Merge branches 'pm-cpuidle', 'pm-core' and 'pm-sleep'
[linux-2.6-microblaze.git] / drivers / nvme / target / fc.c
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
25 struct nvmet_fc_tgtport;
26 struct nvmet_fc_tgt_assoc;
27
28 struct 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
48 struct 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
63 enum nvmet_fcp_datadir {
64         NVMET_FCP_NODATA,
65         NVMET_FCP_WRITE,
66         NVMET_FCP_READ,
67         NVMET_FCP_ABORTED,
68 };
69
70 struct 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
96 struct 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
116 struct 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
124 struct nvmet_fc_defer_fcp_req {
125         struct list_head                req_list;
126         struct nvmefc_tgt_fcp_req       *fcp_req;
127 };
128
129 struct 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 rcu_head                 rcu;
149         struct nvmet_fc_fcp_iod         fod[];          /* array of fcp_iods */
150 } __aligned(sizeof(unsigned long long));
151
152 struct nvmet_fc_hostport {
153         struct nvmet_fc_tgtport         *tgtport;
154         void                            *hosthandle;
155         struct list_head                host_list;
156         struct kref                     ref;
157         u8                              invalid;
158 };
159
160 struct nvmet_fc_tgt_assoc {
161         u64                             association_id;
162         u32                             a_id;
163         atomic_t                        terminating;
164         struct nvmet_fc_tgtport         *tgtport;
165         struct nvmet_fc_hostport        *hostport;
166         struct nvmet_fc_ls_iod          *rcv_disconn;
167         struct list_head                a_list;
168         struct nvmet_fc_tgt_queue __rcu *queues[NVMET_NR_QUEUES + 1];
169         struct kref                     ref;
170         struct work_struct              del_work;
171         struct rcu_head                 rcu;
172 };
173
174
175 static inline int
176 nvmet_fc_iodnum(struct nvmet_fc_ls_iod *iodptr)
177 {
178         return (iodptr - iodptr->tgtport->iod);
179 }
180
181 static inline int
182 nvmet_fc_fodnum(struct nvmet_fc_fcp_iod *fodptr)
183 {
184         return (fodptr - fodptr->queue->fod);
185 }
186
187
188 /*
189  * Association and Connection IDs:
190  *
191  * Association ID will have random number in upper 6 bytes and zero
192  *   in lower 2 bytes
193  *
194  * Connection IDs will be Association ID with QID or'd in lower 2 bytes
195  *
196  * note: Association ID = Connection ID for queue 0
197  */
198 #define BYTES_FOR_QID                   sizeof(u16)
199 #define BYTES_FOR_QID_SHIFT             (BYTES_FOR_QID * 8)
200 #define NVMET_FC_QUEUEID_MASK           ((u64)((1 << BYTES_FOR_QID_SHIFT) - 1))
201
202 static inline u64
203 nvmet_fc_makeconnid(struct nvmet_fc_tgt_assoc *assoc, u16 qid)
204 {
205         return (assoc->association_id | qid);
206 }
207
208 static inline u64
209 nvmet_fc_getassociationid(u64 connectionid)
210 {
211         return connectionid & ~NVMET_FC_QUEUEID_MASK;
212 }
213
214 static inline u16
215 nvmet_fc_getqueueid(u64 connectionid)
216 {
217         return (u16)(connectionid & NVMET_FC_QUEUEID_MASK);
218 }
219
220 static inline struct nvmet_fc_tgtport *
221 targetport_to_tgtport(struct nvmet_fc_target_port *targetport)
222 {
223         return container_of(targetport, struct nvmet_fc_tgtport,
224                                  fc_target_port);
225 }
226
227 static inline struct nvmet_fc_fcp_iod *
228 nvmet_req_to_fod(struct nvmet_req *nvme_req)
229 {
230         return container_of(nvme_req, struct nvmet_fc_fcp_iod, req);
231 }
232
233
234 /* *************************** Globals **************************** */
235
236
237 static DEFINE_SPINLOCK(nvmet_fc_tgtlock);
238
239 static LIST_HEAD(nvmet_fc_target_list);
240 static DEFINE_IDA(nvmet_fc_tgtport_cnt);
241 static LIST_HEAD(nvmet_fc_portentry_list);
242
243
244 static void nvmet_fc_handle_ls_rqst_work(struct work_struct *work);
245 static void nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work);
246 static void nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc);
247 static int nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc);
248 static void nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue);
249 static int nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue);
250 static void nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport);
251 static int nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport);
252 static void nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
253                                         struct nvmet_fc_fcp_iod *fod);
254 static void nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc);
255 static void nvmet_fc_xmt_ls_rsp(struct nvmet_fc_tgtport *tgtport,
256                                 struct nvmet_fc_ls_iod *iod);
257
258
259 /* *********************** FC-NVME DMA Handling **************************** */
260
261 /*
262  * The fcloop device passes in a NULL device pointer. Real LLD's will
263  * pass in a valid device pointer. If NULL is passed to the dma mapping
264  * routines, depending on the platform, it may or may not succeed, and
265  * may crash.
266  *
267  * As such:
268  * Wrapper all the dma routines and check the dev pointer.
269  *
270  * If simple mappings (return just a dma address, we'll noop them,
271  * returning a dma address of 0.
272  *
273  * On more complex mappings (dma_map_sg), a pseudo routine fills
274  * in the scatter list, setting all dma addresses to 0.
275  */
276
277 static inline dma_addr_t
278 fc_dma_map_single(struct device *dev, void *ptr, size_t size,
279                 enum dma_data_direction dir)
280 {
281         return dev ? dma_map_single(dev, ptr, size, dir) : (dma_addr_t)0L;
282 }
283
284 static inline int
285 fc_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
286 {
287         return dev ? dma_mapping_error(dev, dma_addr) : 0;
288 }
289
290 static inline void
291 fc_dma_unmap_single(struct device *dev, dma_addr_t addr, size_t size,
292         enum dma_data_direction dir)
293 {
294         if (dev)
295                 dma_unmap_single(dev, addr, size, dir);
296 }
297
298 static inline void
299 fc_dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
300                 enum dma_data_direction dir)
301 {
302         if (dev)
303                 dma_sync_single_for_cpu(dev, addr, size, dir);
304 }
305
306 static inline void
307 fc_dma_sync_single_for_device(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_device(dev, addr, size, dir);
312 }
313
314 /* pseudo dma_map_sg call */
315 static int
316 fc_map_sg(struct scatterlist *sg, int nents)
317 {
318         struct scatterlist *s;
319         int i;
320
321         WARN_ON(nents == 0 || sg[0].length == 0);
322
323         for_each_sg(sg, s, nents, i) {
324                 s->dma_address = 0L;
325 #ifdef CONFIG_NEED_SG_DMA_LENGTH
326                 s->dma_length = s->length;
327 #endif
328         }
329         return nents;
330 }
331
332 static inline int
333 fc_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
334                 enum dma_data_direction dir)
335 {
336         return dev ? dma_map_sg(dev, sg, nents, dir) : fc_map_sg(sg, nents);
337 }
338
339 static inline void
340 fc_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
341                 enum dma_data_direction dir)
342 {
343         if (dev)
344                 dma_unmap_sg(dev, sg, nents, dir);
345 }
346
347
348 /* ********************** FC-NVME LS XMT Handling ************************* */
349
350
351 static void
352 __nvmet_fc_finish_ls_req(struct nvmet_fc_ls_req_op *lsop)
353 {
354         struct nvmet_fc_tgtport *tgtport = lsop->tgtport;
355         struct nvmefc_ls_req *lsreq = &lsop->ls_req;
356         unsigned long flags;
357
358         spin_lock_irqsave(&tgtport->lock, flags);
359
360         if (!lsop->req_queued) {
361                 spin_unlock_irqrestore(&tgtport->lock, flags);
362                 return;
363         }
364
365         list_del(&lsop->lsreq_list);
366
367         lsop->req_queued = false;
368
369         spin_unlock_irqrestore(&tgtport->lock, flags);
370
371         fc_dma_unmap_single(tgtport->dev, lsreq->rqstdma,
372                                   (lsreq->rqstlen + lsreq->rsplen),
373                                   DMA_BIDIRECTIONAL);
374
375         nvmet_fc_tgtport_put(tgtport);
376 }
377
378 static int
379 __nvmet_fc_send_ls_req(struct nvmet_fc_tgtport *tgtport,
380                 struct nvmet_fc_ls_req_op *lsop,
381                 void (*done)(struct nvmefc_ls_req *req, int status))
382 {
383         struct nvmefc_ls_req *lsreq = &lsop->ls_req;
384         unsigned long flags;
385         int ret = 0;
386
387         if (!tgtport->ops->ls_req)
388                 return -EOPNOTSUPP;
389
390         if (!nvmet_fc_tgtport_get(tgtport))
391                 return -ESHUTDOWN;
392
393         lsreq->done = done;
394         lsop->req_queued = false;
395         INIT_LIST_HEAD(&lsop->lsreq_list);
396
397         lsreq->rqstdma = fc_dma_map_single(tgtport->dev, lsreq->rqstaddr,
398                                   lsreq->rqstlen + lsreq->rsplen,
399                                   DMA_BIDIRECTIONAL);
400         if (fc_dma_mapping_error(tgtport->dev, lsreq->rqstdma)) {
401                 ret = -EFAULT;
402                 goto out_puttgtport;
403         }
404         lsreq->rspdma = lsreq->rqstdma + lsreq->rqstlen;
405
406         spin_lock_irqsave(&tgtport->lock, flags);
407
408         list_add_tail(&lsop->lsreq_list, &tgtport->ls_req_list);
409
410         lsop->req_queued = true;
411
412         spin_unlock_irqrestore(&tgtport->lock, flags);
413
414         ret = tgtport->ops->ls_req(&tgtport->fc_target_port, lsop->hosthandle,
415                                    lsreq);
416         if (ret)
417                 goto out_unlink;
418
419         return 0;
420
421 out_unlink:
422         lsop->ls_error = ret;
423         spin_lock_irqsave(&tgtport->lock, flags);
424         lsop->req_queued = false;
425         list_del(&lsop->lsreq_list);
426         spin_unlock_irqrestore(&tgtport->lock, flags);
427         fc_dma_unmap_single(tgtport->dev, lsreq->rqstdma,
428                                   (lsreq->rqstlen + lsreq->rsplen),
429                                   DMA_BIDIRECTIONAL);
430 out_puttgtport:
431         nvmet_fc_tgtport_put(tgtport);
432
433         return ret;
434 }
435
436 static int
437 nvmet_fc_send_ls_req_async(struct nvmet_fc_tgtport *tgtport,
438                 struct nvmet_fc_ls_req_op *lsop,
439                 void (*done)(struct nvmefc_ls_req *req, int status))
440 {
441         /* don't wait for completion */
442
443         return __nvmet_fc_send_ls_req(tgtport, lsop, done);
444 }
445
446 static void
447 nvmet_fc_disconnect_assoc_done(struct nvmefc_ls_req *lsreq, int status)
448 {
449         struct nvmet_fc_ls_req_op *lsop =
450                 container_of(lsreq, struct nvmet_fc_ls_req_op, ls_req);
451
452         __nvmet_fc_finish_ls_req(lsop);
453
454         /* fc-nvme target doesn't care about success or failure of cmd */
455
456         kfree(lsop);
457 }
458
459 /*
460  * This routine sends a FC-NVME LS to disconnect (aka terminate)
461  * the FC-NVME Association.  Terminating the association also
462  * terminates the FC-NVME connections (per queue, both admin and io
463  * queues) that are part of the association. E.g. things are torn
464  * down, and the related FC-NVME Association ID and Connection IDs
465  * become invalid.
466  *
467  * The behavior of the fc-nvme target is such that it's
468  * understanding of the association and connections will implicitly
469  * be torn down. The action is implicit as it may be due to a loss of
470  * connectivity with the fc-nvme host, so the target may never get a
471  * response even if it tried.  As such, the action of this routine
472  * is to asynchronously send the LS, ignore any results of the LS, and
473  * continue on with terminating the association. If the fc-nvme host
474  * is present and receives the LS, it too can tear down.
475  */
476 static void
477 nvmet_fc_xmt_disconnect_assoc(struct nvmet_fc_tgt_assoc *assoc)
478 {
479         struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
480         struct fcnvme_ls_disconnect_assoc_rqst *discon_rqst;
481         struct fcnvme_ls_disconnect_assoc_acc *discon_acc;
482         struct nvmet_fc_ls_req_op *lsop;
483         struct nvmefc_ls_req *lsreq;
484         int ret;
485
486         /*
487          * If ls_req is NULL or no hosthandle, it's an older lldd and no
488          * message is normal. Otherwise, send unless the hostport has
489          * already been invalidated by the lldd.
