?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
bpf: Prevent memory disambiguation attack
[linux/fpc-iii.git] / drivers / nvme / target / rdma.c
blob978e169c11bf9af0f15d1b8fa35dc502e364ad56
1 /*
2 * NVMe over Fabrics RDMA target.
3 * Copyright (c) 2015-2016 HGST, a Western Digital Company.
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
12 * more details.
13 */
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 #include <linux/atomic.h>
16 #include <linux/ctype.h>
17 #include <linux/delay.h>
18 #include <linux/err.h>
19 #include <linux/init.h>
20 #include <linux/module.h>
21 #include <linux/nvme.h>
22 #include <linux/slab.h>
23 #include <linux/string.h>
24 #include <linux/wait.h>
25 #include <linux/inet.h>
26 #include <asm/unaligned.h>
27
28 #include <rdma/ib_verbs.h>
29 #include <rdma/rdma_cm.h>
30 #include <rdma/rw.h>
31
32 #include <linux/nvme-rdma.h>
33 #include "nvmet.h"
34
35 /*
36 * We allow up to a page of inline data to go with the SQE
37 */
38 #define NVMET_RDMA_INLINE_DATA_SIZE PAGE_SIZE
39
40 struct nvmet_rdma_cmd {
41 struct ib_sge sge[2];
42 struct ib_cqe cqe;
43 struct ib_recv_wr wr;
44 struct scatterlist inline_sg;
45 struct page *inline_page;
46 struct nvme_command *nvme_cmd;
47 struct nvmet_rdma_queue *queue;
48 };
49
50 enum {
51 NVMET_RDMA_REQ_INLINE_DATA = (1 << 0),
52 NVMET_RDMA_REQ_INVALIDATE_RKEY = (1 << 1),
53 };
54
55 struct nvmet_rdma_rsp {
56 struct ib_sge send_sge;
57 struct ib_cqe send_cqe;
58 struct ib_send_wr send_wr;
59
60 struct nvmet_rdma_cmd *cmd;
61 struct nvmet_rdma_queue *queue;
62
63 struct ib_cqe read_cqe;
64 struct rdma_rw_ctx rw;
65
66 struct nvmet_req req;
67
68 u8 n_rdma;
69 u32 flags;
70 u32 invalidate_rkey;
71
72 struct list_head wait_list;
73 struct list_head free_list;
74 };
75
76 enum nvmet_rdma_queue_state {
77 NVMET_RDMA_Q_CONNECTING,
78 NVMET_RDMA_Q_LIVE,
79 NVMET_RDMA_Q_DISCONNECTING,
80 NVMET_RDMA_IN_DEVICE_REMOVAL,
81 };
82
83 struct nvmet_rdma_queue {
84 struct rdma_cm_id *cm_id;
85 struct nvmet_port *port;
86 struct ib_cq *cq;
87 atomic_t sq_wr_avail;
88 struct nvmet_rdma_device *dev;
89 spinlock_t state_lock;
90 enum nvmet_rdma_queue_state state;
91 struct nvmet_cq nvme_cq;
92 struct nvmet_sq nvme_sq;
93
94 struct nvmet_rdma_rsp *rsps;
95 struct list_head free_rsps;
96 spinlock_t rsps_lock;
97 struct nvmet_rdma_cmd *cmds;
98
99 struct work_struct release_work;
100 struct list_head rsp_wait_list;
101 struct list_head rsp_wr_wait_list;
102 spinlock_t rsp_wr_wait_lock;
103
104 int idx;
105 int host_qid;
106 int recv_queue_size;
107 int send_queue_size;
108
109 struct list_head queue_list;
110 };
111
112 struct nvmet_rdma_device {
113 struct ib_device *device;
114 struct ib_pd *pd;
115 struct ib_srq *srq;
116 struct nvmet_rdma_cmd *srq_cmds;
117 size_t srq_size;
118 struct kref ref;
119 struct list_head entry;
120 };
121
122 static bool nvmet_rdma_use_srq;
123 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
124 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
125
126 static DEFINE_IDA(nvmet_rdma_queue_ida);
127 static LIST_HEAD(nvmet_rdma_queue_list);
128 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
129
130 static LIST_HEAD(device_list);
131 static DEFINE_MUTEX(device_list_mutex);
132
133 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
134 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
135 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
136 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
137 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
138 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
139
140 static struct nvmet_fabrics_ops nvmet_rdma_ops;
141
142 /* XXX: really should move to a generic header sooner or later.. */
143 static inline u32 get_unaligned_le24(const u8 *p)
144 {
145 return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
146 }
147
148 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
149 {
150 return nvme_is_write(rsp->req.cmd) &&
151 rsp->req.transfer_len &&
152 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
153 }
154
155 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
156 {
157 return !nvme_is_write(rsp->req.cmd) &&
158 rsp->req.transfer_len &&
159 !rsp->req.rsp->status &&
160 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
161 }
162
163 static inline struct nvmet_rdma_rsp *
164 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
165 {
166 struct nvmet_rdma_rsp *rsp;
167 unsigned long flags;
168
169 spin_lock_irqsave(&queue->rsps_lock, flags);
170 rsp = list_first_entry(&queue->free_rsps,
171 struct nvmet_rdma_rsp, free_list);
172 list_del(&rsp->free_list);
173 spin_unlock_irqrestore(&queue->rsps_lock, flags);
174
175 return rsp;
176 }
177
178 static inline void
179 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
180 {
181 unsigned long flags;
182
183 spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
184 list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
185 spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
186 }
187
188 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
189 struct nvmet_rdma_cmd *c, bool admin)
190 {
191 /* NVMe command / RDMA RECV */
192 c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
193 if (!c->nvme_cmd)
