GNU Linux-libre 5.10.217-gnu1
[releases.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 nvmet_fc_fcp_iod         fod[];          /* array of fcp_iods */
149 } __aligned(sizeof(unsigned long long));
150
151 struct nvmet_fc_hostport {
152         struct nvmet_fc_tgtport         *tgtport;
153         void                            *hosthandle;
154         struct list_head                host_list;
155         struct kref                     ref;
156         u8                              invalid;
157 };
158
159 struct nvmet_fc_tgt_assoc {
160         u64                             association_id;
161         u32                             a_id;
162         atomic_t                        terminating;
163         struct nvmet_fc_tgtport         *tgtport;
164         struct nvmet_fc_hostport        *hostport;
165         struct nvmet_fc_ls_iod          *rcv_disconn;
166         struct list_head                a_list;
167         struct nvmet_fc_tgt_queue       *queues[NVMET_NR_QUEUES + 1];
168         struct kref                     ref;
169         struct work_struct              del_work;
170 };
171
172
173 static inline int
174 nvmet_fc_iodnum(struct nvmet_fc_ls_iod *iodptr)
175 {
176         return (iodptr - iodptr->tgtport->iod);
177 }
178
179 static inline int
180 nvmet_fc_fodnum(struct nvmet_fc_fcp_iod *fodptr)
181 {
182         return (fodptr - fodptr->queue->fod);
183 }
184
185
186 /*
187  * Association and Connection IDs:
188  *
189  * Association ID will have random number in upper 6 bytes and zero
190  *   in lower 2 bytes
191  *
192  * Connection IDs will be Association ID with QID or'd in lower 2 bytes
193  *
194  * note: Association ID = Connection ID for queue 0
195  */
196 #define BYTES_FOR_QID                   sizeof(u16)
197 #define BYTES_FOR_QID_SHIFT             (BYTES_FOR_QID * 8)
198 #define NVMET_FC_QUEUEID_MASK           ((u64)((1 << BYTES_FOR_QID_SHIFT) - 1))
199
200 static inline u64
201 nvmet_fc_makeconnid(struct nvmet_fc_tgt_assoc *assoc, u16 qid)
202 {
203         return (assoc->association_id | qid);
204 }
205
206 static inline u64
207 nvmet_fc_getassociationid(u64 connectionid)
208 {
209         return connectionid & ~NVMET_FC_QUEUEID_MASK;
210 }
211
212 static inline u16
213 nvmet_fc_getqueueid(u64 connectionid)
214 {
215         return (u16)(connectionid & NVMET_FC_QUEUEID_MASK);
216 }
217
218 static inline struct nvmet_fc_tgtport *
219 targetport_to_tgtport(struct nvmet_fc_target_port *targetport)
220 {
221         return container_of(targetport, struct nvmet_fc_tgtport,
222                                  fc_target_port);
223 }
224
225 static inline struct nvmet_fc_fcp_iod *
226 nvmet_req_to_fod(struct nvmet_req *nvme_req)
227 {
228         return container_of(nvme_req, struct nvmet_fc_fcp_iod, req);
229 }
230
231
232 /* *************************** Globals **************************** */
233
234
235 static DEFINE_SPINLOCK(nvmet_fc_tgtlock);
236
237 static LIST_HEAD(nvmet_fc_target_list);
238 static DEFINE_IDA(nvmet_fc_tgtport_cnt);
239 static LIST_HEAD(nvmet_fc_portentry_list);
240
241
242 static void nvmet_fc_handle_ls_rqst_work(struct work_struct *work);
243 static void nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work);
244 static void nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc);
245 static int nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc);
246 static void nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue);
247 static int nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue);
248 static void nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport);
249 static int nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport);
250 static void nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
251                                         struct nvmet_fc_fcp_iod *fod);
252 static void nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc);
253 static void nvmet_fc_xmt_ls_rsp(struct nvmet_fc_tgtport *tgtport,
254                                 struct nvmet_fc_ls_iod *iod);
255
256
257 /* *********************** FC-NVME DMA Handling **************************** */
258
259 /*
260  * The fcloop device passes in a NULL device pointer. Real LLD's will
261  * pass in a valid device pointer. If NULL is passed to the dma mapping
262  * routines, depending on the platform, it may or may not succeed, and
263  * may crash.
264  *
265  * As such:
266  * Wrapper all the dma routines and check the dev pointer.
267  *
268  * If simple mappings (return just a dma address, we'll noop them,
269  * returning a dma address of 0.
270  *
271  * On more complex mappings (dma_map_sg), a pseudo routine fills
272  * in the scatter list, setting all dma addresses to 0.
273  */
274
275 static inline dma_addr_t
276 fc_dma_map_single(struct device *dev, void *ptr, size_t size,
277                 enum dma_data_direction dir)
278 {
279         return dev ? dma_map_single(dev, ptr, size, dir) : (dma_addr_t)0L;
280 }
281
282 static inline int
283 fc_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
284 {
285         return dev ? dma_mapping_error(dev, dma_addr) : 0;
286 }
287
288 static inline void
289 fc_dma_unmap_single(struct device *dev, dma_addr_t addr, size_t size,
290         enum dma_data_direction dir)
291 {
292         if (dev)
293                 dma_unmap_single(dev, addr, size, dir);
294 }
295
296 static inline void
297 fc_dma_sync_single_for_cpu(struct device *dev, dma_addr_t addr, size_t size,
298                 enum dma_data_direction dir)
299 {
300         if (dev)
301                 dma_sync_single_for_cpu(dev, addr, size, dir);
302 }
303
304 static inline void
305 fc_dma_sync_single_for_device(struct device *dev, dma_addr_t addr, size_t size,
306                 enum dma_data_direction dir)
307 {
308         if (dev)
309                 dma_sync_single_for_device(dev, addr, size, dir);
310 }
311
312 /* pseudo dma_map_sg call */
313 static int
314 fc_map_sg(struct scatterlist *sg, int nents)
315 {
316         struct scatterlist *s;
317         int i;
318
319         WARN_ON(nents == 0 || sg[0].length == 0);
320
321         for_each_sg(sg, s, nents, i) {
322                 s->dma_address = 0L;
323 #ifdef CONFIG_NEED_SG_DMA_LENGTH
324                 s->dma_length = s->length;
325 #endif
326         }
327         return nents;
328 }
329
330 static inline int
331 fc_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
332                 enum dma_data_direction dir)
333 {
334         return dev ? dma_map_sg(dev, sg, nents, dir) : fc_map_sg(sg, nents);
335 }
336
337 static inline void
338 fc_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
339                 enum dma_data_direction dir)
340 {
341         if (dev)
342                 dma_unmap_sg(dev, sg, nents, dir);
343 }
344
345
346 /* ********************** FC-NVME LS XMT Handling ************************* */
347
348
349 static void
350 __nvmet_fc_finish_ls_req(struct nvmet_fc_ls_req_op *lsop)
351 {
352         struct nvmet_fc_tgtport *tgtport = lsop->tgtport;
353         struct nvmefc_ls_req *lsreq = &lsop->ls_req;
354         unsigned long flags;
355
356         spin_lock_irqsave(&tgtport->lock, flags);
357
358         if (!lsop->req_queued) {
359                 spin_unlock_irqrestore(&tgtport->lock, flags);
360                 goto out_puttgtport;
361         }
362
363         list_del(&lsop->lsreq_list);
364
365         lsop->req_queued = false;
366
367         spin_unlock_irqrestore(&tgtport->lock, flags);
368
369         fc_dma_unmap_single(tgtport->dev, lsreq->rqstdma,
370                                   (lsreq->rqstlen + lsreq->rsplen),
371                                   DMA_BIDIRECTIONAL);
372
373 out_puttgtport:
374         nvmet_fc_tgtport_put(tgtport);
375 }
376
377 static int
378 __nvmet_fc_send_ls_req(struct nvmet_fc_tgtport *tgtport,
379                 struct nvmet_fc_ls_req_op *lsop,
380                 void (*done)(struct nvmefc_ls_req *req, int status))
381 {
382         struct nvmefc_ls_req *lsreq = &lsop->ls_req;
383         unsigned long flags;
384         int ret = 0;
385
386         if (!tgtport->ops->ls_req)
387                 return -EOPNOTSUPP;
388
389         if (!nvmet_fc_tgtport_get(tgtport))
390                 return -ESHUTDOWN;
391
392         lsreq->done = done;
393         lsop->req_queued = false;
394         INIT_LIST_HEAD(&lsop->lsreq_list);
395
396         lsreq->rqstdma = fc_dma_map_single(tgtport->dev, lsreq->rqstaddr,
397                                   lsreq->rqstlen + lsreq->rsplen,
398                                   DMA_BIDIRECTIONAL);
399         if (fc_dma_mapping_error(tgtport->dev, lsreq->rqstdma)) {
400                 ret = -EFAULT;
401                 goto out_puttgtport;
402         }
403         lsreq->rspdma = lsreq->rqstdma + lsreq->rqstlen;
404
405         spin_lock_irqsave(&tgtport->lock, flags);
406
407         list_add_tail(&lsop->lsreq_list, &tgtport->ls_req_list);
408
409         lsop->req_queued = true;
410
411         spin_unlock_irqrestore(&tgtport->lock, flags);
412
413         ret = tgtport->ops->ls_req(&tgtport->fc_target_port, lsop->hosthandle,
414                                    lsreq);
415         if (ret)
416                 goto out_unlink;
417
418         return 0;
419
420 out_unlink:
421         lsop->ls_error = ret;
422         spin_lock_irqsave(&tgtport->lock, flags);
423         lsop->req_queued = false;
424         list_del(&lsop->lsreq_list);
425         spin_unlock_irqrestore(&tgtport->lock, flags);
426         fc_dma_unmap_single(tgtport->dev, lsreq->rqstdma,
427                                   (lsreq->rqstlen + lsreq->rsplen),
428                                   DMA_BIDIRECTIONAL);
429 out_puttgtport:
430         nvmet_fc_tgtport_put(tgtport);
431
432         return ret;
433 }
434
435 static int
436 nvmet_fc_send_ls_req_async(struct nvmet_fc_tgtport *tgtport,
437                 struct nvmet_fc_ls_req_op *lsop,
438                 void (*done)(struct nvmefc_ls_req *req, int status))
439 {
440         /* don't wait for completion */
441
442         return __nvmet_fc_send_ls_req(tgtport, lsop, done);
443 }
444
445 static void
446 nvmet_fc_disconnect_assoc_done(struct nvmefc_ls_req *lsreq, int status)
447 {
448         struct nvmet_fc_ls_req_op *lsop =
449                 container_of(lsreq, struct nvmet_fc_ls_req_op, ls_req);
450
451         __nvmet_fc_finish_ls_req(lsop);
452
453         /* fc-nvme target doesn't care about success or failure of cmd */
454
455         kfree(lsop);
456 }
457
458 /*
459  * This routine sends a FC-NVME LS to disconnect (aka terminate)
460  * the FC-NVME Association.  Terminating the association also
461  * terminates the FC-NVME connections (per queue, both admin and io
462  * queues) that are part of the association. E.g. things are torn
463  * down, and the related FC-NVME Association ID and Connection IDs
464  * become invalid.
465  *
466  * The behavior of the fc-nvme target is such that it's
467  * understanding of the association and connections will implicitly
468  * be torn down. The action is implicit as it may be due to a loss of
469  * connectivity with the fc-nvme host, so the target may never get a
470  * response even if it tried.  As such, the action of this routine
471  * is to asynchronously send the LS, ignore any results of the LS, and
472  * continue on with terminating the association. If the fc-nvme host
473  * is present and receives the LS, it too can tear down.
474  */
475 static void
476 nvmet_fc_xmt_disconnect_assoc(struct nvmet_fc_tgt_assoc *assoc)
477 {
478         struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
479         struct fcnvme_ls_disconnect_assoc_rqst *discon_rqst;
480         struct fcnvme_ls_disconnect_assoc_acc *discon_acc;
481         struct nvmet_fc_ls_req_op *lsop;
482         struct nvmefc_ls_req *lsreq;
483         int ret;
484
485         /*
486          * If ls_req is NULL or no hosthandle, it's an older lldd and no
487          * message is normal. Otherwise, send unless the hostport has
488          * already been invalidated by the lldd.