490          */
491         if (!tgtport->ops->ls_req || !assoc->hostport ||
492             assoc->hostport->invalid)
493                 return;
494
495         lsop = kzalloc((sizeof(*lsop) +
496                         sizeof(*discon_rqst) + sizeof(*discon_acc) +
497                         tgtport->ops->lsrqst_priv_sz), GFP_KERNEL);
498         if (!lsop) {
499                 dev_info(tgtport->dev,
500                         "{%d:%d} send Disconnect Association failed: ENOMEM\n",
501                         tgtport->fc_target_port.port_num, assoc->a_id);
502                 return;
503         }
504
505         discon_rqst = (struct fcnvme_ls_disconnect_assoc_rqst *)&lsop[1];
506         discon_acc = (struct fcnvme_ls_disconnect_assoc_acc *)&discon_rqst[1];
507         lsreq = &lsop->ls_req;
508         if (tgtport->ops->lsrqst_priv_sz)
509                 lsreq->private = (void *)&discon_acc[1];
510         else
511                 lsreq->private = NULL;
512
513         lsop->tgtport = tgtport;
514         lsop->hosthandle = assoc->hostport->hosthandle;
515
516         nvmefc_fmt_lsreq_discon_assoc(lsreq, discon_rqst, discon_acc,
517                                 assoc->association_id);
518
519         ret = nvmet_fc_send_ls_req_async(tgtport, lsop,
520                                 nvmet_fc_disconnect_assoc_done);
521         if (ret) {
522                 dev_info(tgtport->dev,
523                         "{%d:%d} XMT Disconnect Association failed: %d\n",
524                         tgtport->fc_target_port.port_num, assoc->a_id, ret);
525                 kfree(lsop);
526         }
527 }
528
529
530 /* *********************** FC-NVME Port Management ************************ */
531
532
533 static int
534 nvmet_fc_alloc_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
535 {
536         struct nvmet_fc_ls_iod *iod;
537         int i;
538
539         iod = kcalloc(NVMET_LS_CTX_COUNT, sizeof(struct nvmet_fc_ls_iod),
540                         GFP_KERNEL);
541         if (!iod)
542                 return -ENOMEM;
543
544         tgtport->iod = iod;
545
546         for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
547                 INIT_WORK(&iod->work, nvmet_fc_handle_ls_rqst_work);
548                 iod->tgtport = tgtport;
549                 list_add_tail(&iod->ls_rcv_list, &tgtport->ls_rcv_list);
550
551                 iod->rqstbuf = kzalloc(sizeof(union nvmefc_ls_requests) +
552                                        sizeof(union nvmefc_ls_responses),
553                                        GFP_KERNEL);
554                 if (!iod->rqstbuf)
555                         goto out_fail;
556
557                 iod->rspbuf = (union nvmefc_ls_responses *)&iod->rqstbuf[1];
558
559                 iod->rspdma = fc_dma_map_single(tgtport->dev, iod->rspbuf,
560                                                 sizeof(*iod->rspbuf),
561                                                 DMA_TO_DEVICE);
562                 if (fc_dma_mapping_error(tgtport->dev, iod->rspdma))
563                         goto out_fail;
564         }
565
566         return 0;
567
568 out_fail:
569         kfree(iod->rqstbuf);
570         list_del(&iod->ls_rcv_list);
571         for (iod--, i--; i >= 0; iod--, i--) {
572                 fc_dma_unmap_single(tgtport->dev, iod->rspdma,
573                                 sizeof(*iod->rspbuf), DMA_TO_DEVICE);
574                 kfree(iod->rqstbuf);
575                 list_del(&iod->ls_rcv_list);
576         }
577
578         kfree(iod);
579
580         return -EFAULT;
581 }
582
583 static void
584 nvmet_fc_free_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
585 {
586         struct nvmet_fc_ls_iod *iod = tgtport->iod;
587         int i;
588
589         for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
590                 fc_dma_unmap_single(tgtport->dev,
591                                 iod->rspdma, sizeof(*iod->rspbuf),
592                                 DMA_TO_DEVICE);
593                 kfree(iod->rqstbuf);
594                 list_del(&iod->ls_rcv_list);
595         }
596         kfree(tgtport->iod);
597 }
598
599 static struct nvmet_fc_ls_iod *
600 nvmet_fc_alloc_ls_iod(struct nvmet_fc_tgtport *tgtport)
601 {
602         struct nvmet_fc_ls_iod *iod;
603         unsigned long flags;
604
605         spin_lock_irqsave(&tgtport->lock, flags);
606         iod = list_first_entry_or_null(&tgtport->ls_rcv_list,
607                                         struct nvmet_fc_ls_iod, ls_rcv_list);
608         if (iod)
609                 list_move_tail(&iod->ls_rcv_list, &tgtport->ls_busylist);
610         spin_unlock_irqrestore(&tgtport->lock, flags);
611         return iod;
612 }
613
614
615 static void
616 nvmet_fc_free_ls_iod(struct nvmet_fc_tgtport *tgtport,
617                         struct nvmet_fc_ls_iod *iod)
618 {
619         unsigned long flags;
620
621         spin_lock_irqsave(&tgtport->lock, flags);
622         list_move(&iod->ls_rcv_list, &tgtport->ls_rcv_list);
623         spin_unlock_irqrestore(&tgtport->lock, flags);
624 }
625
626 static void
627 nvmet_fc_prep_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
628                                 struct nvmet_fc_tgt_queue *queue)
629 {
630         struct nvmet_fc_fcp_iod *fod = queue->fod;
631         int i;
632
633         for (i = 0; i < queue->sqsize; fod++, i++) {
634                 INIT_WORK(&fod->defer_work, nvmet_fc_fcp_rqst_op_defer_work);
635                 fod->tgtport = tgtport;
636                 fod->queue = queue;
637                 fod->active = false;
638                 fod->abort = false;
639                 fod->aborted = false;
640                 fod->fcpreq = NULL;
641                 list_add_tail(&fod->fcp_list, &queue->fod_list);
642                 spin_lock_init(&fod->flock);
643
644                 fod->rspdma = fc_dma_map_single(tgtport->dev, &fod->rspiubuf,
645                                         sizeof(fod->rspiubuf), DMA_TO_DEVICE);
646                 if (fc_dma_mapping_error(tgtport->dev, fod->rspdma)) {
647                         list_del(&fod->fcp_list);
648                         for (fod--, i--; i >= 0; fod--, i--) {
649                                 fc_dma_unmap_single(tgtport->dev, fod->rspdma,
650                                                 sizeof(fod->rspiubuf),
651                                                 DMA_TO_DEVICE);
652                                 fod->rspdma = 0L;
653                                 list_del(&fod->fcp_list);
654                         }
655
656                         return;
657                 }
658         }
659 }
660
661 static void
662 nvmet_fc_destroy_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
663                                 struct nvmet_fc_tgt_queue *queue)
664 {
665         struct nvmet_fc_fcp_iod *fod = queue->fod;
666         int i;
667
668         for (i = 0; i < queue->sqsize; fod++, i++) {
669                 if (fod->rspdma)
670                         fc_dma_unmap_single(tgtport->dev, fod->rspdma,
671                                 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
672         }
673 }
674
675 static struct nvmet_fc_fcp_iod *
676 nvmet_fc_alloc_fcp_iod(struct nvmet_fc_tgt_queue *queue)
677 {
678         struct nvmet_fc_fcp_iod *fod;
679
680         lockdep_assert_held(&queue->qlock);
681
682         fod = list_first_entry_or_null(&queue->fod_list,
683                                         struct nvmet_fc_fcp_iod, fcp_list);
684         if (fod) {
685                 list_del(&fod->fcp_list);
686                 fod->active = true;
687                 /*
688                  * no queue reference is taken, as it was taken by the
689                  * queue lookup just prior to the allocation. The iod
690                  * will "inherit" that reference.
691                  */
692         }
693         return fod;
694 }
695
696
697 static void
698 nvmet_fc_queue_fcp_req(struct nvmet_fc_tgtport *tgtport,
699                        struct nvmet_fc_tgt_queue *queue,
700                        struct nvmefc_tgt_fcp_req *fcpreq)
701 {
702         struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
703
704         /*
705          * put all admin cmds on hw queue id 0. All io commands go to
706          * the respective hw queue based on a modulo basis
707          */
708         fcpreq->hwqid = queue->qid ?
709                         ((queue->qid - 1) % tgtport->ops->max_hw_queues) : 0;
710
711         nvmet_fc_handle_fcp_rqst(tgtport, fod);
712 }
713
714 static void
715 nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work)
716 {
717         struct nvmet_fc_fcp_iod *fod =
718                 container_of(work, struct nvmet_fc_fcp_iod, defer_work);
719
720         /* Submit deferred IO for processing */
721         nvmet_fc_queue_fcp_req(fod->tgtport, fod->queue, fod->fcpreq);
722
723 }
724
725 static void
726 nvmet_fc_free_fcp_iod(struct nvmet_fc_tgt_queue *queue,
727                         struct nvmet_fc_fcp_iod *fod)
728 {
729         struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
730         struct nvmet_fc_tgtport *tgtport = fod->tgtport;
731         struct nvmet_fc_defer_fcp_req *deferfcp;
732         unsigned long flags;
733
734         fc_dma_sync_single_for_cpu(tgtport->dev, fod->rspdma,
735                                 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
736
737         fcpreq->nvmet_fc_private = NULL;
738
739         fod->active = false;
740         fod->abort = false;
741         fod->aborted = false;
742         fod->writedataactive = false;
743         fod->fcpreq = NULL;
744
745         tgtport->ops->fcp_req_release(&tgtport->fc_target_port, fcpreq);
746
747         /* release the queue lookup reference on the completed IO */
748         nvmet_fc_tgt_q_put(queue);
749
750         spin_lock_irqsave(&queue->qlock, flags);
751         deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
752                                 struct nvmet_fc_defer_fcp_req, req_list);
753         if (!deferfcp) {
754                 list_add_tail(&fod->fcp_list, &fod->queue->fod_list);
755                 spin_unlock_irqrestore(&queue->qlock, flags);
756                 return;
757         }
758
759         /* Re-use the fod for the next pending cmd that was deferred */
760         list_del(&deferfcp->req_list);
761
762         fcpreq = deferfcp->fcp_req;
763
764         /* deferfcp can be reused for another IO at a later date */
765         list_add_tail(&deferfcp->req_list, &queue->avail_defer_list);
766
767         spin_unlock_irqrestore(&queue->qlock, flags);
768
769         /* Save NVME CMD IO in fod */
770         memcpy(&fod->cmdiubuf, fcpreq->rspaddr, fcpreq->rsplen);
771
772         /* Setup new fcpreq to be processed */
773         fcpreq->rspaddr = NULL;
774         fcpreq->rsplen  = 0;
775         fcpreq->nvmet_fc_private = fod;
776         fod->fcpreq = fcpreq;
777         fod->active = true;
778
779         /* inform LLDD IO is now being processed */
780         tgtport->ops->defer_rcv(&tgtport->fc_target_port, fcpreq);
781
782         /*
783          * Leave the queue lookup get reference taken when
784          * fod was originally allocated.
785          */
786
787         queue_work(queue->work_q, &fod->defer_work);
788 }
789
790 static struct nvmet_fc_tgt_queue *
791 nvmet_fc_alloc_target_queue(struct nvmet_fc_tgt_assoc *assoc,
792                         u16 qid, u16 sqsize)
793 {
794         struct nvmet_fc_tgt_queue *queue;
795         int ret;
796
797         if (qid > NVMET_NR_QUEUES)
798                 return NULL;
799
800         queue = kzalloc(struct_size(queue, fod, sqsize), GFP_KERNEL);
801         if (!queue)
802                 return NULL;
803
804         if (!nvmet_fc_tgt_a_get(assoc))
805                 goto out_free_queue;
806
807         queue->work_q = alloc_workqueue("ntfc%d.%d.%d", 0, 0,
808                                 assoc->tgtport->fc_target_port.port_num,
809                                 assoc->a_id, qid);
810         if (!queue->work_q)
811                 goto out_a_put;
812
813         queue->qid = qid;
814         queue->sqsize = sqsize;
815         queue->assoc = assoc;
816         INIT_LIST_HEAD(&queue->fod_list);
817         INIT_LIST_HEAD(&queue->avail_defer_list);
818         INIT_LIST_HEAD(&queue->pending_cmd_list);
819         atomic_set(&queue->connected, 0);
820         atomic_set(&queue->sqtail, 0);
821         atomic_set(&queue->rsn, 1);
822         atomic_set(&queue->zrspcnt, 0);
823         spin_lock_init(&queue->qlock);
824         kref_init(&queue->ref);
825
826         nvmet_fc_prep_fcp_iodlist(assoc->tgtport, queue);
827
828         ret = nvmet_sq_init(&queue->nvme_sq);
829         if (ret)
830                 goto out_fail_iodlist;
831
832         WARN_ON(assoc->queues[qid]);
833         rcu_assign_pointer(assoc->queues[qid], queue);
834
835         return queue;
836
837 out_fail_iodlist:
838         nvmet_fc_destroy_fcp_iodlist(assoc->tgtport, queue);
839         destroy_workqueue(queue->work_q);
840 out_a_put:
841         nvmet_fc_tgt_a_put(assoc);
842 out_free_queue:
843         kfree(queue);
844         return NULL;
845 }
846
847
848 static void
849 nvmet_fc_tgt_queue_free(struct kref *ref)
850 {
851         struct nvmet_fc_tgt_queue *queue =
852                 container_of(ref, struct nvmet_fc_tgt_queue, ref);
853
854         rcu_assign_pointer(queue->assoc->queues[queue->qid], NULL);
855
856         nvmet_fc_destroy_fcp_iodlist(queue->assoc->tgtport, queue);
857
858         nvmet_fc_tgt_a_put(queue->assoc);
859
860         destroy_workqueue(queue->work_q);
861
862         kfree_rcu(queue, rcu);
863 }
864
865 static void
866 nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue)
867 {
868         kref_put(&queue->ref, nvmet_fc_tgt_queue_free);
869 }
870
871 static int
872 nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue)
873 {
874         return kref_get_unless_zero(&queue->ref);
875 }
876
877
878 static void
879 nvmet_fc_delete_target_queue(struct nvmet_fc_tgt_queue *queue)
880 {
881         struct nvmet_fc_tgtport *tgtport = queue->assoc->tgtport;
882         struct nvmet_fc_fcp_iod *fod = queue->fod;
883         struct nvmet_fc_defer_fcp_req *deferfcp, *tempptr;
884         unsigned long flags;
885         int i;
886         bool disconnect;
887
888         disconnect = atomic_xchg(&queue->connected, 0);
889
890         /* if not connected, nothing to do */
891         if (!disconnect)
892                 return;
893
894         spin_lock_irqsave(&queue->qlock, flags);
895         /* abort outstanding io's */
896         for (i = 0; i < queue->sqsize; fod++, i++) {
897                 if (fod->active) {
898                         spin_lock(&fod->flock);
899                         fod->abort = true;
900                         /*
901                          * only call lldd abort routine if waiting for
902                          * writedata. other outstanding ops should finish
903                          * on their own.
