xref: /kernel/linux/linux-5.10/drivers/nvme/target/fc.c (revision 8c2ecf20)
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
116struct nvmet_fc_port_entry {
117	struct nvmet_fc_tgtport		*tgtport;
118	struct nvmet_port		*port;
119	u64				node_name;
120	u64				port_name;
121	struct list_head		pe_list;
122};
123
124struct nvmet_fc_defer_fcp_req {
125	struct list_head		req_list;
126	struct nvmefc_tgt_fcp_req	*fcp_req;
127};
128
129struct nvmet_fc_tgt_queue {
130	bool				ninetypercent;
131	u16				qid;
132	u16				sqsize;
133	u16				ersp_ratio;
134	__le16				sqhd;
135	atomic_t			connected;
136	atomic_t			sqtail;
137	atomic_t			zrspcnt;
138	atomic_t			rsn;
139	spinlock_t			qlock;
140	struct nvmet_cq			nvme_cq;
141	struct nvmet_sq			nvme_sq;
142	struct nvmet_fc_tgt_assoc	*assoc;
143	struct list_head		fod_list;
144	struct list_head		pending_cmd_list;
145	struct list_head		avail_defer_list;
146	struct workqueue_struct		*work_q;
147	struct kref			ref;
148	struct nvmet_fc_fcp_iod		fod[];		/* array of fcp_iods */
149} __aligned(sizeof(unsigned long long));
150
151struct nvmet_fc_hostport {
152	struct nvmet_fc_tgtport		*tgtport;
153	void				*hosthandle;
154	struct list_head		host_list;
155	struct kref			ref;
156	u8				invalid;
157};
158
159struct nvmet_fc_tgt_assoc {
160	u64				association_id;
161	u32				a_id;
162	atomic_t			terminating;
163	struct nvmet_fc_tgtport		*tgtport;
164	struct nvmet_fc_hostport	*hostport;
165	struct nvmet_fc_ls_iod		*rcv_disconn;
166	struct list_head		a_list;
167	struct nvmet_fc_tgt_queue	*queues[NVMET_NR_QUEUES + 1];
168	struct kref			ref;
169	struct work_struct		del_work;
170};
171
172
173static inline int
174nvmet_fc_iodnum(struct nvmet_fc_ls_iod *iodptr)
175{
176	return (iodptr - iodptr->tgtport->iod);
177}
178
179static inline int
180nvmet_fc_fodnum(struct nvmet_fc_fcp_iod *fodptr)
181{
182	return (fodptr - fodptr->queue->fod);
183}
184
185
186/*
187 * Association and Connection IDs:
188 *
189 * Association ID will have random number in upper 6 bytes and zero
190 *   in lower 2 bytes
191 *
192 * Connection IDs will be Association ID with QID or'd in lower 2 bytes
193 *
194 * note: Association ID = Connection ID for queue 0
195 */
196#define BYTES_FOR_QID			sizeof(u16)
197#define BYTES_FOR_QID_SHIFT		(BYTES_FOR_QID * 8)
198#define NVMET_FC_QUEUEID_MASK		((u64)((1 << BYTES_FOR_QID_SHIFT) - 1))
199
200static inline u64
201nvmet_fc_makeconnid(struct nvmet_fc_tgt_assoc *assoc, u16 qid)
202{
203	return (assoc->association_id | qid);
204}
205
206static inline u64
207nvmet_fc_getassociationid(u64 connectionid)
208{
209	return connectionid & ~NVMET_FC_QUEUEID_MASK;
210}
211
212static inline u16
213nvmet_fc_getqueueid(u64 connectionid)
214{
215	return (u16)(connectionid & NVMET_FC_QUEUEID_MASK);
216}
217
218static inline struct nvmet_fc_tgtport *
219targetport_to_tgtport(struct nvmet_fc_target_port *targetport)
220{
221	return container_of(targetport, struct nvmet_fc_tgtport,
222				 fc_target_port);
223}
224
225static inline struct nvmet_fc_fcp_iod *
226nvmet_req_to_fod(struct nvmet_req *nvme_req)
227{
228	return container_of(nvme_req, struct nvmet_fc_fcp_iod, req);
229}
230
231
232/* *************************** Globals **************************** */
233
234
235static DEFINE_SPINLOCK(nvmet_fc_tgtlock);
236
237static LIST_HEAD(nvmet_fc_target_list);
238static DEFINE_IDA(nvmet_fc_tgtport_cnt);
239static LIST_HEAD(nvmet_fc_portentry_list);
240
241
242static void nvmet_fc_handle_ls_rqst_work(struct work_struct *work);
243static void nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work);
244static void nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc);
245static int nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc);
246static void nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue);
247static int nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue);
248static void nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport);
249static int nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport);
250static void nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
251					struct nvmet_fc_fcp_iod *fod);
252static void nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc);
253static void nvmet_fc_xmt_ls_rsp(struct nvmet_fc_tgtport *tgtport,
254				struct nvmet_fc_ls_iod *iod);
255
256
257/* *********************** FC-NVME DMA Handling **************************** */
258
259/*
260 * The fcloop device passes in a NULL device pointer. Real LLD's will
261 * pass in a valid device pointer. If NULL is passed to the dma mapping
262 * routines, depending on the platform, it may or may not succeed, and
263 * may crash.
264 *
265 * As such:
266 * Wrapper all the dma routines and check the dev pointer.
267 *
268 * If simple mappings (return just a dma address, we'll noop them,
269 * returning a dma address of 0.
270 *
271 * On more complex mappings (dma_map_sg), a pseudo routine fills
272 * in the scatter list, setting all dma addresses to 0.
273 */
274
275static inline dma_addr_t
276fc_dma_map_single(struct device *dev, void *ptr, size_t size,
277		enum dma_data_direction dir)
278{
279	return dev ? dma_map_single(dev, ptr, size, dir) : (dma_addr_t)0L;
280}
281
282static inline int
283fc_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
284{
285	return dev ? dma_mapping_error(dev, dma_addr) : 0;
286}
287
288static inline void
289fc_dma_unmap_single(struct device *dev, dma_addr_t addr, size_t size,
290	enum dma_data_direction dir)
291{
292	if (dev)
293		dma_unmap_single(dev, addr, size, dir);
294}
295
296static inline void
297fc_dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
298		enum dma_data_direction dir)
299{
300	if (dev)
301		dma_sync_single_for_cpu(dev, addr, size, dir);
302}
303
304static inline void
305fc_dma_sync_single_for_device(struct device *dev, dma_addr_t addr, size_t size,
306		enum dma_data_direction dir)
307{
308	if (dev)
309		dma_sync_single_for_device(dev, addr, size, dir);
310}
311
312/* pseudo dma_map_sg call */
313static int
314fc_map_sg(struct scatterlist *sg, int nents)
315{
316	struct scatterlist *s;
317	int i;
318
319	WARN_ON(nents == 0 || sg[0].length == 0);
320
321	for_each_sg(sg, s, nents, i) {
322		s->dma_address = 0L;
323#ifdef CONFIG_NEED_SG_DMA_LENGTH
324		s->dma_length = s->length;
325#endif
326	}
327	return nents;
328}
329
330static inline int
331fc_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
332		enum dma_data_direction dir)
333{
334	return dev ? dma_map_sg(dev, sg, nents, dir) : fc_map_sg(sg, nents);
335}
336
337static inline void
338fc_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
339		enum dma_data_direction dir)
340{
341	if (dev)
342		dma_unmap_sg(dev, sg, nents, dir);
343}
344
345
346/* ********************** FC-NVME LS XMT Handling ************************* */
347
348
349static void
350__nvmet_fc_finish_ls_req(struct nvmet_fc_ls_req_op *lsop)
351{
352	struct nvmet_fc_tgtport *tgtport = lsop->tgtport;
353	struct nvmefc_ls_req *lsreq = &lsop->ls_req;
354	unsigned long flags;
355
356	spin_lock_irqsave(&tgtport->lock, flags);
357
358	if (!lsop->req_queued) {
359		spin_unlock_irqrestore(&tgtport->lock, flags);
360		return;
361	}
362
363	list_del(&lsop->lsreq_list);
364
365	lsop->req_queued = false;
366
367	spin_unlock_irqrestore(&tgtport->lock, flags);
368
369	fc_dma_unmap_single(tgtport->dev, lsreq->rqstdma,
370				  (lsreq->rqstlen + lsreq->rsplen),
371				  DMA_BIDIRECTIONAL);
372
373	nvmet_fc_tgtport_put(tgtport);
374}
375
376static int
377__nvmet_fc_send_ls_req(struct nvmet_fc_tgtport *tgtport,
378		struct nvmet_fc_ls_req_op *lsop,
379		void (*done)(struct nvmefc_ls_req *req, int status))
380{
381	struct nvmefc_ls_req *lsreq = &lsop->ls_req;
382	unsigned long flags;
383	int ret = 0;
384
385	if (!tgtport->ops->ls_req)
386		return -EOPNOTSUPP;
387
388	if (!nvmet_fc_tgtport_get(tgtport))
389		return -ESHUTDOWN;
390
391	lsreq->done = done;
392	lsop->req_queued = false;
393	INIT_LIST_HEAD(&lsop->lsreq_list);
394
395	lsreq->rqstdma = fc_dma_map_single(tgtport->dev, lsreq->rqstaddr,
396				  lsreq->rqstlen + lsreq->rsplen,
397				  DMA_BIDIRECTIONAL);
398	if (fc_dma_mapping_error(tgtport->dev, lsreq->rqstdma)) {
399		ret = -EFAULT;
400		goto out_puttgtport;
401	}
402	lsreq->rspdma = lsreq->rqstdma + lsreq->rqstlen;
403
404	spin_lock_irqsave(&tgtport->lock, flags);
405
406	list_add_tail(&lsop->lsreq_list, &tgtport->ls_req_list);
407
408	lsop->req_queued = true;
409
410	spin_unlock_irqrestore(&tgtport->lock, flags);
411
412	ret = tgtport->ops->ls_req(&tgtport->fc_target_port, lsop->hosthandle,
413				   lsreq);
414	if (ret)
415		goto out_unlink;
416
417	return 0;
418
419out_unlink:
420	lsop->ls_error = ret;
421	spin_lock_irqsave(&tgtport->lock, flags);
422	lsop->req_queued = false;
423	list_del(&lsop->lsreq_list);
424	spin_unlock_irqrestore(&tgtport->lock, flags);
425	fc_dma_unmap_single(tgtport->dev, lsreq->rqstdma,
426				  (lsreq->rqstlen + lsreq->rsplen),
427				  DMA_BIDIRECTIONAL);
428out_puttgtport:
429	nvmet_fc_tgtport_put(tgtport);
430
431	return ret;
432}
433
434static int
435nvmet_fc_send_ls_req_async(struct nvmet_fc_tgtport *tgtport,
436		struct nvmet_fc_ls_req_op *lsop,
437		void (*done)(struct nvmefc_ls_req *req, int status))
438{
439	/* don't wait for completion */
440
441	return __nvmet_fc_send_ls_req(tgtport, lsop, done);
442}
443
444static void
445nvmet_fc_disconnect_assoc_done(struct nvmefc_ls_req *lsreq, int status)
446{
447	struct nvmet_fc_ls_req_op *lsop =
448		container_of(lsreq, struct nvmet_fc_ls_req_op, ls_req);
449
450	__nvmet_fc_finish_ls_req(lsop);
451
452	/* fc-nvme target doesn't care about success or failure of cmd */
453
454	kfree(lsop);
455}
456
457/*
458 * This routine sends a FC-NVME LS to disconnect (aka terminate)
459 * the FC-NVME Association.  Terminating the association also
460 * terminates the FC-NVME connections (per queue, both admin and io
461 * queues) that are part of the association. E.g. things are torn
462 * down, and the related FC-NVME Association ID and Connection IDs
463 * become invalid.
464 *
465 * The behavior of the fc-nvme target is such that it's
466 * understanding of the association and connections will implicitly
467 * be torn down. The action is implicit as it may be due to a loss of
468 * connectivity with the fc-nvme host, so the target may never get a
469 * response even if it tried.  As such, the action of this routine
470 * is to asynchronously send the LS, ignore any results of the LS, and
471 * continue on with terminating the association. If the fc-nvme host
472 * is present and receives the LS, it too can tear down.
473 */
474static void
475nvmet_fc_xmt_disconnect_assoc(struct nvmet_fc_tgt_assoc *assoc)
476{
477	struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
478	struct fcnvme_ls_disconnect_assoc_rqst *discon_rqst;
479	struct fcnvme_ls_disconnect_assoc_acc *discon_acc;
480	struct nvmet_fc_ls_req_op *lsop;
481	struct nvmefc_ls_req *lsreq;
482	int ret;
483
484	/*
485	 * If ls_req is NULL or no hosthandle, it's an older lldd and no
486	 * message is normal. Otherwise, send unless the hostport has
487	 * already been invalidated by the lldd.
