genfs_io.c revision 1.102
1/* $NetBSD: genfs_io.c,v 1.102 2022/01/14 21:59:50 riastradh Exp $ */ 2 3/* 4 * Copyright (c) 1982, 1986, 1989, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. Neither the name of the University nor the names of its contributors 16 * may be used to endorse or promote products derived from this software 17 * without specific prior written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 * 31 */ 32 33#include <sys/cdefs.h> 34__KERNEL_RCSID(0, "$NetBSD: genfs_io.c,v 1.102 2022/01/14 21:59:50 riastradh Exp $"); 35 36#include <sys/param.h> 37#include <sys/systm.h> 38#include <sys/proc.h> 39#include <sys/kernel.h> 40#include <sys/mount.h> 41#include <sys/vnode.h> 42#include <sys/kmem.h> 43#include <sys/kauth.h> 44#include <sys/fstrans.h> 45#include <sys/buf.h> 46#include <sys/atomic.h> 47 48#include <miscfs/genfs/genfs.h> 49#include <miscfs/genfs/genfs_node.h> 50#include <miscfs/specfs/specdev.h> 51 52#include <uvm/uvm.h> 53#include <uvm/uvm_pager.h> 54#include <uvm/uvm_page_array.h> 55 56static int genfs_do_directio(struct vmspace *, vaddr_t, size_t, struct vnode *, 57 off_t, enum uio_rw); 58static void genfs_dio_iodone(struct buf *); 59 60static int genfs_getpages_read(struct vnode *, struct vm_page **, int, off_t, 61 off_t, bool, bool, bool, bool); 62static int genfs_do_io(struct vnode *, off_t, vaddr_t, size_t, int, enum uio_rw, 63 void (*)(struct buf *)); 64static void genfs_rel_pages(struct vm_page **, unsigned int); 65 66int genfs_maxdio = MAXPHYS; 67 68static void 69genfs_rel_pages(struct vm_page **pgs, unsigned int npages) 70{ 71 unsigned int i; 72 73 for (i = 0; i < npages; i++) { 74 struct vm_page *pg = pgs[i]; 75 76 if (pg == NULL || pg == PGO_DONTCARE) 77 continue; 78 KASSERT(uvm_page_owner_locked_p(pg, true)); 79 if (pg->flags & PG_FAKE) { 80 pg->flags |= PG_RELEASED; 81 } 82 } 83 uvm_page_unbusy(pgs, npages); 84} 85 86/* 87 * generic VM getpages routine. 88 * Return PG_BUSY pages for the given range, 89 * reading from backing store if necessary. 90 */ 91 92int 93genfs_getpages(void *v) 94{ 95 struct vop_getpages_args /* { 96 struct vnode *a_vp; 97 voff_t a_offset; 98 struct vm_page **a_m; 99 int *a_count; 100 int a_centeridx; 101 vm_prot_t a_access_type; 102 int a_advice; 103 int a_flags; 104 } */ * const ap = v; 105 106 off_t diskeof, memeof; 107 int i, error, npages, iflag; 108 const int flags = ap->a_flags; 109 struct vnode * const vp = ap->a_vp; 110 struct uvm_object * const uobj = &vp->v_uobj; 111 const bool async = (flags & PGO_SYNCIO) == 0; 112 const bool memwrite = (ap->a_access_type & VM_PROT_WRITE) != 0; 113 const bool overwrite = (flags & PGO_OVERWRITE) != 0; 114 const bool blockalloc = memwrite && (flags & PGO_NOBLOCKALLOC) == 0; 115 const bool need_wapbl = (vp->v_mount->mnt_wapbl && 116 (flags & PGO_JOURNALLOCKED) == 0); 117 const bool glocked = (flags & PGO_GLOCKHELD) != 0; 118 bool holds_wapbl = false; 119 struct mount *trans_mount = NULL; 120 UVMHIST_FUNC("genfs_getpages"); UVMHIST_CALLED(ubchist); 121 122 UVMHIST_LOG(ubchist, "vp %#jx off 0x%jx/%jx count %jd", 123 (uintptr_t)vp, ap->a_offset >> 32, ap->a_offset, *ap->a_count); 124 125 KASSERT(memwrite >= overwrite); 126 KASSERT(vp->v_type == VREG || vp->v_type == VDIR || 127 vp->v_type == VLNK || vp->v_type == VBLK); 128 129 /* 130 * the object must be locked. it can only be a read lock when 131 * processing a read fault with PGO_LOCKED. 132 */ 133 134 KASSERT(rw_lock_held(uobj->vmobjlock)); 135 KASSERT(rw_write_held(uobj->vmobjlock) || 136 ((flags & PGO_LOCKED) != 0 && !memwrite)); 137 138#ifdef DIAGNOSTIC 139 if ((flags & PGO_JOURNALLOCKED) && vp->v_mount->mnt_wapbl) 140 WAPBL_JLOCK_ASSERT(vp->v_mount); 141#endif 142 143 /* 144 * check for reclaimed vnode. v_interlock is not held here, but 145 * VI_DEADCHECK is set with vmobjlock held. 146 */ 147 148 iflag = atomic_load_relaxed(&vp->v_iflag); 149 if (__predict_false((iflag & VI_DEADCHECK) != 0)) { 150 mutex_enter(vp->v_interlock); 151 error = vdead_check(vp, VDEAD_NOWAIT); 152 mutex_exit(vp->v_interlock); 153 if (error) { 154 if ((flags & PGO_LOCKED) == 0) 155 rw_exit(uobj->vmobjlock); 156 return error; 157 } 158 } 159 160startover: 161 error = 0; 162 const voff_t origvsize = vp->v_size; 163 const off_t origoffset = ap->a_offset; 164 const int orignpages = *ap->a_count; 165 166 GOP_SIZE(vp, origvsize, &diskeof, 0); 167 if (flags & PGO_PASTEOF) { 168 off_t newsize; 169#if defined(DIAGNOSTIC) 170 off_t writeeof; 171#endif /* defined(DIAGNOSTIC) */ 172 173 newsize = MAX(origvsize, 174 origoffset + (orignpages << PAGE_SHIFT)); 175 GOP_SIZE(vp, newsize, &memeof, GOP_SIZE_MEM); 176#if defined(DIAGNOSTIC) 177 GOP_SIZE(vp, vp->v_writesize, &writeeof, GOP_SIZE_MEM); 178 if (newsize > round_page(writeeof)) { 179 panic("%s: past eof: %" PRId64 " vs. %" PRId64, 180 __func__, newsize, round_page(writeeof)); 181 } 182#endif /* defined(DIAGNOSTIC) */ 183 } else { 184 GOP_SIZE(vp, origvsize, &memeof, GOP_SIZE_MEM); 185 } 186 KASSERT(ap->a_centeridx >= 0 || ap->a_centeridx <= orignpages); 187 KASSERT((origoffset & (PAGE_SIZE - 1)) == 0 && origoffset >= 0); 188 KASSERT(orignpages > 0); 189 190 /* 191 * Bounds-check the request. 192 */ 193 194 if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= memeof) { 195 if ((flags & PGO_LOCKED) == 0) { 196 rw_exit(uobj->vmobjlock); 197 } 198 UVMHIST_LOG(ubchist, "off 0x%jx count %jd goes past EOF 0x%jx", 199 origoffset, *ap->a_count, memeof,0); 200 error = EINVAL; 201 goto out_err; 202 } 203 204 /* uobj is locked */ 205 206 if ((flags & PGO_NOTIMESTAMP) == 0 && 207 (vp->v_type != VBLK || 208 (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) { 209 int updflags = 0; 210 211 if ((vp->v_mount->mnt_flag & MNT_NOATIME) == 0) { 212 updflags = GOP_UPDATE_ACCESSED; 213 } 214 if (memwrite) { 215 updflags |= GOP_UPDATE_MODIFIED; 216 } 217 if (updflags != 0) { 218 GOP_MARKUPDATE(vp, updflags); 219 } 220 } 221 222 /* 223 * For PGO_LOCKED requests, just return whatever's in memory. 224 */ 225 226 if (flags & PGO_LOCKED) { 227 int nfound; 228 struct vm_page *pg; 229 230 KASSERT(!glocked); 231 npages = *ap->a_count; 232#if defined(DEBUG) 233 for (i = 0; i < npages; i++) { 234 pg = ap->a_m[i]; 235 KASSERT(pg == NULL || pg == PGO_DONTCARE); 236 } 237#endif /* defined(DEBUG) */ 238 nfound = uvn_findpages(uobj, origoffset, &npages, 239 ap->a_m, NULL, 240 UFP_NOWAIT | UFP_NOALLOC | UFP_NOBUSY | 241 (memwrite ? UFP_NORDONLY : 0)); 242 KASSERT(npages == *ap->a_count); 243 if (nfound == 0) { 244 error = EBUSY; 245 goto out_err; 246 } 247 /* 248 * lock and unlock g_glock to ensure that no one is truncating 249 * the file behind us. 250 */ 251 if (!genfs_node_rdtrylock(vp)) { 252 /* 253 * restore the array. 