904                          */
905                         if (fod->writedataactive) {
906                                 fod->aborted = true;
907                                 spin_unlock(&fod->flock);
908                                 tgtport->ops->fcp_abort(
909                                         &tgtport->fc_target_port, fod->fcpreq);
910                         } else
911                                 spin_unlock(&fod->flock);
912                 }
913         }
914
915         /* Cleanup defer'ed IOs in queue */
916         list_for_each_entry_safe(deferfcp, tempptr, &queue->avail_defer_list,
917                                 req_list) {
918                 list_del(&deferfcp->req_list);
919                 kfree(deferfcp);
920         }
921
922         for (;;) {
923                 deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
924                                 struct nvmet_fc_defer_fcp_req, req_list);
925                 if (!deferfcp)
926                         break;
927
928                 list_del(&deferfcp->req_list);
929                 spin_unlock_irqrestore(&queue->qlock, flags);
930
931                 tgtport->ops->defer_rcv(&tgtport->fc_target_port,
932                                 deferfcp->fcp_req);
933
934                 tgtport->ops->fcp_abort(&tgtport->fc_target_port,
935                                 deferfcp->fcp_req);
936
937                 tgtport->ops->fcp_req_release(&tgtport->fc_target_port,
938                                 deferfcp->fcp_req);
939
940                 /* release the queue lookup reference */
941                 nvmet_fc_tgt_q_put(queue);
942
943                 kfree(deferfcp);
944
945                 spin_lock_irqsave(&queue->qlock, flags);
946         }
947         spin_unlock_irqrestore(&queue->qlock, flags);
948
949         flush_workqueue(queue->work_q);
950
951         nvmet_sq_destroy(&queue->nvme_sq);
952
953         nvmet_fc_tgt_q_put(queue);
954 }
955
956 static struct nvmet_fc_tgt_queue *
957 nvmet_fc_find_target_queue(struct nvmet_fc_tgtport *tgtport,
958                                 u64 connection_id)
959 {
960         struct nvmet_fc_tgt_assoc *assoc;
961         struct nvmet_fc_tgt_queue *queue;
962         u64 association_id = nvmet_fc_getassociationid(connection_id);
963         u16 qid = nvmet_fc_getqueueid(connection_id);
964
965         if (qid > NVMET_NR_QUEUES)
966                 return NULL;
967
968         rcu_read_lock();
969         list_for_each_entry_rcu(assoc, &tgtport->assoc_list, a_list) {
970                 if (association_id == assoc->association_id) {
971                         queue = rcu_dereference(assoc->queues[qid]);
972                         if (queue &&
973                             (!atomic_read(&queue->connected) ||
974                              !nvmet_fc_tgt_q_get(queue)))
975                                 queue = NULL;
976                         rcu_read_unlock();
977                         return queue;
978                 }
979         }
980         rcu_read_unlock();
981         return NULL;
982 }
983
984 static void
985 nvmet_fc_hostport_free(struct kref *ref)
986 {
987         struct nvmet_fc_hostport *hostport =
988                 container_of(ref, struct nvmet_fc_hostport, ref);
989         struct nvmet_fc_tgtport *tgtport = hostport->tgtport;
990         unsigned long flags;
991
992         spin_lock_irqsave(&tgtport->lock, flags);
993         list_del(&hostport->host_list);
994         spin_unlock_irqrestore(&tgtport->lock, flags);
995         if (tgtport->ops->host_release && hostport->invalid)
996                 tgtport->ops->host_release(hostport->hosthandle);
997         kfree(hostport);
998         nvmet_fc_tgtport_put(tgtport);
999 }
1000
1001 static void
1002 nvmet_fc_hostport_put(struct nvmet_fc_hostport *hostport)
1003 {
1004         kref_put(&hostport->ref, nvmet_fc_hostport_free);
1005 }
1006
1007 static int
1008 nvmet_fc_hostport_get(struct nvmet_fc_hostport *hostport)
1009 {
1010         return kref_get_unless_zero(&hostport->ref);
1011 }
1012
1013 static void
1014 nvmet_fc_free_hostport(struct nvmet_fc_hostport *hostport)
1015 {
1016         /* if LLDD not implemented, leave as NULL */
1017         if (!hostport || !hostport->hosthandle)
1018                 return;
1019
1020         nvmet_fc_hostport_put(hostport);
1021 }
1022
1023 static struct nvmet_fc_hostport *
1024 nvmet_fc_match_hostport(struct nvmet_fc_tgtport *tgtport, void *hosthandle)
1025 {
1026         struct nvmet_fc_hostport *host;
1027
1028         lockdep_assert_held(&tgtport->lock);
1029
1030         list_for_each_entry(host, &tgtport->host_list, host_list) {
1031                 if (host->hosthandle == hosthandle && !host->invalid) {
1032                         if (nvmet_fc_hostport_get(host))
1033                                 return (host);
1034                 }
1035         }
1036
1037         return NULL;
1038 }
1039
1040 static struct nvmet_fc_hostport *
1041 nvmet_fc_alloc_hostport(struct nvmet_fc_tgtport *tgtport, void *hosthandle)
1042 {
1043         struct nvmet_fc_hostport *newhost, *match = NULL;
1044         unsigned long flags;
1045
1046         /* if LLDD not implemented, leave as NULL */
1047         if (!hosthandle)
1048                 return NULL;
1049
1050         /*
1051          * take reference for what will be the newly allocated hostport if
1052          * we end up using a new allocation
1053          */
1054         if (!nvmet_fc_tgtport_get(tgtport))
1055                 return ERR_PTR(-EINVAL);
1056
1057         spin_lock_irqsave(&tgtport->lock, flags);
1058         match = nvmet_fc_match_hostport(tgtport, hosthandle);
1059         spin_unlock_irqrestore(&tgtport->lock, flags);
1060
1061         if (match) {
1062                 /* no new allocation - release reference */
1063                 nvmet_fc_tgtport_put(tgtport);
1064                 return match;
1065         }
1066
1067         newhost = kzalloc(sizeof(*newhost), GFP_KERNEL);
1068         if (!newhost) {
1069                 /* no new allocation - release reference */
1070                 nvmet_fc_tgtport_put(tgtport);
1071                 return ERR_PTR(-ENOMEM);
1072         }
1073
1074         spin_lock_irqsave(&tgtport->lock, flags);
1075         match = nvmet_fc_match_hostport(tgtport, hosthandle);
1076         if (match) {
1077                 /* new allocation not needed */
1078                 kfree(newhost);
1079                 newhost = match;
1080                 /* no new allocation - release reference */
1081                 nvmet_fc_tgtport_put(tgtport);
1082         } else {
1083                 newhost->tgtport = tgtport;
1084                 newhost->hosthandle = hosthandle;
1085                 INIT_LIST_HEAD(&newhost->host_list);
1086                 kref_init(&newhost->ref);
1087
1088                 list_add_tail(&newhost->host_list, &tgtport->host_list);
1089         }
1090         spin_unlock_irqrestore(&tgtport->lock, flags);
1091
1092         return newhost;
1093 }
1094
1095 static void
1096 nvmet_fc_delete_assoc(struct work_struct *work)
1097 {
1098         struct nvmet_fc_tgt_assoc *assoc =
1099                 container_of(work, struct nvmet_fc_tgt_assoc, del_work);
1100
1101         nvmet_fc_delete_target_assoc(assoc);
1102         nvmet_fc_tgt_a_put(assoc);
1103 }
1104
1105 static struct nvmet_fc_tgt_assoc *
1106 nvmet_fc_alloc_target_assoc(struct nvmet_fc_tgtport *tgtport, void *hosthandle)
1107 {
1108         struct nvmet_fc_tgt_assoc *assoc, *tmpassoc;
1109         unsigned long flags;
1110         u64 ran;
1111         int idx;
1112         bool needrandom = true;
1113
1114         assoc = kzalloc(sizeof(*assoc), GFP_KERNEL);
1115         if (!assoc)
1116                 return NULL;
1117
1118         idx = ida_alloc(&tgtport->assoc_cnt, GFP_KERNEL);
1119         if (idx < 0)
1120                 goto out_free_assoc;
1121
1122         if (!nvmet_fc_tgtport_get(tgtport))
1123                 goto out_ida;
1124
1125         assoc->hostport = nvmet_fc_alloc_hostport(tgtport, hosthandle);
1126         if (IS_ERR(assoc->hostport))
1127                 goto out_put;
1128
1129         assoc->tgtport = tgtport;
1130         assoc->a_id = idx;
1131         INIT_LIST_HEAD(&assoc->a_list);
1132         kref_init(&assoc->ref);
1133         INIT_WORK(&assoc->del_work, nvmet_fc_delete_assoc);
1134         atomic_set(&assoc->terminating, 0);
1135
1136         while (needrandom) {
1137                 get_random_bytes(&ran, sizeof(ran) - BYTES_FOR_QID);
1138                 ran = ran << BYTES_FOR_QID_SHIFT;
1139
1140                 spin_lock_irqsave(&tgtport->lock, flags);
1141                 needrandom = false;
1142                 list_for_each_entry(tmpassoc, &tgtport->assoc_list, a_list) {
1143                         if (ran == tmpassoc->association_id) {
1144                                 needrandom = true;
1145                                 break;
1146                         }
1147                 }
1148                 if (!needrandom) {
1149                         assoc->association_id = ran;
1150                         list_add_tail_rcu(&assoc->a_list, &tgtport->assoc_list);
1151                 }
1152                 spin_unlock_irqrestore(&tgtport->lock, flags);
1153         }
1154
1155         return assoc;
1156
1157 out_put:
1158         nvmet_fc_tgtport_put(tgtport);
1159 out_ida:
1160         ida_free(&tgtport->assoc_cnt, idx);
1161 out_free_assoc:
1162         kfree(assoc);
1163         return NULL;
1164 }
1165
1166 static void
1167 nvmet_fc_target_assoc_free(struct kref *ref)
1168 {
1169         struct nvmet_fc_tgt_assoc *assoc =
1170                 container_of(ref, struct nvmet_fc_tgt_assoc, ref);
1171         struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
1172         struct nvmet_fc_ls_iod  *oldls;
1173         unsigned long flags;
1174
1175         /* Send Disconnect now that all i/o has completed */
1176         nvmet_fc_xmt_disconnect_assoc(assoc);
1177
1178         nvmet_fc_free_hostport(assoc->hostport);
1179         spin_lock_irqsave(&tgtport->lock, flags);
1180         list_del_rcu(&assoc->a_list);
1181         oldls = assoc->rcv_disconn;
1182         spin_unlock_irqrestore(&tgtport->lock, flags);
1183         /* if pending Rcv Disconnect Association LS, send rsp now */
1184         if (oldls)
1185                 nvmet_fc_xmt_ls_rsp(tgtport, oldls);
1186         ida_free(&tgtport->assoc_cnt, assoc->a_id);
1187         dev_info(tgtport->dev,
1188                 "{%d:%d} Association freed\n",
1189                 tgtport->fc_target_port.port_num, assoc->a_id);
1190         kfree_rcu(assoc, rcu);
1191         nvmet_fc_tgtport_put(tgtport);
1192 }
1193
1194 static void
1195 nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc)
1196 {
1197         kref_put(&assoc->ref, nvmet_fc_target_assoc_free);
1198 }
1199
1200 static int
1201 nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc)
1202 {
1203         return kref_get_unless_zero(&assoc->ref);
1204 }
1205
1206 static void
1207 nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc)
1208 {
1209         struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
1210         struct nvmet_fc_tgt_queue *queue;
1211         int i, terminating;
1212
1213         terminating = atomic_xchg(&assoc->terminating, 1);
1214
1215         /* if already terminating, do nothing */
1216         if (terminating)
1217                 return;
1218
1219
1220         for (i = NVMET_NR_QUEUES; i >= 0; i--) {
1221                 rcu_read_lock();
1222                 queue = rcu_dereference(assoc->queues[i]);
1223                 if (!queue) {
1224                         rcu_read_unlock();
1225                         continue;
1226                 }
1227
1228                 if (!nvmet_fc_tgt_q_get(queue)) {
1229                         rcu_read_unlock();
1230                         continue;
1231                 }
1232                 rcu_read_unlock();
1233                 nvmet_fc_delete_target_queue(queue);
1234                 nvmet_fc_tgt_q_put(queue);
1235         }
1236
1237         dev_info(tgtport->dev,
1238                 "{%d:%d} Association deleted\n",
1239                 tgtport->fc_target_port.port_num, assoc->a_id);
1240
1241         nvmet_fc_tgt_a_put(assoc);
1242 }
1243
1244 static struct nvmet_fc_tgt_assoc *
1245 nvmet_fc_find_target_assoc(struct nvmet_fc_tgtport *tgtport,
1246                                 u64 association_id)
1247 {
1248         struct nvmet_fc_tgt_assoc *assoc;
1249         struct nvmet_fc_tgt_assoc *ret = NULL;
1250
1251         rcu_read_lock();
1252         list_for_each_entry_rcu(assoc, &tgtport->assoc_list, a_list) {
1253                 if (association_id == assoc->association_id) {
1254                         ret = assoc;
1255                         if (!nvmet_fc_tgt_a_get(assoc))
1256                                 ret = NULL;
1257                         break;
1258                 }
1259         }
1260         rcu_read_unlock();
1261
1262         return ret;
1263 }
1264
1265 static void
1266 nvmet_fc_portentry_bind(struct nvmet_fc_tgtport *tgtport,
1267                         struct nvmet_fc_port_entry *pe,
1268                         struct nvmet_port *port)
1269 {
1270         lockdep_assert_held(&nvmet_fc_tgtlock);
1271
1272         pe->tgtport = tgtport;
1273         tgtport->pe = pe;
1274
1275         pe->port = port;
1276         port->priv = pe;
1277
1278         pe->node_name = tgtport->fc_target_port.node_name;
1279         pe->port_name = tgtport->fc_target_port.port_name;
1280         INIT_LIST_HEAD(&pe->pe_list);
1281
1282         list_add_tail(&pe->pe_list, &nvmet_fc_portentry_list);
1283 }
1284
1285 static void
1286 nvmet_fc_portentry_unbind(struct nvmet_fc_port_entry *pe)
1287 {
1288         unsigned long flags;
1289
1290         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1291         if (pe->tgtport)
1292                 pe->tgtport->pe = NULL;
1293         list_del(&pe->pe_list);
1294         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1295 }
1296
1297 /*
1298  * called when a targetport deregisters. Breaks the relationship
1299  * with the nvmet port, but leaves the port_entry in place so that
1300  * re-registration can resume operation.