194 goto out;
195
196 c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
197 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
198 if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
199 goto out_free_cmd;
200
201 c->sge[0].length = sizeof(*c->nvme_cmd);
202 c->sge[0].lkey = ndev->pd->local_dma_lkey;
203
204 if (!admin) {
205 c->inline_page = alloc_pages(GFP_KERNEL,
206 get_order(NVMET_RDMA_INLINE_DATA_SIZE));
207 if (!c->inline_page)
208 goto out_unmap_cmd;
209 c->sge[1].addr = ib_dma_map_page(ndev->device,
210 c->inline_page, 0, NVMET_RDMA_INLINE_DATA_SIZE,
211 DMA_FROM_DEVICE);
212 if (ib_dma_mapping_error(ndev->device, c->sge[1].addr))
213 goto out_free_inline_page;
214 c->sge[1].length = NVMET_RDMA_INLINE_DATA_SIZE;
215 c->sge[1].lkey = ndev->pd->local_dma_lkey;
216 }
217
218 c->cqe.done = nvmet_rdma_recv_done;
219
220 c->wr.wr_cqe = &c->cqe;
221 c->wr.sg_list = c->sge;
222 c->wr.num_sge = admin ? 1 : 2;
223
224 return 0;
225
226 out_free_inline_page:
227 if (!admin) {
228 __free_pages(c->inline_page,
229 get_order(NVMET_RDMA_INLINE_DATA_SIZE));
230 }
231 out_unmap_cmd:
232 ib_dma_unmap_single(ndev->device, c->sge[0].addr,
233 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
234 out_free_cmd:
235 kfree(c->nvme_cmd);
236
237 out:
238 return -ENOMEM;
239 }
240
241 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
242 struct nvmet_rdma_cmd *c, bool admin)
243 {
244 if (!admin) {
245 ib_dma_unmap_page(ndev->device, c->sge[1].addr,
246 NVMET_RDMA_INLINE_DATA_SIZE, DMA_FROM_DEVICE);
247 __free_pages(c->inline_page,
248 get_order(NVMET_RDMA_INLINE_DATA_SIZE));
249 }
250 ib_dma_unmap_single(ndev->device, c->sge[0].addr,
251 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
252 kfree(c->nvme_cmd);
253 }
254
255 static struct nvmet_rdma_cmd *
256 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
257 int nr_cmds, bool admin)
258 {
259 struct nvmet_rdma_cmd *cmds;
260 int ret = -EINVAL, i;
261
262 cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
263 if (!cmds)
264 goto out;
265
266 for (i = 0; i < nr_cmds; i++) {
267 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
268 if (ret)
269 goto out_free;
270 }
271
272 return cmds;
273
274 out_free:
275 while (--i >= 0)
276 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
277 kfree(cmds);
278 out:
279 return ERR_PTR(ret);
280 }
281
282 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
283 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
284 {
285 int i;
286
287 for (i = 0; i < nr_cmds; i++)
288 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
289 kfree(cmds);
290 }
291
292 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
293 struct nvmet_rdma_rsp *r)
294 {
295 /* NVMe CQE / RDMA SEND */
296 r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
297 if (!r->req.rsp)
298 goto out;
299
300 r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
301 sizeof(*r->req.rsp), DMA_TO_DEVICE);
302 if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
303 goto out_free_rsp;
304
305 r->send_sge.length = sizeof(*r->req.rsp);
306 r->send_sge.lkey = ndev->pd->local_dma_lkey;
307
308 r->send_cqe.done = nvmet_rdma_send_done;
309
310 r->send_wr.wr_cqe = &r->send_cqe;
311 r->send_wr.sg_list = &r->send_sge;
312 r->send_wr.num_sge = 1;
313 r->send_wr.send_flags = IB_SEND_SIGNALED;
314
315 /* Data In / RDMA READ */
316 r->read_cqe.done = nvmet_rdma_read_data_done;
317 return 0;
318
319 out_free_rsp:
320 kfree(r->req.rsp);
321 out:
322 return -ENOMEM;
323 }
324
325 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
326 struct nvmet_rdma_rsp *r)
327 {
328 ib_dma_unmap_single(ndev->device, r->send_sge.addr,
329 sizeof(*r->req.rsp), DMA_TO_DEVICE);
330 kfree(r->req.rsp);
331 }
332
333 static int
334 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
335 {
336 struct nvmet_rdma_device *ndev = queue->dev;
337 int nr_rsps = queue->recv_queue_size * 2;
338 int ret = -EINVAL, i;
339
340 queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
341 GFP_KERNEL);
342 if (!queue->rsps)
343 goto out;
344
345 for (i = 0; i < nr_rsps; i++) {
346 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
347
348 ret = nvmet_rdma_alloc_rsp(ndev, rsp);
349 if (ret)
350 goto out_free;
351
352 list_add_tail(&rsp->free_list, &queue->free_rsps);
353 }
354
355 return 0;
356
357 out_free:
358 while (--i >= 0) {
359 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
360
361 list_del(&rsp->free_list);
362 nvmet_rdma_free_rsp(ndev, rsp);
363 }
364 kfree(queue->rsps);
365 out:
366 return ret;
367 }
368
369 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
370 {
371 struct nvmet_rdma_device *ndev = queue->dev;
372 int i, nr_rsps = queue->recv_queue_size * 2;
373
374 for (i = 0; i < nr_rsps; i++) {
375 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
376
377 list_del(&rsp->free_list);
378 nvmet_rdma_free_rsp(ndev, rsp);
379 }
380 kfree(queue->rsps);
381 }
382
383 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
384 struct nvmet_rdma_cmd *cmd)
385 {
386 struct ib_recv_wr *bad_wr;
387
388 ib_dma_sync_single_for_device(ndev->device,
389 cmd->sge[0].addr, cmd->sge[0].length,
390 DMA_FROM_DEVICE);
391
392 if (ndev->srq)
393 return ib_post_srq_recv(ndev->srq, &cmd->wr, &bad_wr);
