xref: /openbmc/linux/drivers/nvme/host/pci.c (revision 04b3c795)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * NVM Express device driver
4  * Copyright (c) 2011-2014, Intel Corporation.
5  */
6 
7 #include <linux/aer.h>
8 #include <linux/async.h>
9 #include <linux/blkdev.h>
10 #include <linux/blk-mq.h>
11 #include <linux/blk-mq-pci.h>
12 #include <linux/dmi.h>
13 #include <linux/init.h>
14 #include <linux/interrupt.h>
15 #include <linux/io.h>
16 #include <linux/mm.h>
17 #include <linux/module.h>
18 #include <linux/mutex.h>
19 #include <linux/once.h>
20 #include <linux/pci.h>
21 #include <linux/suspend.h>
22 #include <linux/t10-pi.h>
23 #include <linux/types.h>
24 #include <linux/io-64-nonatomic-lo-hi.h>
25 #include <linux/sed-opal.h>
26 #include <linux/pci-p2pdma.h>
27 
28 #include "trace.h"
29 #include "nvme.h"
30 
31 #define SQ_SIZE(q)	((q)->q_depth << (q)->sqes)
32 #define CQ_SIZE(q)	((q)->q_depth * sizeof(struct nvme_completion))
33 
34 #define SGES_PER_PAGE	(PAGE_SIZE / sizeof(struct nvme_sgl_desc))
35 
36 /*
37  * These can be higher, but we need to ensure that any command doesn't
38  * require an sg allocation that needs more than a page of data.
39  */
40 #define NVME_MAX_KB_SZ	4096
41 #define NVME_MAX_SEGS	127
42 
43 static int use_threaded_interrupts;
44 module_param(use_threaded_interrupts, int, 0);
45 
46 static bool use_cmb_sqes = true;
47 module_param(use_cmb_sqes, bool, 0444);
48 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
49 
50 static unsigned int max_host_mem_size_mb = 128;
51 module_param(max_host_mem_size_mb, uint, 0444);
52 MODULE_PARM_DESC(max_host_mem_size_mb,
53 	"Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
54 
55 static unsigned int sgl_threshold = SZ_32K;
56 module_param(sgl_threshold, uint, 0644);
57 MODULE_PARM_DESC(sgl_threshold,
58 		"Use SGLs when average request segment size is larger or equal to "
59 		"this size. Use 0 to disable SGLs.");
60 
61 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
62 static const struct kernel_param_ops io_queue_depth_ops = {
63 	.set = io_queue_depth_set,
64 	.get = param_get_int,
65 };
66 
67 static int io_queue_depth = 1024;
68 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
69 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2");
70 
71 static int io_queue_count_set(const char *val, const struct kernel_param *kp)
72 {
73 	unsigned int n;
74 	int ret;
75 
76 	ret = kstrtouint(val, 10, &n);
77 	if (ret != 0 || n > num_possible_cpus())
78 		return -EINVAL;
79 	return param_set_uint(val, kp);
80 }
81 
82 static const struct kernel_param_ops io_queue_count_ops = {
83 	.set = io_queue_count_set,
84 	.get = param_get_uint,
85 };
86 
87 static unsigned int write_queues;
88 module_param_cb(write_queues, &io_queue_count_ops, &write_queues, 0644);
89 MODULE_PARM_DESC(write_queues,
90 	"Number of queues to use for writes. If not set, reads and writes "
91 	"will share a queue set.");
92 
93 static unsigned int poll_queues;
94 module_param_cb(poll_queues, &io_queue_count_ops, &poll_queues, 0644);
95 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
96 
97 struct nvme_dev;
98 struct nvme_queue;
99 
100 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
101 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode);
102 
103 /*
104  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
105  */
106 struct nvme_dev {
107 	struct nvme_queue *queues;
108 	struct blk_mq_tag_set tagset;
109 	struct blk_mq_tag_set admin_tagset;
110 	u32 __iomem *dbs;
111 	struct device *dev;
112 	struct dma_pool *prp_page_pool;
113 	struct dma_pool *prp_small_pool;
114 	unsigned online_queues;
115 	unsigned max_qid;
116 	unsigned io_queues[HCTX_MAX_TYPES];
117 	unsigned int num_vecs;
118 	int q_depth;
119 	int io_sqes;
120 	u32 db_stride;
121 	void __iomem *bar;
122 	unsigned long bar_mapped_size;
123 	struct work_struct remove_work;
124 	struct mutex shutdown_lock;
125 	bool subsystem;
126 	u64 cmb_size;
127 	bool cmb_use_sqes;
128 	u32 cmbsz;
129 	u32 cmbloc;
130 	struct nvme_ctrl ctrl;
131 	u32 last_ps;
132 
133 	mempool_t *iod_mempool;
134 
135 	/* shadow doorbell buffer support: */
136 	u32 *dbbuf_dbs;
137 	dma_addr_t dbbuf_dbs_dma_addr;
138 	u32 *dbbuf_eis;
139 	dma_addr_t dbbuf_eis_dma_addr;
140 
141 	/* host memory buffer support: */
142 	u64 host_mem_size;
143 	u32 nr_host_mem_descs;
144 	dma_addr_t host_mem_descs_dma;
145 	struct nvme_host_mem_buf_desc *host_mem_descs;
146 	void **host_mem_desc_bufs;
147 	unsigned int nr_allocated_queues;
148 	unsigned int nr_write_queues;
149 	unsigned int nr_poll_queues;
150 };
151 
152 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
153 {
154 	int n = 0, ret;
155 
156 	ret = kstrtoint(val, 10, &n);
157 	if (ret != 0 || n < 2)
158 		return -EINVAL;
159 
160 	return param_set_int(val, kp);
161 }
162 
163 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
164 {
165 	return qid * 2 * stride;
166 }
167 
168 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
169 {
170 	return (qid * 2 + 1) * stride;
171 }
172 
173 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
174 {
175 	return container_of(ctrl, struct nvme_dev, ctrl);
176 }
177 
178 /*
179  * An NVM Express queue.  Each device has at least two (one for admin
180  * commands and one for I/O commands).
181  */
182 struct nvme_queue {
183 	struct nvme_dev *dev;
184 	spinlock_t sq_lock;
185 	void *sq_cmds;
186 	 /* only used for poll queues: */
187 	spinlock_t cq_poll_lock ____cacheline_aligned_in_smp;
188 	struct nvme_completion *cqes;
189 	dma_addr_t sq_dma_addr;
190 	dma_addr_t cq_dma_addr;
191 	u32 __iomem *q_db;
192 	u16 q_depth;
193 	u16 cq_vector;
194 	u16 sq_tail;
195 	u16 cq_head;
196 	u16 qid;
197 	u8 cq_phase;
198 	u8 sqes;
199 	unsigned long flags;
200 #define NVMEQ_ENABLED		0
201 #define NVMEQ_SQ_CMB		1
202 #define NVMEQ_DELETE_ERROR	2
203 #define NVMEQ_POLLED		3
204 	u32 *dbbuf_sq_db;
205 	u32 *dbbuf_cq_db;
206 	u32 *dbbuf_sq_ei;
207 	u32 *dbbuf_cq_ei;
208 	struct completion delete_done;
209 };
210 
211 /*
212  * The nvme_iod describes the data in an I/O.
213  *
214  * The sg pointer contains the list of PRP/SGL chunk allocations in addition
215  * to the actual struct scatterlist.
216  */
217 struct nvme_iod {
218 	struct nvme_request req;
219 	struct nvme_queue *nvmeq;
220 	bool use_sgl;
221 	int aborted;
222 	int npages;		/* In the PRP list. 0 means small pool in use */
223 	int nents;		/* Used in scatterlist */
224 	dma_addr_t first_dma;
225 	unsigned int dma_len;	/* length of single DMA segment mapping */
226 	dma_addr_t meta_dma;
227 	struct scatterlist *sg;
228 };
229 
230 static inline unsigned int nvme_dbbuf_size(struct nvme_dev *dev)
231 {
232 	return dev->nr_allocated_queues * 8 * dev->db_stride;
233 }
234 
235 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
236 {
237 	unsigned int mem_size = nvme_dbbuf_size(dev);
238 
239 	if (dev->dbbuf_dbs)
240 		return 0;
241 
242 	dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
243 					    &dev->dbbuf_dbs_dma_addr,
244 					    GFP_KERNEL);
245 	if (!dev->dbbuf_dbs)
246 		return -ENOMEM;
247 	dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
248 					    &dev->dbbuf_eis_dma_addr,
249 					    GFP_KERNEL);
250 	if (!dev->dbbuf_eis) {
251 		dma_free_coherent(dev->dev, mem_size,
252 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
253 		dev->dbbuf_dbs = NULL;
254 		return -ENOMEM;
255 	}
256 
257 	return 0;
258 }
259 
260 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
261 {
262 	unsigned int mem_size = nvme_dbbuf_size(dev);
263 
264 	if (dev->dbbuf_dbs) {
265 		dma_free_coherent(dev->dev, mem_size,
266 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
267 		dev->dbbuf_dbs = NULL;
268 	}
269 	if (dev->dbbuf_eis) {
270 		dma_free_coherent(dev->dev, mem_size,
271 				  dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
272 		dev->dbbuf_eis = NULL;
273 	}
274 }
275 
276 static void nvme_dbbuf_init(struct nvme_dev *dev,
277 			    struct nvme_queue *nvmeq, int qid)
278 {
279 	if (!dev->dbbuf_dbs || !qid)
280 		return;
281 
282 	nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
283 	nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
284 	nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
285 	nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
286 }
287 
288 static void nvme_dbbuf_set(struct nvme_dev *dev)
289 {
290 	struct nvme_command c;
291 
292 	if (!dev->dbbuf_dbs)
293 		return;
294 
295 	memset(&c, 0, sizeof(c));
296 	c.dbbuf.opcode = nvme_admin_dbbuf;
297 	c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
298 	c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
299 
300 	if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
301 		dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
302 		/* Free memory and continue on */
303 		nvme_dbbuf_dma_free(dev);
304 	}
305 }
306 
307 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
308 {
309 	return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
310 }
311 
312 /* Update dbbuf and return true if an MMIO is required */
313 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
314 					      volatile u32 *dbbuf_ei)
315 {
316 	if (dbbuf_db) {
317 		u16 old_value;
318 
319 		/*
320 		 * Ensure that the queue is written before updating
321 		 * the doorbell in memory
322 		 */
323 		wmb();
324 
325 		old_value = *dbbuf_db;
326 		*dbbuf_db = value;
327 
328 		/*
329 		 * Ensure that the doorbell is updated before reading the event
330 		 * index from memory.  The controller needs to provide similar
331 		 * ordering to ensure the envent index is updated before reading
332 		 * the doorbell.
333 		 */
334 		mb();
335 
336 		if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
337 			return false;
338 	}
339 
340 	return true;
341 }
342 
343 /*
344  * Will slightly overestimate the number of pages needed.  This is OK
345  * as it only leads to a small amount of wasted memory for the lifetime of
346  * the I/O.
347  */
348 static int nvme_npages(unsigned size, struct nvme_dev *dev)
349 {
350 	unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
351 				      dev->ctrl.page_size);
352 	return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
353 }
354 
355 /*
356  * Calculates the number of pages needed for the SGL segments. For example a 4k
357  * page can accommodate 256 SGL descriptors.
