1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2016 Avago Technologies.  All rights reserved.
4 */
5#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
6#include <linux/module.h>
7#include <linux/slab.h>
8#include <linux/blk-mq.h>
9#include <linux/parser.h>
10#include <linux/random.h>
11#include <uapi/scsi/fc/fc_fs.h>
12#include <uapi/scsi/fc/fc_els.h>
13
14#include "nvmet.h"
15#include <linux/nvme-fc-driver.h>
16#include <linux/nvme-fc.h>
17#include "../host/fc.h"
18
19
20/* *************************** Data Structures/Defines ****************** */
21
22
23#define NVMET_LS_CTX_COUNT		256
24
25struct nvmet_fc_tgtport;
26struct nvmet_fc_tgt_assoc;
27
28struct nvmet_fc_ls_iod {		/* for an LS RQST RCV */
29	struct nvmefc_ls_rsp		*lsrsp;
30	struct nvmefc_tgt_fcp_req	*fcpreq;	/* only if RS */
31
32	struct list_head		ls_rcv_list; /* tgtport->ls_rcv_list */
33
34	struct nvmet_fc_tgtport		*tgtport;
35	struct nvmet_fc_tgt_assoc	*assoc;
36	void				*hosthandle;
37
38	union nvmefc_ls_requests	*rqstbuf;
39	union nvmefc_ls_responses	*rspbuf;
40	u16				rqstdatalen;
41	dma_addr_t			rspdma;
42
43	struct scatterlist		sg[2];
44
45	struct work_struct		work;
46} __aligned(sizeof(unsigned long long));
47
48struct nvmet_fc_ls_req_op {		/* for an LS RQST XMT */
49	struct nvmefc_ls_req		ls_req;
50
51	struct nvmet_fc_tgtport		*tgtport;
52	void				*hosthandle;
53
54	int				ls_error;
55	struct list_head		lsreq_list; /* tgtport->ls_req_list */
56	bool				req_queued;
57};
58
59
60/* desired maximum for a single sequence - if sg list allows it */
61#define NVMET_FC_MAX_SEQ_LENGTH		(256 * 1024)
62
63enum nvmet_fcp_datadir {
64	NVMET_FCP_NODATA,
65	NVMET_FCP_WRITE,
66	NVMET_FCP_READ,
67	NVMET_FCP_ABORTED,
68};
69
70struct nvmet_fc_fcp_iod {
71	struct nvmefc_tgt_fcp_req	*fcpreq;
72
73	struct nvme_fc_cmd_iu		cmdiubuf;
74	struct nvme_fc_ersp_iu		rspiubuf;
75	dma_addr_t			rspdma;
76	struct scatterlist		*next_sg;
77	struct scatterlist		*data_sg;
78	int				data_sg_cnt;
79	u32				offset;
80	enum nvmet_fcp_datadir		io_dir;
81	bool				active;
82	bool				abort;
83	bool				aborted;
84	bool				writedataactive;
85	spinlock_t			flock;
86
87	struct nvmet_req		req;
88	struct work_struct		defer_work;
89
90	struct nvmet_fc_tgtport		*tgtport;
91	struct nvmet_fc_tgt_queue	*queue;
92
93	struct list_head		fcp_list;	/* tgtport->fcp_list */
94};
95
96struct nvmet_fc_tgtport {
97	struct nvmet_fc_target_port	fc_target_port;
98
99	struct list_head		tgt_list; /* nvmet_fc_target_list */
100	struct device			*dev;	/* dev for dma mapping */
101	struct nvmet_fc_target_template	*ops;
102
103	struct nvmet_fc_ls_iod		*iod;
104	spinlock_t			lock;
105	struct list_head		ls_rcv_list;
106	struct list_head		ls_req_list;
107	struct list_head		ls_busylist;
108	struct list_head		assoc_list;
109	struct list_head		host_list;
110	struct ida			assoc_cnt;
111	struct nvmet_fc_port_entry	*pe;
112	struct kref			ref;
113	u32				max_sg_cnt;
114
115	struct work_struct		put_work;
116};
117
118struct nvmet_fc_port_entry {
119	struct nvmet_fc_tgtport		*tgtport;
120	struct nvmet_port		*port;
121	u64				node_name;
122	u64				port_name;
123	struct list_head		pe_list;
124};
125
126struct nvmet_fc_defer_fcp_req {
127	struct list_head		req_list;
128	struct nvmefc_tgt_fcp_req	*fcp_req;
129};
130
131struct nvmet_fc_tgt_queue {
132	bool				ninetypercent;
133	u16				qid;
134	u16				sqsize;
135	u16				ersp_ratio;
136	__le16				sqhd;
137	atomic_t			connected;
138	atomic_t			sqtail;
139	atomic_t			zrspcnt;
140	atomic_t			rsn;
141	spinlock_t			qlock;
142	struct nvmet_cq			nvme_cq;
143	struct nvmet_sq			nvme_sq;
144	struct nvmet_fc_tgt_assoc	*assoc;
145	struct list_head		fod_list;
146	struct list_head		pending_cmd_list;
147	struct list_head		avail_defer_list;
148	struct workqueue_struct		*work_q;
149	struct kref			ref;
150	/* array of fcp_iods */
151	struct nvmet_fc_fcp_iod		fod[] __counted_by(sqsize);
152} __aligned(sizeof(unsigned long long));
153
154struct nvmet_fc_hostport {
155	struct nvmet_fc_tgtport		*tgtport;
156	void				*hosthandle;
157	struct list_head		host_list;
158	struct kref			ref;
159	u8				invalid;
160};
161
162struct nvmet_fc_tgt_assoc {
163	u64				association_id;
164	u32				a_id;
165	atomic_t			terminating;
166	struct nvmet_fc_tgtport		*tgtport;
167	struct nvmet_fc_hostport	*hostport;
168	struct nvmet_fc_ls_iod		*rcv_disconn;
169	struct list_head		a_list;
170	struct nvmet_fc_tgt_queue 	*queues[NVMET_NR_QUEUES + 1];
171	struct kref			ref;
172	struct work_struct		del_work;
173};
174
175
176static inline int
177nvmet_fc_iodnum(struct nvmet_fc_ls_iod *iodptr)
178{
179	return (iodptr - iodptr->tgtport->iod);
180}
181
182static inline int
183nvmet_fc_fodnum(struct nvmet_fc_fcp_iod *fodptr)
184{
185	return (fodptr - fodptr->queue->fod);
186}
187
188
189/*
190 * Association and Connection IDs:
191 *
192 * Association ID will have random number in upper 6 bytes and zero
193 *   in lower 2 bytes
194 *
195 * Connection IDs will be Association ID with QID or'd in lower 2 bytes
196 *
197 * note: Association ID = Connection ID for queue 0
198 */
199#define BYTES_FOR_QID			sizeof(u16)
200#define BYTES_FOR_QID_SHIFT		(BYTES_FOR_QID * 8)
201#define NVMET_FC_QUEUEID_MASK		((u64)((1 << BYTES_FOR_QID_SHIFT) - 1))
202
203static inline u64
204nvmet_fc_makeconnid(struct nvmet_fc_tgt_assoc *assoc, u16 qid)
205{
206	return (assoc->association_id | qid);
207}
208
209static inline u64
210nvmet_fc_getassociationid(u64 connectionid)
211{
212	return connectionid & ~NVMET_FC_QUEUEID_MASK;
213}
214
215static inline u16
216nvmet_fc_getqueueid(u64 connectionid)
217{
218	return (u16)(connectionid & NVMET_FC_QUEUEID_MASK);
219}
220
221static inline struct nvmet_fc_tgtport *
222targetport_to_tgtport(struct nvmet_fc_target_port *targetport)
223{
224	return container_of(targetport, struct nvmet_fc_tgtport,
225				 fc_target_port);
226}
227
228static inline struct nvmet_fc_fcp_iod *
229nvmet_req_to_fod(struct nvmet_req *nvme_req)
230{
231	return container_of(nvme_req, struct nvmet_fc_fcp_iod, req);
232}
233
234
235/* *************************** Globals **************************** */
236
237
238static DEFINE_SPINLOCK(nvmet_fc_tgtlock);
239
240static LIST_HEAD(nvmet_fc_target_list);
241static DEFINE_IDA(nvmet_fc_tgtport_cnt);
242static LIST_HEAD(nvmet_fc_portentry_list);
243
244
245static void nvmet_fc_handle_ls_rqst_work(struct work_struct *work);
246static void nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work);
247static void nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc);
248static int nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc);
249static void nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue);
250static int nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue);
251static void nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport);
252static void nvmet_fc_put_tgtport_work(struct work_struct *work)
253{
254	struct nvmet_fc_tgtport *tgtport =
255		container_of(work, struct nvmet_fc_tgtport, put_work);
256
257	nvmet_fc_tgtport_put(tgtport);
258}
259static int nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport);
260static void nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
261					struct nvmet_fc_fcp_iod *fod);
262static void nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc);
263static void nvmet_fc_xmt_ls_rsp(struct nvmet_fc_tgtport *tgtport,
264				struct nvmet_fc_ls_iod *iod);
265
266
267/* *********************** FC-NVME DMA Handling **************************** */
268
269/*
270 * The fcloop device passes in a NULL device pointer. Real LLD's will
271 * pass in a valid device pointer. If NULL is passed to the dma mapping
272 * routines, depending on the platform, it may or may not succeed, and
273 * may crash.
274 *
275 * As such:
276 * Wrapper all the dma routines and check the dev pointer.
277 *
278 * If simple mappings (return just a dma address, we'll noop them,
279 * returning a dma address of 0.
280 *
281 * On more complex mappings (dma_map_sg), a pseudo routine fills
282 * in the scatter list, setting all dma addresses to 0.
283 */
284
285static inline dma_addr_t
286fc_dma_map_single(struct device *dev, void *ptr, size_t size,
287		enum dma_data_direction dir)
288{
289	return dev ? dma_map_single(dev, ptr, size, dir) : (dma_addr_t)0L;
290}
291
292static inline int
293fc_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
294{
295	return dev ? dma_mapping_error(dev, dma_addr) : 0;
296}
297
298static inline void
299fc_dma_unmap_single(struct device *dev, dma_addr_t addr, size_t size,
300	enum dma_data_direction dir)
301{
302	if (dev)
303		dma_unmap_single(dev, addr, size, dir);
304}
305
306static inline void
307fc_dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
308		enum dma_data_direction dir)
309{
310	if (dev)
311		dma_sync_single_for_cpu(dev, addr, size, dir);
312}
313
314static inline void
315fc_dma_sync_single_for_device(struct device *dev, dma_addr_t addr, size_t size,
316		enum dma_data_direction dir)
317{
318	if (dev)
319		dma_sync_single_for_device(dev, addr, size, dir);
320}
321
322/* pseudo dma_map_sg call */
323static int
324fc_map_sg(struct scatterlist *sg, int nents)
325{
326	struct scatterlist *s;
327	int i;
328
329	WARN_ON(nents == 0 || sg[0].length == 0);
330
331	for_each_sg(sg, s, nents, i) {
332		s->dma_address = 0L;
333#ifdef CONFIG_NEED_SG_DMA_LENGTH
334		s->dma_length = s->length;
335#endif
336	}
337	return nents;
338}
339
340static inline int
341fc_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
342		enum dma_data_direction dir)
343{
344	return dev ? dma_map_sg(dev, sg, nents, dir) : fc_map_sg(sg, nents);
345}
346
347static inline void
348fc_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
349		enum dma_data_direction dir)
350{
351	if (dev)
352		dma_unmap_sg(dev, sg, nents, dir);
353}
354
355
356/* ********************** FC-NVME LS XMT Handling ************************* */
357
358
359static void
360__nvmet_fc_finish_ls_req(struct nvmet_fc_ls_req_op *lsop)
361{
362	struct nvmet_fc_tgtport *tgtport = lsop->tgtport;
363	struct nvmefc_ls_req *lsreq = &lsop->ls_req;
364	unsigned long flags;
365
366	spin_lock_irqsave(&tgtport->lock, flags);
367
368	if (!lsop->req_queued) {
369		spin_unlock_irqrestore(&tgtport->lock, flags);
370		goto out_putwork;
371	}
372
373	list_del(&lsop->lsreq_list);
374
375	lsop->req_queued = false;
376
377	spin_unlock_irqrestore(&tgtport->lock, flags);
378
379	fc_dma_unmap_single(tgtport->dev, lsreq->rqstdma,
380				  (lsreq->rqstlen + lsreq->rsplen),
381				  DMA_BIDIRECTIONAL);
382
383out_putwork:
384	queue_work(nvmet_wq, &tgtport->put_work);
385}
386
387static int
388__nvmet_fc_send_ls_req(struct nvmet_fc_tgtport *tgtport,
389		struct nvmet_fc_ls_req_op *lsop,
390		void (*done)(struct nvmefc_ls_req *req, int status))
391{
392	struct nvmefc_ls_req *lsreq = &lsop->ls_req;
393	unsigned long flags;
394	int ret = 0;
395
396	if (!tgtport->ops->ls_req)
397		return -EOPNOTSUPP;
398
399	if (!nvmet_fc_tgtport_get(tgtport))
400		return -ESHUTDOWN;
401
402	lsreq->done = done;
403	lsop->req_queued = false;
404	INIT_LIST_HEAD(&lsop->lsreq_list);
405
406	lsreq->rqstdma = fc_dma_map_single(tgtport->dev, lsreq->rqstaddr,
407				  lsreq->rqstlen + lsreq->rsplen,
408				  DMA_BIDIRECTIONAL);
409	if (fc_dma_mapping_error(tgtport->dev, lsreq->rqstdma)) {
410		ret = -EFAULT;
411		goto out_puttgtport;
412	}
413	lsreq->rspdma = lsreq->rqstdma + lsreq->rqstlen;
414
415	spin_lock_irqsave(&tgtport->lock, flags);
416
417	list_add_tail(&lsop->lsreq_list, &tgtport->ls_req_list);
418
419	lsop->req_queued = true;
420
421	spin_unlock_irqrestore(&tgtport->lock, flags);
422
423	ret = tgtport->ops->ls_req(&tgtport->fc_target_port, lsop->hosthandle,
424				   lsreq);
425	if (ret)
426		goto out_unlink;
427
428	return 0;
429
430out_unlink:
431	lsop->ls_error = ret;
432	spin_lock_irqsave(&tgtport->lock, flags);
433	lsop->req_queued = false;
434	list_del(&lsop->lsreq_list);
435	spin_unlock_irqrestore(&tgtport->lock, flags);
436	fc_dma_unmap_single(tgtport->dev, lsreq->rqstdma,
437				  (lsreq->rqstlen + lsreq->rsplen),
438				  DMA_BIDIRECTIONAL);
439out_puttgtport:
440	nvmet_fc_tgtport_put(tgtport);
441
442	return ret;
443}
444
445static int
446nvmet_fc_send_ls_req_async(struct nvmet_fc_tgtport *tgtport,
447		struct nvmet_fc_ls_req_op *lsop,
448		void (*done)(struct nvmefc_ls_req *req, int status))
449{
450	/* don't wait for completion */
451
452	return __nvmet_fc_send_ls_req(tgtport, lsop, done);
453}
454
455static void
456nvmet_fc_disconnect_assoc_done(struct nvmefc_ls_req *lsreq, int status)
457{
458	struct nvmet_fc_ls_req_op *lsop =
459		container_of(lsreq, struct nvmet_fc_ls_req_op, ls_req);
460
461	__nvmet_fc_finish_ls_req(lsop);
462
463	/* fc-nvme target doesn't care about success or failure of cmd */
464
465	kfree(lsop);
466}
467
468/*
469 * This routine sends a FC-NVME LS to disconnect (aka terminate)
470 * the FC-NVME Association.  Terminating the association also
471 * terminates the FC-NVME connections (per queue, both admin and io
472 * queues) that are part of the association. E.g. things are torn
473 * down, and the related FC-NVME Association ID and Connection IDs
474 * become invalid.