489          */
490         if (!tgtport->ops->ls_req || !assoc->hostport ||
491             assoc->hostport->invalid)
492                 return;
493
494         lsop = kzalloc((sizeof(*lsop) +
495                         sizeof(*discon_rqst) + sizeof(*discon_acc) +
496                         tgtport->ops->lsrqst_priv_sz), GFP_KERNEL);
497         if (!lsop) {
498                 dev_info(tgtport->dev,
499                         "{%d:%d} send Disconnect Association failed: ENOMEM\n",
500                         tgtport->fc_target_port.port_num, assoc->a_id);
501                 return;
502         }
503
504         discon_rqst = (struct fcnvme_ls_disconnect_assoc_rqst *)&lsop[1];
505         discon_acc = (struct fcnvme_ls_disconnect_assoc_acc *)&discon_rqst[1];
506         lsreq = &lsop->ls_req;
507         if (tgtport->ops->lsrqst_priv_sz)
508                 lsreq->private = (void *)&discon_acc[1];
509         else
510                 lsreq->private = NULL;
511
512         lsop->tgtport = tgtport;
513         lsop->hosthandle = assoc->hostport->hosthandle;
514
515         nvmefc_fmt_lsreq_discon_assoc(lsreq, discon_rqst, discon_acc,
516                                 assoc->association_id);
517
518         ret = nvmet_fc_send_ls_req_async(tgtport, lsop,
519                                 nvmet_fc_disconnect_assoc_done);
520         if (ret) {
521                 dev_info(tgtport->dev,
522                         "{%d:%d} XMT Disconnect Association failed: %d\n",
523                         tgtport->fc_target_port.port_num, assoc->a_id, ret);
524                 kfree(lsop);
525         }
526 }
527
528
529 /* *********************** FC-NVME Port Management ************************ */
530
531
532 static int
533 nvmet_fc_alloc_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
534 {
535         struct nvmet_fc_ls_iod *iod;
536         int i;
537
538         iod = kcalloc(NVMET_LS_CTX_COUNT, sizeof(struct nvmet_fc_ls_iod),
539                         GFP_KERNEL);
540         if (!iod)
541                 return -ENOMEM;
542
543         tgtport->iod = iod;
544
545         for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
546                 INIT_WORK(&iod->work, nvmet_fc_handle_ls_rqst_work);
547                 iod->tgtport = tgtport;
548                 list_add_tail(&iod->ls_rcv_list, &tgtport->ls_rcv_list);
549
550                 iod->rqstbuf = kzalloc(sizeof(union nvmefc_ls_requests) +
551                                        sizeof(union nvmefc_ls_responses),
552                                        GFP_KERNEL);
553                 if (!iod->rqstbuf)
554                         goto out_fail;
555
556                 iod->rspbuf = (union nvmefc_ls_responses *)&iod->rqstbuf[1];
557
558                 iod->rspdma = fc_dma_map_single(tgtport->dev, iod->rspbuf,
559                                                 sizeof(*iod->rspbuf),
560                                                 DMA_TO_DEVICE);
561                 if (fc_dma_mapping_error(tgtport->dev, iod->rspdma))
562                         goto out_fail;
563         }
564
565         return 0;
566
567 out_fail:
568         kfree(iod->rqstbuf);
569         list_del(&iod->ls_rcv_list);
570         for (iod--, i--; i >= 0; iod--, i--) {
571                 fc_dma_unmap_single(tgtport->dev, iod->rspdma,
572                                 sizeof(*iod->rspbuf), DMA_TO_DEVICE);
573                 kfree(iod->rqstbuf);
574                 list_del(&iod->ls_rcv_list);
575         }
576
577         kfree(iod);
578
579         return -EFAULT;
580 }
581
582 static void
583 nvmet_fc_free_ls_iodlist(struct nvmet_fc_tgtport *tgtport)
584 {
585         struct nvmet_fc_ls_iod *iod = tgtport->iod;
586         int i;
587
588         for (i = 0; i < NVMET_LS_CTX_COUNT; iod++, i++) {
589                 fc_dma_unmap_single(tgtport->dev,
590                                 iod->rspdma, sizeof(*iod->rspbuf),
591                                 DMA_TO_DEVICE);
592                 kfree(iod->rqstbuf);
593                 list_del(&iod->ls_rcv_list);
594         }
595         kfree(tgtport->iod);
596 }
597
598 static struct nvmet_fc_ls_iod *
599 nvmet_fc_alloc_ls_iod(struct nvmet_fc_tgtport *tgtport)
600 {
601         struct nvmet_fc_ls_iod *iod;
602         unsigned long flags;
603
604         spin_lock_irqsave(&tgtport->lock, flags);
605         iod = list_first_entry_or_null(&tgtport->ls_rcv_list,
606                                         struct nvmet_fc_ls_iod, ls_rcv_list);
607         if (iod)
608                 list_move_tail(&iod->ls_rcv_list, &tgtport->ls_busylist);
609         spin_unlock_irqrestore(&tgtport->lock, flags);
610         return iod;
611 }
612
613
614 static void
615 nvmet_fc_free_ls_iod(struct nvmet_fc_tgtport *tgtport,
616                         struct nvmet_fc_ls_iod *iod)
617 {
618         unsigned long flags;
619
620         spin_lock_irqsave(&tgtport->lock, flags);
621         list_move(&iod->ls_rcv_list, &tgtport->ls_rcv_list);
622         spin_unlock_irqrestore(&tgtport->lock, flags);
623 }
624
625 static void
626 nvmet_fc_prep_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
627                                 struct nvmet_fc_tgt_queue *queue)
628 {
629         struct nvmet_fc_fcp_iod *fod = queue->fod;
630         int i;
631
632         for (i = 0; i < queue->sqsize; fod++, i++) {
633                 INIT_WORK(&fod->defer_work, nvmet_fc_fcp_rqst_op_defer_work);
634                 fod->tgtport = tgtport;
635                 fod->queue = queue;
636                 fod->active = false;
637                 fod->abort = false;
638                 fod->aborted = false;
639                 fod->fcpreq = NULL;
640                 list_add_tail(&fod->fcp_list, &queue->fod_list);
641                 spin_lock_init(&fod->flock);
642
643                 fod->rspdma = fc_dma_map_single(tgtport->dev, &fod->rspiubuf,
644                                         sizeof(fod->rspiubuf), DMA_TO_DEVICE);
645                 if (fc_dma_mapping_error(tgtport->dev, fod->rspdma)) {
646                         list_del(&fod->fcp_list);
647                         for (fod--, i--; i >= 0; fod--, i--) {
648                                 fc_dma_unmap_single(tgtport->dev, fod->rspdma,
649                                                 sizeof(fod->rspiubuf),
650                                                 DMA_TO_DEVICE);
651                                 fod->rspdma = 0L;
652                                 list_del(&fod->fcp_list);
653                         }
654
655                         return;
656                 }
657         }
658 }
659
660 static void
661 nvmet_fc_destroy_fcp_iodlist(struct nvmet_fc_tgtport *tgtport,
662                                 struct nvmet_fc_tgt_queue *queue)
663 {
664         struct nvmet_fc_fcp_iod *fod = queue->fod;
665         int i;
666
667         for (i = 0; i < queue->sqsize; fod++, i++) {
668                 if (fod->rspdma)
669                         fc_dma_unmap_single(tgtport->dev, fod->rspdma,
670                                 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
671         }
672 }
673
674 static struct nvmet_fc_fcp_iod *
675 nvmet_fc_alloc_fcp_iod(struct nvmet_fc_tgt_queue *queue)
676 {
677         struct nvmet_fc_fcp_iod *fod;
678
679         lockdep_assert_held(&queue->qlock);
680
681         fod = list_first_entry_or_null(&queue->fod_list,
682                                         struct nvmet_fc_fcp_iod, fcp_list);
683         if (fod) {
684                 list_del(&fod->fcp_list);
685                 fod->active = true;
686                 /*
687                  * no queue reference is taken, as it was taken by the
688                  * queue lookup just prior to the allocation. The iod
689                  * will "inherit" that reference.
690                  */
691         }
692         return fod;
693 }
694
695
696 static void
697 nvmet_fc_queue_fcp_req(struct nvmet_fc_tgtport *tgtport,
698                        struct nvmet_fc_tgt_queue *queue,
699                        struct nvmefc_tgt_fcp_req *fcpreq)
700 {
701         struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
702
703         /*
704          * put all admin cmds on hw queue id 0. All io commands go to
705          * the respective hw queue based on a modulo basis
706          */
707         fcpreq->hwqid = queue->qid ?
708                         ((queue->qid - 1) % tgtport->ops->max_hw_queues) : 0;
709
710         nvmet_fc_handle_fcp_rqst(tgtport, fod);
711 }
712
713 static void
714 nvmet_fc_fcp_rqst_op_defer_work(struct work_struct *work)
715 {
716         struct nvmet_fc_fcp_iod *fod =
717                 container_of(work, struct nvmet_fc_fcp_iod, defer_work);
718
719         /* Submit deferred IO for processing */
720         nvmet_fc_queue_fcp_req(fod->tgtport, fod->queue, fod->fcpreq);
721
722 }
723
724 static void
725 nvmet_fc_free_fcp_iod(struct nvmet_fc_tgt_queue *queue,
726                         struct nvmet_fc_fcp_iod *fod)
727 {
728         struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
729         struct nvmet_fc_tgtport *tgtport = fod->tgtport;
730         struct nvmet_fc_defer_fcp_req *deferfcp;
731         unsigned long flags;
732
733         fc_dma_sync_single_for_cpu(tgtport->dev, fod->rspdma,
734                                 sizeof(fod->rspiubuf), DMA_TO_DEVICE);
735
736         fcpreq->nvmet_fc_private = NULL;
737
738         fod->active = false;
739         fod->abort = false;
740         fod->aborted = false;
741         fod->writedataactive = false;
742         fod->fcpreq = NULL;
743
744         tgtport->ops->fcp_req_release(&tgtport->fc_target_port, fcpreq);
745
746         /* release the queue lookup reference on the completed IO */
747         nvmet_fc_tgt_q_put(queue);
748
749         spin_lock_irqsave(&queue->qlock, flags);
750         deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
751                                 struct nvmet_fc_defer_fcp_req, req_list);
752         if (!deferfcp) {
753                 list_add_tail(&fod->fcp_list, &fod->queue->fod_list);
754                 spin_unlock_irqrestore(&queue->qlock, flags);
755                 return;
756         }
757
758         /* Re-use the fod for the next pending cmd that was deferred */
759         list_del(&deferfcp->req_list);
760
761         fcpreq = deferfcp->fcp_req;
762
763         /* deferfcp can be reused for another IO at a later date */
764         list_add_tail(&deferfcp->req_list, &queue->avail_defer_list);
765
766         spin_unlock_irqrestore(&queue->qlock, flags);
767
768         /* Save NVME CMD IO in fod */
769         memcpy(&fod->cmdiubuf, fcpreq->rspaddr, fcpreq->rsplen);
770
771         /* Setup new fcpreq to be processed */
772         fcpreq->rspaddr = NULL;
773         fcpreq->rsplen  = 0;
774         fcpreq->nvmet_fc_private = fod;
775         fod->fcpreq = fcpreq;
776         fod->active = true;
777
778         /* inform LLDD IO is now being processed */
779         tgtport->ops->defer_rcv(&tgtport->fc_target_port, fcpreq);
780
781         /*
782          * Leave the queue lookup get reference taken when
783          * fod was originally allocated.
784          */
785
786         queue_work(queue->work_q, &fod->defer_work);
787 }
788
789 static struct nvmet_fc_tgt_queue *
790 nvmet_fc_alloc_target_queue(struct nvmet_fc_tgt_assoc *assoc,
791                         u16 qid, u16 sqsize)
792 {
793         struct nvmet_fc_tgt_queue *queue;
794         unsigned long flags;
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         spin_lock_irqsave(&assoc->tgtport->lock, flags);
834         assoc->queues[qid] = queue;
835         spin_unlock_irqrestore(&assoc->tgtport->lock, flags);
836
837         return queue;
838
839 out_fail_iodlist:
840         nvmet_fc_destroy_fcp_iodlist(assoc->tgtport, queue);
841         destroy_workqueue(queue->work_q);
842 out_a_put:
843         nvmet_fc_tgt_a_put(assoc);
844 out_free_queue:
845         kfree(queue);
846         return NULL;
847 }
848
849
850 static void
851 nvmet_fc_tgt_queue_free(struct kref *ref)
852 {
853         struct nvmet_fc_tgt_queue *queue =
854                 container_of(ref, struct nvmet_fc_tgt_queue, ref);
855         unsigned long flags;
856
857         spin_lock_irqsave(&queue->assoc->tgtport->lock, flags);
858         queue->assoc->queues[queue->qid] = NULL;
859         spin_unlock_irqrestore(&queue->assoc->tgtport->lock, flags);
860
861         nvmet_fc_destroy_fcp_iodlist(queue->assoc->tgtport, queue);
862
863         nvmet_fc_tgt_a_put(queue->assoc);
864
865         destroy_workqueue(queue->work_q);
866
867         kfree(queue);
868 }
869
870 static void
871 nvmet_fc_tgt_q_put(struct nvmet_fc_tgt_queue *queue)
872 {
873         kref_put(&queue->ref, nvmet_fc_tgt_queue_free);
874 }
875
876 static int
877 nvmet_fc_tgt_q_get(struct nvmet_fc_tgt_queue *queue)
878 {
879         return kref_get_unless_zero(&queue->ref);
880 }
881
882
883 static void
884 nvmet_fc_delete_target_queue(struct nvmet_fc_tgt_queue *queue)
885 {
886         struct nvmet_fc_tgtport *tgtport = queue->assoc->tgtport;
887         struct nvmet_fc_fcp_iod *fod = queue->fod;
888         struct nvmet_fc_defer_fcp_req *deferfcp, *tempptr;
889         unsigned long flags;
890         int i;
891         bool disconnect;
892
893         disconnect = atomic_xchg(&queue->connected, 0);
894
895         /* if not connected, nothing to do */
896         if (!disconnect)
897                 return;
898
899         spin_lock_irqsave(&queue->qlock, flags);
900         /* abort outstanding io's */
901         for (i = 0; i < queue->sqsize; fod++, i++) {
902                 if (fod->active) {
903                         spin_lock(&fod->flock);
904                         fod->abort = true;
905                         /*
906                          * only call lldd abort routine if waiting for
907                          * writedata. other outstanding ops should finish
908                          * on their own.