1301  */
1302 static void
1303 nvmet_fc_portentry_unbind_tgt(struct nvmet_fc_tgtport *tgtport)
1304 {
1305         struct nvmet_fc_port_entry *pe;
1306         unsigned long flags;
1307
1308         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1309         pe = tgtport->pe;
1310         if (pe)
1311                 pe->tgtport = NULL;
1312         tgtport->pe = NULL;
1313         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1314 }
1315
1316 /*
1317  * called when a new targetport is registered. Looks in the
1318  * existing nvmet port_entries to see if the nvmet layer is
1319  * configured for the targetport's wwn's. (the targetport existed,
1320  * nvmet configured, the lldd unregistered the tgtport, and is now
1321  * reregistering the same targetport).  If so, set the nvmet port
1322  * port entry on the targetport.
1323  */
1324 static void
1325 nvmet_fc_portentry_rebind_tgt(struct nvmet_fc_tgtport *tgtport)
1326 {
1327         struct nvmet_fc_port_entry *pe;
1328         unsigned long flags;
1329
1330         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1331         list_for_each_entry(pe, &nvmet_fc_portentry_list, pe_list) {
1332                 if (tgtport->fc_target_port.node_name == pe->node_name &&
1333                     tgtport->fc_target_port.port_name == pe->port_name) {
1334                         WARN_ON(pe->tgtport);
1335                         tgtport->pe = pe;
1336                         pe->tgtport = tgtport;
1337                         break;
1338                 }
1339         }
1340         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1341 }
1342
1343 /**
1344  * nvmet_fc_register_targetport - transport entry point called by an
1345  *                              LLDD to register the existence of a local
1346  *                              NVME subystem FC port.
1347  * @pinfo:     pointer to information about the port to be registered
1348  * @template:  LLDD entrypoints and operational parameters for the port
1349  * @dev:       physical hardware device node port corresponds to. Will be
1350  *             used for DMA mappings
1351  * @portptr:   pointer to a local port pointer. Upon success, the routine
1352  *             will allocate a nvme_fc_local_port structure and place its
1353  *             address in the local port pointer. Upon failure, local port
1354  *             pointer will be set to NULL.
1355  *
1356  * Returns:
1357  * a completion status. Must be 0 upon success; a negative errno
1358  * (ex: -ENXIO) upon failure.
1359  */
1360 int
1361 nvmet_fc_register_targetport(struct nvmet_fc_port_info *pinfo,
1362                         struct nvmet_fc_target_template *template,
1363                         struct device *dev,
1364                         struct nvmet_fc_target_port **portptr)
1365 {
1366         struct nvmet_fc_tgtport *newrec;
1367         unsigned long flags;
1368         int ret, idx;
1369
1370         if (!template->xmt_ls_rsp || !template->fcp_op ||
1371             !template->fcp_abort ||
1372             !template->fcp_req_release || !template->targetport_delete ||
1373             !template->max_hw_queues || !template->max_sgl_segments ||
1374             !template->max_dif_sgl_segments || !template->dma_boundary) {
1375                 ret = -EINVAL;
1376                 goto out_regtgt_failed;
1377         }
1378
1379         newrec = kzalloc((sizeof(*newrec) + template->target_priv_sz),
1380                          GFP_KERNEL);
1381         if (!newrec) {
1382                 ret = -ENOMEM;
1383                 goto out_regtgt_failed;
1384         }
1385
1386         idx = ida_alloc(&nvmet_fc_tgtport_cnt, GFP_KERNEL);
1387         if (idx < 0) {
1388                 ret = -ENOSPC;
1389                 goto out_fail_kfree;
1390         }
1391
1392         if (!get_device(dev) && dev) {
1393                 ret = -ENODEV;
1394                 goto out_ida_put;
1395         }
1396
1397         newrec->fc_target_port.node_name = pinfo->node_name;
1398         newrec->fc_target_port.port_name = pinfo->port_name;
1399         if (template->target_priv_sz)
1400                 newrec->fc_target_port.private = &newrec[1];
1401         else
1402                 newrec->fc_target_port.private = NULL;
1403         newrec->fc_target_port.port_id = pinfo->port_id;
1404         newrec->fc_target_port.port_num = idx;
1405         INIT_LIST_HEAD(&newrec->tgt_list);
1406         newrec->dev = dev;
1407         newrec->ops = template;
1408         spin_lock_init(&newrec->lock);
1409         INIT_LIST_HEAD(&newrec->ls_rcv_list);
1410         INIT_LIST_HEAD(&newrec->ls_req_list);
1411         INIT_LIST_HEAD(&newrec->ls_busylist);
1412         INIT_LIST_HEAD(&newrec->assoc_list);
1413         INIT_LIST_HEAD(&newrec->host_list);
1414         kref_init(&newrec->ref);
1415         ida_init(&newrec->assoc_cnt);
1416         newrec->max_sg_cnt = template->max_sgl_segments;
1417
1418         ret = nvmet_fc_alloc_ls_iodlist(newrec);
1419         if (ret) {
1420                 ret = -ENOMEM;
1421                 goto out_free_newrec;
1422         }
1423
1424         nvmet_fc_portentry_rebind_tgt(newrec);
1425
1426         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1427         list_add_tail(&newrec->tgt_list, &nvmet_fc_target_list);
1428         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1429
1430         *portptr = &newrec->fc_target_port;
1431         return 0;
1432
1433 out_free_newrec:
1434         put_device(dev);
1435 out_ida_put:
1436         ida_free(&nvmet_fc_tgtport_cnt, idx);
1437 out_fail_kfree:
1438         kfree(newrec);
1439 out_regtgt_failed:
1440         *portptr = NULL;
1441         return ret;
1442 }
1443 EXPORT_SYMBOL_GPL(nvmet_fc_register_targetport);
1444
1445
1446 static void
1447 nvmet_fc_free_tgtport(struct kref *ref)
1448 {
1449         struct nvmet_fc_tgtport *tgtport =
1450                 container_of(ref, struct nvmet_fc_tgtport, ref);
1451         struct device *dev = tgtport->dev;
1452         unsigned long flags;
1453
1454         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1455         list_del(&tgtport->tgt_list);
1456         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1457
1458         nvmet_fc_free_ls_iodlist(tgtport);
1459
1460         /* let the LLDD know we've finished tearing it down */
1461         tgtport->ops->targetport_delete(&tgtport->fc_target_port);
1462
1463         ida_free(&nvmet_fc_tgtport_cnt,
1464                         tgtport->fc_target_port.port_num);
1465
1466         ida_destroy(&tgtport->assoc_cnt);
1467
1468         kfree(tgtport);
1469
1470         put_device(dev);
1471 }
1472
1473 static void
1474 nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport)
1475 {
1476         kref_put(&tgtport->ref, nvmet_fc_free_tgtport);
1477 }
1478
1479 static int
1480 nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport)
1481 {
1482         return kref_get_unless_zero(&tgtport->ref);
1483 }
1484
1485 static void
1486 __nvmet_fc_free_assocs(struct nvmet_fc_tgtport *tgtport)
1487 {
1488         struct nvmet_fc_tgt_assoc *assoc;
1489
1490         rcu_read_lock();
1491         list_for_each_entry_rcu(assoc, &tgtport->assoc_list, a_list) {
1492                 if (!nvmet_fc_tgt_a_get(assoc))
1493                         continue;
1494                 if (!queue_work(nvmet_wq, &assoc->del_work))
1495                         /* already deleting - release local reference */
1496                         nvmet_fc_tgt_a_put(assoc);
1497         }
1498         rcu_read_unlock();
1499 }
1500
1501 /**
1502  * nvmet_fc_invalidate_host - transport entry point called by an LLDD
1503  *                       to remove references to a hosthandle for LS's.
1504  *
1505  * The nvmet-fc layer ensures that any references to the hosthandle
1506  * on the targetport are forgotten (set to NULL).  The LLDD will
1507  * typically call this when a login with a remote host port has been
1508  * lost, thus LS's for the remote host port are no longer possible.
1509  *
1510  * If an LS request is outstanding to the targetport/hosthandle (or
1511  * issued concurrently with the call to invalidate the host), the
1512  * LLDD is responsible for terminating/aborting the LS and completing
1513  * the LS request. It is recommended that these terminations/aborts
1514  * occur after calling to invalidate the host handle to avoid additional
1515  * retries by the nvmet-fc transport. The nvmet-fc transport may
1516  * continue to reference host handle while it cleans up outstanding
1517  * NVME associations. The nvmet-fc transport will call the
1518  * ops->host_release() callback to notify the LLDD that all references
1519  * are complete and the related host handle can be recovered.
1520  * Note: if there are no references, the callback may be called before
1521  * the invalidate host call returns.
1522  *
1523  * @target_port: pointer to the (registered) target port that a prior
1524  *              LS was received on and which supplied the transport the
1525  *              hosthandle.
1526  * @hosthandle: the handle (pointer) that represents the host port
1527  *              that no longer has connectivity and that LS's should
1528  *              no longer be directed to.
1529  */
1530 void
1531 nvmet_fc_invalidate_host(struct nvmet_fc_target_port *target_port,
1532                         void *hosthandle)
1533 {
1534         struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
1535         struct nvmet_fc_tgt_assoc *assoc, *next;
1536         unsigned long flags;
1537         bool noassoc = true;
1538
1539         spin_lock_irqsave(&tgtport->lock, flags);
1540         list_for_each_entry_safe(assoc, next,
1541                                 &tgtport->assoc_list, a_list) {
1542                 if (!assoc->hostport ||
1543                     assoc->hostport->hosthandle != hosthandle)
1544                         continue;
1545                 if (!nvmet_fc_tgt_a_get(assoc))
1546                         continue;
1547                 assoc->hostport->invalid = 1;
1548                 noassoc = false;
1549                 if (!queue_work(nvmet_wq, &assoc->del_work))
1550                         /* already deleting - release local reference */
1551                         nvmet_fc_tgt_a_put(assoc);
1552         }
1553         spin_unlock_irqrestore(&tgtport->lock, flags);
1554
1555         /* if there's nothing to wait for - call the callback */
1556         if (noassoc && tgtport->ops->host_release)
1557                 tgtport->ops->host_release(hosthandle);
1558 }
1559 EXPORT_SYMBOL_GPL(nvmet_fc_invalidate_host);
1560
1561 /*
1562  * nvmet layer has called to terminate an association
1563  */
1564 static void
1565 nvmet_fc_delete_ctrl(struct nvmet_ctrl *ctrl)
1566 {
1567         struct nvmet_fc_tgtport *tgtport, *next;
1568         struct nvmet_fc_tgt_assoc *assoc;
1569         struct nvmet_fc_tgt_queue *queue;
1570         unsigned long flags;
1571         bool found_ctrl = false;
1572
1573         /* this is a bit ugly, but don't want to make locks layered */
1574         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1575         list_for_each_entry_safe(tgtport, next, &nvmet_fc_target_list,
1576                         tgt_list) {
1577                 if (!nvmet_fc_tgtport_get(tgtport))
1578                         continue;
1579                 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1580
1581                 rcu_read_lock();
1582                 list_for_each_entry_rcu(assoc, &tgtport->assoc_list, a_list) {
1583                         queue = rcu_dereference(assoc->queues[0]);
1584                         if (queue && queue->nvme_sq.ctrl == ctrl) {
1585                                 if (nvmet_fc_tgt_a_get(assoc))
1586                                         found_ctrl = true;
1587                                 break;
1588                         }
1589                 }
1590                 rcu_read_unlock();
1591
1592                 nvmet_fc_tgtport_put(tgtport);
1593
1594                 if (found_ctrl) {
1595                         if (!queue_work(nvmet_wq, &assoc->del_work))
1596                                 /* already deleting - release local reference */
1597                                 nvmet_fc_tgt_a_put(assoc);
1598                         return;
1599                 }
1600
1601                 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1602         }
1603         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1604 }
1605
1606 /**
1607  * nvmet_fc_unregister_targetport - transport entry point called by an
1608  *                              LLDD to deregister/remove a previously
1609  *                              registered a local NVME subsystem FC port.
1610  * @target_port: pointer to the (registered) target port that is to be
1611  *               deregistered.
1612  *
1613  * Returns:
1614  * a completion status. Must be 0 upon success; a negative errno
1615  * (ex: -ENXIO) upon failure.
1616  */
1617 int
1618 nvmet_fc_unregister_targetport(struct nvmet_fc_target_port *target_port)
1619 {
1620         struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
1621
1622         nvmet_fc_portentry_unbind_tgt(tgtport);
1623
1624         /* terminate any outstanding associations */
1625         __nvmet_fc_free_assocs(tgtport);
1626
1627         /*
1628          * should terminate LS's as well. However, LS's will be generated
1629          * at the tail end of association termination, so they likely don't
1630          * exist yet. And even if they did, it's worthwhile to just let
1631          * them finish and targetport ref counting will clean things up.
