xref: /openbmc/linux/drivers/nvme/target/rdma.c (revision a44e4f3a)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * NVMe over Fabrics RDMA target.
4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
5  */
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/atomic.h>
8 #include <linux/ctype.h>
9 #include <linux/delay.h>
10 #include <linux/err.h>
11 #include <linux/init.h>
12 #include <linux/module.h>
13 #include <linux/nvme.h>
14 #include <linux/slab.h>
15 #include <linux/string.h>
16 #include <linux/wait.h>
17 #include <linux/inet.h>
18 #include <asm/unaligned.h>
19 
20 #include <rdma/ib_verbs.h>
21 #include <rdma/rdma_cm.h>
22 #include <rdma/rw.h>
23 
24 #include <linux/nvme-rdma.h>
25 #include "nvmet.h"
26 
27 /*
28  * We allow at least 1 page, up to 4 SGEs, and up to 16KB of inline data
29  */
30 #define NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE	PAGE_SIZE
31 #define NVMET_RDMA_MAX_INLINE_SGE		4
32 #define NVMET_RDMA_MAX_INLINE_DATA_SIZE		max_t(int, SZ_16K, PAGE_SIZE)
33 
34 struct nvmet_rdma_cmd {
35 	struct ib_sge		sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
36 	struct ib_cqe		cqe;
37 	struct ib_recv_wr	wr;
38 	struct scatterlist	inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
39 	struct nvme_command     *nvme_cmd;
40 	struct nvmet_rdma_queue	*queue;
41 };
42 
43 enum {
44 	NVMET_RDMA_REQ_INLINE_DATA	= (1 << 0),
45 	NVMET_RDMA_REQ_INVALIDATE_RKEY	= (1 << 1),
46 };
47 
48 struct nvmet_rdma_rsp {
49 	struct ib_sge		send_sge;
50 	struct ib_cqe		send_cqe;
51 	struct ib_send_wr	send_wr;
52 
53 	struct nvmet_rdma_cmd	*cmd;
54 	struct nvmet_rdma_queue	*queue;
55 
56 	struct ib_cqe		read_cqe;
57 	struct rdma_rw_ctx	rw;
58 
59 	struct nvmet_req	req;
60 
61 	bool			allocated;
62 	u8			n_rdma;
63 	u32			flags;
64 	u32			invalidate_rkey;
65 
66 	struct list_head	wait_list;
67 	struct list_head	free_list;
68 };
69 
70 enum nvmet_rdma_queue_state {
71 	NVMET_RDMA_Q_CONNECTING,
72 	NVMET_RDMA_Q_LIVE,
73 	NVMET_RDMA_Q_DISCONNECTING,
74 };
75 
76 struct nvmet_rdma_queue {
77 	struct rdma_cm_id	*cm_id;
78 	struct nvmet_port	*port;
79 	struct ib_cq		*cq;
80 	atomic_t		sq_wr_avail;
81 	struct nvmet_rdma_device *dev;
82 	spinlock_t		state_lock;
83 	enum nvmet_rdma_queue_state state;
84 	struct nvmet_cq		nvme_cq;
85 	struct nvmet_sq		nvme_sq;
86 
87 	struct nvmet_rdma_rsp	*rsps;
88 	struct list_head	free_rsps;
89 	spinlock_t		rsps_lock;
90 	struct nvmet_rdma_cmd	*cmds;
91 
92 	struct work_struct	release_work;
93 	struct list_head	rsp_wait_list;
94 	struct list_head	rsp_wr_wait_list;
95 	spinlock_t		rsp_wr_wait_lock;
96 
97 	int			idx;
98 	int			host_qid;
99 	int			recv_queue_size;
100 	int			send_queue_size;
101 
102 	struct list_head	queue_list;
103 };
104 
105 struct nvmet_rdma_device {
106 	struct ib_device	*device;
107 	struct ib_pd		*pd;
108 	struct ib_srq		*srq;
109 	struct nvmet_rdma_cmd	*srq_cmds;
110 	size_t			srq_size;
111 	struct kref		ref;
112 	struct list_head	entry;
113 	int			inline_data_size;
114 	int			inline_page_count;
115 };
116 
117 static bool nvmet_rdma_use_srq;
118 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
119 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
120 
121 static DEFINE_IDA(nvmet_rdma_queue_ida);
122 static LIST_HEAD(nvmet_rdma_queue_list);
123 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
124 
125 static LIST_HEAD(device_list);
126 static DEFINE_MUTEX(device_list_mutex);
127 
128 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
129 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
130 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
131 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
132 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
133 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
134 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
135 				struct nvmet_rdma_rsp *r);
136 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
137 				struct nvmet_rdma_rsp *r);
138 
139 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
140 
141 static int num_pages(int len)
142 {
143 	return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
144 }
145 
146 /* XXX: really should move to a generic header sooner or later.. */
147 static inline u32 get_unaligned_le24(const u8 *p)
148 {
149 	return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
150 }
151 
152 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
153 {
154 	return nvme_is_write(rsp->req.cmd) &&
155 		rsp->req.transfer_len &&
156 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
157 }
158 
159 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
160 {
161 	return !nvme_is_write(rsp->req.cmd) &&
162 		rsp->req.transfer_len &&
163 		!rsp->req.cqe->status &&
164 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
165 }
166 
167 static inline struct nvmet_rdma_rsp *
168 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
169 {
170 	struct nvmet_rdma_rsp *rsp;
171 	unsigned long flags;
172 
173 	spin_lock_irqsave(&queue->rsps_lock, flags);
174 	rsp = list_first_entry_or_null(&queue->free_rsps,
175 				struct nvmet_rdma_rsp, free_list);
176 	if (likely(rsp))
177 		list_del(&rsp->free_list);
178 	spin_unlock_irqrestore(&queue->rsps_lock, flags);
179 
180 	if (unlikely(!rsp)) {
181 		int ret;
182 
183 		rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
184 		if (unlikely(!rsp))
185 			return NULL;
186 		ret = nvmet_rdma_alloc_rsp(queue->dev, rsp);
187 		if (unlikely(ret)) {
188 			kfree(rsp);
189 			return NULL;
190 		}
191 
192 		rsp->allocated = true;
193 	}
194 
195 	return rsp;
196 }
197 
198 static inline void
199 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
200 {
201 	unsigned long flags;
202 
203 	if (unlikely(rsp->allocated)) {
204 		nvmet_rdma_free_rsp(rsp->queue->dev, rsp);
205 		kfree(rsp);
206 		return;
207 	}
208 
209 	spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
210 	list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
211 	spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
212 }
213 
214 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
215 				struct nvmet_rdma_cmd *c)
216 {
217 	struct scatterlist *sg;
218 	struct ib_sge *sge;
219 	int i;
220 