358  */
359 static int nvme_pci_npages_sgl(unsigned int num_seg)
360 {
361 	return DIV_ROUND_UP(num_seg * sizeof(struct nvme_sgl_desc), PAGE_SIZE);
362 }
363 
364 static unsigned int nvme_pci_iod_alloc_size(struct nvme_dev *dev,
365 		unsigned int size, unsigned int nseg, bool use_sgl)
366 {
367 	size_t alloc_size;
368 
369 	if (use_sgl)
370 		alloc_size = sizeof(__le64 *) * nvme_pci_npages_sgl(nseg);
371 	else
372 		alloc_size = sizeof(__le64 *) * nvme_npages(size, dev);
373 
374 	return alloc_size + sizeof(struct scatterlist) * nseg;
375 }
376 
377 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
378 				unsigned int hctx_idx)
379 {
380 	struct nvme_dev *dev = data;
381 	struct nvme_queue *nvmeq = &dev->queues[0];
382 
383 	WARN_ON(hctx_idx != 0);
384 	WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
385 
386 	hctx->driver_data = nvmeq;
387 	return 0;
388 }
389 
390 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
391 			  unsigned int hctx_idx)
392 {
393 	struct nvme_dev *dev = data;
394 	struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
395 
396 	WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
397 	hctx->driver_data = nvmeq;
398 	return 0;
399 }
400 
401 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
402 		unsigned int hctx_idx, unsigned int numa_node)
403 {
404 	struct nvme_dev *dev = set->driver_data;
405 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
406 	int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
407 	struct nvme_queue *nvmeq = &dev->queues[queue_idx];
408 
409 	BUG_ON(!nvmeq);
410 	iod->nvmeq = nvmeq;
411 
412 	nvme_req(req)->ctrl = &dev->ctrl;
413 	return 0;
414 }
415 
416 static int queue_irq_offset(struct nvme_dev *dev)
417 {
418 	/* if we have more than 1 vec, admin queue offsets us by 1 */
419 	if (dev->num_vecs > 1)
420 		return 1;
421 
422 	return 0;
423 }
424 
425 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
426 {
427 	struct nvme_dev *dev = set->driver_data;
428 	int i, qoff, offset;
429 
430 	offset = queue_irq_offset(dev);
431 	for (i = 0, qoff = 0; i < set->nr_maps; i++) {
432 		struct blk_mq_queue_map *map = &set->map[i];
433 
434 		map->nr_queues = dev->io_queues[i];
435 		if (!map->nr_queues) {
436 			BUG_ON(i == HCTX_TYPE_DEFAULT);
437 			continue;
438 		}
439 
440 		/*
441 		 * The poll queue(s) doesn't have an IRQ (and hence IRQ
442 		 * affinity), so use the regular blk-mq cpu mapping
443 		 */
444 		map->queue_offset = qoff;
445 		if (i != HCTX_TYPE_POLL && offset)
446 			blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
447 		else
448 			blk_mq_map_queues(map);
449 		qoff += map->nr_queues;
450 		offset += map->nr_queues;
451 	}
452 
453 	return 0;
454 }
455 
456 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq)
457 {
458 	if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
459 			nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
460 		writel(nvmeq->sq_tail, nvmeq->q_db);
461 }
462 
463 /**
464  * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
465  * @nvmeq: The queue to use
466  * @cmd: The command to send
467  * @write_sq: whether to write to the SQ doorbell
468  */
469 static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
470 			    bool write_sq)
471 {
472 	spin_lock(&nvmeq->sq_lock);
473 	memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
474 	       cmd, sizeof(*cmd));
475 	if (++nvmeq->sq_tail == nvmeq->q_depth)
476 		nvmeq->sq_tail = 0;
477 	if (write_sq)
478 		nvme_write_sq_db(nvmeq);
479 	spin_unlock(&nvmeq->sq_lock);
480 }
481 
482 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
483 {
484 	struct nvme_queue *nvmeq = hctx->driver_data;
485 
486 	spin_lock(&nvmeq->sq_lock);
487 	nvme_write_sq_db(nvmeq);
488 	spin_unlock(&nvmeq->sq_lock);
489 }
490 
491 static void **nvme_pci_iod_list(struct request *req)
492 {
493 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
494 	return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
495 }
496 
497 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req)
498 {
499 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
500 	int nseg = blk_rq_nr_phys_segments(req);
501 	unsigned int avg_seg_size;
502 
503 	if (nseg == 0)
504 		return false;
505 
506 	avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
507 
508 	if (!(dev->ctrl.sgls & ((1 << 0) | (1 << 1))))
509 		return false;
510 	if (!iod->nvmeq->qid)
511 		return false;
512 	if (!sgl_threshold || avg_seg_size < sgl_threshold)
513 		return false;
514 	return true;
515 }
516 
517 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
518 {
519 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
520 	const int last_prp = dev->ctrl.page_size / sizeof(__le64) - 1;
521 	dma_addr_t dma_addr = iod->first_dma, next_dma_addr;
522 	int i;
523 
524 	if (iod->dma_len) {
525 		dma_unmap_page(dev->dev, dma_addr, iod->dma_len,
526 			       rq_dma_dir(req));
527 		return;
528 	}
529 
530 	WARN_ON_ONCE(!iod->nents);
531 
532 	if (is_pci_p2pdma_page(sg_page(iod->sg)))
533 		pci_p2pdma_unmap_sg(dev->dev, iod->sg, iod->nents,
534 				    rq_dma_dir(req));
535 	else
536 		dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req));
537 
538 
539 	if (iod->npages == 0)
540 		dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
541 			dma_addr);
542 
543 	for (i = 0; i < iod->npages; i++) {
544 		void *addr = nvme_pci_iod_list(req)[i];
545 
546 		if (iod->use_sgl) {
547 			struct nvme_sgl_desc *sg_list = addr;
548 
549 			next_dma_addr =
550 			    le64_to_cpu((sg_list[SGES_PER_PAGE - 1]).addr);
551 		} else {
552 			__le64 *prp_list = addr;
553 
554 			next_dma_addr = le64_to_cpu(prp_list[last_prp]);
555 		}
556 
557 		dma_pool_free(dev->prp_page_pool, addr, dma_addr);
558 		dma_addr = next_dma_addr;
559 	}
560 
561 	mempool_free(iod->sg, dev->iod_mempool);
562 }
563 
564 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
565 {
566 	int i;
567 	struct scatterlist *sg;
568 
569 	for_each_sg(sgl, sg, nents, i) {
570 		dma_addr_t phys = sg_phys(sg);
571 		pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
572 			"dma_address:%pad dma_length:%d\n",
573 			i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
574 			sg_dma_len(sg));
575 	}
576 }
577 
578 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
579 		struct request *req, struct nvme_rw_command *cmnd)
580 {
581 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
582 	struct dma_pool *pool;
583 	int length = blk_rq_payload_bytes(req);
584 	struct scatterlist *sg = iod->sg;
585 	int dma_len = sg_dma_len(sg);
586 	u64 dma_addr = sg_dma_address(sg);
587 	u32 page_size = dev->ctrl.page_size;
588 	int offset = dma_addr & (page_size - 1);
589 	__le64 *prp_list;
590 	void **list = nvme_pci_iod_list(req);
591 	dma_addr_t prp_dma;
592 	int nprps, i;
593 
594 	length -= (page_size - offset);
595 	if (length <= 0) {
596 		iod->first_dma = 0;
597 		goto done;
598 	}
599 
600 	dma_len -= (page_size - offset);
601 	if (dma_len) {
602 		dma_addr += (page_size - offset);
603 	} else {
604 		sg = sg_next(sg);
605 		dma_addr = sg_dma_address(sg);
606 		dma_len = sg_dma_len(sg);
607 	}
608 
609 	if (length <= page_size) {
610 		iod->first_dma = dma_addr;
611 		goto done;
612 	}
613 
614 	nprps = DIV_ROUND_UP(length, page_size);
615 	if (nprps <= (256 / 8)) {
616 		pool = dev->prp_small_pool;
617 		iod->npages = 0;
618 	} else {
619 		pool = dev->prp_page_pool;
620 		iod->npages = 1;
621 	}
622 
623 	prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
624 	if (!prp_list) {
625 		iod->first_dma = dma_addr;
626 		iod->npages = -1;
627 		return BLK_STS_RESOURCE;
628 	}
629 	list[0] = prp_list;
630 	iod->first_dma = prp_dma;
631 	i = 0;
632 	for (;;) {
633 		if (i == page_size >> 3) {
634 			__le64 *old_prp_list = prp_list;
635 			prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
636 			if (!prp_list)
637 				return BLK_STS_RESOURCE;
638 			list[iod->npages++] = prp_list;
639 			prp_list[0] = old_prp_list[i - 1];
640 			old_prp_list[i - 1] = cpu_to_le64(prp_dma);
641 			i = 1;
642 		}
643 		prp_list[i++] = cpu_to_le64(dma_addr);
644 		dma_len -= page_size;
645 		dma_addr += page_size;
646 		length -= page_size;
647 		if (length <= 0)
648 			break;
649 		if (dma_len > 0)
650 			continue;
651 		if (unlikely(dma_len < 0))
652 			goto bad_sgl;
653 		sg = sg_next(sg);
654 		dma_addr = sg_dma_address(sg);
655 		dma_len = sg_dma_len(sg);
656 	}
657 
658 done:
659 	cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
660 	cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
661 
662 	return BLK_STS_OK;
663 
664  bad_sgl:
665 	WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
666 			"Invalid SGL for payload:%d nents:%d\n",
667 			blk_rq_payload_bytes(req), iod->nents);
668 	return BLK_STS_IOERR;
669 }
670 
671 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
672 		struct scatterlist *sg)
673 {
674 	sge->addr = cpu_to_le64(sg_dma_address(sg));
675 	sge->length = cpu_to_le32(sg_dma_len(sg));
676 	sge->type = NVME_SGL_FMT_DATA_DESC << 4;
677 }
678 
679 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
680 		dma_addr_t dma_addr, int entries)
681 {
682 	sge->addr = cpu_to_le64(dma_addr);
683 	if (entries < SGES_PER_PAGE) {
684 		sge->length = cpu_to_le32(entries * sizeof(*sge));
685 		sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
686 	} else {
687 		sge->length = cpu_to_le32(PAGE_SIZE);
688 		sge->type = NVME_SGL_FMT_SEG_DESC << 4;
689 	}
690 }
691 
692 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
693 		struct request *req, struct nvme_rw_command *cmd, int entries)
694 {
695 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
696 	struct dma_pool *pool;
697 	struct nvme_sgl_desc *sg_list;
698 	struct scatterlist *sg = iod->sg;
699 	dma_addr_t sgl_dma;
700 	int i = 0;
701 
702 	/* setting the transfer type as SGL */
703 	cmd->flags = NVME_CMD_SGL_METABUF;
704 
705 	if (entries == 1) {
706 		nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
707 		return BLK_STS_OK;
708 	}
709 
710 	if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
711 		pool = dev->prp_small_pool;
712 		iod->npages = 0;
713 	} else {
714 		pool = dev->prp_page_pool;
715 		iod->npages = 1;
716 	}
717 
718 	sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
719 	if (!sg_list) {
720 		iod->npages = -1;
721 		return BLK_STS_RESOURCE;
722 	}
723 
724 	nvme_pci_iod_list(req)[0] = sg_list;
725 	iod->first_dma = sgl_dma;
726 
727 	nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
728 
729 	do {
730 		if (i == SGES_PER_PAGE) {
731 			struct nvme_sgl_desc *old_sg_desc = sg_list;
732 			struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
733 
734 			sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
735 			if (!sg_list)
736 				return BLK_STS_RESOURCE;
737 
738 			i = 0;
739 			nvme_pci_iod_list(req)[iod->npages++] = sg_list;
740 			sg_list[i++] = *link;
741 			nvme_pci_sgl_set_seg(link, sgl_dma, entries);
742 		}
743 
744 		nvme_pci_sgl_set_data(&sg_list[i++], sg);
745 		sg = sg_next(sg);
746 	} while (--entries > 0);
747 
748 	return BLK_STS_OK;
749 }
750 
751 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
752 		struct request *req, struct nvme_rw_command *cmnd,
753 		struct bio_vec *bv)
754 {
755 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
756 	unsigned int offset = bv->bv_offset & (dev->ctrl.page_size - 1);
757 	unsigned int first_prp_len = dev->ctrl.page_size - offset;
758 
759 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
760 	if (dma_mapping_error(dev->dev, iod->first_dma))
761 		return BLK_STS_RESOURCE;
762 	iod->dma_len = bv->bv_len;
763 
764 	cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
765 	if (bv->bv_len > first_prp_len)
766 		cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
767 	return 0;
768 }
769 
770 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
771 		struct request *req, struct nvme_rw_command *cmnd,
772 		struct bio_vec *bv)
773 {
774 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
775 
776 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
777 	if (dma_mapping_error(dev->dev, iod->first_dma))
778 		return BLK_STS_RESOURCE;
779 	iod->dma_len = bv->bv_len;
780 
781 	cmnd->flags = NVME_CMD_SGL_METABUF;
782 	cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
783 	cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
784 	cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
785 	return 0;
786 }
787 
788 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
789 		struct nvme_command *cmnd)
790 {
791 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
792 	blk_status_t ret = BLK_STS_RESOURCE;
793 	int nr_mapped;
794 
795 	if (blk_rq_nr_phys_segments(req) == 1) {
796 		struct bio_vec bv = req_bvec(req);
797 
798 		if (!is_pci_p2pdma_page(bv.bv_page)) {
799 			if (bv.bv_offset + bv.bv_len <= dev->ctrl.page_size * 2)
800 				return nvme_setup_prp_simple(dev, req,
801 							     &cmnd->rw, &bv);
802 
803 			if (iod->nvmeq->qid &&
804 			    dev->ctrl.sgls & ((1 << 0) | (1 << 1)))
805 				return nvme_setup_sgl_simple(dev, req,
806 							     &cmnd->rw, &bv);
807 		}
808 	}
809 
810 	iod->dma_len = 0;
811 	iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
812 	if (!iod->sg)
813 		return BLK_STS_RESOURCE;
814 	sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
815 	iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
816 	if (!iod->nents)
817 		goto out;
818 
819 	if (is_pci_p2pdma_page(sg_page(iod->sg)))
820 		nr_mapped = pci_p2pdma_map_sg_attrs(dev->dev, iod->sg,
821 				iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
822 	else
823 		nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
824 					     rq_dma_dir(req), DMA_ATTR_NO_WARN);
825 	if (!nr_mapped)
826 		goto out;
827 
828 	iod->use_sgl = nvme_pci_use_sgls(dev, req);
829 	if (iod->use_sgl)
830 		ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
831 	else
832 		ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
833 out:
834 	if (ret != BLK_STS_OK)
835 		nvme_unmap_data(dev, req);
836 	return ret;
837 }
838 
839 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
840 		struct nvme_command *cmnd)
841 {
842 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
843 
844 	iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
845 			rq_dma_dir(req), 0);
846 	if (dma_mapping_error(dev->dev, iod->meta_dma))
847 		return BLK_STS_IOERR;
848 	cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
849 	return 0;
850 }
851 
852 /*
853  * NOTE: ns is NULL when called on the admin queue.