254 */ 255 256 for (i = 0; i < npages; i++) { 257 pg = ap->a_m[i]; 258 259 if (pg != NULL && pg != PGO_DONTCARE) { 260 ap->a_m[i] = NULL; 261 } 262 KASSERT(ap->a_m[i] == NULL || 263 ap->a_m[i] == PGO_DONTCARE); 264 } 265 } else { 266 genfs_node_unlock(vp); 267 } 268 error = (ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0); 269 if (error == 0 && memwrite) { 270 for (i = 0; i < npages; i++) { 271 pg = ap->a_m[i]; 272 if (pg == NULL || pg == PGO_DONTCARE) { 273 continue; 274 } 275 if (uvm_pagegetdirty(pg) == 276 UVM_PAGE_STATUS_CLEAN) { 277 uvm_pagemarkdirty(pg, 278 UVM_PAGE_STATUS_UNKNOWN); 279 } 280 } 281 } 282 goto out_err; 283 } 284 rw_exit(uobj->vmobjlock); 285 286 /* 287 * find the requested pages and make some simple checks. 288 * leave space in the page array for a whole block. 289 */ 290 291 const int fs_bshift = (vp->v_type != VBLK) ? 292 vp->v_mount->mnt_fs_bshift : DEV_BSHIFT; 293 const int fs_bsize = 1 << fs_bshift; 294#define blk_mask (fs_bsize - 1) 295#define trunc_blk(x) ((x) & ~blk_mask) 296#define round_blk(x) (((x) + blk_mask) & ~blk_mask) 297 298 const int orignmempages = MIN(orignpages, 299 round_page(memeof - origoffset) >> PAGE_SHIFT); 300 npages = orignmempages; 301 const off_t startoffset = trunc_blk(origoffset); 302 const off_t endoffset = MIN( 303 round_page(round_blk(origoffset + (npages << PAGE_SHIFT))), 304 round_page(memeof)); 305 const int ridx = (origoffset - startoffset) >> PAGE_SHIFT; 306 307 const int pgs_size = sizeof(struct vm_page *) * 308 ((endoffset - startoffset) >> PAGE_SHIFT); 309 struct vm_page **pgs, *pgs_onstack[UBC_MAX_PAGES]; 310 311 if (pgs_size > sizeof(pgs_onstack)) { 312 pgs = kmem_zalloc(pgs_size, async ? KM_NOSLEEP : KM_SLEEP); 313 if (pgs == NULL) { 314 pgs = pgs_onstack; 315 error = ENOMEM; 316 goto out_err; 317 } 318 } else { 319 pgs = pgs_onstack; 320 (void)memset(pgs, 0, pgs_size); 321 } 322 323 UVMHIST_LOG(ubchist, "ridx %jd npages %jd startoff %#jx endoff %#jx", 324 ridx, npages, startoffset, endoffset); 325 326 if (trans_mount == NULL) { 327 trans_mount = vp->v_mount; 328 fstrans_start(trans_mount); 329 /* 330 * check if this vnode is still valid. 331 */ 332 mutex_enter(vp->v_interlock); 333 error = vdead_check(vp, 0); 334 mutex_exit(vp->v_interlock); 335 if (error) 336 goto out_err_free; 337 /* 338 * XXX: This assumes that we come here only via 339 * the mmio path 340 */ 341 if (blockalloc && need_wapbl) { 342 error = WAPBL_BEGIN(trans_mount); 343 if (error) 344 goto out_err_free; 345 holds_wapbl = true; 346 } 347 } 348 349 /* 350 * hold g_glock to prevent a race with truncate. 351 * 352 * check if our idea of v_size is still valid. 353 */ 354 355 KASSERT(!glocked || genfs_node_wrlocked(vp)); 356 if (!glocked) { 357 if (blockalloc) { 358 genfs_node_wrlock(vp); 359 } else { 360 genfs_node_rdlock(vp); 361 } 362 } 363 rw_enter(uobj->vmobjlock, RW_WRITER); 364 if (vp->v_size < origvsize) { 365 if (!glocked) { 366 genfs_node_unlock(vp); 367 } 368 if (pgs != pgs_onstack) 369 kmem_free(pgs, pgs_size); 370 goto startover; 371 } 372 373 if (uvn_findpages(uobj, origoffset, &npages, &pgs[ridx], NULL, 374 async ? UFP_NOWAIT : UFP_ALL) != orignmempages) { 375 if (!glocked) { 376 genfs_node_unlock(vp); 377 } 378 KASSERT(async != 0); 379 genfs_rel_pages(&pgs[ridx], orignmempages); 380 rw_exit(uobj->vmobjlock); 381 error = EBUSY; 382 goto out_err_free; 383 } 384 385 /* 386 * if PGO_OVERWRITE is set, don't bother reading the pages. 387 */ 388 389 if (overwrite) { 390 if (!glocked) { 391 genfs_node_unlock(vp); 392 } 393 UVMHIST_LOG(ubchist, "PGO_OVERWRITE",0,0,0,0); 394 395 for (i = 0; i < npages; i++) { 396 struct vm_page *pg = pgs[ridx + i]; 397 398 /* 399 * it's caller's responsibility to allocate blocks 400 * beforehand for the overwrite case. 401 */ 402 403 KASSERT((pg->flags & PG_RDONLY) == 0 || !blockalloc); 404 pg->flags &= ~PG_RDONLY; 405 406 /* 407 * mark the page DIRTY. 408 * otherwise another thread can do putpages and pull 409 * our vnode from syncer's queue before our caller does 410 * ubc_release. note that putpages won't see CLEAN 411 * pages even if they are BUSY. 412 */ 413 414 uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY); 415 } 416 npages += ridx; 417 goto out; 418 } 419 420 /* 421 * if the pages are already resident, just return them. 422 */ 423 424 for (i = 0; i < npages; i++) { 425 struct vm_page *pg = pgs[ridx + i]; 426 427 if ((pg->flags & PG_FAKE) || 428 (blockalloc && (pg->flags & PG_RDONLY) != 0)) { 429 break; 430 } 431 } 432 if (i == npages) { 433 if (!glocked) { 434 genfs_node_unlock(vp); 435 } 436 UVMHIST_LOG(ubchist, "returning cached pages", 0,0,0,0); 437 npages += ridx; 438 goto out; 439 } 440 441 /* 442 * the page wasn't resident and we're not overwriting, 443 * so we're going to have to do some i/o. 444 * find any additional pages needed to cover the expanded range. 445 */ 446 447 npages = (endoffset - startoffset) >> PAGE_SHIFT; 448 if (startoffset != origoffset || npages != orignmempages) { 449 int npgs; 450 451 /* 452 * we need to avoid deadlocks caused by locking 453 * additional pages at lower offsets than pages we 454 * already have locked. unlock them all and start over. 455 */ 456 457 genfs_rel_pages(&pgs[ridx], orignmempages); 458 memset(pgs, 0, pgs_size); 459 460 UVMHIST_LOG(ubchist, "reset npages start 0x%jx end 0x%jx", 461 startoffset, endoffset, 0,0); 462 npgs = npages; 463 if (uvn_findpages(uobj, startoffset, &npgs, pgs, NULL, 464 async ? UFP_NOWAIT : UFP_ALL) != npages) { 465 if (!glocked) { 466 genfs_node_unlock(vp); 467 } 468 KASSERT(async != 0); 469 genfs_rel_pages(pgs, npages); 470 rw_exit(uobj->vmobjlock); 471 error = EBUSY; 472 goto out_err_free; 473 } 474 } 475 476 rw_exit(uobj->vmobjlock); 477 error = genfs_getpages_read(vp, pgs, npages, startoffset, diskeof, 478 async, memwrite, blockalloc, glocked); 479 if (!glocked) { 480 genfs_node_unlock(vp); 481 } 482 if (error == 0 && async) 483 goto out_err_free; 484 rw_enter(uobj->vmobjlock, RW_WRITER); 485 486 /* 487 * we're almost done! release the pages... 488 * for errors, we free the pages. 489 * otherwise we activate them and mark them as valid and clean. 490 * also, unbusy pages that were not actually requested. 491 */ 492 493 if (error) { 494 genfs_rel_pages(pgs, npages); 495 rw_exit(uobj->vmobjlock); 496 UVMHIST_LOG(ubchist, "returning error %jd", error,0,0,0); 497 goto out_err_free; 498 } 499 500out: 501 UVMHIST_LOG(ubchist, "succeeding, npages %jd", npages,0,0,0); 502 error = 0; 503 for (i = 0; i < npages; i++) { 504 struct vm_page *pg = pgs[i]; 505 if (pg == NULL) { 506 continue; 507 } 508 UVMHIST_LOG(ubchist, "examining pg %#jx flags 0x%jx", 509 (uintptr_t)pg, pg->flags, 0,0); 510 if (pg->flags & PG_FAKE && !overwrite) { 511 /* 512 * we've read page's contents from the backing storage. 