475 *
476 * The behavior of the fc-nvme target is such that it's
477 * understanding of the association and connections will implicitly
478 * be torn down. The action is implicit as it may be due to a loss of
479 * connectivity with the fc-nvme host, so the target may never get a
480 * response even if it tried.  As such, the action of this routine
481 * is to asynchronously send the LS, ignore any results of the LS, and
482 * continue on with terminating the association. If the fc-nvme host
483 * is present and receives the LS, it too can tear down.
484 */
485static void
486nvmet_fc_xmt_disconnect_assoc(struct nvmet_fc_tgt_assoc *assoc)
487{
488	struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
489	struct fcnvme_ls_disconnect_assoc_rqst *discon_rqst;
490	struct fcnvme_ls_disconnect_assoc_acc *discon_acc;
491	struct nvmet_fc_ls_req_op *lsop;
492	struct nvmefc_ls_req *lsreq;
493	int ret;
494
495	/*
496	 * If ls_req is NULL or no hosthandle, it's an older lldd and no
497	 * message is normal. Otherwise, send unless the hostport has
498	 * already been invalidated by the lldd.
499	 */
500	if (!tgtport->ops->ls_req || assoc->hostport->invalid)
501		return;
502
503	lsop = kzalloc((sizeof(*lsop) +
504			sizeof(*discon_rqst) + sizeof(*discon_acc) +
505			tgtport->ops->lsrqst_priv_sz), GFP_KERNEL);
506	if (!lsop) {
507		dev_info(tgtport->dev,
508			"{%d:%d} send Disconnect Association failed: ENOMEM\n",
509			tgtport->fc_target_port.port_num, assoc->a_id);
510		return;
511	}
512
513	discon_rqst = (struct fcnvme_ls_disconnect_assoc_rqst *)&lsop[1];
514	discon_acc = (struct fcnvme_ls_disconnect_assoc_acc *)&discon_rqst[1];
515	lsreq = &lsop->ls_req;
516	if (tgtport->ops->lsrqst_priv_sz)
517		lsreq->private = (void *)&discon_acc[1];
518	else
519		lsreq->private = NULL;
520
521	lsop->tgtport = tgtport;
522	lsop->hosthandle = assoc->hostport->hosthandle;
523
524	nvmefc_fmt_lsreq_discon_assoc(lsreq, discon_rqst, discon_acc,
525				assoc->association_id);
526
527	ret = nvmet_fc_send_ls_req_async(tgtport, lsop,
528				nvmet_fc_disconnect_assoc_done);
529	if (ret) {
530		dev_info(tgtport->dev,
531			"{%d:%d} XMT Disconnect Association failed: %d\n",
532			tgtport->fc_target_port.port_num, assoc->a_id, ret);
533		kfree(lsop);
534	}
535}
536
537
538/* *********************** FC-NVME Port Management ************************ */
539
540
541static int
542nvmet_fc_alloc_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
543{
544	struct nvmet_fc_ls_iod *iod;
545	int i;
546
547	iod = kcalloc(NVMET_LS_CTX_COUNT, sizeof(struct nvmet_fc_ls_iod),
548			GFP_KERNEL);
549	if (!iod)
550		return -ENOMEM;
551
552	tgtport->iod = iod;
553
554	for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
555		INIT_WORK(&iod->work, nvmet_fc_handle_ls_rqst_work);
556		iod->tgtport = tgtport;
557		list_add_tail(&iod->ls_rcv_list, &tgtport->ls_rcv_list);
558
559		iod->rqstbuf = kzalloc(sizeof(union nvmefc_ls_requests) +
560				       sizeof(union nvmefc_ls_responses),
561				       GFP_KERNEL);
562		if (!iod->rqstbuf)
563			goto out_fail;
564
565		iod->rspbuf = (union nvmefc_ls_responses *)&iod->rqstbuf[1];
566
567		iod->rspdma = fc_dma_map_single(tgtport->dev, iod->rspbuf,
568						sizeof(*iod->rspbuf),
569						DMA_TO_DEVICE);
570		if (fc_dma_mapping_error(tgtport->dev, iod->rspdma))
571			goto out_fail;
572	}
573
574	return 0;
575
576out_fail:
577	kfree(iod->rqstbuf);
578	list_del(&iod->ls_rcv_list);
579	for (iod--, i--; i >= 0; iod--, i--) {
580		fc_dma_unmap_single(tgtport->dev, iod->rspdma,
581				sizeof(*iod->rspbuf), DMA_TO_DEVICE);
582		kfree(iod->rqstbuf);
583		list_del(&iod->ls_rcv_list);
584	}
585
586	kfree(iod);
587
588	return -EFAULT;
589}
590
591static void
592nvmet_fc_free_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
593{
594	struct nvmet_fc_ls_iod *iod = tgtport->iod;
595	int i;
596
597	for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
598		fc_dma_unmap_single(tgtport->dev,
599				iod->rspdma, sizeof(*iod->rspbuf),
600				DMA_TO_DEVICE);
601		kfree(iod->rqstbuf);
602		list_del(&iod->ls_rcv_list);
603	}
604	kfree(tgtport->iod);
605}
606
607static struct nvmet_fc_ls_iod *
608nvmet_fc_alloc_ls_iod(struct nvmet_fc_tgtport *tgtport)
609{
610	struct nvmet_fc_ls_iod *iod;
611	unsigned long flags;
612
613	spin_lock_irqsave(&tgtport->lock, flags);
614	iod = list_first_entry_or_null(&tgtport->ls_rcv_list,
615					struct nvmet_fc_ls_iod, ls_rcv_list);
616	if (iod)
617		list_move_tail(&iod->ls_rcv_list, &tgtport->ls_busylist);
618	spin_unlock_irqrestore(&tgtport->lock, flags);
619	return iod;
620}
621
622
623static void
624nvmet_fc_free_ls_iod(struct nvmet_fc_tgtport *tgtport,
625			struct nvmet_fc_ls_iod *iod)
626{
627	unsigned long flags;
628
629	spin_lock_irqsave(&tgtport->lock, flags);
630	list_move(&iod->ls_rcv_list, &tgtport->ls_rcv_list);
631	spin_unlock_irqrestore(&tgtport->lock, flags);
632}
633
634static void
635nvmet_fc_prep_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
636				struct nvmet_fc_tgt_queue *queue)
637{
638	struct nvmet_fc_fcp_iod *fod = queue->fod;
639	int i;
640
641	for (i = 0; i < queue->sqsize; fod++, i++) {
642		INIT_WORK(&fod->defer_work, nvmet_fc_fcp_rqst_op_defer_work);
643		fod->tgtport = tgtport;
644		fod->queue = queue;
645		fod->active = false;
646		fod->abort = false;
647		fod->aborted = false;
648		fod->fcpreq = NULL;
649		list_add_tail(&fod->fcp_list, &queue->fod_list);
650		spin_lock_init(&fod->flock);
651
652		fod->rspdma = fc_dma_map_single(tgtport->dev, &fod->rspiubuf,
653					sizeof(fod->rspiubuf), DMA_TO_DEVICE);
654		if (fc_dma_mapping_error(tgtport->dev, fod->rspdma)) {
655			list_del(&fod->fcp_list);
656			for (fod--, i--; i >= 0; fod--, i--) {
657				fc_dma_unmap_single(tgtport->dev, fod->rspdma,
658						sizeof(fod->rspiubuf),
659						DMA_TO_DEVICE);
660				fod->rspdma = 0L;
661				list_del(&fod->fcp_list);
662			}
663
664			return;
665		}
666	}
667}
668
669static void
670nvmet_fc_destroy_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
671				struct nvmet_fc_tgt_queue *queue)
672{
673	struct nvmet_fc_fcp_iod *fod = queue->fod;
674	int i;
675
676	for (i = 0; i < queue->sqsize; fod++, i++) {
677		if (fod->rspdma)
678			fc_dma_unmap_single(tgtport->dev, fod->rspdma,
679				sizeof(fod->rspiubuf), DMA_TO_DEVICE);
680	}
681}
682
683static struct nvmet_fc_fcp_iod *
684nvmet_fc_alloc_fcp_iod(struct nvmet_fc_tgt_queue *queue)
685{
686	struct nvmet_fc_fcp_iod *fod;
687
688	lockdep_assert_held(&queue->qlock);
689
690	fod = list_first_entry_or_null(&queue->fod_list,
691					struct nvmet_fc_fcp_iod, fcp_list);
692	if (fod) {
693		list_del(&fod->fcp_list);
694		fod->active = true;
695		/*
696		 * no queue reference is taken, as it was taken by the
697		 * queue lookup just prior to the allocation. The iod
698		 * will "inherit" that reference.
699		 */
700	}
701	return fod;
702}
703
704
705static void
706nvmet_fc_queue_fcp_req(struct nvmet_fc_tgtport *tgtport,
707		       struct nvmet_fc_tgt_queue *queue,
708		       struct nvmefc_tgt_fcp_req *fcpreq)
709{
710	struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
711
712	/*
713	 * put all admin cmds on hw queue id 0. All io commands go to
714	 * the respective hw queue based on a modulo basis
715	 */
716	fcpreq->hwqid = queue->qid ?
717			((queue->qid - 1) % tgtport->ops->max_hw_queues) : 0;
718
719	nvmet_fc_handle_fcp_rqst(tgtport, fod);
720}
721
722static void
723nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work)
724{
725	struct nvmet_fc_fcp_iod *fod =
726		container_of(work, struct nvmet_fc_fcp_iod, defer_work);
727
728	/* Submit deferred IO for processing */
729	nvmet_fc_queue_fcp_req(fod->tgtport, fod->queue, fod->fcpreq);
730
731}
732
733static void
734nvmet_fc_free_fcp_iod(struct nvmet_fc_tgt_queue *queue,
735			struct nvmet_fc_fcp_iod *fod)
736{
737	struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
738	struct nvmet_fc_tgtport *tgtport = fod->tgtport;
739	struct nvmet_fc_defer_fcp_req *deferfcp;
740	unsigned long flags;
741
742	fc_dma_sync_single_for_cpu(tgtport->dev, fod->rspdma,
743				sizeof(fod->rspiubuf), DMA_TO_DEVICE);
744
745	fcpreq->nvmet_fc_private = NULL;
746
747	fod->active = false;
748	fod->abort = false;
749	fod->aborted = false;
750	fod->writedataactive = false;
751	fod->fcpreq = NULL;
752
753	tgtport->ops->fcp_req_release(&tgtport->fc_target_port, fcpreq);
754
755	/* release the queue lookup reference on the completed IO */
756	nvmet_fc_tgt_q_put(queue);
757
758	spin_lock_irqsave(&queue->qlock, flags);
759	deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
760				struct nvmet_fc_defer_fcp_req, req_list);
761	if (!deferfcp) {
762		list_add_tail(&fod->fcp_list, &fod->queue->fod_list);
763		spin_unlock_irqrestore(&queue->qlock, flags);
764		return;
765	}
766
767	/* Re-use the fod for the next pending cmd that was deferred */
768	list_del(&deferfcp->req_list);
769
770	fcpreq = deferfcp->fcp_req;
771
772	/* deferfcp can be reused for another IO at a later date */
773	list_add_tail(&deferfcp->req_list, &queue->avail_defer_list);
774
775	spin_unlock_irqrestore(&queue->qlock, flags);
776
777	/* Save NVME CMD IO in fod */
778	memcpy(&fod->cmdiubuf, fcpreq->rspaddr, fcpreq->rsplen);
779
780	/* Setup new fcpreq to be processed */
781	fcpreq->rspaddr = NULL;
782	fcpreq->rsplen  = 0;
783	fcpreq->nvmet_fc_private = fod;
784	fod->fcpreq = fcpreq;
785	fod->active = true;
786
787	/* inform LLDD IO is now being processed */
788	tgtport->ops->defer_rcv(&tgtport->fc_target_port, fcpreq);
789
790	/*
791	 * Leave the queue lookup get reference taken when
792	 * fod was originally allocated.