854  */
855 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
856 			 const struct blk_mq_queue_data *bd)
857 {
858 	struct nvme_ns *ns = hctx->queue->queuedata;
859 	struct nvme_queue *nvmeq = hctx->driver_data;
860 	struct nvme_dev *dev = nvmeq->dev;
861 	struct request *req = bd->rq;
862 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
863 	struct nvme_command cmnd;
864 	blk_status_t ret;
865 
866 	iod->aborted = 0;
867 	iod->npages = -1;
868 	iod->nents = 0;
869 
870 	/*
871 	 * We should not need to do this, but we're still using this to
872 	 * ensure we can drain requests on a dying queue.
873 	 */
874 	if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
875 		return BLK_STS_IOERR;
876 
877 	ret = nvme_setup_cmd(ns, req, &cmnd);
878 	if (ret)
879 		return ret;
880 
881 	if (blk_rq_nr_phys_segments(req)) {
882 		ret = nvme_map_data(dev, req, &cmnd);
883 		if (ret)
884 			goto out_free_cmd;
885 	}
886 
887 	if (blk_integrity_rq(req)) {
888 		ret = nvme_map_metadata(dev, req, &cmnd);
889 		if (ret)
890 			goto out_unmap_data;
891 	}
892 
893 	blk_mq_start_request(req);
894 	nvme_submit_cmd(nvmeq, &cmnd, bd->last);
895 	return BLK_STS_OK;
896 out_unmap_data:
897 	nvme_unmap_data(dev, req);
898 out_free_cmd:
899 	nvme_cleanup_cmd(req);
900 	return ret;
901 }
902 
903 static void nvme_pci_complete_rq(struct request *req)
904 {
905 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
906 	struct nvme_dev *dev = iod->nvmeq->dev;
907 
908 	if (blk_integrity_rq(req))
909 		dma_unmap_page(dev->dev, iod->meta_dma,
910 			       rq_integrity_vec(req)->bv_len, rq_data_dir(req));
911 	if (blk_rq_nr_phys_segments(req))
912 		nvme_unmap_data(dev, req);
913 	nvme_complete_rq(req);
914 }
915 
916 /* We read the CQE phase first to check if the rest of the entry is valid */
917 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
918 {
919 	struct nvme_completion *hcqe = &nvmeq->cqes[nvmeq->cq_head];
920 
921 	return (le16_to_cpu(READ_ONCE(hcqe->status)) & 1) == nvmeq->cq_phase;
922 }
923 
924 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
925 {
926 	u16 head = nvmeq->cq_head;
927 
928 	if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
929 					      nvmeq->dbbuf_cq_ei))
930 		writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
931 }
932 
933 static inline struct blk_mq_tags *nvme_queue_tagset(struct nvme_queue *nvmeq)
934 {
935 	if (!nvmeq->qid)
936 		return nvmeq->dev->admin_tagset.tags[0];
937 	return nvmeq->dev->tagset.tags[nvmeq->qid - 1];
938 }
939 
940 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
941 {
942 	struct nvme_completion *cqe = &nvmeq->cqes[idx];
943 	struct request *req;
944 
945 	if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
946 		dev_warn(nvmeq->dev->ctrl.device,
947 			"invalid id %d completed on queue %d\n",
948 			cqe->command_id, le16_to_cpu(cqe->sq_id));
949 		return;
950 	}
951 
952 	/*
953 	 * AEN requests are special as they don't time out and can
954 	 * survive any kind of queue freeze and often don't respond to
955 	 * aborts.  We don't even bother to allocate a struct request
956 	 * for them but rather special case them here.
957 	 */
958 	if (unlikely(nvme_is_aen_req(nvmeq->qid, cqe->command_id))) {
959 		nvme_complete_async_event(&nvmeq->dev->ctrl,
960 				cqe->status, &cqe->result);
961 		return;
962 	}
963 
964 	req = blk_mq_tag_to_rq(nvme_queue_tagset(nvmeq), cqe->command_id);
965 	trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
966 	nvme_end_request(req, cqe->status, cqe->result);
967 }
968 
969 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
970 {
971 	u16 tmp = nvmeq->cq_head + 1;
972 
973 	if (tmp == nvmeq->q_depth) {
974 		nvmeq->cq_head = 0;
975 		nvmeq->cq_phase ^= 1;
976 	} else {
977 		nvmeq->cq_head = tmp;
978 	}
979 }
980 
981 static inline int nvme_process_cq(struct nvme_queue *nvmeq)
982 {
983 	int found = 0;
984 
985 	while (nvme_cqe_pending(nvmeq)) {
986 		found++;
987 		/*
988 		 * load-load control dependency between phase and the rest of
989 		 * the cqe requires a full read memory barrier
990 		 */
991 		dma_rmb();
992 		nvme_handle_cqe(nvmeq, nvmeq->cq_head);
993 		nvme_update_cq_head(nvmeq);
994 	}
995 
996 	if (found)
997 		nvme_ring_cq_doorbell(nvmeq);
998 	return found;
999 }
1000 
1001 static irqreturn_t nvme_irq(int irq, void *data)
1002 {
1003 	struct nvme_queue *nvmeq = data;
1004 	irqreturn_t ret = IRQ_NONE;
1005 
1006 	/*
1007 	 * The rmb/wmb pair ensures we see all updates from a previous run of
1008 	 * the irq handler, even if that was on another CPU.
1009 	 */
1010 	rmb();
1011 	if (nvme_process_cq(nvmeq))
1012 		ret = IRQ_HANDLED;
1013 	wmb();
1014 
1015 	return ret;
1016 }
1017 
1018 static irqreturn_t nvme_irq_check(int irq, void *data)
1019 {
1020 	struct nvme_queue *nvmeq = data;
1021 	if (nvme_cqe_pending(nvmeq))
1022 		return IRQ_WAKE_THREAD;
1023 	return IRQ_NONE;
1024 }
1025 
1026 /*
1027  * Poll for completions for any interrupt driven queue
1028  * Can be called from any context.
1029  */
1030 static void nvme_poll_irqdisable(struct nvme_queue *nvmeq)
1031 {
1032 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1033 
1034 	WARN_ON_ONCE(test_bit(NVMEQ_POLLED, &nvmeq->flags));
1035 
1036 	disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1037 	nvme_process_cq(nvmeq);
1038 	enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1039 }
1040 
1041 static int nvme_poll(struct blk_mq_hw_ctx *hctx)
1042 {
1043 	struct nvme_queue *nvmeq = hctx->driver_data;
1044 	bool found;
1045 
1046 	if (!nvme_cqe_pending(nvmeq))
1047 		return 0;
1048 
1049 	spin_lock(&nvmeq->cq_poll_lock);
1050 	found = nvme_process_cq(nvmeq);
1051 	spin_unlock(&nvmeq->cq_poll_lock);
1052 
1053 	return found;
1054 }
1055 
1056 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1057 {
1058 	struct nvme_dev *dev = to_nvme_dev(ctrl);
1059 	struct nvme_queue *nvmeq = &dev->queues[0];
1060 	struct nvme_command c;
1061 
1062 	memset(&c, 0, sizeof(c));
1063 	c.common.opcode = nvme_admin_async_event;
1064 	c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1065 	nvme_submit_cmd(nvmeq, &c, true);
1066 }
1067 
1068 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1069 {
1070 	struct nvme_command c;
1071 
1072 	memset(&c, 0, sizeof(c));
1073 	c.delete_queue.opcode = opcode;
1074 	c.delete_queue.qid = cpu_to_le16(id);
1075 
1076 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1077 }
1078 
1079 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1080 		struct nvme_queue *nvmeq, s16 vector)
1081 {
1082 	struct nvme_command c;
1083 	int flags = NVME_QUEUE_PHYS_CONTIG;
1084 
1085 	if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1086 		flags |= NVME_CQ_IRQ_ENABLED;
1087 
1088 	/*
1089 	 * Note: we (ab)use the fact that the prp fields survive if no data
1090 	 * is attached to the request.
1091 	 */
1092 	memset(&c, 0, sizeof(c));
1093 	c.create_cq.opcode = nvme_admin_create_cq;
1094 	c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1095 	c.create_cq.cqid = cpu_to_le16(qid);
1096 	c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1097 	c.create_cq.cq_flags = cpu_to_le16(flags);
1098 	c.create_cq.irq_vector = cpu_to_le16(vector);
1099 
1100 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1101 }
1102 
1103 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1104 						struct nvme_queue *nvmeq)
1105 {
1106 	struct nvme_ctrl *ctrl = &dev->ctrl;
1107 	struct nvme_command c;
1108 	int flags = NVME_QUEUE_PHYS_CONTIG;
1109 
1110 	/*
1111 	 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1112 	 * set. Since URGENT priority is zeroes, it makes all queues
1113 	 * URGENT.
1114 	 */
1115 	if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1116 		flags |= NVME_SQ_PRIO_MEDIUM;
1117 
1118 	/*
1119 	 * Note: we (ab)use the fact that the prp fields survive if no data
1120 	 * is attached to the request.
1121 	 */
1122 	memset(&c, 0, sizeof(c));
1123 	c.create_sq.opcode = nvme_admin_create_sq;
1124 	c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1125 	c.create_sq.sqid = cpu_to_le16(qid);
1126 	c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1127 	c.create_sq.sq_flags = cpu_to_le16(flags);
1128 	c.create_sq.cqid = cpu_to_le16(qid);
1129 
1130 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1131 }
1132 
1133 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1134 {
1135 	return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1136 }
1137 
1138 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1139 {
1140 	return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1141 }
1142 
1143 static void abort_endio(struct request *req, blk_status_t error)
1144 {
1145 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1146 	struct nvme_queue *nvmeq = iod->nvmeq;
1147 
1148 	dev_warn(nvmeq->dev->ctrl.device,
1149 		 "Abort status: 0x%x", nvme_req(req)->status);
1150 	atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1151 	blk_mq_free_request(req);
1152 }
1153 
1154 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1155 {
1156 
1157 	/* If true, indicates loss of adapter communication, possibly by a
1158 	 * NVMe Subsystem reset.