909                          */
910                         if (fod->writedataactive) {
911                                 fod->aborted = true;
912                                 spin_unlock(&fod->flock);
913                                 tgtport->ops->fcp_abort(
914                                         &tgtport->fc_target_port, fod->fcpreq);
915                         } else
916                                 spin_unlock(&fod->flock);
917                 }
918         }
919
920         /* Cleanup defer'ed IOs in queue */
921         list_for_each_entry_safe(deferfcp, tempptr, &queue->avail_defer_list,
922                                 req_list) {
923                 list_del(&deferfcp->req_list);
924                 kfree(deferfcp);
925         }
926
927         for (;;) {
928                 deferfcp = list_first_entry_or_null(&queue->pending_cmd_list,
929                                 struct nvmet_fc_defer_fcp_req, req_list);
930                 if (!deferfcp)
931                         break;
932
933                 list_del(&deferfcp->req_list);
934                 spin_unlock_irqrestore(&queue->qlock, flags);
935
936                 tgtport->ops->defer_rcv(&tgtport->fc_target_port,
937                                 deferfcp->fcp_req);
938
939                 tgtport->ops->fcp_abort(&tgtport->fc_target_port,
940                                 deferfcp->fcp_req);
941
942                 tgtport->ops->fcp_req_release(&tgtport->fc_target_port,
943                                 deferfcp->fcp_req);
944
945                 /* release the queue lookup reference */
946                 nvmet_fc_tgt_q_put(queue);
947
948                 kfree(deferfcp);
949
950                 spin_lock_irqsave(&queue->qlock, flags);
951         }
952         spin_unlock_irqrestore(&queue->qlock, flags);
953
954         flush_workqueue(queue->work_q);
955
956         nvmet_sq_destroy(&queue->nvme_sq);
957
958         nvmet_fc_tgt_q_put(queue);
959 }
960
961 static struct nvmet_fc_tgt_queue *
962 nvmet_fc_find_target_queue(struct nvmet_fc_tgtport *tgtport,
963                                 u64 connection_id)
964 {
965         struct nvmet_fc_tgt_assoc *assoc;
966         struct nvmet_fc_tgt_queue *queue;
967         u64 association_id = nvmet_fc_getassociationid(connection_id);
968         u16 qid = nvmet_fc_getqueueid(connection_id);
969         unsigned long flags;
970
971         if (qid > NVMET_NR_QUEUES)
972                 return NULL;
973
974         spin_lock_irqsave(&tgtport->lock, flags);
975         list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
976                 if (association_id == assoc->association_id) {
977                         queue = assoc->queues[qid];
978                         if (queue &&
979                             (!atomic_read(&queue->connected) ||
980                              !nvmet_fc_tgt_q_get(queue)))
981                                 queue = NULL;
982                         spin_unlock_irqrestore(&tgtport->lock, flags);
983                         return queue;
984                 }
985         }
986         spin_unlock_irqrestore(&tgtport->lock, flags);
987         return NULL;
988 }
989
990 static void
991 nvmet_fc_hostport_free(struct kref *ref)
992 {
993         struct nvmet_fc_hostport *hostport =
994                 container_of(ref, struct nvmet_fc_hostport, ref);
995         struct nvmet_fc_tgtport *tgtport = hostport->tgtport;
996         unsigned long flags;
997
998         spin_lock_irqsave(&tgtport->lock, flags);
999         list_del(&hostport->host_list);
1000         spin_unlock_irqrestore(&tgtport->lock, flags);
1001         if (tgtport->ops->host_release && hostport->invalid)
1002                 tgtport->ops->host_release(hostport->hosthandle);
1003         kfree(hostport);
1004         nvmet_fc_tgtport_put(tgtport);
1005 }
1006
1007 static void
1008 nvmet_fc_hostport_put(struct nvmet_fc_hostport *hostport)
1009 {
1010         kref_put(&hostport->ref, nvmet_fc_hostport_free);
1011 }
1012
1013 static int
1014 nvmet_fc_hostport_get(struct nvmet_fc_hostport *hostport)
1015 {
1016         return kref_get_unless_zero(&hostport->ref);
1017 }
1018
1019 static void
1020 nvmet_fc_free_hostport(struct nvmet_fc_hostport *hostport)
1021 {
1022         /* if LLDD not implemented, leave as NULL */
1023         if (!hostport || !hostport->hosthandle)
1024                 return;
1025
1026         nvmet_fc_hostport_put(hostport);
1027 }
1028
1029 static struct nvmet_fc_hostport *
1030 nvmet_fc_alloc_hostport(struct nvmet_fc_tgtport *tgtport, void *hosthandle)
1031 {
1032         struct nvmet_fc_hostport *newhost, *host, *match = NULL;
1033         unsigned long flags;
1034
1035         /* if LLDD not implemented, leave as NULL */
1036         if (!hosthandle)
1037                 return NULL;
1038
1039         /* take reference for what will be the newly allocated hostport */
1040         if (!nvmet_fc_tgtport_get(tgtport))
1041                 return ERR_PTR(-EINVAL);
1042
1043         newhost = kzalloc(sizeof(*newhost), GFP_KERNEL);
1044         if (!newhost) {
1045                 spin_lock_irqsave(&tgtport->lock, flags);
1046                 list_for_each_entry(host, &tgtport->host_list, host_list) {
1047                         if (host->hosthandle == hosthandle && !host->invalid) {
1048                                 if (nvmet_fc_hostport_get(host)) {
1049                                         match = host;
1050                                         break;
1051                                 }
1052                         }
1053                 }
1054                 spin_unlock_irqrestore(&tgtport->lock, flags);
1055                 /* no allocation - release reference */
1056                 nvmet_fc_tgtport_put(tgtport);
1057                 return (match) ? match : ERR_PTR(-ENOMEM);
1058         }
1059
1060         newhost->tgtport = tgtport;
1061         newhost->hosthandle = hosthandle;
1062         INIT_LIST_HEAD(&newhost->host_list);
1063         kref_init(&newhost->ref);
1064
1065         spin_lock_irqsave(&tgtport->lock, flags);
1066         list_for_each_entry(host, &tgtport->host_list, host_list) {
1067                 if (host->hosthandle == hosthandle && !host->invalid) {
1068                         if (nvmet_fc_hostport_get(host)) {
1069                                 match = host;
1070                                 break;
1071                         }
1072                 }
1073         }
1074         if (match) {
1075                 kfree(newhost);
1076                 newhost = NULL;
1077                 /* releasing allocation - release reference */
1078                 nvmet_fc_tgtport_put(tgtport);
1079         } else
1080                 list_add_tail(&newhost->host_list, &tgtport->host_list);
1081         spin_unlock_irqrestore(&tgtport->lock, flags);
1082
1083         return (match) ? match : newhost;
1084 }
1085
1086 static void
1087 nvmet_fc_delete_assoc(struct work_struct *work)
1088 {
1089         struct nvmet_fc_tgt_assoc *assoc =
1090                 container_of(work, struct nvmet_fc_tgt_assoc, del_work);
1091
1092         nvmet_fc_delete_target_assoc(assoc);
1093         nvmet_fc_tgt_a_put(assoc);
1094 }
1095
1096 static struct nvmet_fc_tgt_assoc *
1097 nvmet_fc_alloc_target_assoc(struct nvmet_fc_tgtport *tgtport, void *hosthandle)
1098 {
1099         struct nvmet_fc_tgt_assoc *assoc, *tmpassoc;
1100         unsigned long flags;
1101         u64 ran;
1102         int idx;
1103         bool needrandom = true;
1104
1105         if (!tgtport->pe)
1106                 return NULL;
1107
1108         assoc = kzalloc(sizeof(*assoc), GFP_KERNEL);
1109         if (!assoc)
1110                 return NULL;
1111
1112         idx = ida_simple_get(&tgtport->assoc_cnt, 0, 0, GFP_KERNEL);
1113         if (idx < 0)
1114                 goto out_free_assoc;
1115
1116         if (!nvmet_fc_tgtport_get(tgtport))
1117                 goto out_ida;
1118
1119         assoc->hostport = nvmet_fc_alloc_hostport(tgtport, hosthandle);
1120         if (IS_ERR(assoc->hostport))
1121                 goto out_put;
1122
1123         assoc->tgtport = tgtport;
1124         assoc->a_id = idx;
1125         INIT_LIST_HEAD(&assoc->a_list);
1126         kref_init(&assoc->ref);
1127         INIT_WORK(&assoc->del_work, nvmet_fc_delete_assoc);
1128         atomic_set(&assoc->terminating, 0);
1129
1130         while (needrandom) {
1131                 get_random_bytes(&ran, sizeof(ran) - BYTES_FOR_QID);
1132                 ran = ran << BYTES_FOR_QID_SHIFT;
1133
1134                 spin_lock_irqsave(&tgtport->lock, flags);
1135                 needrandom = false;
1136                 list_for_each_entry(tmpassoc, &tgtport->assoc_list, a_list) {
1137                         if (ran == tmpassoc->association_id) {
1138                                 needrandom = true;
1139                                 break;
1140                         }
1141                 }
1142                 if (!needrandom) {
1143                         assoc->association_id = ran;
1144                         list_add_tail(&assoc->a_list, &tgtport->assoc_list);
1145                 }
1146                 spin_unlock_irqrestore(&tgtport->lock, flags);
1147         }
1148
1149         return assoc;
1150
1151 out_put:
1152         nvmet_fc_tgtport_put(tgtport);
1153 out_ida:
1154         ida_simple_remove(&tgtport->assoc_cnt, idx);
1155 out_free_assoc:
1156         kfree(assoc);
1157         return NULL;
1158 }
1159
1160 static void
1161 nvmet_fc_target_assoc_free(struct kref *ref)
1162 {
1163         struct nvmet_fc_tgt_assoc *assoc =
1164                 container_of(ref, struct nvmet_fc_tgt_assoc, ref);
1165         struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
1166         struct nvmet_fc_ls_iod  *oldls;
1167         unsigned long flags;
1168
1169         /* Send Disconnect now that all i/o has completed */
1170         nvmet_fc_xmt_disconnect_assoc(assoc);
1171
1172         nvmet_fc_free_hostport(assoc->hostport);
1173         spin_lock_irqsave(&tgtport->lock, flags);
1174         list_del(&assoc->a_list);
1175         oldls = assoc->rcv_disconn;
1176         spin_unlock_irqrestore(&tgtport->lock, flags);
1177         /* if pending Rcv Disconnect Association LS, send rsp now */
1178         if (oldls)
1179                 nvmet_fc_xmt_ls_rsp(tgtport, oldls);
1180         ida_simple_remove(&tgtport->assoc_cnt, assoc->a_id);
1181         dev_info(tgtport->dev,
1182                 "{%d:%d} Association freed\n",
1183                 tgtport->fc_target_port.port_num, assoc->a_id);
1184         kfree(assoc);
1185         nvmet_fc_tgtport_put(tgtport);
1186 }
1187
1188 static void
1189 nvmet_fc_tgt_a_put(struct nvmet_fc_tgt_assoc *assoc)
1190 {
1191         kref_put(&assoc->ref, nvmet_fc_target_assoc_free);
1192 }
1193
1194 static int
1195 nvmet_fc_tgt_a_get(struct nvmet_fc_tgt_assoc *assoc)
1196 {
1197         return kref_get_unless_zero(&assoc->ref);
1198 }
1199
1200 static void
1201 nvmet_fc_delete_target_assoc(struct nvmet_fc_tgt_assoc *assoc)
1202 {
1203         struct nvmet_fc_tgtport *tgtport = assoc->tgtport;
1204         struct nvmet_fc_tgt_queue *queue;
1205         unsigned long flags;
1206         int i, terminating;
1207
1208         terminating = atomic_xchg(&assoc->terminating, 1);
1209
1210         /* if already terminating, do nothing */
1211         if (terminating)
1212                 return;
1213
1214         spin_lock_irqsave(&tgtport->lock, flags);
1215         for (i = NVMET_NR_QUEUES; i >= 0; i--) {
1216                 queue = assoc->queues[i];
1217                 if (queue) {
1218                         if (!nvmet_fc_tgt_q_get(queue))
1219                                 continue;
1220                         spin_unlock_irqrestore(&tgtport->lock, flags);
1221                         nvmet_fc_delete_target_queue(queue);
1222                         nvmet_fc_tgt_q_put(queue);
1223                         spin_lock_irqsave(&tgtport->lock, flags);
1224                 }
1225         }
1226         spin_unlock_irqrestore(&tgtport->lock, flags);
1227
1228         dev_info(tgtport->dev,
1229                 "{%d:%d} Association deleted\n",
1230                 tgtport->fc_target_port.port_num, assoc->a_id);
1231
1232         nvmet_fc_tgt_a_put(assoc);
1233 }
1234
1235 static struct nvmet_fc_tgt_assoc *
1236 nvmet_fc_find_target_assoc(struct nvmet_fc_tgtport *tgtport,
1237                                 u64 association_id)
1238 {
1239         struct nvmet_fc_tgt_assoc *assoc;
1240         struct nvmet_fc_tgt_assoc *ret = NULL;
1241         unsigned long flags;
1242
1243         spin_lock_irqsave(&tgtport->lock, flags);
1244         list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
1245                 if (association_id == assoc->association_id) {
1246                         ret = assoc;
1247                         if (!nvmet_fc_tgt_a_get(assoc))
1248                                 ret = NULL;
1249                         break;
1250                 }
1251         }
1252         spin_unlock_irqrestore(&tgtport->lock, flags);
1253
1254         return ret;
1255 }
1256
1257 static void
1258 nvmet_fc_portentry_bind(struct nvmet_fc_tgtport *tgtport,
1259                         struct nvmet_fc_port_entry *pe,
1260                         struct nvmet_port *port)
1261 {
1262         lockdep_assert_held(&nvmet_fc_tgtlock);
1263
1264         pe->tgtport = tgtport;
1265         tgtport->pe = pe;
1266
1267         pe->port = port;
1268         port->priv = pe;
1269
1270         pe->node_name = tgtport->fc_target_port.node_name;
1271         pe->port_name = tgtport->fc_target_port.port_name;
1272         INIT_LIST_HEAD(&pe->pe_list);
1273
1274         list_add_tail(&pe->pe_list, &nvmet_fc_portentry_list);
1275 }
1276
1277 static void
1278 nvmet_fc_portentry_unbind(struct nvmet_fc_port_entry *pe)
1279 {
1280         unsigned long flags;
1281
1282         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1283         if (pe->tgtport)
1284                 pe->tgtport->pe = NULL;
1285         list_del(&pe->pe_list);
1286         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1287 }
1288
1289 /*
1290  * called when a targetport deregisters. Breaks the relationship
1291  * with the nvmet port, but leaves the port_entry in place so that
1292  * re-registration can resume operation.