221 	if (!ndev->inline_data_size)
222 		return;
223 
224 	sg = c->inline_sg;
225 	sge = &c->sge[1];
226 
227 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
228 		if (sge->length)
229 			ib_dma_unmap_page(ndev->device, sge->addr,
230 					sge->length, DMA_FROM_DEVICE);
231 		if (sg_page(sg))
232 			__free_page(sg_page(sg));
233 	}
234 }
235 
236 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
237 				struct nvmet_rdma_cmd *c)
238 {
239 	struct scatterlist *sg;
240 	struct ib_sge *sge;
241 	struct page *pg;
242 	int len;
243 	int i;
244 
245 	if (!ndev->inline_data_size)
246 		return 0;
247 
248 	sg = c->inline_sg;
249 	sg_init_table(sg, ndev->inline_page_count);
250 	sge = &c->sge[1];
251 	len = ndev->inline_data_size;
252 
253 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
254 		pg = alloc_page(GFP_KERNEL);
255 		if (!pg)
256 			goto out_err;
257 		sg_assign_page(sg, pg);
258 		sge->addr = ib_dma_map_page(ndev->device,
259 			pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
260 		if (ib_dma_mapping_error(ndev->device, sge->addr))
261 			goto out_err;
262 		sge->length = min_t(int, len, PAGE_SIZE);
263 		sge->lkey = ndev->pd->local_dma_lkey;
264 		len -= sge->length;
265 	}
266 
267 	return 0;
268 out_err:
269 	for (; i >= 0; i--, sg--, sge--) {
270 		if (sge->length)
271 			ib_dma_unmap_page(ndev->device, sge->addr,
272 					sge->length, DMA_FROM_DEVICE);
273 		if (sg_page(sg))
274 			__free_page(sg_page(sg));
275 	}
276 	return -ENOMEM;
277 }
278 
279 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
280 			struct nvmet_rdma_cmd *c, bool admin)
281 {
282 	/* NVMe command / RDMA RECV */
283 	c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
284 	if (!c->nvme_cmd)
285 		goto out;
286 
287 	c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
288 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
289 	if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
290 		goto out_free_cmd;
291 
292 	c->sge[0].length = sizeof(*c->nvme_cmd);
293 	c->sge[0].lkey = ndev->pd->local_dma_lkey;
294 
295 	if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
296 		goto out_unmap_cmd;
297 
298 	c->cqe.done = nvmet_rdma_recv_done;
299 
300 	c->wr.wr_cqe = &c->cqe;
301 	c->wr.sg_list = c->sge;
302 	c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
303 
304 	return 0;
305 
306 out_unmap_cmd:
307 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
308 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
309 out_free_cmd:
310 	kfree(c->nvme_cmd);
311 
312 out:
313 	return -ENOMEM;
314 }
315 
316 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
317 		struct nvmet_rdma_cmd *c, bool admin)
318 {
319 	if (!admin)
320 		nvmet_rdma_free_inline_pages(ndev, c);
321 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
322 				sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
323 	kfree(c->nvme_cmd);
324 }
325 
326 static struct nvmet_rdma_cmd *
327 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
328 		int nr_cmds, bool admin)
329 {
330 	struct nvmet_rdma_cmd *cmds;
331 	int ret = -EINVAL, i;
332 
333 	cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
334 	if (!cmds)
335 		goto out;
336 
337 	for (i = 0; i < nr_cmds; i++) {
338 		ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
339 		if (ret)
340 			goto out_free;
341 	}
342 
343 	return cmds;
344 
345 out_free:
346 	while (--i >= 0)
347 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
348 	kfree(cmds);
349 out:
350 	return ERR_PTR(ret);
351 }
352 
353 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
354 		struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
355 {
356 	int i;
357 
358 	for (i = 0; i < nr_cmds; i++)
359 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
360 	kfree(cmds);
361 }
362 
363 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
364 		struct nvmet_rdma_rsp *r)
365 {
366 	/* NVMe CQE / RDMA SEND */
367 	r->req.cqe = kmalloc(sizeof(*r->req.cqe), GFP_KERNEL);
368 	if (!r->req.cqe)
369 		goto out;
370 
371 	r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.cqe,
372 			sizeof(*r->req.cqe), DMA_TO_DEVICE);
373 	if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
374 		goto out_free_rsp;
375 
376 	r->req.p2p_client = &ndev->device->dev;
377 	r->send_sge.length = sizeof(*r->req.cqe);
378 	r->send_sge.lkey = ndev->pd->local_dma_lkey;
379 
380 	r->send_cqe.done = nvmet_rdma_send_done;
381 
382 	r->send_wr.wr_cqe = &r->send_cqe;
383 	r->send_wr.sg_list = &r->send_sge;
384 	r->send_wr.num_sge = 1;
385 	r->send_wr.send_flags = IB_SEND_SIGNALED;
386 
387 	/* Data In / RDMA READ */
388 	r->read_cqe.done = nvmet_rdma_read_data_done;
389 	return 0;
390 
391 out_free_rsp:
392 	kfree(r->req.cqe);
393 out:
394 	return -ENOMEM;
395 }
396 
397 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
398 		struct nvmet_rdma_rsp *r)
399 {
400 	ib_dma_unmap_single(ndev->device, r->send_sge.addr,
401 				sizeof(*r->req.cqe), DMA_TO_DEVICE);
402 	kfree(r->req.cqe);
403 }
404 
405 static int
406 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
407 {
408 	struct nvmet_rdma_device *ndev = queue->dev;
409 	int nr_rsps = queue->recv_queue_size * 2;
410 	int ret = -EINVAL, i;
411 
412 	queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
413 			GFP_KERNEL);
414 	if (!queue->rsps)
415 		goto out;
416 
417 	for (i = 0; i < nr_rsps; i++) {
418 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
419 
420 		ret = nvmet_rdma_alloc_rsp(ndev, rsp);
421 		if (ret)
422 			goto out_free;
423 
424 		list_add_tail(&rsp->free_list, &queue->free_rsps);
425 	}
426 
427 	return 0;
428 
429 out_free:
430 	while (--i >= 0) {
431 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
432 
433 		list_del(&rsp->free_list);
434 		nvmet_rdma_free_rsp(ndev, rsp);
435 	}
436 	kfree(queue->rsps);
437 out:
438 	return ret;
439 }
440 
441 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
442 {
443 	struct nvmet_rdma_device *ndev = queue->dev;
444 	int i, nr_rsps = queue->recv_queue_size * 2;
445 
446 	for (i = 0; i < nr_rsps; i++) {
447 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
448 
449 		list_del(&rsp->free_list);
450 		nvmet_rdma_free_rsp(ndev, rsp);
451 	}
452 	kfree(queue->rsps);
453 }
454 