1159 	 */
1160 	bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1161 
1162 	/* If there is a reset/reinit ongoing, we shouldn't reset again. */
1163 	switch (dev->ctrl.state) {
1164 	case NVME_CTRL_RESETTING:
1165 	case NVME_CTRL_CONNECTING:
1166 		return false;
1167 	default:
1168 		break;
1169 	}
1170 
1171 	/* We shouldn't reset unless the controller is on fatal error state
1172 	 * _or_ if we lost the communication with it.
1173 	 */
1174 	if (!(csts & NVME_CSTS_CFS) && !nssro)
1175 		return false;
1176 
1177 	return true;
1178 }
1179 
1180 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1181 {
1182 	/* Read a config register to help see what died. */
1183 	u16 pci_status;
1184 	int result;
1185 
1186 	result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1187 				      &pci_status);
1188 	if (result == PCIBIOS_SUCCESSFUL)
1189 		dev_warn(dev->ctrl.device,
1190 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1191 			 csts, pci_status);
1192 	else
1193 		dev_warn(dev->ctrl.device,
1194 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1195 			 csts, result);
1196 }
1197 
1198 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1199 {
1200 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1201 	struct nvme_queue *nvmeq = iod->nvmeq;
1202 	struct nvme_dev *dev = nvmeq->dev;
1203 	struct request *abort_req;
1204 	struct nvme_command cmd;
1205 	u32 csts = readl(dev->bar + NVME_REG_CSTS);
1206 
1207 	/* If PCI error recovery process is happening, we cannot reset or
1208 	 * the recovery mechanism will surely fail.
1209 	 */
1210 	mb();
1211 	if (pci_channel_offline(to_pci_dev(dev->dev)))
1212 		return BLK_EH_RESET_TIMER;
1213 
1214 	/*
1215 	 * Reset immediately if the controller is failed
1216 	 */
1217 	if (nvme_should_reset(dev, csts)) {
1218 		nvme_warn_reset(dev, csts);
1219 		nvme_dev_disable(dev, false);
1220 		nvme_reset_ctrl(&dev->ctrl);
1221 		return BLK_EH_DONE;
1222 	}
1223 
1224 	/*
1225 	 * Did we miss an interrupt?
1226 	 */
1227 	if (test_bit(NVMEQ_POLLED, &nvmeq->flags))
1228 		nvme_poll(req->mq_hctx);
1229 	else
1230 		nvme_poll_irqdisable(nvmeq);
1231 
1232 	if (blk_mq_request_completed(req)) {
1233 		dev_warn(dev->ctrl.device,
1234 			 "I/O %d QID %d timeout, completion polled\n",
1235 			 req->tag, nvmeq->qid);
1236 		return BLK_EH_DONE;
1237 	}
1238 
1239 	/*
1240 	 * Shutdown immediately if controller times out while starting. The
1241 	 * reset work will see the pci device disabled when it gets the forced
1242 	 * cancellation error. All outstanding requests are completed on
1243 	 * shutdown, so we return BLK_EH_DONE.
1244 	 */
1245 	switch (dev->ctrl.state) {
1246 	case NVME_CTRL_CONNECTING:
1247 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1248 		/* fall through */
1249 	case NVME_CTRL_DELETING:
1250 		dev_warn_ratelimited(dev->ctrl.device,
1251 			 "I/O %d QID %d timeout, disable controller\n",
1252 			 req->tag, nvmeq->qid);
1253 		nvme_dev_disable(dev, true);
1254 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1255 		return BLK_EH_DONE;
1256 	case NVME_CTRL_RESETTING:
1257 		return BLK_EH_RESET_TIMER;
1258 	default:
1259 		break;
1260 	}
1261 
1262 	/*
1263  	 * Shutdown the controller immediately and schedule a reset if the
1264  	 * command was already aborted once before and still hasn't been
1265  	 * returned to the driver, or if this is the admin queue.
1266 	 */
1267 	if (!nvmeq->qid || iod->aborted) {
1268 		dev_warn(dev->ctrl.device,
1269 			 "I/O %d QID %d timeout, reset controller\n",
1270 			 req->tag, nvmeq->qid);
1271 		nvme_dev_disable(dev, false);
1272 		nvme_reset_ctrl(&dev->ctrl);
1273 
1274 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1275 		return BLK_EH_DONE;
1276 	}
1277 
1278 	if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1279 		atomic_inc(&dev->ctrl.abort_limit);
1280 		return BLK_EH_RESET_TIMER;
1281 	}
1282 	iod->aborted = 1;
1283 
1284 	memset(&cmd, 0, sizeof(cmd));
1285 	cmd.abort.opcode = nvme_admin_abort_cmd;
1286 	cmd.abort.cid = req->tag;
1287 	cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1288 
1289 	dev_warn(nvmeq->dev->ctrl.device,
1290 		"I/O %d QID %d timeout, aborting\n",
1291 		 req->tag, nvmeq->qid);
1292 
1293 	abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1294 			BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1295 	if (IS_ERR(abort_req)) {
1296 		atomic_inc(&dev->ctrl.abort_limit);
1297 		return BLK_EH_RESET_TIMER;
1298 	}
1299 
1300 	abort_req->timeout = ADMIN_TIMEOUT;
1301 	abort_req->end_io_data = NULL;
1302 	blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1303 
1304 	/*
1305 	 * The aborted req will be completed on receiving the abort req.
1306 	 * We enable the timer again. If hit twice, it'll cause a device reset,
1307 	 * as the device then is in a faulty state.
1308 	 */
1309 	return BLK_EH_RESET_TIMER;
1310 }
1311 
1312 static void nvme_free_queue(struct nvme_queue *nvmeq)
1313 {
1314 	dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1315 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1316 	if (!nvmeq->sq_cmds)
1317 		return;
1318 
1319 	if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1320 		pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1321 				nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1322 	} else {
1323 		dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1324 				nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1325 	}
1326 }
1327 
1328 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1329 {
1330 	int i;
1331 
1332 	for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1333 		dev->ctrl.queue_count--;
1334 		nvme_free_queue(&dev->queues[i]);
1335 	}
1336 }
1337 
1338 /**
1339  * nvme_suspend_queue - put queue into suspended state
1340  * @nvmeq: queue to suspend
1341  */
1342 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1343 {
1344 	if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1345 		return 1;
1346 
1347 	/* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1348 	mb();
1349 
1350 	nvmeq->dev->online_queues--;
1351 	if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1352 		blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1353 	if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1354 		pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1355 	return 0;
1356 }
1357 
1358 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1359 {
1360 	int i;
1361 
1362 	for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1363 		nvme_suspend_queue(&dev->queues[i]);
1364 }
1365 
1366 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1367 {
1368 	struct nvme_queue *nvmeq = &dev->queues[0];
1369 
1370 	if (shutdown)
1371 		nvme_shutdown_ctrl(&dev->ctrl);
1372 	else
1373 		nvme_disable_ctrl(&dev->ctrl);
1374 
1375 	nvme_poll_irqdisable(nvmeq);
1376 }
1377 
1378 /*
1379  * Called only on a device that has been disabled and after all other threads
1380  * that can check this device's completion queues have synced, except
1381  * nvme_poll(). This is the last chance for the driver to see a natural
1382  * completion before nvme_cancel_request() terminates all incomplete requests.
1383  */
1384 static void nvme_reap_pending_cqes(struct nvme_dev *dev)
1385 {
1386 	int i;
1387 
1388 	for (i = dev->ctrl.queue_count - 1; i > 0; i--) {
1389 		spin_lock(&dev->queues[i].cq_poll_lock);
1390 		nvme_process_cq(&dev->queues[i]);
1391 		spin_unlock(&dev->queues[i].cq_poll_lock);
1392 	}
1393 }
1394 
1395 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1396 				int entry_size)
1397 {
1398 	int q_depth = dev->q_depth;
1399 	unsigned q_size_aligned = roundup(q_depth * entry_size,
1400 					  dev->ctrl.page_size);
1401 
1402 	if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1403 		u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1404 		mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1405 		q_depth = div_u64(mem_per_q, entry_size);
1406 
1407 		/*
1408 		 * Ensure the reduced q_depth is above some threshold where it
1409 		 * would be better to map queues in system memory with the
1410 		 * original depth
1411 		 */
1412 		if (q_depth < 64)
1413 			return -ENOMEM;
1414 	}
1415 
1416 	return q_depth;
1417 }
1418 
1419 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1420 				int qid)
1421 {
1422 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1423 
1424 	if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1425 		nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1426 		if (nvmeq->sq_cmds) {
1427 			nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1428 							nvmeq->sq_cmds);
1429 			if (nvmeq->sq_dma_addr) {
1430 				set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1431 				return 0;
1432 			}
1433 
1434 			pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1435 		}
1436 	}
1437 
1438 	nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1439 				&nvmeq->sq_dma_addr, GFP_KERNEL);
1440 	if (!nvmeq->sq_cmds)
1441 		return -ENOMEM;
1442 	return 0;
1443 }
1444 
1445 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1446 {
1447 	struct nvme_queue *nvmeq = &dev->queues[qid];
1448 
1449 	if (dev->ctrl.queue_count > qid)
1450 		return 0;
1451 
1452 	nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1453 	nvmeq->q_depth = depth;
1454 	nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1455 					 &nvmeq->cq_dma_addr, GFP_KERNEL);
1456 	if (!nvmeq->cqes)
1457 		goto free_nvmeq;
1458 
1459 	if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1460 		goto free_cqdma;
1461 
1462 	nvmeq->dev = dev;
1463 	spin_lock_init(&nvmeq->sq_lock);
1464 	spin_lock_init(&nvmeq->cq_poll_lock);
1465 	nvmeq->cq_head = 0;
1466 	nvmeq->cq_phase = 1;
1467 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1468 	nvmeq->qid = qid;
1469 	dev->ctrl.queue_count++;
1470 
1471 	return 0;
1472 
1473  free_cqdma:
1474 	dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1475 			  nvmeq->cq_dma_addr);
1476  free_nvmeq:
1477 	return -ENOMEM;
1478 }
1479 
1480 static int queue_request_irq(struct nvme_queue *nvmeq)
1481 {
1482 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1483 	int nr = nvmeq->dev->ctrl.instance;
1484 
1485 	if (use_threaded_interrupts) {
1486 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1487 				nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1488 	} else {
1489 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1490 				NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1491 	}
1492 }
1493 
1494 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1495 {
1496 	struct nvme_dev *dev = nvmeq->dev;
1497 
1498 	nvmeq->sq_tail = 0;
1499 	nvmeq->cq_head = 0;
1500 	nvmeq->cq_phase = 1;
1501 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1502 	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1503 	nvme_dbbuf_init(dev, nvmeq, qid);
1504 	dev->online_queues++;
1505 	wmb(); /* ensure the first interrupt sees the initialization */
1506 }
1507 
1508 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1509 {
1510 	struct nvme_dev *dev = nvmeq->dev;
1511 	int result;
1512 	u16 vector = 0;
1513 
1514 	clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1515 
1516 	/*
1517 	 * A queue's vector matches the queue identifier unless the controller
1518 	 * has only one vector available.