394 return ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, &bad_wr);
395 }
396
397 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
398 {
399 spin_lock(&queue->rsp_wr_wait_lock);
400 while (!list_empty(&queue->rsp_wr_wait_list)) {
401 struct nvmet_rdma_rsp *rsp;
402 bool ret;
403
404 rsp = list_entry(queue->rsp_wr_wait_list.next,
405 struct nvmet_rdma_rsp, wait_list);
406 list_del(&rsp->wait_list);
407
408 spin_unlock(&queue->rsp_wr_wait_lock);
409 ret = nvmet_rdma_execute_command(rsp);
410 spin_lock(&queue->rsp_wr_wait_lock);
411
412 if (!ret) {
413 list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
414 break;
415 }
416 }
417 spin_unlock(&queue->rsp_wr_wait_lock);
418 }
419
420
421 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
422 {
423 struct nvmet_rdma_queue *queue = rsp->queue;
424
425 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
426
427 if (rsp->n_rdma) {
428 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
429 queue->cm_id->port_num, rsp->req.sg,
430 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
431 }
432
433 if (rsp->req.sg != &rsp->cmd->inline_sg)
434 sgl_free(rsp->req.sg);
435
436 if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
437 nvmet_rdma_process_wr_wait_list(queue);
438
439 nvmet_rdma_put_rsp(rsp);
440 }
441
442 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
443 {
444 if (queue->nvme_sq.ctrl) {
445 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
446 } else {
447 /*
448 * we didn't setup the controller yet in case
449 * of admin connect error, just disconnect and
450 * cleanup the queue
451 */
452 nvmet_rdma_queue_disconnect(queue);
453 }
454 }
455
456 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
457 {
458 struct nvmet_rdma_rsp *rsp =
459 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
460
461 nvmet_rdma_release_rsp(rsp);
462
463 if (unlikely(wc->status != IB_WC_SUCCESS &&
464 wc->status != IB_WC_WR_FLUSH_ERR)) {
465 pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
466 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
467 nvmet_rdma_error_comp(rsp->queue);
468 }
469 }
470
471 static void nvmet_rdma_queue_response(struct nvmet_req *req)
472 {
473 struct nvmet_rdma_rsp *rsp =
474 container_of(req, struct nvmet_rdma_rsp, req);
475 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
476 struct ib_send_wr *first_wr, *bad_wr;
477
478 if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
479 rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
480 rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
481 } else {
482 rsp->send_wr.opcode = IB_WR_SEND;
483 }
484
485 if (nvmet_rdma_need_data_out(rsp))
486 first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
487 cm_id->port_num, NULL, &rsp->send_wr);
488 else
489 first_wr = &rsp->send_wr;
490
491 nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
492
493 ib_dma_sync_single_for_device(rsp->queue->dev->device,
494 rsp->send_sge.addr, rsp->send_sge.length,
495 DMA_TO_DEVICE);
496
497 if (ib_post_send(cm_id->qp, first_wr, &bad_wr)) {
498 pr_err("sending cmd response failed\n");
499 nvmet_rdma_release_rsp(rsp);
500 }
501 }
502
503 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
504 {
505 struct nvmet_rdma_rsp *rsp =
506 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
507 struct nvmet_rdma_queue *queue = cq->cq_context;
508
509 WARN_ON(rsp->n_rdma <= 0);
510 atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
511 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
512 queue->cm_id->port_num, rsp->req.sg,
513 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
514 rsp->n_rdma = 0;
515
516 if (unlikely(wc->status != IB_WC_SUCCESS)) {
517 nvmet_req_uninit(&rsp->req);
518 nvmet_rdma_release_rsp(rsp);
519 if (wc->status != IB_WC_WR_FLUSH_ERR) {
520 pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
521 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
522 nvmet_rdma_error_comp(queue);
523 }
524 return;
525 }
526
527 nvmet_req_execute(&rsp->req);
528 }
529
530 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
531 u64 off)
532 {
533 sg_init_table(&rsp->cmd->inline_sg, 1);
534 sg_set_page(&rsp->cmd->inline_sg, rsp->cmd->inline_page, len, off);
535 rsp->req.sg = &rsp->cmd->inline_sg;
536 rsp->req.sg_cnt = 1;
537 }
538
539 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
540 {
541 struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
542 u64 off = le64_to_cpu(sgl->addr);
543 u32 len = le32_to_cpu(sgl->length);
544
545 if (!nvme_is_write(rsp->req.cmd))
546 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
547
548 if (off + len > NVMET_RDMA_INLINE_DATA_SIZE) {
549 pr_err("invalid inline data offset!\n");
550 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
551 }
552
553 /* no data command? */
554 if (!len)
555 return 0;
556
557 nvmet_rdma_use_inline_sg(rsp, len, off);
558 rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
559 rsp->req.transfer_len += len;
560 return 0;
561 }
562
563 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
564 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
565 {
566 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
567 u64 addr = le64_to_cpu(sgl->addr);
568 u32 len = get_unaligned_le24(sgl->length);