455 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
456 		struct nvmet_rdma_cmd *cmd)
457 {
458 	int ret;
459 
460 	ib_dma_sync_single_for_device(ndev->device,
461 		cmd->sge[0].addr, cmd->sge[0].length,
462 		DMA_FROM_DEVICE);
463 
464 	if (ndev->srq)
465 		ret = ib_post_srq_recv(ndev->srq, &cmd->wr, NULL);
466 	else
467 		ret = ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, NULL);
468 
469 	if (unlikely(ret))
470 		pr_err("post_recv cmd failed\n");
471 
472 	return ret;
473 }
474 
475 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
476 {
477 	spin_lock(&queue->rsp_wr_wait_lock);
478 	while (!list_empty(&queue->rsp_wr_wait_list)) {
479 		struct nvmet_rdma_rsp *rsp;
480 		bool ret;
481 
482 		rsp = list_entry(queue->rsp_wr_wait_list.next,
483 				struct nvmet_rdma_rsp, wait_list);
484 		list_del(&rsp->wait_list);
485 
486 		spin_unlock(&queue->rsp_wr_wait_lock);
487 		ret = nvmet_rdma_execute_command(rsp);
488 		spin_lock(&queue->rsp_wr_wait_lock);
489 
490 		if (!ret) {
491 			list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
492 			break;
493 		}
494 	}
495 	spin_unlock(&queue->rsp_wr_wait_lock);
496 }
497 
498 
499 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
500 {
501 	struct nvmet_rdma_queue *queue = rsp->queue;
502 
503 	atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
504 
505 	if (rsp->n_rdma) {
506 		rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
507 				queue->cm_id->port_num, rsp->req.sg,
508 				rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
509 	}
510 
511 	if (rsp->req.sg != rsp->cmd->inline_sg)
512 		nvmet_req_free_sgl(&rsp->req);
513 
514 	if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
515 		nvmet_rdma_process_wr_wait_list(queue);
516 
517 	nvmet_rdma_put_rsp(rsp);
518 }
519 
520 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
521 {
522 	if (queue->nvme_sq.ctrl) {
523 		nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
524 	} else {
525 		/*
526 		 * we didn't setup the controller yet in case
527 		 * of admin connect error, just disconnect and
528 		 * cleanup the queue
529 		 */
530 		nvmet_rdma_queue_disconnect(queue);
531 	}
532 }
533 
534 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
535 {
536 	struct nvmet_rdma_rsp *rsp =
537 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
538 	struct nvmet_rdma_queue *queue = cq->cq_context;
539 
540 	nvmet_rdma_release_rsp(rsp);
541 
542 	if (unlikely(wc->status != IB_WC_SUCCESS &&
543 		     wc->status != IB_WC_WR_FLUSH_ERR)) {
544 		pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
545 			wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
546 		nvmet_rdma_error_comp(queue);
547 	}
548 }
549 
550 static void nvmet_rdma_queue_response(struct nvmet_req *req)
551 {
552 	struct nvmet_rdma_rsp *rsp =
553 		container_of(req, struct nvmet_rdma_rsp, req);
554 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
555 	struct ib_send_wr *first_wr;
556 
557 	if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
558 		rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
559 		rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
560 	} else {
561 		rsp->send_wr.opcode = IB_WR_SEND;
562 	}
563 
564 	if (nvmet_rdma_need_data_out(rsp))
565 		first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
566 				cm_id->port_num, NULL, &rsp->send_wr);
567 	else
568 		first_wr = &rsp->send_wr;
569 
570 	nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
571 
572 	ib_dma_sync_single_for_device(rsp->queue->dev->device,
573 		rsp->send_sge.addr, rsp->send_sge.length,
574 		DMA_TO_DEVICE);
575 
576 	if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
577 		pr_err("sending cmd response failed\n");
578 		nvmet_rdma_release_rsp(rsp);
579 	}
580 }
581 
582 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
583 {
584 	struct nvmet_rdma_rsp *rsp =
585 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
586 	struct nvmet_rdma_queue *queue = cq->cq_context;
587 
588 	WARN_ON(rsp->n_rdma <= 0);
589 	atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
590 	rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
591 			queue->cm_id->port_num, rsp->req.sg,
592 			rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
593 	rsp->n_rdma = 0;
594 
595 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
596 		nvmet_req_uninit(&rsp->req);
597 		nvmet_rdma_release_rsp(rsp);
598 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
599 			pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
600 				wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
601 			nvmet_rdma_error_comp(queue);
602 		}
603 		return;
604 	}
605 
606 	nvmet_req_execute(&rsp->req);
607 }
608 
609 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
610 		u64 off)
611 {
612 	int sg_count = num_pages(len);
613 	struct scatterlist *sg;
614 	int i;
615 
616 	sg = rsp->cmd->inline_sg;
617 	for (i = 0; i < sg_count; i++, sg++) {
618 		if (i < sg_count - 1)
619 			sg_unmark_end(sg);
620 		else
621 			sg_mark_end(sg);
622 		sg->offset = off;
623 		sg->length = min_t(int, len, PAGE_SIZE - off);
624 		len -= sg->length;
625 		if (!i)
626 			off = 0;
627 	}
628 
629 	rsp->req.sg = rsp->cmd->inline_sg;
630 	rsp->req.sg_cnt = sg_count;
631 }
632 
633 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
634 {
635 	struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
636 	u64 off = le64_to_cpu(sgl->addr);
637 	u32 len = le32_to_cpu(sgl->length);
638 
639 	if (!nvme_is_write(rsp->req.cmd)) {
640 		rsp->req.error_loc =
641 			offsetof(struct nvme_common_command, opcode);
642 		return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
643 	}
644 
645 	if (off + len > rsp->queue->dev->inline_data_size) {
646 		pr_err("invalid inline data offset!\n");
647 		return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
648 	}
649 
650 	/* no data command? */
651 	if (!len)
652 		return 0;
653 
654 	nvmet_rdma_use_inline_sg(rsp, len, off);
655 	rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
656 	rsp->req.transfer_len += len;
657 	return 0;
658 }
659 
660 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
661 		struct nvme_keyed_sgl_desc *sgl, bool invalidate)
662 {