488	 */
489	if (!tgtport->ops->ls_req || !assoc->hostport ||
490	    assoc->hostport->invalid)
491		return;
492
493	lsop = kzalloc((sizeof(*lsop) +
494			sizeof(*discon_rqst) + sizeof(*discon_acc) +
495			tgtport->ops->lsrqst_priv_sz), GFP_KERNEL);
496	if (!lsop) {
497		dev_info(tgtport->dev,
498			"{%d:%d} send Disconnect Association failed: ENOMEM\n",
499			tgtport->fc_target_port.port_num, assoc->a_id);
500		return;
501	}
502
503	discon_rqst = (struct fcnvme_ls_disconnect_assoc_rqst *)&lsop[1];
504	discon_acc = (struct fcnvme_ls_disconnect_assoc_acc *)&discon_rqst[1];
505	lsreq = &lsop->ls_req;
506	if (tgtport->ops->lsrqst_priv_sz)
507		lsreq->private = (void *)&discon_acc[1];
508	else
509		lsreq->private = NULL;
510
511	lsop->tgtport = tgtport;
512	lsop->hosthandle = assoc->hostport->hosthandle;
513
514	nvmefc_fmt_lsreq_discon_assoc(lsreq, discon_rqst, discon_acc,
515				assoc->association_id);
516
517	ret = nvmet_fc_send_ls_req_async(tgtport, lsop,
518				nvmet_fc_disconnect_assoc_done);
519	if (ret) {
520		dev_info(tgtport->dev,
521			"{%d:%d} XMT Disconnect Association failed: %d\n",
522			tgtport->fc_target_port.port_num, assoc->a_id, ret);
523		kfree(lsop);
524	}
525}
526
527
528/* *********************** FC-NVME Port Management ************************ */
529
530
531static int
532nvmet_fc_alloc_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
533{
534	struct nvmet_fc_ls_iod *iod;
535	int i;
536
537	iod = kcalloc(NVMET_LS_CTX_COUNT, sizeof(struct nvmet_fc_ls_iod),
538			GFP_KERNEL);
539	if (!iod)
540		return -ENOMEM;
541
542	tgtport->iod = iod;
543
544	for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
545		INIT_WORK(&iod->work, nvmet_fc_handle_ls_rqst_work);
546		iod->tgtport = tgtport;
547		list_add_tail(&iod->ls_rcv_list, &tgtport->ls_rcv_list);
548
549		iod->rqstbuf = kzalloc(sizeof(union nvmefc_ls_requests) +
550				       sizeof(union nvmefc_ls_responses),
551				       GFP_KERNEL);
552		if (!iod->rqstbuf)
553			goto out_fail;
554
555		iod->rspbuf = (union nvmefc_ls_responses *)&iod->rqstbuf[1];
556
557		iod->rspdma = fc_dma_map_single(tgtport->dev, iod->rspbuf,
558						sizeof(*iod->rspbuf),
559						DMA_TO_DEVICE);
560		if (fc_dma_mapping_error(tgtport->dev, iod->rspdma))
561			goto out_fail;
562	}
563
564	return 0;
565
566out_fail:
567	kfree(iod->rqstbuf);
568	list_del(&iod->ls_rcv_list);
569	for (iod--, i--; i >= 0; iod--, i--) {
570		fc_dma_unmap_single(tgtport->dev, iod->rspdma,
571				sizeof(*iod->rspbuf), DMA_TO_DEVICE);
572		kfree(iod->rqstbuf);
573		list_del(&iod->ls_rcv_list);
574	}
575
576	kfree(iod);
577
578	return -EFAULT;
579}
580
581static void
582nvmet_fc_free_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
583{
584	struct nvmet_fc_ls_iod *iod = tgtport->iod;
585	int i;
586
587	for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
588		fc_dma_unmap_single(tgtport->dev,
589				iod->rspdma, sizeof(*iod->rspbuf),
590				DMA_TO_DEVICE);
591		kfree(iod->rqstbuf);
592		list_del(&iod->ls_rcv_list);
593	}
594	kfree(tgtport->iod);
595}
596
597static struct nvmet_fc_ls_iod *
598nvmet_fc_alloc_ls_iod(struct nvmet_fc_tgtport *tgtport)
599{
600	struct nvmet_fc_ls_iod *iod;
601	unsigned long flags;
602
603	spin_lock_irqsave(&tgtport->lock, flags);
604	iod = list_first_entry_or_null(&tgtport->ls_rcv_list,
605					struct nvmet_fc_ls_iod, ls_rcv_list);
606	if (iod)
607		list_move_tail(&iod->ls_rcv_list, &tgtport->ls_busylist);
608	spin_unlock_irqrestore(&tgtport->lock, flags);
609	return iod;
610}
611
612
613static void
614nvmet_fc_free_ls_iod(struct nvmet_fc_tgtport *tgtport,
615			struct nvmet_fc_ls_iod *iod)
616{
617	unsigned long flags;
618
619	spin_lock_irqsave(&tgtport->lock, flags);
620	list_move(&iod->ls_rcv_list, &tgtport->ls_rcv_list);
621	spin_unlock_irqrestore(&tgtport->lock, flags);
622}
623
624static void
625nvmet_fc_prep_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
626				struct nvmet_fc_tgt_queue *queue)
627{
628	struct nvmet_fc_fcp_iod *fod = queue->fod;
629	int i;
630
631	for (i = 0; i < queue->sqsize; fod++, i++) {
632		INIT_WORK(&fod->defer_work, nvmet_fc_fcp_rqst_op_defer_work);
633		fod->tgtport = tgtport;
634		fod->queue = queue;
635		fod->active = false;
636		fod->abort = false;
637		fod->aborted = false;
638		fod->fcpreq = NULL;
639		list_add_tail(&fod->fcp_list, &queue->fod_list);
640		spin_lock_init(&fod->flock);
641
642		fod->rspdma = fc_dma_map_single(tgtport->dev, &fod->rspiubuf,
643					sizeof(fod->rspiubuf), DMA_TO_DEVICE);
644		if (fc_dma_mapping_error(tgtport->dev, fod->rspdma)) {
645			list_del(&fod->fcp_list);
646			for (fod--, i--; i >= 0; fod--, i--) {
647				fc_dma_unmap_single(tgtport->dev, fod->rspdma,
648						sizeof(fod->rspiubuf),
649						DMA_TO_DEVICE);
650				fod->rspdma = 0L;
651				list_del(&fod->fcp_list);
652			}
653
654			return;
655		}
656	}
657}
658
659static void
660nvmet_fc_destroy_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
661				struct nvmet_fc_tgt_queue *queue)
662{
663	struct nvmet_fc_fcp_iod *fod = queue->fod;
664	int i;
665
666	for (i = 0; i < queue->sqsize; fod++, i++) {
667		if (fod->rspdma)
668			fc_dma_unmap_single(tgtport->dev, fod->rspdma,
669				sizeof(fod->rspiubuf), DMA_TO_DEVICE);
670	}
671}
672
673static struct nvmet_fc_fcp_iod *
674nvmet_fc_alloc_fcp_iod(struct nvmet_fc_tgt_queue *queue)
675{
676	struct nvmet_fc_fcp_iod *fod;
677
678	lockdep_assert_held(&queue->qlock);
679
680	fod = list_first_entry_or_null(&queue->fod_list,
681					struct nvmet_fc_fcp_iod, fcp_list);
682	if (fod) {
683		list_del(&fod->fcp_list);
684		fod->active = true;
685		/*
686		 * no queue reference is taken, as it was taken by the
687		 * queue lookup just prior to the allocation. The iod
688		 * will "inherit" that reference.
689		 */
690	}
691	return fod;
692}
693
694
695static void
696nvmet_fc_queue_fcp_req(struct nvmet_fc_tgtport *tgtport,
697		       struct nvmet_fc_tgt_queue *queue,
698		       struct nvmefc_tgt_fcp_req *fcpreq)
699{
700	struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
701
702	/*
703	 * put all admin cmds on hw queue id 0. All io commands go to
704	 * the respective hw queue based on a modulo basis
705	 */
706	fcpreq->hwqid = queue->qid ?
707			((queue->qid - 1) % tgtport->ops->max_hw_queues) : 0;
708
709	nvmet_fc_handle_fcp_rqst(tgtport, fod);
710}
711
712static void
713nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work)
714{
715	struct nvmet_fc_fcp_iod *fod =
716		container_of(work, struct nvmet_fc_fcp_iod, defer_work);
717
718	/* Submit deferred IO for processing */
719	nvmet_fc_queue_fcp_req(fod->tgtport, fod->queue, fod->fcpreq);
720
721}
722
723static void
724nvmet_fc_free_fcp_iod(struct nvmet_fc_tgt_queue *queue,
725			struct nvmet_fc_fcp_iod *fod)
726{
727	struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
728	struct nvmet_fc_tgtport *tgtport = fod->tgtport;
729	struct nvmet_fc_defer_fcp_req *deferfcp;
730	unsigned long flags;
731
732	fc_dma_sync_single_for_cpu(tgtport->dev, fod->rspdma,
733				sizeof(fod->rspiubuf), DMA_TO_DEVICE);
734
735	fcpreq->nvmet_fc_private = NULL;
736
737	fod->active = false;
738	fod->abort = false;
739	fod->aborted = false;
740	fod->writedataactive = false;
741	fod->fcpreq = NULL;
742
743	tgtport->ops->fcp_req_release(&tgtport->fc_target_port, fcpreq);
744
745	/* release the queue lookup reference on the completed IO */
746	nvmet_fc_tgt_q_put(queue);
747
748	spin_lock_irqsave(&queue->qlock, flags);
749	deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
750				struct nvmet_fc_defer_fcp_req, req_list);
751	if (!deferfcp) {
752		list_add_tail(&fod->fcp_list, &fod->queue->fod_list);
753		spin_unlock_irqrestore(&queue->qlock, flags);
754		return;
755	}
756
757	/* Re-use the fod for the next pending cmd that was deferred */
758	list_del(&deferfcp->req_list);
759
760	fcpreq = deferfcp->fcp_req;
761
762	/* deferfcp can be reused for another IO at a later date */
763	list_add_tail(&deferfcp->req_list, &queue->avail_defer_list);
764
765	spin_unlock_irqrestore(&queue->qlock, flags);
766
767	/* Save NVME CMD IO in fod */
768	memcpy(&fod->cmdiubuf, fcpreq->rspaddr, fcpreq->rsplen);
769
770	/* Setup new fcpreq to be processed */
771	fcpreq->rspaddr = NULL;
772	fcpreq->rsplen  = 0;
773	fcpreq->nvmet_fc_private = fod;
774	fod->fcpreq = fcpreq;
775	fod->active = true;
776
777	/* inform LLDD IO is now being processed */
778	tgtport->ops->defer_rcv(&tgtport->fc_target_port, fcpreq);
779
780	/*
781	 * Leave the queue lookup get reference taken when
782	 * fod was originally allocated.
783	 */
784
785	queue_work(queue->work_q, &fod->defer_work);
786}
787
788static struct nvmet_fc_tgt_queue *
789nvmet_fc_alloc_target_queue(struct nvmet_fc_tgt_assoc *assoc,
790			u16 qid, u16 sqsize)
791{
792	struct nvmet_fc_tgt_queue *queue;
793	unsigned long flags;
794	int ret;
795
796	if (qid > NVMET_NR_QUEUES)
797		return NULL;
798
799	queue = kzalloc(struct_size(queue, fod, sqsize), GFP_KERNEL);
800	if (!queue)
801		return NULL;
802
803	if (!nvmet_fc_tgt_a_get(assoc))
804		goto out_free_queue;
805
806	queue->work_q = alloc_workqueue("ntfc%d.%d.%d", 0, 0,
807				assoc->tgtport->fc_target_port.port_num,
808				assoc->a_id, qid);
809	if (!queue->work_q)
810		goto out_a_put;
811
812	queue->qid = qid;
813	queue->sqsize = sqsize;
814	queue->assoc = assoc;
815	INIT_LIST_HEAD(&queue->fod_list);
816	INIT_LIST_HEAD(&queue->avail_defer_list);
817	INIT_LIST_HEAD(&queue->pending_cmd_list);
818	atomic_set(&queue->connected, 0);
819	atomic_set(&queue->sqtail, 0);
820	atomic_set(&queue->rsn, 1);
821	atomic_set(&queue->zrspcnt, 0);
822	spin_lock_init(&queue->qlock);
823	kref_init(&queue->ref);
824
825	nvmet_fc_prep_fcp_iodlist(assoc->tgtport, queue);
826
827	ret = nvmet_sq_init(&queue->nvme_sq);
828	if (ret)
829		goto out_fail_iodlist;
830
831	WARN_ON(assoc->queues[qid]);
832	spin_lock_irqsave(&assoc->tgtport->lock, flags);
833	assoc->queues[qid] = queue;
834	spin_unlock_irqrestore(&assoc->tgtport->lock, flags);
835
836	return queue;
837
838out_fail_iodlist:
839	nvmet_fc_destroy_fcp_iodlist(assoc->tgtport, queue);
840	destroy_workqueue(queue->work_q);
841out_a_put:
842	nvmet_fc_tgt_a_put(assoc);
843out_free_queue:
844	kfree(queue);
845	return NULL;
846}
847
848
849static void
850nvmet_fc_tgt_queue_free(struct kref *ref)
851{
852	struct nvmet_fc_tgt_queue *queue =
853		container_of(ref, struct nvmet_fc_tgt_queue, ref);
854	unsigned long flags;
855
856	spin_lock_irqsave(&queue->assoc->tgtport->lock, flags);
857	queue->assoc->queues[queue->qid] = NULL;
858	spin_unlock_irqrestore(&queue->assoc->tgtport->lock, flags);
859
860	nvmet_fc_destroy_fcp_iodlist(queue->assoc->tgtport, queue);
861
862	nvmet_fc_tgt_a_put(queue->assoc);
863
864	destroy_workqueue(queue->work_q);
865
866	kfree(queue);
867}
868
869static void
870nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue)
871{
872	kref_put(&queue->ref, nvmet_fc_tgt_queue_free);
873}
874
875static int
876nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue)
877{
878	return kref_get_unless_zero(&queue->ref);
879}
880
881
882static void
883nvmet_fc_delete_target_queue(struct nvmet_fc_tgt_queue *queue)
884{
885	struct nvmet_fc_tgtport *tgtport = queue->assoc->tgtport;
886	struct nvmet_fc_fcp_iod *fod = queue->fod;
887	struct nvmet_fc_defer_fcp_req *deferfcp, *tempptr;
888	unsigned long flags;
889	int i;
890	bool disconnect;
891
892	disconnect = atomic_xchg(&queue->connected, 0);
893
894	/* if not connected, nothing to do */
895	if (!disconnect)
896		return;
897
898	spin_lock_irqsave(&queue->qlock, flags);
899	/* abort outstanding io's */
900	for (i = 0; i < queue->sqsize; fod++, i++) {
901		if (fod->active) {
902			spin_lock(&fod->flock);
903			fod->abort = true;
904			/*
905			 * only call lldd abort routine if waiting for
906			 * writedata. other outstanding ops should finish
907			 * on their own.
