_bp;
2121 ddt_t *ddt = ddt_select(spa, bp);
2122 ddt_entry_t *dde;
2123 ddt_phys_t *ddp;
2124
2125 ASSERT(BP_GET_DEDUP(bp));
2126 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2127
2128 ddt_enter(ddt);
2129 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2130 ddp = ddt_phys_select(dde, bp);
2131 ddt_phys_decref(ddp);
2132 ddt_exit(ddt);
2133
2134 return (ZIO_PIPELINE_CONTINUE);
2135}
2136
2137/*
2138 * ==========================================================================
2139 * Allocate and free blocks
2140 * ==========================================================================
2141 */
2142static int
2143zio_dva_allocate(zio_t *zio)
2144{
2145 spa_t *spa = zio->io_spa;
2146 metaslab_class_t *mc = spa_normal_class(spa);
2147 blkptr_t *bp = zio->io_bp;
2148 int error;
2149 int flags = 0;
2150
2151 if (zio->io_gang_leader == NULL) {
2152 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2153 zio->io_gang_leader = zio;
2154 }
2155
2156 ASSERT(BP_IS_HOLE(bp));
2157 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2158 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2159 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2160 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2161
2162 /*
2163 * The dump device does not support gang blocks so allocation on
2164 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2165 * the "fast" gang feature.
2166 */
2167 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2168 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2169 METASLAB_GANG_CHILD : 0;
2170 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2171 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2172
2173 if (error) {
2174 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2175 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2176 error);
2177 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2178 return (zio_write_gang_block(zio));
2179 zio->io_error = error;
2180 }
2181
2182 return (ZIO_PIPELINE_CONTINUE);
2183}
2184
2185static int
2186zio_dva_free(zio_t *zio)
2187{
2188 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2189
2190 return (ZIO_PIPELINE_CONTINUE);
2191}
2192
2193static int
2194zio_dva_claim(zio_t *zio)
2195{
2196 int error;
2197
2198 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2199 if (error)
2200 zio->io_error = error;
2201
2202 return (ZIO_PIPELINE_CONTINUE);
2203}
2204
2205/*
2206 * Undo an allocation. This is used by zio_done() when an I/O fails
2207 * and we want to give back the block we just allocated.
2208 * This handles both normal blocks and gang blocks.
2209 */
2210static void
2211zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2212{
2213 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2214 ASSERT(zio->io_bp_override == NULL);
2215
2216 if (!BP_IS_HOLE(bp))
2217 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2218
2219 if (gn != NULL) {
2220 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2221 zio_dva_unallocate(zio, gn->gn_child[g],
2222 &gn->gn_gbh->zg_blkptr[g]);
2223 }
2224 }
2225}
2226
2227/*
2228 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2229 */
2230int
2231zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2232 uint64_t size, boolean_t use_slog)
2233{
2234 int error = 1;
2235
2236 ASSERT(txg > spa_syncing_txg(spa));
2237
2238 /*
2239 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2240 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2241 * when allocating them.
2242 */
2243 if (use_slog) {
2244 error = metaslab_alloc(spa, spa_log_class(spa), size,
2245 new_bp, 1, txg, old_bp,
2246 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2247 }
2248
2249 if (error) {
2250 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2251 new_bp, 1, txg, old_bp,
2252 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2253 }
2254
2255 if (error == 0) {
2256 BP_SET_LSIZE(new_bp, size);
2257 BP_SET_PSIZE(new_bp, size);
2258 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2259 BP_SET_CHECKSUM(new_bp,
2260 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2261 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2262 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2263 BP_SET_LEVEL(new_bp, 0);
2264 BP_SET_DEDUP(new_bp, 0);
2265 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2266 }
2267
2268 return (error);
2269}
2270
2271/*
2272 * Free an intent log block.
2273 */
2274void
2275zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2276{
2277 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2278 ASSERT(!BP_IS_GANG(bp));
2279
2280 zio_free(spa, txg, bp);
2281}
2282
2283/*
2284 * ==========================================================================
2285 * Read and write to physical devices
2286 * ==========================================================================
2287 */
2288static int
2289zio_vdev_io_start(zio_t *zio)
2290{
2291 vdev_t *vd = zio->io_vd;
2292 uint64_t align;
2293 spa_t *spa = zio->io_spa;
2294
2295 ASSERT(zio->io_error == 0);
2296 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2297
2298 if (vd == NULL) {
2299 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2300 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2301
2302 /*
2303 * The mirror_ops handle multiple DVAs in a single BP.
2304 */
2305 return (vdev_mirror_ops.vdev_op_io_start(zio));
2306 }
2307
2308 /*
2309 * We keep track of time-sensitive I/Os so that the scan thread
2310 * can quickly react to certain workloads. In particular, we care
2311 * about non-scrubbing, top-level reads and writes with the following
2312 * characteristics:
2313 * - synchronous writes of user data to non-slog devices
2314 * - any reads of user data
2315 * When these conditions are met, adjust the timestamp of spa_last_io
2316 * which allows the scan thread to adjust its workload accordingly.
2317 */
2318 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2319 vd == vd->vdev_top && !vd->vdev_islog &&
2320 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2321 zio->io_txg != spa_syncing_txg(spa)) {
2322 uint64_t old = spa->spa_last_io;
2323 uint64_t new = ddi_get_lbolt64();
2324 if (old != new)
2325 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2326 }
2327
2328 align = 1ULL << vd->vdev_top->vdev_ashift;
2329
2330 if (P2PHASE(zio->io_size, align) != 0) {
2331 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2332 char *abuf = zio_buf_alloc(asize);
2333 ASSERT(vd == vd->vdev_top);
2334 if (zio->io_type == ZIO_TYPE_WRITE) {
2335 bcopy(zio->io_data, abuf, zio->io_size);
2336 bzero(abuf + zio->io_size, asize - zio->io_size);
2337 }
2338 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2339 }
2340
2341 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2342 ASSERT(P2PHASE(zio->io_size, align) == 0);
2343 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2344
2345 /*
2346 * If this is a repair I/O, and there's no self-healing involved --
2347 * that is, we're just resilvering what we expect to resilver --
2348 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2349 * This prevents spurious resilvering with nested replication.
2350 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2351 * A is out of date, we'll read from C+D, then use the data to
2352 * resilver A+B -- but we don't actually want to resilver B, just A.
2353 * The top-level mirror has no way to know this, so instead we just
2354 * discard unnecessary repairs as we work our way down the vdev tree.
2355 * The same logic applies to any form of nested replication:
2356 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2357 */
2358 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2359 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2360 zio->io_txg != 0 && /* not a delegated i/o */
2361 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2362 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2363 zio_vdev_io_bypass(zio);
2364 return (ZIO_PIPELINE_CONTINUE);
2365 }
2366
2367 if (vd->vdev_ops->vdev_op_leaf &&
2368 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2369
2370 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2371 return (ZIO_PIPELINE_CONTINUE);
2372
2373 if ((zio = vdev_queue_io(zio)) == NULL)
2374 return (ZIO_PIPELINE_STOP);
2375
2376 if (!vdev_accessible(vd, zio)) {
2377 zio->io_error = ENXIO;
2378 zio_interrupt(zio);
2379 return (ZIO_PIPELINE_STOP);
2380 }
2381 }
2382
2383 return (vd->vdev_ops->vdev_op_io_start(zio));
2384}
2385
2386static int
2387zio_vdev_io_done(zio_t *zio)
2388{
2389 vdev_t *vd = zio->io_vd;
2390 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2391 boolean_t unexpected_error = B_FALSE;
2392
2393 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2394 return (ZIO_PIPELINE_STOP);
2395
2396 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2397
2398 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2399
2400 vdev_queue_io_done(zio);
2401
2402 if (zio->io_type == ZIO_TYPE_WRITE)
2403 vdev_cache_write(zio);
2404
2405 if (zio_injection_enabled && zio->io_error == 0)
2406 zio->io_error = zio_handle_device_injection(vd,
2407 zio, EIO);
2408
2409 if (zio_injection_enabled && zio->io_error == 0)
2410 zio->io_error = zio_handle_label_injection(zio, EIO);
2411
2412 if (zio->io_error) {
2413 if (!vdev_accessible(vd, zio)) {
2414 zio->io_error = ENXIO;
2415 } else {
2416 unexpected_error = B_TRUE;
2417 }
2418 }
2419 }
2420
2421 ops->vdev_op_io_done(zio);
2422
2423 if (unexpected_error)
2424 VERIFY(vdev_probe(vd, zio) == NULL);
2425
2426 return (ZIO_PIPELINE_CONTINUE);
2427}
2428
2429/*
2430 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2431 * disk, and use that to finish the checksum ereport later.
2432 */
2433static void
2434zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2435 const void *good_buf)
2436{
2437 /* no processing needed */
2438 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2439}
2440
2441/*ARGSUSED*/
2442void
2443zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2444{
2445 void *buf = zio_buf_alloc(zio->io_size);
2446
2447 bcopy(zio->io_data, buf, zio->io_size);
2448
2449 zcr->zcr_cbinfo = zio->io_size;
2450 zcr->zcr_cbdata = buf;
2451 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2452 zcr->zcr_free = zio_buf_free;
2453}
2454
2455static int
2456zio_vdev_io_assess(zio_t *zio)
2457{
2458 vdev_t *vd = zio->io_vd;
2459
2460 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2461 return (ZIO_PIPELINE_STOP);
2462
2463 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2464 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2465
2466 if (zio->io_vsd != NULL) {
2467 zio->io_vsd_ops->vsd_free(zio);
2468 zio->io_vsd = NULL;
2469 }
2470
2471 if (zio_injection_enabled && zio->io_error == 0)
2472 zio->io_error = zio_handle_fault_injection(zio, EIO);
2473
2474 /*
2475 * If the I/O failed, determine whether we should attempt to retry it.
2476 *
2477 * On retry, we cut in line in the issue queue, since we don't want
2478 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2479 */
2480 if (zio->io_error && vd == NULL &&
2481 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2482 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2483 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2484 zio->io_error = 0;
2485 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2486 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2487 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2488 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2489 zio_requeue_io_start_cut_in_line);
2490 return (ZIO_PIPELINE_STOP);
2491 }
2492
2493 /*
2494 * If we got an error on a leaf device, convert it to ENXIO
2495 * if the device is not accessible at all.
2496 */
2497 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2498 !vdev_accessible(vd, zio))
2499 zio->io_error = ENXIO;
2500
2501 /*
2502 * If we can't write to an interior vdev (mirror or RAID-Z),
2503 * set vdev_cant_write so that we stop trying to allocate from it.
2504 */
2505 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2506 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2507 vd->vdev_cant_write = B_TRUE;
2508
2509 if (zio->io_error)
2510 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2511
2512 return (ZIO_PIPELINE_CONTINUE);
2513}
2514
2515void
2516zio_vdev_io_reissue(zio_t *zio)
2517{
2518 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2519 ASSERT(zio->io_error == 0);
2520
2521 zio->io_stage >>= 1;
2522}
2523
2524void
2525zio_vdev_io_redone(zio_t *zio)
2526{
2527 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2528
2529 zio->io_stage >>= 1;
2530}
2531
2532void
2533zio_vdev_io_bypass(zio_t *zio)
2534{
2535 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2536 ASSERT(zio->io_error == 0);
2537
2538 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2539 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2540}
2541
2542/*
2543 * ==========================================================================
2544 * Generate and verify checksums
2545 * ==========================================================================
2546 */
2547static int
2548zio_checksum_generate(zio_t *zio)
2549{
2550 blkptr_t *bp = zio->io_bp;
2551 enum zio_checksum checksum;
2552
2553 if (bp == NULL) {
2554 /*
2555 * This is zio_write_phys().
2556 * We're either generating a label checksum, or none at all.
2557 */
2558 checksum = zio->io_prop.zp_checksum;
2559
2560 if (checksum == ZIO_CHECKSUM_OFF)
2561 return (ZIO_PIPELINE_CONTINUE);
2562
2563 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2564 } else {
2565 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2566 ASSERT(!IO_IS_ALLOCATING(zio));
2567 checksum = ZIO_CHECKSUM_GANG_HEADER;
2568 } else {
2569 checksum = BP_GET_CHECKSUM(bp);
2570 }
2571 }
2572
2573 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2574
2575 return (ZIO_PIPELINE_CONTINUE);
2576}
2577
2578static int
2579zio_checksum_verify(zio_t *zio)
2580{
2581 zio_bad_cksum_t info;
2582 blkptr_t *bp = zio->io_bp;
2583 int error;
2584
2585 ASSERT(zio->io_vd != NULL);
2586
2587 if (bp == NULL) {
2588 /*
2589 * This is zio_read_phys().
2590 * We're either verifying a label checksum, or nothing at all.