569 u32 key = get_unaligned_le32(sgl->key);
570 int ret;
571
572 /* no data command? */
573 if (!len)
574 return 0;
575
576 rsp->req.sg = sgl_alloc(len, GFP_KERNEL, &rsp->req.sg_cnt);
577 if (!rsp->req.sg)
578 return NVME_SC_INTERNAL;
579
580 ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
581 rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
582 nvmet_data_dir(&rsp->req));
583 if (ret < 0)
584 return NVME_SC_INTERNAL;
585 rsp->req.transfer_len += len;
586 rsp->n_rdma += ret;
587
588 if (invalidate) {
589 rsp->invalidate_rkey = key;
590 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
591 }
592
593 return 0;
594 }
595
596 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
597 {
598 struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
599
600 switch (sgl->type >> 4) {
601 case NVME_SGL_FMT_DATA_DESC:
602 switch (sgl->type & 0xf) {
603 case NVME_SGL_FMT_OFFSET:
604 return nvmet_rdma_map_sgl_inline(rsp);
605 default:
606 pr_err("invalid SGL subtype: %#x\n", sgl->type);
607 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
608 }
609 case NVME_KEY_SGL_FMT_DATA_DESC:
610 switch (sgl->type & 0xf) {
611 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
612 return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
613 case NVME_SGL_FMT_ADDRESS:
614 return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
615 default:
616 pr_err("invalid SGL subtype: %#x\n", sgl->type);
617 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
618 }
619 default:
620 pr_err("invalid SGL type: %#x\n", sgl->type);
621 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
622 }
623 }
624
625 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
626 {
627 struct nvmet_rdma_queue *queue = rsp->queue;
628
629 if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
630 &queue->sq_wr_avail) < 0)) {
631 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
632 1 + rsp->n_rdma, queue->idx,
633 queue->nvme_sq.ctrl->cntlid);
634 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
635 return false;
636 }
637
638 if (nvmet_rdma_need_data_in(rsp)) {
639 if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
640 queue->cm_id->port_num, &rsp->read_cqe, NULL))
641 nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
642 } else {
643 nvmet_req_execute(&rsp->req);
644 }
645
646 return true;
647 }
648
649 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
650 struct nvmet_rdma_rsp *cmd)
651 {
652 u16 status;
653
654 ib_dma_sync_single_for_cpu(queue->dev->device,
655 cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
656 DMA_FROM_DEVICE);
657 ib_dma_sync_single_for_cpu(queue->dev->device,
658 cmd->send_sge.addr, cmd->send_sge.length,
659 DMA_TO_DEVICE);
660
661 if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
662 &queue->nvme_sq, &nvmet_rdma_ops))
663 return;
664
665 status = nvmet_rdma_map_sgl(cmd);
666 if (status)
667 goto out_err;
668
669 if (unlikely(!nvmet_rdma_execute_command(cmd))) {
670 spin_lock(&queue->rsp_wr_wait_lock);
671 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
672 spin_unlock(&queue->rsp_wr_wait_lock);
673 }
674
675 return;
676
677 out_err:
678 nvmet_req_complete(&cmd->req, status);
679 }
680
681 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
682 {
683 struct nvmet_rdma_cmd *cmd =
684 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
685 struct nvmet_rdma_queue *queue = cq->cq_context;
686 struct nvmet_rdma_rsp *rsp;
687
688 if (unlikely(wc->status != IB_WC_SUCCESS)) {
689 if (wc->status != IB_WC_WR_FLUSH_ERR) {
690 pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
691 wc->wr_cqe, ib_wc_status_msg(wc->status),
692 wc->status);
693 nvmet_rdma_error_comp(queue);
694 }
695 return;
696 }
697
698 if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
699 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
700 nvmet_rdma_error_comp(queue);
701 return;
702 }
703
704 cmd->queue = queue;
705 rsp = nvmet_rdma_get_rsp(queue);
706 rsp->queue = queue;
707 rsp->cmd = cmd;
708 rsp->flags = 0;
709 rsp->req.cmd = cmd->nvme_cmd;
710 rsp->req.port = queue->port;
711 rsp->n_rdma = 0;
712
713 if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
714 unsigned long flags;
715
716 spin_lock_irqsave(&queue->state_lock, flags);
717 if (queue->state == NVMET_RDMA_Q_CONNECTING)
718 list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
719 else
720 nvmet_rdma_put_rsp(rsp);
721 spin_unlock_irqrestore(&queue->state_lock, flags);
722 return;
723 }
724
725 nvmet_rdma_handle_command(queue, rsp);
726 }
727
728 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
729 {
730 if (!ndev->srq)
731 return;
732
733 nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
734 ib_destroy_srq(ndev->srq);
735 }
736
737 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
738 {
739 struct ib_srq_init_attr srq_attr = { NULL, };
740 struct ib_srq *srq;
741 size_t srq_size;
742 int ret, i;
743
744 srq_size = 4095; /* XXX: tune */
745
746 srq_attr.attr.max_wr = srq_size;
747 srq_attr.attr.max_sge = 2;
748 srq_attr.attr.srq_limit = 0;
749 srq_attr.srq_type = IB_SRQT_BASIC;
750 srq = ib_create_srq(ndev->pd, &srq_attr);
751 if (IS_ERR(srq)) {
752 /*
753 * If SRQs aren't supported we just go ahead and use normal
754 * non-shared receive queues.