908			 */
909			if (fod->writedataactive) {
910				fod->aborted = true;
911				spin_unlock(&fod->flock);
912				tgtport->ops->fcp_abort(
913					&tgtport->fc_target_port, fod->fcpreq);
914			} else
915				spin_unlock(&fod->flock);
916		}
917	}
918
919	/* Cleanup defer'ed IOs in queue */
920	list_for_each_entry_safe(deferfcp, tempptr, &queue->avail_defer_list,
921				req_list) {
922		list_del(&deferfcp->req_list);
923		kfree(deferfcp);
924	}
925
926	for (;;) {
927		deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
928				struct nvmet_fc_defer_fcp_req, req_list);
929		if (!deferfcp)
930			break;
931
932		list_del(&deferfcp->req_list);
933		spin_unlock_irqrestore(&queue->qlock, flags);
934
935		tgtport->ops->defer_rcv(&tgtport->fc_target_port,
936				deferfcp->fcp_req);
937
938		tgtport->ops->fcp_abort(&tgtport->fc_target_port,
939				deferfcp->fcp_req);
940
941		tgtport->ops->fcp_req_release(&tgtport->fc_target_port,
942				deferfcp->fcp_req);
943
944		/* release the queue lookup reference */
945		nvmet_fc_tgt_q_put(queue);
946
947		kfree(deferfcp);
948
949		spin_lock_irqsave(&queue->qlock, flags);
950	}
951	spin_unlock_irqrestore(&queue->qlock, flags);
952
953	flush_workqueue(queue->work_q);
954
955	nvmet_sq_destroy(&queue->nvme_sq);
956
957	nvmet_fc_tgt_q_put(queue);
958}
959
960static struct nvmet_fc_tgt_queue *
961nvmet_fc_find_target_queue(struct nvmet_fc_tgtport *tgtport,
962				u64 connection_id)
963{
964	struct nvmet_fc_tgt_assoc *assoc;
965	struct nvmet_fc_tgt_queue *queue;
966	u64 association_id = nvmet_fc_getassociationid(connection_id);
967	u16 qid = nvmet_fc_getqueueid(connection_id);
968	unsigned long flags;
969
970	if (qid > NVMET_NR_QUEUES)
971		return NULL;
972
973	spin_lock_irqsave(&tgtport->lock, flags);
974	list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
975		if (association_id == assoc->association_id) {
976			queue = assoc->queues[qid];
977			if (queue &&
978			    (!atomic_read(&queue->connected) ||
979			     !nvmet_fc_tgt_q_get(queue)))
980				queue = NULL;
981			spin_unlock_irqrestore(&tgtport->lock, flags);
982			return queue;
983		}
984	}
985	spin_unlock_irqrestore(&tgtport->lock, flags);
986	return NULL;
987}
988
989static void
990nvmet_fc_hostport_free(struct kref *ref)
991{
992	struct nvmet_fc_hostport *hostport =
993		container_of(ref, struct nvmet_fc_hostport, ref);
994	struct nvmet_fc_tgtport *tgtport = hostport->tgtport;
995	unsigned long flags;
996
997	spin_lock_irqsave(&tgtport->lock, flags);
998	list_del(&hostport->host_list);
999	spin_unlock_irqrestore(&tgtport->lock, flags);
1000	if (tgtport->ops->host_release && hostport->invalid)
1001		tgtport->ops->host_release(hostport->hosthandle);
1002	kfree(hostport);
1003	nvmet_fc_tgtport_put(tgtport);
1004}
1005
1006static void
1007nvmet_fc_hostport_put(struct nvmet_fc_hostport *hostport)
1008{
1009	kref_put(&hostport->ref, nvmet_fc_hostport_free);
1010}
1011
1012static int
1013nvmet_fc_hostport_get(struct nvmet_fc_hostport *hostport)
1014{
1015	return kref_get_unless_zero(&hostport->ref);
1016}
1017
1018static void
1019nvmet_fc_free_hostport(struct nvmet_fc_hostport *hostport)
1020{
1021	/* if LLDD not implemented, leave as NULL */
1022	if (!hostport || !hostport->hosthandle)
1023		return;
1024
1025	nvmet_fc_hostport_put(hostport);
1026}
1027
1028static struct nvmet_fc_hostport *
1029nvmet_fc_alloc_hostport(struct nvmet_fc_tgtport *tgtport, void *hosthandle)
1030{
1031	struct nvmet_fc_hostport *newhost, *host, *match = NULL;
1032	unsigned long flags;
1033
1034	/* if LLDD not implemented, leave as NULL */
1035	if (!hosthandle)
1036		return NULL;
1037
1038	/* take reference for what will be the newly allocated hostport */
1039	if (!nvmet_fc_tgtport_get(tgtport))
1040		return ERR_PTR(-EINVAL);
1041
1042	newhost = kzalloc(sizeof(*newhost), GFP_KERNEL);
1043	if (!newhost) {
1044		spin_lock_irqsave(&tgtport->lock, flags);
1045		list_for_each_entry(host, &tgtport->host_list, host_list) {
1046			if (host->hosthandle == hosthandle && !host->invalid) {
1047				if (nvmet_fc_hostport_get(host)) {
1048					match = host;
1049					break;
1050				}
1051			}
1052		}
1053		spin_unlock_irqrestore(&tgtport->lock, flags);
1054		/* no allocation - release reference */
1055		nvmet_fc_tgtport_put(tgtport);
1056		return (match) ? match : ERR_PTR(-ENOMEM);
1057	}
1058
1059	newhost->tgtport = tgtport;
1060	newhost->hosthandle = hosthandle;
1061	INIT_LIST_HEAD(&newhost->host_list);
1062	kref_init(&newhost->ref);
1063
1064	spin_lock_irqsave(&tgtport->lock, flags);
1065	list_for_each_entry(host, &tgtport->host_list, host_list) {
1066		if (host->hosthandle == hosthandle && !host->invalid) {
1067			if (nvmet_fc_hostport_get(host)) {
1068				match = host;
1069				break;
1070			}
1071		}
1072	}
1073	if (match) {
1074		kfree(newhost);
1075		newhost = NULL;
1076		/* releasing allocation - release reference */
1077		nvmet_fc_tgtport_put(tgtport);
1078	} else
1079		list_add_tail(&newhost->host_list, &tgtport->host_list);
1080	spin_unlock_irqrestore(&tgtport->lock, flags);
1081
1082	return (match) ? match : newhost;
1083}
1084
1085static void
1086nvmet_fc_delete_assoc(struct work_struct *work)
1087{
1088	struct nvmet_fc_tgt_assoc *assoc =
1089		container_of(work, struct nvmet_fc_tgt_assoc, del_work);
1090
1091	nvmet_fc_delete_target_assoc(assoc);
1092	nvmet_fc_tgt_a_put(assoc);
1093}
1094
1095static struct nvmet_fc_tgt_assoc *
1096nvmet_fc_alloc_target_assoc(struct nvmet_fc_tgtport *tgtport, void *hosthandle)
1097{
1098	struct nvmet_fc_tgt_assoc *assoc, *tmpassoc;
1099	unsigned long flags;
1100	u64 ran;
1101	int idx;
1102	bool needrandom = true;
1103
1104	assoc = kzalloc(sizeof(*assoc), GFP_KERNEL);
1105	if (!assoc)
1106		return NULL;
1107
1108	idx = ida_simple_get(&tgtport->assoc_cnt, 0, 0, GFP_KERNEL);
1109	if (idx < 0)
1110		goto out_free_assoc;
1111
1112	if (!nvmet_fc_tgtport_get(tgtport))
1113		goto out_ida;
1114
1115	assoc->hostport = nvmet_fc_alloc_hostport(tgtport, hosthandle);
1116	if (IS_ERR(assoc->hostport))
1117		goto out_put;
1118
1119	assoc->tgtport = tgtport;
1120	assoc->a_id = idx;
1121	INIT_LIST_HEAD(&assoc->a_list);
1122	kref_init(&assoc->ref);
1123	INIT_WORK(&assoc->del_work, nvmet_fc_delete_assoc);
1124	atomic_set(&assoc->terminating, 0);
1125
1126	while (needrandom) {
1127		get_random_bytes(&ran, sizeof(ran) - BYTES_FOR_QID);
1128		ran = ran << BYTES_FOR_QID_SHIFT;
1129
1130		spin_lock_irqsave(&tgtport->lock, flags);
1131		needrandom = false;
1132		list_for_each_entry(tmpassoc, &tgtport->assoc_list, a_list) {
1133			if (ran == tmpassoc->association_id) {
1134				needrandom = true;
1135				break;
1136			}
1137		}
1138		if (!needrandom) {
1139			assoc->association_id = ran;
1140			list_add_tail(&assoc->a_list, &tgtport->assoc_list);
1141		}
1142		spin_unlock_irqrestore(&tgtport->lock, flags);
1143	}
1144
1145	return assoc;
1146
1147out_put:
1148	nvmet_fc_tgtport_put(tgtport);
1149out_ida:
1150	ida_simple_remove(&tgtport->assoc_cnt, idx);
1151out_free_assoc:
1152	kfree(assoc);
1153	return NULL;
1154}
1155
1156static void
1157nvmet_fc_target_assoc_free(struct kref *ref)
1158{
1159	struct nvmet_fc_tgt_assoc *assoc =
1160		container_of(ref, struct nvmet_fc_tgt_assoc, ref);
1161	struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
1162	struct nvmet_fc_ls_iod	*oldls;
1163	unsigned long flags;
1164
1165	/* Send Disconnect now that all i/o has completed */
1166	nvmet_fc_xmt_disconnect_assoc(assoc);
1167
1168	nvmet_fc_free_hostport(assoc->hostport);
1169	spin_lock_irqsave(&tgtport->lock, flags);
1170	list_del(&assoc->a_list);
1171	oldls = assoc->rcv_disconn;
1172	spin_unlock_irqrestore(&tgtport->lock, flags);
1173	/* if pending Rcv Disconnect Association LS, send rsp now */
1174	if (oldls)
1175		nvmet_fc_xmt_ls_rsp(tgtport, oldls);
1176	ida_simple_remove(&tgtport->assoc_cnt, assoc->a_id);
1177	dev_info(tgtport->dev,
1178		"{%d:%d} Association freed\n",
1179		tgtport->fc_target_port.port_num, assoc->a_id);
1180	kfree(assoc);
1181	nvmet_fc_tgtport_put(tgtport);
1182}
1183
1184static void
1185nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc)
1186{
1187	kref_put(&assoc->ref, nvmet_fc_target_assoc_free);
1188}
1189
1190static int
1191nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc)
1192{
1193	return kref_get_unless_zero(&assoc->ref);
1194}
1195
1196static void
1197nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc)
1198{
1199	struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
1200	struct nvmet_fc_tgt_queue *queue;
1201	unsigned long flags;
1202	int i, terminating;
1203
1204	terminating = atomic_xchg(&assoc->terminating, 1);
1205
1206	/* if already terminating, do nothing */
1207	if (terminating)
1208		return;
1209
1210	spin_lock_irqsave(&tgtport->lock, flags);
1211	for (i = NVMET_NR_QUEUES; i >= 0; i--) {
1212		queue = assoc->queues[i];
1213		if (queue) {
1214			if (!nvmet_fc_tgt_q_get(queue))
1215				continue;
1216			spin_unlock_irqrestore(&tgtport->lock, flags);
1217			nvmet_fc_delete_target_queue(queue);
1218			nvmet_fc_tgt_q_put(queue);
1219			spin_lock_irqsave(&tgtport->lock, flags);
1220		}
1221	}
1222	spin_unlock_irqrestore(&tgtport->lock, flags);
1223
1224	dev_info(tgtport->dev,
1225		"{%d:%d} Association deleted\n",
1226		tgtport->fc_target_port.port_num, assoc->a_id);
1227
1228	nvmet_fc_tgt_a_put(assoc);
1229}
1230
1231static struct nvmet_fc_tgt_assoc *
1232nvmet_fc_find_target_assoc(struct nvmet_fc_tgtport *tgtport,
1233				u64 association_id)
1234{
1235	struct nvmet_fc_tgt_assoc *assoc;
1236	struct nvmet_fc_tgt_assoc *ret = NULL;
1237	unsigned long flags;
1238
1239	spin_lock_irqsave(&tgtport->lock, flags);
1240	list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
1241		if (association_id == assoc->association_id) {
1242			ret = assoc;
1243			if (!nvmet_fc_tgt_a_get(assoc))
1244				ret = NULL;
1245			break;
1246		}
1247	}
1248	spin_unlock_irqrestore(&tgtport->lock, flags);
1249
1250	return ret;
1251}
1252
1253static void
1254nvmet_fc_portentry_bind(struct nvmet_fc_tgtport *tgtport,
1255			struct nvmet_fc_port_entry *pe,
1256			struct nvmet_port *port)
1257{
1258	lockdep_assert_held(&nvmet_fc_tgtlock);
1259
1260	pe->tgtport = tgtport;
1261	tgtport->pe = pe;
1262
1263	pe->port = port;
1264	port->priv = pe;
1265
1266	pe->node_name = tgtport->fc_target_port.node_name;
1267	pe->port_name = tgtport->fc_target_port.port_name;
1268	INIT_LIST_HEAD(&pe->pe_list);
1269
1270	list_add_tail(&pe->pe_list, &nvmet_fc_portentry_list);
1271}
1272
1273static void
1274nvmet_fc_portentry_unbind(struct nvmet_fc_port_entry *pe)
1275{
1276	unsigned long flags;
1277
1278	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1279	if (pe->tgtport)
1280		pe->tgtport->pe = NULL;
1281	list_del(&pe->pe_list);
1282	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1283}
1284
1285/*
1286 * called when a targetport deregisters. Breaks the relationship
1287 * with the nvmet port, but leaves the port_entry in place so that
1288 * re-registration can resume operation.