2591 */
2592 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2593 return (ZIO_PIPELINE_CONTINUE);
2594
2595 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2596 }
2597
2598 if ((error = zio_checksum_error(zio, &info)) != 0) {
2599 zio->io_error = error;
2600 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2601 zfs_ereport_start_checksum(zio->io_spa,
2602 zio->io_vd, zio, zio->io_offset,
2603 zio->io_size, NULL, &info);
2604 }
2605 }
2606
2607 return (ZIO_PIPELINE_CONTINUE);
2608}
2609
2610/*
2611 * Called by RAID-Z to ensure we don't compute the checksum twice.
2612 */
2613void
2614zio_checksum_verified(zio_t *zio)
2615{
2616 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2617}
2618
2619/*
2620 * ==========================================================================
2621 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2622 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2623 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2624 * indicate errors that are specific to one I/O, and most likely permanent.
2625 * Any other error is presumed to be worse because we weren't expecting it.
2626 * ==========================================================================
2627 */
2628int
2629zio_worst_error(int e1, int e2)
2630{
2631 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2632 int r1, r2;
2633
2634 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2635 if (e1 == zio_error_rank[r1])
2636 break;
2637
2638 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2639 if (e2 == zio_error_rank[r2])
2640 break;
2641
2642 return (r1 > r2 ? e1 : e2);
2643}
2644
2645/*
2646 * ==========================================================================
2647 * I/O completion
2648 * ==========================================================================
2649 */
2650static int
2651zio_ready(zio_t *zio)
2652{
2653 blkptr_t *bp = zio->io_bp;
2654 zio_t *pio, *pio_next;
2655
2656 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2657 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2658 return (ZIO_PIPELINE_STOP);
2659
2660 if (zio->io_ready) {
2661 ASSERT(IO_IS_ALLOCATING(zio));
2662 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2663 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2664
2665 zio->io_ready(zio);
2666 }
2667
2668 if (bp != NULL && bp != &zio->io_bp_copy)
2669 zio->io_bp_copy = *bp;
2670
2671 if (zio->io_error)
2672 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2673
2674 mutex_enter(&zio->io_lock);
2675 zio->io_state[ZIO_WAIT_READY] = 1;
2676 pio = zio_walk_parents(zio);
2677 mutex_exit(&zio->io_lock);
2678
2679 /*
2680 * As we notify zio's parents, new parents could be added.
2681 * New parents go to the head of zio's io_parent_list, however,
2682 * so we will (correctly) not notify them. The remainder of zio's
2683 * io_parent_list, from 'pio_next' onward, cannot change because
2684 * all parents must wait for us to be done before they can be done.
2685 */
2686 for (; pio != NULL; pio = pio_next) {
2687 pio_next = zio_walk_parents(zio);
2688 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2689 }
2690
2691 if (zio->io_flags & ZIO_FLAG_NODATA) {
2692 if (BP_IS_GANG(bp)) {
2693 zio->io_flags &= ~ZIO_FLAG_NODATA;
2694 } else {
2695 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2696 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2697 }
2698 }
2699
2700 if (zio_injection_enabled &&
2701 zio->io_spa->spa_syncing_txg == zio->io_txg)
2702 zio_handle_ignored_writes(zio);
2703
2704 return (ZIO_PIPELINE_CONTINUE);
2705}
2706
2707static int
2708zio_done(zio_t *zio)
2709{
2710 spa_t *spa = zio->io_spa;
2711 zio_t *lio = zio->io_logical;
2712 blkptr_t *bp = zio->io_bp;
2713 vdev_t *vd = zio->io_vd;
2714 uint64_t psize = zio->io_size;
2715 zio_t *pio, *pio_next;
2716
2717 /*
2718 * If our children haven't all completed,
2719 * wait for them and then repeat this pipeline stage.
2720 */
2721 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2722 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2723 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2724 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2725 return (ZIO_PIPELINE_STOP);
2726
2727 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2728 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2729 ASSERT(zio->io_children[c][w] == 0);
2730
2731 if (bp != NULL) {
2732 ASSERT(bp->blk_pad[0] == 0);
2733 ASSERT(bp->blk_pad[1] == 0);
2734 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2735 (bp == zio_unique_parent(zio)->io_bp));
2736 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2737 zio->io_bp_override == NULL &&
2738 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2739 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2740 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2741 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2742 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2743 }
2744 }
2745
2746 /*
2747 * If there were child vdev/gang/ddt errors, they apply to us now.
2748 */
2749 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2750 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2751 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2752
2753 /*
2754 * If the I/O on the transformed data was successful, generate any
2755 * checksum reports now while we still have the transformed data.
2756 */
2757 if (zio->io_error == 0) {
2758 while (zio->io_cksum_report != NULL) {
2759 zio_cksum_report_t *zcr = zio->io_cksum_report;
2760 uint64_t align = zcr->zcr_align;
2761 uint64_t asize = P2ROUNDUP(psize, align);
2762 char *abuf = zio->io_data;
2763
2764 if (asize != psize) {
2765 abuf = zio_buf_alloc(asize);
2766 bcopy(zio->io_data, abuf, psize);
2767 bzero(abuf + psize, asize - psize);
2768 }
2769
2770 zio->io_cksum_report = zcr->zcr_next;
2771 zcr->zcr_next = NULL;
2772 zcr->zcr_finish(zcr, abuf);
2773 zfs_ereport_free_checksum(zcr);
2774
2775 if (asize != psize)
2776 zio_buf_free(abuf, asize);
2777 }
2778 }
2779
2780 zio_pop_transforms(zio); /* note: may set zio->io_error */
2781
2782 vdev_stat_update(zio, psize);
2783
2784 if (zio->io_error) {
2785 /*
2786 * If this I/O is attached to a particular vdev,
2787 * generate an error message describing the I/O failure
2788 * at the block level. We ignore these errors if the
2789 * device is currently unavailable.
2790 */
2791 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2792 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2793
2794 if ((zio->io_error == EIO || !(zio->io_flags &
2795 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2796 zio == lio) {
2797 /*
2798 * For logical I/O requests, tell the SPA to log the
2799 * error and generate a logical data ereport.
2800 */
2801 spa_log_error(spa, zio);
2802 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2803 0, 0);
2804 }
2805 }
2806
2807 if (zio->io_error && zio == lio) {
2808 /*
2809 * Determine whether zio should be reexecuted. This will
2810 * propagate all the way to the root via zio_notify_parent().
2811 */
2812 ASSERT(vd == NULL && bp != NULL);
2813 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2814
2815 if (IO_IS_ALLOCATING(zio) &&
2816 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2817 if (zio->io_error != ENOSPC)
2818 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2819 else
2820 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2821 }
2822
2823 if ((zio->io_type == ZIO_TYPE_READ ||
2824 zio->io_type == ZIO_TYPE_FREE) &&
2825 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2826 zio->io_error == ENXIO &&
2827 spa_load_state(spa) == SPA_LOAD_NONE &&
2828 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2829 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2830
2831 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2832 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2833
2834 /*
2835 * Here is a possibly good place to attempt to do
2836 * either combinatorial reconstruction or error correction
2837 * based on checksums. It also might be a good place
2838 * to send out preliminary ereports before we suspend
2839 * processing.
2840 */
2841 }
2842
2843 /*
2844 * If there were logical child errors, they apply to us now.
2845 * We defer this until now to avoid conflating logical child
2846 * errors with errors that happened to the zio itself when
2847 * updating vdev stats and reporting FMA events above.
2848 */
2849 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2850
2851 if ((zio->io_error || zio->io_reexecute) &&
2852 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2853 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2854 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2855
2856 zio_gang_tree_free(&zio->io_gang_tree);
2857
2858 /*
2859 * Godfather I/Os should never suspend.
2860 */
2861 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2862 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2863 zio->io_reexecute = 0;
2864
2865 if (zio->io_reexecute) {
2866 /*
2867 * This is a logical I/O that wants to reexecute.
2868 *
2869 * Reexecute is top-down. When an i/o fails, if it's not
2870 * the root, it simply notifies its parent and sticks around.
2871 * The parent, seeing that it still has children in zio_done(),
2872 * does the same. This percolates all the way up to the root.
2873 * The root i/o will reexecute or suspend the entire tree.
2874 *
2875 * This approach ensures that zio_reexecute() honors
2876 * all the original i/o dependency relationships, e.g.
2877 * parents not executing until children are ready.
2878 */
2879 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2880
2881 zio->io_gang_leader = NULL;
2882
2883 mutex_enter(&zio->io_lock);
2884 zio->io_state[ZIO_WAIT_DONE] = 1;
2885 mutex_exit(&zio->io_lock);
2886
2887 /*
2888 * "The Godfather" I/O monitors its children but is
2889 * not a true parent to them. It will track them through
2890 * the pipeline but severs its ties whenever they get into
2891 * trouble (e.g. suspended). This allows "The Godfather"
2892 * I/O to return status without blocking.
2893 */
2894 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2895 zio_link_t *zl = zio->io_walk_link;
2896 pio_next = zio_walk_parents(zio);
2897
2898 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2899 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2900 zio_remove_child(pio, zio, zl);
2901 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2902 }
2903 }
2904
2905 if ((pio = zio_unique_parent(zio)) != NULL) {
2906 /*
2907 * We're not a root i/o, so there's nothing to do
2908 * but notify our parent. Don't propagate errors
2909 * upward since we haven't permanently failed yet.
2910 */
2911 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2912 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2913 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2914 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2915 /*
2916 * We'd fail again if we reexecuted now, so suspend
2917 * until conditions improve (e.g. device comes online).
2918 */
2919 zio_suspend(spa, zio);
2920 } else {
2921 /*
2922 * Reexecution is potentially a huge amount of work.
2923 * Hand it off to the otherwise-unused claim taskq.
2924 */
2925#ifdef _KERNEL
2926 (void) taskq_dispatch_safe(
2927 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2928 (task_func_t *)zio_reexecute, zio, TQ_SLEEP,
2929 &zio->io_task);
2930#else
2931 (void) taskq_dispatch(
2932 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2933 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2934#endif
2935 }
2936 return (ZIO_PIPELINE_STOP);
2937 }
2938
2939 ASSERT(zio->io_child_count == 0);
2940 ASSERT(zio->io_reexecute == 0);
2941 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2942
2943 /*
2944 * Report any checksum errors, since the I/O is complete.
2945 */
2946 while (zio->io_cksum_report != NULL) {
2947 zio_cksum_report_t *zcr = zio->io_cksum_report;
2948 zio->io_cksum_report = zcr->zcr_next;
2949 zcr->zcr_next = NULL;
2950 zcr->zcr_finish(zcr, NULL);
2951 zfs_ereport_free_checksum(zcr);
2952 }
2953
2954 /*
2955 * It is the responsibility of the done callback to ensure that this
2956 * particular zio is no longer discoverable for adoption, and as
2957 * such, cannot acquire any new parents.
