xref: /linux/drivers/nvme/host/pci.c (revision c411ed854584a71b0e86ac3019b60e4789d88086)
1 /*
2  * NVM Express device driver
3  * Copyright (c) 2011-2014, Intel Corporation.
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 
15 #include <linux/aer.h>
16 #include <linux/bitops.h>
17 #include <linux/blkdev.h>
18 #include <linux/blk-mq.h>
19 #include <linux/blk-mq-pci.h>
20 #include <linux/dmi.h>
21 #include <linux/init.h>
22 #include <linux/interrupt.h>
23 #include <linux/io.h>
24 #include <linux/mm.h>
25 #include <linux/module.h>
26 #include <linux/mutex.h>
27 #include <linux/pci.h>
28 #include <linux/poison.h>
29 #include <linux/t10-pi.h>
30 #include <linux/timer.h>
31 #include <linux/types.h>
32 #include <linux/io-64-nonatomic-lo-hi.h>
33 #include <asm/unaligned.h>
34 #include <linux/sed-opal.h>
35 
36 #include "nvme.h"
37 
38 #define SQ_SIZE(depth)		(depth * sizeof(struct nvme_command))
39 #define CQ_SIZE(depth)		(depth * sizeof(struct nvme_completion))
40 
41 /*
42  * We handle AEN commands ourselves and don't even let the
43  * block layer know about them.
44  */
45 #define NVME_AQ_BLKMQ_DEPTH	(NVME_AQ_DEPTH - NVME_NR_AERS)
46 
47 static int use_threaded_interrupts;
48 module_param(use_threaded_interrupts, int, 0);
49 
50 static bool use_cmb_sqes = true;
51 module_param(use_cmb_sqes, bool, 0644);
52 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
53 
54 static unsigned int max_host_mem_size_mb = 128;
55 module_param(max_host_mem_size_mb, uint, 0444);
56 MODULE_PARM_DESC(max_host_mem_size_mb,
57 	"Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
58 
59 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
60 static const struct kernel_param_ops io_queue_depth_ops = {
61 	.set = io_queue_depth_set,
62 	.get = param_get_int,
63 };
64 
65 static int io_queue_depth = 1024;
66 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
67 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2");
68 
69 struct nvme_dev;
70 struct nvme_queue;
71 
72 static void nvme_process_cq(struct nvme_queue *nvmeq);
73 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
74 
75 /*
76  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
77  */
78 struct nvme_dev {
79 	struct nvme_queue **queues;
80 	struct blk_mq_tag_set tagset;
81 	struct blk_mq_tag_set admin_tagset;
82 	u32 __iomem *dbs;
83 	struct device *dev;
84 	struct dma_pool *prp_page_pool;
85 	struct dma_pool *prp_small_pool;
86 	unsigned online_queues;
87 	unsigned max_qid;
88 	int q_depth;
89 	u32 db_stride;
90 	void __iomem *bar;
91 	unsigned long bar_mapped_size;
92 	struct work_struct remove_work;
93 	struct mutex shutdown_lock;
94 	bool subsystem;
95 	void __iomem *cmb;
96 	dma_addr_t cmb_dma_addr;
97 	u64 cmb_size;
98 	u32 cmbsz;
99 	u32 cmbloc;
100 	struct nvme_ctrl ctrl;
101 	struct completion ioq_wait;
102 
103 	/* shadow doorbell buffer support: */
104 	u32 *dbbuf_dbs;
105 	dma_addr_t dbbuf_dbs_dma_addr;
106 	u32 *dbbuf_eis;
107 	dma_addr_t dbbuf_eis_dma_addr;
108 
109 	/* host memory buffer support: */
110 	u64 host_mem_size;
111 	u32 nr_host_mem_descs;
112 	struct nvme_host_mem_buf_desc *host_mem_descs;
113 	void **host_mem_desc_bufs;
114 };
115 
116 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
117 {
118 	int n = 0, ret;
119 
120 	ret = kstrtoint(val, 10, &n);
121 	if (ret != 0 || n < 2)
122 		return -EINVAL;
123 
124 	return param_set_int(val, kp);
125 }
126 
127 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
128 {
129 	return qid * 2 * stride;
130 }
131 
132 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
133 {
134 	return (qid * 2 + 1) * stride;
135 }
136 
137 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
138 {
139 	return container_of(ctrl, struct nvme_dev, ctrl);
140 }
141 
142 /*
143  * An NVM Express queue.  Each device has at least two (one for admin
144  * commands and one for I/O commands).
145  */
146 struct nvme_queue {
147 	struct device *q_dmadev;
148 	struct nvme_dev *dev;
149 	spinlock_t q_lock;
150 	struct nvme_command *sq_cmds;
151 	struct nvme_command __iomem *sq_cmds_io;
152 	volatile struct nvme_completion *cqes;
153 	struct blk_mq_tags **tags;
154 	dma_addr_t sq_dma_addr;
155 	dma_addr_t cq_dma_addr;
156 	u32 __iomem *q_db;
157 	u16 q_depth;
158 	s16 cq_vector;
159 	u16 sq_tail;
160 	u16 cq_head;
161 	u16 qid;
162 	u8 cq_phase;
163 	u8 cqe_seen;
164 	u32 *dbbuf_sq_db;
165 	u32 *dbbuf_cq_db;
166 	u32 *dbbuf_sq_ei;
167 	u32 *dbbuf_cq_ei;
168 };
169 
170 /*
171  * The nvme_iod describes the data in an I/O, including the list of PRP
172  * entries.  You can't see it in this data structure because C doesn't let
173  * me express that.  Use nvme_init_iod to ensure there's enough space
174  * allocated to store the PRP list.
175  */
176 struct nvme_iod {
177 	struct nvme_request req;
178 	struct nvme_queue *nvmeq;
179 	int aborted;
180 	int npages;		/* In the PRP list. 0 means small pool in use */
181 	int nents;		/* Used in scatterlist */
182 	int length;		/* Of data, in bytes */
183 	dma_addr_t first_dma;
184 	struct scatterlist meta_sg; /* metadata requires single contiguous buffer */
185 	struct scatterlist *sg;
186 	struct scatterlist inline_sg[0];
187 };
188 
189 /*
190  * Check we didin't inadvertently grow the command struct
191  */
192 static inline void _nvme_check_size(void)
193 {
194 	BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
195 	BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
196 	BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
197 	BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
198 	BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
199 	BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
200 	BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
201 	BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
202 	BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != NVME_IDENTIFY_DATA_SIZE);
203 	BUILD_BUG_ON(sizeof(struct nvme_id_ns) != NVME_IDENTIFY_DATA_SIZE);
204 	BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
205 	BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
206 	BUILD_BUG_ON(sizeof(struct nvme_dbbuf) != 64);
207 }
208 
209 static inline unsigned int nvme_dbbuf_size(u32 stride)
210 {
211 	return ((num_possible_cpus() + 1) * 8 * stride);
212 }
213 
214 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
215 {
216 	unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
217 
218 	if (dev->dbbuf_dbs)
219 		return 0;
220 
221 	dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
222 					    &dev->dbbuf_dbs_dma_addr,
223 					    GFP_KERNEL);
224 	if (!dev->dbbuf_dbs)
225 		return -ENOMEM;
226 	dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
227 					    &dev->dbbuf_eis_dma_addr,
228 					    GFP_KERNEL);
229 	if (!dev->dbbuf_eis) {
230 		dma_free_coherent(dev->dev, mem_size,
231 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
232 		dev->dbbuf_dbs = NULL;
233 		return -ENOMEM;
234 	}
235 
236 	return 0;
237 }
238 
239 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
240 {
241 	unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
242 
243 	if (dev->dbbuf_dbs) {
244 		dma_free_coherent(dev->dev, mem_size,
245 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
246 		dev->dbbuf_dbs = NULL;
247 	}
248 	if (dev->dbbuf_eis) {
249 		dma_free_coherent(dev->dev, mem_size,
250 				  dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
251 		dev->dbbuf_eis = NULL;
252 	}
253 }
254 
255 static void nvme_dbbuf_init(struct nvme_dev *dev,
256 			    struct nvme_queue *nvmeq, int qid)
257 {
258 	if (!dev->dbbuf_dbs || !qid)
259 		return;
260 
261 	nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
262 	nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
263 	nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
264 	nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
265 }
266 
267 static void nvme_dbbuf_set(struct nvme_dev *dev)
268 {
269 	struct nvme_command c;
270 
271 	if (!dev->dbbuf_dbs)
272 		return;
273 
274 	memset(&c, 0, sizeof(c));
275 	c.dbbuf.opcode = nvme_admin_dbbuf;
276 	c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
277 	c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
278 
279 	if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
280 		dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
281 		/* Free memory and continue on */
282 		nvme_dbbuf_dma_free(dev);
283 	}
284 }
285 
286 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
287 {
288 	return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
289 }
290 
291 /* Update dbbuf and return true if an MMIO is required */
292 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
293 					      volatile u32 *dbbuf_ei)
294 {
295 	if (dbbuf_db) {
296 		u16 old_value;
297 
298 		/*
299 		 * Ensure that the queue is written before updating
300 		 * the doorbell in memory
301 		 */
302 		wmb();
303 
304 		old_value = *dbbuf_db;
305 		*dbbuf_db = value;
306 
307 		if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
308 			return false;
309 	}
310 
311 	return true;
312 }
313 
314 /*
315  * Max size of iod being embedded in the request payload
316  */
317 #define NVME_INT_PAGES		2
318 #define NVME_INT_BYTES(dev)	(NVME_INT_PAGES * (dev)->ctrl.page_size)
319 
320 /*
321  * Will slightly overestimate the number of pages needed.  This is OK
322  * as it only leads to a small amount of wasted memory for the lifetime of
323  * the I/O.