755 */
756 pr_info("SRQ requested but not supported.\n");
757 return 0;
758 }
759
760 ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
761 if (IS_ERR(ndev->srq_cmds)) {
762 ret = PTR_ERR(ndev->srq_cmds);
763 goto out_destroy_srq;
764 }
765
766 ndev->srq = srq;
767 ndev->srq_size = srq_size;
768
769 for (i = 0; i < srq_size; i++)
770 nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
771
772 return 0;
773
774 out_destroy_srq:
775 ib_destroy_srq(srq);
776 return ret;
777 }
778
779 static void nvmet_rdma_free_dev(struct kref *ref)
780 {
781 struct nvmet_rdma_device *ndev =
782 container_of(ref, struct nvmet_rdma_device, ref);
783
784 mutex_lock(&device_list_mutex);
785 list_del(&ndev->entry);
786 mutex_unlock(&device_list_mutex);
787
788 nvmet_rdma_destroy_srq(ndev);
789 ib_dealloc_pd(ndev->pd);
790
791 kfree(ndev);
792 }
793
794 static struct nvmet_rdma_device *
795 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
796 {
797 struct nvmet_rdma_device *ndev;
798 int ret;
799
800 mutex_lock(&device_list_mutex);
801 list_for_each_entry(ndev, &device_list, entry) {
802 if (ndev->device->node_guid == cm_id->device->node_guid &&
803 kref_get_unless_zero(&ndev->ref))
804 goto out_unlock;
805 }
806
807 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
808 if (!ndev)
809 goto out_err;
810
811 ndev->device = cm_id->device;
812 kref_init(&ndev->ref);
813
814 ndev->pd = ib_alloc_pd(ndev->device, 0);
815 if (IS_ERR(ndev->pd))
816 goto out_free_dev;
817
818 if (nvmet_rdma_use_srq) {
819 ret = nvmet_rdma_init_srq(ndev);
820 if (ret)
821 goto out_free_pd;
822 }
823
824 list_add(&ndev->entry, &device_list);
825 out_unlock:
826 mutex_unlock(&device_list_mutex);
827 pr_debug("added %s.\n", ndev->device->name);
828 return ndev;
829
830 out_free_pd:
831 ib_dealloc_pd(ndev->pd);
832 out_free_dev:
833 kfree(ndev);
834 out_err:
835 mutex_unlock(&device_list_mutex);
836 return NULL;
837 }
838
839 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
840 {
841 struct ib_qp_init_attr qp_attr;
842 struct nvmet_rdma_device *ndev = queue->dev;
843 int comp_vector, nr_cqe, ret, i;
844
845 /*
846 * Spread the io queues across completion vectors,
847 * but still keep all admin queues on vector 0.
848 */
849 comp_vector = !queue->host_qid ? 0 :
850 queue->idx % ndev->device->num_comp_vectors;
851
852 /*
853 * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
854 */
855 nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
856
857 queue->cq = ib_alloc_cq(ndev->device, queue,
858 nr_cqe + 1, comp_vector,
859 IB_POLL_WORKQUEUE);
860 if (IS_ERR(queue->cq)) {
861 ret = PTR_ERR(queue->cq);
862 pr_err("failed to create CQ cqe= %d ret= %d\n",
863 nr_cqe + 1, ret);
864 goto out;
865 }
866
867 memset(&qp_attr, 0, sizeof(qp_attr));
868 qp_attr.qp_context = queue;
869 qp_attr.event_handler = nvmet_rdma_qp_event;
870 qp_attr.send_cq = queue->cq;
871 qp_attr.recv_cq = queue->cq;
872 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
873 qp_attr.qp_type = IB_QPT_RC;
874 /* +1 for drain */
875 qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
876 qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
877 qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
878 ndev->device->attrs.max_sge);
879
880 if (ndev->srq) {
881 qp_attr.srq = ndev->srq;
882 } else {
883 /* +1 for drain */
884 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
885 qp_attr.cap.max_recv_sge = 2;
886 }
887
888 ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
889 if (ret) {
890 pr_err("failed to create_qp ret= %d\n", ret);
891 goto err_destroy_cq;
892 }
893
894 atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
895
896 pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
897 __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
898 qp_attr.cap.max_send_wr, queue->cm_id);
899
900 if (!ndev->srq) {
901 for (i = 0; i < queue->recv_queue_size; i++) {
902 queue->cmds[i].queue = queue;
903 nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
904 }
905 }
906
907 out:
908 return ret;
909
910 err_destroy_cq:
911 ib_free_cq(queue->cq);
912 goto out;
913 }
914
915 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
916 {
917 ib_drain_qp(queue->cm_id->qp);
918 rdma_destroy_qp(queue->cm_id);
919 ib_free_cq(queue->cq);
920 }
921
922 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
923 {
924 pr_debug("freeing queue %d\n", queue->idx);
925
926 nvmet_sq_destroy(&queue->nvme_sq);
927
928 nvmet_rdma_destroy_queue_ib(queue);
929 if (!queue->dev->srq) {
930 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
931 queue->recv_queue_size,
932 !queue->host_qid);
933 }
934 nvmet_rdma_free_rsps(queue);
935 ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
936 kfree(queue);
937 }
938
939 static void nvmet_rdma_release_queue_work(struct work_struct *w)
940 {
941 struct nvmet_rdma_queue *queue =
942 container_of(w, struct nvmet_rdma_queue, release_work);
943 struct rdma_cm_id *cm_id = queue->cm_id;
944 struct nvmet_rdma_device *dev = queue->dev;
945 enum nvmet_rdma_queue_state state = queue->state;
946
947 nvmet_rdma_free_queue(queue);
948
949 if (state != NVMET_RDMA_IN_DEVICE_REMOVAL)
950 rdma_destroy_id(cm_id);
951
952 kref_put(&dev->ref, nvmet_rdma_free_dev);
953 }
954
955 static int
956 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
957 struct nvmet_rdma_queue *queue)
958 {
959 struct nvme_rdma_cm_req *req;
960
961 req = (struct nvme_rdma_cm_req *)conn->private_data;
962 if (!req || conn->private_data_len == 0)
963 return NVME_RDMA_CM_INVALID_LEN;
964
965 if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
966 return NVME_RDMA_CM_INVALID_RECFMT;
967
968 queue->host_qid = le16_to_cpu(req->qid);
969
970 /*
971 * req->hsqsize corresponds to our recv queue size plus 1
972 * req->hrqsize corresponds to our send queue size
973 */
974 queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
975 queue->send_queue_size = le16_to_cpu(req->hrqsize);
976
977 if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
978 return NVME_RDMA_CM_INVALID_HSQSIZE;
979
980 /* XXX: Should we enforce some kind of max for IO queues? */
981
982 return 0;
983 }
984
985 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
986 enum nvme_rdma_cm_status status)
987 {
988 struct nvme_rdma_cm_rej rej;
989
990 pr_debug("rejecting connect request: status %d (%s)\n",
991 status, nvme_rdma_cm_msg(status));
992
993 rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
994 rej.sts = cpu_to_le16(status);
995
996 return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
997 }
998
999 static struct nvmet_rdma_queue *
1000 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1001 struct rdma_cm_id *cm_id,
1002 struct rdma_cm_event *event)
1003 {
1004 struct nvmet_rdma_queue *queue;
1005 int ret;
1006
1007 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1008 if (!queue) {
1009 ret = NVME_RDMA_CM_NO_RSC;
1010 goto out_reject;
1011 }
1012
1013 ret = nvmet_sq_init(&queue->nvme_sq);
1014 if (ret) {
1015 ret = NVME_RDMA_CM_NO_RSC;
1016 goto out_free_queue;
1017 }
1018
1019 ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1020 if (ret)
1021 goto out_destroy_sq;
1022
1023 /*
1024 * Schedules the actual release because calling rdma_destroy_id from
1025 * inside a CM callback would trigger a deadlock. (great API design..)