1632          */
1633
1634         nvmet_fc_tgtport_put(tgtport);
1635
1636         return 0;
1637 }
1638 EXPORT_SYMBOL_GPL(nvmet_fc_unregister_targetport);
1639
1640
1641 /* ********************** FC-NVME LS RCV Handling ************************* */
1642
1643
1644 static void
1645 nvmet_fc_ls_create_association(struct nvmet_fc_tgtport *tgtport,
1646                         struct nvmet_fc_ls_iod *iod)
1647 {
1648         struct fcnvme_ls_cr_assoc_rqst *rqst = &iod->rqstbuf->rq_cr_assoc;
1649         struct fcnvme_ls_cr_assoc_acc *acc = &iod->rspbuf->rsp_cr_assoc;
1650         struct nvmet_fc_tgt_queue *queue;
1651         int ret = 0;
1652
1653         memset(acc, 0, sizeof(*acc));
1654
1655         /*
1656          * FC-NVME spec changes. There are initiators sending different
1657          * lengths as padding sizes for Create Association Cmd descriptor
1658          * was incorrect.
1659          * Accept anything of "minimum" length. Assume format per 1.15
1660          * spec (with HOSTID reduced to 16 bytes), ignore how long the
1661          * trailing pad length is.
1662          */
1663         if (iod->rqstdatalen < FCNVME_LSDESC_CRA_RQST_MINLEN)
1664                 ret = VERR_CR_ASSOC_LEN;
1665         else if (be32_to_cpu(rqst->desc_list_len) <
1666                         FCNVME_LSDESC_CRA_RQST_MIN_LISTLEN)
1667                 ret = VERR_CR_ASSOC_RQST_LEN;
1668         else if (rqst->assoc_cmd.desc_tag !=
1669                         cpu_to_be32(FCNVME_LSDESC_CREATE_ASSOC_CMD))
1670                 ret = VERR_CR_ASSOC_CMD;
1671         else if (be32_to_cpu(rqst->assoc_cmd.desc_len) <
1672                         FCNVME_LSDESC_CRA_CMD_DESC_MIN_DESCLEN)
1673                 ret = VERR_CR_ASSOC_CMD_LEN;
1674         else if (!rqst->assoc_cmd.ersp_ratio ||
1675                  (be16_to_cpu(rqst->assoc_cmd.ersp_ratio) >=
1676                                 be16_to_cpu(rqst->assoc_cmd.sqsize)))
1677                 ret = VERR_ERSP_RATIO;
1678
1679         else {
1680                 /* new association w/ admin queue */
1681                 iod->assoc = nvmet_fc_alloc_target_assoc(
1682                                                 tgtport, iod->hosthandle);
1683                 if (!iod->assoc)
1684                         ret = VERR_ASSOC_ALLOC_FAIL;
1685                 else {
1686                         queue = nvmet_fc_alloc_target_queue(iod->assoc, 0,
1687                                         be16_to_cpu(rqst->assoc_cmd.sqsize));
1688                         if (!queue) {
1689                                 ret = VERR_QUEUE_ALLOC_FAIL;
1690                                 nvmet_fc_tgt_a_put(iod->assoc);
1691                         }
1692                 }
1693         }
1694
1695         if (ret) {
1696                 dev_err(tgtport->dev,
1697                         "Create Association LS failed: %s\n",
1698                         validation_errors[ret]);
1699                 iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1700                                 sizeof(*acc), rqst->w0.ls_cmd,
1701                                 FCNVME_RJT_RC_LOGIC,
1702                                 FCNVME_RJT_EXP_NONE, 0);
1703                 return;
1704         }
1705
1706         queue->ersp_ratio = be16_to_cpu(rqst->assoc_cmd.ersp_ratio);
1707         atomic_set(&queue->connected, 1);
1708         queue->sqhd = 0;        /* best place to init value */
1709
1710         dev_info(tgtport->dev,
1711                 "{%d:%d} Association created\n",
1712                 tgtport->fc_target_port.port_num, iod->assoc->a_id);
1713
1714         /* format a response */
1715
1716         iod->lsrsp->rsplen = sizeof(*acc);
1717
1718         nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1719                         fcnvme_lsdesc_len(
1720                                 sizeof(struct fcnvme_ls_cr_assoc_acc)),
1721                         FCNVME_LS_CREATE_ASSOCIATION);
1722         acc->associd.desc_tag = cpu_to_be32(FCNVME_LSDESC_ASSOC_ID);
1723         acc->associd.desc_len =
1724                         fcnvme_lsdesc_len(
1725                                 sizeof(struct fcnvme_lsdesc_assoc_id));
1726         acc->associd.association_id =
1727                         cpu_to_be64(nvmet_fc_makeconnid(iod->assoc, 0));
1728         acc->connectid.desc_tag = cpu_to_be32(FCNVME_LSDESC_CONN_ID);
1729         acc->connectid.desc_len =
1730                         fcnvme_lsdesc_len(
1731                                 sizeof(struct fcnvme_lsdesc_conn_id));
1732         acc->connectid.connection_id = acc->associd.association_id;
1733 }
1734
1735 static void
1736 nvmet_fc_ls_create_connection(struct nvmet_fc_tgtport *tgtport,
1737                         struct nvmet_fc_ls_iod *iod)
1738 {
1739         struct fcnvme_ls_cr_conn_rqst *rqst = &iod->rqstbuf->rq_cr_conn;
1740         struct fcnvme_ls_cr_conn_acc *acc = &iod->rspbuf->rsp_cr_conn;
1741         struct nvmet_fc_tgt_queue *queue;
1742         int ret = 0;
1743
1744         memset(acc, 0, sizeof(*acc));
1745
1746         if (iod->rqstdatalen < sizeof(struct fcnvme_ls_cr_conn_rqst))
1747                 ret = VERR_CR_CONN_LEN;
1748         else if (rqst->desc_list_len !=
1749                         fcnvme_lsdesc_len(
1750                                 sizeof(struct fcnvme_ls_cr_conn_rqst)))
1751                 ret = VERR_CR_CONN_RQST_LEN;
1752         else if (rqst->associd.desc_tag != cpu_to_be32(FCNVME_LSDESC_ASSOC_ID))
1753                 ret = VERR_ASSOC_ID;
1754         else if (rqst->associd.desc_len !=
1755                         fcnvme_lsdesc_len(
1756                                 sizeof(struct fcnvme_lsdesc_assoc_id)))
1757                 ret = VERR_ASSOC_ID_LEN;
1758         else if (rqst->connect_cmd.desc_tag !=
1759                         cpu_to_be32(FCNVME_LSDESC_CREATE_CONN_CMD))
1760                 ret = VERR_CR_CONN_CMD;
1761         else if (rqst->connect_cmd.desc_len !=
1762                         fcnvme_lsdesc_len(
1763                                 sizeof(struct fcnvme_lsdesc_cr_conn_cmd)))
1764                 ret = VERR_CR_CONN_CMD_LEN;
1765         else if (!rqst->connect_cmd.ersp_ratio ||
1766                  (be16_to_cpu(rqst->connect_cmd.ersp_ratio) >=
1767                                 be16_to_cpu(rqst->connect_cmd.sqsize)))
1768                 ret = VERR_ERSP_RATIO;
1769
1770         else {
1771                 /* new io queue */
1772                 iod->assoc = nvmet_fc_find_target_assoc(tgtport,
1773                                 be64_to_cpu(rqst->associd.association_id));
1774                 if (!iod->assoc)
1775                         ret = VERR_NO_ASSOC;
1776                 else {
1777                         queue = nvmet_fc_alloc_target_queue(iod->assoc,
1778                                         be16_to_cpu(rqst->connect_cmd.qid),
1779                                         be16_to_cpu(rqst->connect_cmd.sqsize));
1780                         if (!queue)
1781                                 ret = VERR_QUEUE_ALLOC_FAIL;
1782
1783                         /* release get taken in nvmet_fc_find_target_assoc */
1784                         nvmet_fc_tgt_a_put(iod->assoc);
1785                 }
1786         }
1787
1788         if (ret) {
1789                 dev_err(tgtport->dev,
1790                         "Create Connection LS failed: %s\n",
1791                         validation_errors[ret]);
1792                 iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1793                                 sizeof(*acc), rqst->w0.ls_cmd,
1794                                 (ret == VERR_NO_ASSOC) ?
1795                                         FCNVME_RJT_RC_INV_ASSOC :
1796                                         FCNVME_RJT_RC_LOGIC,
1797                                 FCNVME_RJT_EXP_NONE, 0);
1798                 return;
1799         }
1800
1801         queue->ersp_ratio = be16_to_cpu(rqst->connect_cmd.ersp_ratio);
1802         atomic_set(&queue->connected, 1);
1803         queue->sqhd = 0;        /* best place to init value */
1804
1805         /* format a response */
1806
1807         iod->lsrsp->rsplen = sizeof(*acc);
1808
1809         nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1810                         fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_cr_conn_acc)),
1811                         FCNVME_LS_CREATE_CONNECTION);
1812         acc->connectid.desc_tag = cpu_to_be32(FCNVME_LSDESC_CONN_ID);
1813         acc->connectid.desc_len =
1814                         fcnvme_lsdesc_len(
1815                                 sizeof(struct fcnvme_lsdesc_conn_id));
1816         acc->connectid.connection_id =
1817                         cpu_to_be64(nvmet_fc_makeconnid(iod->assoc,
1818                                 be16_to_cpu(rqst->connect_cmd.qid)));
1819 }
1820
1821 /*
1822  * Returns true if the LS response is to be transmit
1823  * Returns false if the LS response is to be delayed
1824  */
1825 static int
1826 nvmet_fc_ls_disconnect(struct nvmet_fc_tgtport *tgtport,
1827                         struct nvmet_fc_ls_iod *iod)
1828 {
1829         struct fcnvme_ls_disconnect_assoc_rqst *rqst =
1830                                                 &iod->rqstbuf->rq_dis_assoc;
1831         struct fcnvme_ls_disconnect_assoc_acc *acc =
1832                                                 &iod->rspbuf->rsp_dis_assoc;
1833         struct nvmet_fc_tgt_assoc *assoc = NULL;
1834         struct nvmet_fc_ls_iod *oldls = NULL;
1835         unsigned long flags;
1836         int ret = 0;
1837
1838         memset(acc, 0, sizeof(*acc));
1839
1840         ret = nvmefc_vldt_lsreq_discon_assoc(iod->rqstdatalen, rqst);
1841         if (!ret) {
1842                 /* match an active association - takes an assoc ref if !NULL */
1843                 assoc = nvmet_fc_find_target_assoc(tgtport,
1844                                 be64_to_cpu(rqst->associd.association_id));
1845                 iod->assoc = assoc;
1846                 if (!assoc)
1847                         ret = VERR_NO_ASSOC;
1848         }
1849
1850         if (ret || !assoc) {
1851                 dev_err(tgtport->dev,
1852                         "Disconnect LS failed: %s\n",
1853                         validation_errors[ret]);
1854                 iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1855                                 sizeof(*acc), rqst->w0.ls_cmd,
1856                                 (ret == VERR_NO_ASSOC) ?
1857                                         FCNVME_RJT_RC_INV_ASSOC :
1858                                         FCNVME_RJT_RC_LOGIC,
1859                                 FCNVME_RJT_EXP_NONE, 0);
1860                 return true;
1861         }
1862
1863         /* format a response */
1864
1865         iod->lsrsp->rsplen = sizeof(*acc);
1866
1867         nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1868                         fcnvme_lsdesc_len(
1869                                 sizeof(struct fcnvme_ls_disconnect_assoc_acc)),
1870                         FCNVME_LS_DISCONNECT_ASSOC);
1871
1872         /* release get taken in nvmet_fc_find_target_assoc */
1873         nvmet_fc_tgt_a_put(assoc);
1874
1875         /*
1876          * The rules for LS response says the response cannot
1877          * go back until ABTS's have been sent for all outstanding
1878          * I/O and a Disconnect Association LS has been sent.
1879          * So... save off the Disconnect LS to send the response
1880          * later. If there was a prior LS already saved, replace
1881          * it with the newer one and send a can't perform reject
1882          * on the older one.