513 * 514 * for a read fault, we keep them CLEAN; if we 515 * encountered a hole while reading, the pages can 516 * already been dirtied with zeros. 517 */ 518 KASSERTMSG(blockalloc || uvm_pagegetdirty(pg) == 519 UVM_PAGE_STATUS_CLEAN, "page %p not clean", pg); 520 pg->flags &= ~PG_FAKE; 521 } 522 KASSERT(!memwrite || !blockalloc || (pg->flags & PG_RDONLY) == 0); 523 if (i < ridx || i >= ridx + orignmempages || async) { 524 UVMHIST_LOG(ubchist, "unbusy pg %#jx offset 0x%jx", 525 (uintptr_t)pg, pg->offset,0,0); 526 if (pg->flags & PG_FAKE) { 527 KASSERT(overwrite); 528 uvm_pagezero(pg); 529 } 530 if (pg->flags & PG_RELEASED) { 531 uvm_pagefree(pg); 532 continue; 533 } 534 uvm_pagelock(pg); 535 uvm_pageenqueue(pg); 536 uvm_pagewakeup(pg); 537 uvm_pageunlock(pg); 538 pg->flags &= ~(PG_BUSY|PG_FAKE); 539 UVM_PAGE_OWN(pg, NULL); 540 } else if (memwrite && !overwrite && 541 uvm_pagegetdirty(pg) == UVM_PAGE_STATUS_CLEAN) { 542 /* 543 * for a write fault, start dirtiness tracking of 544 * requested pages. 545 */ 546 uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_UNKNOWN); 547 } 548 } 549 rw_exit(uobj->vmobjlock); 550 if (ap->a_m != NULL) { 551 memcpy(ap->a_m, &pgs[ridx], 552 orignmempages * sizeof(struct vm_page *)); 553 } 554 555out_err_free: 556 if (pgs != NULL && pgs != pgs_onstack) 557 kmem_free(pgs, pgs_size); 558out_err: 559 if (trans_mount != NULL) { 560 if (holds_wapbl) 561 WAPBL_END(trans_mount); 562 fstrans_done(trans_mount); 563 } 564 return error; 565} 566 567/* 568 * genfs_getpages_read: Read the pages in with VOP_BMAP/VOP_STRATEGY. 569 * 570 * "glocked" (which is currently not actually used) tells us not whether 571 * the genfs_node is locked on entry (it always is) but whether it was 572 * locked on entry to genfs_getpages. 573 */ 574static int 575genfs_getpages_read(struct vnode *vp, struct vm_page **pgs, int npages, 576 off_t startoffset, off_t diskeof, 577 bool async, bool memwrite, bool blockalloc, bool glocked) 578{ 579 struct uvm_object * const uobj = &vp->v_uobj; 580 const int fs_bshift = (vp->v_type != VBLK) ? 581 vp->v_mount->mnt_fs_bshift : DEV_BSHIFT; 582 const int dev_bshift = (vp->v_type != VBLK) ? 583 vp->v_mount->mnt_dev_bshift : DEV_BSHIFT; 584 kauth_cred_t const cred = curlwp->l_cred; /* XXXUBC curlwp */ 585 size_t bytes, iobytes, tailstart, tailbytes, totalbytes, skipbytes; 586 vaddr_t kva; 587 struct buf *bp, *mbp; 588 bool sawhole = false; 589 int i; 590 int error = 0; 591 592 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 593 594 /* 595 * read the desired page(s). 596 */ 597 598 totalbytes = npages << PAGE_SHIFT; 599 bytes = MIN(totalbytes, MAX(diskeof - startoffset, 0)); 600 tailbytes = totalbytes - bytes; 601 skipbytes = 0; 602 603 kva = uvm_pagermapin(pgs, npages, 604 UVMPAGER_MAPIN_READ | (async ? 0 : UVMPAGER_MAPIN_WAITOK)); 605 if (kva == 0) 606 return EBUSY; 607 608 mbp = getiobuf(vp, true); 609 mbp->b_bufsize = totalbytes; 610 mbp->b_data = (void *)kva; 611 mbp->b_resid = mbp->b_bcount = bytes; 612 mbp->b_cflags |= BC_BUSY; 613 if (async) { 614 mbp->b_flags = B_READ | B_ASYNC; 615 mbp->b_iodone = uvm_aio_aiodone; 616 } else { 617 mbp->b_flags = B_READ; 618 mbp->b_iodone = NULL; 619 } 620 if (async) 621 BIO_SETPRIO(mbp, BPRIO_TIMELIMITED); 622 else 623 BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL); 624 625 /* 626 * if EOF is in the middle of the range, zero the part past EOF. 627 * skip over pages which are not PG_FAKE since in that case they have 628 * valid data that we need to preserve. 629 */ 630 631 tailstart = bytes; 632 while (tailbytes > 0) { 633 const int len = PAGE_SIZE - (tailstart & PAGE_MASK); 634 635 KASSERT(len <= tailbytes); 636 if ((pgs[tailstart >> PAGE_SHIFT]->flags & PG_FAKE) != 0) { 637 memset((void *)(kva + tailstart), 0, len); 638 UVMHIST_LOG(ubchist, "tailbytes %#jx 0x%jx 0x%jx", 639 (uintptr_t)kva, tailstart, len, 0); 640 } 641 tailstart += len; 642 tailbytes -= len; 643 } 644 645 /* 646 * now loop over the pages, reading as needed. 647 */ 648 649 bp = NULL; 650 off_t offset; 651 for (offset = startoffset; 652 bytes > 0; 653 offset += iobytes, bytes -= iobytes) { 654 int run; 655 daddr_t lbn, blkno; 656 int pidx; 657 struct vnode *devvp; 658 659 /* 660 * skip pages which don't need to be read. 661 */ 662 663 pidx = (offset - startoffset) >> PAGE_SHIFT; 664 while ((pgs[pidx]->flags & PG_FAKE) == 0) { 665 size_t b; 666 667 KASSERT((offset & (PAGE_SIZE - 1)) == 0); 668 if ((pgs[pidx]->flags & PG_RDONLY)) { 669 sawhole = true; 670 } 671 b = MIN(PAGE_SIZE, bytes); 672 offset += b; 673 bytes -= b; 674 skipbytes += b; 675 pidx++; 676 UVMHIST_LOG(ubchist, "skipping, new offset 0x%jx", 677 offset, 0,0,0); 678 if (bytes == 0) { 679 goto loopdone; 680 } 681 } 682 683 /* 684 * bmap the file to find out the blkno to read from and 685 * how much we can read in one i/o. if bmap returns an error, 686 * skip the rest of the top-level i/o. 687 */ 688 689 lbn = offset >> fs_bshift; 690 error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run); 691 if (error) { 692 UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%jx -> %jd", 693 lbn,error,0,0); 694 skipbytes += bytes; 695 bytes = 0; 696 goto loopdone; 697 } 698 699 /* 700 * see how many pages can be read with this i/o. 701 * reduce the i/o size if necessary to avoid 702 * overwriting pages with valid data. 703 */ 704 705 iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset, 706 bytes); 707 if (offset + iobytes > round_page(offset)) { 708 int pcount; 709 710 pcount = 1; 711 while (pidx + pcount < npages && 712 pgs[pidx + pcount]->flags & PG_FAKE) { 713 pcount++; 714 } 715 iobytes = MIN(iobytes, (pcount << PAGE_SHIFT) - 716 (offset - trunc_page(offset))); 717 } 718 719 /* 720 * if this block isn't allocated, zero it instead of 721 * reading it. unless we are going to allocate blocks, 722 * mark the pages we zeroed PG_RDONLY. 723 */ 724 725 if (blkno == (daddr_t)-1) { 726 int holepages = (round_page(offset + iobytes) - 727 trunc_page(offset)) >> PAGE_SHIFT; 728 UVMHIST_LOG(ubchist, "lbn 0x%jx -> HOLE", lbn,0,0,0); 729 730 sawhole = true; 731 memset((char *)kva + (offset - startoffset), 0, 732 iobytes); 733 skipbytes += iobytes; 734 735 if (!blockalloc) { 736 rw_enter(uobj->vmobjlock, RW_WRITER); 737 for (i = 0; i < holepages; i++) { 738 pgs[pidx + i]->flags |= PG_RDONLY; 739 } 740 rw_exit(uobj->vmobjlock); 741 } 742 continue; 743 } 744 745 /* 746 * allocate a sub-buf for this piece of the i/o 747 * (or just use mbp if there's only 1 piece), 748 * and start it going. 