793	 */
794
795	queue_work(queue->work_q, &fod->defer_work);
796}
797
798static struct nvmet_fc_tgt_queue *
799nvmet_fc_alloc_target_queue(struct nvmet_fc_tgt_assoc *assoc,
800			u16 qid, u16 sqsize)
801{
802	struct nvmet_fc_tgt_queue *queue;
803	int ret;
804
805	if (qid > NVMET_NR_QUEUES)
806		return NULL;
807
808	queue = kzalloc(struct_size(queue, fod, sqsize), GFP_KERNEL);
809	if (!queue)
810		return NULL;
811
812	queue->work_q = alloc_workqueue("ntfc%d.%d.%d", 0, 0,
813				assoc->tgtport->fc_target_port.port_num,
814				assoc->a_id, qid);
815	if (!queue->work_q)
816		goto out_free_queue;
817
818	queue->qid = qid;
819	queue->sqsize = sqsize;
820	queue->assoc = assoc;
821	INIT_LIST_HEAD(&queue->fod_list);
822	INIT_LIST_HEAD(&queue->avail_defer_list);
823	INIT_LIST_HEAD(&queue->pending_cmd_list);
824	atomic_set(&queue->connected, 0);
825	atomic_set(&queue->sqtail, 0);
826	atomic_set(&queue->rsn, 1);
827	atomic_set(&queue->zrspcnt, 0);
828	spin_lock_init(&queue->qlock);
829	kref_init(&queue->ref);
830
831	nvmet_fc_prep_fcp_iodlist(assoc->tgtport, queue);
832
833	ret = nvmet_sq_init(&queue->nvme_sq);
834	if (ret)
835		goto out_fail_iodlist;
836
837	WARN_ON(assoc->queues[qid]);
838	assoc->queues[qid] = queue;
839
840	return queue;
841
842out_fail_iodlist:
843	nvmet_fc_destroy_fcp_iodlist(assoc->tgtport, queue);
844	destroy_workqueue(queue->work_q);
845out_free_queue:
846	kfree(queue);
847	return NULL;
848}
849
850
851static void
852nvmet_fc_tgt_queue_free(struct kref *ref)
853{
854	struct nvmet_fc_tgt_queue *queue =
855		container_of(ref, struct nvmet_fc_tgt_queue, ref);
856
857	nvmet_fc_destroy_fcp_iodlist(queue->assoc->tgtport, queue);
858
859	destroy_workqueue(queue->work_q);
860
861	kfree(queue);
862}
863
864static void
865nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue)
866{
867	kref_put(&queue->ref, nvmet_fc_tgt_queue_free);
868}
869
870static int
871nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue)
872{
873	return kref_get_unless_zero(&queue->ref);
874}
875
876
877static void
878nvmet_fc_delete_target_queue(struct nvmet_fc_tgt_queue *queue)
879{
880	struct nvmet_fc_tgtport *tgtport = queue->assoc->tgtport;
881	struct nvmet_fc_fcp_iod *fod = queue->fod;
882	struct nvmet_fc_defer_fcp_req *deferfcp, *tempptr;
883	unsigned long flags;
884	int i;
885	bool disconnect;
886
887	disconnect = atomic_xchg(&queue->connected, 0);
888
889	/* if not connected, nothing to do */
890	if (!disconnect)
891		return;
892
893	spin_lock_irqsave(&queue->qlock, flags);
894	/* abort outstanding io's */
895	for (i = 0; i < queue->sqsize; fod++, i++) {
896		if (fod->active) {
897			spin_lock(&fod->flock);
898			fod->abort = true;
899			/*
900			 * only call lldd abort routine if waiting for
901			 * writedata. other outstanding ops should finish
902			 * on their own.
903			 */
904			if (fod->writedataactive) {
905				fod->aborted = true;
906				spin_unlock(&fod->flock);
907				tgtport->ops->fcp_abort(
908					&tgtport->fc_target_port, fod->fcpreq);
909			} else
910				spin_unlock(&fod->flock);
911		}
912	}
913
914	/* Cleanup defer'ed IOs in queue */
915	list_for_each_entry_safe(deferfcp, tempptr, &queue->avail_defer_list,
916				req_list) {
917		list_del(&deferfcp->req_list);
918		kfree(deferfcp);
919	}
920
921	for (;;) {
922		deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
923				struct nvmet_fc_defer_fcp_req, req_list);
924		if (!deferfcp)
925			break;
926
927		list_del(&deferfcp->req_list);
928		spin_unlock_irqrestore(&queue->qlock, flags);
929
930		tgtport->ops->defer_rcv(&tgtport->fc_target_port,
931				deferfcp->fcp_req);
932
933		tgtport->ops->fcp_abort(&tgtport->fc_target_port,
934				deferfcp->fcp_req);
935
936		tgtport->ops->fcp_req_release(&tgtport->fc_target_port,
937				deferfcp->fcp_req);
938
939		/* release the queue lookup reference */
940		nvmet_fc_tgt_q_put(queue);
941
942		kfree(deferfcp);
943
944		spin_lock_irqsave(&queue->qlock, flags);
945	}
946	spin_unlock_irqrestore(&queue->qlock, flags);
947
948	flush_workqueue(queue->work_q);
949
950	nvmet_sq_destroy(&queue->nvme_sq);
951
952	nvmet_fc_tgt_q_put(queue);
953}
954
955static struct nvmet_fc_tgt_queue *
956nvmet_fc_find_target_queue(struct nvmet_fc_tgtport *tgtport,
957				u64 connection_id)
958{
959	struct nvmet_fc_tgt_assoc *assoc;
960	struct nvmet_fc_tgt_queue *queue;
961	u64 association_id = nvmet_fc_getassociationid(connection_id);
962	u16 qid = nvmet_fc_getqueueid(connection_id);
963
964	if (qid > NVMET_NR_QUEUES)
965		return NULL;
966
967	rcu_read_lock();
968	list_for_each_entry_rcu(assoc, &tgtport->assoc_list, a_list) {
969		if (association_id == assoc->association_id) {
970			queue = assoc->queues[qid];
971			if (queue &&
972			    (!atomic_read(&queue->connected) ||
973			     !nvmet_fc_tgt_q_get(queue)))
974				queue = NULL;
975			rcu_read_unlock();
976			return queue;
977		}
978	}
979	rcu_read_unlock();
980	return NULL;
981}
982
983static void
984nvmet_fc_hostport_free(struct kref *ref)
985{
986	struct nvmet_fc_hostport *hostport =
987		container_of(ref, struct nvmet_fc_hostport, ref);
988	struct nvmet_fc_tgtport *tgtport = hostport->tgtport;
989	unsigned long flags;
990
991	spin_lock_irqsave(&tgtport->lock, flags);
992	list_del(&hostport->host_list);
993	spin_unlock_irqrestore(&tgtport->lock, flags);
994	if (tgtport->ops->host_release && hostport->invalid)
995		tgtport->ops->host_release(hostport->hosthandle);
996	kfree(hostport);
997	nvmet_fc_tgtport_put(tgtport);
998}
999
1000static void
1001nvmet_fc_hostport_put(struct nvmet_fc_hostport *hostport)
1002{
1003	kref_put(&hostport->ref, nvmet_fc_hostport_free);
1004}
1005
1006static int
1007nvmet_fc_hostport_get(struct nvmet_fc_hostport *hostport)
1008{
1009	return kref_get_unless_zero(&hostport->ref);
1010}
1011
1012static void
1013nvmet_fc_free_hostport(struct nvmet_fc_hostport *hostport)
1014{
1015	/* if LLDD not implemented, leave as NULL */
1016	if (!hostport || !hostport->hosthandle)
1017		return;
1018
1019	nvmet_fc_hostport_put(hostport);
1020}
1021
1022static struct nvmet_fc_hostport *
1023nvmet_fc_match_hostport(struct nvmet_fc_tgtport *tgtport, void *hosthandle)
1024{
1025	struct nvmet_fc_hostport *host;
1026
1027	lockdep_assert_held(&tgtport->lock);
1028
1029	list_for_each_entry(host, &tgtport->host_list, host_list) {
1030		if (host->hosthandle == hosthandle && !host->invalid) {
1031			if (nvmet_fc_hostport_get(host))
1032				return host;
1033		}
1034	}
1035
1036	return NULL;
1037}
1038
1039static struct nvmet_fc_hostport *
1040nvmet_fc_alloc_hostport(struct nvmet_fc_tgtport *tgtport, void *hosthandle)
1041{
1042	struct nvmet_fc_hostport *newhost, *match = NULL;
1043	unsigned long flags;
1044
1045	/* if LLDD not implemented, leave as NULL */
1046	if (!hosthandle)
1047		return NULL;
1048
1049	/*
1050	 * take reference for what will be the newly allocated hostport if
1051	 * we end up using a new allocation
1052	 */
1053	if (!nvmet_fc_tgtport_get(tgtport))
1054		return ERR_PTR(-EINVAL);
1055
1056	spin_lock_irqsave(&tgtport->lock, flags);
1057	match = nvmet_fc_match_hostport(tgtport, hosthandle);
1058	spin_unlock_irqrestore(&tgtport->lock, flags);
1059
1060	if (match) {
1061		/* no new allocation - release reference */
1062		nvmet_fc_tgtport_put(tgtport);
1063		return match;
1064	}
1065
1066	newhost = kzalloc(sizeof(*newhost), GFP_KERNEL);
1067	if (!newhost) {
1068		/* no new allocation - release reference */
1069		nvmet_fc_tgtport_put(tgtport);
1070		return ERR_PTR(-ENOMEM);
1071	}
1072
1073	spin_lock_irqsave(&tgtport->lock, flags);
1074	match = nvmet_fc_match_hostport(tgtport, hosthandle);
1075	if (match) {
1076		/* new allocation not needed */
1077		kfree(newhost);
1078		newhost = match;
1079	} else {
1080		newhost->tgtport = tgtport;
1081		newhost->hosthandle = hosthandle;
1082		INIT_LIST_HEAD(&newhost->host_list);
1083		kref_init(&newhost->ref);
1084
1085		list_add_tail(&newhost->host_list, &tgtport->host_list);
1086	}
1087	spin_unlock_irqrestore(&tgtport->lock, flags);
1088
1089	return newhost;
1090}
1091
1092static void
1093nvmet_fc_delete_assoc(struct nvmet_fc_tgt_assoc *assoc)
1094{
1095	nvmet_fc_delete_target_assoc(assoc);
1096	nvmet_fc_tgt_a_put(assoc);
1097}
1098
1099static void
1100nvmet_fc_delete_assoc_work(struct work_struct *work)
1101{
1102	struct nvmet_fc_tgt_assoc *assoc =
1103		container_of(work, struct nvmet_fc_tgt_assoc, del_work);
1104	struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
1105
1106	nvmet_fc_delete_assoc(assoc);
1107	nvmet_fc_tgtport_put(tgtport);
1108}
1109
1110static void
1111nvmet_fc_schedule_delete_assoc(struct nvmet_fc_tgt_assoc *assoc)
1112{
1113	nvmet_fc_tgtport_get(assoc->tgtport);
1114	queue_work(nvmet_wq, &assoc->del_work);
1115}
1116
1117static bool
1118nvmet_fc_assoc_exists(struct nvmet_fc_tgtport *tgtport, u64 association_id)
1119{
1120	struct nvmet_fc_tgt_assoc *a;
1121	bool found = false;
1122
1123	rcu_read_lock();
1124	list_for_each_entry_rcu(a, &tgtport->assoc_list, a_list) {
1125		if (association_id == a->association_id) {
1126			found = true;
1127			break;
1128		}
1129	}
1130	rcu_read_unlock();
1131
1132	return found;
1133}
1134
1135static struct nvmet_fc_tgt_assoc *
1136nvmet_fc_alloc_target_assoc(struct nvmet_fc_tgtport *tgtport, void *hosthandle)
1137{
1138	struct nvmet_fc_tgt_assoc *assoc;
1139	unsigned long flags;
1140	bool done;
1141	u64 ran;
1142	int idx;
1143
1144	if (!tgtport->pe)
1145		return NULL;
1146
1147	assoc = kzalloc(sizeof(*assoc), GFP_KERNEL);
1148	if (!assoc)
1149		return NULL;
1150
1151	idx = ida_alloc(&tgtport->assoc_cnt, GFP_KERNEL);
1152	if (idx < 0)
1153		goto out_free_assoc;
1154
1155	assoc->hostport = nvmet_fc_alloc_hostport(tgtport, hosthandle);
1156	if (IS_ERR(assoc->hostport))
1157		goto out_ida;
1158
1159	assoc->tgtport = tgtport;
1160	assoc->a_id = idx;
1161	INIT_LIST_HEAD(&assoc->a_list);
1162	kref_init(&assoc->ref);
1163	INIT_WORK(&assoc->del_work, nvmet_fc_delete_assoc_work);
1164	atomic_set(&assoc->terminating, 0);
1165
1166	done = false;
1167	do {
1168		get_random_bytes(&ran, sizeof(ran) - BYTES_FOR_QID);
1169		ran = ran << BYTES_FOR_QID_SHIFT;
1170
1171		spin_lock_irqsave(&tgtport->lock, flags);
1172		if (!nvmet_fc_assoc_exists(tgtport, ran)) {
1173			assoc->association_id = ran;
1174			list_add_tail_rcu(&assoc->a_list, &tgtport->assoc_list);
1175			done = true;
1176		}
1177		spin_unlock_irqrestore(&tgtport->lock, flags);
1178	} while (!done);
1179
1180	return assoc;
1181
1182out_ida:
1183	ida_free(&tgtport->assoc_cnt, idx);
1184out_free_assoc:
1185	kfree(assoc);
1186	return NULL;
1187}
1188
1189static void
1190nvmet_fc_target_assoc_free(struct kref *ref)
1191{
1192	struct nvmet_fc_tgt_assoc *assoc =
1193		container_of(ref, struct nvmet_fc_tgt_assoc, ref);
1194	struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
1195	struct nvmet_fc_ls_iod	*oldls;
1196	unsigned long flags;
1197	int i;
1198
1199	for (i = NVMET_NR_QUEUES; i >= 0; i--) {
1200		if (assoc->queues[i])
1201			nvmet_fc_delete_target_queue(assoc->queues[i]);
1202	}
1203
1204	/* Send Disconnect now that all i/o has completed */
1205	nvmet_fc_xmt_disconnect_assoc(assoc);
1206
1207	nvmet_fc_free_hostport(assoc->hostport);
1208	spin_lock_irqsave(&tgtport->lock, flags);
1209	oldls = assoc->rcv_disconn;
1210	spin_unlock_irqrestore(&tgtport->lock, flags);
1211	/* if pending Rcv Disconnect Association LS, send rsp now */
1212	if (oldls)
1213		nvmet_fc_xmt_ls_rsp(tgtport, oldls);
1214	ida_free(&tgtport->assoc_cnt, assoc->a_id);
1215	dev_info(tgtport->dev,
1216		"{%d:%d} Association freed\n",
1217		tgtport->fc_target_port.port_num, assoc->a_id);
1218	kfree(assoc);
1219}
1220
1221static void
1222nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc)
1223{
1224	kref_put(&assoc->ref, nvmet_fc_target_assoc_free);
1225}
1226
1227static int
1228nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc)
1229{
1230	return kref_get_unless_zero(&assoc->ref);
1231}
1232
1233static void
1234nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc)
1235{
1236	struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
1237	unsigned long flags;
1238	int i, terminating;
1239
1240	terminating = atomic_xchg(&assoc->terminating, 1);
1241
1242	/* if already terminating, do nothing */
1243	if (terminating)
1244		return;
1245
1246	spin_lock_irqsave(&tgtport->lock, flags);
1247	list_del_rcu(&assoc->a_list);
1248	spin_unlock_irqrestore(&tgtport->lock, flags);
1249
1250	synchronize_rcu();
1251
1252	/* ensure all in-flight I/Os have been processed */
1253	for (i = NVMET_NR_QUEUES; i >= 0; i--) {
1254		if (assoc->queues[i])
1255			flush_workqueue(assoc->queues[i]->work_q);
1256	}
1257
1258	dev_info(tgtport->dev,
1259		"{%d:%d} Association deleted\n",
1260		tgtport->fc_target_port.port_num, assoc->a_id);
1261}
1262
1263static struct nvmet_fc_tgt_assoc *
1264nvmet_fc_find_target_assoc(struct nvmet_fc_tgtport *tgtport,
1265				u64 association_id)
1266{
1267	struct nvmet_fc_tgt_assoc *assoc;
1268	struct nvmet_fc_tgt_assoc *ret = NULL;
1269
1270	rcu_read_lock();
1271	list_for_each_entry_rcu(assoc, &tgtport->assoc_list, a_list) {
1272		if (association_id == assoc->association_id) {
1273			ret = assoc;
1274			if (!nvmet_fc_tgt_a_get(assoc))
1275				ret = NULL;
1276			break;
1277		}
1278	}
1279	rcu_read_unlock();
1280
1281	return ret;
1282}
1283
1284static void
1285nvmet_fc_portentry_bind(struct nvmet_fc_tgtport *tgtport,
1286			struct nvmet_fc_port_entry *pe,
1287			struct nvmet_port *port)
1288{
1289	lockdep_assert_held(&nvmet_fc_tgtlock);
1290
1291	pe->tgtport = tgtport;
1292	tgtport->pe = pe;
1293
1294	pe->port = port;
1295	port->priv = pe;
1296
1297	pe->node_name = tgtport->fc_target_port.node_name;
1298	pe->port_name = tgtport->fc_target_port.port_name;
1299	INIT_LIST_HEAD(&pe->pe_list);
1300
1301	list_add_tail(&pe->pe_list, &nvmet_fc_portentry_list);
1302}
1303
1304static void
1305nvmet_fc_portentry_unbind(struct nvmet_fc_port_entry *pe)
1306{
1307	unsigned long flags;
1308
1309	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1310	if (pe->tgtport)
1311		pe->tgtport->pe = NULL;
1312	list_del(&pe->pe_list);
1313	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1314}
1315
1316/*
1317 * called when a targetport deregisters. Breaks the relationship
1318 * with the nvmet port, but leaves the port_entry in place so that
1319 * re-registration can resume operation.