1883          */
1884         spin_lock_irqsave(&tgtport->lock, flags);
1885         oldls = assoc->rcv_disconn;
1886         assoc->rcv_disconn = iod;
1887         spin_unlock_irqrestore(&tgtport->lock, flags);
1888
1889         nvmet_fc_delete_target_assoc(assoc);
1890
1891         if (oldls) {
1892                 dev_info(tgtport->dev,
1893                         "{%d:%d} Multiple Disconnect Association LS's "
1894                         "received\n",
1895                         tgtport->fc_target_port.port_num, assoc->a_id);
1896                 /* overwrite good response with bogus failure */
1897                 oldls->lsrsp->rsplen = nvme_fc_format_rjt(oldls->rspbuf,
1898                                                 sizeof(*iod->rspbuf),
1899                                                 /* ok to use rqst, LS is same */
1900                                                 rqst->w0.ls_cmd,
1901                                                 FCNVME_RJT_RC_UNAB,
1902                                                 FCNVME_RJT_EXP_NONE, 0);
1903                 nvmet_fc_xmt_ls_rsp(tgtport, oldls);
1904         }
1905
1906         return false;
1907 }
1908
1909
1910 /* *********************** NVME Ctrl Routines **************************** */
1911
1912
1913 static void nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req);
1914
1915 static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops;
1916
1917 static void
1918 nvmet_fc_xmt_ls_rsp_done(struct nvmefc_ls_rsp *lsrsp)
1919 {
1920         struct nvmet_fc_ls_iod *iod = lsrsp->nvme_fc_private;
1921         struct nvmet_fc_tgtport *tgtport = iod->tgtport;
1922
1923         fc_dma_sync_single_for_cpu(tgtport->dev, iod->rspdma,
1924                                 sizeof(*iod->rspbuf), DMA_TO_DEVICE);
1925         nvmet_fc_free_ls_iod(tgtport, iod);
1926         nvmet_fc_tgtport_put(tgtport);
1927 }
1928
1929 static void
1930 nvmet_fc_xmt_ls_rsp(struct nvmet_fc_tgtport *tgtport,
1931                                 struct nvmet_fc_ls_iod *iod)
1932 {
1933         int ret;
1934
1935         fc_dma_sync_single_for_device(tgtport->dev, iod->rspdma,
1936                                   sizeof(*iod->rspbuf), DMA_TO_DEVICE);
1937
1938         ret = tgtport->ops->xmt_ls_rsp(&tgtport->fc_target_port, iod->lsrsp);
1939         if (ret)
1940                 nvmet_fc_xmt_ls_rsp_done(iod->lsrsp);
1941 }
1942
1943 /*
1944  * Actual processing routine for received FC-NVME LS Requests from the LLD
1945  */
1946 static void
1947 nvmet_fc_handle_ls_rqst(struct nvmet_fc_tgtport *tgtport,
1948                         struct nvmet_fc_ls_iod *iod)
1949 {
1950         struct fcnvme_ls_rqst_w0 *w0 = &iod->rqstbuf->rq_cr_assoc.w0;
1951         bool sendrsp = true;
1952
1953         iod->lsrsp->nvme_fc_private = iod;
1954         iod->lsrsp->rspbuf = iod->rspbuf;
1955         iod->lsrsp->rspdma = iod->rspdma;
1956         iod->lsrsp->done = nvmet_fc_xmt_ls_rsp_done;
1957         /* Be preventative. handlers will later set to valid length */
1958         iod->lsrsp->rsplen = 0;
1959
1960         iod->assoc = NULL;
1961
1962         /*
1963          * handlers:
1964          *   parse request input, execute the request, and format the
1965          *   LS response
1966          */
1967         switch (w0->ls_cmd) {
1968         case FCNVME_LS_CREATE_ASSOCIATION:
1969                 /* Creates Association and initial Admin Queue/Connection */
1970                 nvmet_fc_ls_create_association(tgtport, iod);
1971                 break;
1972         case FCNVME_LS_CREATE_CONNECTION:
1973                 /* Creates an IO Queue/Connection */
1974                 nvmet_fc_ls_create_connection(tgtport, iod);
1975                 break;
1976         case FCNVME_LS_DISCONNECT_ASSOC:
1977                 /* Terminate a Queue/Connection or the Association */
1978                 sendrsp = nvmet_fc_ls_disconnect(tgtport, iod);
1979                 break;
1980         default:
1981                 iod->lsrsp->rsplen = nvme_fc_format_rjt(iod->rspbuf,
1982                                 sizeof(*iod->rspbuf), w0->ls_cmd,
1983                                 FCNVME_RJT_RC_INVAL, FCNVME_RJT_EXP_NONE, 0);
1984         }
1985
1986         if (sendrsp)
1987                 nvmet_fc_xmt_ls_rsp(tgtport, iod);
1988 }
1989
1990 /*
1991  * Actual processing routine for received FC-NVME LS Requests from the LLD
1992  */
1993 static void
1994 nvmet_fc_handle_ls_rqst_work(struct work_struct *work)
1995 {
1996         struct nvmet_fc_ls_iod *iod =
1997                 container_of(work, struct nvmet_fc_ls_iod, work);
1998         struct nvmet_fc_tgtport *tgtport = iod->tgtport;
1999
2000         nvmet_fc_handle_ls_rqst(tgtport, iod);
2001 }
2002
2003
2004 /**
2005  * nvmet_fc_rcv_ls_req - transport entry point called by an LLDD
2006  *                       upon the reception of a NVME LS request.
2007  *
2008  * The nvmet-fc layer will copy payload to an internal structure for
2009  * processing.  As such, upon completion of the routine, the LLDD may
2010  * immediately free/reuse the LS request buffer passed in the call.
2011  *
2012  * If this routine returns error, the LLDD should abort the exchange.
2013  *
2014  * @target_port: pointer to the (registered) target port the LS was
2015  *              received on.
2016  * @hosthandle: pointer to the host specific data, gets stored in iod.
2017  * @lsrsp:      pointer to a lsrsp structure to be used to reference
2018  *              the exchange corresponding to the LS.
2019  * @lsreqbuf:   pointer to the buffer containing the LS Request
2020  * @lsreqbuf_len: length, in bytes, of the received LS request
2021  */
2022 int
2023 nvmet_fc_rcv_ls_req(struct nvmet_fc_target_port *target_port,
2024                         void *hosthandle,
2025                         struct nvmefc_ls_rsp *lsrsp,
2026                         void *lsreqbuf, u32 lsreqbuf_len)
2027 {
2028         struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
2029         struct nvmet_fc_ls_iod *iod;
2030         struct fcnvme_ls_rqst_w0 *w0 = (struct fcnvme_ls_rqst_w0 *)lsreqbuf;
2031
2032         if (lsreqbuf_len > sizeof(union nvmefc_ls_requests)) {
2033                 dev_info(tgtport->dev,
2034                         "RCV %s LS failed: payload too large (%d)\n",
2035                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2036                                 nvmefc_ls_names[w0->ls_cmd] : "",
2037                         lsreqbuf_len);
2038                 return -E2BIG;
2039         }
2040
2041         if (!nvmet_fc_tgtport_get(tgtport)) {
2042                 dev_info(tgtport->dev,
2043                         "RCV %s LS failed: target deleting\n",
2044                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2045                                 nvmefc_ls_names[w0->ls_cmd] : "");
2046                 return -ESHUTDOWN;
2047         }
2048
2049         iod = nvmet_fc_alloc_ls_iod(tgtport);
2050         if (!iod) {
2051                 dev_info(tgtport->dev,
2052                         "RCV %s LS failed: context allocation failed\n",
2053                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2054                                 nvmefc_ls_names[w0->ls_cmd] : "");
2055                 nvmet_fc_tgtport_put(tgtport);
2056                 return -ENOENT;
2057         }
2058
2059         iod->lsrsp = lsrsp;
2060         iod->fcpreq = NULL;
2061         memcpy(iod->rqstbuf, lsreqbuf, lsreqbuf_len);
2062         iod->rqstdatalen = lsreqbuf_len;
2063         iod->hosthandle = hosthandle;
2064
2065         queue_work(nvmet_wq, &iod->work);
2066
2067         return 0;
2068 }
2069 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_ls_req);
2070
2071
2072 /*
2073  * **********************
2074  * Start of FCP handling
2075  * **********************
2076  */
2077
2078 static int
2079 nvmet_fc_alloc_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
2080 {
2081         struct scatterlist *sg;
2082         unsigned int nent;
2083
2084         sg = sgl_alloc(fod->req.transfer_len, GFP_KERNEL, &nent);
2085         if (!sg)
2086                 goto out;
2087
2088         fod->data_sg = sg;
2089         fod->data_sg_cnt = nent;
2090         fod->data_sg_cnt = fc_dma_map_sg(fod->tgtport->dev, sg, nent,
2091                                 ((fod->io_dir == NVMET_FCP_WRITE) ?
2092                                         DMA_FROM_DEVICE : DMA_TO_DEVICE));
2093                                 /* note: write from initiator perspective */
2094         fod->next_sg = fod->data_sg;
2095
2096         return 0;
2097
2098 out:
2099         return NVME_SC_INTERNAL;
2100 }
2101
2102 static void
2103 nvmet_fc_free_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
2104 {
2105         if (!fod->data_sg || !fod->data_sg_cnt)
2106                 return;
2107
2108         fc_dma_unmap_sg(fod->tgtport->dev, fod->data_sg, fod->data_sg_cnt,
2109                                 ((fod->io_dir == NVMET_FCP_WRITE) ?
2110                                         DMA_FROM_DEVICE : DMA_TO_DEVICE));
2111         sgl_free(fod->data_sg);
2112         fod->data_sg = NULL;
2113         fod->data_sg_cnt = 0;
2114 }
2115
2116
2117 static bool
2118 queue_90percent_full(struct nvmet_fc_tgt_queue *q, u32 sqhd)
2119 {
2120         u32 sqtail, used;
2121
2122         /* egad, this is ugly. And sqtail is just a best guess */
2123         sqtail = atomic_read(&q->sqtail) % q->sqsize;
2124
2125         used = (sqtail < sqhd) ? (sqtail + q->sqsize - sqhd) : (sqtail - sqhd);
2126         return ((used * 10) >= (((u32)(q->sqsize - 1) * 9)));
2127 }
2128
2129 /*
2130  * Prep RSP payload.
2131  * May be a NVMET_FCOP_RSP or NVMET_FCOP_READDATA_RSP op
2132  */
2133 static void
2134 nvmet_fc_prep_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
2135                                 struct nvmet_fc_fcp_iod *fod)
2136 {
2137         struct nvme_fc_ersp_iu *ersp = &fod->rspiubuf;
2138         struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
2139         struct nvme_completion *cqe = &ersp->cqe;
2140         u32 *cqewd = (u32 *)cqe;
2141         bool send_ersp = false;
2142         u32 rsn, rspcnt, xfr_length;
2143
2144         if (fod->fcpreq->op == NVMET_FCOP_READDATA_RSP)
2145                 xfr_length = fod->req.transfer_len;
2146         else
2147                 xfr_length = fod->offset;
2148
2149         /*
2150          * check to see if we can send a 0's rsp.
2151          *   Note: to send a 0's response, the NVME-FC host transport will
2152          *   recreate the CQE. The host transport knows: sq id, SQHD (last
2153          *   seen in an ersp), and command_id. Thus it will create a
2154          *   zero-filled CQE with those known fields filled in. Transport
2155          *   must send an ersp for any condition where the cqe won't match
2156          *   this.
2157          *
2158          * Here are the FC-NVME mandated cases where we must send an ersp:
2159          *  every N responses, where N=ersp_ratio
2160          *  force fabric commands to send ersp's (not in FC-NVME but good
2161          *    practice)
2162          *  normal cmds: any time status is non-zero, or status is zero
2163          *     but words 0 or 1 are non-zero.
2164          *  the SQ is 90% or more full
2165          *  the cmd is a fused command
2166          *  transferred data length not equal to cmd iu length
2167          */
2168         rspcnt = atomic_inc_return(&fod->queue->zrspcnt);
2169         if (!(rspcnt % fod->queue->ersp_ratio) ||
2170             nvme_is_fabrics((struct nvme_command *) sqe) ||
2171             xfr_length != fod->req.transfer_len ||
2172             (le16_to_cpu(cqe->status) & 0xFFFE) || cqewd[0] || cqewd[1] ||
2173             (sqe->flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND)) ||
2174             queue_90percent_full(fod->queue, le16_to_cpu(cqe->sq_head)))
2175                 send_ersp = true;
2176
2177         /* re-set the fields */
2178         fod->fcpreq->rspaddr = ersp;
2179         fod->fcpreq->rspdma = fod->rspdma;
2180
2181         if (!send_ersp) {
2182                 memset(ersp, 0, NVME_FC_SIZEOF_ZEROS_RSP);
2183                 fod->fcpreq->rsplen = NVME_FC_SIZEOF_ZEROS_RSP;
2184         } else {
2185                 ersp->iu_len = cpu_to_be16(sizeof(*ersp)/sizeof(u32));
2186                 rsn = atomic_inc_return(&fod->queue->rsn);
2187                 ersp->rsn = cpu_to_be32(rsn);
2188                 ersp->xfrd_len = cpu_to_be32(xfr_length);
2189                 fod->fcpreq->rsplen = sizeof(*ersp);
2190         }
2191
2192         fc_dma_sync_single_for_device(tgtport->dev, fod->rspdma,
2193                                   sizeof(fod->rspiubuf), DMA_TO_DEVICE);
2194 }
2195
2196 static void nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq);
2197
2198 static void
2199 nvmet_fc_abort_op(struct nvmet_fc_tgtport *tgtport,
2200                                 struct nvmet_fc_fcp_iod *fod)
2201 {
2202         struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2203
2204         /* data no longer needed */
2205         nvmet_fc_free_tgt_pgs(fod);
2206
2207         /*
2208          * if an ABTS was received or we issued the fcp_abort early
2209          * don't call abort routine again.
2210          */
2211         /* no need to take lock - lock was taken earlier to get here */
2212         if (!fod->aborted)
2213                 tgtport->ops->fcp_abort(&tgtport->fc_target_port, fcpreq);
2214
2215         nvmet_fc_free_fcp_iod(fod->queue, fod);
2216 }
2217
2218 static void
2219 nvmet_fc_xmt_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
2220                                 struct nvmet_fc_fcp_iod *fod)
2221 {
2222         int ret;
2223
2224         fod->fcpreq->op = NVMET_FCOP_RSP;
2225         fod->fcpreq->timeout = 0;
2226
2227         nvmet_fc_prep_fcp_rsp(tgtport, fod);
2228
2229         ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
2230         if (ret)
2231                 nvmet_fc_abort_op(tgtport, fod);
2232 }
2233
2234 static void
2235 nvmet_fc_transfer_fcp_data(struct nvmet_fc_tgtport *tgtport,
2236                                 struct nvmet_fc_fcp_iod *fod, u8 op)
2237 {
2238         struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2239         struct scatterlist *sg = fod->next_sg;
2240         unsigned long flags;
2241         u32 remaininglen = fod->req.transfer_len - fod->offset;
2242         u32 tlen = 0;
2243         int ret;
2244
2245         fcpreq->op = op;
2246         fcpreq->offset = fod->offset;
2247         fcpreq->timeout = NVME_FC_TGTOP_TIMEOUT_SEC;
2248
2249         /*
2250          * for next sequence:
2251          *  break at a sg element boundary
2252          *  attempt to keep sequence length capped at
2253          *    NVMET_FC_MAX_SEQ_LENGTH but allow sequence to
2254          *    be longer if a single sg element is larger
2255          *    than that amount. This is done to avoid creating
2256          *    a new sg list to use for the tgtport api.