1289 */
1290static void
1291nvmet_fc_portentry_unbind_tgt(struct nvmet_fc_tgtport *tgtport)
1292{
1293	struct nvmet_fc_port_entry *pe;
1294	unsigned long flags;
1295
1296	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1297	pe = tgtport->pe;
1298	if (pe)
1299		pe->tgtport = NULL;
1300	tgtport->pe = NULL;
1301	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1302}
1303
1304/*
1305 * called when a new targetport is registered. Looks in the
1306 * existing nvmet port_entries to see if the nvmet layer is
1307 * configured for the targetport's wwn's. (the targetport existed,
1308 * nvmet configured, the lldd unregistered the tgtport, and is now
1309 * reregistering the same targetport).  If so, set the nvmet port
1310 * port entry on the targetport.
1311 */
1312static void
1313nvmet_fc_portentry_rebind_tgt(struct nvmet_fc_tgtport *tgtport)
1314{
1315	struct nvmet_fc_port_entry *pe;
1316	unsigned long flags;
1317
1318	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1319	list_for_each_entry(pe, &nvmet_fc_portentry_list, pe_list) {
1320		if (tgtport->fc_target_port.node_name == pe->node_name &&
1321		    tgtport->fc_target_port.port_name == pe->port_name) {
1322			WARN_ON(pe->tgtport);
1323			tgtport->pe = pe;
1324			pe->tgtport = tgtport;
1325			break;
1326		}
1327	}
1328	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1329}
1330
1331/**
1332 * nvme_fc_register_targetport - transport entry point called by an
1333 *                              LLDD to register the existence of a local
1334 *                              NVME subystem FC port.
1335 * @pinfo:     pointer to information about the port to be registered
1336 * @template:  LLDD entrypoints and operational parameters for the port
1337 * @dev:       physical hardware device node port corresponds to. Will be
1338 *             used for DMA mappings
1339 * @portptr:   pointer to a local port pointer. Upon success, the routine
1340 *             will allocate a nvme_fc_local_port structure and place its
1341 *             address in the local port pointer. Upon failure, local port
1342 *             pointer will be set to NULL.
1343 *
1344 * Returns:
1345 * a completion status. Must be 0 upon success; a negative errno
1346 * (ex: -ENXIO) upon failure.
1347 */
1348int
1349nvmet_fc_register_targetport(struct nvmet_fc_port_info *pinfo,
1350			struct nvmet_fc_target_template *template,
1351			struct device *dev,
1352			struct nvmet_fc_target_port **portptr)
1353{
1354	struct nvmet_fc_tgtport *newrec;
1355	unsigned long flags;
1356	int ret, idx;
1357
1358	if (!template->xmt_ls_rsp || !template->fcp_op ||
1359	    !template->fcp_abort ||
1360	    !template->fcp_req_release || !template->targetport_delete ||
1361	    !template->max_hw_queues || !template->max_sgl_segments ||
1362	    !template->max_dif_sgl_segments || !template->dma_boundary) {
1363		ret = -EINVAL;
1364		goto out_regtgt_failed;
1365	}
1366
1367	newrec = kzalloc((sizeof(*newrec) + template->target_priv_sz),
1368			 GFP_KERNEL);
1369	if (!newrec) {
1370		ret = -ENOMEM;
1371		goto out_regtgt_failed;
1372	}
1373
1374	idx = ida_simple_get(&nvmet_fc_tgtport_cnt, 0, 0, GFP_KERNEL);
1375	if (idx < 0) {
1376		ret = -ENOSPC;
1377		goto out_fail_kfree;
1378	}
1379
1380	if (!get_device(dev) && dev) {
1381		ret = -ENODEV;
1382		goto out_ida_put;
1383	}
1384
1385	newrec->fc_target_port.node_name = pinfo->node_name;
1386	newrec->fc_target_port.port_name = pinfo->port_name;
1387	if (template->target_priv_sz)
1388		newrec->fc_target_port.private = &newrec[1];
1389	else
1390		newrec->fc_target_port.private = NULL;
1391	newrec->fc_target_port.port_id = pinfo->port_id;
1392	newrec->fc_target_port.port_num = idx;
1393	INIT_LIST_HEAD(&newrec->tgt_list);
1394	newrec->dev = dev;
1395	newrec->ops = template;
1396	spin_lock_init(&newrec->lock);
1397	INIT_LIST_HEAD(&newrec->ls_rcv_list);
1398	INIT_LIST_HEAD(&newrec->ls_req_list);
1399	INIT_LIST_HEAD(&newrec->ls_busylist);
1400	INIT_LIST_HEAD(&newrec->assoc_list);
1401	INIT_LIST_HEAD(&newrec->host_list);
1402	kref_init(&newrec->ref);
1403	ida_init(&newrec->assoc_cnt);
1404	newrec->max_sg_cnt = template->max_sgl_segments;
1405
1406	ret = nvmet_fc_alloc_ls_iodlist(newrec);
1407	if (ret) {
1408		ret = -ENOMEM;
1409		goto out_free_newrec;
1410	}
1411
1412	nvmet_fc_portentry_rebind_tgt(newrec);
1413
1414	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1415	list_add_tail(&newrec->tgt_list, &nvmet_fc_target_list);
1416	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1417
1418	*portptr = &newrec->fc_target_port;
1419	return 0;
1420
1421out_free_newrec:
1422	put_device(dev);
1423out_ida_put:
1424	ida_simple_remove(&nvmet_fc_tgtport_cnt, idx);
1425out_fail_kfree:
1426	kfree(newrec);
1427out_regtgt_failed:
1428	*portptr = NULL;
1429	return ret;
1430}
1431EXPORT_SYMBOL_GPL(nvmet_fc_register_targetport);
1432
1433
1434static void
1435nvmet_fc_free_tgtport(struct kref *ref)
1436{
1437	struct nvmet_fc_tgtport *tgtport =
1438		container_of(ref, struct nvmet_fc_tgtport, ref);
1439	struct device *dev = tgtport->dev;
1440	unsigned long flags;
1441
1442	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1443	list_del(&tgtport->tgt_list);
1444	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1445
1446	nvmet_fc_free_ls_iodlist(tgtport);
1447
1448	/* let the LLDD know we've finished tearing it down */
1449	tgtport->ops->targetport_delete(&tgtport->fc_target_port);
1450
1451	ida_simple_remove(&nvmet_fc_tgtport_cnt,
1452			tgtport->fc_target_port.port_num);
1453
1454	ida_destroy(&tgtport->assoc_cnt);
1455
1456	kfree(tgtport);
1457
1458	put_device(dev);
1459}
1460
1461static void
1462nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport)
1463{
1464	kref_put(&tgtport->ref, nvmet_fc_free_tgtport);
1465}
1466
1467static int
1468nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport)
1469{
1470	return kref_get_unless_zero(&tgtport->ref);
1471}
1472
1473static void
1474__nvmet_fc_free_assocs(struct nvmet_fc_tgtport *tgtport)
1475{
1476	struct nvmet_fc_tgt_assoc *assoc, *next;
1477	unsigned long flags;
1478
1479	spin_lock_irqsave(&tgtport->lock, flags);
1480	list_for_each_entry_safe(assoc, next,
1481				&tgtport->assoc_list, a_list) {
1482		if (!nvmet_fc_tgt_a_get(assoc))
1483			continue;
1484		if (!schedule_work(&assoc->del_work))
1485			/* already deleting - release local reference */
1486			nvmet_fc_tgt_a_put(assoc);
1487	}
1488	spin_unlock_irqrestore(&tgtport->lock, flags);
1489}
1490
1491/**
1492 * nvmet_fc_invalidate_host - transport entry point called by an LLDD
1493 *                       to remove references to a hosthandle for LS's.
1494 *
1495 * The nvmet-fc layer ensures that any references to the hosthandle
1496 * on the targetport are forgotten (set to NULL).  The LLDD will
1497 * typically call this when a login with a remote host port has been
1498 * lost, thus LS's for the remote host port are no longer possible.
1499 *
1500 * If an LS request is outstanding to the targetport/hosthandle (or
1501 * issued concurrently with the call to invalidate the host), the
1502 * LLDD is responsible for terminating/aborting the LS and completing
1503 * the LS request. It is recommended that these terminations/aborts
1504 * occur after calling to invalidate the host handle to avoid additional
1505 * retries by the nvmet-fc transport. The nvmet-fc transport may
1506 * continue to reference host handle while it cleans up outstanding
1507 * NVME associations. The nvmet-fc transport will call the
1508 * ops->host_release() callback to notify the LLDD that all references
1509 * are complete and the related host handle can be recovered.
1510 * Note: if there are no references, the callback may be called before
1511 * the invalidate host call returns.
1512 *
1513 * @target_port: pointer to the (registered) target port that a prior
1514 *              LS was received on and which supplied the transport the
1515 *              hosthandle.
1516 * @hosthandle: the handle (pointer) that represents the host port
1517 *              that no longer has connectivity and that LS's should
1518 *              no longer be directed to.
1519 */
1520void
1521nvmet_fc_invalidate_host(struct nvmet_fc_target_port *target_port,
1522			void *hosthandle)
1523{
1524	struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
1525	struct nvmet_fc_tgt_assoc *assoc, *next;
1526	unsigned long flags;
1527	bool noassoc = true;
1528
1529	spin_lock_irqsave(&tgtport->lock, flags);
1530	list_for_each_entry_safe(assoc, next,
1531				&tgtport->assoc_list, a_list) {
1532		if (!assoc->hostport ||
1533		    assoc->hostport->hosthandle != hosthandle)
1534			continue;
1535		if (!nvmet_fc_tgt_a_get(assoc))
1536			continue;
1537		assoc->hostport->invalid = 1;
1538		noassoc = false;
1539		if (!schedule_work(&assoc->del_work))
1540			/* already deleting - release local reference */
1541			nvmet_fc_tgt_a_put(assoc);
1542	}
1543	spin_unlock_irqrestore(&tgtport->lock, flags);
1544
1545	/* if there's nothing to wait for - call the callback */
1546	if (noassoc && tgtport->ops->host_release)
1547		tgtport->ops->host_release(hosthandle);
1548}
1549EXPORT_SYMBOL_GPL(nvmet_fc_invalidate_host);
1550
1551/*
1552 * nvmet layer has called to terminate an association
1553 */
1554static void
1555nvmet_fc_delete_ctrl(struct nvmet_ctrl *ctrl)
1556{
1557	struct nvmet_fc_tgtport *tgtport, *next;
1558	struct nvmet_fc_tgt_assoc *assoc;
1559	struct nvmet_fc_tgt_queue *queue;
1560	unsigned long flags;
1561	bool found_ctrl = false;
1562
1563	/* this is a bit ugly, but don't want to make locks layered */
1564	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1565	list_for_each_entry_safe(tgtport, next, &nvmet_fc_target_list,
1566			tgt_list) {
1567		if (!nvmet_fc_tgtport_get(tgtport))
1568			continue;
1569		spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1570
1571		spin_lock_irqsave(&tgtport->lock, flags);
1572		list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
1573			queue = assoc->queues[0];
1574			if (queue && queue->nvme_sq.ctrl == ctrl) {
1575				if (nvmet_fc_tgt_a_get(assoc))
1576					found_ctrl = true;
1577				break;
1578			}
1579		}
1580		spin_unlock_irqrestore(&tgtport->lock, flags);
1581
1582		nvmet_fc_tgtport_put(tgtport);
1583
1584		if (found_ctrl) {
1585			if (!schedule_work(&assoc->del_work))
1586				/* already deleting - release local reference */
1587				nvmet_fc_tgt_a_put(assoc);
1588			return;
1589		}
1590
1591		spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1592	}
1593	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1594}
1595
1596/**
1597 * nvme_fc_unregister_targetport - transport entry point called by an
1598 *                              LLDD to deregister/remove a previously
1599 *                              registered a local NVME subsystem FC port.
1600 * @target_port: pointer to the (registered) target port that is to be
1601 *               deregistered.
1602 *
1603 * Returns:
1604 * a completion status. Must be 0 upon success; a negative errno
1605 * (ex: -ENXIO) upon failure.
1606 */
1607int
1608nvmet_fc_unregister_targetport(struct nvmet_fc_target_port *target_port)
1609{
1610	struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
1611
1612	nvmet_fc_portentry_unbind_tgt(tgtport);
1613
1614	/* terminate any outstanding associations */
1615	__nvmet_fc_free_assocs(tgtport);
1616
1617	/*
1618	 * should terminate LS's as well. However, LS's will be generated
1619	 * at the tail end of association termination, so they likely don't
1620	 * exist yet. And even if they did, it's worthwhile to just let
1621	 * them finish and targetport ref counting will clean things up.