1293  */
1294 static void
1295 nvmet_fc_portentry_unbind_tgt(struct nvmet_fc_tgtport *tgtport)
1296 {
1297         struct nvmet_fc_port_entry *pe;
1298         unsigned long flags;
1299
1300         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1301         pe = tgtport->pe;
1302         if (pe)
1303                 pe->tgtport = NULL;
1304         tgtport->pe = NULL;
1305         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1306 }
1307
1308 /*
1309  * called when a new targetport is registered. Looks in the
1310  * existing nvmet port_entries to see if the nvmet layer is
1311  * configured for the targetport's wwn's. (the targetport existed,
1312  * nvmet configured, the lldd unregistered the tgtport, and is now
1313  * reregistering the same targetport).  If so, set the nvmet port
1314  * port entry on the targetport.
1315  */
1316 static void
1317 nvmet_fc_portentry_rebind_tgt(struct nvmet_fc_tgtport *tgtport)
1318 {
1319         struct nvmet_fc_port_entry *pe;
1320         unsigned long flags;
1321
1322         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1323         list_for_each_entry(pe, &nvmet_fc_portentry_list, pe_list) {
1324                 if (tgtport->fc_target_port.node_name == pe->node_name &&
1325                     tgtport->fc_target_port.port_name == pe->port_name) {
1326                         WARN_ON(pe->tgtport);
1327                         tgtport->pe = pe;
1328                         pe->tgtport = tgtport;
1329                         break;
1330                 }
1331         }
1332         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1333 }
1334
1335 /**
1336  * nvme_fc_register_targetport - transport entry point called by an
1337  *                              LLDD to register the existence of a local
1338  *                              NVME subystem FC port.
1339  * @pinfo:     pointer to information about the port to be registered
1340  * @template:  LLDD entrypoints and operational parameters for the port
1341  * @dev:       physical hardware device node port corresponds to. Will be
1342  *             used for DMA mappings
1343  * @portptr:   pointer to a local port pointer. Upon success, the routine
1344  *             will allocate a nvme_fc_local_port structure and place its
1345  *             address in the local port pointer. Upon failure, local port
1346  *             pointer will be set to NULL.
1347  *
1348  * Returns:
1349  * a completion status. Must be 0 upon success; a negative errno
1350  * (ex: -ENXIO) upon failure.
1351  */
1352 int
1353 nvmet_fc_register_targetport(struct nvmet_fc_port_info *pinfo,
1354                         struct nvmet_fc_target_template *template,
1355                         struct device *dev,
1356                         struct nvmet_fc_target_port **portptr)
1357 {
1358         struct nvmet_fc_tgtport *newrec;
1359         unsigned long flags;
1360         int ret, idx;
1361
1362         if (!template->xmt_ls_rsp || !template->fcp_op ||
1363             !template->fcp_abort ||
1364             !template->fcp_req_release || !template->targetport_delete ||
1365             !template->max_hw_queues || !template->max_sgl_segments ||
1366             !template->max_dif_sgl_segments || !template->dma_boundary) {
1367                 ret = -EINVAL;
1368                 goto out_regtgt_failed;
1369         }
1370
1371         newrec = kzalloc((sizeof(*newrec) + template->target_priv_sz),
1372                          GFP_KERNEL);
1373         if (!newrec) {
1374                 ret = -ENOMEM;
1375                 goto out_regtgt_failed;
1376         }
1377
1378         idx = ida_simple_get(&nvmet_fc_tgtport_cnt, 0, 0, GFP_KERNEL);
1379         if (idx < 0) {
1380                 ret = -ENOSPC;
1381                 goto out_fail_kfree;
1382         }
1383
1384         if (!get_device(dev) && dev) {
1385                 ret = -ENODEV;
1386                 goto out_ida_put;
1387         }
1388
1389         newrec->fc_target_port.node_name = pinfo->node_name;
1390         newrec->fc_target_port.port_name = pinfo->port_name;
1391         if (template->target_priv_sz)
1392                 newrec->fc_target_port.private = &newrec[1];
1393         else
1394                 newrec->fc_target_port.private = NULL;
1395         newrec->fc_target_port.port_id = pinfo->port_id;
1396         newrec->fc_target_port.port_num = idx;
1397         INIT_LIST_HEAD(&newrec->tgt_list);
1398         newrec->dev = dev;
1399         newrec->ops = template;
1400         spin_lock_init(&newrec->lock);
1401         INIT_LIST_HEAD(&newrec->ls_rcv_list);
1402         INIT_LIST_HEAD(&newrec->ls_req_list);
1403         INIT_LIST_HEAD(&newrec->ls_busylist);
1404         INIT_LIST_HEAD(&newrec->assoc_list);
1405         INIT_LIST_HEAD(&newrec->host_list);
1406         kref_init(&newrec->ref);
1407         ida_init(&newrec->assoc_cnt);
1408         newrec->max_sg_cnt = template->max_sgl_segments;
1409
1410         ret = nvmet_fc_alloc_ls_iodlist(newrec);
1411         if (ret) {
1412                 ret = -ENOMEM;
1413                 goto out_free_newrec;
1414         }
1415
1416         nvmet_fc_portentry_rebind_tgt(newrec);
1417
1418         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1419         list_add_tail(&newrec->tgt_list, &nvmet_fc_target_list);
1420         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1421
1422         *portptr = &newrec->fc_target_port;
1423         return 0;
1424
1425 out_free_newrec:
1426         put_device(dev);
1427 out_ida_put:
1428         ida_simple_remove(&nvmet_fc_tgtport_cnt, idx);
1429 out_fail_kfree:
1430         kfree(newrec);
1431 out_regtgt_failed:
1432         *portptr = NULL;
1433         return ret;
1434 }
1435 EXPORT_SYMBOL_GPL(nvmet_fc_register_targetport);
1436
1437
1438 static void
1439 nvmet_fc_free_tgtport(struct kref *ref)
1440 {
1441         struct nvmet_fc_tgtport *tgtport =
1442                 container_of(ref, struct nvmet_fc_tgtport, ref);
1443         struct device *dev = tgtport->dev;
1444         unsigned long flags;
1445
1446         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1447         list_del(&tgtport->tgt_list);
1448         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1449
1450         nvmet_fc_free_ls_iodlist(tgtport);
1451
1452         /* let the LLDD know we've finished tearing it down */
1453         tgtport->ops->targetport_delete(&tgtport->fc_target_port);
1454
1455         ida_simple_remove(&nvmet_fc_tgtport_cnt,
1456                         tgtport->fc_target_port.port_num);
1457
1458         ida_destroy(&tgtport->assoc_cnt);
1459
1460         kfree(tgtport);
1461
1462         put_device(dev);
1463 }
1464
1465 static void
1466 nvmet_fc_tgtport_put(struct nvmet_fc_tgtport *tgtport)
1467 {
1468         kref_put(&tgtport->ref, nvmet_fc_free_tgtport);
1469 }
1470
1471 static int
1472 nvmet_fc_tgtport_get(struct nvmet_fc_tgtport *tgtport)
1473 {
1474         return kref_get_unless_zero(&tgtport->ref);
1475 }
1476
1477 static void
1478 __nvmet_fc_free_assocs(struct nvmet_fc_tgtport *tgtport)
1479 {
1480         struct nvmet_fc_tgt_assoc *assoc, *next;
1481         unsigned long flags;
1482
1483         spin_lock_irqsave(&tgtport->lock, flags);
1484         list_for_each_entry_safe(assoc, next,
1485                                 &tgtport->assoc_list, a_list) {
1486                 if (!nvmet_fc_tgt_a_get(assoc))
1487                         continue;
1488                 if (!schedule_work(&assoc->del_work))
1489                         /* already deleting - release local reference */
1490                         nvmet_fc_tgt_a_put(assoc);
1491         }
1492         spin_unlock_irqrestore(&tgtport->lock, flags);
1493 }
1494
1495 /**
1496  * nvmet_fc_invalidate_host - transport entry point called by an LLDD
1497  *                       to remove references to a hosthandle for LS's.
1498  *
1499  * The nvmet-fc layer ensures that any references to the hosthandle
1500  * on the targetport are forgotten (set to NULL).  The LLDD will
1501  * typically call this when a login with a remote host port has been
1502  * lost, thus LS's for the remote host port are no longer possible.
1503  *
1504  * If an LS request is outstanding to the targetport/hosthandle (or
1505  * issued concurrently with the call to invalidate the host), the
1506  * LLDD is responsible for terminating/aborting the LS and completing
1507  * the LS request. It is recommended that these terminations/aborts
1508  * occur after calling to invalidate the host handle to avoid additional
1509  * retries by the nvmet-fc transport. The nvmet-fc transport may
1510  * continue to reference host handle while it cleans up outstanding
1511  * NVME associations. The nvmet-fc transport will call the
1512  * ops->host_release() callback to notify the LLDD that all references
1513  * are complete and the related host handle can be recovered.
1514  * Note: if there are no references, the callback may be called before
1515  * the invalidate host call returns.
1516  *
1517  * @target_port: pointer to the (registered) target port that a prior
1518  *              LS was received on and which supplied the transport the
1519  *              hosthandle.
1520  * @hosthandle: the handle (pointer) that represents the host port
1521  *              that no longer has connectivity and that LS's should
1522  *              no longer be directed to.
1523  */
1524 void
1525 nvmet_fc_invalidate_host(struct nvmet_fc_target_port *target_port,
1526                         void *hosthandle)
1527 {
1528         struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
1529         struct nvmet_fc_tgt_assoc *assoc, *next;
1530         unsigned long flags;
1531         bool noassoc = true;
1532
1533         spin_lock_irqsave(&tgtport->lock, flags);
1534         list_for_each_entry_safe(assoc, next,
1535                                 &tgtport->assoc_list, a_list) {
1536                 if (!assoc->hostport ||
1537                     assoc->hostport->hosthandle != hosthandle)
1538                         continue;
1539                 if (!nvmet_fc_tgt_a_get(assoc))
1540                         continue;
1541                 assoc->hostport->invalid = 1;
1542                 noassoc = false;
1543                 if (!schedule_work(&assoc->del_work))
1544                         /* already deleting - release local reference */
1545                         nvmet_fc_tgt_a_put(assoc);
1546         }
1547         spin_unlock_irqrestore(&tgtport->lock, flags);
1548
1549         /* if there's nothing to wait for - call the callback */
1550         if (noassoc && tgtport->ops->host_release)
1551                 tgtport->ops->host_release(hosthandle);
1552 }
1553 EXPORT_SYMBOL_GPL(nvmet_fc_invalidate_host);
1554
1555 /*
1556  * nvmet layer has called to terminate an association
1557  */
1558 static void
1559 nvmet_fc_delete_ctrl(struct nvmet_ctrl *ctrl)
1560 {
1561         struct nvmet_fc_tgtport *tgtport, *next;
1562         struct nvmet_fc_tgt_assoc *assoc;
1563         struct nvmet_fc_tgt_queue *queue;
1564         unsigned long flags;
1565         bool found_ctrl = false;
1566
1567         /* this is a bit ugly, but don't want to make locks layered */
1568         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1569         list_for_each_entry_safe(tgtport, next, &nvmet_fc_target_list,
1570                         tgt_list) {
1571                 if (!nvmet_fc_tgtport_get(tgtport))
1572                         continue;
1573                 spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1574
1575                 spin_lock_irqsave(&tgtport->lock, flags);
1576                 list_for_each_entry(assoc, &tgtport->assoc_list, a_list) {
1577                         queue = assoc->queues[0];
1578                         if (queue && queue->nvme_sq.ctrl == ctrl) {
1579                                 if (nvmet_fc_tgt_a_get(assoc))
1580                                         found_ctrl = true;
1581                                 break;
1582                         }
1583                 }
1584                 spin_unlock_irqrestore(&tgtport->lock, flags);
1585
1586                 nvmet_fc_tgtport_put(tgtport);
1587
1588                 if (found_ctrl) {
1589                         if (!schedule_work(&assoc->del_work))
1590                                 /* already deleting - release local reference */
1591                                 nvmet_fc_tgt_a_put(assoc);
1592                         return;
1593                 }
1594
1595                 spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
1596         }
1597         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
1598 }
1599
1600 /**
1601  * nvme_fc_unregister_targetport - transport entry point called by an
1602  *                              LLDD to deregister/remove a previously
1603  *                              registered a local NVME subsystem FC port.
1604  * @target_port: pointer to the (registered) target port that is to be
1605  *               deregistered.
1606  *
1607  * Returns:
1608  * a completion status. Must be 0 upon success; a negative errno
1609  * (ex: -ENXIO) upon failure.
1610  */
1611 int
1612 nvmet_fc_unregister_targetport(struct nvmet_fc_target_port *target_port)
1613 {
1614         struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
1615
1616         nvmet_fc_portentry_unbind_tgt(tgtport);
1617
1618         /* terminate any outstanding associations */
1619         __nvmet_fc_free_assocs(tgtport);
1620
1621         /*
1622          * should terminate LS's as well. However, LS's will be generated
1623          * at the tail end of association termination, so they likely don't
1624          * exist yet. And even if they did, it's worthwhile to just let
1625          * them finish and targetport ref counting will clean things up.
1626          */
1627
1628         nvmet_fc_tgtport_put(tgtport);
1629
1630         return 0;
1631 }
1632 EXPORT_SYMBOL_GPL(nvmet_fc_unregister_targetport);
1633
1634
1635 /* ********************** FC-NVME LS RCV Handling ************************* */
1636
1637
1638 static void
1639 nvmet_fc_ls_create_association(struct nvmet_fc_tgtport *tgtport,
1640                         struct nvmet_fc_ls_iod *iod)
1641 {
1642         struct fcnvme_ls_cr_assoc_rqst *rqst = &iod->rqstbuf->rq_cr_assoc;
1643         struct fcnvme_ls_cr_assoc_acc *acc = &iod->rspbuf->rsp_cr_assoc;
1644         struct nvmet_fc_tgt_queue *queue;
1645         int ret = 0;
1646
1647         memset(acc, 0, sizeof(*acc));
1648
1649         /*
1650          * FC-NVME spec changes. There are initiators sending different
1651          * lengths as padding sizes for Create Association Cmd descriptor
1652          * was incorrect.