1320 */
1321static void
1322nvmet_fc_portentry_unbind_tgt(struct nvmet_fc_tgtport *tgtport)
1323{
1324	struct nvmet_fc_port_entry *pe;
1325	unsigned long flags;
1326
1327	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1328	pe = tgtport->pe;
1329	if (pe)
1330		pe->tgtport = NULL;
1331	tgtport->pe = NULL;
1332	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1333}
1334
1335/*
1336 * called when a new targetport is registered. Looks in the
1337 * existing nvmet port_entries to see if the nvmet layer is
1338 * configured for the targetport's wwn's. (the targetport existed,
1339 * nvmet configured, the lldd unregistered the tgtport, and is now
1340 * reregistering the same targetport).  If so, set the nvmet port
1341 * port entry on the targetport.
1342 */
1343static void
1344nvmet_fc_portentry_rebind_tgt(struct nvmet_fc_tgtport *tgtport)
1345{
1346	struct nvmet_fc_port_entry *pe;
1347	unsigned long flags;
1348
1349	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1350	list_for_each_entry(pe, &nvmet_fc_portentry_list, pe_list) {
1351		if (tgtport->fc_target_port.node_name == pe->node_name &&
1352		    tgtport->fc_target_port.port_name == pe->port_name) {
1353			WARN_ON(pe->tgtport);
1354			tgtport->pe = pe;
1355			pe->tgtport = tgtport;
1356			break;
1357		}
1358	}
1359	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1360}
1361
1362/**
1363 * nvmet_fc_register_targetport - transport entry point called by an
1364 *                              LLDD to register the existence of a local
1365 *                              NVME subystem FC port.
1366 * @pinfo:     pointer to information about the port to be registered
1367 * @template:  LLDD entrypoints and operational parameters for the port
1368 * @dev:       physical hardware device node port corresponds to. Will be
1369 *             used for DMA mappings
1370 * @portptr:   pointer to a local port pointer. Upon success, the routine
1371 *             will allocate a nvme_fc_local_port structure and place its
1372 *             address in the local port pointer. Upon failure, local port
1373 *             pointer will be set to NULL.
1374 *
1375 * Returns:
1376 * a completion status. Must be 0 upon success; a negative errno
1377 * (ex: -ENXIO) upon failure.
1378 */
1379int
1380nvmet_fc_register_targetport(struct nvmet_fc_port_info *pinfo,
1381			struct nvmet_fc_target_template *template,
1382			struct device *dev,
1383			struct nvmet_fc_target_port **portptr)
1384{
1385	struct nvmet_fc_tgtport *newrec;
1386	unsigned long flags;
1387	int ret, idx;
1388
1389	if (!template->xmt_ls_rsp || !template->fcp_op ||
1390	    !template->fcp_abort ||
1391	    !template->fcp_req_release || !template->targetport_delete ||
1392	    !template->max_hw_queues || !template->max_sgl_segments ||
1393	    !template->max_dif_sgl_segments || !template->dma_boundary) {
1394		ret = -EINVAL;
1395		goto out_regtgt_failed;
1396	}
1397
1398	newrec = kzalloc((sizeof(*newrec) + template->target_priv_sz),
1399			 GFP_KERNEL);
1400	if (!newrec) {
1401		ret = -ENOMEM;
1402		goto out_regtgt_failed;
1403	}
1404
1405	idx = ida_alloc(&nvmet_fc_tgtport_cnt, GFP_KERNEL);
1406	if (idx < 0) {
1407		ret = -ENOSPC;
1408		goto out_fail_kfree;
1409	}
1410
1411	if (!get_device(dev) && dev) {
1412		ret = -ENODEV;
1413		goto out_ida_put;
1414	}
1415
1416	newrec->fc_target_port.node_name = pinfo->node_name;
1417	newrec->fc_target_port.port_name = pinfo->port_name;
1418	if (template->target_priv_sz)
1419		newrec->fc_target_port.private = &newrec[1];
1420	else
1421		newrec->fc_target_port.private = NULL;
1422	newrec->fc_target_port.port_id = pinfo->port_id;
1423	newrec->fc_target_port.port_num = idx;
1424	INIT_LIST_HEAD(&newrec->tgt_list);
1425	newrec->dev = dev;
1426	newrec->ops = template;
1427	spin_lock_init(&newrec->lock);
1428	INIT_LIST_HEAD(&newrec->ls_rcv_list);
1429	INIT_LIST_HEAD(&newrec->ls_req_list);
1430	INIT_LIST_HEAD(&newrec->ls_busylist);
1431	INIT_LIST_HEAD(&newrec->assoc_list);
1432	INIT_LIST_HEAD(&newrec->host_list);
1433	kref_init(&newrec->ref);
1434	ida_init(&newrec->assoc_cnt);
1435	newrec->max_sg_cnt = template->max_sgl_segments;
1436	INIT_WORK(&newrec->put_work, nvmet_fc_put_tgtport_work);
1437
1438	ret = nvmet_fc_alloc_ls_iodlist(newrec);
1439	if (ret) {
1440		ret = -ENOMEM;
1441		goto out_free_newrec;
1442	}
1443
1444	nvmet_fc_portentry_rebind_tgt(newrec);
1445
1446	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1447	list_add_tail(&newrec->tgt_list, &nvmet_fc_target_list);
1448	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1449
1450	*portptr = &newrec->fc_target_port;
1451	return 0;
1452
1453out_free_newrec:
1454	put_device(dev);
1455out_ida_put:
1456	ida_free(&nvmet_fc_tgtport_cnt, idx);
1457out_fail_kfree:
1458	kfree(newrec);
1459out_regtgt_failed:
1460	*portptr = NULL;
1461	return ret;
1462}
1463EXPORT_SYMBOL_GPL(nvmet_fc_register_targetport);
1464
1465
1466static void
1467nvmet_fc_free_tgtport(struct kref *ref)
1468{
1469	struct nvmet_fc_tgtport *tgtport =
1470		container_of(ref, struct nvmet_fc_tgtport, ref);
1471	struct device *dev = tgtport->dev;
1472	unsigned long flags;
1473
1474	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1475	list_del(&tgtport->tgt_list);
1476	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1477
1478	nvmet_fc_free_ls_iodlist(tgtport);
1479
1480	/* let the LLDD know we've finished tearing it down */
1481	tgtport->ops->targetport_delete(&tgtport->fc_target_port);
1482
1483	ida_free(&nvmet_fc_tgtport_cnt,
1484			tgtport->fc_target_port.port_num);
1485
1486	ida_destroy(&tgtport->assoc_cnt);
1487
1488	kfree(tgtport);
1489
1490	put_device(dev);
1491}
1492
1493static void
1494nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport)
1495{
1496	kref_put(&tgtport->ref, nvmet_fc_free_tgtport);
1497}
1498
1499static int
1500nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport)
1501{
1502	return kref_get_unless_zero(&tgtport->ref);
1503}
1504
1505static void
1506__nvmet_fc_free_assocs(struct nvmet_fc_tgtport *tgtport)
1507{
1508	struct nvmet_fc_tgt_assoc *assoc;
1509
1510	rcu_read_lock();
1511	list_for_each_entry_rcu(assoc, &tgtport->assoc_list, a_list) {
1512		if (!nvmet_fc_tgt_a_get(assoc))
1513			continue;
1514		nvmet_fc_schedule_delete_assoc(assoc);
1515		nvmet_fc_tgt_a_put(assoc);
1516	}
1517	rcu_read_unlock();
1518}
1519
1520/**
1521 * nvmet_fc_invalidate_host - transport entry point called by an LLDD
1522 *                       to remove references to a hosthandle for LS's.
1523 *
1524 * The nvmet-fc layer ensures that any references to the hosthandle
1525 * on the targetport are forgotten (set to NULL).  The LLDD will
1526 * typically call this when a login with a remote host port has been
1527 * lost, thus LS's for the remote host port are no longer possible.
1528 *
1529 * If an LS request is outstanding to the targetport/hosthandle (or
1530 * issued concurrently with the call to invalidate the host), the
1531 * LLDD is responsible for terminating/aborting the LS and completing
1532 * the LS request. It is recommended that these terminations/aborts
1533 * occur after calling to invalidate the host handle to avoid additional
1534 * retries by the nvmet-fc transport. The nvmet-fc transport may
1535 * continue to reference host handle while it cleans up outstanding
1536 * NVME associations. The nvmet-fc transport will call the
1537 * ops->host_release() callback to notify the LLDD that all references
1538 * are complete and the related host handle can be recovered.
1539 * Note: if there are no references, the callback may be called before
1540 * the invalidate host call returns.
1541 *
1542 * @target_port: pointer to the (registered) target port that a prior
1543 *              LS was received on and which supplied the transport the
1544 *              hosthandle.
1545 * @hosthandle: the handle (pointer) that represents the host port
1546 *              that no longer has connectivity and that LS's should
1547 *              no longer be directed to.
1548 */
1549void
1550nvmet_fc_invalidate_host(struct nvmet_fc_target_port *target_port,
1551			void *hosthandle)
1552{
1553	struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
1554	struct nvmet_fc_tgt_assoc *assoc, *next;
1555	unsigned long flags;
1556	bool noassoc = true;
1557
1558	spin_lock_irqsave(&tgtport->lock, flags);
1559	list_for_each_entry_safe(assoc, next,
1560				&tgtport->assoc_list, a_list) {
1561		if (assoc->hostport->hosthandle != hosthandle)
1562			continue;
1563		if (!nvmet_fc_tgt_a_get(assoc))
1564			continue;
1565		assoc->hostport->invalid = 1;
1566		noassoc = false;
1567		nvmet_fc_schedule_delete_assoc(assoc);
1568		nvmet_fc_tgt_a_put(assoc);
1569	}
1570	spin_unlock_irqrestore(&tgtport->lock, flags);
1571
1572	/* if there's nothing to wait for - call the callback */
1573	if (noassoc && tgtport->ops->host_release)
1574		tgtport->ops->host_release(hosthandle);
1575}
1576EXPORT_SYMBOL_GPL(nvmet_fc_invalidate_host);
1577
1578/*
1579 * nvmet layer has called to terminate an association
1580 */
1581static void
1582nvmet_fc_delete_ctrl(struct nvmet_ctrl *ctrl)
1583{
1584	struct nvmet_fc_tgtport *tgtport, *next;
1585	struct nvmet_fc_tgt_assoc *assoc;
1586	struct nvmet_fc_tgt_queue *queue;
1587	unsigned long flags;
1588	bool found_ctrl = false;
1589
1590	/* this is a bit ugly, but don't want to make locks layered */
1591	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1592	list_for_each_entry_safe(tgtport, next, &nvmet_fc_target_list,
1593			tgt_list) {
1594		if (!nvmet_fc_tgtport_get(tgtport))
1595			continue;
1596		spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1597
1598		rcu_read_lock();
1599		list_for_each_entry_rcu(assoc, &tgtport->assoc_list, a_list) {
1600			queue = assoc->queues[0];
1601			if (queue && queue->nvme_sq.ctrl == ctrl) {
1602				if (nvmet_fc_tgt_a_get(assoc))
1603					found_ctrl = true;
1604				break;
1605			}
1606		}
1607		rcu_read_unlock();
1608
1609		nvmet_fc_tgtport_put(tgtport);
1610
1611		if (found_ctrl) {
1612			nvmet_fc_schedule_delete_assoc(assoc);
1613			nvmet_fc_tgt_a_put(assoc);
1614			return;
1615		}
1616
1617		spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1618	}
1619	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1620}
1621
1622/**
1623 * nvmet_fc_unregister_targetport - transport entry point called by an
1624 *                              LLDD to deregister/remove a previously
1625 *                              registered a local NVME subsystem FC port.