663 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
664 	u64 addr = le64_to_cpu(sgl->addr);
665 	u32 key = get_unaligned_le32(sgl->key);
666 	int ret;
667 
668 	rsp->req.transfer_len = get_unaligned_le24(sgl->length);
669 
670 	/* no data command? */
671 	if (!rsp->req.transfer_len)
672 		return 0;
673 
674 	ret = nvmet_req_alloc_sgl(&rsp->req);
675 	if (ret < 0)
676 		goto error_out;
677 
678 	ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
679 			rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
680 			nvmet_data_dir(&rsp->req));
681 	if (ret < 0)
682 		goto error_out;
683 	rsp->n_rdma += ret;
684 
685 	if (invalidate) {
686 		rsp->invalidate_rkey = key;
687 		rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
688 	}
689 
690 	return 0;
691 
692 error_out:
693 	rsp->req.transfer_len = 0;
694 	return NVME_SC_INTERNAL;
695 }
696 
697 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
698 {
699 	struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
700 
701 	switch (sgl->type >> 4) {
702 	case NVME_SGL_FMT_DATA_DESC:
703 		switch (sgl->type & 0xf) {
704 		case NVME_SGL_FMT_OFFSET:
705 			return nvmet_rdma_map_sgl_inline(rsp);
706 		default:
707 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
708 			rsp->req.error_loc =
709 				offsetof(struct nvme_common_command, dptr);
710 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
711 		}
712 	case NVME_KEY_SGL_FMT_DATA_DESC:
713 		switch (sgl->type & 0xf) {
714 		case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
715 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
716 		case NVME_SGL_FMT_ADDRESS:
717 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
718 		default:
719 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
720 			rsp->req.error_loc =
721 				offsetof(struct nvme_common_command, dptr);
722 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
723 		}
724 	default:
725 		pr_err("invalid SGL type: %#x\n", sgl->type);
726 		rsp->req.error_loc = offsetof(struct nvme_common_command, dptr);
727 		return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
728 	}
729 }
730 
731 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
732 {
733 	struct nvmet_rdma_queue *queue = rsp->queue;
734 
735 	if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
736 			&queue->sq_wr_avail) < 0)) {
737 		pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
738 				1 + rsp->n_rdma, queue->idx,
739 				queue->nvme_sq.ctrl->cntlid);
740 		atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
741 		return false;
742 	}
743 
744 	if (nvmet_rdma_need_data_in(rsp)) {
745 		if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
746 				queue->cm_id->port_num, &rsp->read_cqe, NULL))
747 			nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
748 	} else {
749 		nvmet_req_execute(&rsp->req);
750 	}
751 
752 	return true;
753 }
754 
755 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
756 		struct nvmet_rdma_rsp *cmd)
757 {
758 	u16 status;
759 
760 	ib_dma_sync_single_for_cpu(queue->dev->device,
761 		cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
762 		DMA_FROM_DEVICE);
763 	ib_dma_sync_single_for_cpu(queue->dev->device,
764 		cmd->send_sge.addr, cmd->send_sge.length,
765 		DMA_TO_DEVICE);
766 
767 	if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
768 			&queue->nvme_sq, &nvmet_rdma_ops))
769 		return;
770 
771 	status = nvmet_rdma_map_sgl(cmd);
772 	if (status)
773 		goto out_err;
774 
775 	if (unlikely(!nvmet_rdma_execute_command(cmd))) {
776 		spin_lock(&queue->rsp_wr_wait_lock);
777 		list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
778 		spin_unlock(&queue->rsp_wr_wait_lock);
779 	}
780 
781 	return;
782 
783 out_err:
784 	nvmet_req_complete(&cmd->req, status);
785 }
786 
787 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
788 {
789 	struct nvmet_rdma_cmd *cmd =
790 		container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
791 	struct nvmet_rdma_queue *queue = cq->cq_context;
792 	struct nvmet_rdma_rsp *rsp;
793 
794 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
795 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
796 			pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
797 				wc->wr_cqe, ib_wc_status_msg(wc->status),
798 				wc->status);
799 			nvmet_rdma_error_comp(queue);
800 		}
801 		return;
802 	}
803 
804 	if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
805 		pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
806 		nvmet_rdma_error_comp(queue);
807 		return;
808 	}
809 
810 	cmd->queue = queue;
811 	rsp = nvmet_rdma_get_rsp(queue);
812 	if (unlikely(!rsp)) {
813 		/*
814 		 * we get here only under memory pressure,
815 		 * silently drop and have the host retry
816 		 * as we can't even fail it.
817 		 */
818 		nvmet_rdma_post_recv(queue->dev, cmd);
819 		return;
820 	}
821 	rsp->queue = queue;
822 	rsp->cmd = cmd;
823 	rsp->flags = 0;
824 	rsp->req.cmd = cmd->nvme_cmd;
825 	rsp->req.port = queue->port;
826 	rsp->n_rdma = 0;
827 
828 	if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
829 		unsigned long flags;
830 
831 		spin_lock_irqsave(&queue->state_lock, flags);
832 		if (queue->state == NVMET_RDMA_Q_CONNECTING)
833 			list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
834 		else
835 			nvmet_rdma_put_rsp(rsp);
836 		spin_unlock_irqrestore(&queue->state_lock, flags);
837 		return;
838 	}
839 
840 	nvmet_rdma_handle_command(queue, rsp);
841 }
842 
843 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
844 {
845 	if (!ndev->srq)
846 		return;
847 
848 	nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
849 	ib_destroy_srq(ndev->srq);
850 }
851 
852 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
853 {
854 	struct ib_srq_init_attr srq_attr = { NULL, };
855 	struct ib_srq *srq;
856 	size_t srq_size;
857 	int ret, i;
858 
859 	srq_size = 4095;	/* XXX: tune */
860 
861 	srq_attr.attr.max_wr = srq_size;
862 	srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
863 	srq_attr.attr.srq_limit = 0;
864 	srq_attr.srq_type = IB_SRQT_BASIC;
865 	srq = ib_create_srq(ndev->pd, &srq_attr);
866 	if (IS_ERR(srq)) {
867 		/*
868 		 * If SRQs aren't supported we just go ahead and use normal
869 		 * non-shared receive queues.