1653          * Accept anything of "minimum" length. Assume format per 1.15
1654          * spec (with HOSTID reduced to 16 bytes), ignore how long the
1655          * trailing pad length is.
1656          */
1657         if (iod->rqstdatalen < FCNVME_LSDESC_CRA_RQST_MINLEN)
1658                 ret = VERR_CR_ASSOC_LEN;
1659         else if (be32_to_cpu(rqst->desc_list_len) <
1660                         FCNVME_LSDESC_CRA_RQST_MIN_LISTLEN)
1661                 ret = VERR_CR_ASSOC_RQST_LEN;
1662         else if (rqst->assoc_cmd.desc_tag !=
1663                         cpu_to_be32(FCNVME_LSDESC_CREATE_ASSOC_CMD))
1664                 ret = VERR_CR_ASSOC_CMD;
1665         else if (be32_to_cpu(rqst->assoc_cmd.desc_len) <
1666                         FCNVME_LSDESC_CRA_CMD_DESC_MIN_DESCLEN)
1667                 ret = VERR_CR_ASSOC_CMD_LEN;
1668         else if (!rqst->assoc_cmd.ersp_ratio ||
1669                  (be16_to_cpu(rqst->assoc_cmd.ersp_ratio) >=
1670                                 be16_to_cpu(rqst->assoc_cmd.sqsize)))
1671                 ret = VERR_ERSP_RATIO;
1672
1673         else {
1674                 /* new association w/ admin queue */
1675                 iod->assoc = nvmet_fc_alloc_target_assoc(
1676                                                 tgtport, iod->hosthandle);
1677                 if (!iod->assoc)
1678                         ret = VERR_ASSOC_ALLOC_FAIL;
1679                 else {
1680                         queue = nvmet_fc_alloc_target_queue(iod->assoc, 0,
1681                                         be16_to_cpu(rqst->assoc_cmd.sqsize));
1682                         if (!queue) {
1683                                 ret = VERR_QUEUE_ALLOC_FAIL;
1684                                 nvmet_fc_tgt_a_put(iod->assoc);
1685                         }
1686                 }
1687         }
1688
1689         if (ret) {
1690                 dev_err(tgtport->dev,
1691                         "Create Association LS failed: %s\n",
1692                         validation_errors[ret]);
1693                 iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1694                                 sizeof(*acc), rqst->w0.ls_cmd,
1695                                 FCNVME_RJT_RC_LOGIC,
1696                                 FCNVME_RJT_EXP_NONE, 0);
1697                 return;
1698         }
1699
1700         queue->ersp_ratio = be16_to_cpu(rqst->assoc_cmd.ersp_ratio);
1701         atomic_set(&queue->connected, 1);
1702         queue->sqhd = 0;        /* best place to init value */
1703
1704         dev_info(tgtport->dev,
1705                 "{%d:%d} Association created\n",
1706                 tgtport->fc_target_port.port_num, iod->assoc->a_id);
1707
1708         /* format a response */
1709
1710         iod->lsrsp->rsplen = sizeof(*acc);
1711
1712         nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1713                         fcnvme_lsdesc_len(
1714                                 sizeof(struct fcnvme_ls_cr_assoc_acc)),
1715                         FCNVME_LS_CREATE_ASSOCIATION);
1716         acc->associd.desc_tag = cpu_to_be32(FCNVME_LSDESC_ASSOC_ID);
1717         acc->associd.desc_len =
1718                         fcnvme_lsdesc_len(
1719                                 sizeof(struct fcnvme_lsdesc_assoc_id));
1720         acc->associd.association_id =
1721                         cpu_to_be64(nvmet_fc_makeconnid(iod->assoc, 0));
1722         acc->connectid.desc_tag = cpu_to_be32(FCNVME_LSDESC_CONN_ID);
1723         acc->connectid.desc_len =
1724                         fcnvme_lsdesc_len(
1725                                 sizeof(struct fcnvme_lsdesc_conn_id));
1726         acc->connectid.connection_id = acc->associd.association_id;
1727 }
1728
1729 static void
1730 nvmet_fc_ls_create_connection(struct nvmet_fc_tgtport *tgtport,
1731                         struct nvmet_fc_ls_iod *iod)
1732 {
1733         struct fcnvme_ls_cr_conn_rqst *rqst = &iod->rqstbuf->rq_cr_conn;
1734         struct fcnvme_ls_cr_conn_acc *acc = &iod->rspbuf->rsp_cr_conn;
1735         struct nvmet_fc_tgt_queue *queue;
1736         int ret = 0;
1737
1738         memset(acc, 0, sizeof(*acc));
1739
1740         if (iod->rqstdatalen < sizeof(struct fcnvme_ls_cr_conn_rqst))
1741                 ret = VERR_CR_CONN_LEN;
1742         else if (rqst->desc_list_len !=
1743                         fcnvme_lsdesc_len(
1744                                 sizeof(struct fcnvme_ls_cr_conn_rqst)))
1745                 ret = VERR_CR_CONN_RQST_LEN;
1746         else if (rqst->associd.desc_tag != cpu_to_be32(FCNVME_LSDESC_ASSOC_ID))
1747                 ret = VERR_ASSOC_ID;
1748         else if (rqst->associd.desc_len !=
1749                         fcnvme_lsdesc_len(
1750                                 sizeof(struct fcnvme_lsdesc_assoc_id)))
1751                 ret = VERR_ASSOC_ID_LEN;
1752         else if (rqst->connect_cmd.desc_tag !=
1753                         cpu_to_be32(FCNVME_LSDESC_CREATE_CONN_CMD))
1754                 ret = VERR_CR_CONN_CMD;
1755         else if (rqst->connect_cmd.desc_len !=
1756                         fcnvme_lsdesc_len(
1757                                 sizeof(struct fcnvme_lsdesc_cr_conn_cmd)))
1758                 ret = VERR_CR_CONN_CMD_LEN;
1759         else if (!rqst->connect_cmd.ersp_ratio ||
1760                  (be16_to_cpu(rqst->connect_cmd.ersp_ratio) >=
1761                                 be16_to_cpu(rqst->connect_cmd.sqsize)))
1762                 ret = VERR_ERSP_RATIO;
1763
1764         else {
1765                 /* new io queue */
1766                 iod->assoc = nvmet_fc_find_target_assoc(tgtport,
1767                                 be64_to_cpu(rqst->associd.association_id));
1768                 if (!iod->assoc)
1769                         ret = VERR_NO_ASSOC;
1770                 else {
1771                         queue = nvmet_fc_alloc_target_queue(iod->assoc,
1772                                         be16_to_cpu(rqst->connect_cmd.qid),
1773                                         be16_to_cpu(rqst->connect_cmd.sqsize));
1774                         if (!queue)
1775                                 ret = VERR_QUEUE_ALLOC_FAIL;
1776
1777                         /* release get taken in nvmet_fc_find_target_assoc */
1778                         nvmet_fc_tgt_a_put(iod->assoc);
1779                 }
1780         }
1781
1782         if (ret) {
1783                 dev_err(tgtport->dev,
1784                         "Create Connection LS failed: %s\n",
1785                         validation_errors[ret]);
1786                 iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1787                                 sizeof(*acc), rqst->w0.ls_cmd,
1788                                 (ret == VERR_NO_ASSOC) ?
1789                                         FCNVME_RJT_RC_INV_ASSOC :
1790                                         FCNVME_RJT_RC_LOGIC,
1791                                 FCNVME_RJT_EXP_NONE, 0);
1792                 return;
1793         }
1794
1795         queue->ersp_ratio = be16_to_cpu(rqst->connect_cmd.ersp_ratio);
1796         atomic_set(&queue->connected, 1);
1797         queue->sqhd = 0;        /* best place to init value */
1798
1799         /* format a response */
1800
1801         iod->lsrsp->rsplen = sizeof(*acc);
1802
1803         nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1804                         fcnvme_lsdesc_len(sizeof(struct fcnvme_ls_cr_conn_acc)),
1805                         FCNVME_LS_CREATE_CONNECTION);
1806         acc->connectid.desc_tag = cpu_to_be32(FCNVME_LSDESC_CONN_ID);
1807         acc->connectid.desc_len =
1808                         fcnvme_lsdesc_len(
1809                                 sizeof(struct fcnvme_lsdesc_conn_id));
1810         acc->connectid.connection_id =
1811                         cpu_to_be64(nvmet_fc_makeconnid(iod->assoc,
1812                                 be16_to_cpu(rqst->connect_cmd.qid)));
1813 }
1814
1815 /*
1816  * Returns true if the LS response is to be transmit
1817  * Returns false if the LS response is to be delayed
1818  */
1819 static int
1820 nvmet_fc_ls_disconnect(struct nvmet_fc_tgtport *tgtport,
1821                         struct nvmet_fc_ls_iod *iod)
1822 {
1823         struct fcnvme_ls_disconnect_assoc_rqst *rqst =
1824                                                 &iod->rqstbuf->rq_dis_assoc;
1825         struct fcnvme_ls_disconnect_assoc_acc *acc =
1826                                                 &iod->rspbuf->rsp_dis_assoc;
1827         struct nvmet_fc_tgt_assoc *assoc = NULL;
1828         struct nvmet_fc_ls_iod *oldls = NULL;
1829         unsigned long flags;
1830         int ret = 0;
1831
1832         memset(acc, 0, sizeof(*acc));
1833
1834         ret = nvmefc_vldt_lsreq_discon_assoc(iod->rqstdatalen, rqst);
1835         if (!ret) {
1836                 /* match an active association - takes an assoc ref if !NULL */
1837                 assoc = nvmet_fc_find_target_assoc(tgtport,
1838                                 be64_to_cpu(rqst->associd.association_id));
1839                 iod->assoc = assoc;
1840                 if (!assoc)
1841                         ret = VERR_NO_ASSOC;
1842         }
1843
1844         if (ret || !assoc) {
1845                 dev_err(tgtport->dev,
1846                         "Disconnect LS failed: %s\n",
1847                         validation_errors[ret]);
1848                 iod->lsrsp->rsplen = nvme_fc_format_rjt(acc,
1849                                 sizeof(*acc), rqst->w0.ls_cmd,
1850                                 (ret == VERR_NO_ASSOC) ?
1851                                         FCNVME_RJT_RC_INV_ASSOC :
1852                                         FCNVME_RJT_RC_LOGIC,
1853                                 FCNVME_RJT_EXP_NONE, 0);
1854                 return true;
1855         }
1856
1857         /* format a response */
1858
1859         iod->lsrsp->rsplen = sizeof(*acc);
1860
1861         nvme_fc_format_rsp_hdr(acc, FCNVME_LS_ACC,
1862                         fcnvme_lsdesc_len(
1863                                 sizeof(struct fcnvme_ls_disconnect_assoc_acc)),
1864                         FCNVME_LS_DISCONNECT_ASSOC);
1865
1866         /* release get taken in nvmet_fc_find_target_assoc */
1867         nvmet_fc_tgt_a_put(assoc);
1868
1869         /*
1870          * The rules for LS response says the response cannot
1871          * go back until ABTS's have been sent for all outstanding
1872          * I/O and a Disconnect Association LS has been sent.
1873          * So... save off the Disconnect LS to send the response
1874          * later. If there was a prior LS already saved, replace
1875          * it with the newer one and send a can't perform reject
1876          * on the older one.
1877          */
1878         spin_lock_irqsave(&tgtport->lock, flags);
1879         oldls = assoc->rcv_disconn;
1880         assoc->rcv_disconn = iod;
1881         spin_unlock_irqrestore(&tgtport->lock, flags);
1882
1883         nvmet_fc_delete_target_assoc(assoc);
1884
1885         if (oldls) {
1886                 dev_info(tgtport->dev,
1887                         "{%d:%d} Multiple Disconnect Association LS's "
1888                         "received\n",
1889                         tgtport->fc_target_port.port_num, assoc->a_id);
1890                 /* overwrite good response with bogus failure */
1891                 oldls->lsrsp->rsplen = nvme_fc_format_rjt(oldls->rspbuf,
1892                                                 sizeof(*iod->rspbuf),
1893                                                 /* ok to use rqst, LS is same */
1894                                                 rqst->w0.ls_cmd,
1895                                                 FCNVME_RJT_RC_UNAB,
1896                                                 FCNVME_RJT_EXP_NONE, 0);
1897                 nvmet_fc_xmt_ls_rsp(tgtport, oldls);
1898         }
1899
1900         return false;
1901 }
1902
1903
1904 /* *********************** NVME Ctrl Routines **************************** */
1905
1906
1907 static void nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req);
1908
1909 static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops;
1910
1911 static void
1912 nvmet_fc_xmt_ls_rsp_done(struct nvmefc_ls_rsp *lsrsp)
1913 {
1914         struct nvmet_fc_ls_iod *iod = lsrsp->nvme_fc_private;
1915         struct nvmet_fc_tgtport *tgtport = iod->tgtport;
1916
1917         fc_dma_sync_single_for_cpu(tgtport->dev, iod->rspdma,
1918                                 sizeof(*iod->rspbuf), DMA_TO_DEVICE);
1919         nvmet_fc_free_ls_iod(tgtport, iod);
1920         nvmet_fc_tgtport_put(tgtport);
1921 }
1922
1923 static void
1924 nvmet_fc_xmt_ls_rsp(struct nvmet_fc_tgtport *tgtport,
1925                                 struct nvmet_fc_ls_iod *iod)
1926 {
1927         int ret;
1928
1929         fc_dma_sync_single_for_device(tgtport->dev, iod->rspdma,
1930                                   sizeof(*iod->rspbuf), DMA_TO_DEVICE);
1931
1932         ret = tgtport->ops->xmt_ls_rsp(&tgtport->fc_target_port, iod->lsrsp);
1933         if (ret)
1934                 nvmet_fc_xmt_ls_rsp_done(iod->lsrsp);
1935 }
1936
1937 /*
1938  * Actual processing routine for received FC-NVME LS Requests from the LLD
1939  */
1940 static void
1941 nvmet_fc_handle_ls_rqst(struct nvmet_fc_tgtport *tgtport,
1942                         struct nvmet_fc_ls_iod *iod)
1943 {
1944         struct fcnvme_ls_rqst_w0 *w0 = &iod->rqstbuf->rq_cr_assoc.w0;
1945         bool sendrsp = true;
1946
1947         iod->lsrsp->nvme_fc_private = iod;
1948         iod->lsrsp->rspbuf = iod->rspbuf;
1949         iod->lsrsp->rspdma = iod->rspdma;
1950         iod->lsrsp->done = nvmet_fc_xmt_ls_rsp_done;
1951         /* Be preventative. handlers will later set to valid length */
1952         iod->lsrsp->rsplen = 0;
1953
1954         iod->assoc = NULL;
1955
1956         /*
1957          * handlers:
1958          *   parse request input, execute the request, and format the
1959          *   LS response
1960          */
1961         switch (w0->ls_cmd) {
1962         case FCNVME_LS_CREATE_ASSOCIATION:
1963                 /* Creates Association and initial Admin Queue/Connection */
1964                 nvmet_fc_ls_create_association(tgtport, iod);
1965                 break;
1966         case FCNVME_LS_CREATE_CONNECTION:
1967                 /* Creates an IO Queue/Connection */
1968                 nvmet_fc_ls_create_connection(tgtport, iod);
1969                 break;
1970         case FCNVME_LS_DISCONNECT_ASSOC:
1971                 /* Terminate a Queue/Connection or the Association */
1972                 sendrsp = nvmet_fc_ls_disconnect(tgtport, iod);
1973                 break;
1974         default:
1975                 iod->lsrsp->rsplen = nvme_fc_format_rjt(iod->rspbuf,
1976                                 sizeof(*iod->rspbuf), w0->ls_cmd,
1977                                 FCNVME_RJT_RC_INVAL, FCNVME_RJT_EXP_NONE, 0);
1978         }
1979
1980         if (sendrsp)
1981                 nvmet_fc_xmt_ls_rsp(tgtport, iod);
1982 }
1983
1984 /*
1985  * Actual processing routine for received FC-NVME LS Requests from the LLD
1986  */
1987 static void
1988 nvmet_fc_handle_ls_rqst_work(struct work_struct *work)
1989 {
1990         struct nvmet_fc_ls_iod *iod =
1991                 container_of(work, struct nvmet_fc_ls_iod, work);
1992         struct nvmet_fc_tgtport *tgtport = iod->tgtport;
1993
1994         nvmet_fc_handle_ls_rqst(tgtport, iod);
1995 }
1996
1997
1998 /**
1999  * nvmet_fc_rcv_ls_req - transport entry point called by an LLDD
2000  *                       upon the reception of a NVME LS request.