1626 * @target_port: pointer to the (registered) target port that is to be
1627 *               deregistered.
1628 *
1629 * Returns:
1630 * a completion status. Must be 0 upon success; a negative errno
1631 * (ex: -ENXIO) upon failure.
1632 */
1633int
1634nvmet_fc_unregister_targetport(struct nvmet_fc_target_port *target_port)
1635{
1636	struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
1637
1638	nvmet_fc_portentry_unbind_tgt(tgtport);
1639
1640	/* terminate any outstanding associations */
1641	__nvmet_fc_free_assocs(tgtport);
1642
1643	flush_workqueue(nvmet_wq);
1644
1645	/*
1646	 * should terminate LS's as well. However, LS's will be generated
1647	 * at the tail end of association termination, so they likely don't
1648	 * exist yet. And even if they did, it's worthwhile to just let
1649	 * them finish and targetport ref counting will clean things up.
1650	 */
1651
1652	nvmet_fc_tgtport_put(tgtport);
1653
1654	return 0;
1655}
1656EXPORT_SYMBOL_GPL(nvmet_fc_unregister_targetport);
1657
1658
1659/* ********************** FC-NVME LS RCV Handling ************************* */
1660
1661
1662static void
1663nvmet_fc_ls_create_association(struct nvmet_fc_tgtport *tgtport,
1664			struct nvmet_fc_ls_iod *iod)
1665{
1666	struct fcnvme_ls_cr_assoc_rqst *rqst = &iod->rqstbuf->rq_cr_assoc;
1667	struct fcnvme_ls_cr_assoc_acc *acc = &iod->rspbuf->rsp_cr_assoc;
1668	struct nvmet_fc_tgt_queue *queue;
1669	int ret = 0;
1670
1671	memset(acc, 0, sizeof(*acc));
1672
1673	/*
1674	 * FC-NVME spec changes. There are initiators sending different
1675	 * lengths as padding sizes for Create Association Cmd descriptor
1676	 * was incorrect.
1677	 * Accept anything of "minimum" length. Assume format per 1.15
1678	 * spec (with HOSTID reduced to 16 bytes), ignore how long the
1679	 * trailing pad length is.
1680	 */
1681	if (iod->rqstdatalen < FCNVME_LSDESC_CRA_RQST_MINLEN)
1682		ret = VERR_CR_ASSOC_LEN;
1683	else if (be32_to_cpu(rqst->desc_list_len) <
1684			FCNVME_LSDESC_CRA_RQST_MIN_LISTLEN)
1685		ret = VERR_CR_ASSOC_RQST_LEN;
1686	else if (rqst->assoc_cmd.desc_tag !=
1687			cpu_to_be32(FCNVME_LSDESC_CREATE_ASSOC_CMD))
1688		ret = VERR_CR_ASSOC_CMD;
1689	else if (be32_to_cpu(rqst->assoc_cmd.desc_len) <
1690			FCNVME_LSDESC_CRA_CMD_DESC_MIN_DESCLEN)
1691		ret = VERR_CR_ASSOC_CMD_LEN;
1692	else if (!rqst->assoc_cmd.ersp_ratio ||
1693		 (be16_to_cpu(rqst->assoc_cmd.ersp_ratio) >=
1694				be16_to_cpu(rqst->assoc_cmd.sqsize)))
1695		ret = VERR_ERSP_RATIO;
1696
1697	else {
1698		/* new association w/ admin queue */
1699		iod->assoc = nvmet_fc_alloc_target_assoc(
1700						tgtport, iod->hosthandle);
1701		if (!iod->assoc)
1702			ret = VERR_ASSOC_ALLOC_FAIL;
1703		else {
1704			queue = nvmet_fc_alloc_target_queue(iod->assoc, 0,
1705					be16_to_cpu(rqst->assoc_cmd.sqsize));
1706			if (!queue) {
1707				ret = VERR_QUEUE_ALLOC_FAIL;
1708				nvmet_fc_tgt_a_put(iod->assoc);
1709			}
1710		}
1711	}
1712
1713	if (ret) {
1714		dev_err(tgtport->dev,
1715			"Create Association LS failed: %s\n",
1716			validation_errors[ret]);
1717		iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1718				sizeof(*acc), rqst->w0.ls_cmd,
1719				FCNVME_RJT_RC_LOGIC,
1720				FCNVME_RJT_EXP_NONE, 0);
1721		return;
1722	}
1723
1724	queue->ersp_ratio = be16_to_cpu(rqst->assoc_cmd.ersp_ratio);
1725	atomic_set(&queue->connected, 1);
1726	queue->sqhd = 0;	/* best place to init value */
1727
1728	dev_info(tgtport->dev,
1729		"{%d:%d} Association created\n",
1730		tgtport->fc_target_port.port_num, iod->assoc->a_id);
1731
1732	/* format a response */
1733
1734	iod->lsrsp->rsplen = sizeof(*acc);
1735
1736	nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1737			fcnvme_lsdesc_len(
1738				sizeof(struct fcnvme_ls_cr_assoc_acc)),
1739			FCNVME_LS_CREATE_ASSOCIATION);
1740	acc->associd.desc_tag = cpu_to_be32(FCNVME_LSDESC_ASSOC_ID);
1741	acc->associd.desc_len =
1742			fcnvme_lsdesc_len(
1743				sizeof(struct fcnvme_lsdesc_assoc_id));
1744	acc->associd.association_id =
1745			cpu_to_be64(nvmet_fc_makeconnid(iod->assoc, 0));
1746	acc->connectid.desc_tag = cpu_to_be32(FCNVME_LSDESC_CONN_ID);
1747	acc->connectid.desc_len =
1748			fcnvme_lsdesc_len(
1749				sizeof(struct fcnvme_lsdesc_conn_id));
1750	acc->connectid.connection_id = acc->associd.association_id;
1751}
1752
1753static void
1754nvmet_fc_ls_create_connection(struct nvmet_fc_tgtport *tgtport,
1755			struct nvmet_fc_ls_iod *iod)
1756{
1757	struct fcnvme_ls_cr_conn_rqst *rqst = &iod->rqstbuf->rq_cr_conn;
1758	struct fcnvme_ls_cr_conn_acc *acc = &iod->rspbuf->rsp_cr_conn;
1759	struct nvmet_fc_tgt_queue *queue;
1760	int ret = 0;
1761
1762	memset(acc, 0, sizeof(*acc));
1763
1764	if (iod->rqstdatalen < sizeof(struct fcnvme_ls_cr_conn_rqst))
1765		ret = VERR_CR_CONN_LEN;
1766	else if (rqst->desc_list_len !=
1767			fcnvme_lsdesc_len(
1768				sizeof(struct fcnvme_ls_cr_conn_rqst)))
1769		ret = VERR_CR_CONN_RQST_LEN;
1770	else if (rqst->associd.desc_tag != cpu_to_be32(FCNVME_LSDESC_ASSOC_ID))
1771		ret = VERR_ASSOC_ID;
1772	else if (rqst->associd.desc_len !=
1773			fcnvme_lsdesc_len(
1774				sizeof(struct fcnvme_lsdesc_assoc_id)))
1775		ret = VERR_ASSOC_ID_LEN;
1776	else if (rqst->connect_cmd.desc_tag !=
1777			cpu_to_be32(FCNVME_LSDESC_CREATE_CONN_CMD))
1778		ret = VERR_CR_CONN_CMD;
1779	else if (rqst->connect_cmd.desc_len !=
1780			fcnvme_lsdesc_len(
1781				sizeof(struct fcnvme_lsdesc_cr_conn_cmd)))
1782		ret = VERR_CR_CONN_CMD_LEN;
1783	else if (!rqst->connect_cmd.ersp_ratio ||
1784		 (be16_to_cpu(rqst->connect_cmd.ersp_ratio) >=
1785				be16_to_cpu(rqst->connect_cmd.sqsize)))
1786		ret = VERR_ERSP_RATIO;
1787
1788	else {
1789		/* new io queue */
1790		iod->assoc = nvmet_fc_find_target_assoc(tgtport,
1791				be64_to_cpu(rqst->associd.association_id));
1792		if (!iod->assoc)
1793			ret = VERR_NO_ASSOC;
1794		else {
1795			queue = nvmet_fc_alloc_target_queue(iod->assoc,
1796					be16_to_cpu(rqst->connect_cmd.qid),
1797					be16_to_cpu(rqst->connect_cmd.sqsize));
1798			if (!queue)
1799				ret = VERR_QUEUE_ALLOC_FAIL;
1800
1801			/* release get taken in nvmet_fc_find_target_assoc */
1802			nvmet_fc_tgt_a_put(iod->assoc);
1803		}
1804	}
1805
1806	if (ret) {
1807		dev_err(tgtport->dev,
1808			"Create Connection LS failed: %s\n",
1809			validation_errors[ret]);
1810		iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1811				sizeof(*acc), rqst->w0.ls_cmd,
1812				(ret == VERR_NO_ASSOC) ?
1813					FCNVME_RJT_RC_INV_ASSOC :
1814					FCNVME_RJT_RC_LOGIC,
1815				FCNVME_RJT_EXP_NONE, 0);
1816		return;
1817	}
1818
1819	queue->ersp_ratio = be16_to_cpu(rqst->connect_cmd.ersp_ratio);
1820	atomic_set(&queue->connected, 1);
1821	queue->sqhd = 0;	/* best place to init value */
1822
1823	/* format a response */
1824
1825	iod->lsrsp->rsplen = sizeof(*acc);
1826
1827	nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1828			fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_cr_conn_acc)),
1829			FCNVME_LS_CREATE_CONNECTION);
1830	acc->connectid.desc_tag = cpu_to_be32(FCNVME_LSDESC_CONN_ID);
1831	acc->connectid.desc_len =
1832			fcnvme_lsdesc_len(
1833				sizeof(struct fcnvme_lsdesc_conn_id));
1834	acc->connectid.connection_id =
1835			cpu_to_be64(nvmet_fc_makeconnid(iod->assoc,
1836				be16_to_cpu(rqst->connect_cmd.qid)));
1837}
1838
1839/*
1840 * Returns true if the LS response is to be transmit
1841 * Returns false if the LS response is to be delayed
1842 */
1843static int
1844nvmet_fc_ls_disconnect(struct nvmet_fc_tgtport *tgtport,
1845			struct nvmet_fc_ls_iod *iod)
1846{
1847	struct fcnvme_ls_disconnect_assoc_rqst *rqst =
1848						&iod->rqstbuf->rq_dis_assoc;
1849	struct fcnvme_ls_disconnect_assoc_acc *acc =
1850						&iod->rspbuf->rsp_dis_assoc;
1851	struct nvmet_fc_tgt_assoc *assoc = NULL;
1852	struct nvmet_fc_ls_iod *oldls = NULL;
1853	unsigned long flags;
1854	int ret = 0;
1855
1856	memset(acc, 0, sizeof(*acc));
1857
1858	ret = nvmefc_vldt_lsreq_discon_assoc(iod->rqstdatalen, rqst);
1859	if (!ret) {
1860		/* match an active association - takes an assoc ref if !NULL */
1861		assoc = nvmet_fc_find_target_assoc(tgtport,
1862				be64_to_cpu(rqst->associd.association_id));
1863		iod->assoc = assoc;
1864		if (!assoc)
1865			ret = VERR_NO_ASSOC;
1866	}
1867
1868	if (ret || !assoc) {
1869		dev_err(tgtport->dev,
1870			"Disconnect LS failed: %s\n",
1871			validation_errors[ret]);
1872		iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1873				sizeof(*acc), rqst->w0.ls_cmd,
1874				(ret == VERR_NO_ASSOC) ?
1875					FCNVME_RJT_RC_INV_ASSOC :
1876					FCNVME_RJT_RC_LOGIC,
1877				FCNVME_RJT_EXP_NONE, 0);
1878		return true;
1879	}
1880
1881	/* format a response */
1882
1883	iod->lsrsp->rsplen = sizeof(*acc);
1884
1885	nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1886			fcnvme_lsdesc_len(
1887				sizeof(struct fcnvme_ls_disconnect_assoc_acc)),
1888			FCNVME_LS_DISCONNECT_ASSOC);
1889
1890	/*
1891	 * The rules for LS response says the response cannot
1892	 * go back until ABTS's have been sent for all outstanding
1893	 * I/O and a Disconnect Association LS has been sent.
1894	 * So... save off the Disconnect LS to send the response
1895	 * later. If there was a prior LS already saved, replace
1896	 * it with the newer one and send a can't perform reject
1897	 * on the older one.