1519 	 */
1520 	if (!polled)
1521 		vector = dev->num_vecs == 1 ? 0 : qid;
1522 	else
1523 		set_bit(NVMEQ_POLLED, &nvmeq->flags);
1524 
1525 	result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1526 	if (result)
1527 		return result;
1528 
1529 	result = adapter_alloc_sq(dev, qid, nvmeq);
1530 	if (result < 0)
1531 		return result;
1532 	if (result)
1533 		goto release_cq;
1534 
1535 	nvmeq->cq_vector = vector;
1536 	nvme_init_queue(nvmeq, qid);
1537 
1538 	if (!polled) {
1539 		result = queue_request_irq(nvmeq);
1540 		if (result < 0)
1541 			goto release_sq;
1542 	}
1543 
1544 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1545 	return result;
1546 
1547 release_sq:
1548 	dev->online_queues--;
1549 	adapter_delete_sq(dev, qid);
1550 release_cq:
1551 	adapter_delete_cq(dev, qid);
1552 	return result;
1553 }
1554 
1555 static const struct blk_mq_ops nvme_mq_admin_ops = {
1556 	.queue_rq	= nvme_queue_rq,
1557 	.complete	= nvme_pci_complete_rq,
1558 	.init_hctx	= nvme_admin_init_hctx,
1559 	.init_request	= nvme_init_request,
1560 	.timeout	= nvme_timeout,
1561 };
1562 
1563 static const struct blk_mq_ops nvme_mq_ops = {
1564 	.queue_rq	= nvme_queue_rq,
1565 	.complete	= nvme_pci_complete_rq,
1566 	.commit_rqs	= nvme_commit_rqs,
1567 	.init_hctx	= nvme_init_hctx,
1568 	.init_request	= nvme_init_request,
1569 	.map_queues	= nvme_pci_map_queues,
1570 	.timeout	= nvme_timeout,
1571 	.poll		= nvme_poll,
1572 };
1573 
1574 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1575 {
1576 	if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1577 		/*
1578 		 * If the controller was reset during removal, it's possible
1579 		 * user requests may be waiting on a stopped queue. Start the
1580 		 * queue to flush these to completion.
1581 		 */
1582 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1583 		blk_cleanup_queue(dev->ctrl.admin_q);
1584 		blk_mq_free_tag_set(&dev->admin_tagset);
1585 	}
1586 }
1587 
1588 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1589 {
1590 	if (!dev->ctrl.admin_q) {
1591 		dev->admin_tagset.ops = &nvme_mq_admin_ops;
1592 		dev->admin_tagset.nr_hw_queues = 1;
1593 
1594 		dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1595 		dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1596 		dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1597 		dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1598 		dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1599 		dev->admin_tagset.driver_data = dev;
1600 
1601 		if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1602 			return -ENOMEM;
1603 		dev->ctrl.admin_tagset = &dev->admin_tagset;
1604 
1605 		dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1606 		if (IS_ERR(dev->ctrl.admin_q)) {
1607 			blk_mq_free_tag_set(&dev->admin_tagset);
1608 			return -ENOMEM;
1609 		}
1610 		if (!blk_get_queue(dev->ctrl.admin_q)) {
1611 			nvme_dev_remove_admin(dev);
1612 			dev->ctrl.admin_q = NULL;
1613 			return -ENODEV;
1614 		}
1615 	} else
1616 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1617 
1618 	return 0;
1619 }
1620 
1621 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1622 {
1623 	return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1624 }
1625 
1626 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1627 {
1628 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1629 
1630 	if (size <= dev->bar_mapped_size)
1631 		return 0;
1632 	if (size > pci_resource_len(pdev, 0))
1633 		return -ENOMEM;
1634 	if (dev->bar)
1635 		iounmap(dev->bar);
1636 	dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1637 	if (!dev->bar) {
1638 		dev->bar_mapped_size = 0;
1639 		return -ENOMEM;
1640 	}
1641 	dev->bar_mapped_size = size;
1642 	dev->dbs = dev->bar + NVME_REG_DBS;
1643 
1644 	return 0;
1645 }
1646 
1647 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1648 {
1649 	int result;
1650 	u32 aqa;
1651 	struct nvme_queue *nvmeq;
1652 
1653 	result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1654 	if (result < 0)
1655 		return result;
1656 
1657 	dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1658 				NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1659 
1660 	if (dev->subsystem &&
1661 	    (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1662 		writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1663 
1664 	result = nvme_disable_ctrl(&dev->ctrl);
1665 	if (result < 0)
1666 		return result;
1667 
1668 	result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1669 	if (result)
1670 		return result;
1671 
1672 	nvmeq = &dev->queues[0];
1673 	aqa = nvmeq->q_depth - 1;
1674 	aqa |= aqa << 16;
1675 
1676 	writel(aqa, dev->bar + NVME_REG_AQA);
1677 	lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1678 	lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1679 
1680 	result = nvme_enable_ctrl(&dev->ctrl);
1681 	if (result)
1682 		return result;
1683 
1684 	nvmeq->cq_vector = 0;
1685 	nvme_init_queue(nvmeq, 0);
1686 	result = queue_request_irq(nvmeq);
1687 	if (result) {
1688 		dev->online_queues--;
1689 		return result;
1690 	}
1691 
1692 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1693 	return result;
1694 }
1695 
1696 static int nvme_create_io_queues(struct nvme_dev *dev)
1697 {
1698 	unsigned i, max, rw_queues;
1699 	int ret = 0;
1700 
1701 	for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1702 		if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1703 			ret = -ENOMEM;
1704 			break;
1705 		}
1706 	}
1707 
1708 	max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1709 	if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1710 		rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1711 				dev->io_queues[HCTX_TYPE_READ];
1712 	} else {
1713 		rw_queues = max;
1714 	}
1715 
1716 	for (i = dev->online_queues; i <= max; i++) {
1717 		bool polled = i > rw_queues;
1718 
1719 		ret = nvme_create_queue(&dev->queues[i], i, polled);
1720 		if (ret)
1721 			break;
1722 	}
1723 
1724 	/*
1725 	 * Ignore failing Create SQ/CQ commands, we can continue with less
1726 	 * than the desired amount of queues, and even a controller without
1727 	 * I/O queues can still be used to issue admin commands.  This might
1728 	 * be useful to upgrade a buggy firmware for example.
1729 	 */
1730 	return ret >= 0 ? 0 : ret;
1731 }
1732 
1733 static ssize_t nvme_cmb_show(struct device *dev,
1734 			     struct device_attribute *attr,
1735 			     char *buf)
1736 {
1737 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1738 
1739 	return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz  : x%08x\n",
1740 		       ndev->cmbloc, ndev->cmbsz);
1741 }
1742 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1743 
1744 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1745 {
1746 	u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1747 
1748 	return 1ULL << (12 + 4 * szu);
1749 }
1750 
1751 static u32 nvme_cmb_size(struct nvme_dev *dev)
1752 {
1753 	return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1754 }
1755 
1756 static void nvme_map_cmb(struct nvme_dev *dev)
1757 {
1758 	u64 size, offset;
1759 	resource_size_t bar_size;
1760 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1761 	int bar;
1762 
1763 	if (dev->cmb_size)
1764 		return;
1765 
1766 	dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1767 	if (!dev->cmbsz)
1768 		return;
1769 	dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1770 
1771 	size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1772 	offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1773 	bar = NVME_CMB_BIR(dev->cmbloc);
1774 	bar_size = pci_resource_len(pdev, bar);
1775 
1776 	if (offset > bar_size)
1777 		return;
1778 
1779 	/*
1780 	 * Controllers may support a CMB size larger than their BAR,
1781 	 * for example, due to being behind a bridge. Reduce the CMB to
1782 	 * the reported size of the BAR
1783 	 */
1784 	if (size > bar_size - offset)
1785 		size = bar_size - offset;
1786 
1787 	if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1788 		dev_warn(dev->ctrl.device,
1789 			 "failed to register the CMB\n");
1790 		return;
1791 	}
1792 
1793 	dev->cmb_size = size;
1794 	dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1795 
1796 	if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1797 			(NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1798 		pci_p2pmem_publish(pdev, true);
1799 
1800 	if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1801 				    &dev_attr_cmb.attr, NULL))
1802 		dev_warn(dev->ctrl.device,
1803 			 "failed to add sysfs attribute for CMB\n");
1804 }
1805 
1806 static inline void nvme_release_cmb(struct nvme_dev *dev)
1807 {
1808 	if (dev->cmb_size) {
1809 		sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1810 					     &dev_attr_cmb.attr, NULL);
1811 		dev->cmb_size = 0;
1812 	}
1813 }
1814 
1815 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1816 {
1817 	u64 dma_addr = dev->host_mem_descs_dma;
1818 	struct nvme_command c;
1819 	int ret;
1820 
1821 	memset(&c, 0, sizeof(c));
1822 	c.features.opcode	= nvme_admin_set_features;
1823 	c.features.fid		= cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1824 	c.features.dword11	= cpu_to_le32(bits);
1825 	c.features.dword12	= cpu_to_le32(dev->host_mem_size >>
1826 					      ilog2(dev->ctrl.page_size));
1827 	c.features.dword13	= cpu_to_le32(lower_32_bits(dma_addr));
1828 	c.features.dword14	= cpu_to_le32(upper_32_bits(dma_addr));
1829 	c.features.dword15	= cpu_to_le32(dev->nr_host_mem_descs);
1830 
1831 	ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1832 	if (ret) {
1833 		dev_warn(dev->ctrl.device,
1834 			 "failed to set host mem (err %d, flags %#x).\n",
1835 			 ret, bits);
1836 	}
1837 	return ret;
1838 }
1839 
1840 static void nvme_free_host_mem(struct nvme_dev *dev)
1841 {
1842 	int i;
1843 
1844 	for (i = 0; i < dev->nr_host_mem_descs; i++) {
1845 		struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1846 		size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1847 
1848 		dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
1849 			       le64_to_cpu(desc->addr),
1850 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1851 	}
1852 
1853 	kfree(dev->host_mem_desc_bufs);
1854 	dev->host_mem_desc_bufs = NULL;
1855 	dma_free_coherent(dev->dev,
1856 			dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1857 			dev->host_mem_descs, dev->host_mem_descs_dma);
1858 	dev->host_mem_descs = NULL;
1859 	dev->nr_host_mem_descs = 0;
1860 }
1861 
1862 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1863 		u32 chunk_size)
1864 {
1865 	struct nvme_host_mem_buf_desc *descs;
1866 	u32 max_entries, len;
1867 	dma_addr_t descs_dma;
1868 	int i = 0;
1869 	void **bufs;
1870 	u64 size, tmp;
1871 
1872 	tmp = (preferred + chunk_size - 1);
1873 	do_div(tmp, chunk_size);
1874 	max_entries = tmp;
1875 
1876 	if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1877 		max_entries = dev->ctrl.hmmaxd;
1878 
1879 	descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
1880 				   &descs_dma, GFP_KERNEL);
1881 	if (!descs)
1882 		goto out;
1883 
1884 	bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1885 	if (!bufs)
1886 		goto out_free_descs;
1887 
1888 	for (size = 0; size < preferred && i < max_entries; size += len) {
1889 		dma_addr_t dma_addr;
1890 
1891 		len = min_t(u64, chunk_size, preferred - size);
1892 		bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1893 				DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1894 		if (!bufs[i])
1895 			break;
1896 
1897 		descs[i].addr = cpu_to_le64(dma_addr);
1898 		descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1899 		i++;
1900 	}
1901 
1902 	if (!size)
1903 		goto out_free_bufs;
1904 
1905 	dev->nr_host_mem_descs = i;
1906 	dev->host_mem_size = size;
1907 	dev->host_mem_descs = descs;
1908 	dev->host_mem_descs_dma = descs_dma;
1909 	dev->host_mem_desc_bufs = bufs;
1910 	return 0;
1911 
1912 out_free_bufs:
1913 	while (--i >= 0) {
1914 		size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1915 
1916 		dma_free_attrs(dev->dev, size, bufs[i],
1917 			       le64_to_cpu(descs[i].addr),
1918 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1919 	}
1920 
1921 	kfree(bufs);
1922 out_free_descs:
1923 	dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
1924 			descs_dma);
1925 out:
1926 	dev->host_mem_descs = NULL;
1927 	return -ENOMEM;
1928 }
1929 
1930 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1931 {
1932 	u32 chunk_size;
1933 
1934 	/* start big and work our way down */
1935 	for (chunk_size = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
1936 	     chunk_size >= max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
1937 	     chunk_size /= 2) {
1938 		if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
1939 			if (!min || dev->host_mem_size >= min)
1940 				return 0;
1941 			nvme_free_host_mem(dev);
1942 		}
1943 	}
1944 
1945 	return -ENOMEM;
1946 }
1947 
1948 static int nvme_setup_host_mem(struct nvme_dev *dev)
1949 {
1950 	u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1951 	u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1952 	u64 min = (u64)dev->ctrl.hmmin * 4096;
1953 	u32 enable_bits = NVME_HOST_MEM_ENABLE;
1954 	int ret;
1955 
1956 	preferred = min(preferred, max);
1957 	if (min > max) {
1958 		dev_warn(dev->ctrl.device,
1959 			"min host memory (%lld MiB) above limit (%d MiB).\n",
1960 			min >> ilog2(SZ_1M), max_host_mem_size_mb);
1961 		nvme_free_host_mem(dev);
1962 		return 0;
1963 	}
1964 
1965 	/*
1966 	 * If we already have a buffer allocated check if we can reuse it.