324  */
325 static int nvme_npages(unsigned size, struct nvme_dev *dev)
326 {
327 	unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
328 				      dev->ctrl.page_size);
329 	return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
330 }
331 
332 static unsigned int nvme_iod_alloc_size(struct nvme_dev *dev,
333 		unsigned int size, unsigned int nseg)
334 {
335 	return sizeof(__le64 *) * nvme_npages(size, dev) +
336 			sizeof(struct scatterlist) * nseg;
337 }
338 
339 static unsigned int nvme_cmd_size(struct nvme_dev *dev)
340 {
341 	return sizeof(struct nvme_iod) +
342 		nvme_iod_alloc_size(dev, NVME_INT_BYTES(dev), NVME_INT_PAGES);
343 }
344 
345 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
346 				unsigned int hctx_idx)
347 {
348 	struct nvme_dev *dev = data;
349 	struct nvme_queue *nvmeq = dev->queues[0];
350 
351 	WARN_ON(hctx_idx != 0);
352 	WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
353 	WARN_ON(nvmeq->tags);
354 
355 	hctx->driver_data = nvmeq;
356 	nvmeq->tags = &dev->admin_tagset.tags[0];
357 	return 0;
358 }
359 
360 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
361 {
362 	struct nvme_queue *nvmeq = hctx->driver_data;
363 
364 	nvmeq->tags = NULL;
365 }
366 
367 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
368 			  unsigned int hctx_idx)
369 {
370 	struct nvme_dev *dev = data;
371 	struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
372 
373 	if (!nvmeq->tags)
374 		nvmeq->tags = &dev->tagset.tags[hctx_idx];
375 
376 	WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
377 	hctx->driver_data = nvmeq;
378 	return 0;
379 }
380 
381 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
382 		unsigned int hctx_idx, unsigned int numa_node)
383 {
384 	struct nvme_dev *dev = set->driver_data;
385 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
386 	int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
387 	struct nvme_queue *nvmeq = dev->queues[queue_idx];
388 
389 	BUG_ON(!nvmeq);
390 	iod->nvmeq = nvmeq;
391 	return 0;
392 }
393 
394 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
395 {
396 	struct nvme_dev *dev = set->driver_data;
397 
398 	return blk_mq_pci_map_queues(set, to_pci_dev(dev->dev));
399 }
400 
401 /**
402  * __nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
403  * @nvmeq: The queue to use
404  * @cmd: The command to send
405  *
406  * Safe to use from interrupt context
407  */
408 static void __nvme_submit_cmd(struct nvme_queue *nvmeq,
409 						struct nvme_command *cmd)
410 {
411 	u16 tail = nvmeq->sq_tail;
412 
413 	if (nvmeq->sq_cmds_io)
414 		memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd));
415 	else
416 		memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
417 
418 	if (++tail == nvmeq->q_depth)
419 		tail = 0;
420 	if (nvme_dbbuf_update_and_check_event(tail, nvmeq->dbbuf_sq_db,
421 					      nvmeq->dbbuf_sq_ei))
422 		writel(tail, nvmeq->q_db);
423 	nvmeq->sq_tail = tail;
424 }
425 
426 static __le64 **iod_list(struct request *req)
427 {
428 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
429 	return (__le64 **)(iod->sg + blk_rq_nr_phys_segments(req));
430 }
431 
432 static blk_status_t nvme_init_iod(struct request *rq, struct nvme_dev *dev)
433 {
434 	struct nvme_iod *iod = blk_mq_rq_to_pdu(rq);
435 	int nseg = blk_rq_nr_phys_segments(rq);
436 	unsigned int size = blk_rq_payload_bytes(rq);
437 
438 	if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) {
439 		iod->sg = kmalloc(nvme_iod_alloc_size(dev, size, nseg), GFP_ATOMIC);
440 		if (!iod->sg)
441 			return BLK_STS_RESOURCE;
442 	} else {
443 		iod->sg = iod->inline_sg;
444 	}
445 
446 	iod->aborted = 0;
447 	iod->npages = -1;
448 	iod->nents = 0;
449 	iod->length = size;
450 
451 	return BLK_STS_OK;
452 }
453 
454 static void nvme_free_iod(struct nvme_dev *dev, struct request *req)
455 {
456 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
457 	const int last_prp = dev->ctrl.page_size / 8 - 1;
458 	int i;
459 	__le64 **list = iod_list(req);
460 	dma_addr_t prp_dma = iod->first_dma;
461 
462 	if (iod->npages == 0)
463 		dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
464 	for (i = 0; i < iod->npages; i++) {
465 		__le64 *prp_list = list[i];
466 		dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
467 		dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
468 		prp_dma = next_prp_dma;
469 	}
470 
471 	if (iod->sg != iod->inline_sg)
472 		kfree(iod->sg);
473 }
474 
475 #ifdef CONFIG_BLK_DEV_INTEGRITY
476 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
477 {
478 	if (be32_to_cpu(pi->ref_tag) == v)
479 		pi->ref_tag = cpu_to_be32(p);
480 }
481 
482 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
483 {
484 	if (be32_to_cpu(pi->ref_tag) == p)
485 		pi->ref_tag = cpu_to_be32(v);
486 }
487 
488 /**
489  * nvme_dif_remap - remaps ref tags to bip seed and physical lba
490  *
491  * The virtual start sector is the one that was originally submitted by the
492  * block layer.	Due to partitioning, MD/DM cloning, etc. the actual physical
493  * start sector may be different. Remap protection information to match the
494  * physical LBA on writes, and back to the original seed on reads.
495  *
496  * Type 0 and 3 do not have a ref tag, so no remapping required.
497  */
498 static void nvme_dif_remap(struct request *req,
499 			void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
500 {
501 	struct nvme_ns *ns = req->rq_disk->private_data;
502 	struct bio_integrity_payload *bip;
503 	struct t10_pi_tuple *pi;
504 	void *p, *pmap;
505 	u32 i, nlb, ts, phys, virt;
506 
507 	if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
508 		return;
509 
510 	bip = bio_integrity(req->bio);
511 	if (!bip)
512 		return;
513 
514 	pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
515 
516 	p = pmap;
517 	virt = bip_get_seed(bip);
518 	phys = nvme_block_nr(ns, blk_rq_pos(req));
519 	nlb = (blk_rq_bytes(req) >> ns->lba_shift);
520 	ts = ns->disk->queue->integrity.tuple_size;
521 
522 	for (i = 0; i < nlb; i++, virt++, phys++) {
523 		pi = (struct t10_pi_tuple *)p;
524 		dif_swap(phys, virt, pi);
525 		p += ts;
526 	}
527 	kunmap_atomic(pmap);
528 }
529 #else /* CONFIG_BLK_DEV_INTEGRITY */
530 static void nvme_dif_remap(struct request *req,
531 			void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
532 {
533 }
534 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
535 {
536 }
537 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
538 {
539 }
540 #endif
541 
542 static blk_status_t nvme_setup_prps(struct nvme_dev *dev, struct request *req)
543 {
544 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
545 	struct dma_pool *pool;
546 	int length = blk_rq_payload_bytes(req);
547 	struct scatterlist *sg = iod->sg;
548 	int dma_len = sg_dma_len(sg);
549 	u64 dma_addr = sg_dma_address(sg);
550 	u32 page_size = dev->ctrl.page_size;
551 	int offset = dma_addr & (page_size - 1);
552 	__le64 *prp_list;
553 	__le64 **list = iod_list(req);
554 	dma_addr_t prp_dma;
555 	int nprps, i;
556 
557 	length -= (page_size - offset);
558 	if (length <= 0)
559 		return BLK_STS_OK;
560 
561 	dma_len -= (page_size - offset);
562 	if (dma_len) {
563 		dma_addr += (page_size - offset);
564 	} else {
565 		sg = sg_next(sg);
566 		dma_addr = sg_dma_address(sg);
567 		dma_len = sg_dma_len(sg);
568 	}
569 
570 	if (length <= page_size) {
571 		iod->first_dma = dma_addr;
572 		return BLK_STS_OK;
573 	}
574 
575 	nprps = DIV_ROUND_UP(length, page_size);
576 	if (nprps <= (256 / 8)) {
577 		pool = dev->prp_small_pool;
578 		iod->npages = 0;
579 	} else {
580 		pool = dev->prp_page_pool;
581 		iod->npages = 1;
582 	}
583 
584 	prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
585 	if (!prp_list) {
586 		iod->first_dma = dma_addr;
587 		iod->npages = -1;
588 		return BLK_STS_RESOURCE;
589 	}
590 	list[0] = prp_list;
591 	iod->first_dma = prp_dma;
592 	i = 0;
593 	for (;;) {
594 		if (i == page_size >> 3) {
595 			__le64 *old_prp_list = prp_list;
596 			prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
597 			if (!prp_list)
598 				return BLK_STS_RESOURCE;
599 			list[iod->npages++] = prp_list;
600 			prp_list[0] = old_prp_list[i - 1];
601 			old_prp_list[i - 1] = cpu_to_le64(prp_dma);
602 			i = 1;
603 		}
604 		prp_list[i++] = cpu_to_le64(dma_addr);
605 		dma_len -= page_size;
606 		dma_addr += page_size;
607 		length -= page_size;
608 		if (length <= 0)
609 			break;
610 		if (dma_len > 0)
611 			continue;
612 		if (unlikely(dma_len < 0))
613 			goto bad_sgl;
614 		sg = sg_next(sg);
615 		dma_addr = sg_dma_address(sg);
616 		dma_len = sg_dma_len(sg);
617 	}
618 
619 	return BLK_STS_OK;
620 
621  bad_sgl:
622 	if (WARN_ONCE(1, "Invalid SGL for payload:%d nents:%d\n",
623 				blk_rq_payload_bytes(req), iod->nents)) {
624 		for_each_sg(iod->sg, sg, iod->nents, i) {
625 			dma_addr_t phys = sg_phys(sg);
626 			pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
627 			       "dma_address:%pad dma_length:%d\n", i, &phys,
628 					sg->offset, sg->length,
629 					&sg_dma_address(sg),
630 					sg_dma_len(sg));
631 		}
632 	}
633 	return BLK_STS_IOERR;
634 
635 }
636 
637 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
638 		struct nvme_command *cmnd)
639 {
640 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
641 	struct request_queue *q = req->q;
642 	enum dma_data_direction dma_dir = rq_data_dir(req) ?
643 			DMA_TO_DEVICE : DMA_FROM_DEVICE;
644 	blk_status_t ret = BLK_STS_IOERR;
645 
646 	sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
647 	iod->nents = blk_rq_map_sg(q, req, iod->sg);
648 	if (!iod->nents)
649 		goto out;
650 
651 	ret = BLK_STS_RESOURCE;
652 	if (!dma_map_sg_attrs(dev->dev, iod->sg, iod->nents, dma_dir,
653 				DMA_ATTR_NO_WARN))
654 		goto out;
655 
656 	ret = nvme_setup_prps(dev, req);
657 	if (ret != BLK_STS_OK)
658 		goto out_unmap;
659 
660 	ret = BLK_STS_IOERR;
661 	if (blk_integrity_rq(req)) {
662 		if (blk_rq_count_integrity_sg(q, req->bio) != 1)
663 			goto out_unmap;
664 
665 		sg_init_table(&iod->meta_sg, 1);
666 		if (blk_rq_map_integrity_sg(q, req->bio, &iod->meta_sg) != 1)
667 			goto out_unmap;
668 
669 		if (rq_data_dir(req))
670 			nvme_dif_remap(req, nvme_dif_prep);
671 
672 		if (!dma_map_sg(dev->dev, &iod->meta_sg, 1, dma_dir))
673 			goto out_unmap;
674 	}
675 
676 	cmnd->rw.dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
677 	cmnd->rw.dptr.prp2 = cpu_to_le64(iod->first_dma);
678 	if (blk_integrity_rq(req))
679 		cmnd->rw.metadata = cpu_to_le64(sg_dma_address(&iod->meta_sg));
680 	return BLK_STS_OK;
681 
682 out_unmap:
683 	dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
684 out:
685 	return ret;
686 }
687 
688 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
689 {
690 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
691 	enum dma_data_direction dma_dir = rq_data_dir(req) ?