1898	 */
1899	spin_lock_irqsave(&tgtport->lock, flags);
1900	oldls = assoc->rcv_disconn;
1901	assoc->rcv_disconn = iod;
1902	spin_unlock_irqrestore(&tgtport->lock, flags);
1903
1904	if (oldls) {
1905		dev_info(tgtport->dev,
1906			"{%d:%d} Multiple Disconnect Association LS's "
1907			"received\n",
1908			tgtport->fc_target_port.port_num, assoc->a_id);
1909		/* overwrite good response with bogus failure */
1910		oldls->lsrsp->rsplen = nvme_fc_format_rjt(oldls->rspbuf,
1911						sizeof(*iod->rspbuf),
1912						/* ok to use rqst, LS is same */
1913						rqst->w0.ls_cmd,
1914						FCNVME_RJT_RC_UNAB,
1915						FCNVME_RJT_EXP_NONE, 0);
1916		nvmet_fc_xmt_ls_rsp(tgtport, oldls);
1917	}
1918
1919	nvmet_fc_schedule_delete_assoc(assoc);
1920	nvmet_fc_tgt_a_put(assoc);
1921
1922	return false;
1923}
1924
1925
1926/* *********************** NVME Ctrl Routines **************************** */
1927
1928
1929static void nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req);
1930
1931static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops;
1932
1933static void
1934nvmet_fc_xmt_ls_rsp_done(struct nvmefc_ls_rsp *lsrsp)
1935{
1936	struct nvmet_fc_ls_iod *iod = lsrsp->nvme_fc_private;
1937	struct nvmet_fc_tgtport *tgtport = iod->tgtport;
1938
1939	fc_dma_sync_single_for_cpu(tgtport->dev, iod->rspdma,
1940				sizeof(*iod->rspbuf), DMA_TO_DEVICE);
1941	nvmet_fc_free_ls_iod(tgtport, iod);
1942	nvmet_fc_tgtport_put(tgtport);
1943}
1944
1945static void
1946nvmet_fc_xmt_ls_rsp(struct nvmet_fc_tgtport *tgtport,
1947				struct nvmet_fc_ls_iod *iod)
1948{
1949	int ret;
1950
1951	fc_dma_sync_single_for_device(tgtport->dev, iod->rspdma,
1952				  sizeof(*iod->rspbuf), DMA_TO_DEVICE);
1953
1954	ret = tgtport->ops->xmt_ls_rsp(&tgtport->fc_target_port, iod->lsrsp);
1955	if (ret)
1956		nvmet_fc_xmt_ls_rsp_done(iod->lsrsp);
1957}
1958
1959/*
1960 * Actual processing routine for received FC-NVME LS Requests from the LLD
1961 */
1962static void
1963nvmet_fc_handle_ls_rqst(struct nvmet_fc_tgtport *tgtport,
1964			struct nvmet_fc_ls_iod *iod)
1965{
1966	struct fcnvme_ls_rqst_w0 *w0 = &iod->rqstbuf->rq_cr_assoc.w0;
1967	bool sendrsp = true;
1968
1969	iod->lsrsp->nvme_fc_private = iod;
1970	iod->lsrsp->rspbuf = iod->rspbuf;
1971	iod->lsrsp->rspdma = iod->rspdma;
1972	iod->lsrsp->done = nvmet_fc_xmt_ls_rsp_done;
1973	/* Be preventative. handlers will later set to valid length */
1974	iod->lsrsp->rsplen = 0;
1975
1976	iod->assoc = NULL;
1977
1978	/*
1979	 * handlers:
1980	 *   parse request input, execute the request, and format the
1981	 *   LS response
1982	 */
1983	switch (w0->ls_cmd) {
1984	case FCNVME_LS_CREATE_ASSOCIATION:
1985		/* Creates Association and initial Admin Queue/Connection */
1986		nvmet_fc_ls_create_association(tgtport, iod);
1987		break;
1988	case FCNVME_LS_CREATE_CONNECTION:
1989		/* Creates an IO Queue/Connection */
1990		nvmet_fc_ls_create_connection(tgtport, iod);
1991		break;
1992	case FCNVME_LS_DISCONNECT_ASSOC:
1993		/* Terminate a Queue/Connection or the Association */
1994		sendrsp = nvmet_fc_ls_disconnect(tgtport, iod);
1995		break;
1996	default:
1997		iod->lsrsp->rsplen = nvme_fc_format_rjt(iod->rspbuf,
1998				sizeof(*iod->rspbuf), w0->ls_cmd,
1999				FCNVME_RJT_RC_INVAL, FCNVME_RJT_EXP_NONE, 0);
2000	}
2001
2002	if (sendrsp)
2003		nvmet_fc_xmt_ls_rsp(tgtport, iod);
2004}
2005
2006/*
2007 * Actual processing routine for received FC-NVME LS Requests from the LLD
2008 */
2009static void
2010nvmet_fc_handle_ls_rqst_work(struct work_struct *work)
2011{
2012	struct nvmet_fc_ls_iod *iod =
2013		container_of(work, struct nvmet_fc_ls_iod, work);
2014	struct nvmet_fc_tgtport *tgtport = iod->tgtport;
2015
2016	nvmet_fc_handle_ls_rqst(tgtport, iod);
2017}
2018
2019
2020/**
2021 * nvmet_fc_rcv_ls_req - transport entry point called by an LLDD
2022 *                       upon the reception of a NVME LS request.
2023 *
2024 * The nvmet-fc layer will copy payload to an internal structure for
2025 * processing.  As such, upon completion of the routine, the LLDD may
2026 * immediately free/reuse the LS request buffer passed in the call.
2027 *
2028 * If this routine returns error, the LLDD should abort the exchange.
2029 *
2030 * @target_port: pointer to the (registered) target port the LS was
2031 *              received on.
2032 * @hosthandle: pointer to the host specific data, gets stored in iod.
2033 * @lsrsp:      pointer to a lsrsp structure to be used to reference
2034 *              the exchange corresponding to the LS.
2035 * @lsreqbuf:   pointer to the buffer containing the LS Request
2036 * @lsreqbuf_len: length, in bytes, of the received LS request
2037 */
2038int
2039nvmet_fc_rcv_ls_req(struct nvmet_fc_target_port *target_port,
2040			void *hosthandle,
2041			struct nvmefc_ls_rsp *lsrsp,
2042			void *lsreqbuf, u32 lsreqbuf_len)
2043{
2044	struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
2045	struct nvmet_fc_ls_iod *iod;
2046	struct fcnvme_ls_rqst_w0 *w0 = (struct fcnvme_ls_rqst_w0 *)lsreqbuf;
2047
2048	if (lsreqbuf_len > sizeof(union nvmefc_ls_requests)) {
2049		dev_info(tgtport->dev,
2050			"RCV %s LS failed: payload too large (%d)\n",
2051			(w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2052				nvmefc_ls_names[w0->ls_cmd] : "",
2053			lsreqbuf_len);
2054		return -E2BIG;
2055	}
2056
2057	if (!nvmet_fc_tgtport_get(tgtport)) {
2058		dev_info(tgtport->dev,
2059			"RCV %s LS failed: target deleting\n",
2060			(w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2061				nvmefc_ls_names[w0->ls_cmd] : "");
2062		return -ESHUTDOWN;
2063	}
2064
2065	iod = nvmet_fc_alloc_ls_iod(tgtport);
2066	if (!iod) {
2067		dev_info(tgtport->dev,
2068			"RCV %s LS failed: context allocation failed\n",
2069			(w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2070				nvmefc_ls_names[w0->ls_cmd] : "");
2071		nvmet_fc_tgtport_put(tgtport);
2072		return -ENOENT;
2073	}
2074
2075	iod->lsrsp = lsrsp;
2076	iod->fcpreq = NULL;
2077	memcpy(iod->rqstbuf, lsreqbuf, lsreqbuf_len);
2078	iod->rqstdatalen = lsreqbuf_len;
2079	iod->hosthandle = hosthandle;
2080
2081	queue_work(nvmet_wq, &iod->work);
2082
2083	return 0;
2084}
2085EXPORT_SYMBOL_GPL(nvmet_fc_rcv_ls_req);
2086
2087
2088/*
2089 * **********************
2090 * Start of FCP handling
2091 * **********************
2092 */
2093
2094static int
2095nvmet_fc_alloc_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
2096{
2097	struct scatterlist *sg;
2098	unsigned int nent;
2099
2100	sg = sgl_alloc(fod->req.transfer_len, GFP_KERNEL, &nent);
2101	if (!sg)
2102		goto out;
2103
2104	fod->data_sg = sg;
2105	fod->data_sg_cnt = nent;
2106	fod->data_sg_cnt = fc_dma_map_sg(fod->tgtport->dev, sg, nent,
2107				((fod->io_dir == NVMET_FCP_WRITE) ?
2108					DMA_FROM_DEVICE : DMA_TO_DEVICE));
2109				/* note: write from initiator perspective */
2110	fod->next_sg = fod->data_sg;
2111
2112	return 0;
2113
2114out:
2115	return NVME_SC_INTERNAL;
2116}
2117
2118static void
2119nvmet_fc_free_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
2120{
2121	if (!fod->data_sg || !fod->data_sg_cnt)
2122		return;
2123
2124	fc_dma_unmap_sg(fod->tgtport->dev, fod->data_sg, fod->data_sg_cnt,
2125				((fod->io_dir == NVMET_FCP_WRITE) ?
2126					DMA_FROM_DEVICE : DMA_TO_DEVICE));
2127	sgl_free(fod->data_sg);
2128	fod->data_sg = NULL;
2129	fod->data_sg_cnt = 0;
2130}
2131
2132
2133static bool
2134queue_90percent_full(struct nvmet_fc_tgt_queue *q, u32 sqhd)
2135{
2136	u32 sqtail, used;
2137
2138	/* egad, this is ugly. And sqtail is just a best guess */
2139	sqtail = atomic_read(&q->sqtail) % q->sqsize;
2140
2141	used = (sqtail < sqhd) ? (sqtail + q->sqsize - sqhd) : (sqtail - sqhd);
2142	return ((used * 10) >= (((u32)(q->sqsize - 1) * 9)));
2143}
2144
2145/*
2146 * Prep RSP payload.
2147 * May be a NVMET_FCOP_RSP or NVMET_FCOP_READDATA_RSP op
2148 */
2149static void
2150nvmet_fc_prep_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
2151				struct nvmet_fc_fcp_iod *fod)
2152{
2153	struct nvme_fc_ersp_iu *ersp = &fod->rspiubuf;
2154	struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
2155	struct nvme_completion *cqe = &ersp->cqe;
2156	u32 *cqewd = (u32 *)cqe;
2157	bool send_ersp = false;
2158	u32 rsn, rspcnt, xfr_length;
2159
2160	if (fod->fcpreq->op == NVMET_FCOP_READDATA_RSP)
2161		xfr_length = fod->req.transfer_len;
2162	else
2163		xfr_length = fod->offset;
2164
2165	/*
2166	 * check to see if we can send a 0's rsp.
2167	 *   Note: to send a 0's response, the NVME-FC host transport will
2168	 *   recreate the CQE. The host transport knows: sq id, SQHD (last
2169	 *   seen in an ersp), and command_id. Thus it will create a
2170	 *   zero-filled CQE with those known fields filled in. Transport
2171	 *   must send an ersp for any condition where the cqe won't match
2172	 *   this.
2173	 *
2174	 * Here are the FC-NVME mandated cases where we must send an ersp:
2175	 *  every N responses, where N=ersp_ratio
2176	 *  force fabric commands to send ersp's (not in FC-NVME but good
2177	 *    practice)
2178	 *  normal cmds: any time status is non-zero, or status is zero
2179	 *     but words 0 or 1 are non-zero.
2180	 *  the SQ is 90% or more full
2181	 *  the cmd is a fused command
2182	 *  transferred data length not equal to cmd iu length
2183	 */
2184	rspcnt = atomic_inc_return(&fod->queue->zrspcnt);
2185	if (!(rspcnt % fod->queue->ersp_ratio) ||
2186	    nvme_is_fabrics((struct nvme_command *) sqe) ||
2187	    xfr_length != fod->req.transfer_len ||
2188	    (le16_to_cpu(cqe->status) & 0xFFFE) || cqewd[0] || cqewd[1] ||
2189	    (sqe->flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND)) ||
2190	    queue_90percent_full(fod->queue, le16_to_cpu(cqe->sq_head)))
2191		send_ersp = true;
2192
2193	/* re-set the fields */
2194	fod->fcpreq->rspaddr = ersp;
2195	fod->fcpreq->rspdma = fod->rspdma;
2196
2197	if (!send_ersp) {
2198		memset(ersp, 0, NVME_FC_SIZEOF_ZEROS_RSP);
2199		fod->fcpreq->rsplen = NVME_FC_SIZEOF_ZEROS_RSP;
2200	} else {
2201		ersp->iu_len = cpu_to_be16(sizeof(*ersp)/sizeof(u32));
2202		rsn = atomic_inc_return(&fod->queue->rsn);
2203		ersp->rsn = cpu_to_be32(rsn);
2204		ersp->xfrd_len = cpu_to_be32(xfr_length);
2205		fod->fcpreq->rsplen = sizeof(*ersp);
2206	}
2207
2208	fc_dma_sync_single_for_device(tgtport->dev, fod->rspdma,
2209				  sizeof(fod->rspiubuf), DMA_TO_DEVICE);
2210}
2211
2212static void nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq);
2213
2214static void
2215nvmet_fc_abort_op(struct nvmet_fc_tgtport *tgtport,
2216				struct nvmet_fc_fcp_iod *fod)
2217{
2218	struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2219
2220	/* data no longer needed */
2221	nvmet_fc_free_tgt_pgs(fod);
2222
2223	/*
2224	 * if an ABTS was received or we issued the fcp_abort early
2225	 * don't call abort routine again.