1967 	 */
1968 	if (dev->host_mem_descs) {
1969 		if (dev->host_mem_size >= min)
1970 			enable_bits |= NVME_HOST_MEM_RETURN;
1971 		else
1972 			nvme_free_host_mem(dev);
1973 	}
1974 
1975 	if (!dev->host_mem_descs) {
1976 		if (nvme_alloc_host_mem(dev, min, preferred)) {
1977 			dev_warn(dev->ctrl.device,
1978 				"failed to allocate host memory buffer.\n");
1979 			return 0; /* controller must work without HMB */
1980 		}
1981 
1982 		dev_info(dev->ctrl.device,
1983 			"allocated %lld MiB host memory buffer.\n",
1984 			dev->host_mem_size >> ilog2(SZ_1M));
1985 	}
1986 
1987 	ret = nvme_set_host_mem(dev, enable_bits);
1988 	if (ret)
1989 		nvme_free_host_mem(dev);
1990 	return ret;
1991 }
1992 
1993 /*
1994  * nirqs is the number of interrupts available for write and read
1995  * queues. The core already reserved an interrupt for the admin queue.
1996  */
1997 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
1998 {
1999 	struct nvme_dev *dev = affd->priv;
2000 	unsigned int nr_read_queues, nr_write_queues = dev->nr_write_queues;
2001 
2002 	/*
2003 	 * If there is no interupt available for queues, ensure that
2004 	 * the default queue is set to 1. The affinity set size is
2005 	 * also set to one, but the irq core ignores it for this case.
2006 	 *
2007 	 * If only one interrupt is available or 'write_queue' == 0, combine
2008 	 * write and read queues.
2009 	 *
2010 	 * If 'write_queues' > 0, ensure it leaves room for at least one read
2011 	 * queue.
2012 	 */
2013 	if (!nrirqs) {
2014 		nrirqs = 1;
2015 		nr_read_queues = 0;
2016 	} else if (nrirqs == 1 || !nr_write_queues) {
2017 		nr_read_queues = 0;
2018 	} else if (nr_write_queues >= nrirqs) {
2019 		nr_read_queues = 1;
2020 	} else {
2021 		nr_read_queues = nrirqs - nr_write_queues;
2022 	}
2023 
2024 	dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2025 	affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2026 	dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2027 	affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2028 	affd->nr_sets = nr_read_queues ? 2 : 1;
2029 }
2030 
2031 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2032 {
2033 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2034 	struct irq_affinity affd = {
2035 		.pre_vectors	= 1,
2036 		.calc_sets	= nvme_calc_irq_sets,
2037 		.priv		= dev,
2038 	};
2039 	unsigned int irq_queues, this_p_queues;
2040 
2041 	/*
2042 	 * Poll queues don't need interrupts, but we need at least one IO
2043 	 * queue left over for non-polled IO.
2044 	 */
2045 	this_p_queues = dev->nr_poll_queues;
2046 	if (this_p_queues >= nr_io_queues) {
2047 		this_p_queues = nr_io_queues - 1;
2048 		irq_queues = 1;
2049 	} else {
2050 		irq_queues = nr_io_queues - this_p_queues + 1;
2051 	}
2052 	dev->io_queues[HCTX_TYPE_POLL] = this_p_queues;
2053 
2054 	/* Initialize for the single interrupt case */
2055 	dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2056 	dev->io_queues[HCTX_TYPE_READ] = 0;
2057 
2058 	/*
2059 	 * Some Apple controllers require all queues to use the
2060 	 * first vector.
2061 	 */
2062 	if (dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR)
2063 		irq_queues = 1;
2064 
2065 	return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2066 			      PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2067 }
2068 
2069 static void nvme_disable_io_queues(struct nvme_dev *dev)
2070 {
2071 	if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2072 		__nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2073 }
2074 
2075 static unsigned int nvme_max_io_queues(struct nvme_dev *dev)
2076 {
2077 	return num_possible_cpus() + dev->nr_write_queues + dev->nr_poll_queues;
2078 }
2079 
2080 static int nvme_setup_io_queues(struct nvme_dev *dev)
2081 {
2082 	struct nvme_queue *adminq = &dev->queues[0];
2083 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2084 	unsigned int nr_io_queues;
2085 	unsigned long size;
2086 	int result;
2087 
2088 	/*
2089 	 * Sample the module parameters once at reset time so that we have
2090 	 * stable values to work with.
2091 	 */
2092 	dev->nr_write_queues = write_queues;
2093 	dev->nr_poll_queues = poll_queues;
2094 
2095 	/*
2096 	 * If tags are shared with admin queue (Apple bug), then
2097 	 * make sure we only use one IO queue.
2098 	 */
2099 	if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2100 		nr_io_queues = 1;
2101 	else
2102 		nr_io_queues = min(nvme_max_io_queues(dev),
2103 				   dev->nr_allocated_queues - 1);
2104 
2105 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2106 	if (result < 0)
2107 		return result;
2108 
2109 	if (nr_io_queues == 0)
2110 		return 0;
2111 
2112 	clear_bit(NVMEQ_ENABLED, &adminq->flags);
2113 
2114 	if (dev->cmb_use_sqes) {
2115 		result = nvme_cmb_qdepth(dev, nr_io_queues,
2116 				sizeof(struct nvme_command));
2117 		if (result > 0)
2118 			dev->q_depth = result;
2119 		else
2120 			dev->cmb_use_sqes = false;
2121 	}
2122 
2123 	do {
2124 		size = db_bar_size(dev, nr_io_queues);
2125 		result = nvme_remap_bar(dev, size);
2126 		if (!result)
2127 			break;
2128 		if (!--nr_io_queues)
2129 			return -ENOMEM;
2130 	} while (1);
2131 	adminq->q_db = dev->dbs;
2132 
2133  retry:
2134 	/* Deregister the admin queue's interrupt */
2135 	pci_free_irq(pdev, 0, adminq);
2136 
2137 	/*
2138 	 * If we enable msix early due to not intx, disable it again before
2139 	 * setting up the full range we need.
2140 	 */
2141 	pci_free_irq_vectors(pdev);
2142 
2143 	result = nvme_setup_irqs(dev, nr_io_queues);
2144 	if (result <= 0)
2145 		return -EIO;
2146 
2147 	dev->num_vecs = result;
2148 	result = max(result - 1, 1);
2149 	dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2150 
2151 	/*
2152 	 * Should investigate if there's a performance win from allocating
2153 	 * more queues than interrupt vectors; it might allow the submission
2154 	 * path to scale better, even if the receive path is limited by the
2155 	 * number of interrupts.
2156 	 */
2157 	result = queue_request_irq(adminq);
2158 	if (result)
2159 		return result;
2160 	set_bit(NVMEQ_ENABLED, &adminq->flags);
2161 
2162 	result = nvme_create_io_queues(dev);
2163 	if (result || dev->online_queues < 2)
2164 		return result;
2165 
2166 	if (dev->online_queues - 1 < dev->max_qid) {
2167 		nr_io_queues = dev->online_queues - 1;
2168 		nvme_disable_io_queues(dev);
2169 		nvme_suspend_io_queues(dev);
2170 		goto retry;
2171 	}
2172 	dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2173 					dev->io_queues[HCTX_TYPE_DEFAULT],
2174 					dev->io_queues[HCTX_TYPE_READ],
2175 					dev->io_queues[HCTX_TYPE_POLL]);
2176 	return 0;
2177 }
2178 
2179 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2180 {
2181 	struct nvme_queue *nvmeq = req->end_io_data;
2182 
2183 	blk_mq_free_request(req);
2184 	complete(&nvmeq->delete_done);
2185 }
2186 
2187 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2188 {
2189 	struct nvme_queue *nvmeq = req->end_io_data;
2190 
2191 	if (error)
2192 		set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2193 
2194 	nvme_del_queue_end(req, error);
2195 }
2196 
2197 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2198 {
2199 	struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2200 	struct request *req;
2201 	struct nvme_command cmd;
2202 
2203 	memset(&cmd, 0, sizeof(cmd));
2204 	cmd.delete_queue.opcode = opcode;
2205 	cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2206 
2207 	req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
2208 	if (IS_ERR(req))
2209 		return PTR_ERR(req);
2210 
2211 	req->timeout = ADMIN_TIMEOUT;
2212 	req->end_io_data = nvmeq;
2213 
2214 	init_completion(&nvmeq->delete_done);
2215 	blk_execute_rq_nowait(q, NULL, req, false,
2216 			opcode == nvme_admin_delete_cq ?
2217 				nvme_del_cq_end : nvme_del_queue_end);
2218 	return 0;
2219 }
2220 
2221 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2222 {
2223 	int nr_queues = dev->online_queues - 1, sent = 0;
2224 	unsigned long timeout;
2225 
2226  retry:
2227 	timeout = ADMIN_TIMEOUT;
2228 	while (nr_queues > 0) {
2229 		if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2230 			break;
2231 		nr_queues--;
2232 		sent++;
2233 	}
2234 	while (sent) {
2235 		struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2236 
2237 		timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2238 				timeout);
2239 		if (timeout == 0)
2240 			return false;
2241 
2242 		sent--;
2243 		if (nr_queues)
2244 			goto retry;
2245 	}
2246 	return true;
2247 }
2248 
2249 static void nvme_dev_add(struct nvme_dev *dev)
2250 {
2251 	int ret;
2252 
2253 	if (!dev->ctrl.tagset) {
2254 		dev->tagset.ops = &nvme_mq_ops;
2255 		dev->tagset.nr_hw_queues = dev->online_queues - 1;
2256 		dev->tagset.nr_maps = 2; /* default + read */
2257 		if (dev->io_queues[HCTX_TYPE_POLL])
2258 			dev->tagset.nr_maps++;
2259 		dev->tagset.timeout = NVME_IO_TIMEOUT;
2260 		dev->tagset.numa_node = dev_to_node(dev->dev);
2261 		dev->tagset.queue_depth =
2262 				min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2263 		dev->tagset.cmd_size = sizeof(struct nvme_iod);
2264 		dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2265 		dev->tagset.driver_data = dev;
2266 
2267 		/*
2268 		 * Some Apple controllers requires tags to be unique
2269 		 * across admin and IO queue, so reserve the first 32
2270 		 * tags of the IO queue.
2271 		 */
2272 		if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2273 			dev->tagset.reserved_tags = NVME_AQ_DEPTH;
2274 
2275 		ret = blk_mq_alloc_tag_set(&dev->tagset);
2276 		if (ret) {
2277 			dev_warn(dev->ctrl.device,
2278 				"IO queues tagset allocation failed %d\n", ret);
2279 			return;
2280 		}
2281 		dev->ctrl.tagset = &dev->tagset;
2282 	} else {
2283 		blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2284 
2285 		/* Free previously allocated queues that are no longer usable */
2286 		nvme_free_queues(dev, dev->online_queues);
2287 	}
2288 
2289 	nvme_dbbuf_set(dev);
2290 }
2291 
2292 static int nvme_pci_enable(struct nvme_dev *dev)
2293 {
2294 	int result = -ENOMEM;
2295 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2296 
2297 	if (pci_enable_device_mem(pdev))
2298 		return result;
2299 
2300 	pci_set_master(pdev);
2301 
2302 	if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)))
2303 		goto disable;
2304 
2305 	if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2306 		result = -ENODEV;
2307 		goto disable;
2308 	}
2309 
2310 	/*
2311 	 * Some devices and/or platforms don't advertise or work with INTx
2312 	 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2313 	 * adjust this later.
2314 	 */
2315 	result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2316 	if (result < 0)
2317 		return result;
2318 
2319 	dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2320 
2321 	dev->q_depth = min_t(int, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2322 				io_queue_depth);
2323 	dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2324 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2325 	dev->dbs = dev->bar + 4096;
2326 
2327 	/*
2328 	 * Some Apple controllers require a non-standard SQE size.