870 		 */
871 		pr_info("SRQ requested but not supported.\n");
872 		return 0;
873 	}
874 
875 	ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
876 	if (IS_ERR(ndev->srq_cmds)) {
877 		ret = PTR_ERR(ndev->srq_cmds);
878 		goto out_destroy_srq;
879 	}
880 
881 	ndev->srq = srq;
882 	ndev->srq_size = srq_size;
883 
884 	for (i = 0; i < srq_size; i++) {
885 		ret = nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
886 		if (ret)
887 			goto out_free_cmds;
888 	}
889 
890 	return 0;
891 
892 out_free_cmds:
893 	nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
894 out_destroy_srq:
895 	ib_destroy_srq(srq);
896 	return ret;
897 }
898 
899 static void nvmet_rdma_free_dev(struct kref *ref)
900 {
901 	struct nvmet_rdma_device *ndev =
902 		container_of(ref, struct nvmet_rdma_device, ref);
903 
904 	mutex_lock(&device_list_mutex);
905 	list_del(&ndev->entry);
906 	mutex_unlock(&device_list_mutex);
907 
908 	nvmet_rdma_destroy_srq(ndev);
909 	ib_dealloc_pd(ndev->pd);
910 
911 	kfree(ndev);
912 }
913 
914 static struct nvmet_rdma_device *
915 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
916 {
917 	struct nvmet_port *port = cm_id->context;
918 	struct nvmet_rdma_device *ndev;
919 	int inline_page_count;
920 	int inline_sge_count;
921 	int ret;
922 
923 	mutex_lock(&device_list_mutex);
924 	list_for_each_entry(ndev, &device_list, entry) {
925 		if (ndev->device->node_guid == cm_id->device->node_guid &&
926 		    kref_get_unless_zero(&ndev->ref))
927 			goto out_unlock;
928 	}
929 
930 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
931 	if (!ndev)
932 		goto out_err;
933 
934 	inline_page_count = num_pages(port->inline_data_size);
935 	inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
936 				cm_id->device->attrs.max_recv_sge) - 1;
937 	if (inline_page_count > inline_sge_count) {
938 		pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
939 			port->inline_data_size, cm_id->device->name,
940 			inline_sge_count * PAGE_SIZE);
941 		port->inline_data_size = inline_sge_count * PAGE_SIZE;
942 		inline_page_count = inline_sge_count;
943 	}
944 	ndev->inline_data_size = port->inline_data_size;
945 	ndev->inline_page_count = inline_page_count;
946 	ndev->device = cm_id->device;
947 	kref_init(&ndev->ref);
948 
949 	ndev->pd = ib_alloc_pd(ndev->device, 0);
950 	if (IS_ERR(ndev->pd))
951 		goto out_free_dev;
952 
953 	if (nvmet_rdma_use_srq) {
954 		ret = nvmet_rdma_init_srq(ndev);
955 		if (ret)
956 			goto out_free_pd;
957 	}
958 
959 	list_add(&ndev->entry, &device_list);
960 out_unlock:
961 	mutex_unlock(&device_list_mutex);
962 	pr_debug("added %s.\n", ndev->device->name);
963 	return ndev;
964 
965 out_free_pd:
966 	ib_dealloc_pd(ndev->pd);
967 out_free_dev:
968 	kfree(ndev);
969 out_err:
970 	mutex_unlock(&device_list_mutex);
971 	return NULL;
972 }
973 
974 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
975 {
976 	struct ib_qp_init_attr qp_attr;
977 	struct nvmet_rdma_device *ndev = queue->dev;
978 	int comp_vector, nr_cqe, ret, i;
979 
980 	/*
981 	 * Spread the io queues across completion vectors,
982 	 * but still keep all admin queues on vector 0.
983 	 */
984 	comp_vector = !queue->host_qid ? 0 :
985 		queue->idx % ndev->device->num_comp_vectors;
986 
987 	/*
988 	 * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
989 	 */
990 	nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
991 
992 	queue->cq = ib_alloc_cq(ndev->device, queue,
993 			nr_cqe + 1, comp_vector,
994 			IB_POLL_WORKQUEUE);
995 	if (IS_ERR(queue->cq)) {
996 		ret = PTR_ERR(queue->cq);
997 		pr_err("failed to create CQ cqe= %d ret= %d\n",
998 		       nr_cqe + 1, ret);
999 		goto out;
1000 	}
1001 
1002 	memset(&qp_attr, 0, sizeof(qp_attr));
1003 	qp_attr.qp_context = queue;
1004 	qp_attr.event_handler = nvmet_rdma_qp_event;
1005 	qp_attr.send_cq = queue->cq;
1006 	qp_attr.recv_cq = queue->cq;
1007 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1008 	qp_attr.qp_type = IB_QPT_RC;
1009 	/* +1 for drain */
1010 	qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
1011 	qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
1012 	qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
1013 					ndev->device->attrs.max_send_sge);
1014 
1015 	if (ndev->srq) {
1016 		qp_attr.srq = ndev->srq;
1017 	} else {
1018 		/* +1 for drain */
1019 		qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
1020 		qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
1021 	}
1022 
1023 	ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
1024 	if (ret) {
1025 		pr_err("failed to create_qp ret= %d\n", ret);
1026 		goto err_destroy_cq;
1027 	}
1028 
1029 	atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
1030 
1031 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
1032 		 __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
1033 		 qp_attr.cap.max_send_wr, queue->cm_id);
1034 
1035 	if (!ndev->srq) {
1036 		for (i = 0; i < queue->recv_queue_size; i++) {
1037 			queue->cmds[i].queue = queue;
1038 			ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1039 			if (ret)
1040 				goto err_destroy_qp;
1041 		}
1042 	}
1043 
1044 out:
1045 	return ret;
1046 
1047 err_destroy_qp:
1048 	rdma_destroy_qp(queue->cm_id);
1049 err_destroy_cq:
1050 	ib_free_cq(queue->cq);
1051 	goto out;
1052 }
1053 
1054 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1055 {
1056 	struct ib_qp *qp = queue->cm_id->qp;
1057 
1058 	ib_drain_qp(qp);
1059 	rdma_destroy_id(queue->cm_id);
1060 	ib_destroy_qp(qp);
1061 	ib_free_cq(queue->cq);
1062 }
1063 
1064 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1065 {
1066 	pr_debug("freeing queue %d\n", queue->idx);
1067 
1068 	nvmet_sq_destroy(&queue->nvme_sq);
1069 
1070 	nvmet_rdma_destroy_queue_ib(queue);
1071 	if (!queue->dev->srq) {
1072 		nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1073 				queue->recv_queue_size,
1074 				!queue->host_qid);
1075 	}
1076 	nvmet_rdma_free_rsps(queue);
1077 	ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1078 	kfree(queue);
1079 }
1080 
1081 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1082 {
1083 	struct nvmet_rdma_queue *queue =
1084 		container_of(w, struct nvmet_rdma_queue, release_work);
1085 	struct nvmet_rdma_device *dev = queue->dev;
1086 
1087 	nvmet_rdma_free_queue(queue);
1088 
1089 	kref_put(&dev->ref, nvmet_rdma_free_dev);
1090 }
1091 
1092 static int
1093 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1094 				struct nvmet_rdma_queue *queue)
1095 {
1096 	struct nvme_rdma_cm_req *req;
1097 
1098 	req = (struct nvme_rdma_cm_req *)conn->private_data;
1099 	if (!req || conn->private_data_len == 0)
1100 		return NVME_RDMA_CM_INVALID_LEN;
1101 
1102 	if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1103 		return NVME_RDMA_CM_INVALID_RECFMT;
1104 
1105 	queue->host_qid = le16_to_cpu(req->qid);
1106 
1107 	/*
1108 	 * req->hsqsize corresponds to our recv queue size plus 1
1109 	 * req->hrqsize corresponds to our send queue size
1110 	 */
1111 	queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1112 	queue->send_queue_size = le16_to_cpu(req->hrqsize);
1113 
1114 	if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1115 		return NVME_RDMA_CM_INVALID_HSQSIZE;
1116 
1117 	/* XXX: Should we enforce some kind of max for IO queues? */
1118 
1119 	return 0;
1120 }
1121 
1122 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1123 				enum nvme_rdma_cm_status status)
1124 {
1125 	struct nvme_rdma_cm_rej rej;
1126 
1127 	pr_debug("rejecting connect request: status %d (%s)\n",
1128 		 status, nvme_rdma_cm_msg(status));
1129 
1130 	rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1131 	rej.sts = cpu_to_le16(status);
1132 
1133 	return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1134 }
1135 
1136 static struct nvmet_rdma_queue *
1137 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1138 		struct rdma_cm_id *cm_id,
1139 		struct rdma_cm_event *event)
1140 {
1141 	struct nvmet_rdma_queue *queue;
1142 	int ret;
1143 
1144 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1145 	if (!queue) {
1146 		ret = NVME_RDMA_CM_NO_RSC;
1147 		goto out_reject;
1148 	}
1149 
1150 	ret = nvmet_sq_init(&queue->nvme_sq);
1151 	if (ret) {
1152 		ret = NVME_RDMA_CM_NO_RSC;
1153 		goto out_free_queue;
1154 	}
1155 
1156 	ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1157 	if (ret)
1158 		goto out_destroy_sq;
1159 
1160 	/*
1161 	 * Schedules the actual release because calling rdma_destroy_id from
1162 	 * inside a CM callback would trigger a deadlock. (great API design..)