749 */ 750 751 if (offset == startoffset && iobytes == bytes) { 752 bp = mbp; 753 } else { 754 UVMHIST_LOG(ubchist, "vp %#jx bp %#jx num now %jd", 755 (uintptr_t)vp, (uintptr_t)bp, vp->v_numoutput, 0); 756 bp = getiobuf(vp, true); 757 nestiobuf_setup(mbp, bp, offset - startoffset, iobytes); 758 } 759 bp->b_lblkno = 0; 760 761 /* adjust physical blkno for partial blocks */ 762 bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >> 763 dev_bshift); 764 765 UVMHIST_LOG(ubchist, 766 "bp %#jx offset 0x%x bcount 0x%x blkno 0x%x", 767 (uintptr_t)bp, offset, bp->b_bcount, bp->b_blkno); 768 769 VOP_STRATEGY(devvp, bp); 770 } 771 772loopdone: 773 nestiobuf_done(mbp, skipbytes, error); 774 if (async) { 775 UVMHIST_LOG(ubchist, "returning 0 (async)",0,0,0,0); 776 return 0; 777 } 778 if (bp != NULL) { 779 error = biowait(mbp); 780 } 781 782 /* Remove the mapping (make KVA available as soon as possible) */ 783 uvm_pagermapout(kva, npages); 784 785 /* 786 * if this we encountered a hole then we have to do a little more work. 787 * for read faults, we marked the page PG_RDONLY so that future 788 * write accesses to the page will fault again. 789 * for write faults, we must make sure that the backing store for 790 * the page is completely allocated while the pages are locked. 791 */ 792 793 if (!error && sawhole && blockalloc) { 794 error = GOP_ALLOC(vp, startoffset, 795 npages << PAGE_SHIFT, 0, cred); 796 UVMHIST_LOG(ubchist, "gop_alloc off 0x%jx/0x%jx -> %jd", 797 startoffset, npages << PAGE_SHIFT, error,0); 798 if (!error) { 799 rw_enter(uobj->vmobjlock, RW_WRITER); 800 for (i = 0; i < npages; i++) { 801 struct vm_page *pg = pgs[i]; 802 803 if (pg == NULL) { 804 continue; 805 } 806 pg->flags &= ~PG_RDONLY; 807 uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY); 808 UVMHIST_LOG(ubchist, "mark dirty pg %#jx", 809 (uintptr_t)pg, 0, 0, 0); 810 } 811 rw_exit(uobj->vmobjlock); 812 } 813 } 814 815 putiobuf(mbp); 816 return error; 817} 818 819/* 820 * generic VM putpages routine. 821 * Write the given range of pages to backing store. 822 * 823 * => "offhi == 0" means flush all pages at or after "offlo". 824 * => object should be locked by caller. we return with the 825 * object unlocked. 826 * => if PGO_CLEANIT or PGO_SYNCIO is set, we may block (due to I/O). 827 * thus, a caller might want to unlock higher level resources 828 * (e.g. vm_map) before calling flush. 829 * => if neither PGO_CLEANIT nor PGO_SYNCIO is set, we will not block 830 * => if PGO_ALLPAGES is set, then all pages in the object will be processed. 831 * 832 * note on "cleaning" object and PG_BUSY pages: 833 * this routine is holding the lock on the object. the only time 834 * that it can run into a PG_BUSY page that it does not own is if 835 * some other process has started I/O on the page (e.g. either 836 * a pagein, or a pageout). if the PG_BUSY page is being paged 837 * in, then it can not be dirty (!UVM_PAGE_STATUS_CLEAN) because no 838 * one has had a chance to modify it yet. if the PG_BUSY page is 839 * being paged out then it means that someone else has already started 840 * cleaning the page for us (how nice!). in this case, if we 841 * have syncio specified, then after we make our pass through the 842 * object we need to wait for the other PG_BUSY pages to clear 843 * off (i.e. we need to do an iosync). also note that once a 844 * page is PG_BUSY it must stay in its object until it is un-busyed. 845 */ 846 847int 848genfs_putpages(void *v) 849{ 850 struct vop_putpages_args /* { 851 struct vnode *a_vp; 852 voff_t a_offlo; 853 voff_t a_offhi; 854 int a_flags; 855 } */ * const ap = v; 856 857 return genfs_do_putpages(ap->a_vp, ap->a_offlo, ap->a_offhi, 858 ap->a_flags, NULL); 859} 860 861int 862genfs_do_putpages(struct vnode *vp, off_t startoff, off_t endoff, 863 int origflags, struct vm_page **busypg) 864{ 865 struct uvm_object * const uobj = &vp->v_uobj; 866 krwlock_t * const slock = uobj->vmobjlock; 867 off_t nextoff; 868 int i, error, npages, nback; 869 int freeflag; 870 /* 871 * This array is larger than it should so that it's size is constant. 872 * The right size is MAXPAGES. 873 */ 874 struct vm_page *pgs[MAXPHYS / MIN_PAGE_SIZE]; 875#define MAXPAGES (MAXPHYS / PAGE_SIZE) 876 struct vm_page *pg, *tpg; 877 struct uvm_page_array a; 878 bool wasclean, needs_clean; 879 bool async = (origflags & PGO_SYNCIO) == 0; 880 bool pagedaemon = curlwp == uvm.pagedaemon_lwp; 881 struct mount *trans_mp; 882 int flags; 883 bool modified; /* if we write out any pages */ 884 bool holds_wapbl; 885 bool cleanall; /* try to pull off from the syncer's list */ 886 bool onworklst; 887 bool nodirty; 888 const bool dirtyonly = (origflags & (PGO_DEACTIVATE|PGO_FREE)) == 0; 889 890 UVMHIST_FUNC("genfs_putpages"); UVMHIST_CALLED(ubchist); 891 892 KASSERT(origflags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)); 893 KASSERT((startoff & PAGE_MASK) == 0 && (endoff & PAGE_MASK) == 0); 894 KASSERT(startoff < endoff || endoff == 0); 895 KASSERT(rw_write_held(slock)); 896 897 UVMHIST_LOG(ubchist, "vp %#jx pages %jd off 0x%jx len 0x%jx", 898 (uintptr_t)vp, uobj->uo_npages, startoff, endoff - startoff); 899 900#ifdef DIAGNOSTIC 901 if ((origflags & PGO_JOURNALLOCKED) && vp->v_mount->mnt_wapbl) 902 WAPBL_JLOCK_ASSERT(vp->v_mount); 903#endif 904 905 trans_mp = NULL; 906 holds_wapbl = false; 907 908retry: 909 modified = false; 910 flags = origflags; 911 912 /* 913 * shortcut if we have no pages to process. 914 */ 915 916 nodirty = uvm_obj_clean_p(uobj); 917#ifdef DIAGNOSTIC 918 mutex_enter(vp->v_interlock); 919 KASSERT((vp->v_iflag & VI_ONWORKLST) != 0 || nodirty); 920 mutex_exit(vp->v_interlock); 921#endif 922 if (uobj->uo_npages == 0 || (dirtyonly && nodirty)) { 923 mutex_enter(vp->v_interlock); 924 if (vp->v_iflag & VI_ONWORKLST && LIST_EMPTY(&vp->v_dirtyblkhd)) { 925 vn_syncer_remove_from_worklist(vp); 926 } 927 mutex_exit(vp->v_interlock); 928 if (trans_mp) { 929 if (holds_wapbl) 930 WAPBL_END(trans_mp); 931 fstrans_done(trans_mp); 932 } 933 rw_exit(slock); 934 return (0); 935 } 936 937 /* 938 * the vnode has pages, set up to process the request. 939 */ 940 941 if (trans_mp == NULL && (flags & PGO_CLEANIT) != 0) { 942 if (pagedaemon) { 943 /* Pagedaemon must not sleep here. */ 944 trans_mp = vp->v_mount; 945 error = fstrans_start_nowait(trans_mp); 946 if (error) { 947 rw_exit(slock); 948 return error; 949 } 950 } else { 951 /* 952 * Cannot use vdeadcheck() here as this operation 953 * usually gets used from VOP_RECLAIM(). Test for 954 * change of v_mount instead and retry on change. 