692 			DMA_TO_DEVICE : DMA_FROM_DEVICE;
693 
694 	if (iod->nents) {
695 		dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
696 		if (blk_integrity_rq(req)) {
697 			if (!rq_data_dir(req))
698 				nvme_dif_remap(req, nvme_dif_complete);
699 			dma_unmap_sg(dev->dev, &iod->meta_sg, 1, dma_dir);
700 		}
701 	}
702 
703 	nvme_cleanup_cmd(req);
704 	nvme_free_iod(dev, req);
705 }
706 
707 /*
708  * NOTE: ns is NULL when called on the admin queue.
709  */
710 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
711 			 const struct blk_mq_queue_data *bd)
712 {
713 	struct nvme_ns *ns = hctx->queue->queuedata;
714 	struct nvme_queue *nvmeq = hctx->driver_data;
715 	struct nvme_dev *dev = nvmeq->dev;
716 	struct request *req = bd->rq;
717 	struct nvme_command cmnd;
718 	blk_status_t ret;
719 
720 	ret = nvme_setup_cmd(ns, req, &cmnd);
721 	if (ret)
722 		return ret;
723 
724 	ret = nvme_init_iod(req, dev);
725 	if (ret)
726 		goto out_free_cmd;
727 
728 	if (blk_rq_nr_phys_segments(req)) {
729 		ret = nvme_map_data(dev, req, &cmnd);
730 		if (ret)
731 			goto out_cleanup_iod;
732 	}
733 
734 	blk_mq_start_request(req);
735 
736 	spin_lock_irq(&nvmeq->q_lock);
737 	if (unlikely(nvmeq->cq_vector < 0)) {
738 		ret = BLK_STS_IOERR;
739 		spin_unlock_irq(&nvmeq->q_lock);
740 		goto out_cleanup_iod;
741 	}
742 	__nvme_submit_cmd(nvmeq, &cmnd);
743 	nvme_process_cq(nvmeq);
744 	spin_unlock_irq(&nvmeq->q_lock);
745 	return BLK_STS_OK;
746 out_cleanup_iod:
747 	nvme_free_iod(dev, req);
748 out_free_cmd:
749 	nvme_cleanup_cmd(req);
750 	return ret;
751 }
752 
753 static void nvme_pci_complete_rq(struct request *req)
754 {
755 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
756 
757 	nvme_unmap_data(iod->nvmeq->dev, req);
758 	nvme_complete_rq(req);
759 }
760 
761 /* We read the CQE phase first to check if the rest of the entry is valid */
762 static inline bool nvme_cqe_valid(struct nvme_queue *nvmeq, u16 head,
763 		u16 phase)
764 {
765 	return (le16_to_cpu(nvmeq->cqes[head].status) & 1) == phase;
766 }
767 
768 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
769 {
770 	u16 head = nvmeq->cq_head;
771 
772 	if (likely(nvmeq->cq_vector >= 0)) {
773 		if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
774 						      nvmeq->dbbuf_cq_ei))
775 			writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
776 	}
777 }
778 
779 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq,
780 		struct nvme_completion *cqe)
781 {
782 	struct request *req;
783 
784 	if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
785 		dev_warn(nvmeq->dev->ctrl.device,
786 			"invalid id %d completed on queue %d\n",
787 			cqe->command_id, le16_to_cpu(cqe->sq_id));
788 		return;
789 	}
790 
791 	/*
792 	 * AEN requests are special as they don't time out and can
793 	 * survive any kind of queue freeze and often don't respond to
794 	 * aborts.  We don't even bother to allocate a struct request
795 	 * for them but rather special case them here.
796 	 */
797 	if (unlikely(nvmeq->qid == 0 &&
798 			cqe->command_id >= NVME_AQ_BLKMQ_DEPTH)) {
799 		nvme_complete_async_event(&nvmeq->dev->ctrl,
800 				cqe->status, &cqe->result);
801 		return;
802 	}
803 
804 	nvmeq->cqe_seen = 1;
805 	req = blk_mq_tag_to_rq(*nvmeq->tags, cqe->command_id);
806 	nvme_end_request(req, cqe->status, cqe->result);
807 }
808 
809 static inline bool nvme_read_cqe(struct nvme_queue *nvmeq,
810 		struct nvme_completion *cqe)
811 {
812 	if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase)) {
813 		*cqe = nvmeq->cqes[nvmeq->cq_head];
814 
815 		if (++nvmeq->cq_head == nvmeq->q_depth) {
816 			nvmeq->cq_head = 0;
817 			nvmeq->cq_phase = !nvmeq->cq_phase;
818 		}
819 		return true;
820 	}
821 	return false;
822 }
823 
824 static void nvme_process_cq(struct nvme_queue *nvmeq)
825 {
826 	struct nvme_completion cqe;
827 	int consumed = 0;
828 
829 	while (nvme_read_cqe(nvmeq, &cqe)) {
830 		nvme_handle_cqe(nvmeq, &cqe);
831 		consumed++;
832 	}
833 
834 	if (consumed)
835 		nvme_ring_cq_doorbell(nvmeq);
836 }
837 
838 static irqreturn_t nvme_irq(int irq, void *data)
839 {
840 	irqreturn_t result;
841 	struct nvme_queue *nvmeq = data;
842 	spin_lock(&nvmeq->q_lock);
843 	nvme_process_cq(nvmeq);
844 	result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
845 	nvmeq->cqe_seen = 0;
846 	spin_unlock(&nvmeq->q_lock);
847 	return result;
848 }
849 
850 static irqreturn_t nvme_irq_check(int irq, void *data)
851 {
852 	struct nvme_queue *nvmeq = data;
853 	if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
854 		return IRQ_WAKE_THREAD;
855 	return IRQ_NONE;
856 }
857 
858 static int __nvme_poll(struct nvme_queue *nvmeq, unsigned int tag)
859 {
860 	struct nvme_completion cqe;
861 	int found = 0, consumed = 0;
862 
863 	if (!nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
864 		return 0;
865 
866 	spin_lock_irq(&nvmeq->q_lock);
867 	while (nvme_read_cqe(nvmeq, &cqe)) {
868 		nvme_handle_cqe(nvmeq, &cqe);
869 		consumed++;
870 
871 		if (tag == cqe.command_id) {
872 			found = 1;
873 			break;
874 		}
875        }
876 
877 	if (consumed)
878 		nvme_ring_cq_doorbell(nvmeq);
879 	spin_unlock_irq(&nvmeq->q_lock);
880 
881 	return found;
882 }
883 
884 static int nvme_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
885 {
886 	struct nvme_queue *nvmeq = hctx->driver_data;
887 
888 	return __nvme_poll(nvmeq, tag);
889 }
890 
891 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl, int aer_idx)
892 {
893 	struct nvme_dev *dev = to_nvme_dev(ctrl);
894 	struct nvme_queue *nvmeq = dev->queues[0];
895 	struct nvme_command c;
896 
897 	memset(&c, 0, sizeof(c));
898 	c.common.opcode = nvme_admin_async_event;
899 	c.common.command_id = NVME_AQ_BLKMQ_DEPTH + aer_idx;
900 
901 	spin_lock_irq(&nvmeq->q_lock);
902 	__nvme_submit_cmd(nvmeq, &c);
903 	spin_unlock_irq(&nvmeq->q_lock);
904 }
905 
906 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
907 {
908 	struct nvme_command c;
909 
910 	memset(&c, 0, sizeof(c));
911 	c.delete_queue.opcode = opcode;
912 	c.delete_queue.qid = cpu_to_le16(id);
913 
914 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
915 }
916 
917 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
918 						struct nvme_queue *nvmeq)
919 {
920 	struct nvme_command c;
921 	int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
922 
923 	/*
924 	 * Note: we (ab)use the fact the the prp fields survive if no data
925 	 * is attached to the request.
926 	 */
927 	memset(&c, 0, sizeof(c));
928 	c.create_cq.opcode = nvme_admin_create_cq;
929 	c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
930 	c.create_cq.cqid = cpu_to_le16(qid);
931 	c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
932 	c.create_cq.cq_flags = cpu_to_le16(flags);
933 	c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
934 
935 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
936 }
937 
938 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
939 						struct nvme_queue *nvmeq)
940 {
941 	struct nvme_command c;
942 	int flags = NVME_QUEUE_PHYS_CONTIG;
943 
944 	/*
945 	 * Note: we (ab)use the fact the the prp fields survive if no data
946 	 * is attached to the request.
947 	 */
948 	memset(&c, 0, sizeof(c));
949 	c.create_sq.opcode = nvme_admin_create_sq;
950 	c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
951 	c.create_sq.sqid = cpu_to_le16(qid);
952 	c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
953 	c.create_sq.sq_flags = cpu_to_le16(flags);
954 	c.create_sq.cqid = cpu_to_le16(qid);
955 
956 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
957 }
958 
959 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
960 {
961 	return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
962 }
963 
964 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
965 {
966 	return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
967 }
968 
969 static void abort_endio(struct request *req, blk_status_t error)
970 {
971 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
972 	struct nvme_queue *nvmeq = iod->nvmeq;
973 
974 	dev_warn(nvmeq->dev->ctrl.device,
975 		 "Abort status: 0x%x", nvme_req(req)->status);
976 	atomic_inc(&nvmeq->dev->ctrl.abort_limit);
977 	blk_mq_free_request(req);
978 }
979 
980 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
981 {
982 
983 	/* If true, indicates loss of adapter communication, possibly by a
984 	 * NVMe Subsystem reset.
985 	 */
986 	bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
987 
988 	/* If there is a reset ongoing, we shouldn't reset again. */
989 	if (dev->ctrl.state == NVME_CTRL_RESETTING)
990 		return false;
991 
992 	/* We shouldn't reset unless the controller is on fatal error state
993 	 * _or_ if we lost the communication with it.