2226	 */
2227	/* no need to take lock - lock was taken earlier to get here */
2228	if (!fod->aborted)
2229		tgtport->ops->fcp_abort(&tgtport->fc_target_port, fcpreq);
2230
2231	nvmet_fc_free_fcp_iod(fod->queue, fod);
2232}
2233
2234static void
2235nvmet_fc_xmt_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
2236				struct nvmet_fc_fcp_iod *fod)
2237{
2238	int ret;
2239
2240	fod->fcpreq->op = NVMET_FCOP_RSP;
2241	fod->fcpreq->timeout = 0;
2242
2243	nvmet_fc_prep_fcp_rsp(tgtport, fod);
2244
2245	ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
2246	if (ret)
2247		nvmet_fc_abort_op(tgtport, fod);
2248}
2249
2250static void
2251nvmet_fc_transfer_fcp_data(struct nvmet_fc_tgtport *tgtport,
2252				struct nvmet_fc_fcp_iod *fod, u8 op)
2253{
2254	struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2255	struct scatterlist *sg = fod->next_sg;
2256	unsigned long flags;
2257	u32 remaininglen = fod->req.transfer_len - fod->offset;
2258	u32 tlen = 0;
2259	int ret;
2260
2261	fcpreq->op = op;
2262	fcpreq->offset = fod->offset;
2263	fcpreq->timeout = NVME_FC_TGTOP_TIMEOUT_SEC;
2264
2265	/*
2266	 * for next sequence:
2267	 *  break at a sg element boundary
2268	 *  attempt to keep sequence length capped at
2269	 *    NVMET_FC_MAX_SEQ_LENGTH but allow sequence to
2270	 *    be longer if a single sg element is larger
2271	 *    than that amount. This is done to avoid creating
2272	 *    a new sg list to use for the tgtport api.
2273	 */
2274	fcpreq->sg = sg;
2275	fcpreq->sg_cnt = 0;
2276	while (tlen < remaininglen &&
2277	       fcpreq->sg_cnt < tgtport->max_sg_cnt &&
2278	       tlen + sg_dma_len(sg) < NVMET_FC_MAX_SEQ_LENGTH) {
2279		fcpreq->sg_cnt++;
2280		tlen += sg_dma_len(sg);
2281		sg = sg_next(sg);
2282	}
2283	if (tlen < remaininglen && fcpreq->sg_cnt == 0) {
2284		fcpreq->sg_cnt++;
2285		tlen += min_t(u32, sg_dma_len(sg), remaininglen);
2286		sg = sg_next(sg);
2287	}
2288	if (tlen < remaininglen)
2289		fod->next_sg = sg;
2290	else
2291		fod->next_sg = NULL;
2292
2293	fcpreq->transfer_length = tlen;
2294	fcpreq->transferred_length = 0;
2295	fcpreq->fcp_error = 0;
2296	fcpreq->rsplen = 0;
2297
2298	/*
2299	 * If the last READDATA request: check if LLDD supports
2300	 * combined xfr with response.
2301	 */
2302	if ((op == NVMET_FCOP_READDATA) &&
2303	    ((fod->offset + fcpreq->transfer_length) == fod->req.transfer_len) &&
2304	    (tgtport->ops->target_features & NVMET_FCTGTFEAT_READDATA_RSP)) {
2305		fcpreq->op = NVMET_FCOP_READDATA_RSP;
2306		nvmet_fc_prep_fcp_rsp(tgtport, fod);
2307	}
2308
2309	ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
2310	if (ret) {
2311		/*
2312		 * should be ok to set w/o lock as its in the thread of
2313		 * execution (not an async timer routine) and doesn't
2314		 * contend with any clearing action
2315		 */
2316		fod->abort = true;
2317
2318		if (op == NVMET_FCOP_WRITEDATA) {
2319			spin_lock_irqsave(&fod->flock, flags);
2320			fod->writedataactive = false;
2321			spin_unlock_irqrestore(&fod->flock, flags);
2322			nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2323		} else /* NVMET_FCOP_READDATA or NVMET_FCOP_READDATA_RSP */ {
2324			fcpreq->fcp_error = ret;
2325			fcpreq->transferred_length = 0;
2326			nvmet_fc_xmt_fcp_op_done(fod->fcpreq);
2327		}
2328	}
2329}
2330
2331static inline bool
2332__nvmet_fc_fod_op_abort(struct nvmet_fc_fcp_iod *fod, bool abort)
2333{
2334	struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2335	struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2336
2337	/* if in the middle of an io and we need to tear down */
2338	if (abort) {
2339		if (fcpreq->op == NVMET_FCOP_WRITEDATA) {
2340			nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2341			return true;
2342		}
2343
2344		nvmet_fc_abort_op(tgtport, fod);
2345		return true;
2346	}
2347
2348	return false;
2349}
2350
2351/*
2352 * actual done handler for FCP operations when completed by the lldd
2353 */
2354static void
2355nvmet_fc_fod_op_done(struct nvmet_fc_fcp_iod *fod)
2356{
2357	struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2358	struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2359	unsigned long flags;
2360	bool abort;
2361
2362	spin_lock_irqsave(&fod->flock, flags);
2363	abort = fod->abort;
2364	fod->writedataactive = false;
2365	spin_unlock_irqrestore(&fod->flock, flags);
2366
2367	switch (fcpreq->op) {
2368
2369	case NVMET_FCOP_WRITEDATA:
2370		if (__nvmet_fc_fod_op_abort(fod, abort))
2371			return;
2372		if (fcpreq->fcp_error ||
2373		    fcpreq->transferred_length != fcpreq->transfer_length) {
2374			spin_lock_irqsave(&fod->flock, flags);
2375			fod->abort = true;
2376			spin_unlock_irqrestore(&fod->flock, flags);
2377
2378			nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2379			return;
2380		}
2381
2382		fod->offset += fcpreq->transferred_length;
2383		if (fod->offset != fod->req.transfer_len) {
2384			spin_lock_irqsave(&fod->flock, flags);
2385			fod->writedataactive = true;
2386			spin_unlock_irqrestore(&fod->flock, flags);
2387
2388			/* transfer the next chunk */
2389			nvmet_fc_transfer_fcp_data(tgtport, fod,
2390						NVMET_FCOP_WRITEDATA);
2391			return;
2392		}
2393
2394		/* data transfer complete, resume with nvmet layer */
2395		fod->req.execute(&fod->req);
2396		break;
2397
2398	case NVMET_FCOP_READDATA:
2399	case NVMET_FCOP_READDATA_RSP:
2400		if (__nvmet_fc_fod_op_abort(fod, abort))
2401			return;
2402		if (fcpreq->fcp_error ||
2403		    fcpreq->transferred_length != fcpreq->transfer_length) {
2404			nvmet_fc_abort_op(tgtport, fod);
2405			return;
2406		}
2407
2408		/* success */
2409
2410		if (fcpreq->op == NVMET_FCOP_READDATA_RSP) {
2411			/* data no longer needed */
2412			nvmet_fc_free_tgt_pgs(fod);
2413			nvmet_fc_free_fcp_iod(fod->queue, fod);
2414			return;
2415		}
2416
2417		fod->offset += fcpreq->transferred_length;
2418		if (fod->offset != fod->req.transfer_len) {
2419			/* transfer the next chunk */
2420			nvmet_fc_transfer_fcp_data(tgtport, fod,
2421						NVMET_FCOP_READDATA);
2422			return;
2423		}
2424
2425		/* data transfer complete, send response */
2426
2427		/* data no longer needed */
2428		nvmet_fc_free_tgt_pgs(fod);
2429
2430		nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2431
2432		break;
2433
2434	case NVMET_FCOP_RSP:
2435		if (__nvmet_fc_fod_op_abort(fod, abort))
2436			return;
2437		nvmet_fc_free_fcp_iod(fod->queue, fod);
2438		break;
2439
2440	default:
2441		break;
2442	}
2443}
2444
2445static void
2446nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq)
2447{
2448	struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2449
2450	nvmet_fc_fod_op_done(fod);
2451}
2452
2453/*
2454 * actual completion handler after execution by the nvmet layer
2455 */
2456static void
2457__nvmet_fc_fcp_nvme_cmd_done(struct nvmet_fc_tgtport *tgtport,
2458			struct nvmet_fc_fcp_iod *fod, int status)
2459{
2460	struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
2461	struct nvme_completion *cqe = &fod->rspiubuf.cqe;
2462	unsigned long flags;
2463	bool abort;
2464
2465	spin_lock_irqsave(&fod->flock, flags);
2466	abort = fod->abort;
2467	spin_unlock_irqrestore(&fod->flock, flags);
2468
2469	/* if we have a CQE, snoop the last sq_head value */
2470	if (!status)
2471		fod->queue->sqhd = cqe->sq_head;
2472
2473	if (abort) {
2474		nvmet_fc_abort_op(tgtport, fod);
2475		return;
2476	}
2477
2478	/* if an error handling the cmd post initial parsing */
2479	if (status) {
2480		/* fudge up a failed CQE status for our transport error */
2481		memset(cqe, 0, sizeof(*cqe));
2482		cqe->sq_head = fod->queue->sqhd;	/* echo last cqe sqhd */
2483		cqe->sq_id = cpu_to_le16(fod->queue->qid);
2484		cqe->command_id = sqe->command_id;
2485		cqe->status = cpu_to_le16(status);
2486	} else {
2487
2488		/*
2489		 * try to push the data even if the SQE status is non-zero.
2490		 * There may be a status where data still was intended to
2491		 * be moved
2492		 */
2493		if ((fod->io_dir == NVMET_FCP_READ) && (fod->data_sg_cnt)) {
2494			/* push the data over before sending rsp */
2495			nvmet_fc_transfer_fcp_data(tgtport, fod,
2496						NVMET_FCOP_READDATA);
2497			return;
2498		}
2499
2500		/* writes & no data - fall thru */
2501	}
2502
2503	/* data no longer needed */
2504	nvmet_fc_free_tgt_pgs(fod);
2505
2506	nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2507}
2508
2509
2510static void
2511nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req)
2512{
2513	struct nvmet_fc_fcp_iod *fod = nvmet_req_to_fod(nvme_req);
2514	struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2515
2516	__nvmet_fc_fcp_nvme_cmd_done(tgtport, fod, 0);
2517}
2518
2519
2520/*
2521 * Actual processing routine for received FC-NVME I/O Requests from the LLD
2522 */
2523static void
2524nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
2525			struct nvmet_fc_fcp_iod *fod)
2526{
2527	struct nvme_fc_cmd_iu *cmdiu = &fod->cmdiubuf;
2528	u32 xfrlen = be32_to_cpu(cmdiu->data_len);
2529	int ret;
2530
2531	/*
2532	 * Fused commands are currently not supported in the linux
2533	 * implementation.
2534	 *
2535	 * As such, the implementation of the FC transport does not
2536	 * look at the fused commands and order delivery to the upper
2537	 * layer until we have both based on csn.
2538	 */
2539
2540	fod->fcpreq->done = nvmet_fc_xmt_fcp_op_done;
2541
2542	if (cmdiu->flags & FCNVME_CMD_FLAGS_WRITE) {
2543		fod->io_dir = NVMET_FCP_WRITE;
2544		if (!nvme_is_write(&cmdiu->sqe))
2545			goto transport_error;
2546	} else if (cmdiu->flags & FCNVME_CMD_FLAGS_READ) {
2547		fod->io_dir = NVMET_FCP_READ;
2548		if (nvme_is_write(&cmdiu->sqe))
2549			goto transport_error;
2550	} else {
2551		fod->io_dir = NVMET_FCP_NODATA;
2552		if (xfrlen)
2553			goto transport_error;
2554	}
2555
2556	fod->req.cmd = &fod->cmdiubuf.sqe;
2557	fod->req.cqe = &fod->rspiubuf.cqe;
2558	if (!tgtport->pe)
2559		goto transport_error;
2560	fod->req.port = tgtport->pe->port;
2561
2562	/* clear any response payload */
2563	memset(&fod->rspiubuf, 0, sizeof(fod->rspiubuf));
2564
2565	fod->data_sg = NULL;
2566	fod->data_sg_cnt = 0;
2567
2568	ret = nvmet_req_init(&fod->req,
2569				&fod->queue->nvme_cq,
2570				&fod->queue->nvme_sq,
2571				&nvmet_fc_tgt_fcp_ops);
2572	if (!ret) {
2573		/* bad SQE content or invalid ctrl state */
2574		/* nvmet layer has already called op done to send rsp. */
2575		return;
2576	}
2577
2578	fod->req.transfer_len = xfrlen;
2579
2580	/* keep a running counter of tail position */
2581	atomic_inc(&fod->queue->sqtail);
2582
2583	if (fod->req.transfer_len) {
2584		ret = nvmet_fc_alloc_tgt_pgs(fod);
2585		if (ret) {
2586			nvmet_req_complete(&fod->req, ret);
2587			return;
2588		}
2589	}
2590	fod->req.sg = fod->data_sg;
2591	fod->req.sg_cnt = fod->data_sg_cnt;
2592	fod->offset = 0;
2593
2594	if (fod->io_dir == NVMET_FCP_WRITE) {
2595		/* pull the data over before invoking nvmet layer */
2596		nvmet_fc_transfer_fcp_data(tgtport, fod, NVMET_FCOP_WRITEDATA);
2597		return;
2598	}
2599
2600	/*
2601	 * Reads or no data:
2602	 *
2603	 * can invoke the nvmet_layer now. If read data, cmd completion will
2604	 * push the data
2605	 */
2606	fod->req.execute(&fod->req);
2607	return;
2608
2609transport_error:
2610	nvmet_fc_abort_op(tgtport, fod);
2611}
2612
2613/**
2614 * nvmet_fc_rcv_fcp_req - transport entry point called by an LLDD
2615 *                       upon the reception of a NVME FCP CMD IU.
2616 *
2617 * Pass a FC-NVME FCP CMD IU received from the FC link to the nvmet-fc
2618 * layer for processing.
2619 *
2620 * The nvmet_fc layer allocates a local job structure (struct
2621 * nvmet_fc_fcp_iod) from the queue for the io and copies the
2622 * CMD IU buffer to the job structure. As such, on a successful
2623 * completion (returns 0), the LLDD may immediately free/reuse
2624 * the CMD IU buffer passed in the call.
2625 *
2626 * However, in some circumstances, due to the packetized nature of FC
2627 * and the api of the FC LLDD which may issue a hw command to send the
2628 * response, but the LLDD may not get the hw completion for that command
2629 * and upcall the nvmet_fc layer before a new command may be
2630 * asynchronously received - its possible for a command to be received
2631 * before the LLDD and nvmet_fc have recycled the job structure. It gives
2632 * the appearance of more commands received than fits in the sq.