2001  *
2002  * The nvmet-fc layer will copy payload to an internal structure for
2003  * processing.  As such, upon completion of the routine, the LLDD may
2004  * immediately free/reuse the LS request buffer passed in the call.
2005  *
2006  * If this routine returns error, the LLDD should abort the exchange.
2007  *
2008  * @target_port: pointer to the (registered) target port the LS was
2009  *              received on.
2010  * @lsrsp:      pointer to a lsrsp structure to be used to reference
2011  *              the exchange corresponding to the LS.
2012  * @lsreqbuf:   pointer to the buffer containing the LS Request
2013  * @lsreqbuf_len: length, in bytes, of the received LS request
2014  */
2015 int
2016 nvmet_fc_rcv_ls_req(struct nvmet_fc_target_port *target_port,
2017                         void *hosthandle,
2018                         struct nvmefc_ls_rsp *lsrsp,
2019                         void *lsreqbuf, u32 lsreqbuf_len)
2020 {
2021         struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
2022         struct nvmet_fc_ls_iod *iod;
2023         struct fcnvme_ls_rqst_w0 *w0 = (struct fcnvme_ls_rqst_w0 *)lsreqbuf;
2024
2025         if (lsreqbuf_len > sizeof(union nvmefc_ls_requests)) {
2026                 dev_info(tgtport->dev,
2027                         "RCV %s LS failed: payload too large (%d)\n",
2028                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2029                                 nvmefc_ls_names[w0->ls_cmd] : "",
2030                         lsreqbuf_len);
2031                 return -E2BIG;
2032         }
2033
2034         if (!nvmet_fc_tgtport_get(tgtport)) {
2035                 dev_info(tgtport->dev,
2036                         "RCV %s LS failed: target deleting\n",
2037                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2038                                 nvmefc_ls_names[w0->ls_cmd] : "");
2039                 return -ESHUTDOWN;
2040         }
2041
2042         iod = nvmet_fc_alloc_ls_iod(tgtport);
2043         if (!iod) {
2044                 dev_info(tgtport->dev,
2045                         "RCV %s LS failed: context allocation failed\n",
2046                         (w0->ls_cmd <= NVME_FC_LAST_LS_CMD_VALUE) ?
2047                                 nvmefc_ls_names[w0->ls_cmd] : "");
2048                 nvmet_fc_tgtport_put(tgtport);
2049                 return -ENOENT;
2050         }
2051
2052         iod->lsrsp = lsrsp;
2053         iod->fcpreq = NULL;
2054         memcpy(iod->rqstbuf, lsreqbuf, lsreqbuf_len);
2055         iod->rqstdatalen = lsreqbuf_len;
2056         iod->hosthandle = hosthandle;
2057
2058         schedule_work(&iod->work);
2059
2060         return 0;
2061 }
2062 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_ls_req);
2063
2064
2065 /*
2066  * **********************
2067  * Start of FCP handling
2068  * **********************
2069  */
2070
2071 static int
2072 nvmet_fc_alloc_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
2073 {
2074         struct scatterlist *sg;
2075         unsigned int nent;
2076
2077         sg = sgl_alloc(fod->req.transfer_len, GFP_KERNEL, &nent);
2078         if (!sg)
2079                 goto out;
2080
2081         fod->data_sg = sg;
2082         fod->data_sg_cnt = nent;
2083         fod->data_sg_cnt = fc_dma_map_sg(fod->tgtport->dev, sg, nent,
2084                                 ((fod->io_dir == NVMET_FCP_WRITE) ?
2085                                         DMA_FROM_DEVICE : DMA_TO_DEVICE));
2086                                 /* note: write from initiator perspective */
2087         fod->next_sg = fod->data_sg;
2088
2089         return 0;
2090
2091 out:
2092         return NVME_SC_INTERNAL;
2093 }
2094
2095 static void
2096 nvmet_fc_free_tgt_pgs(struct nvmet_fc_fcp_iod *fod)
2097 {
2098         if (!fod->data_sg || !fod->data_sg_cnt)
2099                 return;
2100
2101         fc_dma_unmap_sg(fod->tgtport->dev, fod->data_sg, fod->data_sg_cnt,
2102                                 ((fod->io_dir == NVMET_FCP_WRITE) ?
2103                                         DMA_FROM_DEVICE : DMA_TO_DEVICE));
2104         sgl_free(fod->data_sg);
2105         fod->data_sg = NULL;
2106         fod->data_sg_cnt = 0;
2107 }
2108
2109
2110 static bool
2111 queue_90percent_full(struct nvmet_fc_tgt_queue *q, u32 sqhd)
2112 {
2113         u32 sqtail, used;
2114
2115         /* egad, this is ugly. And sqtail is just a best guess */
2116         sqtail = atomic_read(&q->sqtail) % q->sqsize;
2117
2118         used = (sqtail < sqhd) ? (sqtail + q->sqsize - sqhd) : (sqtail - sqhd);
2119         return ((used * 10) >= (((u32)(q->sqsize - 1) * 9)));
2120 }
2121
2122 /*
2123  * Prep RSP payload.
2124  * May be a NVMET_FCOP_RSP or NVMET_FCOP_READDATA_RSP op
2125  */
2126 static void
2127 nvmet_fc_prep_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
2128                                 struct nvmet_fc_fcp_iod *fod)
2129 {
2130         struct nvme_fc_ersp_iu *ersp = &fod->rspiubuf;
2131         struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
2132         struct nvme_completion *cqe = &ersp->cqe;
2133         u32 *cqewd = (u32 *)cqe;
2134         bool send_ersp = false;
2135         u32 rsn, rspcnt, xfr_length;
2136
2137         if (fod->fcpreq->op == NVMET_FCOP_READDATA_RSP)
2138                 xfr_length = fod->req.transfer_len;
2139         else
2140                 xfr_length = fod->offset;
2141
2142         /*
2143          * check to see if we can send a 0's rsp.
2144          *   Note: to send a 0's response, the NVME-FC host transport will
2145          *   recreate the CQE. The host transport knows: sq id, SQHD (last
2146          *   seen in an ersp), and command_id. Thus it will create a
2147          *   zero-filled CQE with those known fields filled in. Transport
2148          *   must send an ersp for any condition where the cqe won't match
2149          *   this.
2150          *
2151          * Here are the FC-NVME mandated cases where we must send an ersp:
2152          *  every N responses, where N=ersp_ratio
2153          *  force fabric commands to send ersp's (not in FC-NVME but good
2154          *    practice)
2155          *  normal cmds: any time status is non-zero, or status is zero
2156          *     but words 0 or 1 are non-zero.
2157          *  the SQ is 90% or more full
2158          *  the cmd is a fused command
2159          *  transferred data length not equal to cmd iu length
2160          */
2161         rspcnt = atomic_inc_return(&fod->queue->zrspcnt);
2162         if (!(rspcnt % fod->queue->ersp_ratio) ||
2163             nvme_is_fabrics((struct nvme_command *) sqe) ||
2164             xfr_length != fod->req.transfer_len ||
2165             (le16_to_cpu(cqe->status) & 0xFFFE) || cqewd[0] || cqewd[1] ||
2166             (sqe->flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND)) ||
2167             queue_90percent_full(fod->queue, le16_to_cpu(cqe->sq_head)))
2168                 send_ersp = true;
2169
2170         /* re-set the fields */
2171         fod->fcpreq->rspaddr = ersp;
2172         fod->fcpreq->rspdma = fod->rspdma;
2173
2174         if (!send_ersp) {
2175                 memset(ersp, 0, NVME_FC_SIZEOF_ZEROS_RSP);
2176                 fod->fcpreq->rsplen = NVME_FC_SIZEOF_ZEROS_RSP;
2177         } else {
2178                 ersp->iu_len = cpu_to_be16(sizeof(*ersp)/sizeof(u32));
2179                 rsn = atomic_inc_return(&fod->queue->rsn);
2180                 ersp->rsn = cpu_to_be32(rsn);
2181                 ersp->xfrd_len = cpu_to_be32(xfr_length);
2182                 fod->fcpreq->rsplen = sizeof(*ersp);
2183         }
2184
2185         fc_dma_sync_single_for_device(tgtport->dev, fod->rspdma,
2186                                   sizeof(fod->rspiubuf), DMA_TO_DEVICE);
2187 }
2188
2189 static void nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq);
2190
2191 static void
2192 nvmet_fc_abort_op(struct nvmet_fc_tgtport *tgtport,
2193                                 struct nvmet_fc_fcp_iod *fod)
2194 {
2195         struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2196
2197         /* data no longer needed */
2198         nvmet_fc_free_tgt_pgs(fod);
2199
2200         /*
2201          * if an ABTS was received or we issued the fcp_abort early
2202          * don't call abort routine again.
2203          */
2204         /* no need to take lock - lock was taken earlier to get here */
2205         if (!fod->aborted)
2206                 tgtport->ops->fcp_abort(&tgtport->fc_target_port, fcpreq);
2207
2208         nvmet_fc_free_fcp_iod(fod->queue, fod);
2209 }
2210
2211 static void
2212 nvmet_fc_xmt_fcp_rsp(struct nvmet_fc_tgtport *tgtport,
2213                                 struct nvmet_fc_fcp_iod *fod)
2214 {
2215         int ret;
2216
2217         fod->fcpreq->op = NVMET_FCOP_RSP;
2218         fod->fcpreq->timeout = 0;
2219
2220         nvmet_fc_prep_fcp_rsp(tgtport, fod);
2221
2222         ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
2223         if (ret)
2224                 nvmet_fc_abort_op(tgtport, fod);
2225 }
2226
2227 static void
2228 nvmet_fc_transfer_fcp_data(struct nvmet_fc_tgtport *tgtport,
2229                                 struct nvmet_fc_fcp_iod *fod, u8 op)
2230 {
2231         struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2232         struct scatterlist *sg = fod->next_sg;
2233         unsigned long flags;
2234         u32 remaininglen = fod->req.transfer_len - fod->offset;
2235         u32 tlen = 0;
2236         int ret;
2237
2238         fcpreq->op = op;
2239         fcpreq->offset = fod->offset;
2240         fcpreq->timeout = NVME_FC_TGTOP_TIMEOUT_SEC;
2241
2242         /*
2243          * for next sequence:
2244          *  break at a sg element boundary
2245          *  attempt to keep sequence length capped at
2246          *    NVMET_FC_MAX_SEQ_LENGTH but allow sequence to
2247          *    be longer if a single sg element is larger
2248          *    than that amount. This is done to avoid creating
2249          *    a new sg list to use for the tgtport api.
2250          */
2251         fcpreq->sg = sg;
2252         fcpreq->sg_cnt = 0;
2253         while (tlen < remaininglen &&
2254                fcpreq->sg_cnt < tgtport->max_sg_cnt &&
2255                tlen + sg_dma_len(sg) < NVMET_FC_MAX_SEQ_LENGTH) {
2256                 fcpreq->sg_cnt++;
2257                 tlen += sg_dma_len(sg);
2258                 sg = sg_next(sg);
2259         }
2260         if (tlen < remaininglen && fcpreq->sg_cnt == 0) {
2261                 fcpreq->sg_cnt++;
2262                 tlen += min_t(u32, sg_dma_len(sg), remaininglen);
2263                 sg = sg_next(sg);
2264         }
2265         if (tlen < remaininglen)
2266                 fod->next_sg = sg;
2267         else
2268                 fod->next_sg = NULL;
2269
2270         fcpreq->transfer_length = tlen;
2271         fcpreq->transferred_length = 0;
2272         fcpreq->fcp_error = 0;
2273         fcpreq->rsplen = 0;
2274
2275         /*
2276          * If the last READDATA request: check if LLDD supports
2277          * combined xfr with response.