994 	 */
995 	if (!(csts & NVME_CSTS_CFS) && !nssro)
996 		return false;
997 
998 	/* If PCI error recovery process is happening, we cannot reset or
999 	 * the recovery mechanism will surely fail.
1000 	 */
1001 	if (pci_channel_offline(to_pci_dev(dev->dev)))
1002 		return false;
1003 
1004 	return true;
1005 }
1006 
1007 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1008 {
1009 	/* Read a config register to help see what died. */
1010 	u16 pci_status;
1011 	int result;
1012 
1013 	result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1014 				      &pci_status);
1015 	if (result == PCIBIOS_SUCCESSFUL)
1016 		dev_warn(dev->ctrl.device,
1017 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1018 			 csts, pci_status);
1019 	else
1020 		dev_warn(dev->ctrl.device,
1021 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1022 			 csts, result);
1023 }
1024 
1025 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1026 {
1027 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1028 	struct nvme_queue *nvmeq = iod->nvmeq;
1029 	struct nvme_dev *dev = nvmeq->dev;
1030 	struct request *abort_req;
1031 	struct nvme_command cmd;
1032 	u32 csts = readl(dev->bar + NVME_REG_CSTS);
1033 
1034 	/*
1035 	 * Reset immediately if the controller is failed
1036 	 */
1037 	if (nvme_should_reset(dev, csts)) {
1038 		nvme_warn_reset(dev, csts);
1039 		nvme_dev_disable(dev, false);
1040 		nvme_reset_ctrl(&dev->ctrl);
1041 		return BLK_EH_HANDLED;
1042 	}
1043 
1044 	/*
1045 	 * Did we miss an interrupt?
1046 	 */
1047 	if (__nvme_poll(nvmeq, req->tag)) {
1048 		dev_warn(dev->ctrl.device,
1049 			 "I/O %d QID %d timeout, completion polled\n",
1050 			 req->tag, nvmeq->qid);
1051 		return BLK_EH_HANDLED;
1052 	}
1053 
1054 	/*
1055 	 * Shutdown immediately if controller times out while starting. The
1056 	 * reset work will see the pci device disabled when it gets the forced
1057 	 * cancellation error. All outstanding requests are completed on
1058 	 * shutdown, so we return BLK_EH_HANDLED.
1059 	 */
1060 	if (dev->ctrl.state == NVME_CTRL_RESETTING) {
1061 		dev_warn(dev->ctrl.device,
1062 			 "I/O %d QID %d timeout, disable controller\n",
1063 			 req->tag, nvmeq->qid);
1064 		nvme_dev_disable(dev, false);
1065 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1066 		return BLK_EH_HANDLED;
1067 	}
1068 
1069 	/*
1070  	 * Shutdown the controller immediately and schedule a reset if the
1071  	 * command was already aborted once before and still hasn't been
1072  	 * returned to the driver, or if this is the admin queue.
1073 	 */
1074 	if (!nvmeq->qid || iod->aborted) {
1075 		dev_warn(dev->ctrl.device,
1076 			 "I/O %d QID %d timeout, reset controller\n",
1077 			 req->tag, nvmeq->qid);
1078 		nvme_dev_disable(dev, false);
1079 		nvme_reset_ctrl(&dev->ctrl);
1080 
1081 		/*
1082 		 * Mark the request as handled, since the inline shutdown
1083 		 * forces all outstanding requests to complete.
1084 		 */
1085 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1086 		return BLK_EH_HANDLED;
1087 	}
1088 
1089 	if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1090 		atomic_inc(&dev->ctrl.abort_limit);
1091 		return BLK_EH_RESET_TIMER;
1092 	}
1093 	iod->aborted = 1;
1094 
1095 	memset(&cmd, 0, sizeof(cmd));
1096 	cmd.abort.opcode = nvme_admin_abort_cmd;
1097 	cmd.abort.cid = req->tag;
1098 	cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1099 
1100 	dev_warn(nvmeq->dev->ctrl.device,
1101 		"I/O %d QID %d timeout, aborting\n",
1102 		 req->tag, nvmeq->qid);
1103 
1104 	abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1105 			BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1106 	if (IS_ERR(abort_req)) {
1107 		atomic_inc(&dev->ctrl.abort_limit);
1108 		return BLK_EH_RESET_TIMER;
1109 	}
1110 
1111 	abort_req->timeout = ADMIN_TIMEOUT;
1112 	abort_req->end_io_data = NULL;
1113 	blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1114 
1115 	/*
1116 	 * The aborted req will be completed on receiving the abort req.
1117 	 * We enable the timer again. If hit twice, it'll cause a device reset,
1118 	 * as the device then is in a faulty state.
1119 	 */
1120 	return BLK_EH_RESET_TIMER;
1121 }
1122 
1123 static void nvme_free_queue(struct nvme_queue *nvmeq)
1124 {
1125 	dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1126 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1127 	if (nvmeq->sq_cmds)
1128 		dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1129 					nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1130 	kfree(nvmeq);
1131 }
1132 
1133 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1134 {
1135 	int i;
1136 
1137 	for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1138 		struct nvme_queue *nvmeq = dev->queues[i];
1139 		dev->ctrl.queue_count--;
1140 		dev->queues[i] = NULL;
1141 		nvme_free_queue(nvmeq);
1142 	}
1143 }
1144 
1145 /**
1146  * nvme_suspend_queue - put queue into suspended state
1147  * @nvmeq - queue to suspend
1148  */
1149 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1150 {
1151 	int vector;
1152 
1153 	spin_lock_irq(&nvmeq->q_lock);
1154 	if (nvmeq->cq_vector == -1) {
1155 		spin_unlock_irq(&nvmeq->q_lock);
1156 		return 1;
1157 	}
1158 	vector = nvmeq->cq_vector;
1159 	nvmeq->dev->online_queues--;
1160 	nvmeq->cq_vector = -1;
1161 	spin_unlock_irq(&nvmeq->q_lock);
1162 
1163 	if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1164 		blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1165 
1166 	pci_free_irq(to_pci_dev(nvmeq->dev->dev), vector, nvmeq);
1167 
1168 	return 0;
1169 }
1170 
1171 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1172 {
1173 	struct nvme_queue *nvmeq = dev->queues[0];
1174 
1175 	if (!nvmeq)
1176 		return;
1177 	if (nvme_suspend_queue(nvmeq))
1178 		return;
1179 
1180 	if (shutdown)
1181 		nvme_shutdown_ctrl(&dev->ctrl);
1182 	else
1183 		nvme_disable_ctrl(&dev->ctrl, dev->ctrl.cap);
1184 
1185 	spin_lock_irq(&nvmeq->q_lock);
1186 	nvme_process_cq(nvmeq);
1187 	spin_unlock_irq(&nvmeq->q_lock);
1188 }
1189 
1190 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1191 				int entry_size)
1192 {
1193 	int q_depth = dev->q_depth;
1194 	unsigned q_size_aligned = roundup(q_depth * entry_size,
1195 					  dev->ctrl.page_size);
1196 
1197 	if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1198 		u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1199 		mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1200 		q_depth = div_u64(mem_per_q, entry_size);
1201 
1202 		/*
1203 		 * Ensure the reduced q_depth is above some threshold where it
1204 		 * would be better to map queues in system memory with the
1205 		 * original depth
1206 		 */
1207 		if (q_depth < 64)
1208 			return -ENOMEM;
1209 	}
1210 
1211 	return q_depth;
1212 }
1213 
1214 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1215 				int qid, int depth)
1216 {
1217 	if (qid && dev->cmb && use_cmb_sqes && NVME_CMB_SQS(dev->cmbsz)) {
1218 		unsigned offset = (qid - 1) * roundup(SQ_SIZE(depth),
1219 						      dev->ctrl.page_size);
1220 		nvmeq->sq_dma_addr = dev->cmb_dma_addr + offset;
1221 		nvmeq->sq_cmds_io = dev->cmb + offset;
1222 	} else {
1223 		nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1224 					&nvmeq->sq_dma_addr, GFP_KERNEL);
1225 		if (!nvmeq->sq_cmds)
1226 			return -ENOMEM;
1227 	}
1228 
1229 	return 0;
1230 }
1231 
1232 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1233 							int depth, int node)
1234 {
1235 	struct nvme_queue *nvmeq = kzalloc_node(sizeof(*nvmeq), GFP_KERNEL,
1236 							node);
1237 	if (!nvmeq)
1238 		return NULL;
1239 
1240 	nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1241 					  &nvmeq->cq_dma_addr, GFP_KERNEL);
1242 	if (!nvmeq->cqes)
1243 		goto free_nvmeq;
1244 
1245 	if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
1246 		goto free_cqdma;
1247 
1248 	nvmeq->q_dmadev = dev->dev;
1249 	nvmeq->dev = dev;
1250 	spin_lock_init(&nvmeq->q_lock);
1251 	nvmeq->cq_head = 0;
1252 	nvmeq->cq_phase = 1;
1253 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1254 	nvmeq->q_depth = depth;
1255 	nvmeq->qid = qid;
1256 	nvmeq->cq_vector = -1;
1257 	dev->queues[qid] = nvmeq;
1258 	dev->ctrl.queue_count++;
1259 
1260 	return nvmeq;
1261 
1262  free_cqdma:
1263 	dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1264 							nvmeq->cq_dma_addr);
1265  free_nvmeq:
1266 	kfree(nvmeq);
1267 	return NULL;
1268 }
1269 
1270 static int queue_request_irq(struct nvme_queue *nvmeq)
1271 {
1272 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1273 	int nr = nvmeq->dev->ctrl.instance;
1274 
1275 	if (use_threaded_interrupts) {
1276 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1277 				nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1278 	} else {
1279 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1280 				NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1281 	}
1282 }
1283 
1284 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1285 {
1286 	struct nvme_dev *dev = nvmeq->dev;
1287 
1288 	spin_lock_irq(&nvmeq->q_lock);
1289 	nvmeq->sq_tail = 0;
1290 	nvmeq->cq_head = 0;
1291 	nvmeq->cq_phase = 1;
1292 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1293 	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1294 	nvme_dbbuf_init(dev, nvmeq, qid);
1295 	dev->online_queues++;
1296 	spin_unlock_irq(&nvmeq->q_lock);
1297 }
1298 
1299 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1300 {
1301 	struct nvme_dev *dev = nvmeq->dev;
1302 	int result;
1303 
1304 	nvmeq->cq_vector = qid - 1;
1305 	result = adapter_alloc_cq(dev, qid, nvmeq);
1306 	if (result < 0)
1307 		return result;
1308 
1309 	result = adapter_alloc_sq(dev, qid, nvmeq);
1310 	if (result < 0)
1311 		goto release_cq;
1312 
1313 	result = queue_request_irq(nvmeq);
1314 	if (result < 0)
1315 		goto release_sq;
1316 
1317 	nvme_init_queue(nvmeq, qid);
1318 	return result;
1319 
1320  release_sq:
1321 	adapter_delete_sq(dev, qid);
1322  release_cq:
1323 	adapter_delete_cq(dev, qid);
1324 	return result;
1325 }
1326 
1327 static const struct blk_mq_ops nvme_mq_admin_ops = {
1328 	.queue_rq	= nvme_queue_rq,
1329 	.complete	= nvme_pci_complete_rq,
1330 	.init_hctx	= nvme_admin_init_hctx,
1331 	.exit_hctx      = nvme_admin_exit_hctx,
1332 	.init_request	= nvme_init_request,
1333 	.timeout	= nvme_timeout,
1334 };
1335 
1336 static const struct blk_mq_ops nvme_mq_ops = {
1337 	.queue_rq	= nvme_queue_rq,
1338 	.complete	= nvme_pci_complete_rq,
1339 	.init_hctx	= nvme_init_hctx,
1340 	.init_request	= nvme_init_request,
1341 	.map_queues	= nvme_pci_map_queues,
1342 	.timeout	= nvme_timeout,
1343 	.poll		= nvme_poll,
1344 };
1345 
1346 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1347 {
1348 	if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1349 		/*
1350 		 * If the controller was reset during removal, it's possible
1351 		 * user requests may be waiting on a stopped queue. Start the
1352 		 * queue to flush these to completion.