2958 */
2959 if (zio->io_done)
2960 zio->io_done(zio);
2961
2962 mutex_enter(&zio->io_lock);
2963 zio->io_state[ZIO_WAIT_DONE] = 1;
2964 mutex_exit(&zio->io_lock);
2965
2966 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2967 zio_link_t *zl = zio->io_walk_link;
2968 pio_next = zio_walk_parents(zio);
2969 zio_remove_child(pio, zio, zl);
2970 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2971 }
2972
2973 if (zio->io_waiter != NULL) {
2974 mutex_enter(&zio->io_lock);
2975 zio->io_executor = NULL;
2976 cv_broadcast(&zio->io_cv);
2977 mutex_exit(&zio->io_lock);
2978 } else {
2979 zio_destroy(zio);
2980 }
2981
2982 return (ZIO_PIPELINE_STOP);
2983}
2984
2985/*
2986 * ==========================================================================
2987 * I/O pipeline definition
2988 * ==========================================================================
2989 */
2990static zio_pipe_stage_t *zio_pipeline[] = {
2991 NULL,
2992 zio_read_bp_init,
2993 zio_free_bp_init,
2994 zio_issue_async,
2995 zio_write_bp_init,
2996 zio_checksum_generate,
2997 zio_ddt_read_start,
2998 zio_ddt_read_done,
2999 zio_ddt_write,
3000 zio_ddt_free,
3001 zio_gang_assemble,
3002 zio_gang_issue,
3003 zio_dva_allocate,
3004 zio_dva_free,
3005 zio_dva_claim,
3006 zio_ready,
3007 zio_vdev_io_start,
3008 zio_vdev_io_done,
3009 zio_vdev_io_assess,
3010 zio_checksum_verify,
3011 zio_done
3012};
| 246}
247
248void
249zio_buf_free(void *buf, size_t size)
250{
251 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
252
253 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
254
255 if (zio_use_uma)
256 kmem_cache_free(zio_buf_cache[c], buf);
257 else
258 kmem_free(buf, size);
259}
260
261void
262zio_data_buf_free(void *buf, size_t size)
263{
264 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
265
266 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
267
268 if (zio_use_uma)
269 kmem_cache_free(zio_data_buf_cache[c], buf);
270 else
271 kmem_free(buf, size);
272}
273
274/*
275 * ==========================================================================
276 * Push and pop I/O transform buffers
277 * ==========================================================================
278 */
279static void
280zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
281 zio_transform_func_t *transform)
282{
283 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
284
285 zt->zt_orig_data = zio->io_data;
286 zt->zt_orig_size = zio->io_size;
287 zt->zt_bufsize = bufsize;
288 zt->zt_transform = transform;
289
290 zt->zt_next = zio->io_transform_stack;
291 zio->io_transform_stack = zt;
292
293 zio->io_data = data;
294 zio->io_size = size;
295}
296
297static void
298zio_pop_transforms(zio_t *zio)
299{
300 zio_transform_t *zt;
301
302 while ((zt = zio->io_transform_stack) != NULL) {
303 if (zt->zt_transform != NULL)
304 zt->zt_transform(zio,
305 zt->zt_orig_data, zt->zt_orig_size);
306
307 if (zt->zt_bufsize != 0)
308 zio_buf_free(zio->io_data, zt->zt_bufsize);
309
310 zio->io_data = zt->zt_orig_data;
311 zio->io_size = zt->zt_orig_size;
312 zio->io_transform_stack = zt->zt_next;
313
314 kmem_free(zt, sizeof (zio_transform_t));
315 }
316}
317
318/*
319 * ==========================================================================
320 * I/O transform callbacks for subblocks and decompression
321 * ==========================================================================
322 */
323static void
324zio_subblock(zio_t *zio, void *data, uint64_t size)
325{
326 ASSERT(zio->io_size > size);
327
328 if (zio->io_type == ZIO_TYPE_READ)
329 bcopy(zio->io_data, data, size);
330}
331
332static void
333zio_decompress(zio_t *zio, void *data, uint64_t size)
334{
335 if (zio->io_error == 0 &&
336 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
337 zio->io_data, data, zio->io_size, size) != 0)
338 zio->io_error = EIO;
339}
340
341/*
342 * ==========================================================================
343 * I/O parent/child relationships and pipeline interlocks
344 * ==========================================================================
345 */
346/*
347 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
348 * continue calling these functions until they return NULL.
349 * Otherwise, the next caller will pick up the list walk in
350 * some indeterminate state. (Otherwise every caller would
351 * have to pass in a cookie to keep the state represented by
352 * io_walk_link, which gets annoying.)
353 */
354zio_t *
355zio_walk_parents(zio_t *cio)
356{
357 zio_link_t *zl = cio->io_walk_link;
358 list_t *pl = &cio->io_parent_list;
359
360 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
361 cio->io_walk_link = zl;
362
363 if (zl == NULL)
364 return (NULL);
365
366 ASSERT(zl->zl_child == cio);
367 return (zl->zl_parent);
368}
369
370zio_t *
371zio_walk_children(zio_t *pio)
372{
373 zio_link_t *zl = pio->io_walk_link;
374 list_t *cl = &pio->io_child_list;
375
376 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
377 pio->io_walk_link = zl;
378
379 if (zl == NULL)
380 return (NULL);
381
382 ASSERT(zl->zl_parent == pio);
383 return (zl->zl_child);
384}
385
386zio_t *
387zio_unique_parent(zio_t *cio)
388{
389 zio_t *pio = zio_walk_parents(cio);
390
391 VERIFY(zio_walk_parents(cio) == NULL);
392 return (pio);
393}
394
395void
396zio_add_child(zio_t *pio, zio_t *cio)
397{
398 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
399
400 /*
401 * Logical I/Os can have logical, gang, or vdev children.
402 * Gang I/Os can have gang or vdev children.
403 * Vdev I/Os can only have vdev children.
404 * The following ASSERT captures all of these constraints.
405 */
406 ASSERT(cio->io_child_type <= pio->io_child_type);
407
408 zl->zl_parent = pio;
409 zl->zl_child = cio;
410
411 mutex_enter(&cio->io_lock);
412 mutex_enter(&pio->io_lock);
413
414 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
415
416 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
417 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
418
419 list_insert_head(&pio->io_child_list, zl);
420 list_insert_head(&cio->io_parent_list, zl);
421
422 pio->io_child_count++;
423 cio->io_parent_count++;
424
425 mutex_exit(&pio->io_lock);
426 mutex_exit(&cio->io_lock);
427}
428
429static void
430zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
431{
432 ASSERT(zl->zl_parent == pio);
433 ASSERT(zl->zl_child == cio);
434
435 mutex_enter(&cio->io_lock);
436 mutex_enter(&pio->io_lock);
437
438 list_remove(&pio->io_child_list, zl);
439 list_remove(&cio->io_parent_list, zl);
440
441 pio->io_child_count--;
442 cio->io_parent_count--;
443
444 mutex_exit(&pio->io_lock);
445 mutex_exit(&cio->io_lock);
446
447 kmem_cache_free(zio_link_cache, zl);
448}
449
450static boolean_t
451zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
452{
453 uint64_t *countp = &zio->io_children[child][wait];
454 boolean_t waiting = B_FALSE;
455
456 mutex_enter(&zio->io_lock);
457 ASSERT(zio->io_stall == NULL);
458 if (*countp != 0) {
459 zio->io_stage >>= 1;
460 zio->io_stall = countp;
461 waiting = B_TRUE;
462 }
463 mutex_exit(&zio->io_lock);
464
465 return (waiting);
466}
467
468static void
469zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
470{
471 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
472 int *errorp = &pio->io_child_error[zio->io_child_type];
473
474 mutex_enter(&pio->io_lock);
475 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
476 *errorp = zio_worst_error(*errorp, zio->io_error);
477 pio->io_reexecute |= zio->io_reexecute;
478 ASSERT3U(*countp, >, 0);
479 if (--*countp == 0 && pio->io_stall == countp) {
480 pio->io_stall = NULL;
481 mutex_exit(&pio->io_lock);
482 zio_execute(pio);
483 } else {
484 mutex_exit(&pio->io_lock);
485 }
486}
487
488static void
489zio_inherit_child_errors(zio_t *zio, enum zio_child c)
490{
491 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
492 zio->io_error = zio->io_child_error[c];
493}
494
495/*
496 * ==========================================================================
497 * Create the various types of I/O (read, write, free, etc)
498 * ==========================================================================
499 */
500static zio_t *
501zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
502 void *data, uint64_t size, zio_done_func_t *done, void *private,
503 zio_type_t type, int priority, enum zio_flag flags,
504 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
505 enum zio_stage stage, enum zio_stage pipeline)
506{
507 zio_t *zio;
508
509 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
510 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
511 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
512
513 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
514 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
515 ASSERT(vd || stage == ZIO_STAGE_OPEN);
516
517 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
518 bzero(zio, sizeof (zio_t));
519
520 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
521 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
522
523 list_create(&zio->io_parent_list, sizeof (zio_link_t),
524 offsetof(zio_link_t, zl_parent_node));
525 list_create(&zio->io_child_list, sizeof (zio_link_t),
526 offsetof(zio_link_t, zl_child_node));
527
528 if (vd != NULL)
529 zio->io_child_type = ZIO_CHILD_VDEV;
530 else if (flags & ZIO_FLAG_GANG_CHILD)
531 zio->io_child_type = ZIO_CHILD_GANG;
532 else if (flags & ZIO_FLAG_DDT_CHILD)
533 zio->io_child_type = ZIO_CHILD_DDT;
534 else
535 zio->io_child_type = ZIO_CHILD_LOGICAL;
536
537 if (bp != NULL) {
538 zio->io_bp = (blkptr_t *)bp;
539 zio->io_bp_copy = *bp;
540 zio->io_bp_orig = *bp;
541 if (type != ZIO_TYPE_WRITE ||
542 zio->io_child_type == ZIO_CHILD_DDT)
543 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
544 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
545 zio->io_logical = zio;
546 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
547 pipeline |= ZIO_GANG_STAGES;
548 }
549
550 zio->io_spa = spa;
551 zio->io_txg = txg;
552 zio->io_done = done;
553 zio->io_private = private;
554 zio->io_type = type;
555 zio->io_priority = priority;
556 zio->io_vd = vd;
557 zio->io_offset = offset;
558 zio->io_orig_data = zio->io_data = data;
559 zio->io_orig_size = zio->io_size = size;
560 zio->io_orig_flags = zio->io_flags = flags;
561 zio->io_orig_stage = zio->io_stage = stage;
562 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
563
564 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
565 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
566
567 if (zb != NULL)
568 zio->io_bookmark = *zb;
569
570 if (pio != NULL) {
571 if (zio->io_logical == NULL)
572 zio->io_logical = pio->io_logical;
573 if (zio->io_child_type == ZIO_CHILD_GANG)
574 zio->io_gang_leader = pio->io_gang_leader;
575 zio_add_child(pio, zio);
576 }
577
578 return (zio);
579}
580
581static void
582zio_destroy(zio_t *zio)
583{
584 list_destroy(&zio->io_parent_list);
585 list_destroy(&zio->io_child_list);
586 mutex_destroy(&zio->io_lock);
587 cv_destroy(&zio->io_cv);
588 kmem_cache_free(zio_cache, zio);
589}
590
591zio_t *
592zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
593 void *private, enum zio_flag flags)
594{
595 zio_t *zio;
596
597 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
598 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
599 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
600
601 return (zio);
602}
603
604zio_t *
605zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
606{
607 return (zio_null(NULL, spa, NULL, done, private, flags));
608}
609
610zio_t *
611zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
612 void *data, uint64_t size, zio_done_func_t *done, void *private,
613 int priority, enum zio_flag flags, const zbookmark_t *zb)
614{
615 zio_t *zio;
616
617 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
618 data, size, done, private,
619 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
620 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
621 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
622
623 return (zio);
624}
625
626zio_t *
627zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
628 void *data, uint64_t size, const zio_prop_t *zp,
629 zio_done_func_t *ready, zio_done_func_t *done, void *private,
630 int priority, enum zio_flag flags, const zbookmark_t *zb)
631{
632 zio_t *zio;
633
634 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
635 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
636 zp->zp_compress >= ZIO_COMPRESS_OFF &&
637 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
638 zp->zp_type < DMU_OT_NUMTYPES &&
639 zp->zp_level < 32 &&
640 zp->zp_copies > 0 &&
641 zp->zp_copies <= spa_max_replication(spa) &&
642 zp->zp_dedup <= 1 &&
643 zp->zp_dedup_verify <= 1);
644
645 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
646 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
647 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
648 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
649
650 zio->io_ready = ready;
651 zio->io_prop = *zp;
652
653 return (zio);
654}
655
656zio_t *
657zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
658 uint64_t size, zio_done_func_t *done, void *private, int priority,
659 enum zio_flag flags, zbookmark_t *zb)
660{
661 zio_t *zio;
662
663 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
664 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
665 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
666
667 return (zio);
668}
669
670void
671zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
672{
673 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
674 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
675 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
676 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
677
678 zio->io_prop.zp_copies = copies;
679 zio->io_bp_override = bp;
680}
681
682void
683zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
684{
685 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
686}
687
688zio_t *
689zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
690 enum zio_flag flags)
691{
692 zio_t *zio;
693
694 dprintf_bp(bp, "freeing in txg %llu, pass %u",
695 (longlong_t)txg, spa->spa_sync_pass);
696
697 ASSERT(!BP_IS_HOLE(bp));
698 ASSERT(spa_syncing_txg(spa) == txg);
699 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
700
701 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
702 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
703 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
704
705 return (zio);
706}
707
708zio_t *
709zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
710 zio_done_func_t *done, void *private, enum zio_flag flags)
711{
712 zio_t *zio;
713
714 /*
715 * A claim is an allocation of a specific block. Claims are needed
716 * to support immediate writes in the intent log. The issue is that
717 * immediate writes contain committed data, but in a txg that was
718 * *not* committed. Upon opening the pool after an unclean shutdown,
719 * the intent log claims all blocks that contain immediate write data
720 * so that the SPA knows they're in use.
721 *
722 * All claims *must* be resolved in the first txg -- before the SPA
723 * starts allocating blocks -- so that nothing is allocated twice.
724 * If txg == 0 we just verify that the block is claimable.