2329 	 * Interestingly they also seem to ignore the CC:IOSQES register
2330 	 * so we don't bother updating it here.
2331 	 */
2332 	if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2333 		dev->io_sqes = 7;
2334 	else
2335 		dev->io_sqes = NVME_NVM_IOSQES;
2336 
2337 	/*
2338 	 * Temporary fix for the Apple controller found in the MacBook8,1 and
2339 	 * some MacBook7,1 to avoid controller resets and data loss.
2340 	 */
2341 	if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2342 		dev->q_depth = 2;
2343 		dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2344 			"set queue depth=%u to work around controller resets\n",
2345 			dev->q_depth);
2346 	} else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2347 		   (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2348 		   NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2349 		dev->q_depth = 64;
2350 		dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2351                         "set queue depth=%u\n", dev->q_depth);
2352 	}
2353 
2354 	/*
2355 	 * Controllers with the shared tags quirk need the IO queue to be
2356 	 * big enough so that we get 32 tags for the admin queue
2357 	 */
2358 	if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2359 	    (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2360 		dev->q_depth = NVME_AQ_DEPTH + 2;
2361 		dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2362 			 dev->q_depth);
2363 	}
2364 
2365 
2366 	nvme_map_cmb(dev);
2367 
2368 	pci_enable_pcie_error_reporting(pdev);
2369 	pci_save_state(pdev);
2370 	return 0;
2371 
2372  disable:
2373 	pci_disable_device(pdev);
2374 	return result;
2375 }
2376 
2377 static void nvme_dev_unmap(struct nvme_dev *dev)
2378 {
2379 	if (dev->bar)
2380 		iounmap(dev->bar);
2381 	pci_release_mem_regions(to_pci_dev(dev->dev));
2382 }
2383 
2384 static void nvme_pci_disable(struct nvme_dev *dev)
2385 {
2386 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2387 
2388 	pci_free_irq_vectors(pdev);
2389 
2390 	if (pci_is_enabled(pdev)) {
2391 		pci_disable_pcie_error_reporting(pdev);
2392 		pci_disable_device(pdev);
2393 	}
2394 }
2395 
2396 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2397 {
2398 	bool dead = true, freeze = false;
2399 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2400 
2401 	mutex_lock(&dev->shutdown_lock);
2402 	if (pci_is_enabled(pdev)) {
2403 		u32 csts = readl(dev->bar + NVME_REG_CSTS);
2404 
2405 		if (dev->ctrl.state == NVME_CTRL_LIVE ||
2406 		    dev->ctrl.state == NVME_CTRL_RESETTING) {
2407 			freeze = true;
2408 			nvme_start_freeze(&dev->ctrl);
2409 		}
2410 		dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2411 			pdev->error_state  != pci_channel_io_normal);
2412 	}
2413 
2414 	/*
2415 	 * Give the controller a chance to complete all entered requests if
2416 	 * doing a safe shutdown.
2417 	 */
2418 	if (!dead && shutdown && freeze)
2419 		nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2420 
2421 	nvme_stop_queues(&dev->ctrl);
2422 
2423 	if (!dead && dev->ctrl.queue_count > 0) {
2424 		nvme_disable_io_queues(dev);
2425 		nvme_disable_admin_queue(dev, shutdown);
2426 	}
2427 	nvme_suspend_io_queues(dev);
2428 	nvme_suspend_queue(&dev->queues[0]);
2429 	nvme_pci_disable(dev);
2430 	nvme_reap_pending_cqes(dev);
2431 
2432 	blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2433 	blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2434 	blk_mq_tagset_wait_completed_request(&dev->tagset);
2435 	blk_mq_tagset_wait_completed_request(&dev->admin_tagset);
2436 
2437 	/*
2438 	 * The driver will not be starting up queues again if shutting down so
2439 	 * must flush all entered requests to their failed completion to avoid
2440 	 * deadlocking blk-mq hot-cpu notifier.
2441 	 */
2442 	if (shutdown) {
2443 		nvme_start_queues(&dev->ctrl);
2444 		if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2445 			blk_mq_unquiesce_queue(dev->ctrl.admin_q);
2446 	}
2447 	mutex_unlock(&dev->shutdown_lock);
2448 }
2449 
2450 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown)
2451 {
2452 	if (!nvme_wait_reset(&dev->ctrl))
2453 		return -EBUSY;
2454 	nvme_dev_disable(dev, shutdown);
2455 	return 0;
2456 }
2457 
2458 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2459 {
2460 	dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2461 						PAGE_SIZE, PAGE_SIZE, 0);
2462 	if (!dev->prp_page_pool)
2463 		return -ENOMEM;
2464 
2465 	/* Optimisation for I/Os between 4k and 128k */
2466 	dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2467 						256, 256, 0);
2468 	if (!dev->prp_small_pool) {
2469 		dma_pool_destroy(dev->prp_page_pool);
2470 		return -ENOMEM;
2471 	}
2472 	return 0;
2473 }
2474 
2475 static void nvme_release_prp_pools(struct nvme_dev *dev)
2476 {
2477 	dma_pool_destroy(dev->prp_page_pool);
2478 	dma_pool_destroy(dev->prp_small_pool);
2479 }
2480 
2481 static void nvme_free_tagset(struct nvme_dev *dev)
2482 {
2483 	if (dev->tagset.tags)
2484 		blk_mq_free_tag_set(&dev->tagset);
2485 	dev->ctrl.tagset = NULL;
2486 }
2487 
2488 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2489 {
2490 	struct nvme_dev *dev = to_nvme_dev(ctrl);
2491 
2492 	nvme_dbbuf_dma_free(dev);
2493 	nvme_free_tagset(dev);
2494 	if (dev->ctrl.admin_q)
2495 		blk_put_queue(dev->ctrl.admin_q);
2496 	free_opal_dev(dev->ctrl.opal_dev);
2497 	mempool_destroy(dev->iod_mempool);
2498 	put_device(dev->dev);
2499 	kfree(dev->queues);
2500 	kfree(dev);
2501 }
2502 
2503 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2504 {
2505 	/*
2506 	 * Set state to deleting now to avoid blocking nvme_wait_reset(), which
2507 	 * may be holding this pci_dev's device lock.
2508 	 */
2509 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2510 	nvme_get_ctrl(&dev->ctrl);
2511 	nvme_dev_disable(dev, false);
2512 	nvme_kill_queues(&dev->ctrl);
2513 	if (!queue_work(nvme_wq, &dev->remove_work))
2514 		nvme_put_ctrl(&dev->ctrl);
2515 }
2516 
2517 static void nvme_reset_work(struct work_struct *work)
2518 {
2519 	struct nvme_dev *dev =
2520 		container_of(work, struct nvme_dev, ctrl.reset_work);
2521 	bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2522 	int result;
2523 
2524 	if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING)) {
2525 		result = -ENODEV;
2526 		goto out;
2527 	}
2528 
2529 	/*
2530 	 * If we're called to reset a live controller first shut it down before
2531 	 * moving on.
2532 	 */
2533 	if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2534 		nvme_dev_disable(dev, false);
2535 	nvme_sync_queues(&dev->ctrl);
2536 
2537 	mutex_lock(&dev->shutdown_lock);
2538 	result = nvme_pci_enable(dev);
2539 	if (result)
2540 		goto out_unlock;
2541 
2542 	result = nvme_pci_configure_admin_queue(dev);
2543 	if (result)
2544 		goto out_unlock;
2545 
2546 	result = nvme_alloc_admin_tags(dev);
2547 	if (result)
2548 		goto out_unlock;
2549 
2550 	/*
2551 	 * Limit the max command size to prevent iod->sg allocations going
2552 	 * over a single page.
2553 	 */
2554 	dev->ctrl.max_hw_sectors = min_t(u32,
2555 		NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2556 	dev->ctrl.max_segments = NVME_MAX_SEGS;
2557 
2558 	/*
2559 	 * Don't limit the IOMMU merged segment size.
2560 	 */
2561 	dma_set_max_seg_size(dev->dev, 0xffffffff);
2562 
2563 	mutex_unlock(&dev->shutdown_lock);
2564 
2565 	/*
2566 	 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2567 	 * initializing procedure here.
2568 	 */
2569 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2570 		dev_warn(dev->ctrl.device,
2571 			"failed to mark controller CONNECTING\n");
2572 		result = -EBUSY;
2573 		goto out;
2574 	}
2575 
2576 	/*
2577 	 * We do not support an SGL for metadata (yet), so we are limited to a
2578 	 * single integrity segment for the separate metadata pointer.
2579 	 */
2580 	dev->ctrl.max_integrity_segments = 1;
2581 
2582 	result = nvme_init_identify(&dev->ctrl);
2583 	if (result)
2584 		goto out;
2585 
2586 	if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2587 		if (!dev->ctrl.opal_dev)
2588 			dev->ctrl.opal_dev =
2589 				init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2590 		else if (was_suspend)
2591 			opal_unlock_from_suspend(dev->ctrl.opal_dev);
2592 	} else {
2593 		free_opal_dev(dev->ctrl.opal_dev);
2594 		dev->ctrl.opal_dev = NULL;
2595 	}
2596 
2597 	if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2598 		result = nvme_dbbuf_dma_alloc(dev);
2599 		if (result)
2600 			dev_warn(dev->dev,
2601 				 "unable to allocate dma for dbbuf\n");
2602 	}
2603 
2604 	if (dev->ctrl.hmpre) {
2605 		result = nvme_setup_host_mem(dev);
2606 		if (result < 0)
2607 			goto out;
2608 	}
2609 
2610 	result = nvme_setup_io_queues(dev);
2611 	if (result)
2612 		goto out;
2613 
2614 	/*
2615 	 * Keep the controller around but remove all namespaces if we don't have
2616 	 * any working I/O queue.
2617 	 */
2618 	if (dev->online_queues < 2) {
2619 		dev_warn(dev->ctrl.device, "IO queues not created\n");
2620 		nvme_kill_queues(&dev->ctrl);
2621 		nvme_remove_namespaces(&dev->ctrl);
2622 		nvme_free_tagset(dev);
2623 	} else {
2624 		nvme_start_queues(&dev->ctrl);
2625 		nvme_wait_freeze(&dev->ctrl);
2626 		nvme_dev_add(dev);
2627 		nvme_unfreeze(&dev->ctrl);
2628 	}
2629 
2630 	/*
2631 	 * If only admin queue live, keep it to do further investigation or
2632 	 * recovery.
2633 	 */
2634 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2635 		dev_warn(dev->ctrl.device,
2636 			"failed to mark controller live state\n");
2637 		result = -ENODEV;
2638 		goto out;
2639 	}
2640 
2641 	nvme_start_ctrl(&dev->ctrl);
2642 	return;
2643 
2644  out_unlock:
2645 	mutex_unlock(&dev->shutdown_lock);
2646  out:
2647 	if (result)
2648 		dev_warn(dev->ctrl.device,
2649 			 "Removing after probe failure status: %d\n", result);
2650 	nvme_remove_dead_ctrl(dev);
2651 }
2652 
2653 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2654 {
2655 	struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2656 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2657 
2658 	if (pci_get_drvdata(pdev))
2659 		device_release_driver(&pdev->dev);
2660 	nvme_put_ctrl(&dev->ctrl);
2661 }
2662 
2663 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2664 {
2665 	*val = readl(to_nvme_dev(ctrl)->bar + off);
2666 	return 0;
2667 }
2668 
2669 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2670 {
2671 	writel(val, to_nvme_dev(ctrl)->bar + off);
2672 	return 0;
2673 }
2674 
2675 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2676 {
2677 	*val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
2678 	return 0;
2679 }
2680 
2681 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2682 {
2683 	struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2684 
2685 	return snprintf(buf, size, "%s\n", dev_name(&pdev->dev));
2686 }
2687 
2688 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2689 	.name			= "pcie",
2690 	.module			= THIS_MODULE,
2691 	.flags			= NVME_F_METADATA_SUPPORTED |
2692 				  NVME_F_PCI_P2PDMA,
2693 	.reg_read32		= nvme_pci_reg_read32,
2694 	.reg_write32		= nvme_pci_reg_write32,
2695 	.reg_read64		= nvme_pci_reg_read64,
2696 	.free_ctrl		= nvme_pci_free_ctrl,
2697 	.submit_async_event	= nvme_pci_submit_async_event,
2698 	.get_address		= nvme_pci_get_address,
2699 };
2700 
2701 static int nvme_dev_map(struct nvme_dev *dev)
2702 {
2703 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2704 
2705 	if (pci_request_mem_regions(pdev, "nvme"))
2706 		return -ENODEV;
2707 
2708 	if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2709 		goto release;
2710 
2711 	return 0;
2712   release:
2713 	pci_release_mem_regions(pdev);
2714 	return -ENODEV;
2715 }
2716 
2717 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2718 {
2719 	if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2720 		/*
2721 		 * Several Samsung devices seem to drop off the PCIe bus
2722 		 * randomly when APST is on and uses the deepest sleep state.