1026 */
1027 INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1028 queue->dev = ndev;
1029 queue->cm_id = cm_id;
1030
1031 spin_lock_init(&queue->state_lock);
1032 queue->state = NVMET_RDMA_Q_CONNECTING;
1033 INIT_LIST_HEAD(&queue->rsp_wait_list);
1034 INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1035 spin_lock_init(&queue->rsp_wr_wait_lock);
1036 INIT_LIST_HEAD(&queue->free_rsps);
1037 spin_lock_init(&queue->rsps_lock);
1038 INIT_LIST_HEAD(&queue->queue_list);
1039
1040 queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1041 if (queue->idx < 0) {
1042 ret = NVME_RDMA_CM_NO_RSC;
1043 goto out_destroy_sq;
1044 }
1045
1046 ret = nvmet_rdma_alloc_rsps(queue);
1047 if (ret) {
1048 ret = NVME_RDMA_CM_NO_RSC;
1049 goto out_ida_remove;
1050 }
1051
1052 if (!ndev->srq) {
1053 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1054 queue->recv_queue_size,
1055 !queue->host_qid);
1056 if (IS_ERR(queue->cmds)) {
1057 ret = NVME_RDMA_CM_NO_RSC;
1058 goto out_free_responses;
1059 }
1060 }
1061
1062 ret = nvmet_rdma_create_queue_ib(queue);
1063 if (ret) {
1064 pr_err("%s: creating RDMA queue failed (%d).\n",
1065 __func__, ret);
1066 ret = NVME_RDMA_CM_NO_RSC;
1067 goto out_free_cmds;
1068 }
1069
1070 return queue;
1071
1072 out_free_cmds:
1073 if (!ndev->srq) {
1074 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1075 queue->recv_queue_size,
1076 !queue->host_qid);
1077 }
1078 out_free_responses:
1079 nvmet_rdma_free_rsps(queue);
1080 out_ida_remove:
1081 ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1082 out_destroy_sq:
1083 nvmet_sq_destroy(&queue->nvme_sq);
1084 out_free_queue:
1085 kfree(queue);
1086 out_reject:
1087 nvmet_rdma_cm_reject(cm_id, ret);
1088 return NULL;
1089 }
1090
1091 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1092 {
1093 struct nvmet_rdma_queue *queue = priv;
1094
1095 switch (event->event) {
1096 case IB_EVENT_COMM_EST:
1097 rdma_notify(queue->cm_id, event->event);
1098 break;
1099 default:
1100 pr_err("received IB QP event: %s (%d)\n",
1101 ib_event_msg(event->event), event->event);
1102 break;
1103 }
1104 }
1105
1106 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1107 struct nvmet_rdma_queue *queue,
1108 struct rdma_conn_param *p)
1109 {
1110 struct rdma_conn_param param = { };
1111 struct nvme_rdma_cm_rep priv = { };
1112 int ret = -ENOMEM;
1113
1114 param.rnr_retry_count = 7;
1115 param.flow_control = 1;
1116 param.initiator_depth = min_t(u8, p->initiator_depth,
1117 queue->dev->device->attrs.max_qp_init_rd_atom);
1118 param.private_data = &priv;
1119 param.private_data_len = sizeof(priv);
1120 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1121 priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1122
1123 ret = rdma_accept(cm_id, &param);
1124 if (ret)
1125 pr_err("rdma_accept failed (error code = %d)\n", ret);
1126
1127 return ret;
1128 }
1129
1130 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1131 struct rdma_cm_event *event)
1132 {
1133 struct nvmet_rdma_device *ndev;
1134 struct nvmet_rdma_queue *queue;
1135 int ret = -EINVAL;
1136
1137 ndev = nvmet_rdma_find_get_device(cm_id);
1138 if (!ndev) {
1139 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1140 return -ECONNREFUSED;
1141 }
1142
1143 queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1144 if (!queue) {
1145 ret = -ENOMEM;
1146 goto put_device;
1147 }
1148 queue->port = cm_id->context;
1149
1150 if (queue->host_qid == 0) {
1151 /* Let inflight controller teardown complete */
1152 flush_scheduled_work();
1153 }
1154
1155 ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1156 if (ret)
1157 goto release_queue;
1158
1159 mutex_lock(&nvmet_rdma_queue_mutex);
1160 list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1161 mutex_unlock(&nvmet_rdma_queue_mutex);
1162
1163 return 0;
1164
1165 release_queue:
1166 nvmet_rdma_free_queue(queue);
1167 put_device:
1168 kref_put(&ndev->ref, nvmet_rdma_free_dev);
1169
1170 return ret;
1171 }
1172
1173 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1174 {
1175 unsigned long flags;
1176
1177 spin_lock_irqsave(&queue->state_lock, flags);
1178 if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1179 pr_warn("trying to establish a connected queue\n");
1180 goto out_unlock;
1181 }
1182 queue->state = NVMET_RDMA_Q_LIVE;
1183
1184 while (!list_empty(&queue->rsp_wait_list)) {