725 */
726 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
727 ASSERT(txg == spa_first_txg(spa) || txg == 0);
728 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
729
730 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
731 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
732 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
733
734 return (zio);
735}
736
737zio_t *
738zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
739 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
740{
741 zio_t *zio;
742 int c;
743
744 if (vd->vdev_children == 0) {
745 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
746 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
747 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
748
749 zio->io_cmd = cmd;
750 } else {
751 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
752
753 for (c = 0; c < vd->vdev_children; c++)
754 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
755 done, private, priority, flags));
756 }
757
758 return (zio);
759}
760
761zio_t *
762zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
763 void *data, int checksum, zio_done_func_t *done, void *private,
764 int priority, enum zio_flag flags, boolean_t labels)
765{
766 zio_t *zio;
767
768 ASSERT(vd->vdev_children == 0);
769 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
770 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
771 ASSERT3U(offset + size, <=, vd->vdev_psize);
772
773 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
774 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
775 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
776
777 zio->io_prop.zp_checksum = checksum;
778
779 return (zio);
780}
781
782zio_t *
783zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
784 void *data, int checksum, zio_done_func_t *done, void *private,
785 int priority, enum zio_flag flags, boolean_t labels)
786{
787 zio_t *zio;
788
789 ASSERT(vd->vdev_children == 0);
790 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
791 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
792 ASSERT3U(offset + size, <=, vd->vdev_psize);
793
794 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
795 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
796 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
797
798 zio->io_prop.zp_checksum = checksum;
799
800 if (zio_checksum_table[checksum].ci_eck) {
801 /*
802 * zec checksums are necessarily destructive -- they modify
803 * the end of the write buffer to hold the verifier/checksum.
804 * Therefore, we must make a local copy in case the data is
805 * being written to multiple places in parallel.
806 */
807 void *wbuf = zio_buf_alloc(size);
808 bcopy(data, wbuf, size);
809 zio_push_transform(zio, wbuf, size, size, NULL);
810 }
811
812 return (zio);
813}
814
815/*
816 * Create a child I/O to do some work for us.
817 */
818zio_t *
819zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
820 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
821 zio_done_func_t *done, void *private)
822{
823 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
824 zio_t *zio;
825
826 ASSERT(vd->vdev_parent ==
827 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
828
829 if (type == ZIO_TYPE_READ && bp != NULL) {
830 /*
831 * If we have the bp, then the child should perform the
832 * checksum and the parent need not. This pushes error
833 * detection as close to the leaves as possible and
834 * eliminates redundant checksums in the interior nodes.
835 */
836 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
837 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
838 }
839
840 if (vd->vdev_children == 0)
841 offset += VDEV_LABEL_START_SIZE;
842
843 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
844
845 /*
846 * If we've decided to do a repair, the write is not speculative --
847 * even if the original read was.
848 */
849 if (flags & ZIO_FLAG_IO_REPAIR)
850 flags &= ~ZIO_FLAG_SPECULATIVE;
851
852 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
853 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
854 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
855
856 return (zio);
857}
858
859zio_t *
860zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
861 int type, int priority, enum zio_flag flags,
862 zio_done_func_t *done, void *private)
863{
864 zio_t *zio;
865
866 ASSERT(vd->vdev_ops->vdev_op_leaf);
867
868 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
869 data, size, done, private, type, priority,
870 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
871 vd, offset, NULL,
872 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
873
874 return (zio);
875}
876
877void
878zio_flush(zio_t *zio, vdev_t *vd)
879{
880 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
881 NULL, NULL, ZIO_PRIORITY_NOW,
882 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
883}
884
885void
886zio_shrink(zio_t *zio, uint64_t size)
887{
888 ASSERT(zio->io_executor == NULL);
889 ASSERT(zio->io_orig_size == zio->io_size);
890 ASSERT(size <= zio->io_size);
891
892 /*
893 * We don't shrink for raidz because of problems with the
894 * reconstruction when reading back less than the block size.
895 * Note, BP_IS_RAIDZ() assumes no compression.
896 */
897 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
898 if (!BP_IS_RAIDZ(zio->io_bp))
899 zio->io_orig_size = zio->io_size = size;
900}
901
902/*
903 * ==========================================================================
904 * Prepare to read and write logical blocks
905 * ==========================================================================
906 */
907
908static int
909zio_read_bp_init(zio_t *zio)
910{
911 blkptr_t *bp = zio->io_bp;
912
913 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
914 zio->io_child_type == ZIO_CHILD_LOGICAL &&
915 !(zio->io_flags & ZIO_FLAG_RAW)) {
916 uint64_t psize = BP_GET_PSIZE(bp);
917 void *cbuf = zio_buf_alloc(psize);
918
919 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
920 }
921
922 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
923 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
924
925 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
926 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
927
928 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
929 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
930
931 return (ZIO_PIPELINE_CONTINUE);
932}
933
934static int
935zio_write_bp_init(zio_t *zio)
936{
937 spa_t *spa = zio->io_spa;
938 zio_prop_t *zp = &zio->io_prop;
939 enum zio_compress compress = zp->zp_compress;
940 blkptr_t *bp = zio->io_bp;
941 uint64_t lsize = zio->io_size;
942 uint64_t psize = lsize;
943 int pass = 1;
944
945 /*
946 * If our children haven't all reached the ready stage,
947 * wait for them and then repeat this pipeline stage.
948 */
949 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
950 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
951 return (ZIO_PIPELINE_STOP);
952
953 if (!IO_IS_ALLOCATING(zio))
954 return (ZIO_PIPELINE_CONTINUE);
955
956 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
957
958 if (zio->io_bp_override) {
959 ASSERT(bp->blk_birth != zio->io_txg);
960 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
961
962 *bp = *zio->io_bp_override;
963 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
964
965 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
966 return (ZIO_PIPELINE_CONTINUE);
967
968 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
969 zp->zp_dedup_verify);
970
971 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
972 BP_SET_DEDUP(bp, 1);
973 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
974 return (ZIO_PIPELINE_CONTINUE);
975 }
976 zio->io_bp_override = NULL;
977 BP_ZERO(bp);
978 }
979
980 if (bp->blk_birth == zio->io_txg) {
981 /*
982 * We're rewriting an existing block, which means we're
983 * working on behalf of spa_sync(). For spa_sync() to
984 * converge, it must eventually be the case that we don't
985 * have to allocate new blocks. But compression changes
986 * the blocksize, which forces a reallocate, and makes
987 * convergence take longer. Therefore, after the first
988 * few passes, stop compressing to ensure convergence.
989 */
990 pass = spa_sync_pass(spa);
991
992 ASSERT(zio->io_txg == spa_syncing_txg(spa));
993 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
994 ASSERT(!BP_GET_DEDUP(bp));
995
996 if (pass > SYNC_PASS_DONT_COMPRESS)
997 compress = ZIO_COMPRESS_OFF;
998
999 /* Make sure someone doesn't change their mind on overwrites */
1000 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1001 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1002 }
1003
1004 if (compress != ZIO_COMPRESS_OFF) {
1005 void *cbuf = zio_buf_alloc(lsize);
1006 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1007 if (psize == 0 || psize == lsize) {
1008 compress = ZIO_COMPRESS_OFF;
1009 zio_buf_free(cbuf, lsize);
1010 } else {
1011 ASSERT(psize < lsize);
1012 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1013 }
1014 }
1015
1016 /*
1017 * The final pass of spa_sync() must be all rewrites, but the first
1018 * few passes offer a trade-off: allocating blocks defers convergence,
1019 * but newly allocated blocks are sequential, so they can be written
1020 * to disk faster. Therefore, we allow the first few passes of
1021 * spa_sync() to allocate new blocks, but force rewrites after that.
1022 * There should only be a handful of blocks after pass 1 in any case.
1023 */
1024 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1025 pass > SYNC_PASS_REWRITE) {
1026 ASSERT(psize != 0);
1027 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1028 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1029 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1030 } else {
1031 BP_ZERO(bp);
1032 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1033 }
1034
1035 if (psize == 0) {
1036 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1037 } else {
1038 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1039 BP_SET_LSIZE(bp, lsize);
1040 BP_SET_PSIZE(bp, psize);
1041 BP_SET_COMPRESS(bp, compress);
1042 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1043 BP_SET_TYPE(bp, zp->zp_type);
1044 BP_SET_LEVEL(bp, zp->zp_level);
1045 BP_SET_DEDUP(bp, zp->zp_dedup);
1046 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1047 if (zp->zp_dedup) {
1048 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1049 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1050 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1051 }
1052 }
1053
1054 return (ZIO_PIPELINE_CONTINUE);
1055}
1056
1057static int
1058zio_free_bp_init(zio_t *zio)
1059{
1060 blkptr_t *bp = zio->io_bp;
1061
1062 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1063 if (BP_GET_DEDUP(bp))
1064 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1065 }
1066
1067 return (ZIO_PIPELINE_CONTINUE);
1068}
1069
1070/*
1071 * ==========================================================================
1072 * Execute the I/O pipeline
1073 * ==========================================================================
1074 */
1075
1076static void
1077zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1078{
1079 spa_t *spa = zio->io_spa;
1080 zio_type_t t = zio->io_type;
1081 int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1082
1083 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1084
1085 /*
1086 * If we're a config writer or a probe, the normal issue and
1087 * interrupt threads may all be blocked waiting for the config lock.
1088 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1089 */
1090 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1091 t = ZIO_TYPE_NULL;
1092
1093 /*
1094 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1095 */
1096 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1097 t = ZIO_TYPE_NULL;
1098
1099 /*
1100 * If this is a high priority I/O, then use the high priority taskq.
1101 */
1102 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1103 spa->spa_zio_taskq[t][q + 1] != NULL)
1104 q++;
1105
1106 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1107#ifdef _KERNEL
1108 (void) taskq_dispatch_safe(spa->spa_zio_taskq[t][q],
1109 (task_func_t *)zio_execute, zio, flags, &zio->io_task);
1110#else
1111 (void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1112 (task_func_t *)zio_execute, zio, flags);
1113#endif
1114}
1115
1116static boolean_t
1117zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1118{
1119 kthread_t *executor = zio->io_executor;
1120 spa_t *spa = zio->io_spa;
1121
1122 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1123 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1124 return (B_TRUE);
1125
1126 return (B_FALSE);
1127}
1128
1129static int
1130zio_issue_async(zio_t *zio)
1131{
1132 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1133
1134 return (ZIO_PIPELINE_STOP);
1135}
1136
1137void
1138zio_interrupt(zio_t *zio)
1139{
1140 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1141}
1142
1143/*
1144 * Execute the I/O pipeline until one of the following occurs:
1145 * (1) the I/O completes; (2) the pipeline stalls waiting for
1146 * dependent child I/Os; (3) the I/O issues, so we're waiting
1147 * for an I/O completion interrupt; (4) the I/O is delegated by
1148 * vdev-level caching or aggregation; (5) the I/O is deferred
1149 * due to vdev-level queueing; (6) the I/O is handed off to
1150 * another thread. In all cases, the pipeline stops whenever
1151 * there's no CPU work; it never burns a thread in cv_wait().
1152 *
1153 * There's no locking on io_stage because there's no legitimate way
1154 * for multiple threads to be attempting to process the same I/O.
1155 */
1156static zio_pipe_stage_t *zio_pipeline[];
1157
1158void
1159zio_execute(zio_t *zio)
1160{
1161 zio->io_executor = curthread;
1162
1163 while (zio->io_stage < ZIO_STAGE_DONE) {
1164 enum zio_stage pipeline = zio->io_pipeline;
1165 enum zio_stage stage = zio->io_stage;
1166 int rv;
1167
1168 ASSERT(!MUTEX_HELD(&zio->io_lock));
1169 ASSERT(ISP2(stage));
1170 ASSERT(zio->io_stall == NULL);
1171
1172 do {
1173 stage <<= 1;
1174 } while ((stage & pipeline) == 0);
1175
1176 ASSERT(stage <= ZIO_STAGE_DONE);
1177
1178 /*
1179 * If we are in interrupt context and this pipeline stage
1180 * will grab a config lock that is held across I/O,
1181 * or may wait for an I/O that needs an interrupt thread
1182 * to complete, issue async to avoid deadlock.
1183 *
1184 * For VDEV_IO_START, we cut in line so that the io will
1185 * be sent to disk promptly.
1186 */
1187 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1188 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1189 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1190 zio_requeue_io_start_cut_in_line : B_FALSE;
1191 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1192 return;
1193 }
1194
1195 zio->io_stage = stage;
1196 rv = zio_pipeline[highbit(stage) - 1](zio);
1197
1198 if (rv == ZIO_PIPELINE_STOP)
1199 return;
1200
1201 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1202 }
1203}
1204
1205/*
1206 * ==========================================================================
1207 * Initiate I/O, either sync or async
1208 * ==========================================================================
1209 */
1210int
1211zio_wait(zio_t *zio)
1212{
1213 int error;
1214
1215 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1216 ASSERT(zio->io_executor == NULL);
1217
1218 zio->io_waiter = curthread;
1219
1220 zio_execute(zio);
1221
1222 mutex_enter(&zio->io_lock);
1223 while (zio->io_executor != NULL)
1224 cv_wait(&zio->io_cv, &zio->io_lock);
1225 mutex_exit(&zio->io_lock);
1226
1227 error = zio->io_error;
1228 zio_destroy(zio);
1229
1230 return (error);
1231}
1232
1233void
1234zio_nowait(zio_t *zio)
1235{
1236 ASSERT(zio->io_executor == NULL);
1237
1238 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1239 zio_unique_parent(zio) == NULL) {
1240 /*
1241 * This is a logical async I/O with no parent to wait for it.