955 */ 956 rw_exit(slock); 957 trans_mp = vp->v_mount; 958 fstrans_start(trans_mp); 959 if (vp->v_mount != trans_mp) { 960 fstrans_done(trans_mp); 961 trans_mp = NULL; 962 } else { 963 holds_wapbl = (trans_mp->mnt_wapbl && 964 (origflags & PGO_JOURNALLOCKED) == 0); 965 if (holds_wapbl) { 966 error = WAPBL_BEGIN(trans_mp); 967 if (error) { 968 fstrans_done(trans_mp); 969 return error; 970 } 971 } 972 } 973 rw_enter(slock, RW_WRITER); 974 goto retry; 975 } 976 } 977 978 error = 0; 979 wasclean = uvm_obj_nowriteback_p(uobj); 980 nextoff = startoff; 981 if (endoff == 0 || flags & PGO_ALLPAGES) { 982 endoff = trunc_page(LLONG_MAX); 983 } 984 985 /* 986 * if this vnode is known not to have dirty pages, 987 * don't bother to clean it out. 988 */ 989 990 if (nodirty) { 991 /* We handled the dirtyonly && nodirty case above. */ 992 KASSERT(!dirtyonly); 993 flags &= ~PGO_CLEANIT; 994 } 995 996 /* 997 * start the loop to scan pages. 998 */ 999 1000 cleanall = true; 1001 freeflag = pagedaemon ? PG_PAGEOUT : PG_RELEASED; 1002 uvm_page_array_init(&a, uobj, dirtyonly ? (UVM_PAGE_ARRAY_FILL_DIRTY | 1003 (!async ? UVM_PAGE_ARRAY_FILL_WRITEBACK : 0)) : 0); 1004 for (;;) { 1005 bool pgprotected; 1006 1007 /* 1008 * if !dirtyonly, iterate over all resident pages in the range. 1009 * 1010 * if dirtyonly, only possibly dirty pages are interesting. 1011 * however, if we are asked to sync for integrity, we should 1012 * wait on pages being written back by other threads as well. 1013 */ 1014 1015 pg = uvm_page_array_fill_and_peek(&a, nextoff, 0); 1016 if (pg == NULL) { 1017 break; 1018 } 1019 1020 KASSERT(pg->uobject == uobj); 1021 KASSERT((pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 || 1022 (pg->flags & (PG_BUSY)) != 0); 1023 KASSERT(pg->offset >= startoff); 1024 KASSERT(pg->offset >= nextoff); 1025 KASSERT(!dirtyonly || 1026 uvm_pagegetdirty(pg) != UVM_PAGE_STATUS_CLEAN || 1027 uvm_obj_page_writeback_p(pg)); 1028 1029 if (pg->offset >= endoff) { 1030 break; 1031 } 1032 1033 /* 1034 * a preempt point. 1035 */ 1036 1037 if (preempt_needed()) { 1038 nextoff = pg->offset; /* visit this page again */ 1039 rw_exit(slock); 1040 preempt(); 1041 /* 1042 * as we dropped the object lock, our cached pages can 1043 * be stale. 1044 */ 1045 uvm_page_array_clear(&a); 1046 rw_enter(slock, RW_WRITER); 1047 continue; 1048 } 1049 1050 /* 1051 * if the current page is busy, wait for it to become unbusy. 1052 */ 1053 1054 if ((pg->flags & PG_BUSY) != 0) { 1055 UVMHIST_LOG(ubchist, "busy %#jx", (uintptr_t)pg, 1056 0, 0, 0); 1057 if ((pg->flags & (PG_RELEASED|PG_PAGEOUT)) != 0 1058 && (flags & PGO_BUSYFAIL) != 0) { 1059 UVMHIST_LOG(ubchist, "busyfail %#jx", 1060 (uintptr_t)pg, 0, 0, 0); 1061 error = EDEADLK; 1062 if (busypg != NULL) 1063 *busypg = pg; 1064 break; 1065 } 1066 if (pagedaemon) { 1067 /* 1068 * someone has taken the page while we 1069 * dropped the lock for fstrans_start. 1070 */ 1071 break; 1072 } 1073 /* 1074 * don't bother to wait on other's activities 1075 * unless we are asked to sync for integrity. 1076 */ 1077 if (!async && (flags & PGO_RECLAIM) == 0) { 1078 wasclean = false; 1079 nextoff = pg->offset + PAGE_SIZE; 1080 uvm_page_array_advance(&a); 1081 continue; 1082 } 1083 nextoff = pg->offset; /* visit this page again */ 1084 uvm_pagewait(pg, slock, "genput"); 1085 /* 1086 * as we dropped the object lock, our cached pages can 1087 * be stale. 1088 */ 1089 uvm_page_array_clear(&a); 1090 rw_enter(slock, RW_WRITER); 1091 continue; 1092 } 1093 1094 nextoff = pg->offset + PAGE_SIZE; 1095 uvm_page_array_advance(&a); 1096 1097 /* 1098 * if we're freeing, remove all mappings of the page now. 1099 * if we're cleaning, check if the page is needs to be cleaned. 1100 */ 1101 1102 pgprotected = false; 1103 if (flags & PGO_FREE) { 1104 pmap_page_protect(pg, VM_PROT_NONE); 1105 pgprotected = true; 1106 } else if (flags & PGO_CLEANIT) { 1107 1108 /* 1109 * if we still have some hope to pull this vnode off 1110 * from the syncer queue, write-protect the page. 1111 */ 1112 1113 if (cleanall && wasclean) { 1114 1115 /* 1116 * uobj pages get wired only by uvm_fault 1117 * where uobj is locked. 1118 */ 1119 1120 if (pg->wire_count == 0) { 1121 pmap_page_protect(pg, 1122 VM_PROT_READ|VM_PROT_EXECUTE); 1123 pgprotected = true; 1124 } else { 1125 cleanall = false; 1126 } 1127 } 1128 } 1129 1130 if (flags & PGO_CLEANIT) { 1131 needs_clean = uvm_pagecheckdirty(pg, pgprotected); 1132 } else { 1133 needs_clean = false; 1134 } 1135 1136 /* 1137 * if we're cleaning, build a cluster. 1138 * the cluster will consist of pages which are currently dirty. 1139 * if not cleaning, just operate on the one page. 1140 */ 1141 1142 if (needs_clean) { 1143 wasclean = false; 1144 memset(pgs, 0, sizeof(pgs)); 1145 pg->flags |= PG_BUSY; 1146 UVM_PAGE_OWN(pg, "genfs_putpages"); 1147 1148 /* 1149 * let the fs constrain the offset range of the cluster. 1150 * we additionally constrain the range here such that 1151 * it fits in the "pgs" pages array. 1152 */ 1153 1154 off_t fslo, fshi, genlo, lo, off = pg->offset; 1155 GOP_PUTRANGE(vp, off, &fslo, &fshi); 1156 KASSERT(fslo == trunc_page(fslo)); 1157 KASSERT(fslo <= off); 1158 KASSERT(fshi == trunc_page(fshi)); 1159 KASSERT(fshi == 0 || off < fshi); 1160 1161 if (off > MAXPHYS / 2) 1162 genlo = trunc_page(off - (MAXPHYS / 2)); 1163 else 1164 genlo = 0; 1165 lo = MAX(fslo, genlo); 1166 1167 /* 1168 * first look backward. 1169 */ 1170 1171 npages = (off - lo) >> PAGE_SHIFT; 1172 nback = npages; 1173 uvn_findpages(uobj, off - PAGE_SIZE, &nback, 1174 &pgs[0], NULL, 1175 UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY|UFP_BACKWARD); 1176 if (nback) { 1177 memmove(&pgs[0], &pgs[npages - nback], 1178 nback * sizeof(pgs[0])); 1179 if (npages - nback < nback) 1180 memset(&pgs[nback], 0, 1181 (npages - nback) * sizeof(pgs[0])); 1182 else 1183 memset(&pgs[npages - nback], 0, 1184 nback * sizeof(pgs[0])); 1185 } 1186 1187 /* 1188 * then plug in our page of interest. 1189 */ 1190 1191 pgs[nback] = pg; 1192 1193 /* 1194 * then look forward to fill in the remaining space in 1195 * the array of pages. 1196 * 1197 * pass our cached array of pages so that hopefully 1198 * uvn_findpages can find some good pages in it. 1199 * the array a was filled above with the one of 1200 * following sets of flags: 1201 * 0 1202 * UVM_PAGE_ARRAY_FILL_DIRTY 1203 * UVM_PAGE_ARRAY_FILL_DIRTY|WRITEBACK 1204 * 1205 * XXX this is fragile but it'll work: the array 1206 * was earlier filled sparsely, but UFP_DIRTYONLY 1207 * implies dense. see corresponding comment in 1208 * uvn_findpages(). 1209 */ 1210 1211 npages = MAXPAGES - nback - 1; 1212 if (fshi) 1213 npages = MIN(npages, 1214 (fshi - off - 1) >> PAGE_SHIFT); 1215 uvn_findpages(uobj, off + PAGE_SIZE, &npages, 1216 &pgs[nback + 1], &a, 1217 UFP_NOWAIT|UFP_NOALLOC|UFP_DIRTYONLY); 1218 npages += nback + 1; 1219 } else { 1220 pgs[0] = pg; 1221 npages = 1; 1222 nback = 0; 1223 } 1224 1225 /* 1226 * apply FREE or DEACTIVATE options if requested. 