1622	 */
1623
1624	nvmet_fc_tgtport_put(tgtport);
1625
1626	return 0;
1627}
1628EXPORT_SYMBOL_GPL(nvmet_fc_unregister_targetport);
1629
1630
1631/* ********************** FC-NVME LS RCV Handling ************************* */
1632
1633
1634static void
1635nvmet_fc_ls_create_association(struct nvmet_fc_tgtport *tgtport,
1636			struct nvmet_fc_ls_iod *iod)
1637{
1638	struct fcnvme_ls_cr_assoc_rqst *rqst = &iod->rqstbuf->rq_cr_assoc;
1639	struct fcnvme_ls_cr_assoc_acc *acc = &iod->rspbuf->rsp_cr_assoc;
1640	struct nvmet_fc_tgt_queue *queue;
1641	int ret = 0;
1642
1643	memset(acc, 0, sizeof(*acc));
1644
1645	/*
1646	 * FC-NVME spec changes. There are initiators sending different
1647	 * lengths as padding sizes for Create Association Cmd descriptor
1648	 * was incorrect.
1649	 * Accept anything of "minimum" length. Assume format per 1.15
1650	 * spec (with HOSTID reduced to 16 bytes), ignore how long the
1651	 * trailing pad length is.
1652	 */
1653	if (iod->rqstdatalen < FCNVME_LSDESC_CRA_RQST_MINLEN)
1654		ret = VERR_CR_ASSOC_LEN;
1655	else if (be32_to_cpu(rqst->desc_list_len) <
1656			FCNVME_LSDESC_CRA_RQST_MIN_LISTLEN)
1657		ret = VERR_CR_ASSOC_RQST_LEN;
1658	else if (rqst->assoc_cmd.desc_tag !=
1659			cpu_to_be32(FCNVME_LSDESC_CREATE_ASSOC_CMD))
1660		ret = VERR_CR_ASSOC_CMD;
1661	else if (be32_to_cpu(rqst->assoc_cmd.desc_len) <
1662			FCNVME_LSDESC_CRA_CMD_DESC_MIN_DESCLEN)
1663		ret = VERR_CR_ASSOC_CMD_LEN;
1664	else if (!rqst->assoc_cmd.ersp_ratio ||
1665		 (be16_to_cpu(rqst->assoc_cmd.ersp_ratio) >=
1666				be16_to_cpu(rqst->assoc_cmd.sqsize)))
1667		ret = VERR_ERSP_RATIO;
1668
1669	else {
1670		/* new association w/ admin queue */
1671		iod->assoc = nvmet_fc_alloc_target_assoc(
1672						tgtport, iod->hosthandle);
1673		if (!iod->assoc)
1674			ret = VERR_ASSOC_ALLOC_FAIL;
1675		else {
1676			queue = nvmet_fc_alloc_target_queue(iod->assoc, 0,
1677					be16_to_cpu(rqst->assoc_cmd.sqsize));
1678			if (!queue) {
1679				ret = VERR_QUEUE_ALLOC_FAIL;
1680				nvmet_fc_tgt_a_put(iod->assoc);
1681			}
1682		}
1683	}
1684
1685	if (ret) {
1686		dev_err(tgtport->dev,
1687			"Create Association LS failed: %s\n",
1688			validation_errors[ret]);
1689		iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1690				sizeof(*acc), rqst->w0.ls_cmd,
1691				FCNVME_RJT_RC_LOGIC,
1692				FCNVME_RJT_EXP_NONE, 0);
1693		return;
1694	}
1695
1696	queue->ersp_ratio = be16_to_cpu(rqst->assoc_cmd.ersp_ratio);
1697	atomic_set(&queue->connected, 1);
1698	queue->sqhd = 0;	/* best place to init value */
1699
1700	dev_info(tgtport->dev,
1701		"{%d:%d} Association created\n",
1702		tgtport->fc_target_port.port_num, iod->assoc->a_id);
1703
1704	/* format a response */
1705
1706	iod->lsrsp->rsplen = sizeof(*acc);
1707
1708	nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1709			fcnvme_lsdesc_len(
1710				sizeof(struct fcnvme_ls_cr_assoc_acc)),
1711			FCNVME_LS_CREATE_ASSOCIATION);
1712	acc->associd.desc_tag = cpu_to_be32(FCNVME_LSDESC_ASSOC_ID);
1713	acc->associd.desc_len =
1714			fcnvme_lsdesc_len(
1715				sizeof(struct fcnvme_lsdesc_assoc_id));
1716	acc->associd.association_id =
1717			cpu_to_be64(nvmet_fc_makeconnid(iod->assoc, 0));
1718	acc->connectid.desc_tag = cpu_to_be32(FCNVME_LSDESC_CONN_ID);
1719	acc->connectid.desc_len =
1720			fcnvme_lsdesc_len(
1721				sizeof(struct fcnvme_lsdesc_conn_id));
1722	acc->connectid.connection_id = acc->associd.association_id;
1723}
1724
1725static void
1726nvmet_fc_ls_create_connection(struct nvmet_fc_tgtport *tgtport,
1727			struct nvmet_fc_ls_iod *iod)
1728{
1729	struct fcnvme_ls_cr_conn_rqst *rqst = &iod->rqstbuf->rq_cr_conn;
1730	struct fcnvme_ls_cr_conn_acc *acc = &iod->rspbuf->rsp_cr_conn;
1731	struct nvmet_fc_tgt_queue *queue;
1732	int ret = 0;
1733
1734	memset(acc, 0, sizeof(*acc));
1735
1736	if (iod->rqstdatalen < sizeof(struct fcnvme_ls_cr_conn_rqst))
1737		ret = VERR_CR_CONN_LEN;
1738	else if (rqst->desc_list_len !=
1739			fcnvme_lsdesc_len(
1740				sizeof(struct fcnvme_ls_cr_conn_rqst)))
1741		ret = VERR_CR_CONN_RQST_LEN;
1742	else if (rqst->associd.desc_tag != cpu_to_be32(FCNVME_LSDESC_ASSOC_ID))
1743		ret = VERR_ASSOC_ID;
1744	else if (rqst->associd.desc_len !=
1745			fcnvme_lsdesc_len(
1746				sizeof(struct fcnvme_lsdesc_assoc_id)))
1747		ret = VERR_ASSOC_ID_LEN;
1748	else if (rqst->connect_cmd.desc_tag !=
1749			cpu_to_be32(FCNVME_LSDESC_CREATE_CONN_CMD))
1750		ret = VERR_CR_CONN_CMD;
1751	else if (rqst->connect_cmd.desc_len !=
1752			fcnvme_lsdesc_len(
1753				sizeof(struct fcnvme_lsdesc_cr_conn_cmd)))
1754		ret = VERR_CR_CONN_CMD_LEN;
1755	else if (!rqst->connect_cmd.ersp_ratio ||
1756		 (be16_to_cpu(rqst->connect_cmd.ersp_ratio) >=
1757				be16_to_cpu(rqst->connect_cmd.sqsize)))
1758		ret = VERR_ERSP_RATIO;
1759
1760	else {
1761		/* new io queue */
1762		iod->assoc = nvmet_fc_find_target_assoc(tgtport,
1763				be64_to_cpu(rqst->associd.association_id));
1764		if (!iod->assoc)
1765			ret = VERR_NO_ASSOC;
1766		else {
1767			queue = nvmet_fc_alloc_target_queue(iod->assoc,
1768					be16_to_cpu(rqst->connect_cmd.qid),
1769					be16_to_cpu(rqst->connect_cmd.sqsize));
1770			if (!queue)
1771				ret = VERR_QUEUE_ALLOC_FAIL;
1772
1773			/* release get taken in nvmet_fc_find_target_assoc */
1774			nvmet_fc_tgt_a_put(iod->assoc);
1775		}
1776	}
1777
1778	if (ret) {
1779		dev_err(tgtport->dev,
1780			"Create Connection LS failed: %s\n",
1781			validation_errors[ret]);
1782		iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1783				sizeof(*acc), rqst->w0.ls_cmd,
1784				(ret == VERR_NO_ASSOC) ?
1785					FCNVME_RJT_RC_INV_ASSOC :
1786					FCNVME_RJT_RC_LOGIC,
1787				FCNVME_RJT_EXP_NONE, 0);
1788		return;
1789	}
1790
1791	queue->ersp_ratio = be16_to_cpu(rqst->connect_cmd.ersp_ratio);
1792	atomic_set(&queue->connected, 1);
1793	queue->sqhd = 0;	/* best place to init value */
1794
1795	/* format a response */
1796
1797	iod->lsrsp->rsplen = sizeof(*acc);
1798
1799	nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1800			fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_cr_conn_acc)),
1801			FCNVME_LS_CREATE_CONNECTION);
1802	acc->connectid.desc_tag = cpu_to_be32(FCNVME_LSDESC_CONN_ID);
1803	acc->connectid.desc_len =
1804			fcnvme_lsdesc_len(
1805				sizeof(struct fcnvme_lsdesc_conn_id));
1806	acc->connectid.connection_id =
1807			cpu_to_be64(nvmet_fc_makeconnid(iod->assoc,
1808				be16_to_cpu(rqst->connect_cmd.qid)));
1809}
1810
1811/*
1812 * Returns true if the LS response is to be transmit
1813 * Returns false if the LS response is to be delayed
1814 */
1815static int
1816nvmet_fc_ls_disconnect(struct nvmet_fc_tgtport *tgtport,
1817			struct nvmet_fc_ls_iod *iod)
1818{
1819	struct fcnvme_ls_disconnect_assoc_rqst *rqst =
1820						&iod->rqstbuf->rq_dis_assoc;
1821	struct fcnvme_ls_disconnect_assoc_acc *acc =
1822						&iod->rspbuf->rsp_dis_assoc;
1823	struct nvmet_fc_tgt_assoc *assoc = NULL;
1824	struct nvmet_fc_ls_iod *oldls = NULL;
1825	unsigned long flags;
1826	int ret = 0;
1827
1828	memset(acc, 0, sizeof(*acc));
1829
1830	ret = nvmefc_vldt_lsreq_discon_assoc(iod->rqstdatalen, rqst);
1831	if (!ret) {
1832		/* match an active association - takes an assoc ref if !NULL */
1833		assoc = nvmet_fc_find_target_assoc(tgtport,
1834				be64_to_cpu(rqst->associd.association_id));
1835		iod->assoc = assoc;
1836		if (!assoc)
1837			ret = VERR_NO_ASSOC;
1838	}
1839
1840	if (ret || !assoc) {
1841		dev_err(tgtport->dev,
1842			"Disconnect LS failed: %s\n",
1843			validation_errors[ret]);
1844		iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1845				sizeof(*acc), rqst->w0.ls_cmd,
1846				(ret == VERR_NO_ASSOC) ?
1847					FCNVME_RJT_RC_INV_ASSOC :
1848					FCNVME_RJT_RC_LOGIC,
1849				FCNVME_RJT_EXP_NONE, 0);
1850		return true;
1851	}
1852
1853	/* format a response */
1854
1855	iod->lsrsp->rsplen = sizeof(*acc);
1856
1857	nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1858			fcnvme_lsdesc_len(
1859				sizeof(struct fcnvme_ls_disconnect_assoc_acc)),
1860			FCNVME_LS_DISCONNECT_ASSOC);
1861
1862	/* release get taken in nvmet_fc_find_target_assoc */
1863	nvmet_fc_tgt_a_put(assoc);
1864
1865	/*
1866	 * The rules for LS response says the response cannot
1867	 * go back until ABTS's have been sent for all outstanding
1868	 * I/O and a Disconnect Association LS has been sent.
1869	 * So... save off the Disconnect LS to send the response
1870	 * later. If there was a prior LS already saved, replace
1871	 * it with the newer one and send a can't perform reject
1872	 * on the older one.