1242 * We add it to the spa_async_root_zio "Godfather" I/O which
1243 * will ensure they complete prior to unloading the pool.
1244 */
1245 spa_t *spa = zio->io_spa;
1246
1247 zio_add_child(spa->spa_async_zio_root, zio);
1248 }
1249
1250 zio_execute(zio);
1251}
1252
1253/*
1254 * ==========================================================================
1255 * Reexecute or suspend/resume failed I/O
1256 * ==========================================================================
1257 */
1258
1259static void
1260zio_reexecute(zio_t *pio)
1261{
1262 zio_t *cio, *cio_next;
1263
1264 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1265 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1266 ASSERT(pio->io_gang_leader == NULL);
1267 ASSERT(pio->io_gang_tree == NULL);
1268
1269 pio->io_flags = pio->io_orig_flags;
1270 pio->io_stage = pio->io_orig_stage;
1271 pio->io_pipeline = pio->io_orig_pipeline;
1272 pio->io_reexecute = 0;
1273 pio->io_error = 0;
1274 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1275 pio->io_state[w] = 0;
1276 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1277 pio->io_child_error[c] = 0;
1278
1279 if (IO_IS_ALLOCATING(pio))
1280 BP_ZERO(pio->io_bp);
1281
1282 /*
1283 * As we reexecute pio's children, new children could be created.
1284 * New children go to the head of pio's io_child_list, however,
1285 * so we will (correctly) not reexecute them. The key is that
1286 * the remainder of pio's io_child_list, from 'cio_next' onward,
1287 * cannot be affected by any side effects of reexecuting 'cio'.
1288 */
1289 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1290 cio_next = zio_walk_children(pio);
1291 mutex_enter(&pio->io_lock);
1292 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1293 pio->io_children[cio->io_child_type][w]++;
1294 mutex_exit(&pio->io_lock);
1295 zio_reexecute(cio);
1296 }
1297
1298 /*
1299 * Now that all children have been reexecuted, execute the parent.
1300 * We don't reexecute "The Godfather" I/O here as it's the
1301 * responsibility of the caller to wait on him.
1302 */
1303 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1304 zio_execute(pio);
1305}
1306
1307void
1308zio_suspend(spa_t *spa, zio_t *zio)
1309{
1310 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1311 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1312 "failure and the failure mode property for this pool "
1313 "is set to panic.", spa_name(spa));
1314
1315 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1316
1317 mutex_enter(&spa->spa_suspend_lock);
1318
1319 if (spa->spa_suspend_zio_root == NULL)
1320 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1321 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1322 ZIO_FLAG_GODFATHER);
1323
1324 spa->spa_suspended = B_TRUE;
1325
1326 if (zio != NULL) {
1327 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1328 ASSERT(zio != spa->spa_suspend_zio_root);
1329 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1330 ASSERT(zio_unique_parent(zio) == NULL);
1331 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1332 zio_add_child(spa->spa_suspend_zio_root, zio);
1333 }
1334
1335 mutex_exit(&spa->spa_suspend_lock);
1336}
1337
1338int
1339zio_resume(spa_t *spa)
1340{
1341 zio_t *pio;
1342
1343 /*
1344 * Reexecute all previously suspended i/o.
1345 */
1346 mutex_enter(&spa->spa_suspend_lock);
1347 spa->spa_suspended = B_FALSE;
1348 cv_broadcast(&spa->spa_suspend_cv);
1349 pio = spa->spa_suspend_zio_root;
1350 spa->spa_suspend_zio_root = NULL;
1351 mutex_exit(&spa->spa_suspend_lock);
1352
1353 if (pio == NULL)
1354 return (0);
1355
1356 zio_reexecute(pio);
1357 return (zio_wait(pio));
1358}
1359
1360void
1361zio_resume_wait(spa_t *spa)
1362{
1363 mutex_enter(&spa->spa_suspend_lock);
1364 while (spa_suspended(spa))
1365 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1366 mutex_exit(&spa->spa_suspend_lock);
1367}
1368
1369/*
1370 * ==========================================================================
1371 * Gang blocks.
1372 *
1373 * A gang block is a collection of small blocks that looks to the DMU
1374 * like one large block. When zio_dva_allocate() cannot find a block
1375 * of the requested size, due to either severe fragmentation or the pool
1376 * being nearly full, it calls zio_write_gang_block() to construct the
1377 * block from smaller fragments.
1378 *
1379 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1380 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1381 * an indirect block: it's an array of block pointers. It consumes
1382 * only one sector and hence is allocatable regardless of fragmentation.
1383 * The gang header's bps point to its gang members, which hold the data.
1384 *
1385 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1386 * as the verifier to ensure uniqueness of the SHA256 checksum.
1387 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1388 * not the gang header. This ensures that data block signatures (needed for
1389 * deduplication) are independent of how the block is physically stored.
1390 *
1391 * Gang blocks can be nested: a gang member may itself be a gang block.
1392 * Thus every gang block is a tree in which root and all interior nodes are
1393 * gang headers, and the leaves are normal blocks that contain user data.
1394 * The root of the gang tree is called the gang leader.
1395 *
1396 * To perform any operation (read, rewrite, free, claim) on a gang block,
1397 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1398 * in the io_gang_tree field of the original logical i/o by recursively
1399 * reading the gang leader and all gang headers below it. This yields
1400 * an in-core tree containing the contents of every gang header and the
1401 * bps for every constituent of the gang block.
1402 *
1403 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1404 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1405 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1406 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1407 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1408 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1409 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1410 * of the gang header plus zio_checksum_compute() of the data to update the
1411 * gang header's blk_cksum as described above.
1412 *
1413 * The two-phase assemble/issue model solves the problem of partial failure --
1414 * what if you'd freed part of a gang block but then couldn't read the
1415 * gang header for another part? Assembling the entire gang tree first
1416 * ensures that all the necessary gang header I/O has succeeded before
1417 * starting the actual work of free, claim, or write. Once the gang tree
1418 * is assembled, free and claim are in-memory operations that cannot fail.
1419 *
1420 * In the event that a gang write fails, zio_dva_unallocate() walks the
1421 * gang tree to immediately free (i.e. insert back into the space map)
1422 * everything we've allocated. This ensures that we don't get ENOSPC
1423 * errors during repeated suspend/resume cycles due to a flaky device.
1424 *
1425 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1426 * the gang tree, we won't modify the block, so we can safely defer the free
1427 * (knowing that the block is still intact). If we *can* assemble the gang
1428 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1429 * each constituent bp and we can allocate a new block on the next sync pass.
1430 *
1431 * In all cases, the gang tree allows complete recovery from partial failure.
1432 * ==========================================================================
1433 */
1434
1435static zio_t *
1436zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1437{
1438 if (gn != NULL)
1439 return (pio);
1440
1441 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1442 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1443 &pio->io_bookmark));
1444}
1445
1446zio_t *
1447zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1448{
1449 zio_t *zio;
1450
1451 if (gn != NULL) {
1452 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1453 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1454 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1455 /*
1456 * As we rewrite each gang header, the pipeline will compute
1457 * a new gang block header checksum for it; but no one will
1458 * compute a new data checksum, so we do that here. The one
1459 * exception is the gang leader: the pipeline already computed
1460 * its data checksum because that stage precedes gang assembly.
1461 * (Presently, nothing actually uses interior data checksums;
1462 * this is just good hygiene.)
1463 */
1464 if (gn != pio->io_gang_leader->io_gang_tree) {
1465 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1466 data, BP_GET_PSIZE(bp));
1467 }
1468 /*
1469 * If we are here to damage data for testing purposes,
1470 * leave the GBH alone so that we can detect the damage.
1471 */
1472 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1473 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1474 } else {
1475 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1476 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1477 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1478 }
1479
1480 return (zio);
1481}
1482
1483/* ARGSUSED */
1484zio_t *
1485zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1486{
1487 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1488 ZIO_GANG_CHILD_FLAGS(pio)));
1489}
1490
1491/* ARGSUSED */
1492zio_t *
1493zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1494{
1495 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1496 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1497}
1498
1499static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1500 NULL,
1501 zio_read_gang,
1502 zio_rewrite_gang,
1503 zio_free_gang,
1504 zio_claim_gang,
1505 NULL
1506};
1507
1508static void zio_gang_tree_assemble_done(zio_t *zio);
1509
1510static zio_gang_node_t *
1511zio_gang_node_alloc(zio_gang_node_t **gnpp)
1512{
1513 zio_gang_node_t *gn;
1514
1515 ASSERT(*gnpp == NULL);
1516
1517 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1518 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1519 *gnpp = gn;
1520
1521 return (gn);
1522}
1523
1524static void
1525zio_gang_node_free(zio_gang_node_t **gnpp)
1526{
1527 zio_gang_node_t *gn = *gnpp;
1528
1529 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1530 ASSERT(gn->gn_child[g] == NULL);
1531
1532 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1533 kmem_free(gn, sizeof (*gn));
1534 *gnpp = NULL;
1535}
1536
1537static void
1538zio_gang_tree_free(zio_gang_node_t **gnpp)
1539{
1540 zio_gang_node_t *gn = *gnpp;
1541
1542 if (gn == NULL)
1543 return;
1544
1545 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1546 zio_gang_tree_free(&gn->gn_child[g]);
1547
1548 zio_gang_node_free(gnpp);
1549}
1550
1551static void
1552zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1553{
1554 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1555
1556 ASSERT(gio->io_gang_leader == gio);
1557 ASSERT(BP_IS_GANG(bp));
1558
1559 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1560 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1561 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1562}
1563
1564static void
1565zio_gang_tree_assemble_done(zio_t *zio)
1566{
1567 zio_t *gio = zio->io_gang_leader;
1568 zio_gang_node_t *gn = zio->io_private;
1569 blkptr_t *bp = zio->io_bp;
1570
1571 ASSERT(gio == zio_unique_parent(zio));
1572 ASSERT(zio->io_child_count == 0);
1573
1574 if (zio->io_error)
1575 return;
1576
1577 if (BP_SHOULD_BYTESWAP(bp))
1578 byteswap_uint64_array(zio->io_data, zio->io_size);
1579
1580 ASSERT(zio->io_data == gn->gn_gbh);
1581 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1582 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1583
1584 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1585 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1586 if (!BP_IS_GANG(gbp))
1587 continue;
1588 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1589 }
1590}
1591
1592static void
1593zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1594{
1595 zio_t *gio = pio->io_gang_leader;
1596 zio_t *zio;
1597
1598 ASSERT(BP_IS_GANG(bp) == !!gn);
1599 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1600 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1601
1602 /*
1603 * If you're a gang header, your data is in gn->gn_gbh.
1604 * If you're a gang member, your data is in 'data' and gn == NULL.
1605 */
1606 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1607
1608 if (gn != NULL) {
1609 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1610
1611 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1612 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1613 if (BP_IS_HOLE(gbp))
1614 continue;
1615 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1616 data = (char *)data + BP_GET_PSIZE(gbp);
1617 }
1618 }
1619
1620 if (gn == gio->io_gang_tree)
1621 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1622
1623 if (zio != pio)
1624 zio_nowait(zio);
1625}
1626
1627static int
1628zio_gang_assemble(zio_t *zio)
1629{
1630 blkptr_t *bp = zio->io_bp;
1631
1632 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1633 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1634
1635 zio->io_gang_leader = zio;
1636
1637 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1638
1639 return (ZIO_PIPELINE_CONTINUE);
1640}
1641
1642static int
1643zio_gang_issue(zio_t *zio)
1644{
1645 blkptr_t *bp = zio->io_bp;
1646
1647 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1648 return (ZIO_PIPELINE_STOP);
1649
1650 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1651 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1652
1653 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1654 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1655 else
1656 zio_gang_tree_free(&zio->io_gang_tree);
1657
1658 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1659
1660 return (ZIO_PIPELINE_CONTINUE);
1661}
1662
1663static void
1664zio_write_gang_member_ready(zio_t *zio)
1665{
1666 zio_t *pio = zio_unique_parent(zio);
1667 zio_t *gio = zio->io_gang_leader;
1668 dva_t *cdva = zio->io_bp->blk_dva;
1669 dva_t *pdva = pio->io_bp->blk_dva;
1670 uint64_t asize;
1671
1672 if (BP_IS_HOLE(zio->io_bp))
1673 return;
1674
1675 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1676
1677 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1678 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1679 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1680 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1681 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1682
1683 mutex_enter(&pio->io_lock);
1684 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1685 ASSERT(DVA_GET_GANG(&pdva[d]));
1686 asize = DVA_GET_ASIZE(&pdva[d]);
1687 asize += DVA_GET_ASIZE(&cdva[d]);
1688 DVA_SET_ASIZE(&pdva[d], asize);
1689 }
1690 mutex_exit(&pio->io_lock);
1691}
1692
1693static int
1694zio_write_gang_block(zio_t *pio)
1695{
1696 spa_t *spa = pio->io_spa;
1697 blkptr_t *bp = pio->io_bp;
1698 zio_t *gio = pio->io_gang_leader;
1699 zio_t *zio;
1700 zio_gang_node_t *gn, **gnpp;
1701 zio_gbh_phys_t *gbh;
1702 uint64_t txg = pio->io_txg;
1703 uint64_t resid = pio->io_size;
1704 uint64_t lsize;
1705 int copies = gio->io_prop.zp_copies;
1706 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1707 zio_prop_t zp;
1708 int error;
1709
1710 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1711 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1712 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1713 if (error) {
1714 pio->io_error = error;
1715 return (ZIO_PIPELINE_CONTINUE);
1716 }
1717
1718 if (pio == gio) {
1719 gnpp = &gio->io_gang_tree;
1720 } else {
1721 gnpp = pio->io_private;
1722 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1723 }
1724
1725 gn = zio_gang_node_alloc(gnpp);
1726 gbh = gn->gn_gbh;
1727 bzero(gbh, SPA_GANGBLOCKSIZE);
1728
1729 /*
1730 * Create the gang header.