2257          */
2258         fcpreq->sg = sg;
2259         fcpreq->sg_cnt = 0;
2260         while (tlen < remaininglen &&
2261                fcpreq->sg_cnt < tgtport->max_sg_cnt &&
2262                tlen + sg_dma_len(sg) < NVMET_FC_MAX_SEQ_LENGTH) {
2263                 fcpreq->sg_cnt++;
2264                 tlen += sg_dma_len(sg);
2265                 sg = sg_next(sg);
2266         }
2267         if (tlen < remaininglen && fcpreq->sg_cnt == 0) {
2268                 fcpreq->sg_cnt++;
2269                 tlen += min_t(u32, sg_dma_len(sg), remaininglen);
2270                 sg = sg_next(sg);
2271         }
2272         if (tlen < remaininglen)
2273                 fod->next_sg = sg;
2274         else
2275                 fod->next_sg = NULL;
2276
2277         fcpreq->transfer_length = tlen;
2278         fcpreq->transferred_length = 0;
2279         fcpreq->fcp_error = 0;
2280         fcpreq->rsplen = 0;
2281
2282         /*
2283          * If the last READDATA request: check if LLDD supports
2284          * combined xfr with response.
2285          */
2286         if ((op == NVMET_FCOP_READDATA) &&
2287             ((fod->offset + fcpreq->transfer_length) == fod->req.transfer_len) &&
2288             (tgtport->ops->target_features & NVMET_FCTGTFEAT_READDATA_RSP)) {
2289                 fcpreq->op = NVMET_FCOP_READDATA_RSP;
2290                 nvmet_fc_prep_fcp_rsp(tgtport, fod);
2291         }
2292
2293         ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
2294         if (ret) {
2295                 /*
2296                  * should be ok to set w/o lock as its in the thread of
2297                  * execution (not an async timer routine) and doesn't
2298                  * contend with any clearing action
2299                  */
2300                 fod->abort = true;
2301
2302                 if (op == NVMET_FCOP_WRITEDATA) {
2303                         spin_lock_irqsave(&fod->flock, flags);
2304                         fod->writedataactive = false;
2305                         spin_unlock_irqrestore(&fod->flock, flags);
2306                         nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2307                 } else /* NVMET_FCOP_READDATA or NVMET_FCOP_READDATA_RSP */ {
2308                         fcpreq->fcp_error = ret;
2309                         fcpreq->transferred_length = 0;
2310                         nvmet_fc_xmt_fcp_op_done(fod->fcpreq);
2311                 }
2312         }
2313 }
2314
2315 static inline bool
2316 __nvmet_fc_fod_op_abort(struct nvmet_fc_fcp_iod *fod, bool abort)
2317 {
2318         struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2319         struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2320
2321         /* if in the middle of an io and we need to tear down */
2322         if (abort) {
2323                 if (fcpreq->op == NVMET_FCOP_WRITEDATA) {
2324                         nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2325                         return true;
2326                 }
2327
2328                 nvmet_fc_abort_op(tgtport, fod);
2329                 return true;
2330         }
2331
2332         return false;
2333 }
2334
2335 /*
2336  * actual done handler for FCP operations when completed by the lldd
2337  */
2338 static void
2339 nvmet_fc_fod_op_done(struct nvmet_fc_fcp_iod *fod)
2340 {
2341         struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2342         struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2343         unsigned long flags;
2344         bool abort;
2345
2346         spin_lock_irqsave(&fod->flock, flags);
2347         abort = fod->abort;
2348         fod->writedataactive = false;
2349         spin_unlock_irqrestore(&fod->flock, flags);
2350
2351         switch (fcpreq->op) {
2352
2353         case NVMET_FCOP_WRITEDATA:
2354                 if (__nvmet_fc_fod_op_abort(fod, abort))
2355                         return;
2356                 if (fcpreq->fcp_error ||
2357                     fcpreq->transferred_length != fcpreq->transfer_length) {
2358                         spin_lock_irqsave(&fod->flock, flags);
2359                         fod->abort = true;
2360                         spin_unlock_irqrestore(&fod->flock, flags);
2361
2362                         nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2363                         return;
2364                 }
2365
2366                 fod->offset += fcpreq->transferred_length;
2367                 if (fod->offset != fod->req.transfer_len) {
2368                         spin_lock_irqsave(&fod->flock, flags);
2369                         fod->writedataactive = true;
2370                         spin_unlock_irqrestore(&fod->flock, flags);
2371
2372                         /* transfer the next chunk */
2373                         nvmet_fc_transfer_fcp_data(tgtport, fod,
2374                                                 NVMET_FCOP_WRITEDATA);
2375                         return;
2376                 }
2377
2378                 /* data transfer complete, resume with nvmet layer */
2379                 fod->req.execute(&fod->req);
2380                 break;
2381
2382         case NVMET_FCOP_READDATA:
2383         case NVMET_FCOP_READDATA_RSP:
2384                 if (__nvmet_fc_fod_op_abort(fod, abort))
2385                         return;
2386                 if (fcpreq->fcp_error ||
2387                     fcpreq->transferred_length != fcpreq->transfer_length) {
2388                         nvmet_fc_abort_op(tgtport, fod);
2389                         return;
2390                 }
2391
2392                 /* success */
2393
2394                 if (fcpreq->op == NVMET_FCOP_READDATA_RSP) {
2395                         /* data no longer needed */
2396                         nvmet_fc_free_tgt_pgs(fod);
2397                         nvmet_fc_free_fcp_iod(fod->queue, fod);
2398                         return;
2399                 }
2400
2401                 fod->offset += fcpreq->transferred_length;
2402                 if (fod->offset != fod->req.transfer_len) {
2403                         /* transfer the next chunk */
2404                         nvmet_fc_transfer_fcp_data(tgtport, fod,
2405                                                 NVMET_FCOP_READDATA);
2406                         return;
2407                 }
2408
2409                 /* data transfer complete, send response */
2410
2411                 /* data no longer needed */
2412                 nvmet_fc_free_tgt_pgs(fod);
2413
2414                 nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2415
2416                 break;
2417
2418         case NVMET_FCOP_RSP:
2419                 if (__nvmet_fc_fod_op_abort(fod, abort))
2420                         return;
2421                 nvmet_fc_free_fcp_iod(fod->queue, fod);
2422                 break;
2423
2424         default:
2425                 break;
2426         }
2427 }
2428
2429 static void
2430 nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq)
2431 {
2432         struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2433
2434         nvmet_fc_fod_op_done(fod);
2435 }
2436
2437 /*
2438  * actual completion handler after execution by the nvmet layer
2439  */
2440 static void
2441 __nvmet_fc_fcp_nvme_cmd_done(struct nvmet_fc_tgtport *tgtport,
2442                         struct nvmet_fc_fcp_iod *fod, int status)
2443 {
2444         struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
2445         struct nvme_completion *cqe = &fod->rspiubuf.cqe;
2446         unsigned long flags;
2447         bool abort;
2448
2449         spin_lock_irqsave(&fod->flock, flags);
2450         abort = fod->abort;
2451         spin_unlock_irqrestore(&fod->flock, flags);
2452
2453         /* if we have a CQE, snoop the last sq_head value */
2454         if (!status)
2455                 fod->queue->sqhd = cqe->sq_head;
2456
2457         if (abort) {
2458                 nvmet_fc_abort_op(tgtport, fod);
2459                 return;
2460         }
2461
2462         /* if an error handling the cmd post initial parsing */
2463         if (status) {
2464                 /* fudge up a failed CQE status for our transport error */
2465                 memset(cqe, 0, sizeof(*cqe));
2466                 cqe->sq_head = fod->queue->sqhd;        /* echo last cqe sqhd */
2467                 cqe->sq_id = cpu_to_le16(fod->queue->qid);
2468                 cqe->command_id = sqe->command_id;
2469                 cqe->status = cpu_to_le16(status);
2470         } else {
2471
2472                 /*
2473                  * try to push the data even if the SQE status is non-zero.
2474                  * There may be a status where data still was intended to
2475                  * be moved
2476                  */
2477                 if ((fod->io_dir == NVMET_FCP_READ) && (fod->data_sg_cnt)) {
2478                         /* push the data over before sending rsp */
2479                         nvmet_fc_transfer_fcp_data(tgtport, fod,
2480                                                 NVMET_FCOP_READDATA);
2481                         return;
2482                 }
2483
2484                 /* writes & no data - fall thru */
2485         }
2486
2487         /* data no longer needed */
2488         nvmet_fc_free_tgt_pgs(fod);
2489
2490         nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2491 }
2492
2493
2494 static void
2495 nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req)
2496 {
2497         struct nvmet_fc_fcp_iod *fod = nvmet_req_to_fod(nvme_req);
2498         struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2499
2500         __nvmet_fc_fcp_nvme_cmd_done(tgtport, fod, 0);
2501 }
2502
2503
2504 /*
2505  * Actual processing routine for received FC-NVME I/O Requests from the LLD
2506  */
2507 static void
2508 nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
2509                         struct nvmet_fc_fcp_iod *fod)
2510 {
2511         struct nvme_fc_cmd_iu *cmdiu = &fod->cmdiubuf;
2512         u32 xfrlen = be32_to_cpu(cmdiu->data_len);
2513         int ret;
2514
2515         /*
2516          * Fused commands are currently not supported in the linux
2517          * implementation.
2518          *
2519          * As such, the implementation of the FC transport does not
2520          * look at the fused commands and order delivery to the upper
2521          * layer until we have both based on csn.
2522          */
2523
2524         fod->fcpreq->done = nvmet_fc_xmt_fcp_op_done;
2525
2526         if (cmdiu->flags & FCNVME_CMD_FLAGS_WRITE) {
2527                 fod->io_dir = NVMET_FCP_WRITE;
2528                 if (!nvme_is_write(&cmdiu->sqe))
2529                         goto transport_error;
2530         } else if (cmdiu->flags & FCNVME_CMD_FLAGS_READ) {
2531                 fod->io_dir = NVMET_FCP_READ;
2532                 if (nvme_is_write(&cmdiu->sqe))
2533                         goto transport_error;
2534         } else {
2535                 fod->io_dir = NVMET_FCP_NODATA;
2536                 if (xfrlen)
2537                         goto transport_error;
2538         }
2539
2540         fod->req.cmd = &fod->cmdiubuf.sqe;
2541         fod->req.cqe = &fod->rspiubuf.cqe;
2542         if (tgtport->pe)
2543                 fod->req.port = tgtport->pe->port;
2544
2545         /* clear any response payload */
2546         memset(&fod->rspiubuf, 0, sizeof(fod->rspiubuf));
2547
2548         fod->data_sg = NULL;
2549         fod->data_sg_cnt = 0;
2550
2551         ret = nvmet_req_init(&fod->req,
2552                                 &fod->queue->nvme_cq,
2553                                 &fod->queue->nvme_sq,
2554                                 &nvmet_fc_tgt_fcp_ops);
2555         if (!ret) {
2556                 /* bad SQE content or invalid ctrl state */
2557                 /* nvmet layer has already called op done to send rsp. */
2558                 return;
2559         }
2560
2561         fod->req.transfer_len = xfrlen;
2562
2563         /* keep a running counter of tail position */
2564         atomic_inc(&fod->queue->sqtail);
2565
2566         if (fod->req.transfer_len) {
2567                 ret = nvmet_fc_alloc_tgt_pgs(fod);
2568                 if (ret) {
2569                         nvmet_req_complete(&fod->req, ret);
2570                         return;
2571                 }
2572         }
2573         fod->req.sg = fod->data_sg;
2574         fod->req.sg_cnt = fod->data_sg_cnt;
2575         fod->offset = 0;
2576
2577         if (fod->io_dir == NVMET_FCP_WRITE) {
2578                 /* pull the data over before invoking nvmet layer */
2579                 nvmet_fc_transfer_fcp_data(tgtport, fod, NVMET_FCOP_WRITEDATA);
2580                 return;
2581         }
2582
2583         /*
2584          * Reads or no data:
2585          *
2586          * can invoke the nvmet_layer now. If read data, cmd completion will
2587          * push the data
2588          */
2589         fod->req.execute(&fod->req);
2590         return;
2591
2592 transport_error:
2593         nvmet_fc_abort_op(tgtport, fod);
2594 }
2595
2596 /**
2597  * nvmet_fc_rcv_fcp_req - transport entry point called by an LLDD
2598  *                       upon the reception of a NVME FCP CMD IU.
2599  *
2600  * Pass a FC-NVME FCP CMD IU received from the FC link to the nvmet-fc
2601  * layer for processing.
2602  *
2603  * The nvmet_fc layer allocates a local job structure (struct
2604  * nvmet_fc_fcp_iod) from the queue for the io and copies the
2605  * CMD IU buffer to the job structure. As such, on a successful
2606  * completion (returns 0), the LLDD may immediately free/reuse
2607  * the CMD IU buffer passed in the call.
2608  *
2609  * However, in some circumstances, due to the packetized nature of FC
2610  * and the api of the FC LLDD which may issue a hw command to send the
2611  * response, but the LLDD may not get the hw completion for that command
2612  * and upcall the nvmet_fc layer before a new command may be
2613  * asynchronously received - its possible for a command to be received
2614  * before the LLDD and nvmet_fc have recycled the job structure. It gives
2615  * the appearance of more commands received than fits in the sq.
2616  * To alleviate this scenario, a temporary queue is maintained in the
2617  * transport for pending LLDD requests waiting for a queue job structure.