1873	 */
1874	spin_lock_irqsave(&tgtport->lock, flags);
1875	oldls = assoc->rcv_disconn;
1876	assoc->rcv_disconn = iod;
1877	spin_unlock_irqrestore(&tgtport->lock, flags);
1878
1879	nvmet_fc_delete_target_assoc(assoc);
1880
1881	if (oldls) {
1882		dev_info(tgtport->dev,
1883			"{%d:%d} Multiple Disconnect Association LS's "
1884			"received\n",
1885			tgtport->fc_target_port.port_num, assoc->a_id);
1886		/* overwrite good response with bogus failure */
1887		oldls->lsrsp->rsplen = nvme_fc_format_rjt(oldls->rspbuf,
1888						sizeof(*iod->rspbuf),
1889						/* ok to use rqst, LS is same */
1890						rqst->w0.ls_cmd,
1891						FCNVME_RJT_RC_UNAB,
1892						FCNVME_RJT_EXP_NONE, 0);
1893		nvmet_fc_xmt_ls_rsp(tgtport, oldls);
1894	}
1895
1896	return false;
1897}
1898
1899
1900/* *********************** NVME Ctrl Routines **************************** */
1901
1902
1903static void nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req);
1904
1905static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops;
1906
1907static void
1908nvmet_fc_xmt_ls_rsp_done(struct nvmefc_ls_rsp *lsrsp)
1909{
1910	struct nvmet_fc_ls_iod *iod = lsrsp->nvme_fc_private;
1911	struct nvmet_fc_tgtport *tgtport = iod->tgtport;
1912
1913	fc_dma_sync_single_for_cpu(tgtport->dev, iod->rspdma,
1914				sizeof(*iod->rspbuf), DMA_TO_DEVICE);
1915	nvmet_fc_free_ls_iod(tgtport, iod);
1916	nvmet_fc_tgtport_put(tgtport);
1917}
1918
1919static void
1920nvmet_fc_xmt_ls_rsp(struct nvmet_fc_tgtport *tgtport,
1921				struct nvmet_fc_ls_iod *iod)
1922{
1923	int ret;
1924
1925	fc_dma_sync_single_for_device(tgtport->dev, iod->rspdma,
1926				  sizeof(*iod->rspbuf), DMA_TO_DEVICE);
1927
1928	ret = tgtport->ops->xmt_ls_rsp(&tgtport->fc_target_port, iod->lsrsp);
1929	if (ret)
1930		nvmet_fc_xmt_ls_rsp_done(iod->lsrsp);
1931}
1932
1933/*
1934 * Actual processing routine for received FC-NVME LS Requests from the LLD
1935 */
1936static void
1937nvmet_fc_handle_ls_rqst(struct nvmet_fc_tgtport *tgtport,
1938			struct nvmet_fc_ls_iod *iod)
1939{
1940	struct fcnvme_ls_rqst_w0 *w0 = &iod->rqstbuf->rq_cr_assoc.w0;
1941	bool sendrsp = true;
1942
1943	iod->lsrsp->nvme_fc_private = iod;
1944	iod->lsrsp->rspbuf = iod->rspbuf;
1945	iod->lsrsp->rspdma = iod->rspdma;
1946	iod->lsrsp->done = nvmet_fc_xmt_ls_rsp_done;
1947	/* Be preventative. handlers will later set to valid length */
1948	iod->lsrsp->rsplen = 0;
1949
1950	iod->assoc = NULL;
1951
1952	/*
1953	 * handlers:
1954	 *   parse request input, execute the request, and format the
1955	 *   LS response
1956	 */
1957	switch (w0->ls_cmd) {
1958	case FCNVME_LS_CREATE_ASSOCIATION:
1959		/* Creates Association and initial Admin Queue/Connection */
1960		nvmet_fc_ls_create_association(tgtport, iod);
1961		break;
1962	case FCNVME_LS_CREATE_CONNECTION:
1963		/* Creates an IO Queue/Connection */
1964		nvmet_fc_ls_create_connection(tgtport, iod);
1965		break;
1966	case FCNVME_LS_DISCONNECT_ASSOC:
1967		/* Terminate a Queue/Connection or the Association */
1968		sendrsp = nvmet_fc_ls_disconnect(tgtport, iod);
1969		break;
1970	default:
1971		iod->lsrsp->rsplen = nvme_fc_format_rjt(iod->rspbuf,
1972				sizeof(*iod->rspbuf), w0->ls_cmd,
1973				FCNVME_RJT_RC_INVAL, FCNVME_RJT_EXP_NONE, 0);
1974	}
1975
1976	if (sendrsp)
1977		nvmet_fc_xmt_ls_rsp(tgtport, iod);
1978}
1979
1980/*
1981 * Actual processing routine for received FC-NVME LS Requests from the LLD
1982 */
1983static void
1984nvmet_fc_handle_ls_rqst_work(struct work_struct *work)
1985{
1986	struct nvmet_fc_ls_iod *iod =
1987		container_of(work, struct nvmet_fc_ls_iod, work);
1988	struct nvmet_fc_tgtport *tgtport = iod->tgtport;
1989
1990	nvmet_fc_handle_ls_rqst(tgtport, iod);
1991}
1992
1993
1994/**
1995 * nvmet_fc_rcv_ls_req - transport entry point called by an LLDD
1996 *                       upon the reception of a NVME LS request.
1997 *
1998 * The nvmet-fc layer will copy payload to an internal structure for
1999 * processing.  As such, upon completion of the routine, the LLDD may
2000 * immediately free/reuse the LS request buffer passed in the call.
2001 *
2002 * If this routine returns error, the LLDD should abort the exchange.
2003 *
2004 * @target_port: pointer to the (registered) target port the LS was
2005 *              received on.
2006 * @lsrsp:      pointer to a lsrsp structure to be used to reference
2007 *              the exchange corresponding to the LS.
2008 * @lsreqbuf:   pointer to the buffer containing the LS Request
2009 * @lsreqbuf_len: length, in bytes, of the received LS request
2010 */
2011int
2012nvmet_fc_rcv_ls_req(struct nvmet_fc_target_port *target_port,
2013			void *hosthandle,
2014			struct nvmefc_ls_rsp *lsrsp,
2015			void *lsreqbuf, u32 lsreqbuf_len)
2016{
2017	struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
2018	struct nvmet_fc_ls_iod *iod;
2019	struct fcnvme_ls_rqst_w0 *w0 = (struct fcnvme_ls_rqst_w0 *)lsreqbuf;
2020
2021	if (lsreqbuf_len > sizeof(union nvmefc_ls_requests)) {
2022		dev_info(tgtport->dev,
2023			"RCV %s LS failed: payload too large (%d)\n",
2024			(w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2025				nvmefc_ls_names[w0->ls_cmd] : "",
2026			lsreqbuf_len);
2027		return -E2BIG;
2028	}
2029
2030	if (!nvmet_fc_tgtport_get(tgtport)) {
2031		dev_info(tgtport->dev,
2032			"RCV %s LS failed: target deleting\n",
2033			(w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2034				nvmefc_ls_names[w0->ls_cmd] : "");
2035		return -ESHUTDOWN;
2036	}
2037
2038	iod = nvmet_fc_alloc_ls_iod(tgtport);
2039	if (!iod) {
2040		dev_info(tgtport->dev,
2041			"RCV %s LS failed: context allocation failed\n",
2042			(w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2043				nvmefc_ls_names[w0->ls_cmd] : "");
2044		nvmet_fc_tgtport_put(tgtport);
2045		return -ENOENT;
2046	}
2047
2048	iod->lsrsp = lsrsp;
2049	iod->fcpreq = NULL;
2050	memcpy(iod->rqstbuf, lsreqbuf, lsreqbuf_len);
2051	iod->rqstdatalen = lsreqbuf_len;
2052	iod->hosthandle = hosthandle;
2053
2054	schedule_work(&iod->work);
2055
2056	return 0;
2057}
2058EXPORT_SYMBOL_GPL(nvmet_fc_rcv_ls_req);
2059
2060
2061/*
2062 * **********************
2063 * Start of FCP handling
2064 * **********************
2065 */
2066
2067static int
2068nvmet_fc_alloc_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
2069{
2070	struct scatterlist *sg;
2071	unsigned int nent;
2072
2073	sg = sgl_alloc(fod->req.transfer_len, GFP_KERNEL, &nent);
2074	if (!sg)
2075		goto out;
2076
2077	fod->data_sg = sg;
2078	fod->data_sg_cnt = nent;
2079	fod->data_sg_cnt = fc_dma_map_sg(fod->tgtport->dev, sg, nent,
2080				((fod->io_dir == NVMET_FCP_WRITE) ?
2081					DMA_FROM_DEVICE : DMA_TO_DEVICE));
2082				/* note: write from initiator perspective */
2083	fod->next_sg = fod->data_sg;
2084
2085	return 0;
2086
2087out:
2088	return NVME_SC_INTERNAL;
2089}
2090
2091static void
2092nvmet_fc_free_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
2093{
2094	if (!fod->data_sg || !fod->data_sg_cnt)
2095		return;
2096
2097	fc_dma_unmap_sg(fod->tgtport->dev, fod->data_sg, fod->data_sg_cnt,
2098				((fod->io_dir == NVMET_FCP_WRITE) ?
2099					DMA_FROM_DEVICE : DMA_TO_DEVICE));
2100	sgl_free(fod->data_sg);
2101	fod->data_sg = NULL;
2102	fod->data_sg_cnt = 0;
2103}
2104
2105
2106static bool
2107queue_90percent_full(struct nvmet_fc_tgt_queue *q, u32 sqhd)
2108{
2109	u32 sqtail, used;
2110
2111	/* egad, this is ugly. And sqtail is just a best guess */
2112	sqtail = atomic_read(&q->sqtail) % q->sqsize;
2113
2114	used = (sqtail < sqhd) ? (sqtail + q->sqsize - sqhd) : (sqtail - sqhd);
2115	return ((used * 10) >= (((u32)(q->sqsize - 1) * 9)));
2116}
2117
2118/*
2119 * Prep RSP payload.
2120 * May be a NVMET_FCOP_RSP or NVMET_FCOP_READDATA_RSP op
2121 */
2122static void
2123nvmet_fc_prep_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
2124				struct nvmet_fc_fcp_iod *fod)
2125{
2126	struct nvme_fc_ersp_iu *ersp = &fod->rspiubuf;
2127	struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
2128	struct nvme_completion *cqe = &ersp->cqe;
2129	u32 *cqewd = (u32 *)cqe;
2130	bool send_ersp = false;
2131	u32 rsn, rspcnt, xfr_length;
2132
2133	if (fod->fcpreq->op == NVMET_FCOP_READDATA_RSP)
2134		xfr_length = fod->req.transfer_len;
2135	else
2136		xfr_length = fod->offset;
2137
2138	/*
2139	 * check to see if we can send a 0's rsp.
2140	 *   Note: to send a 0's response, the NVME-FC host transport will
2141	 *   recreate the CQE. The host transport knows: sq id, SQHD (last
2142	 *   seen in an ersp), and command_id. Thus it will create a
2143	 *   zero-filled CQE with those known fields filled in. Transport
2144	 *   must send an ersp for any condition where the cqe won't match
2145	 *   this.
2146	 *
2147	 * Here are the FC-NVME mandated cases where we must send an ersp:
2148	 *  every N responses, where N=ersp_ratio
2149	 *  force fabric commands to send ersp's (not in FC-NVME but good
2150	 *    practice)
2151	 *  normal cmds: any time status is non-zero, or status is zero
2152	 *     but words 0 or 1 are non-zero.
2153	 *  the SQ is 90% or more full
2154	 *  the cmd is a fused command
2155	 *  transferred data length not equal to cmd iu length
2156	 */
2157	rspcnt = atomic_inc_return(&fod->queue->zrspcnt);
2158	if (!(rspcnt % fod->queue->ersp_ratio) ||
2159	    nvme_is_fabrics((struct nvme_command *) sqe) ||
2160	    xfr_length != fod->req.transfer_len ||
2161	    (le16_to_cpu(cqe->status) & 0xFFFE) || cqewd[0] || cqewd[1] ||
2162	    (sqe->flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND)) ||
2163	    queue_90percent_full(fod->queue, le16_to_cpu(cqe->sq_head)))
2164		send_ersp = true;
2165
2166	/* re-set the fields */
2167	fod->fcpreq->rspaddr = ersp;
2168	fod->fcpreq->rspdma = fod->rspdma;
2169
2170	if (!send_ersp) {
2171		memset(ersp, 0, NVME_FC_SIZEOF_ZEROS_RSP);
2172		fod->fcpreq->rsplen = NVME_FC_SIZEOF_ZEROS_RSP;
2173	} else {
2174		ersp->iu_len = cpu_to_be16(sizeof(*ersp)/sizeof(u32));
2175		rsn = atomic_inc_return(&fod->queue->rsn);
2176		ersp->rsn = cpu_to_be32(rsn);
2177		ersp->xfrd_len = cpu_to_be32(xfr_length);
2178		fod->fcpreq->rsplen = sizeof(*ersp);
2179	}
2180
2181	fc_dma_sync_single_for_device(tgtport->dev, fod->rspdma,
2182				  sizeof(fod->rspiubuf), DMA_TO_DEVICE);
2183}
2184
2185static void nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq);
2186
2187static void
2188nvmet_fc_abort_op(struct nvmet_fc_tgtport *tgtport,
2189				struct nvmet_fc_fcp_iod *fod)
2190{
2191	struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2192
2193	/* data no longer needed */
2194	nvmet_fc_free_tgt_pgs(fod);
2195
2196	/*
2197	 * if an ABTS was received or we issued the fcp_abort early
2198	 * don't call abort routine again.
2199	 */
2200	/* no need to take lock - lock was taken earlier to get here */
2201	if (!fod->aborted)
2202		tgtport->ops->fcp_abort(&tgtport->fc_target_port, fcpreq);
2203
2204	nvmet_fc_free_fcp_iod(fod->queue, fod);
2205}
2206
2207static void
2208nvmet_fc_xmt_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
2209				struct nvmet_fc_fcp_iod *fod)
2210{
2211	int ret;
2212
2213	fod->fcpreq->op = NVMET_FCOP_RSP;
2214	fod->fcpreq->timeout = 0;
2215
2216	nvmet_fc_prep_fcp_rsp(tgtport, fod);
2217
2218	ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
2219	if (ret)
2220		nvmet_fc_abort_op(tgtport, fod);
2221}
2222
2223static void
2224nvmet_fc_transfer_fcp_data(struct nvmet_fc_tgtport *tgtport,
2225				struct nvmet_fc_fcp_iod *fod, u8 op)
2226{
2227	struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2228	struct scatterlist *sg = fod->next_sg;
2229	unsigned long flags;
2230	u32 remaininglen = fod->req.transfer_len - fod->offset;
2231	u32 tlen = 0;
2232	int ret;
2233
2234	fcpreq->op = op;
2235	fcpreq->offset = fod->offset;
2236	fcpreq->timeout = NVME_FC_TGTOP_TIMEOUT_SEC;
2237
2238	/*
2239	 * for next sequence:
2240	 *  break at a sg element boundary
2241	 *  attempt to keep sequence length capped at
2242	 *    NVMET_FC_MAX_SEQ_LENGTH but allow sequence to
2243	 *    be longer if a single sg element is larger
2244	 *    than that amount. This is done to avoid creating
2245	 *    a new sg list to use for the tgtport api.