1731 */
1732 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1733 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1734
1735 /*
1736 * Create and nowait the gang children.
1737 */
1738 for (int g = 0; resid != 0; resid -= lsize, g++) {
1739 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1740 SPA_MINBLOCKSIZE);
1741 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1742
1743 zp.zp_checksum = gio->io_prop.zp_checksum;
1744 zp.zp_compress = ZIO_COMPRESS_OFF;
1745 zp.zp_type = DMU_OT_NONE;
1746 zp.zp_level = 0;
1747 zp.zp_copies = gio->io_prop.zp_copies;
1748 zp.zp_dedup = 0;
1749 zp.zp_dedup_verify = 0;
1750
1751 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1752 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1753 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1754 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1755 &pio->io_bookmark));
1756 }
1757
1758 /*
1759 * Set pio's pipeline to just wait for zio to finish.
1760 */
1761 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1762
1763 zio_nowait(zio);
1764
1765 return (ZIO_PIPELINE_CONTINUE);
1766}
1767
1768/*
1769 * ==========================================================================
1770 * Dedup
1771 * ==========================================================================
1772 */
1773static void
1774zio_ddt_child_read_done(zio_t *zio)
1775{
1776 blkptr_t *bp = zio->io_bp;
1777 ddt_entry_t *dde = zio->io_private;
1778 ddt_phys_t *ddp;
1779 zio_t *pio = zio_unique_parent(zio);
1780
1781 mutex_enter(&pio->io_lock);
1782 ddp = ddt_phys_select(dde, bp);
1783 if (zio->io_error == 0)
1784 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1785 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1786 dde->dde_repair_data = zio->io_data;
1787 else
1788 zio_buf_free(zio->io_data, zio->io_size);
1789 mutex_exit(&pio->io_lock);
1790}
1791
1792static int
1793zio_ddt_read_start(zio_t *zio)
1794{
1795 blkptr_t *bp = zio->io_bp;
1796
1797 ASSERT(BP_GET_DEDUP(bp));
1798 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1799 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1800
1801 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1802 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1803 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1804 ddt_phys_t *ddp = dde->dde_phys;
1805 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1806 blkptr_t blk;
1807
1808 ASSERT(zio->io_vsd == NULL);
1809 zio->io_vsd = dde;
1810
1811 if (ddp_self == NULL)
1812 return (ZIO_PIPELINE_CONTINUE);
1813
1814 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1815 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1816 continue;
1817 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1818 &blk);
1819 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1820 zio_buf_alloc(zio->io_size), zio->io_size,
1821 zio_ddt_child_read_done, dde, zio->io_priority,
1822 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1823 &zio->io_bookmark));
1824 }
1825 return (ZIO_PIPELINE_CONTINUE);
1826 }
1827
1828 zio_nowait(zio_read(zio, zio->io_spa, bp,
1829 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1830 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1831
1832 return (ZIO_PIPELINE_CONTINUE);
1833}
1834
1835static int
1836zio_ddt_read_done(zio_t *zio)
1837{
1838 blkptr_t *bp = zio->io_bp;
1839
1840 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1841 return (ZIO_PIPELINE_STOP);
1842
1843 ASSERT(BP_GET_DEDUP(bp));
1844 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1845 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1846
1847 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1848 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1849 ddt_entry_t *dde = zio->io_vsd;
1850 if (ddt == NULL) {
1851 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1852 return (ZIO_PIPELINE_CONTINUE);
1853 }
1854 if (dde == NULL) {
1855 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1856 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1857 return (ZIO_PIPELINE_STOP);
1858 }
1859 if (dde->dde_repair_data != NULL) {
1860 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1861 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1862 }
1863 ddt_repair_done(ddt, dde);
1864 zio->io_vsd = NULL;
1865 }
1866
1867 ASSERT(zio->io_vsd == NULL);
1868
1869 return (ZIO_PIPELINE_CONTINUE);
1870}
1871
1872static boolean_t
1873zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1874{
1875 spa_t *spa = zio->io_spa;
1876
1877 /*
1878 * Note: we compare the original data, not the transformed data,
1879 * because when zio->io_bp is an override bp, we will not have
1880 * pushed the I/O transforms. That's an important optimization
1881 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1882 */
1883 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1884 zio_t *lio = dde->dde_lead_zio[p];
1885
1886 if (lio != NULL) {
1887 return (lio->io_orig_size != zio->io_orig_size ||
1888 bcmp(zio->io_orig_data, lio->io_orig_data,
1889 zio->io_orig_size) != 0);
1890 }
1891 }
1892
1893 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1894 ddt_phys_t *ddp = &dde->dde_phys[p];
1895
1896 if (ddp->ddp_phys_birth != 0) {
1897 arc_buf_t *abuf = NULL;
1898 uint32_t aflags = ARC_WAIT;
1899 blkptr_t blk = *zio->io_bp;
1900 int error;
1901
1902 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1903
1904 ddt_exit(ddt);
1905
1906 error = arc_read_nolock(NULL, spa, &blk,
1907 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1908 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1909 &aflags, &zio->io_bookmark);
1910
1911 if (error == 0) {
1912 if (arc_buf_size(abuf) != zio->io_orig_size ||
1913 bcmp(abuf->b_data, zio->io_orig_data,
1914 zio->io_orig_size) != 0)
1915 error = EEXIST;
1916 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1917 }
1918
1919 ddt_enter(ddt);
1920 return (error != 0);
1921 }
1922 }
1923
1924 return (B_FALSE);
1925}
1926
1927static void
1928zio_ddt_child_write_ready(zio_t *zio)
1929{
1930 int p = zio->io_prop.zp_copies;
1931 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1932 ddt_entry_t *dde = zio->io_private;
1933 ddt_phys_t *ddp = &dde->dde_phys[p];
1934 zio_t *pio;
1935
1936 if (zio->io_error)
1937 return;
1938
1939 ddt_enter(ddt);
1940
1941 ASSERT(dde->dde_lead_zio[p] == zio);
1942
1943 ddt_phys_fill(ddp, zio->io_bp);
1944
1945 while ((pio = zio_walk_parents(zio)) != NULL)
1946 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1947
1948 ddt_exit(ddt);
1949}
1950
1951static void
1952zio_ddt_child_write_done(zio_t *zio)
1953{
1954 int p = zio->io_prop.zp_copies;
1955 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1956 ddt_entry_t *dde = zio->io_private;
1957 ddt_phys_t *ddp = &dde->dde_phys[p];
1958
1959 ddt_enter(ddt);
1960
1961 ASSERT(ddp->ddp_refcnt == 0);
1962 ASSERT(dde->dde_lead_zio[p] == zio);
1963 dde->dde_lead_zio[p] = NULL;
1964
1965 if (zio->io_error == 0) {
1966 while (zio_walk_parents(zio) != NULL)
1967 ddt_phys_addref(ddp);
1968 } else {
1969 ddt_phys_clear(ddp);
1970 }
1971
1972 ddt_exit(ddt);
1973}
1974
1975static void
1976zio_ddt_ditto_write_done(zio_t *zio)
1977{
1978 int p = DDT_PHYS_DITTO;
1979 zio_prop_t *zp = &zio->io_prop;
1980 blkptr_t *bp = zio->io_bp;
1981 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1982 ddt_entry_t *dde = zio->io_private;
1983 ddt_phys_t *ddp = &dde->dde_phys[p];
1984 ddt_key_t *ddk = &dde->dde_key;
1985
1986 ddt_enter(ddt);
1987
1988 ASSERT(ddp->ddp_refcnt == 0);
1989 ASSERT(dde->dde_lead_zio[p] == zio);
1990 dde->dde_lead_zio[p] = NULL;
1991
1992 if (zio->io_error == 0) {
1993 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1994 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1995 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1996 if (ddp->ddp_phys_birth != 0)
1997 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1998 ddt_phys_fill(ddp, bp);
1999 }
2000
2001 ddt_exit(ddt);
2002}
2003
2004static int
2005zio_ddt_write(zio_t *zio)
2006{
2007 spa_t *spa = zio->io_spa;
2008 blkptr_t *bp = zio->io_bp;
2009 uint64_t txg = zio->io_txg;
2010 zio_prop_t *zp = &zio->io_prop;
2011 int p = zp->zp_copies;
2012 int ditto_copies;
2013 zio_t *cio = NULL;
2014 zio_t *dio = NULL;
2015 ddt_t *ddt = ddt_select(spa, bp);
2016 ddt_entry_t *dde;
2017 ddt_phys_t *ddp;
2018
2019 ASSERT(BP_GET_DEDUP(bp));
2020 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2021 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2022
2023 ddt_enter(ddt);
2024 dde = ddt_lookup(ddt, bp, B_TRUE);
2025 ddp = &dde->dde_phys[p];
2026
2027 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2028 /*
2029 * If we're using a weak checksum, upgrade to a strong checksum
2030 * and try again. If we're already using a strong checksum,
2031 * we can't resolve it, so just convert to an ordinary write.
2032 * (And automatically e-mail a paper to Nature?)
2033 */
2034 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2035 zp->zp_checksum = spa_dedup_checksum(spa);
2036 zio_pop_transforms(zio);
2037 zio->io_stage = ZIO_STAGE_OPEN;
2038 BP_ZERO(bp);
2039 } else {
2040 zp->zp_dedup = 0;
2041 }
2042 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2043 ddt_exit(ddt);
2044 return (ZIO_PIPELINE_CONTINUE);
2045 }
2046
2047 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2048 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2049
2050 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2051 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2052 zio_prop_t czp = *zp;
2053
2054 czp.zp_copies = ditto_copies;
2055
2056 /*
2057 * If we arrived here with an override bp, we won't have run
2058 * the transform stack, so we won't have the data we need to
2059 * generate a child i/o. So, toss the override bp and restart.
2060 * This is safe, because using the override bp is just an
2061 * optimization; and it's rare, so the cost doesn't matter.