1227 */ 1228 1229 for (i = 0; i < npages; i++) { 1230 tpg = pgs[i]; 1231 KASSERT(tpg->uobject == uobj); 1232 KASSERT(i == 0 || 1233 pgs[i-1]->offset + PAGE_SIZE == tpg->offset); 1234 KASSERT(!needs_clean || uvm_pagegetdirty(pgs[i]) != 1235 UVM_PAGE_STATUS_DIRTY); 1236 if (needs_clean) { 1237 /* 1238 * mark pages as WRITEBACK so that concurrent 1239 * fsync can find and wait for our activities. 1240 */ 1241 uvm_obj_page_set_writeback(pgs[i]); 1242 } 1243 if (tpg->offset < startoff || tpg->offset >= endoff) 1244 continue; 1245 if (flags & PGO_DEACTIVATE && tpg->wire_count == 0) { 1246 uvm_pagelock(tpg); 1247 uvm_pagedeactivate(tpg); 1248 uvm_pageunlock(tpg); 1249 } else if (flags & PGO_FREE) { 1250 pmap_page_protect(tpg, VM_PROT_NONE); 1251 if (tpg->flags & PG_BUSY) { 1252 tpg->flags |= freeflag; 1253 if (pagedaemon) { 1254 uvm_pageout_start(1); 1255 uvm_pagelock(tpg); 1256 uvm_pagedequeue(tpg); 1257 uvm_pageunlock(tpg); 1258 } 1259 } else { 1260 1261 /* 1262 * ``page is not busy'' 1263 * implies that npages is 1 1264 * and needs_clean is false. 1265 */ 1266 1267 KASSERT(npages == 1); 1268 KASSERT(!needs_clean); 1269 KASSERT(pg == tpg); 1270 KASSERT(nextoff == 1271 tpg->offset + PAGE_SIZE); 1272 uvm_pagefree(tpg); 1273 if (pagedaemon) 1274 uvmexp.pdfreed++; 1275 } 1276 } 1277 } 1278 if (needs_clean) { 1279 modified = true; 1280 KASSERT(nextoff == pg->offset + PAGE_SIZE); 1281 KASSERT(nback < npages); 1282 nextoff = pg->offset + ((npages - nback) << PAGE_SHIFT); 1283 KASSERT(pgs[nback] == pg); 1284 KASSERT(nextoff == pgs[npages - 1]->offset + PAGE_SIZE); 1285 1286 /* 1287 * start the i/o. 1288 */ 1289 rw_exit(slock); 1290 error = GOP_WRITE(vp, pgs, npages, flags); 1291 /* 1292 * as we dropped the object lock, our cached pages can 1293 * be stale. 1294 */ 1295 uvm_page_array_clear(&a); 1296 rw_enter(slock, RW_WRITER); 1297 if (error) { 1298 break; 1299 } 1300 } 1301 } 1302 uvm_page_array_fini(&a); 1303 1304 /* 1305 * update ctime/mtime if the modification we started writing out might 1306 * be from mmap'ed write. 1307 * 1308 * this is necessary when an application keeps a file mmaped and 1309 * repeatedly modifies it via the window. note that, because we 1310 * don't always write-protect pages when cleaning, such modifications 1311 * might not involve any page faults. 1312 */ 1313 1314 mutex_enter(vp->v_interlock); 1315 if (modified && (vp->v_iflag & VI_WRMAP) != 0 && 1316 (vp->v_type != VBLK || 1317 (vp->v_mount->mnt_flag & MNT_NODEVMTIME) == 0)) { 1318 GOP_MARKUPDATE(vp, GOP_UPDATE_MODIFIED); 1319 } 1320 1321 /* 1322 * if we no longer have any possibly dirty pages, take us off the 1323 * syncer list. 1324 */ 1325 1326 if ((vp->v_iflag & VI_ONWORKLST) != 0 && uvm_obj_clean_p(uobj) && 1327 LIST_EMPTY(&vp->v_dirtyblkhd)) { 1328 vn_syncer_remove_from_worklist(vp); 1329 } 1330 1331 /* Wait for output to complete. */ 1332 rw_exit(slock); 1333 if (!wasclean && !async && vp->v_numoutput != 0) { 1334 while (vp->v_numoutput != 0) 1335 cv_wait(&vp->v_cv, vp->v_interlock); 1336 } 1337 onworklst = (vp->v_iflag & VI_ONWORKLST) != 0; 1338 mutex_exit(vp->v_interlock); 1339 1340 if ((flags & PGO_RECLAIM) != 0 && onworklst) { 1341 /* 1342 * in the case of PGO_RECLAIM, ensure to make the vnode clean. 1343 * retrying is not a big deal because, in many cases, 1344 * uobj->uo_npages is already 0 here. 1345 */ 1346 rw_enter(slock, RW_WRITER); 1347 goto retry; 1348 } 1349 1350 if (trans_mp) { 1351 if (holds_wapbl) 1352 WAPBL_END(trans_mp); 1353 fstrans_done(trans_mp); 1354 } 1355 1356 return (error); 1357} 1358 1359/* 1360 * Default putrange method for file systems that do not care 1361 * how many pages are given to one GOP_WRITE() call. 1362 */ 1363void 1364genfs_gop_putrange(struct vnode *vp, off_t off, off_t *lop, off_t *hip) 1365{ 1366 1367 *lop = 0; 1368 *hip = 0; 1369} 1370 1371int 1372genfs_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, int flags) 1373{ 1374 off_t off; 1375 vaddr_t kva; 1376 size_t len; 1377 int error; 1378 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 1379 1380 UVMHIST_LOG(ubchist, "vp %#jx pgs %#jx npages %jd flags 0x%jx", 1381 (uintptr_t)vp, (uintptr_t)pgs, npages, flags); 1382 1383 off = pgs[0]->offset; 1384 kva = uvm_pagermapin(pgs, npages, 1385 UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK); 1386 len = npages << PAGE_SHIFT; 1387 1388 error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE, 1389 uvm_aio_aiodone); 1390 1391 return error; 1392} 1393 1394/* 1395 * genfs_gop_write_rwmap: 1396 * 1397 * a variant of genfs_gop_write. it's used by UDF for its directory buffers. 1398 * this maps pages with PROT_WRITE so that VOP_STRATEGY can modifies 1399 * the contents before writing it out to the underlying storage. 1400 */ 1401 1402int 1403genfs_gop_write_rwmap(struct vnode *vp, struct vm_page **pgs, int npages, 1404 int flags) 1405{ 1406 off_t off; 1407 vaddr_t kva; 1408 size_t len; 1409 int error; 1410 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 1411 1412 UVMHIST_LOG(ubchist, "vp %#jx pgs %#jx npages %jd flags 0x%jx", 1413 (uintptr_t)vp, (uintptr_t)pgs, npages, flags); 1414 1415 off = pgs[0]->offset; 1416 kva = uvm_pagermapin(pgs, npages, 1417 UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK); 1418 len = npages << PAGE_SHIFT; 1419 1420 error = genfs_do_io(vp, off, kva, len, flags, UIO_WRITE, 1421 uvm_aio_aiodone); 1422 1423 return error; 1424} 1425 1426/* 1427 * Backend routine for doing I/O to vnode pages. Pages are already locked 1428 * and mapped into kernel memory. Here we just look up the underlying 1429 * device block addresses and call the strategy routine. 1430 */ 1431 1432static int 1433genfs_do_io(struct vnode *vp, off_t off, vaddr_t kva, size_t len, int flags, 1434 enum uio_rw rw, void (*iodone)(struct buf *)) 1435{ 1436 int s, error; 1437 int fs_bshift, dev_bshift; 1438 off_t eof, offset, startoffset; 1439 size_t bytes, iobytes, skipbytes; 1440 struct buf *mbp, *bp; 1441 const bool async = (flags & PGO_SYNCIO) == 0; 1442 const bool lazy = (flags & PGO_LAZY) == 0; 1443 const bool iowrite = rw == UIO_WRITE; 1444 const int brw = iowrite ? B_WRITE : B_READ; 1445 UVMHIST_FUNC(__func__); UVMHIST_CALLED(ubchist); 1446 1447 UVMHIST_LOG(ubchist, "vp %#jx kva %#jx len 0x%jx flags 0x%jx", 1448 (uintptr_t)vp, (uintptr_t)kva, len, flags); 1449 1450 KASSERT(vp->v_size <= vp->v_writesize); 1451 GOP_SIZE(vp, vp->v_writesize, &eof, 0); 1452 if (vp->v_type != VBLK) { 1453 fs_bshift = vp->v_mount->mnt_fs_bshift; 1454 dev_bshift = vp->v_mount->mnt_dev_bshift; 1455 } else { 1456 fs_bshift = DEV_BSHIFT; 1457 dev_bshift = DEV_BSHIFT; 1458 } 1459 error = 0; 1460 startoffset = off; 1461 bytes = MIN(len, eof - startoffset); 1462 skipbytes = 0; 1463 KASSERT(bytes != 0); 1464 1465 if (iowrite) { 1466 /* 1467 * why += 2? 