1185 struct nvmet_rdma_rsp *cmd;
1186
1187 cmd = list_first_entry(&queue->rsp_wait_list,
1188 struct nvmet_rdma_rsp, wait_list);
1189 list_del(&cmd->wait_list);
1190
1191 spin_unlock_irqrestore(&queue->state_lock, flags);
1192 nvmet_rdma_handle_command(queue, cmd);
1193 spin_lock_irqsave(&queue->state_lock, flags);
1194 }
1195
1196 out_unlock:
1197 spin_unlock_irqrestore(&queue->state_lock, flags);
1198 }
1199
1200 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1201 {
1202 bool disconnect = false;
1203 unsigned long flags;
1204
1205 pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1206
1207 spin_lock_irqsave(&queue->state_lock, flags);
1208 switch (queue->state) {
1209 case NVMET_RDMA_Q_CONNECTING:
1210 case NVMET_RDMA_Q_LIVE:
1211 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1212 case NVMET_RDMA_IN_DEVICE_REMOVAL:
1213 disconnect = true;
1214 break;
1215 case NVMET_RDMA_Q_DISCONNECTING:
1216 break;
1217 }
1218 spin_unlock_irqrestore(&queue->state_lock, flags);
1219
1220 if (disconnect) {
1221 rdma_disconnect(queue->cm_id);
1222 schedule_work(&queue->release_work);
1223 }
1224 }
1225
1226 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1227 {
1228 bool disconnect = false;
1229
1230 mutex_lock(&nvmet_rdma_queue_mutex);
1231 if (!list_empty(&queue->queue_list)) {
1232 list_del_init(&queue->queue_list);
1233 disconnect = true;
1234 }
1235 mutex_unlock(&nvmet_rdma_queue_mutex);
1236
1237 if (disconnect)
1238 __nvmet_rdma_queue_disconnect(queue);
1239 }
1240
1241 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1242 struct nvmet_rdma_queue *queue)
1243 {
1244 WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1245
1246 mutex_lock(&nvmet_rdma_queue_mutex);
1247 if (!list_empty(&queue->queue_list))
1248 list_del_init(&queue->queue_list);
1249 mutex_unlock(&nvmet_rdma_queue_mutex);
1250
1251 pr_err("failed to connect queue %d\n", queue->idx);
1252 schedule_work(&queue->release_work);
1253 }
1254
1255 /**
1256 * nvme_rdma_device_removal() - Handle RDMA device removal
1257 * @cm_id: rdma_cm id, used for nvmet port
1258 * @queue: nvmet rdma queue (cm id qp_context)
1259 *
1260 * DEVICE_REMOVAL event notifies us that the RDMA device is about
1261 * to unplug. Note that this event can be generated on a normal
1262 * queue cm_id and/or a device bound listener cm_id (where in this
1263 * case queue will be null).
1264 *
1265 * We registered an ib_client to handle device removal for queues,
1266 * so we only need to handle the listening port cm_ids. In this case
1267 * we nullify the priv to prevent double cm_id destruction and destroying
1268 * the cm_id implicitely by returning a non-zero rc to the callout.
1269 */
1270 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1271 struct nvmet_rdma_queue *queue)
1272 {
1273 struct nvmet_port *port;
1274
1275 if (queue) {
1276 /*
1277 * This is a queue cm_id. we have registered
1278 * an ib_client to handle queues removal
1279 * so don't interfear and just return.
1280 */
1281 return 0;
1282 }
1283
1284 port = cm_id->context;
1285
1286 /*
1287 * This is a listener cm_id. Make sure that
1288 * future remove_port won't invoke a double
1289 * cm_id destroy. use atomic xchg to make sure
1290 * we don't compete with remove_port.
1291 */
1292 if (xchg(&port->priv, NULL) != cm_id)
1293 return 0;
1294
1295 /*
1296 * We need to return 1 so that the core will destroy
1297 * it's own ID. What a great API design..
1298 */
1299 return 1;
1300 }
1301
1302 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1303 struct rdma_cm_event *event)
1304 {
1305 struct nvmet_rdma_queue *queue = NULL;
1306 int ret = 0;
1307
1308 if (cm_id->qp)
1309 queue = cm_id->qp->qp_context;
1310
1311 pr_debug("%s (%d): status %d id %p\n",
1312 rdma_event_msg(event->event), event->event,
1313 event->status, cm_id);
1314
1315 switch (event->event) {
1316 case RDMA_CM_EVENT_CONNECT_REQUEST:
1317 ret = nvmet_rdma_queue_connect(cm_id, event);
1318 break;
1319 case RDMA_CM_EVENT_ESTABLISHED:
1320 nvmet_rdma_queue_established(queue);
1321 break;
1322 case RDMA_CM_EVENT_ADDR_CHANGE:
1323 case RDMA_CM_EVENT_DISCONNECTED:
1324 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1325 /*
1326 * We might end up here when we already freed the qp
1327 * which means queue release sequence is in progress,
1328 * so don't get in the way...