1163 	 */
1164 	INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1165 	queue->dev = ndev;
1166 	queue->cm_id = cm_id;
1167 
1168 	spin_lock_init(&queue->state_lock);
1169 	queue->state = NVMET_RDMA_Q_CONNECTING;
1170 	INIT_LIST_HEAD(&queue->rsp_wait_list);
1171 	INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1172 	spin_lock_init(&queue->rsp_wr_wait_lock);
1173 	INIT_LIST_HEAD(&queue->free_rsps);
1174 	spin_lock_init(&queue->rsps_lock);
1175 	INIT_LIST_HEAD(&queue->queue_list);
1176 
1177 	queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1178 	if (queue->idx < 0) {
1179 		ret = NVME_RDMA_CM_NO_RSC;
1180 		goto out_destroy_sq;
1181 	}
1182 
1183 	ret = nvmet_rdma_alloc_rsps(queue);
1184 	if (ret) {
1185 		ret = NVME_RDMA_CM_NO_RSC;
1186 		goto out_ida_remove;
1187 	}
1188 
1189 	if (!ndev->srq) {
1190 		queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1191 				queue->recv_queue_size,
1192 				!queue->host_qid);
1193 		if (IS_ERR(queue->cmds)) {
1194 			ret = NVME_RDMA_CM_NO_RSC;
1195 			goto out_free_responses;
1196 		}
1197 	}
1198 
1199 	ret = nvmet_rdma_create_queue_ib(queue);
1200 	if (ret) {
1201 		pr_err("%s: creating RDMA queue failed (%d).\n",
1202 			__func__, ret);
1203 		ret = NVME_RDMA_CM_NO_RSC;
1204 		goto out_free_cmds;
1205 	}
1206 
1207 	return queue;
1208 
1209 out_free_cmds:
1210 	if (!ndev->srq) {
1211 		nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1212 				queue->recv_queue_size,
1213 				!queue->host_qid);
1214 	}
1215 out_free_responses:
1216 	nvmet_rdma_free_rsps(queue);
1217 out_ida_remove:
1218 	ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1219 out_destroy_sq:
1220 	nvmet_sq_destroy(&queue->nvme_sq);
1221 out_free_queue:
1222 	kfree(queue);
1223 out_reject:
1224 	nvmet_rdma_cm_reject(cm_id, ret);
1225 	return NULL;
1226 }
1227 
1228 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1229 {
1230 	struct nvmet_rdma_queue *queue = priv;
1231 
1232 	switch (event->event) {
1233 	case IB_EVENT_COMM_EST:
1234 		rdma_notify(queue->cm_id, event->event);
1235 		break;
1236 	default:
1237 		pr_err("received IB QP event: %s (%d)\n",
1238 		       ib_event_msg(event->event), event->event);
1239 		break;
1240 	}
1241 }
1242 
1243 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1244 		struct nvmet_rdma_queue *queue,
1245 		struct rdma_conn_param *p)
1246 {
1247 	struct rdma_conn_param  param = { };
1248 	struct nvme_rdma_cm_rep priv = { };
1249 	int ret = -ENOMEM;
1250 
1251 	param.rnr_retry_count = 7;
1252 	param.flow_control = 1;
1253 	param.initiator_depth = min_t(u8, p->initiator_depth,
1254 		queue->dev->device->attrs.max_qp_init_rd_atom);
1255 	param.private_data = &priv;
1256 	param.private_data_len = sizeof(priv);
1257 	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1258 	priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1259 
1260 	ret = rdma_accept(cm_id, &param);
1261 	if (ret)
1262 		pr_err("rdma_accept failed (error code = %d)\n", ret);
1263 
1264 	return ret;
1265 }
1266 
1267 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1268 		struct rdma_cm_event *event)
1269 {
1270 	struct nvmet_rdma_device *ndev;
1271 	struct nvmet_rdma_queue *queue;
1272 	int ret = -EINVAL;
1273 
1274 	ndev = nvmet_rdma_find_get_device(cm_id);
1275 	if (!ndev) {
1276 		nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1277 		return -ECONNREFUSED;
1278 	}
1279 
1280 	queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1281 	if (!queue) {
1282 		ret = -ENOMEM;
1283 		goto put_device;
1284 	}
1285 	queue->port = cm_id->context;
1286 
1287 	if (queue->host_qid == 0) {
1288 		/* Let inflight controller teardown complete */
1289 		flush_scheduled_work();
1290 	}
1291 
1292 	ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1293 	if (ret) {
1294 		schedule_work(&queue->release_work);
1295 		/* Destroying rdma_cm id is not needed here */
1296 		return 0;
1297 	}
1298 
1299 	mutex_lock(&nvmet_rdma_queue_mutex);
1300 	list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1301 	mutex_unlock(&nvmet_rdma_queue_mutex);
1302 
1303 	return 0;
1304 
1305 put_device:
1306 	kref_put(&ndev->ref, nvmet_rdma_free_dev);
1307 
1308 	return ret;
1309 }
1310 
1311 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1312 {
1313 	unsigned long flags;
1314 
1315 	spin_lock_irqsave(&queue->state_lock, flags);
1316 	if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1317 		pr_warn("trying to establish a connected queue\n");
1318 		goto out_unlock;
1319 	}
1320 	queue->state = NVMET_RDMA_Q_LIVE;
1321 
1322 	while (!list_empty(&queue->rsp_wait_list)) {
1323 		struct nvmet_rdma_rsp *cmd;
1324 
1325 		cmd = list_first_entry(&queue->rsp_wait_list,
1326 					struct nvmet_rdma_rsp, wait_list);
1327 		list_del(&cmd->wait_list);
1328 
1329 		spin_unlock_irqrestore(&queue->state_lock, flags);
1330 		nvmet_rdma_handle_command(queue, cmd);
1331 		spin_lock_irqsave(&queue->state_lock, flags);
1332 	}
1333 
1334 out_unlock:
1335 	spin_unlock_irqrestore(&queue->state_lock, flags);
1336 }
1337 
1338 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1339 {
1340 	bool disconnect = false;
1341 	unsigned long flags;
1342 
1343 	pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1344 
1345 	spin_lock_irqsave(&queue->state_lock, flags);
1346 	switch (queue->state) {
1347 	case NVMET_RDMA_Q_CONNECTING:
1348 	case NVMET_RDMA_Q_LIVE:
1349 		queue->state = NVMET_RDMA_Q_DISCONNECTING;
1350 		disconnect = true;
1351 		break;
1352 	case NVMET_RDMA_Q_DISCONNECTING:
1353 		break;
1354 	}
1355 	spin_unlock_irqrestore(&queue->state_lock, flags);
1356 
1357 	if (disconnect) {
1358 		rdma_disconnect(queue->cm_id);
1359 		schedule_work(&queue->release_work);
1360 	}
1361 }
1362 
1363 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1364 {