2062 */
2063 if (zio->io_bp_override) {
2064 zio_pop_transforms(zio);
2065 zio->io_stage = ZIO_STAGE_OPEN;
2066 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2067 zio->io_bp_override = NULL;
2068 BP_ZERO(bp);
2069 ddt_exit(ddt);
2070 return (ZIO_PIPELINE_CONTINUE);
2071 }
2072
2073 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2074 zio->io_orig_size, &czp, NULL,
2075 zio_ddt_ditto_write_done, dde, zio->io_priority,
2076 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2077
2078 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2079 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2080 }
2081
2082 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2083 if (ddp->ddp_phys_birth != 0)
2084 ddt_bp_fill(ddp, bp, txg);
2085 if (dde->dde_lead_zio[p] != NULL)
2086 zio_add_child(zio, dde->dde_lead_zio[p]);
2087 else
2088 ddt_phys_addref(ddp);
2089 } else if (zio->io_bp_override) {
2090 ASSERT(bp->blk_birth == txg);
2091 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2092 ddt_phys_fill(ddp, bp);
2093 ddt_phys_addref(ddp);
2094 } else {
2095 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2096 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2097 zio_ddt_child_write_done, dde, zio->io_priority,
2098 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2099
2100 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2101 dde->dde_lead_zio[p] = cio;
2102 }
2103
2104 ddt_exit(ddt);
2105
2106 if (cio)
2107 zio_nowait(cio);
2108 if (dio)
2109 zio_nowait(dio);
2110
2111 return (ZIO_PIPELINE_CONTINUE);
2112}
2113
2114ddt_entry_t *freedde; /* for debugging */
2115
2116static int
2117zio_ddt_free(zio_t *zio)
2118{
2119 spa_t *spa = zio->io_spa;
2120 blkptr_t *bp = zio->io_bp;
2121 ddt_t *ddt = ddt_select(spa, bp);
2122 ddt_entry_t *dde;
2123 ddt_phys_t *ddp;
2124
2125 ASSERT(BP_GET_DEDUP(bp));
2126 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2127
2128 ddt_enter(ddt);
2129 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2130 ddp = ddt_phys_select(dde, bp);
2131 ddt_phys_decref(ddp);
2132 ddt_exit(ddt);
2133
2134 return (ZIO_PIPELINE_CONTINUE);
2135}
2136
2137/*
2138 * ==========================================================================
2139 * Allocate and free blocks
2140 * ==========================================================================
2141 */
2142static int
2143zio_dva_allocate(zio_t *zio)
2144{
2145 spa_t *spa = zio->io_spa;
2146 metaslab_class_t *mc = spa_normal_class(spa);
2147 blkptr_t *bp = zio->io_bp;
2148 int error;
2149 int flags = 0;
2150
2151 if (zio->io_gang_leader == NULL) {
2152 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2153 zio->io_gang_leader = zio;
2154 }
2155
2156 ASSERT(BP_IS_HOLE(bp));
2157 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2158 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2159 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2160 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2161
2162 /*
2163 * The dump device does not support gang blocks so allocation on
2164 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2165 * the "fast" gang feature.
2166 */
2167 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2168 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2169 METASLAB_GANG_CHILD : 0;
2170 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2171 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2172
2173 if (error) {
2174 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2175 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2176 error);
2177 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2178 return (zio_write_gang_block(zio));
2179 zio->io_error = error;
2180 }
2181
2182 return (ZIO_PIPELINE_CONTINUE);
2183}
2184
2185static int
2186zio_dva_free(zio_t *zio)
2187{
2188 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2189
2190 return (ZIO_PIPELINE_CONTINUE);
2191}
2192
2193static int
2194zio_dva_claim(zio_t *zio)
2195{
2196 int error;
2197
2198 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2199 if (error)
2200 zio->io_error = error;
2201
2202 return (ZIO_PIPELINE_CONTINUE);
2203}
2204
2205/*
2206 * Undo an allocation. This is used by zio_done() when an I/O fails
2207 * and we want to give back the block we just allocated.
2208 * This handles both normal blocks and gang blocks.
2209 */
2210static void
2211zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2212{
2213 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2214 ASSERT(zio->io_bp_override == NULL);
2215
2216 if (!BP_IS_HOLE(bp))
2217 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2218
2219 if (gn != NULL) {
2220 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2221 zio_dva_unallocate(zio, gn->gn_child[g],
2222 &gn->gn_gbh->zg_blkptr[g]);
2223 }
2224 }
2225}
2226
2227/*
2228 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2229 */
2230int
2231zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2232 uint64_t size, boolean_t use_slog)
2233{
2234 int error = 1;
2235
2236 ASSERT(txg > spa_syncing_txg(spa));
2237
2238 /*
2239 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2240 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2241 * when allocating them.
2242 */
2243 if (use_slog) {
2244 error = metaslab_alloc(spa, spa_log_class(spa), size,
2245 new_bp, 1, txg, old_bp,
2246 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2247 }
2248
2249 if (error) {
2250 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2251 new_bp, 1, txg, old_bp,
2252 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2253 }
2254
2255 if (error == 0) {
2256 BP_SET_LSIZE(new_bp, size);
2257 BP_SET_PSIZE(new_bp, size);
2258 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2259 BP_SET_CHECKSUM(new_bp,
2260 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2261 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2262 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2263 BP_SET_LEVEL(new_bp, 0);
2264 BP_SET_DEDUP(new_bp, 0);
2265 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2266 }
2267
2268 return (error);
2269}
2270
2271/*
2272 * Free an intent log block.
2273 */
2274void
2275zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2276{
2277 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2278 ASSERT(!BP_IS_GANG(bp));
2279
2280 zio_free(spa, txg, bp);
2281}
2282
2283/*
2284 * ==========================================================================
2285 * Read and write to physical devices
2286 * ==========================================================================
2287 */
2288static int
2289zio_vdev_io_start(zio_t *zio)
2290{
2291 vdev_t *vd = zio->io_vd;
2292 uint64_t align;
2293 spa_t *spa = zio->io_spa;
2294
2295 ASSERT(zio->io_error == 0);
2296 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2297
2298 if (vd == NULL) {
2299 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2300 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2301
2302 /*
2303 * The mirror_ops handle multiple DVAs in a single BP.
2304 */
2305 return (vdev_mirror_ops.vdev_op_io_start(zio));
2306 }
2307
2308 /*
2309 * We keep track of time-sensitive I/Os so that the scan thread
2310 * can quickly react to certain workloads. In particular, we care
2311 * about non-scrubbing, top-level reads and writes with the following
2312 * characteristics:
2313 * - synchronous writes of user data to non-slog devices
2314 * - any reads of user data
2315 * When these conditions are met, adjust the timestamp of spa_last_io
2316 * which allows the scan thread to adjust its workload accordingly.
2317 */
2318 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2319 vd == vd->vdev_top && !vd->vdev_islog &&
2320 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2321 zio->io_txg != spa_syncing_txg(spa)) {
2322 uint64_t old = spa->spa_last_io;
2323 uint64_t new = ddi_get_lbolt64();
2324 if (old != new)
2325 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2326 }
2327
2328 align = 1ULL << vd->vdev_top->vdev_ashift;
2329
2330 if (P2PHASE(zio->io_size, align) != 0) {
2331 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2332 char *abuf = zio_buf_alloc(asize);
2333 ASSERT(vd == vd->vdev_top);
2334 if (zio->io_type == ZIO_TYPE_WRITE) {
2335 bcopy(zio->io_data, abuf, zio->io_size);
2336 bzero(abuf + zio->io_size, asize - zio->io_size);
2337 }
2338 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2339 }
2340
2341 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2342 ASSERT(P2PHASE(zio->io_size, align) == 0);
2343 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2344
2345 /*
2346 * If this is a repair I/O, and there's no self-healing involved --
2347 * that is, we're just resilvering what we expect to resilver --
2348 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2349 * This prevents spurious resilvering with nested replication.
2350 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2351 * A is out of date, we'll read from C+D, then use the data to
2352 * resilver A+B -- but we don't actually want to resilver B, just A.
2353 * The top-level mirror has no way to know this, so instead we just
2354 * discard unnecessary repairs as we work our way down the vdev tree.
2355 * The same logic applies to any form of nested replication:
2356 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2357 */
2358 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2359 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2360 zio->io_txg != 0 && /* not a delegated i/o */
2361 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2362 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2363 zio_vdev_io_bypass(zio);
2364 return (ZIO_PIPELINE_CONTINUE);
2365 }
2366
2367 if (vd->vdev_ops->vdev_op_leaf &&
2368 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2369
2370 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2371 return (ZIO_PIPELINE_CONTINUE);
2372
2373 if ((zio = vdev_queue_io(zio)) == NULL)
2374 return (ZIO_PIPELINE_STOP);
2375
2376 if (!vdev_accessible(vd, zio)) {
2377 zio->io_error = ENXIO;
2378 zio_interrupt(zio);
2379 return (ZIO_PIPELINE_STOP);
2380 }
2381 }
2382
2383 return (vd->vdev_ops->vdev_op_io_start(zio));
2384}
2385
2386static int
2387zio_vdev_io_done(zio_t *zio)
2388{
2389 vdev_t *vd = zio->io_vd;
2390 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2391 boolean_t unexpected_error = B_FALSE;
2392
2393 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2394 return (ZIO_PIPELINE_STOP);
2395
2396 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2397
2398 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2399
2400 vdev_queue_io_done(zio);
2401
2402 if (zio->io_type == ZIO_TYPE_WRITE)
2403 vdev_cache_write(zio);
2404
2405 if (zio_injection_enabled && zio->io_error == 0)
2406 zio->io_error = zio_handle_device_injection(vd,
2407 zio, EIO);
2408
2409 if (zio_injection_enabled && zio->io_error == 0)
2410 zio->io_error = zio_handle_label_injection(zio, EIO);
2411
2412 if (zio->io_error) {
2413 if (!vdev_accessible(vd, zio)) {
2414 zio->io_error = ENXIO;
2415 } else {
2416 unexpected_error = B_TRUE;
2417 }
2418 }
2419 }
2420
2421 ops->vdev_op_io_done(zio);
2422
2423 if (unexpected_error)
2424 VERIFY(vdev_probe(vd, zio) == NULL);
2425
2426 return (ZIO_PIPELINE_CONTINUE);
2427}
2428
2429/*
2430 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2431 * disk, and use that to finish the checksum ereport later.
2432 */
2433static void
2434zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2435 const void *good_buf)
2436{
2437 /* no processing needed */
2438 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2439}
2440
2441/*ARGSUSED*/
2442void
2443zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2444{
2445 void *buf = zio_buf_alloc(zio->io_size);
2446
2447 bcopy(zio->io_data, buf, zio->io_size);
2448
2449 zcr->zcr_cbinfo = zio->io_size;
2450 zcr->zcr_cbdata = buf;
2451 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2452 zcr->zcr_free = zio_buf_free;
2453}
2454
2455static int
2456zio_vdev_io_assess(zio_t *zio)
2457{
2458 vdev_t *vd = zio->io_vd;
2459
2460 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2461 return (ZIO_PIPELINE_STOP);
2462
2463 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2464 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2465
2466 if (zio->io_vsd != NULL) {
2467 zio->io_vsd_ops->vsd_free(zio);
2468 zio->io_vsd = NULL;
2469 }
2470
2471 if (zio_injection_enabled && zio->io_error == 0)
2472 zio->io_error = zio_handle_fault_injection(zio, EIO);
2473
2474 /*
2475 * If the I/O failed, determine whether we should attempt to retry it.
2476 *
2477 * On retry, we cut in line in the issue queue, since we don't want
2478 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2479 */
2480 if (zio->io_error && vd == NULL &&
2481 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2482 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2483 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2484 zio->io_error = 0;
2485 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2486 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2487 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2488 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2489 zio_requeue_io_start_cut_in_line);
2490 return (ZIO_PIPELINE_STOP);
2491 }
2492
2493 /*
2494 * If we got an error on a leaf device, convert it to ENXIO
2495 * if the device is not accessible at all.
2496 */
2497 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2498 !vdev_accessible(vd, zio))
2499 zio->io_error = ENXIO;
2500
2501 /*
2502 * If we can't write to an interior vdev (mirror or RAID-Z),
2503 * set vdev_cant_write so that we stop trying to allocate from it.
2504 */
2505 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2506 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2507 vd->vdev_cant_write = B_TRUE;
2508
2509 if (zio->io_error)
2510 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2511
2512 return (ZIO_PIPELINE_CONTINUE);
2513}
2514
2515void
2516zio_vdev_io_reissue(zio_t *zio)
2517{
2518 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2519 ASSERT(zio->io_error == 0);
2520
2521 zio->io_stage >>= 1;
2522}
2523
2524void
2525zio_vdev_io_redone(zio_t *zio)
2526{
2527 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2528
2529 zio->io_stage >>= 1;
2530}
2531
2532void
2533zio_vdev_io_bypass(zio_t *zio)
2534{
2535 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2536 ASSERT(zio->io_error == 0);
2537
2538 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2539 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2540}
2541
2542/*
2543 * ==========================================================================
2544 * Generate and verify checksums
2545 * ==========================================================================
2546 */
2547static int
2548zio_checksum_generate(zio_t *zio)
2549{
2550 blkptr_t *bp = zio->io_bp;
2551 enum zio_checksum checksum;
2552
2553 if (bp == NULL) {
2554 /*
2555 * This is zio_write_phys().
2556 * We're either generating a label checksum, or none at all.
2557 */
2558 checksum = zio->io_prop.zp_checksum;
2559
2560 if (checksum == ZIO_CHECKSUM_OFF)
2561 return (ZIO_PIPELINE_CONTINUE);
2562
2563 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2564 } else {
2565 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2566 ASSERT(!IO_IS_ALLOCATING(zio));
2567 checksum = ZIO_CHECKSUM_GANG_HEADER;
2568 } else {
2569 checksum = BP_GET_CHECKSUM(bp);
2570 }
2571 }
2572
2573 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2574
2575 return (ZIO_PIPELINE_CONTINUE);
2576}
2577
2578static int
2579zio_checksum_verify(zio_t *zio)
2580{
2581 zio_bad_cksum_t info;
2582 blkptr_t *bp = zio->io_bp;
2583 int error;
2584
2585 ASSERT(zio->io_vd != NULL);
2586
2587 if (bp == NULL) {
2588 /*
2589 * This is zio_read_phys().