2633 * To alleviate this scenario, a temporary queue is maintained in the
2634 * transport for pending LLDD requests waiting for a queue job structure.
2635 * In these "overrun" cases, a temporary queue element is allocated
2636 * the LLDD request and CMD iu buffer information remembered, and the
2637 * routine returns a -EOVERFLOW status. Subsequently, when a queue job
2638 * structure is freed, it is immediately reallocated for anything on the
2639 * pending request list. The LLDDs defer_rcv() callback is called,
2640 * informing the LLDD that it may reuse the CMD IU buffer, and the io
2641 * is then started normally with the transport.
2642 *
2643 * The LLDD, when receiving an -EOVERFLOW completion status, is to treat
2644 * the completion as successful but must not reuse the CMD IU buffer
2645 * until the LLDD's defer_rcv() callback has been called for the
2646 * corresponding struct nvmefc_tgt_fcp_req pointer.
2647 *
2648 * If there is any other condition in which an error occurs, the
2649 * transport will return a non-zero status indicating the error.
2650 * In all cases other than -EOVERFLOW, the transport has not accepted the
2651 * request and the LLDD should abort the exchange.
2652 *
2653 * @target_port: pointer to the (registered) target port the FCP CMD IU
2654 *              was received on.
2655 * @fcpreq:     pointer to a fcpreq request structure to be used to reference
2656 *              the exchange corresponding to the FCP Exchange.
2657 * @cmdiubuf:   pointer to the buffer containing the FCP CMD IU
2658 * @cmdiubuf_len: length, in bytes, of the received FCP CMD IU
2659 */
2660int
2661nvmet_fc_rcv_fcp_req(struct nvmet_fc_target_port *target_port,
2662			struct nvmefc_tgt_fcp_req *fcpreq,
2663			void *cmdiubuf, u32 cmdiubuf_len)
2664{
2665	struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
2666	struct nvme_fc_cmd_iu *cmdiu = cmdiubuf;
2667	struct nvmet_fc_tgt_queue *queue;
2668	struct nvmet_fc_fcp_iod *fod;
2669	struct nvmet_fc_defer_fcp_req *deferfcp;
2670	unsigned long flags;
2671
2672	/* validate iu, so the connection id can be used to find the queue */
2673	if ((cmdiubuf_len != sizeof(*cmdiu)) ||
2674			(cmdiu->format_id != NVME_CMD_FORMAT_ID) ||
2675			(cmdiu->fc_id != NVME_CMD_FC_ID) ||
2676			(be16_to_cpu(cmdiu->iu_len) != (sizeof(*cmdiu)/4)))
2677		return -EIO;
2678
2679	queue = nvmet_fc_find_target_queue(tgtport,
2680				be64_to_cpu(cmdiu->connection_id));
2681	if (!queue)
2682		return -ENOTCONN;
2683
2684	/*
2685	 * note: reference taken by find_target_queue
2686	 * After successful fod allocation, the fod will inherit the
2687	 * ownership of that reference and will remove the reference
2688	 * when the fod is freed.
2689	 */
2690
2691	spin_lock_irqsave(&queue->qlock, flags);
2692
2693	fod = nvmet_fc_alloc_fcp_iod(queue);
2694	if (fod) {
2695		spin_unlock_irqrestore(&queue->qlock, flags);
2696
2697		fcpreq->nvmet_fc_private = fod;
2698		fod->fcpreq = fcpreq;
2699
2700		memcpy(&fod->cmdiubuf, cmdiubuf, cmdiubuf_len);
2701
2702		nvmet_fc_queue_fcp_req(tgtport, queue, fcpreq);
2703
2704		return 0;
2705	}
2706
2707	if (!tgtport->ops->defer_rcv) {
2708		spin_unlock_irqrestore(&queue->qlock, flags);
2709		/* release the queue lookup reference */
2710		nvmet_fc_tgt_q_put(queue);
2711		return -ENOENT;
2712	}
2713
2714	deferfcp = list_first_entry_or_null(&queue->avail_defer_list,
2715			struct nvmet_fc_defer_fcp_req, req_list);
2716	if (deferfcp) {
2717		/* Just re-use one that was previously allocated */
2718		list_del(&deferfcp->req_list);
2719	} else {
2720		spin_unlock_irqrestore(&queue->qlock, flags);
2721
2722		/* Now we need to dynamically allocate one */
2723		deferfcp = kmalloc(sizeof(*deferfcp), GFP_KERNEL);
2724		if (!deferfcp) {
2725			/* release the queue lookup reference */
2726			nvmet_fc_tgt_q_put(queue);
2727			return -ENOMEM;
2728		}
2729		spin_lock_irqsave(&queue->qlock, flags);
2730	}
2731
2732	/* For now, use rspaddr / rsplen to save payload information */
2733	fcpreq->rspaddr = cmdiubuf;
2734	fcpreq->rsplen  = cmdiubuf_len;
2735	deferfcp->fcp_req = fcpreq;
2736
2737	/* defer processing till a fod becomes available */
2738	list_add_tail(&deferfcp->req_list, &queue->pending_cmd_list);
2739
2740	/* NOTE: the queue lookup reference is still valid */
2741
2742	spin_unlock_irqrestore(&queue->qlock, flags);
2743
2744	return -EOVERFLOW;
2745}
2746EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_req);
2747
2748/**
2749 * nvmet_fc_rcv_fcp_abort - transport entry point called by an LLDD
2750 *                       upon the reception of an ABTS for a FCP command
2751 *
2752 * Notify the transport that an ABTS has been received for a FCP command
2753 * that had been given to the transport via nvmet_fc_rcv_fcp_req(). The
2754 * LLDD believes the command is still being worked on
2755 * (template_ops->fcp_req_release() has not been called).
2756 *
2757 * The transport will wait for any outstanding work (an op to the LLDD,
2758 * which the lldd should complete with error due to the ABTS; or the
2759 * completion from the nvmet layer of the nvme command), then will
2760 * stop processing and call the nvmet_fc_rcv_fcp_req() callback to
2761 * return the i/o context to the LLDD.  The LLDD may send the BA_ACC
2762 * to the ABTS either after return from this function (assuming any
2763 * outstanding op work has been terminated) or upon the callback being
2764 * called.
2765 *
2766 * @target_port: pointer to the (registered) target port the FCP CMD IU
2767 *              was received on.
2768 * @fcpreq:     pointer to the fcpreq request structure that corresponds
2769 *              to the exchange that received the ABTS.
2770 */
2771void
2772nvmet_fc_rcv_fcp_abort(struct nvmet_fc_target_port *target_port,
2773			struct nvmefc_tgt_fcp_req *fcpreq)
2774{
2775	struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2776	struct nvmet_fc_tgt_queue *queue;
2777	unsigned long flags;
2778
2779	if (!fod || fod->fcpreq != fcpreq)
2780		/* job appears to have already completed, ignore abort */
2781		return;
2782
2783	queue = fod->queue;
2784
2785	spin_lock_irqsave(&queue->qlock, flags);
2786	if (fod->active) {
2787		/*
2788		 * mark as abort. The abort handler, invoked upon completion
2789		 * of any work, will detect the aborted status and do the
2790		 * callback.
2791		 */
2792		spin_lock(&fod->flock);
2793		fod->abort = true;
2794		fod->aborted = true;
2795		spin_unlock(&fod->flock);
2796	}
2797	spin_unlock_irqrestore(&queue->qlock, flags);
2798}
2799EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_abort);
2800
2801
2802struct nvmet_fc_traddr {
2803	u64	nn;
2804	u64	pn;
2805};
2806
2807static int
2808__nvme_fc_parse_u64(substring_t *sstr, u64 *val)
2809{
2810	u64 token64;
2811
2812	if (match_u64(sstr, &token64))
2813		return -EINVAL;
2814	*val = token64;
2815
2816	return 0;
2817}
2818
2819/*
2820 * This routine validates and extracts the WWN's from the TRADDR string.
2821 * As kernel parsers need the 0x to determine number base, universally
2822 * build string to parse with 0x prefix before parsing name strings.
2823 */
2824static int
2825nvme_fc_parse_traddr(struct nvmet_fc_traddr *traddr, char *buf, size_t blen)
2826{
2827	char name[2 + NVME_FC_TRADDR_HEXNAMELEN + 1];
2828	substring_t wwn = { name, &name[sizeof(name)-1] };
2829	int nnoffset, pnoffset;
2830
2831	/* validate if string is one of the 2 allowed formats */
2832	if (strnlen(buf, blen) == NVME_FC_TRADDR_MAXLENGTH &&
2833			!strncmp(buf, "nn-0x", NVME_FC_TRADDR_OXNNLEN) &&
2834			!strncmp(&buf[NVME_FC_TRADDR_MAX_PN_OFFSET],
2835				"pn-0x", NVME_FC_TRADDR_OXNNLEN)) {
2836		nnoffset = NVME_FC_TRADDR_OXNNLEN;
2837		pnoffset = NVME_FC_TRADDR_MAX_PN_OFFSET +
2838						NVME_FC_TRADDR_OXNNLEN;
2839	} else if ((strnlen(buf, blen) == NVME_FC_TRADDR_MINLENGTH &&
2840			!strncmp(buf, "nn-", NVME_FC_TRADDR_NNLEN) &&
2841			!strncmp(&buf[NVME_FC_TRADDR_MIN_PN_OFFSET],
2842				"pn-", NVME_FC_TRADDR_NNLEN))) {
2843		nnoffset = NVME_FC_TRADDR_NNLEN;
2844		pnoffset = NVME_FC_TRADDR_MIN_PN_OFFSET + NVME_FC_TRADDR_NNLEN;
2845	} else
2846		goto out_einval;
2847
2848	name[0] = '0';
2849	name[1] = 'x';
2850	name[2 + NVME_FC_TRADDR_HEXNAMELEN] = 0;
2851
2852	memcpy(&name[2], &buf[nnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2853	if (__nvme_fc_parse_u64(&wwn, &traddr->nn))
2854		goto out_einval;
2855
2856	memcpy(&name[2], &buf[pnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2857	if (__nvme_fc_parse_u64(&wwn, &traddr->pn))
2858		goto out_einval;
2859
2860	return 0;
2861
2862out_einval:
2863	pr_warn("%s: bad traddr string\n", __func__);
2864	return -EINVAL;
2865}
2866
2867static int
2868nvmet_fc_add_port(struct nvmet_port *port)
2869{
2870	struct nvmet_fc_tgtport *tgtport;
2871	struct nvmet_fc_port_entry *pe;
2872	struct nvmet_fc_traddr traddr = { 0L, 0L };
2873	unsigned long flags;
2874	int ret;
2875
2876	/* validate the address info */
2877	if ((port->disc_addr.trtype != NVMF_TRTYPE_FC) ||
2878	    (port->disc_addr.adrfam != NVMF_ADDR_FAMILY_FC))
2879		return -EINVAL;
2880
2881	/* map the traddr address info to a target port */
2882
2883	ret = nvme_fc_parse_traddr(&traddr, port->disc_addr.traddr,
2884			sizeof(port->disc_addr.traddr));
2885	if (ret)
2886		return ret;
2887
2888	pe = kzalloc(sizeof(*pe), GFP_KERNEL);
2889	if (!pe)
2890		return -ENOMEM;
2891
2892	ret = -ENXIO;
2893	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
2894	list_for_each_entry(tgtport, &nvmet_fc_target_list, tgt_list) {
2895		if ((tgtport->fc_target_port.node_name == traddr.nn) &&
2896		    (tgtport->fc_target_port.port_name == traddr.pn)) {
2897			/* a FC port can only be 1 nvmet port id */
2898			if (!tgtport->pe) {
2899				nvmet_fc_portentry_bind(tgtport, pe, port);
2900				ret = 0;
2901			} else
2902				ret = -EALREADY;
2903			break;
2904		}
2905	}
2906	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
2907
2908	if (ret)
2909		kfree(pe);
2910
2911	return ret;
2912}
2913
2914static void
2915nvmet_fc_remove_port(struct nvmet_port *port)
2916{
2917	struct nvmet_fc_port_entry *pe = port->priv;
2918
2919	nvmet_fc_portentry_unbind(pe);
2920
2921	/* terminate any outstanding associations */
2922	__nvmet_fc_free_assocs(pe->tgtport);
2923
2924	kfree(pe);
2925}
2926
2927static void
2928nvmet_fc_discovery_chg(struct nvmet_port *port)
2929{
2930	struct nvmet_fc_port_entry *pe = port->priv;
2931	struct nvmet_fc_tgtport *tgtport = pe->tgtport;
2932
2933	if (tgtport && tgtport->ops->discovery_event)
2934		tgtport->ops->discovery_event(&tgtport->fc_target_port);
2935}
2936
2937static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops = {
2938	.owner			= THIS_MODULE,
2939	.type			= NVMF_TRTYPE_FC,
2940	.msdbd			= 1,
2941	.add_port		= nvmet_fc_add_port,
2942	.remove_port		= nvmet_fc_remove_port,
2943	.queue_response		= nvmet_fc_fcp_nvme_cmd_done,
2944	.delete_ctrl		= nvmet_fc_delete_ctrl,
2945	.discovery_chg		= nvmet_fc_discovery_chg,
2946};
2947
2948static int __init nvmet_fc_init_module(void)
2949{
2950	return nvmet_register_transport(&nvmet_fc_tgt_fcp_ops);
2951}
2952
2953static void __exit nvmet_fc_exit_module(void)
2954{
2955	/* ensure any shutdown operation, e.g. delete ctrls have finished */
2956	flush_workqueue(nvmet_wq);
2957
2958	/* sanity check - all lports should be removed */
2959	if (!list_empty(&nvmet_fc_target_list))
2960		pr_warn("%s: targetport list not empty\n", __func__);
2961
2962	nvmet_unregister_transport(&nvmet_fc_tgt_fcp_ops);
2963
2964	ida_destroy(&nvmet_fc_tgtport_cnt);
2965}
2966
2967module_init(nvmet_fc_init_module);
2968module_exit(nvmet_fc_exit_module);
2969
2970MODULE_DESCRIPTION("NVMe target FC transport driver");
2971MODULE_LICENSE("GPL v2");
2972