1365 	bool disconnect = false;
1366 
1367 	mutex_lock(&nvmet_rdma_queue_mutex);
1368 	if (!list_empty(&queue->queue_list)) {
1369 		list_del_init(&queue->queue_list);
1370 		disconnect = true;
1371 	}
1372 	mutex_unlock(&nvmet_rdma_queue_mutex);
1373 
1374 	if (disconnect)
1375 		__nvmet_rdma_queue_disconnect(queue);
1376 }
1377 
1378 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1379 		struct nvmet_rdma_queue *queue)
1380 {
1381 	WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1382 
1383 	mutex_lock(&nvmet_rdma_queue_mutex);
1384 	if (!list_empty(&queue->queue_list))
1385 		list_del_init(&queue->queue_list);
1386 	mutex_unlock(&nvmet_rdma_queue_mutex);
1387 
1388 	pr_err("failed to connect queue %d\n", queue->idx);
1389 	schedule_work(&queue->release_work);
1390 }
1391 
1392 /**
1393  * nvme_rdma_device_removal() - Handle RDMA device removal
1394  * @cm_id:	rdma_cm id, used for nvmet port
1395  * @queue:      nvmet rdma queue (cm id qp_context)
1396  *
1397  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1398  * to unplug. Note that this event can be generated on a normal
1399  * queue cm_id and/or a device bound listener cm_id (where in this
1400  * case queue will be null).
1401  *
1402  * We registered an ib_client to handle device removal for queues,
1403  * so we only need to handle the listening port cm_ids. In this case
1404  * we nullify the priv to prevent double cm_id destruction and destroying
1405  * the cm_id implicitely by returning a non-zero rc to the callout.
1406  */
1407 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1408 		struct nvmet_rdma_queue *queue)
1409 {
1410 	struct nvmet_port *port;
1411 
1412 	if (queue) {
1413 		/*
1414 		 * This is a queue cm_id. we have registered
1415 		 * an ib_client to handle queues removal
1416 		 * so don't interfear and just return.
1417 		 */
1418 		return 0;
1419 	}
1420 
1421 	port = cm_id->context;
1422 
1423 	/*
1424 	 * This is a listener cm_id. Make sure that
1425 	 * future remove_port won't invoke a double
1426 	 * cm_id destroy. use atomic xchg to make sure
1427 	 * we don't compete with remove_port.
1428 	 */
1429 	if (xchg(&port->priv, NULL) != cm_id)
1430 		return 0;
1431 
1432 	/*
1433 	 * We need to return 1 so that the core will destroy
1434 	 * it's own ID.  What a great API design..
1435 	 */
1436 	return 1;
1437 }
1438 
1439 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1440 		struct rdma_cm_event *event)
1441 {
1442 	struct nvmet_rdma_queue *queue = NULL;
1443 	int ret = 0;
1444 
1445 	if (cm_id->qp)
1446 		queue = cm_id->qp->qp_context;
1447 
1448 	pr_debug("%s (%d): status %d id %p\n",
1449 		rdma_event_msg(event->event), event->event,
1450 		event->status, cm_id);
1451 
1452 	switch (event->event) {
1453 	case RDMA_CM_EVENT_CONNECT_REQUEST:
1454 		ret = nvmet_rdma_queue_connect(cm_id, event);
1455 		break;
1456 	case RDMA_CM_EVENT_ESTABLISHED:
1457 		nvmet_rdma_queue_established(queue);
1458 		break;
1459 	case RDMA_CM_EVENT_ADDR_CHANGE:
1460 	case RDMA_CM_EVENT_DISCONNECTED:
1461 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1462 		nvmet_rdma_queue_disconnect(queue);
1463 		break;
1464 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1465 		ret = nvmet_rdma_device_removal(cm_id, queue);
1466 		break;
1467 	case RDMA_CM_EVENT_REJECTED:
1468 		pr_debug("Connection rejected: %s\n",
1469 			 rdma_reject_msg(cm_id, event->status));
1470 		/* FALLTHROUGH */
1471 	case RDMA_CM_EVENT_UNREACHABLE:
1472 	case RDMA_CM_EVENT_CONNECT_ERROR:
1473 		nvmet_rdma_queue_connect_fail(cm_id, queue);
1474 		break;
1475 	default:
1476 		pr_err("received unrecognized RDMA CM event %d\n",
1477 			event->event);
1478 		break;
1479 	}
1480 
1481 	return ret;
1482 }
1483 
1484 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1485 {
1486 	struct nvmet_rdma_queue *queue;
1487 
1488 restart:
1489 	mutex_lock(&nvmet_rdma_queue_mutex);
1490 	list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1491 		if (queue->nvme_sq.ctrl == ctrl) {
1492 			list_del_init(&queue->queue_list);
1493 			mutex_unlock(&nvmet_rdma_queue_mutex);
1494 
1495 			__nvmet_rdma_queue_disconnect(queue);
1496 			goto restart;
1497 		}
1498 	}
1499 	mutex_unlock(&nvmet_rdma_queue_mutex);
1500 }
1501 
1502 static int nvmet_rdma_add_port(struct nvmet_port *port)
1503 {
1504 	struct rdma_cm_id *cm_id;
1505 	struct sockaddr_storage addr = { };
1506 	__kernel_sa_family_t af;
1507 	int ret;
1508 
1509 	switch (port->disc_addr.adrfam) {
1510 	case NVMF_ADDR_FAMILY_IP4:
1511 		af = AF_INET;
1512 		break;
1513 	case NVMF_ADDR_FAMILY_IP6:
1514 		af = AF_INET6;
1515 		break;
1516 	default:
1517 		pr_err("address family %d not supported\n",
1518 				port->disc_addr.adrfam);
1519 		return -EINVAL;
1520 	}
1521 
1522 	if (port->inline_data_size < 0) {
1523 		port->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1524 	} else if (port->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1525 		pr_warn("inline_data_size %u is too large, reducing to %u\n",
1526 			port->inline_data_size,
1527 			NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1528 		port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1529 	}
1530 
1531 	ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1532 			port->disc_addr.trsvcid, &addr);
1533 	if (ret) {
1534 		pr_err("malformed ip/port passed: %s:%s\n",
1535 			port->disc_addr.traddr, port->disc_addr.trsvcid);
1536 		return ret;
1537 	}
1538 
1539 	cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1540 			RDMA_PS_TCP, IB_QPT_RC);
1541 	if (IS_ERR(cm_id)) {
1542 		pr_err("CM ID creation failed\n");
1543 		return PTR_ERR(cm_id);
1544 	}
1545 
1546 	/*
1547 	 * Allow both IPv4 and IPv6 sockets to bind a single port
1548 	 * at the same time.