2278          */
2279         if ((op == NVMET_FCOP_READDATA) &&
2280             ((fod->offset + fcpreq->transfer_length) == fod->req.transfer_len) &&
2281             (tgtport->ops->target_features & NVMET_FCTGTFEAT_READDATA_RSP)) {
2282                 fcpreq->op = NVMET_FCOP_READDATA_RSP;
2283                 nvmet_fc_prep_fcp_rsp(tgtport, fod);
2284         }
2285
2286         ret = tgtport->ops->fcp_op(&tgtport->fc_target_port, fod->fcpreq);
2287         if (ret) {
2288                 /*
2289                  * should be ok to set w/o lock as its in the thread of
2290                  * execution (not an async timer routine) and doesn't
2291                  * contend with any clearing action
2292                  */
2293                 fod->abort = true;
2294
2295                 if (op == NVMET_FCOP_WRITEDATA) {
2296                         spin_lock_irqsave(&fod->flock, flags);
2297                         fod->writedataactive = false;
2298                         spin_unlock_irqrestore(&fod->flock, flags);
2299                         nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2300                 } else /* NVMET_FCOP_READDATA or NVMET_FCOP_READDATA_RSP */ {
2301                         fcpreq->fcp_error = ret;
2302                         fcpreq->transferred_length = 0;
2303                         nvmet_fc_xmt_fcp_op_done(fod->fcpreq);
2304                 }
2305         }
2306 }
2307
2308 static inline bool
2309 __nvmet_fc_fod_op_abort(struct nvmet_fc_fcp_iod *fod, bool abort)
2310 {
2311         struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2312         struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2313
2314         /* if in the middle of an io and we need to tear down */
2315         if (abort) {
2316                 if (fcpreq->op == NVMET_FCOP_WRITEDATA) {
2317                         nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2318                         return true;
2319                 }
2320
2321                 nvmet_fc_abort_op(tgtport, fod);
2322                 return true;
2323         }
2324
2325         return false;
2326 }
2327
2328 /*
2329  * actual done handler for FCP operations when completed by the lldd
2330  */
2331 static void
2332 nvmet_fc_fod_op_done(struct nvmet_fc_fcp_iod *fod)
2333 {
2334         struct nvmefc_tgt_fcp_req *fcpreq = fod->fcpreq;
2335         struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2336         unsigned long flags;
2337         bool abort;
2338
2339         spin_lock_irqsave(&fod->flock, flags);
2340         abort = fod->abort;
2341         fod->writedataactive = false;
2342         spin_unlock_irqrestore(&fod->flock, flags);
2343
2344         switch (fcpreq->op) {
2345
2346         case NVMET_FCOP_WRITEDATA:
2347                 if (__nvmet_fc_fod_op_abort(fod, abort))
2348                         return;
2349                 if (fcpreq->fcp_error ||
2350                     fcpreq->transferred_length != fcpreq->transfer_length) {
2351                         spin_lock_irqsave(&fod->flock, flags);
2352                         fod->abort = true;
2353                         spin_unlock_irqrestore(&fod->flock, flags);
2354
2355                         nvmet_req_complete(&fod->req, NVME_SC_INTERNAL);
2356                         return;
2357                 }
2358
2359                 fod->offset += fcpreq->transferred_length;
2360                 if (fod->offset != fod->req.transfer_len) {
2361                         spin_lock_irqsave(&fod->flock, flags);
2362                         fod->writedataactive = true;
2363                         spin_unlock_irqrestore(&fod->flock, flags);
2364
2365                         /* transfer the next chunk */
2366                         nvmet_fc_transfer_fcp_data(tgtport, fod,
2367                                                 NVMET_FCOP_WRITEDATA);
2368                         return;
2369                 }
2370
2371                 /* data transfer complete, resume with nvmet layer */
2372                 fod->req.execute(&fod->req);
2373                 break;
2374
2375         case NVMET_FCOP_READDATA:
2376         case NVMET_FCOP_READDATA_RSP:
2377                 if (__nvmet_fc_fod_op_abort(fod, abort))
2378                         return;
2379                 if (fcpreq->fcp_error ||
2380                     fcpreq->transferred_length != fcpreq->transfer_length) {
2381                         nvmet_fc_abort_op(tgtport, fod);
2382                         return;
2383                 }
2384
2385                 /* success */
2386
2387                 if (fcpreq->op == NVMET_FCOP_READDATA_RSP) {
2388                         /* data no longer needed */
2389                         nvmet_fc_free_tgt_pgs(fod);
2390                         nvmet_fc_free_fcp_iod(fod->queue, fod);
2391                         return;
2392                 }
2393
2394                 fod->offset += fcpreq->transferred_length;
2395                 if (fod->offset != fod->req.transfer_len) {
2396                         /* transfer the next chunk */
2397                         nvmet_fc_transfer_fcp_data(tgtport, fod,
2398                                                 NVMET_FCOP_READDATA);
2399                         return;
2400                 }
2401
2402                 /* data transfer complete, send response */
2403
2404                 /* data no longer needed */
2405                 nvmet_fc_free_tgt_pgs(fod);
2406
2407                 nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2408
2409                 break;
2410
2411         case NVMET_FCOP_RSP:
2412                 if (__nvmet_fc_fod_op_abort(fod, abort))
2413                         return;
2414                 nvmet_fc_free_fcp_iod(fod->queue, fod);
2415                 break;
2416
2417         default:
2418                 break;
2419         }
2420 }
2421
2422 static void
2423 nvmet_fc_xmt_fcp_op_done(struct nvmefc_tgt_fcp_req *fcpreq)
2424 {
2425         struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2426
2427         nvmet_fc_fod_op_done(fod);
2428 }
2429
2430 /*
2431  * actual completion handler after execution by the nvmet layer
2432  */
2433 static void
2434 __nvmet_fc_fcp_nvme_cmd_done(struct nvmet_fc_tgtport *tgtport,
2435                         struct nvmet_fc_fcp_iod *fod, int status)
2436 {
2437         struct nvme_common_command *sqe = &fod->cmdiubuf.sqe.common;
2438         struct nvme_completion *cqe = &fod->rspiubuf.cqe;
2439         unsigned long flags;
2440         bool abort;
2441
2442         spin_lock_irqsave(&fod->flock, flags);
2443         abort = fod->abort;
2444         spin_unlock_irqrestore(&fod->flock, flags);
2445
2446         /* if we have a CQE, snoop the last sq_head value */
2447         if (!status)
2448                 fod->queue->sqhd = cqe->sq_head;
2449
2450         if (abort) {
2451                 nvmet_fc_abort_op(tgtport, fod);
2452                 return;
2453         }
2454
2455         /* if an error handling the cmd post initial parsing */
2456         if (status) {
2457                 /* fudge up a failed CQE status for our transport error */
2458                 memset(cqe, 0, sizeof(*cqe));
2459                 cqe->sq_head = fod->queue->sqhd;        /* echo last cqe sqhd */
2460                 cqe->sq_id = cpu_to_le16(fod->queue->qid);
2461                 cqe->command_id = sqe->command_id;
2462                 cqe->status = cpu_to_le16(status);
2463         } else {
2464
2465                 /*
2466                  * try to push the data even if the SQE status is non-zero.
2467                  * There may be a status where data still was intended to
2468                  * be moved
2469                  */
2470                 if ((fod->io_dir == NVMET_FCP_READ) && (fod->data_sg_cnt)) {
2471                         /* push the data over before sending rsp */
2472                         nvmet_fc_transfer_fcp_data(tgtport, fod,
2473                                                 NVMET_FCOP_READDATA);
2474                         return;
2475                 }
2476
2477                 /* writes & no data - fall thru */
2478         }
2479
2480         /* data no longer needed */
2481         nvmet_fc_free_tgt_pgs(fod);
2482
2483         nvmet_fc_xmt_fcp_rsp(tgtport, fod);
2484 }
2485
2486
2487 static void
2488 nvmet_fc_fcp_nvme_cmd_done(struct nvmet_req *nvme_req)
2489 {
2490         struct nvmet_fc_fcp_iod *fod = nvmet_req_to_fod(nvme_req);
2491         struct nvmet_fc_tgtport *tgtport = fod->tgtport;
2492
2493         __nvmet_fc_fcp_nvme_cmd_done(tgtport, fod, 0);
2494 }
2495
2496
2497 /*
2498  * Actual processing routine for received FC-NVME I/O Requests from the LLD
2499  */
2500 static void
2501 nvmet_fc_handle_fcp_rqst(struct nvmet_fc_tgtport *tgtport,
2502                         struct nvmet_fc_fcp_iod *fod)
2503 {
2504         struct nvme_fc_cmd_iu *cmdiu = &fod->cmdiubuf;
2505         u32 xfrlen = be32_to_cpu(cmdiu->data_len);
2506         int ret;
2507
2508         /*
2509          * Fused commands are currently not supported in the linux
2510          * implementation.
2511          *
2512          * As such, the implementation of the FC transport does not
2513          * look at the fused commands and order delivery to the upper
2514          * layer until we have both based on csn.
2515          */
2516
2517         fod->fcpreq->done = nvmet_fc_xmt_fcp_op_done;
2518
2519         if (cmdiu->flags & FCNVME_CMD_FLAGS_WRITE) {
2520                 fod->io_dir = NVMET_FCP_WRITE;
2521                 if (!nvme_is_write(&cmdiu->sqe))
2522                         goto transport_error;
2523         } else if (cmdiu->flags & FCNVME_CMD_FLAGS_READ) {
2524                 fod->io_dir = NVMET_FCP_READ;
2525                 if (nvme_is_write(&cmdiu->sqe))
2526                         goto transport_error;
2527         } else {
2528                 fod->io_dir = NVMET_FCP_NODATA;
2529                 if (xfrlen)
2530                         goto transport_error;
2531         }
2532
2533         fod->req.cmd = &fod->cmdiubuf.sqe;
2534         fod->req.cqe = &fod->rspiubuf.cqe;
2535         if (!tgtport->pe)
2536                 goto transport_error;
2537         fod->req.port = tgtport->pe->port;
2538
2539         /* clear any response payload */
2540         memset(&fod->rspiubuf, 0, sizeof(fod->rspiubuf));
2541
2542         fod->data_sg = NULL;
2543         fod->data_sg_cnt = 0;
2544
2545         ret = nvmet_req_init(&fod->req,
2546                                 &fod->queue->nvme_cq,
2547                                 &fod->queue->nvme_sq,
2548                                 &nvmet_fc_tgt_fcp_ops);
2549         if (!ret) {
2550                 /* bad SQE content or invalid ctrl state */
2551                 /* nvmet layer has already called op done to send rsp. */
2552                 return;
2553         }
2554
2555         fod->req.transfer_len = xfrlen;
2556
2557         /* keep a running counter of tail position */
2558         atomic_inc(&fod->queue->sqtail);
2559
2560         if (fod->req.transfer_len) {
2561                 ret = nvmet_fc_alloc_tgt_pgs(fod);
2562                 if (ret) {
2563                         nvmet_req_complete(&fod->req, ret);
2564                         return;
2565                 }
2566         }
2567         fod->req.sg = fod->data_sg;
2568         fod->req.sg_cnt = fod->data_sg_cnt;
2569         fod->offset = 0;
2570
2571         if (fod->io_dir == NVMET_FCP_WRITE) {
2572                 /* pull the data over before invoking nvmet layer */
2573                 nvmet_fc_transfer_fcp_data(tgtport, fod, NVMET_FCOP_WRITEDATA);
2574                 return;
2575         }
2576
2577         /*
2578          * Reads or no data:
2579          *
2580          * can invoke the nvmet_layer now. If read data, cmd completion will
2581          * push the data
2582          */
2583         fod->req.execute(&fod->req);
2584         return;
2585
2586 transport_error:
2587         nvmet_fc_abort_op(tgtport, fod);
2588 }
2589
2590 /**
2591  * nvmet_fc_rcv_fcp_req - transport entry point called by an LLDD
2592  *                       upon the reception of a NVME FCP CMD IU.
2593  *
2594  * Pass a FC-NVME FCP CMD IU received from the FC link to the nvmet-fc
2595  * layer for processing.
2596  *
2597  * The nvmet_fc layer allocates a local job structure (struct
2598  * nvmet_fc_fcp_iod) from the queue for the io and copies the
2599  * CMD IU buffer to the job structure. As such, on a successful
2600  * completion (returns 0), the LLDD may immediately free/reuse
2601  * the CMD IU buffer passed in the call.
2602  *
2603  * However, in some circumstances, due to the packetized nature of FC
2604  * and the api of the FC LLDD which may issue a hw command to send the
2605  * response, but the LLDD may not get the hw completion for that command
2606  * and upcall the nvmet_fc layer before a new command may be
2607  * asynchronously received - its possible for a command to be received
2608  * before the LLDD and nvmet_fc have recycled the job structure. It gives
2609  * the appearance of more commands received than fits in the sq.
2610  * To alleviate this scenario, a temporary queue is maintained in the
2611  * transport for pending LLDD requests waiting for a queue job structure.
2612  * In these "overrun" cases, a temporary queue element is allocated
2613  * the LLDD request and CMD iu buffer information remembered, and the
2614  * routine returns a -EOVERFLOW status. Subsequently, when a queue job
2615  * structure is freed, it is immediately reallocated for anything on the
2616  * pending request list. The LLDDs defer_rcv() callback is called,
2617  * informing the LLDD that it may reuse the CMD IU buffer, and the io
2618  * is then started normally with the transport.