1468 * 1 for biodone, 1 for uvm_aio_aiodone. 1469 */ 1470 mutex_enter(vp->v_interlock); 1471 vp->v_numoutput += 2; 1472 mutex_exit(vp->v_interlock); 1473 } 1474 mbp = getiobuf(vp, true); 1475 UVMHIST_LOG(ubchist, "vp %#jx mbp %#jx num now %jd bytes 0x%jx", 1476 (uintptr_t)vp, (uintptr_t)mbp, vp->v_numoutput, bytes); 1477 mbp->b_bufsize = len; 1478 mbp->b_data = (void *)kva; 1479 mbp->b_resid = mbp->b_bcount = bytes; 1480 mbp->b_cflags |= BC_BUSY | BC_AGE; 1481 if (async) { 1482 mbp->b_flags = brw | B_ASYNC; 1483 mbp->b_iodone = iodone; 1484 } else { 1485 mbp->b_flags = brw; 1486 mbp->b_iodone = NULL; 1487 } 1488 if (curlwp == uvm.pagedaemon_lwp) 1489 BIO_SETPRIO(mbp, BPRIO_TIMELIMITED); 1490 else if (async || lazy) 1491 BIO_SETPRIO(mbp, BPRIO_TIMENONCRITICAL); 1492 else 1493 BIO_SETPRIO(mbp, BPRIO_TIMECRITICAL); 1494 1495 bp = NULL; 1496 for (offset = startoffset; 1497 bytes > 0; 1498 offset += iobytes, bytes -= iobytes) { 1499 int run; 1500 daddr_t lbn, blkno; 1501 struct vnode *devvp; 1502 1503 /* 1504 * bmap the file to find out the blkno to read from and 1505 * how much we can read in one i/o. if bmap returns an error, 1506 * skip the rest of the top-level i/o. 1507 */ 1508 1509 lbn = offset >> fs_bshift; 1510 error = VOP_BMAP(vp, lbn, &devvp, &blkno, &run); 1511 if (error) { 1512 UVMHIST_LOG(ubchist, "VOP_BMAP lbn 0x%jx -> %jd", 1513 lbn, error, 0, 0); 1514 skipbytes += bytes; 1515 bytes = 0; 1516 goto loopdone; 1517 } 1518 1519 /* 1520 * see how many pages can be read with this i/o. 1521 * reduce the i/o size if necessary to avoid 1522 * overwriting pages with valid data. 1523 */ 1524 1525 iobytes = MIN((((off_t)lbn + 1 + run) << fs_bshift) - offset, 1526 bytes); 1527 1528 /* 1529 * if this block isn't allocated, zero it instead of 1530 * reading it. unless we are going to allocate blocks, 1531 * mark the pages we zeroed PG_RDONLY. 1532 */ 1533 1534 if (blkno == (daddr_t)-1) { 1535 if (!iowrite) { 1536 memset((char *)kva + (offset - startoffset), 0, 1537 iobytes); 1538 } 1539 skipbytes += iobytes; 1540 continue; 1541 } 1542 1543 /* 1544 * allocate a sub-buf for this piece of the i/o 1545 * (or just use mbp if there's only 1 piece), 1546 * and start it going. 1547 */ 1548 1549 if (offset == startoffset && iobytes == bytes) { 1550 bp = mbp; 1551 } else { 1552 UVMHIST_LOG(ubchist, "vp %#jx bp %#jx num now %jd", 1553 (uintptr_t)vp, (uintptr_t)bp, vp->v_numoutput, 0); 1554 bp = getiobuf(vp, true); 1555 nestiobuf_setup(mbp, bp, offset - startoffset, iobytes); 1556 } 1557 bp->b_lblkno = 0; 1558 1559 /* adjust physical blkno for partial blocks */ 1560 bp->b_blkno = blkno + ((offset - ((off_t)lbn << fs_bshift)) >> 1561 dev_bshift); 1562 1563 UVMHIST_LOG(ubchist, 1564 "bp %#jx offset 0x%jx bcount 0x%jx blkno 0x%jx", 1565 (uintptr_t)bp, offset, bp->b_bcount, bp->b_blkno); 1566 1567 VOP_STRATEGY(devvp, bp); 1568 } 1569 1570loopdone: 1571 if (skipbytes) { 1572 UVMHIST_LOG(ubchist, "skipbytes %jd", skipbytes, 0,0,0); 1573 } 1574 nestiobuf_done(mbp, skipbytes, error); 1575 if (async) { 1576 UVMHIST_LOG(ubchist, "returning 0 (async)", 0,0,0,0); 1577 return (0); 1578 } 1579 UVMHIST_LOG(ubchist, "waiting for mbp %#jx", (uintptr_t)mbp, 0, 0, 0); 1580 error = biowait(mbp); 1581 s = splbio(); 1582 (*iodone)(mbp); 1583 splx(s); 1584 UVMHIST_LOG(ubchist, "returning, error %jd", error, 0, 0, 0); 1585 return (error); 1586} 1587 1588int 1589genfs_compat_getpages(void *v) 1590{ 1591 struct vop_getpages_args /* { 1592 struct vnode *a_vp; 1593 voff_t a_offset; 1594 struct vm_page **a_m; 1595 int *a_count; 1596 int a_centeridx; 1597 vm_prot_t a_access_type; 1598 int a_advice; 1599 int a_flags; 1600 } */ *ap = v; 1601 1602 off_t origoffset; 1603 struct vnode *vp = ap->a_vp; 1604 struct uvm_object *uobj = &vp->v_uobj; 1605 struct vm_page *pg, **pgs; 1606 vaddr_t kva; 1607 int i, error, orignpages, npages; 1608 struct iovec iov; 1609 struct uio uio; 1610 kauth_cred_t cred = curlwp->l_cred; 1611 const bool memwrite = (ap->a_access_type & VM_PROT_WRITE) != 0; 1612 1613 error = 0; 1614 origoffset = ap->a_offset; 1615 orignpages = *ap->a_count; 1616 pgs = ap->a_m; 1617 1618 if (ap->a_flags & PGO_LOCKED) { 1619 uvn_findpages(uobj, origoffset, ap->a_count, ap->a_m, NULL, 1620 UFP_NOWAIT|UFP_NOALLOC| (memwrite ? UFP_NORDONLY : 0)); 1621 1622 error = ap->a_m[ap->a_centeridx] == NULL ? EBUSY : 0; 1623 return error; 1624 } 1625 if (origoffset + (ap->a_centeridx << PAGE_SHIFT) >= vp->v_size) { 1626 rw_exit(uobj->vmobjlock); 1627 return EINVAL; 1628 } 1629 if ((ap->a_flags & PGO_SYNCIO) == 0) { 1630 rw_exit(uobj->vmobjlock); 1631 return 0; 1632 } 1633 npages = orignpages; 1634 uvn_findpages(uobj, origoffset, &npages, pgs, NULL, UFP_ALL); 1635 rw_exit(uobj->vmobjlock); 1636 kva = uvm_pagermapin(pgs, npages, 1637 UVMPAGER_MAPIN_READ | UVMPAGER_MAPIN_WAITOK); 1638 for (i = 0; i < npages; i++) { 1639 pg = pgs[i]; 1640 if ((pg->flags & PG_FAKE) == 0) { 1641 continue; 1642 } 1643 iov.iov_base = (char *)kva + (i << PAGE_SHIFT); 1644 iov.iov_len = PAGE_SIZE; 1645 uio.uio_iov = &iov; 1646 uio.uio_iovcnt = 1; 1647 uio.uio_offset = origoffset + (i << PAGE_SHIFT); 1648 uio.uio_rw = UIO_READ; 1649 uio.uio_resid = PAGE_SIZE; 1650 UIO_SETUP_SYSSPACE(&uio); 1651 /* XXX vn_lock */ 1652 error = VOP_READ(vp, &uio, 0, cred); 1653 if (error) { 1654 break; 1655 } 1656 if (uio.uio_resid) { 1657 memset(iov.iov_base, 0, uio.uio_resid); 1658 } 1659 } 1660 uvm_pagermapout(kva, npages); 1661 rw_enter(uobj->vmobjlock, RW_WRITER); 1662 for (i = 0; i < npages; i++) { 1663 pg = pgs[i]; 1664 if (error && (pg->flags & PG_FAKE) != 0) { 1665 pg->flags |= PG_RELEASED; 1666 } else { 1667 uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_UNKNOWN); 1668 uvm_pagelock(pg); 1669 uvm_pageactivate(pg); 1670 uvm_pageunlock(pg); 1671 } 1672 } 1673 if (error) { 1674 uvm_page_unbusy(pgs, npages); 1675 } 1676 rw_exit(uobj->vmobjlock); 1677 return error; 1678} 1679 1680int 1681genfs_compat_gop_write(struct vnode *vp, struct vm_page **pgs, int npages, 1682 int flags) 1683{ 1684 off_t offset; 1685 struct iovec iov; 1686 struct uio uio; 1687 kauth_cred_t cred = curlwp->l_cred; 1688 struct buf *bp; 1689 vaddr_t kva; 1690 int error; 1691 1692 offset = pgs[0]->offset; 1693 kva = uvm_pagermapin(pgs, npages, 1694 UVMPAGER_MAPIN_WRITE | UVMPAGER_MAPIN_WAITOK); 1695 1696 iov.iov_base = (void *)kva; 1697 iov.iov_len = npages << PAGE_SHIFT; 1698 uio.uio_iov = &iov; 1699 uio.uio_iovcnt = 1; 1700 uio.uio_offset = offset; 1701 uio.uio_rw = UIO_WRITE; 1702 uio.uio_resid = npages << PAGE_SHIFT; 1703 UIO_SETUP_SYSSPACE(&uio); 1704 /* XXX vn_lock */ 1705 error = VOP_WRITE(vp, &uio, 0, cred); 1706 1707 mutex_enter(vp->v_interlock); 1708 vp->v_numoutput++; 1709 mutex_exit(vp->v_interlock); 1710 1711 bp = getiobuf(vp, true); 1712 bp->b_cflags |= BC_BUSY | BC_AGE; 1713 bp->b_lblkno = offset >> vp->v_mount->mnt_fs_bshift; 1714 bp->b_data = (char *)kva; 1715 bp->b_bcount = npages << PAGE_SHIFT; 1716 bp->b_bufsize = npages << PAGE_SHIFT; 1717 bp->b_resid = 0; 1718 bp->b_error = error; 1719 uvm_aio_aiodone(bp); 1720 return (error); 1721} 1722 1723/* 1724 * Process a uio using direct I/O. If we reach a part of the request 1725 * which cannot be processed in this fashion for some reason, just return. 