1549 	 */
1550 	ret = rdma_set_afonly(cm_id, 1);
1551 	if (ret) {
1552 		pr_err("rdma_set_afonly failed (%d)\n", ret);
1553 		goto out_destroy_id;
1554 	}
1555 
1556 	ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1557 	if (ret) {
1558 		pr_err("binding CM ID to %pISpcs failed (%d)\n",
1559 			(struct sockaddr *)&addr, ret);
1560 		goto out_destroy_id;
1561 	}
1562 
1563 	ret = rdma_listen(cm_id, 128);
1564 	if (ret) {
1565 		pr_err("listening to %pISpcs failed (%d)\n",
1566 			(struct sockaddr *)&addr, ret);
1567 		goto out_destroy_id;
1568 	}
1569 
1570 	pr_info("enabling port %d (%pISpcs)\n",
1571 		le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1572 	port->priv = cm_id;
1573 	return 0;
1574 
1575 out_destroy_id:
1576 	rdma_destroy_id(cm_id);
1577 	return ret;
1578 }
1579 
1580 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1581 {
1582 	struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1583 
1584 	if (cm_id)
1585 		rdma_destroy_id(cm_id);
1586 }
1587 
1588 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1589 		struct nvmet_port *port, char *traddr)
1590 {
1591 	struct rdma_cm_id *cm_id = port->priv;
1592 
1593 	if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1594 		struct nvmet_rdma_rsp *rsp =
1595 			container_of(req, struct nvmet_rdma_rsp, req);
1596 		struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1597 		struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1598 
1599 		sprintf(traddr, "%pISc", addr);
1600 	} else {
1601 		memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
1602 	}
1603 }
1604 
1605 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1606 	.owner			= THIS_MODULE,
1607 	.type			= NVMF_TRTYPE_RDMA,
1608 	.msdbd			= 1,
1609 	.has_keyed_sgls		= 1,
1610 	.add_port		= nvmet_rdma_add_port,
1611 	.remove_port		= nvmet_rdma_remove_port,
1612 	.queue_response		= nvmet_rdma_queue_response,
1613 	.delete_ctrl		= nvmet_rdma_delete_ctrl,
1614 	.disc_traddr		= nvmet_rdma_disc_port_addr,
1615 };
1616 
1617 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1618 {
1619 	struct nvmet_rdma_queue *queue, *tmp;
1620 	struct nvmet_rdma_device *ndev;
1621 	bool found = false;
1622 
1623 	mutex_lock(&device_list_mutex);
1624 	list_for_each_entry(ndev, &device_list, entry) {
1625 		if (ndev->device == ib_device) {
1626 			found = true;
1627 			break;
1628 		}
1629 	}
1630 	mutex_unlock(&device_list_mutex);
1631 
1632 	if (!found)
1633 		return;
1634 
1635 	/*
1636 	 * IB Device that is used by nvmet controllers is being removed,
1637 	 * delete all queues using this device.
1638 	 */
1639 	mutex_lock(&nvmet_rdma_queue_mutex);
1640 	list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1641 				 queue_list) {
1642 		if (queue->dev->device != ib_device)
1643 			continue;
1644 
1645 		pr_info("Removing queue %d\n", queue->idx);
1646 		list_del_init(&queue->queue_list);
1647 		__nvmet_rdma_queue_disconnect(queue);
1648 	}
1649 	mutex_unlock(&nvmet_rdma_queue_mutex);
1650 
1651 	flush_scheduled_work();
1652 }
1653 
1654 static struct ib_client nvmet_rdma_ib_client = {
1655 	.name   = "nvmet_rdma",
1656 	.remove = nvmet_rdma_remove_one
1657 };
1658 
1659 static int __init nvmet_rdma_init(void)
1660 {
1661 	int ret;
1662 
1663 	ret = ib_register_client(&nvmet_rdma_ib_client);
1664 	if (ret)
1665 		return ret;
1666 
1667 	ret = nvmet_register_transport(&nvmet_rdma_ops);
1668 	if (ret)
1669 		goto err_ib_client;
1670 
1671 	return 0;
1672 
1673 err_ib_client:
1674 	ib_unregister_client(&nvmet_rdma_ib_client);
1675 	return ret;
1676 }
1677 
1678 static void __exit nvmet_rdma_exit(void)
1679 {
1680 	nvmet_unregister_transport(&nvmet_rdma_ops);
1681 	ib_unregister_client(&nvmet_rdma_ib_client);
1682 	WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
1683 	ida_destroy(&nvmet_rdma_queue_ida);
1684 }
1685 
1686 module_init(nvmet_rdma_init);
1687 module_exit(nvmet_rdma_exit);
1688 
1689 MODULE_LICENSE("GPL v2");
1690 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */
1691