1353 		 */
1354 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1355 		blk_cleanup_queue(dev->ctrl.admin_q);
1356 		blk_mq_free_tag_set(&dev->admin_tagset);
1357 	}
1358 }
1359 
1360 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1361 {
1362 	if (!dev->ctrl.admin_q) {
1363 		dev->admin_tagset.ops = &nvme_mq_admin_ops;
1364 		dev->admin_tagset.nr_hw_queues = 1;
1365 
1366 		/*
1367 		 * Subtract one to leave an empty queue entry for 'Full Queue'
1368 		 * condition. See NVM-Express 1.2 specification, section 4.1.2.
1369 		 */
1370 		dev->admin_tagset.queue_depth = NVME_AQ_BLKMQ_DEPTH - 1;
1371 		dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1372 		dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1373 		dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1374 		dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1375 		dev->admin_tagset.driver_data = dev;
1376 
1377 		if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1378 			return -ENOMEM;
1379 
1380 		dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1381 		if (IS_ERR(dev->ctrl.admin_q)) {
1382 			blk_mq_free_tag_set(&dev->admin_tagset);
1383 			return -ENOMEM;
1384 		}
1385 		if (!blk_get_queue(dev->ctrl.admin_q)) {
1386 			nvme_dev_remove_admin(dev);
1387 			dev->ctrl.admin_q = NULL;
1388 			return -ENODEV;
1389 		}
1390 	} else
1391 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1392 
1393 	return 0;
1394 }
1395 
1396 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1397 {
1398 	return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1399 }
1400 
1401 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1402 {
1403 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1404 
1405 	if (size <= dev->bar_mapped_size)
1406 		return 0;
1407 	if (size > pci_resource_len(pdev, 0))
1408 		return -ENOMEM;
1409 	if (dev->bar)
1410 		iounmap(dev->bar);
1411 	dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1412 	if (!dev->bar) {
1413 		dev->bar_mapped_size = 0;
1414 		return -ENOMEM;
1415 	}
1416 	dev->bar_mapped_size = size;
1417 	dev->dbs = dev->bar + NVME_REG_DBS;
1418 
1419 	return 0;
1420 }
1421 
1422 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1423 {
1424 	int result;
1425 	u32 aqa;
1426 	struct nvme_queue *nvmeq;
1427 
1428 	result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1429 	if (result < 0)
1430 		return result;
1431 
1432 	dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1433 				NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1434 
1435 	if (dev->subsystem &&
1436 	    (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1437 		writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1438 
1439 	result = nvme_disable_ctrl(&dev->ctrl, dev->ctrl.cap);
1440 	if (result < 0)
1441 		return result;
1442 
1443 	nvmeq = dev->queues[0];
1444 	if (!nvmeq) {
1445 		nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH,
1446 					dev_to_node(dev->dev));
1447 		if (!nvmeq)
1448 			return -ENOMEM;
1449 	}
1450 
1451 	aqa = nvmeq->q_depth - 1;
1452 	aqa |= aqa << 16;
1453 
1454 	writel(aqa, dev->bar + NVME_REG_AQA);
1455 	lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1456 	lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1457 
1458 	result = nvme_enable_ctrl(&dev->ctrl, dev->ctrl.cap);
1459 	if (result)
1460 		return result;
1461 
1462 	nvmeq->cq_vector = 0;
1463 	result = queue_request_irq(nvmeq);
1464 	if (result) {
1465 		nvmeq->cq_vector = -1;
1466 		return result;
1467 	}
1468 
1469 	return result;
1470 }
1471 
1472 static int nvme_create_io_queues(struct nvme_dev *dev)
1473 {
1474 	unsigned i, max;
1475 	int ret = 0;
1476 
1477 	for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1478 		/* vector == qid - 1, match nvme_create_queue */
1479 		if (!nvme_alloc_queue(dev, i, dev->q_depth,
1480 		     pci_irq_get_node(to_pci_dev(dev->dev), i - 1))) {
1481 			ret = -ENOMEM;
1482 			break;
1483 		}
1484 	}
1485 
1486 	max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1487 	for (i = dev->online_queues; i <= max; i++) {
1488 		ret = nvme_create_queue(dev->queues[i], i);
1489 		if (ret)
1490 			break;
1491 	}
1492 
1493 	/*
1494 	 * Ignore failing Create SQ/CQ commands, we can continue with less
1495 	 * than the desired aount of queues, and even a controller without
1496 	 * I/O queues an still be used to issue admin commands.  This might
1497 	 * be useful to upgrade a buggy firmware for example.
1498 	 */
1499 	return ret >= 0 ? 0 : ret;
1500 }
1501 
1502 static ssize_t nvme_cmb_show(struct device *dev,
1503 			     struct device_attribute *attr,
1504 			     char *buf)
1505 {
1506 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1507 
1508 	return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz  : x%08x\n",
1509 		       ndev->cmbloc, ndev->cmbsz);
1510 }
1511 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1512 
1513 static void __iomem *nvme_map_cmb(struct nvme_dev *dev)
1514 {
1515 	u64 szu, size, offset;
1516 	resource_size_t bar_size;
1517 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1518 	void __iomem *cmb;
1519 	dma_addr_t dma_addr;
1520 
1521 	dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1522 	if (!(NVME_CMB_SZ(dev->cmbsz)))
1523 		return NULL;
1524 	dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1525 
1526 	if (!use_cmb_sqes)
1527 		return NULL;
1528 
1529 	szu = (u64)1 << (12 + 4 * NVME_CMB_SZU(dev->cmbsz));
1530 	size = szu * NVME_CMB_SZ(dev->cmbsz);
1531 	offset = szu * NVME_CMB_OFST(dev->cmbloc);
1532 	bar_size = pci_resource_len(pdev, NVME_CMB_BIR(dev->cmbloc));
1533 
1534 	if (offset > bar_size)
1535 		return NULL;
1536 
1537 	/*
1538 	 * Controllers may support a CMB size larger than their BAR,
1539 	 * for example, due to being behind a bridge. Reduce the CMB to
1540 	 * the reported size of the BAR
1541 	 */
1542 	if (size > bar_size - offset)
1543 		size = bar_size - offset;
1544 
1545 	dma_addr = pci_resource_start(pdev, NVME_CMB_BIR(dev->cmbloc)) + offset;
1546 	cmb = ioremap_wc(dma_addr, size);
1547 	if (!cmb)
1548 		return NULL;
1549 
1550 	dev->cmb_dma_addr = dma_addr;
1551 	dev->cmb_size = size;
1552 	return cmb;
1553 }
1554 
1555 static inline void nvme_release_cmb(struct nvme_dev *dev)
1556 {
1557 	if (dev->cmb) {
1558 		iounmap(dev->cmb);
1559 		dev->cmb = NULL;
1560 		sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1561 					     &dev_attr_cmb.attr, NULL);
1562 		dev->cmbsz = 0;
1563 	}
1564 }
1565 
1566 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1567 {
1568 	size_t len = dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs);
1569 	struct nvme_command c;
1570 	u64 dma_addr;
1571 	int ret;
1572 
1573 	dma_addr = dma_map_single(dev->dev, dev->host_mem_descs, len,
1574 			DMA_TO_DEVICE);
1575 	if (dma_mapping_error(dev->dev, dma_addr))
1576 		return -ENOMEM;
1577 
1578 	memset(&c, 0, sizeof(c));
1579 	c.features.opcode	= nvme_admin_set_features;
1580 	c.features.fid		= cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1581 	c.features.dword11	= cpu_to_le32(bits);
1582 	c.features.dword12	= cpu_to_le32(dev->host_mem_size >>
1583 					      ilog2(dev->ctrl.page_size));
1584 	c.features.dword13	= cpu_to_le32(lower_32_bits(dma_addr));
1585 	c.features.dword14	= cpu_to_le32(upper_32_bits(dma_addr));
1586 	c.features.dword15	= cpu_to_le32(dev->nr_host_mem_descs);
1587 
1588 	ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1589 	if (ret) {
1590 		dev_warn(dev->ctrl.device,
1591 			 "failed to set host mem (err %d, flags %#x).\n",
1592 			 ret, bits);
1593 	}
1594 	dma_unmap_single(dev->dev, dma_addr, len, DMA_TO_DEVICE);
1595 	return ret;
1596 }
1597 
1598 static void nvme_free_host_mem(struct nvme_dev *dev)
1599 {
1600 	int i;
1601 
1602 	for (i = 0; i < dev->nr_host_mem_descs; i++) {
1603 		struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1604 		size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1605 
1606 		dma_free_coherent(dev->dev, size, dev->host_mem_desc_bufs[i],
1607 				le64_to_cpu(desc->addr));
1608 	}
1609 
1610 	kfree(dev->host_mem_desc_bufs);
1611 	dev->host_mem_desc_bufs = NULL;
1612 	kfree(dev->host_mem_descs);
1613 	dev->host_mem_descs = NULL;
1614 }
1615 
1616 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1617 {
1618 	struct nvme_host_mem_buf_desc *descs;
1619 	u32 chunk_size, max_entries, len;
1620 	int i = 0;
1621 	void **bufs;
1622 	u64 size = 0, tmp;
1623 
1624 	/* start big and work our way down */
1625 	chunk_size = min(preferred, (u64)PAGE_SIZE << MAX_ORDER);
1626 retry:
1627 	tmp = (preferred + chunk_size - 1);
1628 	do_div(tmp, chunk_size);
1629 	max_entries = tmp;