2619  *
2620  * The LLDD, when receiving an -EOVERFLOW completion status, is to treat
2621  * the completion as successful but must not reuse the CMD IU buffer
2622  * until the LLDD's defer_rcv() callback has been called for the
2623  * corresponding struct nvmefc_tgt_fcp_req pointer.
2624  *
2625  * If there is any other condition in which an error occurs, the
2626  * transport will return a non-zero status indicating the error.
2627  * In all cases other than -EOVERFLOW, the transport has not accepted the
2628  * request and the LLDD should abort the exchange.
2629  *
2630  * @target_port: pointer to the (registered) target port the FCP CMD IU
2631  *              was received on.
2632  * @fcpreq:     pointer to a fcpreq request structure to be used to reference
2633  *              the exchange corresponding to the FCP Exchange.
2634  * @cmdiubuf:   pointer to the buffer containing the FCP CMD IU
2635  * @cmdiubuf_len: length, in bytes, of the received FCP CMD IU
2636  */
2637 int
2638 nvmet_fc_rcv_fcp_req(struct nvmet_fc_target_port *target_port,
2639                         struct nvmefc_tgt_fcp_req *fcpreq,
2640                         void *cmdiubuf, u32 cmdiubuf_len)
2641 {
2642         struct nvmet_fc_tgtport *tgtport = targetport_to_tgtport(target_port);
2643         struct nvme_fc_cmd_iu *cmdiu = cmdiubuf;
2644         struct nvmet_fc_tgt_queue *queue;
2645         struct nvmet_fc_fcp_iod *fod;
2646         struct nvmet_fc_defer_fcp_req *deferfcp;
2647         unsigned long flags;
2648
2649         /* validate iu, so the connection id can be used to find the queue */
2650         if ((cmdiubuf_len != sizeof(*cmdiu)) ||
2651                         (cmdiu->format_id != NVME_CMD_FORMAT_ID) ||
2652                         (cmdiu->fc_id != NVME_CMD_FC_ID) ||
2653                         (be16_to_cpu(cmdiu->iu_len) != (sizeof(*cmdiu)/4)))
2654                 return -EIO;
2655
2656         queue = nvmet_fc_find_target_queue(tgtport,
2657                                 be64_to_cpu(cmdiu->connection_id));
2658         if (!queue)
2659                 return -ENOTCONN;
2660
2661         /*
2662          * note: reference taken by find_target_queue
2663          * After successful fod allocation, the fod will inherit the
2664          * ownership of that reference and will remove the reference
2665          * when the fod is freed.
2666          */
2667
2668         spin_lock_irqsave(&queue->qlock, flags);
2669
2670         fod = nvmet_fc_alloc_fcp_iod(queue);
2671         if (fod) {
2672                 spin_unlock_irqrestore(&queue->qlock, flags);
2673
2674                 fcpreq->nvmet_fc_private = fod;
2675                 fod->fcpreq = fcpreq;
2676
2677                 memcpy(&fod->cmdiubuf, cmdiubuf, cmdiubuf_len);
2678
2679                 nvmet_fc_queue_fcp_req(tgtport, queue, fcpreq);
2680
2681                 return 0;
2682         }
2683
2684         if (!tgtport->ops->defer_rcv) {
2685                 spin_unlock_irqrestore(&queue->qlock, flags);
2686                 /* release the queue lookup reference */
2687                 nvmet_fc_tgt_q_put(queue);
2688                 return -ENOENT;
2689         }
2690
2691         deferfcp = list_first_entry_or_null(&queue->avail_defer_list,
2692                         struct nvmet_fc_defer_fcp_req, req_list);
2693         if (deferfcp) {
2694                 /* Just re-use one that was previously allocated */
2695                 list_del(&deferfcp->req_list);
2696         } else {
2697                 spin_unlock_irqrestore(&queue->qlock, flags);
2698
2699                 /* Now we need to dynamically allocate one */
2700                 deferfcp = kmalloc(sizeof(*deferfcp), GFP_KERNEL);
2701                 if (!deferfcp) {
2702                         /* release the queue lookup reference */
2703                         nvmet_fc_tgt_q_put(queue);
2704                         return -ENOMEM;
2705                 }
2706                 spin_lock_irqsave(&queue->qlock, flags);
2707         }
2708
2709         /* For now, use rspaddr / rsplen to save payload information */
2710         fcpreq->rspaddr = cmdiubuf;
2711         fcpreq->rsplen  = cmdiubuf_len;
2712         deferfcp->fcp_req = fcpreq;
2713
2714         /* defer processing till a fod becomes available */
2715         list_add_tail(&deferfcp->req_list, &queue->pending_cmd_list);
2716
2717         /* NOTE: the queue lookup reference is still valid */
2718
2719         spin_unlock_irqrestore(&queue->qlock, flags);
2720
2721         return -EOVERFLOW;
2722 }
2723 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_req);
2724
2725 /**
2726  * nvmet_fc_rcv_fcp_abort - transport entry point called by an LLDD
2727  *                       upon the reception of an ABTS for a FCP command
2728  *
2729  * Notify the transport that an ABTS has been received for a FCP command
2730  * that had been given to the transport via nvmet_fc_rcv_fcp_req(). The
2731  * LLDD believes the command is still being worked on
2732  * (template_ops->fcp_req_release() has not been called).
2733  *
2734  * The transport will wait for any outstanding work (an op to the LLDD,
2735  * which the lldd should complete with error due to the ABTS; or the
2736  * completion from the nvmet layer of the nvme command), then will
2737  * stop processing and call the nvmet_fc_rcv_fcp_req() callback to
2738  * return the i/o context to the LLDD.  The LLDD may send the BA_ACC
2739  * to the ABTS either after return from this function (assuming any
2740  * outstanding op work has been terminated) or upon the callback being
2741  * called.
2742  *
2743  * @target_port: pointer to the (registered) target port the FCP CMD IU
2744  *              was received on.
2745  * @fcpreq:     pointer to the fcpreq request structure that corresponds
2746  *              to the exchange that received the ABTS.
2747  */
2748 void
2749 nvmet_fc_rcv_fcp_abort(struct nvmet_fc_target_port *target_port,
2750                         struct nvmefc_tgt_fcp_req *fcpreq)
2751 {
2752         struct nvmet_fc_fcp_iod *fod = fcpreq->nvmet_fc_private;
2753         struct nvmet_fc_tgt_queue *queue;
2754         unsigned long flags;
2755
2756         if (!fod || fod->fcpreq != fcpreq)
2757                 /* job appears to have already completed, ignore abort */
2758                 return;
2759
2760         queue = fod->queue;
2761
2762         spin_lock_irqsave(&queue->qlock, flags);
2763         if (fod->active) {
2764                 /*
2765                  * mark as abort. The abort handler, invoked upon completion
2766                  * of any work, will detect the aborted status and do the
2767                  * callback.
2768                  */
2769                 spin_lock(&fod->flock);
2770                 fod->abort = true;
2771                 fod->aborted = true;
2772                 spin_unlock(&fod->flock);
2773         }
2774         spin_unlock_irqrestore(&queue->qlock, flags);
2775 }
2776 EXPORT_SYMBOL_GPL(nvmet_fc_rcv_fcp_abort);
2777
2778
2779 struct nvmet_fc_traddr {
2780         u64     nn;
2781         u64     pn;
2782 };
2783
2784 static int
2785 __nvme_fc_parse_u64(substring_t *sstr, u64 *val)
2786 {
2787         u64 token64;
2788
2789         if (match_u64(sstr, &token64))
2790                 return -EINVAL;
2791         *val = token64;
2792
2793         return 0;
2794 }
2795
2796 /*
2797  * This routine validates and extracts the WWN's from the TRADDR string.
2798  * As kernel parsers need the 0x to determine number base, universally
2799  * build string to parse with 0x prefix before parsing name strings.
2800  */
2801 static int
2802 nvme_fc_parse_traddr(struct nvmet_fc_traddr *traddr, char *buf, size_t blen)
2803 {
2804         char name[2 + NVME_FC_TRADDR_HEXNAMELEN + 1];
2805         substring_t wwn = { name, &name[sizeof(name)-1] };
2806         int nnoffset, pnoffset;
2807
2808         /* validate if string is one of the 2 allowed formats */
2809         if (strnlen(buf, blen) == NVME_FC_TRADDR_MAXLENGTH &&
2810                         !strncmp(buf, "nn-0x", NVME_FC_TRADDR_OXNNLEN) &&
2811                         !strncmp(&buf[NVME_FC_TRADDR_MAX_PN_OFFSET],
2812                                 "pn-0x", NVME_FC_TRADDR_OXNNLEN)) {
2813                 nnoffset = NVME_FC_TRADDR_OXNNLEN;
2814                 pnoffset = NVME_FC_TRADDR_MAX_PN_OFFSET +
2815                                                 NVME_FC_TRADDR_OXNNLEN;
2816         } else if ((strnlen(buf, blen) == NVME_FC_TRADDR_MINLENGTH &&
2817                         !strncmp(buf, "nn-", NVME_FC_TRADDR_NNLEN) &&
2818                         !strncmp(&buf[NVME_FC_TRADDR_MIN_PN_OFFSET],
2819                                 "pn-", NVME_FC_TRADDR_NNLEN))) {
2820                 nnoffset = NVME_FC_TRADDR_NNLEN;
2821                 pnoffset = NVME_FC_TRADDR_MIN_PN_OFFSET + NVME_FC_TRADDR_NNLEN;
2822         } else
2823                 goto out_einval;
2824
2825         name[0] = '0';
2826         name[1] = 'x';
2827         name[2 + NVME_FC_TRADDR_HEXNAMELEN] = 0;
2828
2829         memcpy(&name[2], &buf[nnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2830         if (__nvme_fc_parse_u64(&wwn, &traddr->nn))
2831                 goto out_einval;
2832
2833         memcpy(&name[2], &buf[pnoffset], NVME_FC_TRADDR_HEXNAMELEN);
2834         if (__nvme_fc_parse_u64(&wwn, &traddr->pn))
2835                 goto out_einval;
2836
2837         return 0;
2838
2839 out_einval:
2840         pr_warn("%s: bad traddr string\n", __func__);
2841         return -EINVAL;
2842 }
2843
2844 static int
2845 nvmet_fc_add_port(struct nvmet_port *port)
2846 {
2847         struct nvmet_fc_tgtport *tgtport;
2848         struct nvmet_fc_port_entry *pe;
2849         struct nvmet_fc_traddr traddr = { 0L, 0L };
2850         unsigned long flags;
2851         int ret;
2852
2853         /* validate the address info */
2854         if ((port->disc_addr.trtype != NVMF_TRTYPE_FC) ||
2855             (port->disc_addr.adrfam != NVMF_ADDR_FAMILY_FC))
2856                 return -EINVAL;
2857
2858         /* map the traddr address info to a target port */
2859
2860         ret = nvme_fc_parse_traddr(&traddr, port->disc_addr.traddr,
2861                         sizeof(port->disc_addr.traddr));
2862         if (ret)
2863                 return ret;
2864
2865         pe = kzalloc(sizeof(*pe), GFP_KERNEL);
2866         if (!pe)
2867                 return -ENOMEM;
2868
2869         ret = -ENXIO;
2870         spin_lock_irqsave(&nvmet_fc_tgtlock, flags);
2871         list_for_each_entry(tgtport, &nvmet_fc_target_list, tgt_list) {
2872                 if ((tgtport->fc_target_port.node_name == traddr.nn) &&
2873                     (tgtport->fc_target_port.port_name == traddr.pn)) {
2874                         /* a FC port can only be 1 nvmet port id */
2875                         if (!tgtport->pe) {
2876                                 nvmet_fc_portentry_bind(tgtport, pe, port);
2877                                 ret = 0;
2878                         } else
2879                                 ret = -EALREADY;
2880                         break;
2881                 }
2882         }
2883         spin_unlock_irqrestore(&nvmet_fc_tgtlock, flags);
2884
2885         if (ret)
2886                 kfree(pe);
2887
2888         return ret;
2889 }
2890
2891 static void
2892 nvmet_fc_remove_port(struct nvmet_port *port)
2893 {
2894         struct nvmet_fc_port_entry *pe = port->priv;
2895
2896         nvmet_fc_portentry_unbind(pe);
2897
2898         kfree(pe);
2899 }
2900
2901 static void
2902 nvmet_fc_discovery_chg(struct nvmet_port *port)
2903 {
2904         struct nvmet_fc_port_entry *pe = port->priv;
2905         struct nvmet_fc_tgtport *tgtport = pe->tgtport;
2906
2907         if (tgtport && tgtport->ops->discovery_event)
2908                 tgtport->ops->discovery_event(&tgtport->fc_target_port);
2909 }
2910
2911 static const struct nvmet_fabrics_ops nvmet_fc_tgt_fcp_ops = {
2912         .owner                  = THIS_MODULE,
2913         .type                   = NVMF_TRTYPE_FC,
2914         .msdbd                  = 1,
2915         .add_port               = nvmet_fc_add_port,
2916         .remove_port            = nvmet_fc_remove_port,
2917         .queue_response         = nvmet_fc_fcp_nvme_cmd_done,
2918         .delete_ctrl            = nvmet_fc_delete_ctrl,
2919         .discovery_chg          = nvmet_fc_discovery_chg,
2920 };
2921
2922 static int __init nvmet_fc_init_module(void)
2923 {
2924         return nvmet_register_transport(&nvmet_fc_tgt_fcp_ops);
2925 }
2926
2927 static void __exit nvmet_fc_exit_module(void)
2928 {
2929         /* sanity check - all lports should be removed */
2930         if (!list_empty(&nvmet_fc_target_list))
2931                 pr_warn("%s: targetport list not empty\n", __func__);
2932
2933         nvmet_unregister_transport(&nvmet_fc_tgt_fcp_ops);
2934
2935         ida_destroy(&nvmet_fc_tgtport_cnt);
2936 }
2937
2938 module_init(nvmet_fc_init_module);
2939 module_exit(nvmet_fc_exit_module);
2940
2941 MODULE_LICENSE("GPL v2");