2618  * In these "overrun" cases, a temporary queue element is allocated
2619  * the LLDD request and CMD iu buffer information remembered, and the
2620  * routine returns a -EOVERFLOW status. Subsequently, when a queue job
2621  * structure is freed, it is immediately reallocated for anything on the
2622  * pending request list. The LLDDs defer_rcv() callback is called,
2623  * informing the LLDD that it may reuse the CMD IU buffer, and the io
2624  * is then started normally with the transport.
2625  *
2626  * The LLDD, when receiving an -EOVERFLOW completion status, is to treat
2627  * the completion as successful but must not reuse the CMD IU buffer
2628  * until the LLDD's defer_rcv() callback has been called for the
2629  * corresponding struct nvmefc_tgt_fcp_req pointer.
2630  *
2631  * If there is any other condition in which an error occurs, the
2632  * transport will return a non-zero status indicating the error.
2633  * In all cases other than -EOVERFLOW, the transport has not accepted the
2634  * request and the LLDD should abort the exchange.
2635  *
2636  * @target_port: pointer to the (registered) target port the FCP CMD IU
2637  *              was received on.
2638  * @fcpreq:     pointer to a fcpreq request structure to be used to reference
2639  *              the exchange corresponding to the FCP Exchange.
2640  * @cmdiubuf:   pointer to the buffer containing the FCP CMD IU
2641  * @cmdiubuf_len: length, in bytes, of the received FCP CMD IU
2642  */
2643 int
2644 nvmet_fc_rcv_fcp_req(struct nvmet_fc_target_port *target_port,
2645                         struct nvmefc_tgt_fcp_req *fcpreq,
2646                         void *cmdiubuf, u32 cmdiubuf_len)
2647 {
2648         struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
2649         struct nvme_fc_cmd_iu *cmdiu = cmdiubuf;
2650         struct nvmet_fc_tgt_queue *queue;
2651         struct nvmet_fc_fcp_iod *fod;
2652         struct nvmet_fc_defer_fcp_req *deferfcp;
2653         unsigned long flags;
2654
2655         /* validate iu, so the connection id can be used to find the queue */
2656         if ((cmdiubuf_len != sizeof(*cmdiu)) ||
2657                         (cmdiu->format_id != NVME_CMD_FORMAT_ID) ||
2658                         (cmdiu->fc_id != NVME_CMD_FC_ID) ||
2659                         (be16_to_cpu(cmdiu->iu_len) != (sizeof(*cmdiu)/4)))
2660                 return -EIO;
2661
2662         queue = nvmet_fc_find_target_queue(tgtport,
2663                                 be64_to_cpu(cmdiu->connection_id));
2664         if (!queue)
2665                 return -ENOTCONN;
2666
2667         /*
2668          * note: reference taken by find_target_queue
2669          * After successful fod allocation, the fod will inherit the
2670          * ownership of that reference and will remove the reference
2671          * when the fod is freed.
2672          */
2673
2674         spin_lock_irqsave(&queue->qlock, flags);
2675
2676         fod = nvmet_fc_alloc_fcp_iod(queue);
2677         if (fod) {
2678                 spin_unlock_irqrestore(&queue->qlock, flags);
2679
2680                 fcpreq->nvmet_fc_private = fod;
2681                 fod->fcpreq = fcpreq;
2682
2683                 memcpy(&fod->cmdiubuf, cmdiubuf, cmdiubuf_len);
2684
2685                 nvmet_fc_queue_fcp_req(tgtport, queue, fcpreq);
2686
2687                 return 0;
2688         }
2689
2690         if (!tgtport->ops->defer_rcv) {
2691                 spin_unlock_irqrestore(&queue->qlock, flags);
2692                 /* release the queue lookup reference */
2693                 nvmet_fc_tgt_q_put(queue);
2694                 return -ENOENT;
2695         }
2696
2697         deferfcp = list_first_entry_or_null(&queue->avail_defer_list,
2698                         struct nvmet_fc_defer_fcp_req, req_list);
2699         if (deferfcp) {
2700                 /* Just re-use one that was previously allocated */
2701                 list_del(&deferfcp->req_list);
2702         } else {
2703                 spin_unlock_irqrestore(&queue->qlock, flags);
2704
2705                 /* Now we need to dynamically allocate one */
2706                 deferfcp = kmalloc(sizeof(*deferfcp), GFP_KERNEL);
2707                 if (!deferfcp) {
2708                         /* release the queue lookup reference */
2709                         nvmet_fc_tgt_q_put(queue);
2710                         return -ENOMEM;
2711                 }
2712                 spin_lock_irqsave(&queue->qlock, flags);
2713         }
2714
2715         /* For now, use rspaddr / rsplen to save payload information */
2716         fcpreq->rspaddr = cmdiubuf;
2717         fcpreq->rsplen  = cmdiubuf_len;
2718         deferfcp->fcp_req = fcpreq;
2719
2720         /* defer processing till a fod becomes available */
2721         list_add_tail(&deferfcp->req_list, &queue->pending_cmd_list);
2722
2723         /* NOTE: the queue lookup reference is still valid */
2724
2725         spin_unlock_irqrestore(&queue->qlock, flags);
2726
2727         return -EOVERFLOW;
2728 }
2729 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_req);
2730
2731 /**
2732  * nvmet_fc_rcv_fcp_abort - transport entry point called by an LLDD
2733  *                       upon the reception of an ABTS for a FCP command
2734  *
2735  * Notify the transport that an ABTS has been received for a FCP command
2736  * that had been given to the transport via nvmet_fc_rcv_fcp_req(). The
2737  * LLDD believes the command is still being worked on
2738  * (template_ops->fcp_req_release() has not been called).
2739  *
2740  * The transport will wait for any outstanding work (an op to the LLDD,
2741  * which the lldd should complete with error due to the ABTS; or the
2742  * completion from the nvmet layer of the nvme command), then will
2743  * stop processing and call the nvmet_fc_rcv_fcp_req() callback to
2744  * return the i/o context to the LLDD.  The LLDD may send the BA_ACC
2745  * to the ABTS either after return from this function (assuming any
2746  * outstanding op work has been terminated) or upon the callback being
2747  * called.
2748  *
2749  * @target_port: pointer to the (registered) target port the FCP CMD IU
2750  *              was received on.
2751  * @fcpreq:     pointer to the fcpreq request structure that corresponds
2752  *              to the exchange that received the ABTS.
2753  */
2754 void
2755 nvmet_fc_rcv_fcp_abort(struct nvmet_fc_target_port *target_port,
2756                         struct nvmefc_tgt_fcp_req *fcpreq)
2757 {
2758         struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2759         struct nvmet_fc_tgt_queue *queue;
2760         unsigned long flags;
2761
2762         if (!fod || fod->fcpreq != fcpreq)
2763                 /* job appears to have already completed, ignore abort */
2764                 return;
2765
2766         queue = fod->queue;
2767
2768         spin_lock_irqsave(&queue->qlock, flags);
2769         if (fod->active) {
2770                 /*
2771                  * mark as abort. The abort handler, invoked upon completion
2772                  * of any work, will detect the aborted status and do the
2773                  * callback.
2774                  */
2775                 spin_lock(&fod->flock);
2776                 fod->abort = true;
2777                 fod->aborted = true;
2778                 spin_unlock(&fod->flock);
2779         }
2780         spin_unlock_irqrestore(&queue->qlock, flags);
2781 }
2782 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_abort);
2783
2784
2785 struct nvmet_fc_traddr {
2786         u64     nn;
2787         u64     pn;
2788 };
2789
2790 static int
2791 __nvme_fc_parse_u64(substring_t *sstr, u64 *val)
2792 {
2793         u64 token64;
2794
2795         if (match_u64(sstr, &token64))
2796                 return -EINVAL;
2797         *val = token64;
2798
2799         return 0;
2800 }
2801
2802 /*
2803  * This routine validates and extracts the WWN's from the TRADDR string.
2804  * As kernel parsers need the 0x to determine number base, universally
2805  * build string to parse with 0x prefix before parsing name strings.
2806  */
2807 static int
2808 nvme_fc_parse_traddr(struct nvmet_fc_traddr *traddr, char *buf, size_t blen)
2809 {
2810         char name[2 + NVME_FC_TRADDR_HEXNAMELEN + 1];
2811         substring_t wwn = { name, &name[sizeof(name)-1] };
2812         int nnoffset, pnoffset;
2813
2814         /* validate if string is one of the 2 allowed formats */
2815         if (strnlen(buf, blen) == NVME_FC_TRADDR_MAXLENGTH &&
2816                         !strncmp(buf, "nn-0x", NVME_FC_TRADDR_OXNNLEN) &&
2817                         !strncmp(&buf[NVME_FC_TRADDR_MAX_PN_OFFSET],
2818                                 "pn-0x", NVME_FC_TRADDR_OXNNLEN)) {
2819                 nnoffset = NVME_FC_TRADDR_OXNNLEN;
2820                 pnoffset = NVME_FC_TRADDR_MAX_PN_OFFSET +
2821                                                 NVME_FC_TRADDR_OXNNLEN;
2822         } else if ((strnlen(buf, blen) == NVME_FC_TRADDR_MINLENGTH &&
2823                         !strncmp(buf, "nn-", NVME_FC_TRADDR_NNLEN) &&
2824                         !strncmp(&buf[NVME_FC_TRADDR_MIN_PN_OFFSET],
2825                                 "pn-", NVME_FC_TRADDR_NNLEN))) {
2826                 nnoffset = NVME_FC_TRADDR_NNLEN;
2827                 pnoffset = NVME_FC_TRADDR_MIN_PN_OFFSET + NVME_FC_TRADDR_NNLEN;
2828         } else
2829                 goto out_einval;
2830
2831         name[0] = '0';
2832         name[1] = 'x';
2833         name[2 + NVME_FC_TRADDR_HEXNAMELEN] = 0;
2834
2835         memcpy(&name[2], &buf[nnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2836         if (__nvme_fc_parse_u64(&wwn, &traddr->nn))
2837                 goto out_einval;
2838
2839         memcpy(&name[2], &buf[pnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2840         if (__nvme_fc_parse_u64(&wwn, &traddr->pn))
2841                 goto out_einval;
2842
2843         return 0;
2844
2845 out_einval:
2846         pr_warn("%s: bad traddr string\n", __func__);
2847         return -EINVAL;
2848 }
2849
2850 static int
2851 nvmet_fc_add_port(struct nvmet_port *port)
2852 {
2853         struct nvmet_fc_tgtport *tgtport;
2854         struct nvmet_fc_port_entry *pe;
2855         struct nvmet_fc_traddr traddr = { 0L, 0L };
2856         unsigned long flags;
2857         int ret;
2858
2859         /* validate the address info */
2860         if ((port->disc_addr.trtype != NVMF_TRTYPE_FC) ||
2861             (port->disc_addr.adrfam != NVMF_ADDR_FAMILY_FC))
2862                 return -EINVAL;
2863
2864         /* map the traddr address info to a target port */
2865
2866         ret = nvme_fc_parse_traddr(&traddr, port->disc_addr.traddr,
2867                         sizeof(port->disc_addr.traddr));
2868         if (ret)
2869                 return ret;
2870
2871         pe = kzalloc(sizeof(*pe), GFP_KERNEL);
2872         if (!pe)
2873                 return -ENOMEM;
2874
2875         ret = -ENXIO;
2876         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
2877         list_for_each_entry(tgtport, &nvmet_fc_target_list, tgt_list) {
2878                 if ((tgtport->fc_target_port.node_name == traddr.nn) &&
2879                     (tgtport->fc_target_port.port_name == traddr.pn)) {
2880                         /* a FC port can only be 1 nvmet port id */
2881                         if (!tgtport->pe) {
2882                                 nvmet_fc_portentry_bind(tgtport, pe, port);
2883                                 ret = 0;
2884                         } else
2885                                 ret = -EALREADY;
2886                         break;
2887                 }
2888         }
2889         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
2890
2891         if (ret)
2892                 kfree(pe);
2893
2894         return ret;
2895 }
2896
2897 static void
2898 nvmet_fc_remove_port(struct nvmet_port *port)
2899 {
2900         struct nvmet_fc_port_entry *pe = port->priv;
2901
2902         nvmet_fc_portentry_unbind(pe);
2903
2904         kfree(pe);
2905 }
2906
2907 static void
2908 nvmet_fc_discovery_chg(struct nvmet_port *port)
2909 {
2910         struct nvmet_fc_port_entry *pe = port->priv;
2911         struct nvmet_fc_tgtport *tgtport = pe->tgtport;
2912
2913         if (tgtport && tgtport->ops->discovery_event)
2914                 tgtport->ops->discovery_event(&tgtport->fc_target_port);
2915 }
2916
2917 static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops = {
2918         .owner                  = THIS_MODULE,
2919         .type                   = NVMF_TRTYPE_FC,
2920         .msdbd                  = 1,
2921         .add_port               = nvmet_fc_add_port,
2922         .remove_port            = nvmet_fc_remove_port,
2923         .queue_response         = nvmet_fc_fcp_nvme_cmd_done,
2924         .delete_ctrl            = nvmet_fc_delete_ctrl,
2925         .discovery_chg          = nvmet_fc_discovery_chg,
2926 };
2927
2928 static int __init nvmet_fc_init_module(void)
2929 {
2930         return nvmet_register_transport(&nvmet_fc_tgt_fcp_ops);
2931 }
2932
2933 static void __exit nvmet_fc_exit_module(void)
2934 {
2935         /* sanity check - all lports should be removed */
2936         if (!list_empty(&nvmet_fc_target_list))
2937                 pr_warn("%s: targetport list not empty\n", __func__);
2938
2939         nvmet_unregister_transport(&nvmet_fc_tgt_fcp_ops);
2940
2941         ida_destroy(&nvmet_fc_tgtport_cnt);
2942 }
2943
2944 module_init(nvmet_fc_init_module);
2945 module_exit(nvmet_fc_exit_module);
2946
2947 MODULE_LICENSE("GPL v2");