2246	 */
2247	fcpreq->sg = sg;
2248	fcpreq->sg_cnt = 0;
2249	while (tlen < remaininglen &&
2250	       fcpreq->sg_cnt < tgtport->max_sg_cnt &&
2251	       tlen + sg_dma_len(sg) < NVMET_FC_MAX_SEQ_LENGTH) {
2252		fcpreq->sg_cnt++;
2253		tlen += sg_dma_len(sg);
2254		sg = sg_next(sg);
2255	}
2256	if (tlen < remaininglen && fcpreq->sg_cnt == 0) {
2257		fcpreq->sg_cnt++;
2258		tlen += min_t(u32, sg_dma_len(sg), remaininglen);
2259		sg = sg_next(sg);
2260	}
2261	if (tlen < remaininglen)
2262		fod->next_sg = sg;
2263	else
2264		fod->next_sg = NULL;
2265
2266	fcpreq->transfer_length = tlen;
2267	fcpreq->transferred_length = 0;
2268	fcpreq->fcp_error = 0;
2269	fcpreq->rsplen = 0;
2270
2271	/*
2272	 * If the last READDATA request: check if LLDD supports
2273	 * combined xfr with response.
2274	 */
2275	if ((op == NVMET_FCOP_READDATA) &&
2276	    ((fod->offset + fcpreq->transfer_length) == fod->req.transfer_len) &&
2277	    (tgtport->ops->target_features & NVMET_FCTGTFEAT_READDATA_RSP)) {
2278		fcpreq->op = NVMET_FCOP_READDATA_RSP;
2279		nvmet_fc_prep_fcp_rsp(tgtport, fod);
2280	}
2281
2282	ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
2283	if (ret) {
2284		/*
2285		 * should be ok to set w/o lock as its in the thread of
2286		 * execution (not an async timer routine) and doesn't
2287		 * contend with any clearing action
2288		 */
2289		fod->abort = true;
2290
2291		if (op == NVMET_FCOP_WRITEDATA) {
2292			spin_lock_irqsave(&fod->flock, flags);
2293			fod->writedataactive = false;
2294			spin_unlock_irqrestore(&fod->flock, flags);
2295			nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2296		} else /* NVMET_FCOP_READDATA or NVMET_FCOP_READDATA_RSP */ {
2297			fcpreq->fcp_error = ret;
2298			fcpreq->transferred_length = 0;
2299			nvmet_fc_xmt_fcp_op_done(fod->fcpreq);
2300		}
2301	}
2302}
2303
2304static inline bool
2305__nvmet_fc_fod_op_abort(struct nvmet_fc_fcp_iod *fod, bool abort)
2306{
2307	struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2308	struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2309
2310	/* if in the middle of an io and we need to tear down */
2311	if (abort) {
2312		if (fcpreq->op == NVMET_FCOP_WRITEDATA) {
2313			nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2314			return true;
2315		}
2316
2317		nvmet_fc_abort_op(tgtport, fod);
2318		return true;
2319	}
2320
2321	return false;
2322}
2323
2324/*
2325 * actual done handler for FCP operations when completed by the lldd
2326 */
2327static void
2328nvmet_fc_fod_op_done(struct nvmet_fc_fcp_iod *fod)
2329{
2330	struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2331	struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2332	unsigned long flags;
2333	bool abort;
2334
2335	spin_lock_irqsave(&fod->flock, flags);
2336	abort = fod->abort;
2337	fod->writedataactive = false;
2338	spin_unlock_irqrestore(&fod->flock, flags);
2339
2340	switch (fcpreq->op) {
2341
2342	case NVMET_FCOP_WRITEDATA:
2343		if (__nvmet_fc_fod_op_abort(fod, abort))
2344			return;
2345		if (fcpreq->fcp_error ||
2346		    fcpreq->transferred_length != fcpreq->transfer_length) {
2347			spin_lock_irqsave(&fod->flock, flags);
2348			fod->abort = true;
2349			spin_unlock_irqrestore(&fod->flock, flags);
2350
2351			nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2352			return;
2353		}
2354
2355		fod->offset += fcpreq->transferred_length;
2356		if (fod->offset != fod->req.transfer_len) {
2357			spin_lock_irqsave(&fod->flock, flags);
2358			fod->writedataactive = true;
2359			spin_unlock_irqrestore(&fod->flock, flags);
2360
2361			/* transfer the next chunk */
2362			nvmet_fc_transfer_fcp_data(tgtport, fod,
2363						NVMET_FCOP_WRITEDATA);
2364			return;
2365		}
2366
2367		/* data transfer complete, resume with nvmet layer */
2368		fod->req.execute(&fod->req);
2369		break;
2370
2371	case NVMET_FCOP_READDATA:
2372	case NVMET_FCOP_READDATA_RSP:
2373		if (__nvmet_fc_fod_op_abort(fod, abort))
2374			return;
2375		if (fcpreq->fcp_error ||
2376		    fcpreq->transferred_length != fcpreq->transfer_length) {
2377			nvmet_fc_abort_op(tgtport, fod);
2378			return;
2379		}
2380
2381		/* success */
2382
2383		if (fcpreq->op == NVMET_FCOP_READDATA_RSP) {
2384			/* data no longer needed */
2385			nvmet_fc_free_tgt_pgs(fod);
2386			nvmet_fc_free_fcp_iod(fod->queue, fod);
2387			return;
2388		}
2389
2390		fod->offset += fcpreq->transferred_length;
2391		if (fod->offset != fod->req.transfer_len) {
2392			/* transfer the next chunk */
2393			nvmet_fc_transfer_fcp_data(tgtport, fod,
2394						NVMET_FCOP_READDATA);
2395			return;
2396		}
2397
2398		/* data transfer complete, send response */
2399
2400		/* data no longer needed */
2401		nvmet_fc_free_tgt_pgs(fod);
2402
2403		nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2404
2405		break;
2406
2407	case NVMET_FCOP_RSP:
2408		if (__nvmet_fc_fod_op_abort(fod, abort))
2409			return;
2410		nvmet_fc_free_fcp_iod(fod->queue, fod);
2411		break;
2412
2413	default:
2414		break;
2415	}
2416}
2417
2418static void
2419nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq)
2420{
2421	struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2422
2423	nvmet_fc_fod_op_done(fod);
2424}
2425
2426/*
2427 * actual completion handler after execution by the nvmet layer
2428 */
2429static void
2430__nvmet_fc_fcp_nvme_cmd_done(struct nvmet_fc_tgtport *tgtport,
2431			struct nvmet_fc_fcp_iod *fod, int status)
2432{
2433	struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
2434	struct nvme_completion *cqe = &fod->rspiubuf.cqe;
2435	unsigned long flags;
2436	bool abort;
2437
2438	spin_lock_irqsave(&fod->flock, flags);
2439	abort = fod->abort;
2440	spin_unlock_irqrestore(&fod->flock, flags);
2441
2442	/* if we have a CQE, snoop the last sq_head value */
2443	if (!status)
2444		fod->queue->sqhd = cqe->sq_head;
2445
2446	if (abort) {
2447		nvmet_fc_abort_op(tgtport, fod);
2448		return;
2449	}
2450
2451	/* if an error handling the cmd post initial parsing */
2452	if (status) {
2453		/* fudge up a failed CQE status for our transport error */
2454		memset(cqe, 0, sizeof(*cqe));
2455		cqe->sq_head = fod->queue->sqhd;	/* echo last cqe sqhd */
2456		cqe->sq_id = cpu_to_le16(fod->queue->qid);
2457		cqe->command_id = sqe->command_id;
2458		cqe->status = cpu_to_le16(status);
2459	} else {
2460
2461		/*
2462		 * try to push the data even if the SQE status is non-zero.
2463		 * There may be a status where data still was intended to
2464		 * be moved
2465		 */
2466		if ((fod->io_dir == NVMET_FCP_READ) && (fod->data_sg_cnt)) {
2467			/* push the data over before sending rsp */
2468			nvmet_fc_transfer_fcp_data(tgtport, fod,
2469						NVMET_FCOP_READDATA);
2470			return;
2471		}
2472
2473		/* writes & no data - fall thru */
2474	}
2475
2476	/* data no longer needed */
2477	nvmet_fc_free_tgt_pgs(fod);
2478
2479	nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2480}
2481
2482
2483static void
2484nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req)
2485{
2486	struct nvmet_fc_fcp_iod *fod = nvmet_req_to_fod(nvme_req);
2487	struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2488
2489	__nvmet_fc_fcp_nvme_cmd_done(tgtport, fod, 0);
2490}
2491
2492
2493/*
2494 * Actual processing routine for received FC-NVME I/O Requests from the LLD
2495 */
2496static void
2497nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
2498			struct nvmet_fc_fcp_iod *fod)
2499{
2500	struct nvme_fc_cmd_iu *cmdiu = &fod->cmdiubuf;
2501	u32 xfrlen = be32_to_cpu(cmdiu->data_len);
2502	int ret;
2503
2504	/*
2505	 * Fused commands are currently not supported in the linux
2506	 * implementation.
2507	 *
2508	 * As such, the implementation of the FC transport does not
2509	 * look at the fused commands and order delivery to the upper
2510	 * layer until we have both based on csn.
2511	 */
2512
2513	fod->fcpreq->done = nvmet_fc_xmt_fcp_op_done;
2514
2515	if (cmdiu->flags & FCNVME_CMD_FLAGS_WRITE) {
2516		fod->io_dir = NVMET_FCP_WRITE;
2517		if (!nvme_is_write(&cmdiu->sqe))
2518			goto transport_error;
2519	} else if (cmdiu->flags & FCNVME_CMD_FLAGS_READ) {
2520		fod->io_dir = NVMET_FCP_READ;
2521		if (nvme_is_write(&cmdiu->sqe))
2522			goto transport_error;
2523	} else {
2524		fod->io_dir = NVMET_FCP_NODATA;
2525		if (xfrlen)
2526			goto transport_error;
2527	}
2528
2529	fod->req.cmd = &fod->cmdiubuf.sqe;
2530	fod->req.cqe = &fod->rspiubuf.cqe;
2531	if (tgtport->pe)
2532		fod->req.port = tgtport->pe->port;
2533
2534	/* clear any response payload */
2535	memset(&fod->rspiubuf, 0, sizeof(fod->rspiubuf));
2536
2537	fod->data_sg = NULL;
2538	fod->data_sg_cnt = 0;
2539
2540	ret = nvmet_req_init(&fod->req,
2541				&fod->queue->nvme_cq,
2542				&fod->queue->nvme_sq,
2543				&nvmet_fc_tgt_fcp_ops);
2544	if (!ret) {
2545		/* bad SQE content or invalid ctrl state */
2546		/* nvmet layer has already called op done to send rsp. */
2547		return;
2548	}
2549
2550	fod->req.transfer_len = xfrlen;
2551
2552	/* keep a running counter of tail position */
2553	atomic_inc(&fod->queue->sqtail);
2554
2555	if (fod->req.transfer_len) {
2556		ret = nvmet_fc_alloc_tgt_pgs(fod);
2557		if (ret) {
2558			nvmet_req_complete(&fod->req, ret);
2559			return;
2560		}
2561	}
2562	fod->req.sg = fod->data_sg;
2563	fod->req.sg_cnt = fod->data_sg_cnt;
2564	fod->offset = 0;
2565
2566	if (fod->io_dir == NVMET_FCP_WRITE) {
2567		/* pull the data over before invoking nvmet layer */
2568		nvmet_fc_transfer_fcp_data(tgtport, fod, NVMET_FCOP_WRITEDATA);
2569		return;
2570	}
2571
2572	/*
2573	 * Reads or no data:
2574	 *
2575	 * can invoke the nvmet_layer now. If read data, cmd completion will
2576	 * push the data
2577	 */
2578	fod->req.execute(&fod->req);
2579	return;
2580
2581transport_error:
2582	nvmet_fc_abort_op(tgtport, fod);
2583}
2584
2585/**
2586 * nvmet_fc_rcv_fcp_req - transport entry point called by an LLDD
2587 *                       upon the reception of a NVME FCP CMD IU.
2588 *
2589 * Pass a FC-NVME FCP CMD IU received from the FC link to the nvmet-fc
2590 * layer for processing.
2591 *
2592 * The nvmet_fc layer allocates a local job structure (struct
2593 * nvmet_fc_fcp_iod) from the queue for the io and copies the
2594 * CMD IU buffer to the job structure. As such, on a successful
2595 * completion (returns 0), the LLDD may immediately free/reuse
2596 * the CMD IU buffer passed in the call.
2597 *
2598 * However, in some circumstances, due to the packetized nature of FC
2599 * and the api of the FC LLDD which may issue a hw command to send the
2600 * response, but the LLDD may not get the hw completion for that command
2601 * and upcall the nvmet_fc layer before a new command may be
2602 * asynchronously received - its possible for a command to be received
2603 * before the LLDD and nvmet_fc have recycled the job structure. It gives
2604 * the appearance of more commands received than fits in the sq.
2605 * To alleviate this scenario, a temporary queue is maintained in the
2606 * transport for pending LLDD requests waiting for a queue job structure.