1630 	descs = kcalloc(max_entries, sizeof(*descs), GFP_KERNEL);
1631 	if (!descs)
1632 		goto out;
1633 
1634 	bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1635 	if (!bufs)
1636 		goto out_free_descs;
1637 
1638 	for (size = 0; size < preferred; size += len) {
1639 		dma_addr_t dma_addr;
1640 
1641 		len = min_t(u64, chunk_size, preferred - size);
1642 		bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1643 				DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1644 		if (!bufs[i])
1645 			break;
1646 
1647 		descs[i].addr = cpu_to_le64(dma_addr);
1648 		descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1649 		i++;
1650 	}
1651 
1652 	if (!size || (min && size < min)) {
1653 		dev_warn(dev->ctrl.device,
1654 			"failed to allocate host memory buffer.\n");
1655 		goto out_free_bufs;
1656 	}
1657 
1658 	dev_info(dev->ctrl.device,
1659 		"allocated %lld MiB host memory buffer.\n",
1660 		size >> ilog2(SZ_1M));
1661 	dev->nr_host_mem_descs = i;
1662 	dev->host_mem_size = size;
1663 	dev->host_mem_descs = descs;
1664 	dev->host_mem_desc_bufs = bufs;
1665 	return 0;
1666 
1667 out_free_bufs:
1668 	while (--i >= 0) {
1669 		size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1670 
1671 		dma_free_coherent(dev->dev, size, bufs[i],
1672 				le64_to_cpu(descs[i].addr));
1673 	}
1674 
1675 	kfree(bufs);
1676 out_free_descs:
1677 	kfree(descs);
1678 out:
1679 	/* try a smaller chunk size if we failed early */
1680 	if (chunk_size >= PAGE_SIZE * 2 && (i == 0 || size < min)) {
1681 		chunk_size /= 2;
1682 		goto retry;
1683 	}
1684 	dev->host_mem_descs = NULL;
1685 	return -ENOMEM;
1686 }
1687 
1688 static void nvme_setup_host_mem(struct nvme_dev *dev)
1689 {
1690 	u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1691 	u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1692 	u64 min = (u64)dev->ctrl.hmmin * 4096;
1693 	u32 enable_bits = NVME_HOST_MEM_ENABLE;
1694 
1695 	preferred = min(preferred, max);
1696 	if (min > max) {
1697 		dev_warn(dev->ctrl.device,
1698 			"min host memory (%lld MiB) above limit (%d MiB).\n",
1699 			min >> ilog2(SZ_1M), max_host_mem_size_mb);
1700 		nvme_free_host_mem(dev);
1701 		return;
1702 	}
1703 
1704 	/*
1705 	 * If we already have a buffer allocated check if we can reuse it.
1706 	 */
1707 	if (dev->host_mem_descs) {
1708 		if (dev->host_mem_size >= min)
1709 			enable_bits |= NVME_HOST_MEM_RETURN;
1710 		else
1711 			nvme_free_host_mem(dev);
1712 	}
1713 
1714 	if (!dev->host_mem_descs) {
1715 		if (nvme_alloc_host_mem(dev, min, preferred))
1716 			return;
1717 	}
1718 
1719 	if (nvme_set_host_mem(dev, enable_bits))
1720 		nvme_free_host_mem(dev);
1721 }
1722 
1723 static int nvme_setup_io_queues(struct nvme_dev *dev)
1724 {
1725 	struct nvme_queue *adminq = dev->queues[0];
1726 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1727 	int result, nr_io_queues;
1728 	unsigned long size;
1729 
1730 	nr_io_queues = num_present_cpus();
1731 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
1732 	if (result < 0)
1733 		return result;
1734 
1735 	if (nr_io_queues == 0)
1736 		return 0;
1737 
1738 	if (dev->cmb && NVME_CMB_SQS(dev->cmbsz)) {
1739 		result = nvme_cmb_qdepth(dev, nr_io_queues,
1740 				sizeof(struct nvme_command));
1741 		if (result > 0)
1742 			dev->q_depth = result;
1743 		else
1744 			nvme_release_cmb(dev);
1745 	}
1746 
1747 	do {
1748 		size = db_bar_size(dev, nr_io_queues);
1749 		result = nvme_remap_bar(dev, size);
1750 		if (!result)
1751 			break;
1752 		if (!--nr_io_queues)
1753 			return -ENOMEM;
1754 	} while (1);
1755 	adminq->q_db = dev->dbs;
1756 
1757 	/* Deregister the admin queue's interrupt */
1758 	pci_free_irq(pdev, 0, adminq);
1759 
1760 	/*
1761 	 * If we enable msix early due to not intx, disable it again before
1762 	 * setting up the full range we need.
1763 	 */
1764 	pci_free_irq_vectors(pdev);
1765 	nr_io_queues = pci_alloc_irq_vectors(pdev, 1, nr_io_queues,
1766 			PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY);
1767 	if (nr_io_queues <= 0)
1768 		return -EIO;
1769 	dev->max_qid = nr_io_queues;
1770 
1771 	/*
1772 	 * Should investigate if there's a performance win from allocating
1773 	 * more queues than interrupt vectors; it might allow the submission
1774 	 * path to scale better, even if the receive path is limited by the
1775 	 * number of interrupts.
1776 	 */
1777 
1778 	result = queue_request_irq(adminq);
1779 	if (result) {
1780 		adminq->cq_vector = -1;
1781 		return result;
1782 	}
1783 	return nvme_create_io_queues(dev);
1784 }
1785 
1786 static void nvme_del_queue_end(struct request *req, blk_status_t error)
1787 {
1788 	struct nvme_queue *nvmeq = req->end_io_data;
1789 
1790 	blk_mq_free_request(req);
1791 	complete(&nvmeq->dev->ioq_wait);
1792 }
1793 
1794 static void nvme_del_cq_end(struct request *req, blk_status_t error)
1795 {
1796 	struct nvme_queue *nvmeq = req->end_io_data;
1797 
1798 	if (!error) {
1799 		unsigned long flags;
1800 
1801 		/*
1802 		 * We might be called with the AQ q_lock held
1803 		 * and the I/O queue q_lock should always
1804 		 * nest inside the AQ one.
1805 		 */
1806 		spin_lock_irqsave_nested(&nvmeq->q_lock, flags,
1807 					SINGLE_DEPTH_NESTING);
1808 		nvme_process_cq(nvmeq);
1809 		spin_unlock_irqrestore(&nvmeq->q_lock, flags);
1810 	}
1811 
1812 	nvme_del_queue_end(req, error);
1813 }
1814 
1815 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
1816 {
1817 	struct request_queue *q = nvmeq->dev->ctrl.admin_q;
1818 	struct request *req;
1819 	struct nvme_command cmd;
1820 
1821 	memset(&cmd, 0, sizeof(cmd));
1822 	cmd.delete_queue.opcode = opcode;
1823 	cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
1824 
1825 	req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1826 	if (IS_ERR(req))
1827 		return PTR_ERR(req);
1828 
1829 	req->timeout = ADMIN_TIMEOUT;
1830 	req->end_io_data = nvmeq;
1831 
1832 	blk_execute_rq_nowait(q, NULL, req, false,
1833 			opcode == nvme_admin_delete_cq ?
1834 				nvme_del_cq_end : nvme_del_queue_end);
1835 	return 0;
1836 }
1837 
1838 static void nvme_disable_io_queues(struct nvme_dev *dev, int queues)
1839 {
1840 	int pass;
1841 	unsigned long timeout;
1842 	u8 opcode = nvme_admin_delete_sq;
1843 
1844 	for (pass = 0; pass < 2; pass++) {
1845 		int sent = 0, i = queues;
1846 
1847 		reinit_completion(&dev->ioq_wait);
1848  retry:
1849 		timeout = ADMIN_TIMEOUT;
1850 		for (; i > 0; i--, sent++)
1851 			if (nvme_delete_queue(dev->queues[i], opcode))
1852 				break;
1853 
1854 		while (sent--) {
1855 			timeout = wait_for_completion_io_timeout(&dev->ioq_wait, timeout);
1856 			if (timeout == 0)
1857 				return;
1858 			if (i)
1859 				goto retry;
1860 		}
1861 		opcode = nvme_admin_delete_cq;
1862 	}
1863 }
1864 
1865 /*
1866  * Return: error value if an error occurred setting up the queues or calling
1867  * Identify Device.  0 if these succeeded, even if adding some of the
1868  * namespaces failed.  At the moment, these failures are silent.  TBD which
1869  * failures should be reported.
1870  */
1871 static int nvme_dev_add(struct nvme_dev *dev)
1872 {
1873 	if (!dev->ctrl.tagset) {
1874 		dev->tagset.ops = &nvme_mq_ops;
1875 		dev->tagset.nr_hw_queues = dev->online_queues - 1;
1876 		dev->tagset.timeout = NVME_IO_TIMEOUT;
1877 		dev->tagset.numa_node = dev_to_node(dev->dev);
1878 		dev->tagset.queue_depth =
1879 				min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
1880 		dev->tagset.cmd_size = nvme_cmd_size(dev);
1881 		dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
1882 		dev->tagset.driver_data = dev;
1883 
1884 		if (blk_mq_alloc_tag_set(&dev->tagset))
1885 			return 0;
1886 		dev->ctrl.tagset = &dev->tagset;
1887 
1888 		nvme_dbbuf_set(dev);
1889 	} else {
1890 		blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
1891 
1892 		/* Free previously allocated queues that are no longer usable */
1893 		nvme_free_queues(dev, dev->online_queues);
1894 	}
1895 
1896 	return 0;
1897 }
1898 
1899 static int nvme_pci_enable(struct nvme_dev *dev)
1900 {
1901 	int result = -ENOMEM;
1902 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1903 
1904 	if (pci_enable_device_mem(pdev))
1905 		return result;
1906 
1907 	pci_set_master(pdev);
1908 
1909 	if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
1910 	    dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
1911 		goto disable;
1912 
1913 	if (readl(dev->bar + NVME_REG_CSTS) == -1) {
1914 		result = -ENODEV;
1915 		goto disable;
1916 	}
1917 
1918 	/*
1919 	 * Some devices and/or platforms don't advertise or work with INTx
1920 	 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
1921 	 * adjust this later.