1329 */
1330 if (queue)
1331 nvmet_rdma_queue_disconnect(queue);
1332 break;
1333 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1334 ret = nvmet_rdma_device_removal(cm_id, queue);
1335 break;
1336 case RDMA_CM_EVENT_REJECTED:
1337 pr_debug("Connection rejected: %s\n",
1338 rdma_reject_msg(cm_id, event->status));
1339 /* FALLTHROUGH */
1340 case RDMA_CM_EVENT_UNREACHABLE:
1341 case RDMA_CM_EVENT_CONNECT_ERROR:
1342 nvmet_rdma_queue_connect_fail(cm_id, queue);
1343 break;
1344 default:
1345 pr_err("received unrecognized RDMA CM event %d\n",
1346 event->event);
1347 break;
1348 }
1349
1350 return ret;
1351 }
1352
1353 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1354 {
1355 struct nvmet_rdma_queue *queue;
1356
1357 restart:
1358 mutex_lock(&nvmet_rdma_queue_mutex);
1359 list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1360 if (queue->nvme_sq.ctrl == ctrl) {
1361 list_del_init(&queue->queue_list);
1362 mutex_unlock(&nvmet_rdma_queue_mutex);
1363
1364 __nvmet_rdma_queue_disconnect(queue);
1365 goto restart;
1366 }
1367 }
1368 mutex_unlock(&nvmet_rdma_queue_mutex);
1369 }
1370
1371 static int nvmet_rdma_add_port(struct nvmet_port *port)
1372 {
1373 struct rdma_cm_id *cm_id;
1374 struct sockaddr_storage addr = { };
1375 __kernel_sa_family_t af;
1376 int ret;
1377
1378 switch (port->disc_addr.adrfam) {
1379 case NVMF_ADDR_FAMILY_IP4:
1380 af = AF_INET;
1381 break;
1382 case NVMF_ADDR_FAMILY_IP6:
1383 af = AF_INET6;
1384 break;
1385 default:
1386 pr_err("address family %d not supported\n",
1387 port->disc_addr.adrfam);
1388 return -EINVAL;
1389 }
1390
1391 ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1392 port->disc_addr.trsvcid, &addr);
1393 if (ret) {
1394 pr_err("malformed ip/port passed: %s:%s\n",
1395 port->disc_addr.traddr, port->disc_addr.trsvcid);
1396 return ret;
1397 }
1398
1399 cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1400 RDMA_PS_TCP, IB_QPT_RC);
1401 if (IS_ERR(cm_id)) {
1402 pr_err("CM ID creation failed\n");
1403 return PTR_ERR(cm_id);
1404 }
1405
1406 /*
1407 * Allow both IPv4 and IPv6 sockets to bind a single port
1408 * at the same time.
1409 */
1410 ret = rdma_set_afonly(cm_id, 1);
1411 if (ret) {
1412 pr_err("rdma_set_afonly failed (%d)\n", ret);
1413 goto out_destroy_id;
1414 }
1415
1416 ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1417 if (ret) {
1418 pr_err("binding CM ID to %pISpcs failed (%d)\n",
1419 (struct sockaddr *)&addr, ret);
1420 goto out_destroy_id;
1421 }
1422
1423 ret = rdma_listen(cm_id, 128);
1424 if (ret) {
1425 pr_err("listening to %pISpcs failed (%d)\n",
1426 (struct sockaddr *)&addr, ret);
1427 goto out_destroy_id;
1428 }
1429
1430 pr_info("enabling port %d (%pISpcs)\n",
1431 le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1432 port->priv = cm_id;
1433 return 0;
1434
1435 out_destroy_id:
1436 rdma_destroy_id(cm_id);
1437 return ret;
1438 }
1439
1440 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1441 {
1442 struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1443
1444 if (cm_id)
1445 rdma_destroy_id(cm_id);
1446 }
1447
1448 static struct nvmet_fabrics_ops nvmet_rdma_ops = {
1449 .owner = THIS_MODULE,
1450 .type = NVMF_TRTYPE_RDMA,
1451 .sqe_inline_size = NVMET_RDMA_INLINE_DATA_SIZE,
1452 .msdbd = 1,
1453 .has_keyed_sgls = 1,
1454 .add_port = nvmet_rdma_add_port,
1455 .remove_port = nvmet_rdma_remove_port,
1456 .queue_response = nvmet_rdma_queue_response,
1457 .delete_ctrl = nvmet_rdma_delete_ctrl,
1458 };
1459
1460 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1461 {
1462 struct nvmet_rdma_queue *queue, *tmp;
1463
1464 /* Device is being removed, delete all queues using this device */
1465 mutex_lock(&nvmet_rdma_queue_mutex);
1466 list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1467 queue_list) {
1468 if (queue->dev->device != ib_device)
1469 continue;
1470
1471 pr_info("Removing queue %d\n", queue->idx);
1472 list_del_init(&queue->queue_list);
1473 __nvmet_rdma_queue_disconnect(queue);
1474 }
1475 mutex_unlock(&nvmet_rdma_queue_mutex);
1476
1477 flush_scheduled_work();
1478 }
1479
1480 static struct ib_client nvmet_rdma_ib_client = {
1481 .name = "nvmet_rdma",
1482 .remove = nvmet_rdma_remove_one
1483 };
1484
1485 static int __init nvmet_rdma_init(void)
1486 {
1487 int ret;
1488
1489 ret = ib_register_client(&nvmet_rdma_ib_client);
1490 if (ret)
1491 return ret;
1492
1493 ret = nvmet_register_transport(&nvmet_rdma_ops);
1494 if (ret)
1495 goto err_ib_client;
1496
1497 return 0;
1498
1499 err_ib_client:
1500 ib_unregister_client(&nvmet_rdma_ib_client);
1501 return ret;
1502 }
1503
1504 static void __exit nvmet_rdma_exit(void)
1505 {
1506 nvmet_unregister_transport(&nvmet_rdma_ops);
1507 ib_unregister_client(&nvmet_rdma_ib_client);
1508 WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
1509 ida_destroy(&nvmet_rdma_queue_ida);
1510 }
1511
1512 module_init(nvmet_rdma_init);
1513 module_exit(nvmet_rdma_exit);
1514
1515 MODULE_LICENSE("GPL v2");
1516 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */
Linux with added FPC-III support