2590 * We're either verifying a label checksum, or nothing at all.
2591 */
2592 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2593 return (ZIO_PIPELINE_CONTINUE);
2594
2595 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2596 }
2597
2598 if ((error = zio_checksum_error(zio, &info)) != 0) {
2599 zio->io_error = error;
2600 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2601 zfs_ereport_start_checksum(zio->io_spa,
2602 zio->io_vd, zio, zio->io_offset,
2603 zio->io_size, NULL, &info);
2604 }
2605 }
2606
2607 return (ZIO_PIPELINE_CONTINUE);
2608}
2609
2610/*
2611 * Called by RAID-Z to ensure we don't compute the checksum twice.
2612 */
2613void
2614zio_checksum_verified(zio_t *zio)
2615{
2616 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2617}
2618
2619/*
2620 * ==========================================================================
2621 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2622 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2623 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2624 * indicate errors that are specific to one I/O, and most likely permanent.
2625 * Any other error is presumed to be worse because we weren't expecting it.
2626 * ==========================================================================
2627 */
2628int
2629zio_worst_error(int e1, int e2)
2630{
2631 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2632 int r1, r2;
2633
2634 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2635 if (e1 == zio_error_rank[r1])
2636 break;
2637
2638 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2639 if (e2 == zio_error_rank[r2])
2640 break;
2641
2642 return (r1 > r2 ? e1 : e2);
2643}
2644
2645/*
2646 * ==========================================================================
2647 * I/O completion
2648 * ==========================================================================
2649 */
2650static int
2651zio_ready(zio_t *zio)
2652{
2653 blkptr_t *bp = zio->io_bp;
2654 zio_t *pio, *pio_next;
2655
2656 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2657 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2658 return (ZIO_PIPELINE_STOP);
2659
2660 if (zio->io_ready) {
2661 ASSERT(IO_IS_ALLOCATING(zio));
2662 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2663 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2664
2665 zio->io_ready(zio);
2666 }
2667
2668 if (bp != NULL && bp != &zio->io_bp_copy)
2669 zio->io_bp_copy = *bp;
2670
2671 if (zio->io_error)
2672 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2673
2674 mutex_enter(&zio->io_lock);
2675 zio->io_state[ZIO_WAIT_READY] = 1;
2676 pio = zio_walk_parents(zio);
2677 mutex_exit(&zio->io_lock);
2678
2679 /*
2680 * As we notify zio's parents, new parents could be added.
2681 * New parents go to the head of zio's io_parent_list, however,
2682 * so we will (correctly) not notify them. The remainder of zio's
2683 * io_parent_list, from 'pio_next' onward, cannot change because
2684 * all parents must wait for us to be done before they can be done.
2685 */
2686 for (; pio != NULL; pio = pio_next) {
2687 pio_next = zio_walk_parents(zio);
2688 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2689 }
2690
2691 if (zio->io_flags & ZIO_FLAG_NODATA) {
2692 if (BP_IS_GANG(bp)) {
2693 zio->io_flags &= ~ZIO_FLAG_NODATA;
2694 } else {
2695 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2696 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2697 }
2698 }
2699
2700 if (zio_injection_enabled &&
2701 zio->io_spa->spa_syncing_txg == zio->io_txg)
2702 zio_handle_ignored_writes(zio);
2703
2704 return (ZIO_PIPELINE_CONTINUE);
2705}
2706
2707static int
2708zio_done(zio_t *zio)
2709{
2710 spa_t *spa = zio->io_spa;
2711 zio_t *lio = zio->io_logical;
2712 blkptr_t *bp = zio->io_bp;
2713 vdev_t *vd = zio->io_vd;
2714 uint64_t psize = zio->io_size;
2715 zio_t *pio, *pio_next;
2716
2717 /*
2718 * If our children haven't all completed,
2719 * wait for them and then repeat this pipeline stage.
2720 */
2721 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2722 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2723 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2724 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2725 return (ZIO_PIPELINE_STOP);
2726
2727 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2728 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2729 ASSERT(zio->io_children[c][w] == 0);
2730
2731 if (bp != NULL) {
2732 ASSERT(bp->blk_pad[0] == 0);
2733 ASSERT(bp->blk_pad[1] == 0);
2734 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2735 (bp == zio_unique_parent(zio)->io_bp));
2736 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2737 zio->io_bp_override == NULL &&
2738 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2739 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2740 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2741 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2742 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2743 }
2744 }
2745
2746 /*
2747 * If there were child vdev/gang/ddt errors, they apply to us now.
2748 */
2749 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2750 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2751 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2752
2753 /*
2754 * If the I/O on the transformed data was successful, generate any
2755 * checksum reports now while we still have the transformed data.
2756 */
2757 if (zio->io_error == 0) {
2758 while (zio->io_cksum_report != NULL) {
2759 zio_cksum_report_t *zcr = zio->io_cksum_report;
2760 uint64_t align = zcr->zcr_align;
2761 uint64_t asize = P2ROUNDUP(psize, align);
2762 char *abuf = zio->io_data;
2763
2764 if (asize != psize) {
2765 abuf = zio_buf_alloc(asize);
2766 bcopy(zio->io_data, abuf, psize);
2767 bzero(abuf + psize, asize - psize);
2768 }
2769
2770 zio->io_cksum_report = zcr->zcr_next;
2771 zcr->zcr_next = NULL;
2772 zcr->zcr_finish(zcr, abuf);
2773 zfs_ereport_free_checksum(zcr);
2774
2775 if (asize != psize)
2776 zio_buf_free(abuf, asize);
2777 }
2778 }
2779
2780 zio_pop_transforms(zio); /* note: may set zio->io_error */
2781
2782 vdev_stat_update(zio, psize);
2783
2784 if (zio->io_error) {
2785 /*
2786 * If this I/O is attached to a particular vdev,
2787 * generate an error message describing the I/O failure
2788 * at the block level. We ignore these errors if the
2789 * device is currently unavailable.
2790 */
2791 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2792 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2793
2794 if ((zio->io_error == EIO || !(zio->io_flags &
2795 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2796 zio == lio) {
2797 /*
2798 * For logical I/O requests, tell the SPA to log the
2799 * error and generate a logical data ereport.
2800 */
2801 spa_log_error(spa, zio);
2802 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2803 0, 0);
2804 }
2805 }
2806
2807 if (zio->io_error && zio == lio) {
2808 /*
2809 * Determine whether zio should be reexecuted. This will
2810 * propagate all the way to the root via zio_notify_parent().
2811 */
2812 ASSERT(vd == NULL && bp != NULL);
2813 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2814
2815 if (IO_IS_ALLOCATING(zio) &&
2816 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2817 if (zio->io_error != ENOSPC)
2818 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2819 else
2820 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2821 }
2822
2823 if ((zio->io_type == ZIO_TYPE_READ ||
2824 zio->io_type == ZIO_TYPE_FREE) &&
2825 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2826 zio->io_error == ENXIO &&
2827 spa_load_state(spa) == SPA_LOAD_NONE &&
2828 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2829 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2830
2831 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2832 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2833
2834 /*
2835 * Here is a possibly good place to attempt to do
2836 * either combinatorial reconstruction or error correction
2837 * based on checksums. It also might be a good place
2838 * to send out preliminary ereports before we suspend
2839 * processing.
2840 */
2841 }
2842
2843 /*
2844 * If there were logical child errors, they apply to us now.
2845 * We defer this until now to avoid conflating logical child
2846 * errors with errors that happened to the zio itself when
2847 * updating vdev stats and reporting FMA events above.
2848 */
2849 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2850
2851 if ((zio->io_error || zio->io_reexecute) &&
2852 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2853 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2854 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2855
2856 zio_gang_tree_free(&zio->io_gang_tree);
2857
2858 /*
2859 * Godfather I/Os should never suspend.
2860 */
2861 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2862 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2863 zio->io_reexecute = 0;
2864
2865 if (zio->io_reexecute) {
2866 /*
2867 * This is a logical I/O that wants to reexecute.
2868 *
2869 * Reexecute is top-down. When an i/o fails, if it's not
2870 * the root, it simply notifies its parent and sticks around.
2871 * The parent, seeing that it still has children in zio_done(),
2872 * does the same. This percolates all the way up to the root.
2873 * The root i/o will reexecute or suspend the entire tree.
2874 *
2875 * This approach ensures that zio_reexecute() honors
2876 * all the original i/o dependency relationships, e.g.
2877 * parents not executing until children are ready.
2878 */
2879 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2880
2881 zio->io_gang_leader = NULL;
2882
2883 mutex_enter(&zio->io_lock);
2884 zio->io_state[ZIO_WAIT_DONE] = 1;
2885 mutex_exit(&zio->io_lock);
2886
2887 /*
2888 * "The Godfather" I/O monitors its children but is
2889 * not a true parent to them. It will track them through
2890 * the pipeline but severs its ties whenever they get into
2891 * trouble (e.g. suspended). This allows "The Godfather"
2892 * I/O to return status without blocking.
2893 */
2894 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2895 zio_link_t *zl = zio->io_walk_link;
2896 pio_next = zio_walk_parents(zio);
2897
2898 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2899 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2900 zio_remove_child(pio, zio, zl);
2901 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2902 }
2903 }
2904
2905 if ((pio = zio_unique_parent(zio)) != NULL) {
2906 /*
2907 * We're not a root i/o, so there's nothing to do
2908 * but notify our parent. Don't propagate errors
2909 * upward since we haven't permanently failed yet.
2910 */
2911 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2912 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2913 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2914 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2915 /*
2916 * We'd fail again if we reexecuted now, so suspend
2917 * until conditions improve (e.g. device comes online).
2918 */
2919 zio_suspend(spa, zio);
2920 } else {
2921 /*
2922 * Reexecution is potentially a huge amount of work.
2923 * Hand it off to the otherwise-unused claim taskq.
2924 */
2925#ifdef _KERNEL
2926 (void) taskq_dispatch_safe(
2927 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2928 (task_func_t *)zio_reexecute, zio, TQ_SLEEP,
2929 &zio->io_task);
2930#else
2931 (void) taskq_dispatch(
2932 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2933 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2934#endif
2935 }
2936 return (ZIO_PIPELINE_STOP);
2937 }
2938
2939 ASSERT(zio->io_child_count == 0);
2940 ASSERT(zio->io_reexecute == 0);
2941 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2942
2943 /*
2944 * Report any checksum errors, since the I/O is complete.
2945 */
2946 while (zio->io_cksum_report != NULL) {
2947 zio_cksum_report_t *zcr = zio->io_cksum_report;
2948 zio->io_cksum_report = zcr->zcr_next;
2949 zcr->zcr_next = NULL;
2950 zcr->zcr_finish(zcr, NULL);
2951 zfs_ereport_free_checksum(zcr);
2952 }
2953
2954 /*
2955 * It is the responsibility of the done callback to ensure that this
2956 * particular zio is no longer discoverable for adoption, and as
2957 * such, cannot acquire any new parents.
2958 */
2959 if (zio->io_done)
2960 zio->io_done(zio);
2961
2962 mutex_enter(&zio->io_lock);
2963 zio->io_state[ZIO_WAIT_DONE] = 1;
2964 mutex_exit(&zio->io_lock);
2965
2966 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2967 zio_link_t *zl = zio->io_walk_link;
2968 pio_next = zio_walk_parents(zio);
2969 zio_remove_child(pio, zio, zl);
2970 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2971 }
2972
2973 if (zio->io_waiter != NULL) {
2974 mutex_enter(&zio->io_lock);
2975 zio->io_executor = NULL;
2976 cv_broadcast(&zio->io_cv);
2977 mutex_exit(&zio->io_lock);
2978 } else {
2979 zio_destroy(zio);
2980 }
2981
2982 return (ZIO_PIPELINE_STOP);
2983}
2984
2985/*
2986 * ==========================================================================
2987 * I/O pipeline definition
2988 * ==========================================================================
2989 */
2990static zio_pipe_stage_t *zio_pipeline[] = {
2991 NULL,
2992 zio_read_bp_init,
2993 zio_free_bp_init,
2994 zio_issue_async,
2995 zio_write_bp_init,
2996 zio_checksum_generate,
2997 zio_ddt_read_start,
2998 zio_ddt_read_done,
2999 zio_ddt_write,
3000 zio_ddt_free,
3001 zio_gang_assemble,
3002 zio_gang_issue,
3003 zio_dva_allocate,
3004 zio_dva_free,
3005 zio_dva_claim,
3006 zio_ready,
3007 zio_vdev_io_start,
3008 zio_vdev_io_done,
3009 zio_vdev_io_assess,
3010 zio_checksum_verify,
3011 zio_done
3012};
|