2607 * In these "overrun" cases, a temporary queue element is allocated
2608 * the LLDD request and CMD iu buffer information remembered, and the
2609 * routine returns a -EOVERFLOW status. Subsequently, when a queue job
2610 * structure is freed, it is immediately reallocated for anything on the
2611 * pending request list. The LLDDs defer_rcv() callback is called,
2612 * informing the LLDD that it may reuse the CMD IU buffer, and the io
2613 * is then started normally with the transport.
2614 *
2615 * The LLDD, when receiving an -EOVERFLOW completion status, is to treat
2616 * the completion as successful but must not reuse the CMD IU buffer
2617 * until the LLDD's defer_rcv() callback has been called for the
2618 * corresponding struct nvmefc_tgt_fcp_req pointer.
2619 *
2620 * If there is any other condition in which an error occurs, the
2621 * transport will return a non-zero status indicating the error.
2622 * In all cases other than -EOVERFLOW, the transport has not accepted the
2623 * request and the LLDD should abort the exchange.
2624 *
2625 * @target_port: pointer to the (registered) target port the FCP CMD IU
2626 *              was received on.
2627 * @fcpreq:     pointer to a fcpreq request structure to be used to reference
2628 *              the exchange corresponding to the FCP Exchange.
2629 * @cmdiubuf:   pointer to the buffer containing the FCP CMD IU
2630 * @cmdiubuf_len: length, in bytes, of the received FCP CMD IU
2631 */
2632int
2633nvmet_fc_rcv_fcp_req(struct nvmet_fc_target_port *target_port,
2634			struct nvmefc_tgt_fcp_req *fcpreq,
2635			void *cmdiubuf, u32 cmdiubuf_len)
2636{
2637	struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
2638	struct nvme_fc_cmd_iu *cmdiu = cmdiubuf;
2639	struct nvmet_fc_tgt_queue *queue;
2640	struct nvmet_fc_fcp_iod *fod;
2641	struct nvmet_fc_defer_fcp_req *deferfcp;
2642	unsigned long flags;
2643
2644	/* validate iu, so the connection id can be used to find the queue */
2645	if ((cmdiubuf_len != sizeof(*cmdiu)) ||
2646			(cmdiu->format_id != NVME_CMD_FORMAT_ID) ||
2647			(cmdiu->fc_id != NVME_CMD_FC_ID) ||
2648			(be16_to_cpu(cmdiu->iu_len) != (sizeof(*cmdiu)/4)))
2649		return -EIO;
2650
2651	queue = nvmet_fc_find_target_queue(tgtport,
2652				be64_to_cpu(cmdiu->connection_id));
2653	if (!queue)
2654		return -ENOTCONN;
2655
2656	/*
2657	 * note: reference taken by find_target_queue
2658	 * After successful fod allocation, the fod will inherit the
2659	 * ownership of that reference and will remove the reference
2660	 * when the fod is freed.
2661	 */
2662
2663	spin_lock_irqsave(&queue->qlock, flags);
2664
2665	fod = nvmet_fc_alloc_fcp_iod(queue);
2666	if (fod) {
2667		spin_unlock_irqrestore(&queue->qlock, flags);
2668
2669		fcpreq->nvmet_fc_private = fod;
2670		fod->fcpreq = fcpreq;
2671
2672		memcpy(&fod->cmdiubuf, cmdiubuf, cmdiubuf_len);
2673
2674		nvmet_fc_queue_fcp_req(tgtport, queue, fcpreq);
2675
2676		return 0;
2677	}
2678
2679	if (!tgtport->ops->defer_rcv) {
2680		spin_unlock_irqrestore(&queue->qlock, flags);
2681		/* release the queue lookup reference */
2682		nvmet_fc_tgt_q_put(queue);
2683		return -ENOENT;
2684	}
2685
2686	deferfcp = list_first_entry_or_null(&queue->avail_defer_list,
2687			struct nvmet_fc_defer_fcp_req, req_list);
2688	if (deferfcp) {
2689		/* Just re-use one that was previously allocated */
2690		list_del(&deferfcp->req_list);
2691	} else {
2692		spin_unlock_irqrestore(&queue->qlock, flags);
2693
2694		/* Now we need to dynamically allocate one */
2695		deferfcp = kmalloc(sizeof(*deferfcp), GFP_KERNEL);
2696		if (!deferfcp) {
2697			/* release the queue lookup reference */
2698			nvmet_fc_tgt_q_put(queue);
2699			return -ENOMEM;
2700		}
2701		spin_lock_irqsave(&queue->qlock, flags);
2702	}
2703
2704	/* For now, use rspaddr / rsplen to save payload information */
2705	fcpreq->rspaddr = cmdiubuf;
2706	fcpreq->rsplen  = cmdiubuf_len;
2707	deferfcp->fcp_req = fcpreq;
2708
2709	/* defer processing till a fod becomes available */
2710	list_add_tail(&deferfcp->req_list, &queue->pending_cmd_list);
2711
2712	/* NOTE: the queue lookup reference is still valid */
2713
2714	spin_unlock_irqrestore(&queue->qlock, flags);
2715
2716	return -EOVERFLOW;
2717}
2718EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_req);
2719
2720/**
2721 * nvmet_fc_rcv_fcp_abort - transport entry point called by an LLDD
2722 *                       upon the reception of an ABTS for a FCP command
2723 *
2724 * Notify the transport that an ABTS has been received for a FCP command
2725 * that had been given to the transport via nvmet_fc_rcv_fcp_req(). The
2726 * LLDD believes the command is still being worked on
2727 * (template_ops->fcp_req_release() has not been called).
2728 *
2729 * The transport will wait for any outstanding work (an op to the LLDD,
2730 * which the lldd should complete with error due to the ABTS; or the
2731 * completion from the nvmet layer of the nvme command), then will
2732 * stop processing and call the nvmet_fc_rcv_fcp_req() callback to
2733 * return the i/o context to the LLDD.  The LLDD may send the BA_ACC
2734 * to the ABTS either after return from this function (assuming any
2735 * outstanding op work has been terminated) or upon the callback being
2736 * called.
2737 *
2738 * @target_port: pointer to the (registered) target port the FCP CMD IU
2739 *              was received on.
2740 * @fcpreq:     pointer to the fcpreq request structure that corresponds
2741 *              to the exchange that received the ABTS.
2742 */
2743void
2744nvmet_fc_rcv_fcp_abort(struct nvmet_fc_target_port *target_port,
2745			struct nvmefc_tgt_fcp_req *fcpreq)
2746{
2747	struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2748	struct nvmet_fc_tgt_queue *queue;
2749	unsigned long flags;
2750
2751	if (!fod || fod->fcpreq != fcpreq)
2752		/* job appears to have already completed, ignore abort */
2753		return;
2754
2755	queue = fod->queue;
2756
2757	spin_lock_irqsave(&queue->qlock, flags);
2758	if (fod->active) {
2759		/*
2760		 * mark as abort. The abort handler, invoked upon completion
2761		 * of any work, will detect the aborted status and do the
2762		 * callback.
2763		 */
2764		spin_lock(&fod->flock);
2765		fod->abort = true;
2766		fod->aborted = true;
2767		spin_unlock(&fod->flock);
2768	}
2769	spin_unlock_irqrestore(&queue->qlock, flags);
2770}
2771EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_abort);
2772
2773
2774struct nvmet_fc_traddr {
2775	u64	nn;
2776	u64	pn;
2777};
2778
2779static int
2780__nvme_fc_parse_u64(substring_t *sstr, u64 *val)
2781{
2782	u64 token64;
2783
2784	if (match_u64(sstr, &token64))
2785		return -EINVAL;
2786	*val = token64;
2787
2788	return 0;
2789}
2790
2791/*
2792 * This routine validates and extracts the WWN's from the TRADDR string.
2793 * As kernel parsers need the 0x to determine number base, universally
2794 * build string to parse with 0x prefix before parsing name strings.
2795 */
2796static int
2797nvme_fc_parse_traddr(struct nvmet_fc_traddr *traddr, char *buf, size_t blen)
2798{
2799	char name[2 + NVME_FC_TRADDR_HEXNAMELEN + 1];
2800	substring_t wwn = { name, &name[sizeof(name)-1] };
2801	int nnoffset, pnoffset;
2802
2803	/* validate if string is one of the 2 allowed formats */
2804	if (strnlen(buf, blen) == NVME_FC_TRADDR_MAXLENGTH &&
2805			!strncmp(buf, "nn-0x", NVME_FC_TRADDR_OXNNLEN) &&
2806			!strncmp(&buf[NVME_FC_TRADDR_MAX_PN_OFFSET],
2807				"pn-0x", NVME_FC_TRADDR_OXNNLEN)) {
2808		nnoffset = NVME_FC_TRADDR_OXNNLEN;
2809		pnoffset = NVME_FC_TRADDR_MAX_PN_OFFSET +
2810						NVME_FC_TRADDR_OXNNLEN;
2811	} else if ((strnlen(buf, blen) == NVME_FC_TRADDR_MINLENGTH &&
2812			!strncmp(buf, "nn-", NVME_FC_TRADDR_NNLEN) &&
2813			!strncmp(&buf[NVME_FC_TRADDR_MIN_PN_OFFSET],
2814				"pn-", NVME_FC_TRADDR_NNLEN))) {
2815		nnoffset = NVME_FC_TRADDR_NNLEN;
2816		pnoffset = NVME_FC_TRADDR_MIN_PN_OFFSET + NVME_FC_TRADDR_NNLEN;
2817	} else
2818		goto out_einval;
2819
2820	name[0] = '0';
2821	name[1] = 'x';
2822	name[2 + NVME_FC_TRADDR_HEXNAMELEN] = 0;
2823
2824	memcpy(&name[2], &buf[nnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2825	if (__nvme_fc_parse_u64(&wwn, &traddr->nn))
2826		goto out_einval;
2827
2828	memcpy(&name[2], &buf[pnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2829	if (__nvme_fc_parse_u64(&wwn, &traddr->pn))
2830		goto out_einval;
2831
2832	return 0;
2833
2834out_einval:
2835	pr_warn("%s: bad traddr string\n", __func__);
2836	return -EINVAL;
2837}
2838
2839static int
2840nvmet_fc_add_port(struct nvmet_port *port)
2841{
2842	struct nvmet_fc_tgtport *tgtport;
2843	struct nvmet_fc_port_entry *pe;
2844	struct nvmet_fc_traddr traddr = { 0L, 0L };
2845	unsigned long flags;
2846	int ret;
2847
2848	/* validate the address info */
2849	if ((port->disc_addr.trtype != NVMF_TRTYPE_FC) ||
2850	    (port->disc_addr.adrfam != NVMF_ADDR_FAMILY_FC))
2851		return -EINVAL;
2852
2853	/* map the traddr address info to a target port */
2854
2855	ret = nvme_fc_parse_traddr(&traddr, port->disc_addr.traddr,
2856			sizeof(port->disc_addr.traddr));
2857	if (ret)
2858		return ret;
2859
2860	pe = kzalloc(sizeof(*pe), GFP_KERNEL);
2861	if (!pe)
2862		return -ENOMEM;
2863
2864	ret = -ENXIO;
2865	spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
2866	list_for_each_entry(tgtport, &nvmet_fc_target_list, tgt_list) {
2867		if ((tgtport->fc_target_port.node_name == traddr.nn) &&
2868		    (tgtport->fc_target_port.port_name == traddr.pn)) {
2869			/* a FC port can only be 1 nvmet port id */
2870			if (!tgtport->pe) {
2871				nvmet_fc_portentry_bind(tgtport, pe, port);
2872				ret = 0;
2873			} else
2874				ret = -EALREADY;
2875			break;
2876		}
2877	}
2878	spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
2879
2880	if (ret)
2881		kfree(pe);
2882
2883	return ret;
2884}
2885
2886static void
2887nvmet_fc_remove_port(struct nvmet_port *port)
2888{
2889	struct nvmet_fc_port_entry *pe = port->priv;
2890
2891	nvmet_fc_portentry_unbind(pe);
2892
2893	kfree(pe);
2894}
2895
2896static void
2897nvmet_fc_discovery_chg(struct nvmet_port *port)
2898{
2899	struct nvmet_fc_port_entry *pe = port->priv;
2900	struct nvmet_fc_tgtport *tgtport = pe->tgtport;
2901
2902	if (tgtport && tgtport->ops->discovery_event)
2903		tgtport->ops->discovery_event(&tgtport->fc_target_port);
2904}
2905
2906static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops = {
2907	.owner			= THIS_MODULE,
2908	.type			= NVMF_TRTYPE_FC,
2909	.msdbd			= 1,
2910	.add_port		= nvmet_fc_add_port,
2911	.remove_port		= nvmet_fc_remove_port,
2912	.queue_response		= nvmet_fc_fcp_nvme_cmd_done,
2913	.delete_ctrl		= nvmet_fc_delete_ctrl,
2914	.discovery_chg		= nvmet_fc_discovery_chg,
2915};
2916
2917static int __init nvmet_fc_init_module(void)
2918{
2919	return nvmet_register_transport(&nvmet_fc_tgt_fcp_ops);
2920}
2921
2922static void __exit nvmet_fc_exit_module(void)
2923{
2924	/* sanity check - all lports should be removed */
2925	if (!list_empty(&nvmet_fc_target_list))
2926		pr_warn("%s: targetport list not empty\n", __func__);
2927
2928	nvmet_unregister_transport(&nvmet_fc_tgt_fcp_ops);
2929
2930	ida_destroy(&nvmet_fc_tgtport_cnt);
2931}
2932
2933module_init(nvmet_fc_init_module);
2934module_exit(nvmet_fc_exit_module);
2935
2936MODULE_LICENSE("GPL v2");
2937