1922 	 */
1923 	result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
1924 	if (result < 0)
1925 		return result;
1926 
1927 	dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1928 
1929 	dev->q_depth = min_t(int, NVME_CAP_MQES(dev->ctrl.cap) + 1,
1930 				io_queue_depth);
1931 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
1932 	dev->dbs = dev->bar + 4096;
1933 
1934 	/*
1935 	 * Temporary fix for the Apple controller found in the MacBook8,1 and
1936 	 * some MacBook7,1 to avoid controller resets and data loss.
1937 	 */
1938 	if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
1939 		dev->q_depth = 2;
1940 		dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
1941 			"set queue depth=%u to work around controller resets\n",
1942 			dev->q_depth);
1943 	} else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
1944 		   (pdev->device == 0xa821 || pdev->device == 0xa822) &&
1945 		   NVME_CAP_MQES(dev->ctrl.cap) == 0) {
1946 		dev->q_depth = 64;
1947 		dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
1948                         "set queue depth=%u\n", dev->q_depth);
1949 	}
1950 
1951 	/*
1952 	 * CMBs can currently only exist on >=1.2 PCIe devices. We only
1953 	 * populate sysfs if a CMB is implemented. Since nvme_dev_attrs_group
1954 	 * has no name we can pass NULL as final argument to
1955 	 * sysfs_add_file_to_group.
1956 	 */
1957 
1958 	if (readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 2, 0)) {
1959 		dev->cmb = nvme_map_cmb(dev);
1960 		if (dev->cmb) {
1961 			if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1962 						    &dev_attr_cmb.attr, NULL))
1963 				dev_warn(dev->ctrl.device,
1964 					 "failed to add sysfs attribute for CMB\n");
1965 		}
1966 	}
1967 
1968 	pci_enable_pcie_error_reporting(pdev);
1969 	pci_save_state(pdev);
1970 	return 0;
1971 
1972  disable:
1973 	pci_disable_device(pdev);
1974 	return result;
1975 }
1976 
1977 static void nvme_dev_unmap(struct nvme_dev *dev)
1978 {
1979 	if (dev->bar)
1980 		iounmap(dev->bar);
1981 	pci_release_mem_regions(to_pci_dev(dev->dev));
1982 }
1983 
1984 static void nvme_pci_disable(struct nvme_dev *dev)
1985 {
1986 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1987 
1988 	nvme_release_cmb(dev);
1989 	pci_free_irq_vectors(pdev);
1990 
1991 	if (pci_is_enabled(pdev)) {
1992 		pci_disable_pcie_error_reporting(pdev);
1993 		pci_disable_device(pdev);
1994 	}
1995 }
1996 
1997 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
1998 {
1999 	int i, queues;
2000 	bool dead = true;
2001 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2002 
2003 	mutex_lock(&dev->shutdown_lock);
2004 	if (pci_is_enabled(pdev)) {
2005 		u32 csts = readl(dev->bar + NVME_REG_CSTS);
2006 
2007 		if (dev->ctrl.state == NVME_CTRL_LIVE ||
2008 		    dev->ctrl.state == NVME_CTRL_RESETTING)
2009 			nvme_start_freeze(&dev->ctrl);
2010 		dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2011 			pdev->error_state  != pci_channel_io_normal);
2012 	}
2013 
2014 	/*
2015 	 * Give the controller a chance to complete all entered requests if
2016 	 * doing a safe shutdown.
2017 	 */
2018 	if (!dead) {
2019 		if (shutdown)
2020 			nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2021 
2022 		/*
2023 		 * If the controller is still alive tell it to stop using the
2024 		 * host memory buffer.  In theory the shutdown / reset should
2025 		 * make sure that it doesn't access the host memoery anymore,
2026 		 * but I'd rather be safe than sorry..
2027 		 */
2028 		if (dev->host_mem_descs)
2029 			nvme_set_host_mem(dev, 0);
2030 
2031 	}
2032 	nvme_stop_queues(&dev->ctrl);
2033 
2034 	queues = dev->online_queues - 1;
2035 	for (i = dev->ctrl.queue_count - 1; i > 0; i--)
2036 		nvme_suspend_queue(dev->queues[i]);
2037 
2038 	if (dead) {
2039 		/* A device might become IO incapable very soon during
2040 		 * probe, before the admin queue is configured. Thus,
2041 		 * queue_count can be 0 here.
2042 		 */
2043 		if (dev->ctrl.queue_count)
2044 			nvme_suspend_queue(dev->queues[0]);
2045 	} else {
2046 		nvme_disable_io_queues(dev, queues);
2047 		nvme_disable_admin_queue(dev, shutdown);
2048 	}
2049 	nvme_pci_disable(dev);
2050 
2051 	blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2052 	blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2053 
2054 	/*
2055 	 * The driver will not be starting up queues again if shutting down so
2056 	 * must flush all entered requests to their failed completion to avoid
2057 	 * deadlocking blk-mq hot-cpu notifier.
2058 	 */
2059 	if (shutdown)
2060 		nvme_start_queues(&dev->ctrl);
2061 	mutex_unlock(&dev->shutdown_lock);
2062 }
2063 
2064 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2065 {
2066 	dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2067 						PAGE_SIZE, PAGE_SIZE, 0);
2068 	if (!dev->prp_page_pool)
2069 		return -ENOMEM;
2070 
2071 	/* Optimisation for I/Os between 4k and 128k */
2072 	dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2073 						256, 256, 0);
2074 	if (!dev->prp_small_pool) {
2075 		dma_pool_destroy(dev->prp_page_pool);
2076 		return -ENOMEM;
2077 	}
2078 	return 0;
2079 }
2080 
2081 static void nvme_release_prp_pools(struct nvme_dev *dev)
2082 {
2083 	dma_pool_destroy(dev->prp_page_pool);
2084 	dma_pool_destroy(dev->prp_small_pool);
2085 }
2086 
2087 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2088 {
2089 	struct nvme_dev *dev = to_nvme_dev(ctrl);
2090 
2091 	nvme_dbbuf_dma_free(dev);
2092 	put_device(dev->dev);
2093 	if (dev->tagset.tags)
2094 		blk_mq_free_tag_set(&dev->tagset);
2095 	if (dev->ctrl.admin_q)
2096 		blk_put_queue(dev->ctrl.admin_q);
2097 	kfree(dev->queues);
2098 	free_opal_dev(dev->ctrl.opal_dev);
2099 	kfree(dev);
2100 }
2101 
2102 static void nvme_remove_dead_ctrl(struct nvme_dev *dev, int status)
2103 {
2104 	dev_warn(dev->ctrl.device, "Removing after probe failure status: %d\n", status);
2105 
2106 	kref_get(&dev->ctrl.kref);
2107 	nvme_dev_disable(dev, false);
2108 	if (!schedule_work(&dev->remove_work))
2109 		nvme_put_ctrl(&dev->ctrl);
2110 }
2111 
2112 static void nvme_reset_work(struct work_struct *work)
2113 {
2114 	struct nvme_dev *dev =
2115 		container_of(work, struct nvme_dev, ctrl.reset_work);
2116 	bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2117 	int result = -ENODEV;
2118 
2119 	if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING))
2120 		goto out;
2121 
2122 	/*
2123 	 * If we're called to reset a live controller first shut it down before
2124 	 * moving on.
2125 	 */
2126 	if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2127 		nvme_dev_disable(dev, false);
2128 
2129 	result = nvme_pci_enable(dev);
2130 	if (result)
2131 		goto out;
2132 
2133 	result = nvme_pci_configure_admin_queue(dev);
2134 	if (result)
2135 		goto out;
2136 
2137 	nvme_init_queue(dev->queues[0], 0);
2138 	result = nvme_alloc_admin_tags(dev);
2139 	if (result)
2140 		goto out;
2141 
2142 	result = nvme_init_identify(&dev->ctrl);
2143 	if (result)
2144 		goto out;
2145 
2146 	if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2147 		if (!dev->ctrl.opal_dev)
2148 			dev->ctrl.opal_dev =
2149 				init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2150 		else if (was_suspend)
2151 			opal_unlock_from_suspend(dev->ctrl.opal_dev);
2152 	} else {
2153 		free_opal_dev(dev->ctrl.opal_dev);
2154 		dev->ctrl.opal_dev = NULL;
2155 	}
2156 
2157 	if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2158 		result = nvme_dbbuf_dma_alloc(dev);
2159 		if (result)
2160 			dev_warn(dev->dev,
2161 				 "unable to allocate dma for dbbuf\n");
2162 	}
2163 
2164 	if (dev->ctrl.hmpre)
2165 		nvme_setup_host_mem(dev);
2166 
2167 	result = nvme_setup_io_queues(dev);
2168 	if (result)
2169 		goto out;
2170 
2171 	/*
2172 	 * Keep the controller around but remove all namespaces if we don't have
2173 	 * any working I/O queue.
2174 	 */
2175 	if (dev->online_queues < 2) {
2176 		dev_warn(dev->ctrl.device, "IO queues not created\n");
2177 		nvme_kill_queues(&dev->ctrl);
2178 		nvme_remove_namespaces(&dev->ctrl);
2179 	} else {
2180 		nvme_start_queues(&dev->ctrl);
2181 		nvme_wait_freeze(&dev->ctrl);
2182 		nvme_dev_add(dev);
2183 		nvme_unfreeze(&dev->ctrl);
2184 	}
2185 
2186 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2187 		dev_warn(dev->ctrl.device, "failed to mark controller live\n");
2188 		goto out;
2189 	}
2190 
2191 	nvme_start_ctrl(&dev->ctrl);
2192 	return;
2193 
2194  out:
2195 	nvme_remove_dead_ctrl(dev, result);
2196 }
2197 
2198 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2199 {
2200 	struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2201 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2202 
2203 	nvme_kill_queues(&dev->ctrl);
2204 	if (pci_get_drvdata(pdev))
2205 		device_release_driver(&pdev->dev);
2206 	nvme_put_ctrl(&dev->ctrl);
2207 }
2208 
2209 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2210 {
2211 	*val = readl(to_nvme_dev(ctrl)->bar + off);
2212 	return 0;
2213 }
2214 
2215 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2216 {
2217 	writel(val, to_nvme_dev(ctrl)->bar + off);
2218 	return 0;
2219 }
2220 
2221 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2222 {
2223 	*val = readq(to_nvme_dev(ctrl)->bar + off);
2224 	return 0;
2225 }
2226 
2227 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2228 	.name			= "pcie",
2229 	.module			= THIS_MODULE,
2230 	.flags			= NVME_F_METADATA_SUPPORTED,
2231 	.reg_read32		= nvme_pci_reg_read32,
2232 	.reg_write32		= nvme_pci_reg_write32,
2233 	.reg_read64		= nvme_pci_reg_read64,
2234 	.free_ctrl		= nvme_pci_free_ctrl,
2235 	.submit_async_event	= nvme_pci_submit_async_event,
2236 };
2237 
2238 static int nvme_dev_map(struct nvme_dev *dev)
2239 {
2240 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2241 
2242 	if (pci_request_mem_regions(pdev, "nvme"))
2243 		return -ENODEV;
2244 
2245 	if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2246 		goto release;
2247 
2248 	return 0;
2249   release:
2250 	pci_release_mem_regions(pdev);
2251 	return -ENODEV;
2252 }
2253 
2254 static unsigned long check_dell_samsung_bug(struct pci_dev *pdev)
2255 {
2256 	if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2257 		/*
2258 		 * Several Samsung devices seem to drop off the PCIe bus
2259 		 * randomly when APST is on and uses the deepest sleep state.