1726 * The caller must handle some additional part of the request using 1727 * buffered I/O before trying direct I/O again. 1728 */ 1729 1730void 1731genfs_directio(struct vnode *vp, struct uio *uio, int ioflag) 1732{ 1733 struct vmspace *vs; 1734 struct iovec *iov; 1735 vaddr_t va; 1736 size_t len; 1737 const int mask = DEV_BSIZE - 1; 1738 int error; 1739 bool need_wapbl = (vp->v_mount && vp->v_mount->mnt_wapbl && 1740 (ioflag & IO_JOURNALLOCKED) == 0); 1741 1742#ifdef DIAGNOSTIC 1743 if ((ioflag & IO_JOURNALLOCKED) && vp->v_mount->mnt_wapbl) 1744 WAPBL_JLOCK_ASSERT(vp->v_mount); 1745#endif 1746 1747 /* 1748 * We only support direct I/O to user space for now. 1749 */ 1750 1751 if (VMSPACE_IS_KERNEL_P(uio->uio_vmspace)) { 1752 return; 1753 } 1754 1755 /* 1756 * If the vnode is mapped, we would need to get the getpages lock 1757 * to stabilize the bmap, but then we would get into trouble while 1758 * locking the pages if the pages belong to this same vnode (or a 1759 * multi-vnode cascade to the same effect). Just fall back to 1760 * buffered I/O if the vnode is mapped to avoid this mess. 1761 */ 1762 1763 if (vp->v_vflag & VV_MAPPED) { 1764 return; 1765 } 1766 1767 if (need_wapbl) { 1768 error = WAPBL_BEGIN(vp->v_mount); 1769 if (error) 1770 return; 1771 } 1772 1773 /* 1774 * Do as much of the uio as possible with direct I/O. 1775 */ 1776 1777 vs = uio->uio_vmspace; 1778 while (uio->uio_resid) { 1779 iov = uio->uio_iov; 1780 if (iov->iov_len == 0) { 1781 uio->uio_iov++; 1782 uio->uio_iovcnt--; 1783 continue; 1784 } 1785 va = (vaddr_t)iov->iov_base; 1786 len = MIN(iov->iov_len, genfs_maxdio); 1787 len &= ~mask; 1788 1789 /* 1790 * If the next chunk is smaller than DEV_BSIZE or extends past 1791 * the current EOF, then fall back to buffered I/O. 1792 */ 1793 1794 if (len == 0 || uio->uio_offset + len > vp->v_size) { 1795 break; 1796 } 1797 1798 /* 1799 * Check alignment. The file offset must be at least 1800 * sector-aligned. The exact constraint on memory alignment 1801 * is very hardware-dependent, but requiring sector-aligned 1802 * addresses there too is safe. 1803 */ 1804 1805 if (uio->uio_offset & mask || va & mask) { 1806 break; 1807 } 1808 error = genfs_do_directio(vs, va, len, vp, uio->uio_offset, 1809 uio->uio_rw); 1810 if (error) { 1811 break; 1812 } 1813 iov->iov_base = (char *)iov->iov_base + len; 1814 iov->iov_len -= len; 1815 uio->uio_offset += len; 1816 uio->uio_resid -= len; 1817 } 1818 1819 if (need_wapbl) 1820 WAPBL_END(vp->v_mount); 1821} 1822 1823/* 1824 * Iodone routine for direct I/O. We don't do much here since the request is 1825 * always synchronous, so the caller will do most of the work after biowait(). 1826 */ 1827 1828static void 1829genfs_dio_iodone(struct buf *bp) 1830{ 1831 1832 KASSERT((bp->b_flags & B_ASYNC) == 0); 1833 if ((bp->b_flags & B_READ) == 0 && (bp->b_cflags & BC_AGE) != 0) { 1834 mutex_enter(bp->b_objlock); 1835 vwakeup(bp); 1836 mutex_exit(bp->b_objlock); 1837 } 1838 putiobuf(bp); 1839} 1840 1841/* 1842 * Process one chunk of a direct I/O request. 1843 */ 1844 1845static int 1846genfs_do_directio(struct vmspace *vs, vaddr_t uva, size_t len, struct vnode *vp, 1847 off_t off, enum uio_rw rw) 1848{ 1849 struct vm_map *map; 1850 struct pmap *upm, *kpm __unused; 1851 size_t klen = round_page(uva + len) - trunc_page(uva); 1852 off_t spoff, epoff; 1853 vaddr_t kva, puva; 1854 paddr_t pa; 1855 vm_prot_t prot; 1856 int error, rv __diagused, poff, koff; 1857 const int pgoflags = PGO_CLEANIT | PGO_SYNCIO | PGO_JOURNALLOCKED | 1858 (rw == UIO_WRITE ? PGO_FREE : 0); 1859 1860 /* 1861 * For writes, verify that this range of the file already has fully 1862 * allocated backing store. If there are any holes, just punt and 1863 * make the caller take the buffered write path. 1864 */ 1865 1866 if (rw == UIO_WRITE) { 1867 daddr_t lbn, elbn, blkno; 1868 int bsize, bshift, run; 1869 1870 bshift = vp->v_mount->mnt_fs_bshift; 1871 bsize = 1 << bshift; 1872 lbn = off >> bshift; 1873 elbn = (off + len + bsize - 1) >> bshift; 1874 while (lbn < elbn) { 1875 error = VOP_BMAP(vp, lbn, NULL, &blkno, &run); 1876 if (error) { 1877 return error; 1878 } 1879 if (blkno == (daddr_t)-1) { 1880 return ENOSPC; 1881 } 1882 lbn += 1 + run; 1883 } 1884 } 1885 1886 /* 1887 * Flush any cached pages for parts of the file that we're about to 1888 * access. If we're writing, invalidate pages as well. 1889 */ 1890 1891 spoff = trunc_page(off); 1892 epoff = round_page(off + len); 1893 rw_enter(vp->v_uobj.vmobjlock, RW_WRITER); 1894 error = VOP_PUTPAGES(vp, spoff, epoff, pgoflags); 1895 if (error) { 1896 return error; 1897 } 1898 1899 /* 1900 * Wire the user pages and remap them into kernel memory. 1901 */ 1902 1903 prot = rw == UIO_READ ? VM_PROT_READ | VM_PROT_WRITE : VM_PROT_READ; 1904 error = uvm_vslock(vs, (void *)uva, len, prot); 1905 if (error) { 1906 return error; 1907 } 1908 1909 map = &vs->vm_map; 1910 upm = vm_map_pmap(map); 1911 kpm = vm_map_pmap(kernel_map); 1912 puva = trunc_page(uva); 1913 kva = uvm_km_alloc(kernel_map, klen, atop(puva) & uvmexp.colormask, 1914 UVM_KMF_VAONLY | UVM_KMF_WAITVA | UVM_KMF_COLORMATCH); 1915 for (poff = 0; poff < klen; poff += PAGE_SIZE) { 1916 rv = pmap_extract(upm, puva + poff, &pa); 1917 KASSERT(rv); 1918 pmap_kenter_pa(kva + poff, pa, prot, PMAP_WIRED); 1919 } 1920 pmap_update(kpm); 1921 1922 /* 1923 * Do the I/O. 1924 */ 1925 1926 koff = uva - trunc_page(uva); 1927 error = genfs_do_io(vp, off, kva + koff, len, PGO_SYNCIO, rw, 1928 genfs_dio_iodone); 1929 1930 /* 1931 * Tear down the kernel mapping. 1932 */ 1933 1934 pmap_kremove(kva, klen); 1935 pmap_update(kpm); 1936 uvm_km_free(kernel_map, kva, klen, UVM_KMF_VAONLY); 1937 1938 /* 1939 * Unwire the user pages. 1940 */ 1941 1942 uvm_vsunlock(vs, (void *)uva, len); 1943 return error; 1944} 1945