2260 		 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2261 		 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2262 		 * 950 PRO 256GB", but it seems to be restricted to two Dell
2263 		 * laptops.
2264 		 */
2265 		if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2266 		    (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2267 		     dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2268 			return NVME_QUIRK_NO_DEEPEST_PS;
2269 	}
2270 
2271 	return 0;
2272 }
2273 
2274 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2275 {
2276 	int node, result = -ENOMEM;
2277 	struct nvme_dev *dev;
2278 	unsigned long quirks = id->driver_data;
2279 
2280 	node = dev_to_node(&pdev->dev);
2281 	if (node == NUMA_NO_NODE)
2282 		set_dev_node(&pdev->dev, first_memory_node);
2283 
2284 	dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2285 	if (!dev)
2286 		return -ENOMEM;
2287 	dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
2288 							GFP_KERNEL, node);
2289 	if (!dev->queues)
2290 		goto free;
2291 
2292 	dev->dev = get_device(&pdev->dev);
2293 	pci_set_drvdata(pdev, dev);
2294 
2295 	result = nvme_dev_map(dev);
2296 	if (result)
2297 		goto put_pci;
2298 
2299 	INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2300 	INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2301 	mutex_init(&dev->shutdown_lock);
2302 	init_completion(&dev->ioq_wait);
2303 
2304 	result = nvme_setup_prp_pools(dev);
2305 	if (result)
2306 		goto unmap;
2307 
2308 	quirks |= check_dell_samsung_bug(pdev);
2309 
2310 	result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2311 			quirks);
2312 	if (result)
2313 		goto release_pools;
2314 
2315 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING);
2316 	dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2317 
2318 	queue_work(nvme_wq, &dev->ctrl.reset_work);
2319 	return 0;
2320 
2321  release_pools:
2322 	nvme_release_prp_pools(dev);
2323  unmap:
2324 	nvme_dev_unmap(dev);
2325  put_pci:
2326 	put_device(dev->dev);
2327  free:
2328 	kfree(dev->queues);
2329 	kfree(dev);
2330 	return result;
2331 }
2332 
2333 static void nvme_reset_prepare(struct pci_dev *pdev)
2334 {
2335 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2336 	nvme_dev_disable(dev, false);
2337 }
2338 
2339 static void nvme_reset_done(struct pci_dev *pdev)
2340 {
2341 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2342 	nvme_reset_ctrl(&dev->ctrl);
2343 }
2344 
2345 static void nvme_shutdown(struct pci_dev *pdev)
2346 {
2347 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2348 	nvme_dev_disable(dev, true);
2349 }
2350 
2351 /*
2352  * The driver's remove may be called on a device in a partially initialized
2353  * state. This function must not have any dependencies on the device state in
2354  * order to proceed.
2355  */
2356 static void nvme_remove(struct pci_dev *pdev)
2357 {
2358 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2359 
2360 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2361 
2362 	cancel_work_sync(&dev->ctrl.reset_work);
2363 	pci_set_drvdata(pdev, NULL);
2364 
2365 	if (!pci_device_is_present(pdev)) {
2366 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2367 		nvme_dev_disable(dev, false);
2368 	}
2369 
2370 	flush_work(&dev->ctrl.reset_work);
2371 	nvme_stop_ctrl(&dev->ctrl);
2372 	nvme_remove_namespaces(&dev->ctrl);
2373 	nvme_dev_disable(dev, true);
2374 	nvme_free_host_mem(dev);
2375 	nvme_dev_remove_admin(dev);
2376 	nvme_free_queues(dev, 0);
2377 	nvme_uninit_ctrl(&dev->ctrl);
2378 	nvme_release_prp_pools(dev);
2379 	nvme_dev_unmap(dev);
2380 	nvme_put_ctrl(&dev->ctrl);
2381 }
2382 
2383 static int nvme_pci_sriov_configure(struct pci_dev *pdev, int numvfs)
2384 {
2385 	int ret = 0;
2386 
2387 	if (numvfs == 0) {
2388 		if (pci_vfs_assigned(pdev)) {
2389 			dev_warn(&pdev->dev,
2390 				"Cannot disable SR-IOV VFs while assigned\n");
2391 			return -EPERM;
2392 		}
2393 		pci_disable_sriov(pdev);
2394 		return 0;
2395 	}
2396 
2397 	ret = pci_enable_sriov(pdev, numvfs);
2398 	return ret ? ret : numvfs;
2399 }
2400 
2401 #ifdef CONFIG_PM_SLEEP
2402 static int nvme_suspend(struct device *dev)
2403 {
2404 	struct pci_dev *pdev = to_pci_dev(dev);
2405 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2406 
2407 	nvme_dev_disable(ndev, true);
2408 	return 0;
2409 }
2410 
2411 static int nvme_resume(struct device *dev)
2412 {
2413 	struct pci_dev *pdev = to_pci_dev(dev);
2414 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2415 
2416 	nvme_reset_ctrl(&ndev->ctrl);
2417 	return 0;
2418 }
2419 #endif
2420 
2421 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2422 
2423 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
2424 						pci_channel_state_t state)
2425 {
2426 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2427 
2428 	/*
2429 	 * A frozen channel requires a reset. When detected, this method will
2430 	 * shutdown the controller to quiesce. The controller will be restarted
2431 	 * after the slot reset through driver's slot_reset callback.
2432 	 */
2433 	switch (state) {
2434 	case pci_channel_io_normal:
2435 		return PCI_ERS_RESULT_CAN_RECOVER;
2436 	case pci_channel_io_frozen:
2437 		dev_warn(dev->ctrl.device,
2438 			"frozen state error detected, reset controller\n");
2439 		nvme_dev_disable(dev, false);
2440 		return PCI_ERS_RESULT_NEED_RESET;
2441 	case pci_channel_io_perm_failure:
2442 		dev_warn(dev->ctrl.device,
2443 			"failure state error detected, request disconnect\n");
2444 		return PCI_ERS_RESULT_DISCONNECT;
2445 	}
2446 	return PCI_ERS_RESULT_NEED_RESET;
2447 }
2448 
2449 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
2450 {
2451 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2452 
2453 	dev_info(dev->ctrl.device, "restart after slot reset\n");
2454 	pci_restore_state(pdev);
2455 	nvme_reset_ctrl(&dev->ctrl);
2456 	return PCI_ERS_RESULT_RECOVERED;
2457 }
2458 
2459 static void nvme_error_resume(struct pci_dev *pdev)
2460 {
2461 	pci_cleanup_aer_uncorrect_error_status(pdev);
2462 }
2463 
2464 static const struct pci_error_handlers nvme_err_handler = {
2465 	.error_detected	= nvme_error_detected,
2466 	.slot_reset	= nvme_slot_reset,
2467 	.resume		= nvme_error_resume,
2468 	.reset_prepare	= nvme_reset_prepare,
2469 	.reset_done	= nvme_reset_done,
2470 };
2471 
2472 static const struct pci_device_id nvme_id_table[] = {
2473 	{ PCI_VDEVICE(INTEL, 0x0953),
2474 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2475 				NVME_QUIRK_DEALLOCATE_ZEROES, },
2476 	{ PCI_VDEVICE(INTEL, 0x0a53),
2477 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2478 				NVME_QUIRK_DEALLOCATE_ZEROES, },
2479 	{ PCI_VDEVICE(INTEL, 0x0a54),
2480 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2481 				NVME_QUIRK_DEALLOCATE_ZEROES, },
2482 	{ PCI_VDEVICE(INTEL, 0x0a55),
2483 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2484 				NVME_QUIRK_DEALLOCATE_ZEROES, },
2485 	{ PCI_VDEVICE(INTEL, 0xf1a5),	/* Intel 600P/P3100 */
2486 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS },
2487 	{ PCI_VDEVICE(INTEL, 0x5845),	/* Qemu emulated controller */
2488 		.driver_data = NVME_QUIRK_IDENTIFY_CNS, },
2489 	{ PCI_DEVICE(0x1c58, 0x0003),	/* HGST adapter */
2490 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2491 	{ PCI_DEVICE(0x1c5f, 0x0540),	/* Memblaze Pblaze4 adapter */
2492 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2493 	{ PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
2494 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2495 	{ PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
2496 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2497 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2498 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
2499 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
2500 	{ 0, }
2501 };
2502 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2503 
2504 static struct pci_driver nvme_driver = {
2505 	.name		= "nvme",
2506 	.id_table	= nvme_id_table,
2507 	.probe		= nvme_probe,
2508 	.remove		= nvme_remove,
2509 	.shutdown	= nvme_shutdown,
2510 	.driver		= {
2511 		.pm	= &nvme_dev_pm_ops,
2512 	},
2513 	.sriov_configure = nvme_pci_sriov_configure,
2514 	.err_handler	= &nvme_err_handler,
2515 };
2516 
2517 static int __init nvme_init(void)
2518 {
2519 	return pci_register_driver(&nvme_driver);
2520 }
2521 
2522 static void __exit nvme_exit(void)
2523 {
2524 	pci_unregister_driver(&nvme_driver);
2525 	_nvme_check_size();
2526 }
2527 
2528 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2529 MODULE_LICENSE("GPL");
2530 MODULE_VERSION("1.0");
2531 module_init(nvme_init);
2532 module_exit(nvme_exit);
2533