2723 		 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2724 		 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2725 		 * 950 PRO 256GB", but it seems to be restricted to two Dell
2726 		 * laptops.
2727 		 */
2728 		if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2729 		    (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2730 		     dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2731 			return NVME_QUIRK_NO_DEEPEST_PS;
2732 	} else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2733 		/*
2734 		 * Samsung SSD 960 EVO drops off the PCIe bus after system
2735 		 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2736 		 * within few minutes after bootup on a Coffee Lake board -
2737 		 * ASUS PRIME Z370-A
2738 		 */
2739 		if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2740 		    (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2741 		     dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2742 			return NVME_QUIRK_NO_APST;
2743 	} else if ((pdev->vendor == 0x144d && (pdev->device == 0xa801 ||
2744 		    pdev->device == 0xa808 || pdev->device == 0xa809)) ||
2745 		   (pdev->vendor == 0x1e0f && pdev->device == 0x0001)) {
2746 		/*
2747 		 * Forcing to use host managed nvme power settings for
2748 		 * lowest idle power with quick resume latency on
2749 		 * Samsung and Toshiba SSDs based on suspend behavior
2750 		 * on Coffee Lake board for LENOVO C640
2751 		 */
2752 		if ((dmi_match(DMI_BOARD_VENDOR, "LENOVO")) &&
2753 		     dmi_match(DMI_BOARD_NAME, "LNVNB161216"))
2754 			return NVME_QUIRK_SIMPLE_SUSPEND;
2755 	}
2756 
2757 	return 0;
2758 }
2759 
2760 static void nvme_async_probe(void *data, async_cookie_t cookie)
2761 {
2762 	struct nvme_dev *dev = data;
2763 
2764 	flush_work(&dev->ctrl.reset_work);
2765 	flush_work(&dev->ctrl.scan_work);
2766 	nvme_put_ctrl(&dev->ctrl);
2767 }
2768 
2769 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2770 {
2771 	int node, result = -ENOMEM;
2772 	struct nvme_dev *dev;
2773 	unsigned long quirks = id->driver_data;
2774 	size_t alloc_size;
2775 
2776 	node = dev_to_node(&pdev->dev);
2777 	if (node == NUMA_NO_NODE)
2778 		set_dev_node(&pdev->dev, first_memory_node);
2779 
2780 	dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2781 	if (!dev)
2782 		return -ENOMEM;
2783 
2784 	dev->nr_write_queues = write_queues;
2785 	dev->nr_poll_queues = poll_queues;
2786 	dev->nr_allocated_queues = nvme_max_io_queues(dev) + 1;
2787 	dev->queues = kcalloc_node(dev->nr_allocated_queues,
2788 			sizeof(struct nvme_queue), GFP_KERNEL, node);
2789 	if (!dev->queues)
2790 		goto free;
2791 
2792 	dev->dev = get_device(&pdev->dev);
2793 	pci_set_drvdata(pdev, dev);
2794 
2795 	result = nvme_dev_map(dev);
2796 	if (result)
2797 		goto put_pci;
2798 
2799 	INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2800 	INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2801 	mutex_init(&dev->shutdown_lock);
2802 
2803 	result = nvme_setup_prp_pools(dev);
2804 	if (result)
2805 		goto unmap;
2806 
2807 	quirks |= check_vendor_combination_bug(pdev);
2808 
2809 	/*
2810 	 * Double check that our mempool alloc size will cover the biggest
2811 	 * command we support.
2812 	 */
2813 	alloc_size = nvme_pci_iod_alloc_size(dev, NVME_MAX_KB_SZ,
2814 						NVME_MAX_SEGS, true);
2815 	WARN_ON_ONCE(alloc_size > PAGE_SIZE);
2816 
2817 	dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
2818 						mempool_kfree,
2819 						(void *) alloc_size,
2820 						GFP_KERNEL, node);
2821 	if (!dev->iod_mempool) {
2822 		result = -ENOMEM;
2823 		goto release_pools;
2824 	}
2825 
2826 	result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2827 			quirks);
2828 	if (result)
2829 		goto release_mempool;
2830 
2831 	dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2832 
2833 	nvme_reset_ctrl(&dev->ctrl);
2834 	async_schedule(nvme_async_probe, dev);
2835 
2836 	return 0;
2837 
2838  release_mempool:
2839 	mempool_destroy(dev->iod_mempool);
2840  release_pools:
2841 	nvme_release_prp_pools(dev);
2842  unmap:
2843 	nvme_dev_unmap(dev);
2844  put_pci:
2845 	put_device(dev->dev);
2846  free:
2847 	kfree(dev->queues);
2848 	kfree(dev);
2849 	return result;
2850 }
2851 
2852 static void nvme_reset_prepare(struct pci_dev *pdev)
2853 {
2854 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2855 
2856 	/*
2857 	 * We don't need to check the return value from waiting for the reset
2858 	 * state as pci_dev device lock is held, making it impossible to race
2859 	 * with ->remove().
2860 	 */
2861 	nvme_disable_prepare_reset(dev, false);
2862 	nvme_sync_queues(&dev->ctrl);
2863 }
2864 
2865 static void nvme_reset_done(struct pci_dev *pdev)
2866 {
2867 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2868 
2869 	if (!nvme_try_sched_reset(&dev->ctrl))
2870 		flush_work(&dev->ctrl.reset_work);
2871 }
2872 
2873 static void nvme_shutdown(struct pci_dev *pdev)
2874 {
2875 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2876 	nvme_disable_prepare_reset(dev, true);
2877 }
2878 
2879 /*
2880  * The driver's remove may be called on a device in a partially initialized
2881  * state. This function must not have any dependencies on the device state in
2882  * order to proceed.
2883  */
2884 static void nvme_remove(struct pci_dev *pdev)
2885 {
2886 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2887 
2888 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2889 	pci_set_drvdata(pdev, NULL);
2890 
2891 	if (!pci_device_is_present(pdev)) {
2892 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2893 		nvme_dev_disable(dev, true);
2894 		nvme_dev_remove_admin(dev);
2895 	}
2896 
2897 	flush_work(&dev->ctrl.reset_work);
2898 	nvme_stop_ctrl(&dev->ctrl);
2899 	nvme_remove_namespaces(&dev->ctrl);
2900 	nvme_dev_disable(dev, true);
2901 	nvme_release_cmb(dev);
2902 	nvme_free_host_mem(dev);
2903 	nvme_dev_remove_admin(dev);
2904 	nvme_free_queues(dev, 0);
2905 	nvme_release_prp_pools(dev);
2906 	nvme_dev_unmap(dev);
2907 	nvme_uninit_ctrl(&dev->ctrl);
2908 }
2909 
2910 #ifdef CONFIG_PM_SLEEP
2911 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
2912 {
2913 	return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
2914 }
2915 
2916 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
2917 {
2918 	return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
2919 }
2920 
2921 static int nvme_resume(struct device *dev)
2922 {
2923 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
2924 	struct nvme_ctrl *ctrl = &ndev->ctrl;
2925 
2926 	if (ndev->last_ps == U32_MAX ||
2927 	    nvme_set_power_state(ctrl, ndev->last_ps) != 0)
2928 		return nvme_try_sched_reset(&ndev->ctrl);
2929 	return 0;
2930 }
2931 
2932 static int nvme_suspend(struct device *dev)
2933 {
2934 	struct pci_dev *pdev = to_pci_dev(dev);
2935 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2936 	struct nvme_ctrl *ctrl = &ndev->ctrl;
2937 	int ret = -EBUSY;
2938 
2939 	ndev->last_ps = U32_MAX;
2940 
2941 	/*
2942 	 * The platform does not remove power for a kernel managed suspend so
2943 	 * use host managed nvme power settings for lowest idle power if
2944 	 * possible. This should have quicker resume latency than a full device
2945 	 * shutdown.  But if the firmware is involved after the suspend or the
2946 	 * device does not support any non-default power states, shut down the
2947 	 * device fully.
2948 	 *
2949 	 * If ASPM is not enabled for the device, shut down the device and allow
2950 	 * the PCI bus layer to put it into D3 in order to take the PCIe link
2951 	 * down, so as to allow the platform to achieve its minimum low-power
2952 	 * state (which may not be possible if the link is up).
2953 	 *
2954 	 * If a host memory buffer is enabled, shut down the device as the NVMe
2955 	 * specification allows the device to access the host memory buffer in
2956 	 * host DRAM from all power states, but hosts will fail access to DRAM
2957 	 * during S3.
2958 	 */
2959 	if (pm_suspend_via_firmware() || !ctrl->npss ||
2960 	    !pcie_aspm_enabled(pdev) ||
2961 	    ndev->nr_host_mem_descs ||
2962 	    (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND))
2963 		return nvme_disable_prepare_reset(ndev, true);
2964 
2965 	nvme_start_freeze(ctrl);
2966 	nvme_wait_freeze(ctrl);
2967 	nvme_sync_queues(ctrl);
2968 
2969 	if (ctrl->state != NVME_CTRL_LIVE)
2970 		goto unfreeze;
2971 
2972 	ret = nvme_get_power_state(ctrl, &ndev->last_ps);
2973 	if (ret < 0)
2974 		goto unfreeze;
2975 
2976 	/*
2977 	 * A saved state prevents pci pm from generically controlling the
2978 	 * device's power. If we're using protocol specific settings, we don't
2979 	 * want pci interfering.
2980 	 */
2981 	pci_save_state(pdev);
2982 
2983 	ret = nvme_set_power_state(ctrl, ctrl->npss);
2984 	if (ret < 0)
2985 		goto unfreeze;
2986 
2987 	if (ret) {
2988 		/* discard the saved state */
2989 		pci_load_saved_state(pdev, NULL);
2990 
2991 		/*
2992 		 * Clearing npss forces a controller reset on resume. The
2993 		 * correct value will be rediscovered then.
2994 		 */
2995 		ret = nvme_disable_prepare_reset(ndev, true);
2996 		ctrl->npss = 0;
2997 	}
2998 unfreeze:
2999 	nvme_unfreeze(ctrl);
3000 	return ret;
3001 }
3002 
3003 static int nvme_simple_suspend(struct device *dev)
3004 {
3005 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3006 	return nvme_disable_prepare_reset(ndev, true);
3007 }
3008 
3009 static int nvme_simple_resume(struct device *dev)
3010 {
3011 	struct pci_dev *pdev = to_pci_dev(dev);
3012 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
3013 
3014 	return nvme_try_sched_reset(&ndev->ctrl);
3015 }
3016 
3017 static const struct dev_pm_ops nvme_dev_pm_ops = {
3018 	.suspend	= nvme_suspend,
3019 	.resume		= nvme_resume,
3020 	.freeze		= nvme_simple_suspend,
3021 	.thaw		= nvme_simple_resume,
3022 	.poweroff	= nvme_simple_suspend,
3023 	.restore	= nvme_simple_resume,
3024 };
3025 #endif /* CONFIG_PM_SLEEP */
3026 
3027 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3028 						pci_channel_state_t state)
3029 {
3030 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3031 
3032 	/*
3033 	 * A frozen channel requires a reset. When detected, this method will
3034 	 * shutdown the controller to quiesce. The controller will be restarted
3035 	 * after the slot reset through driver's slot_reset callback.
3036 	 */
3037 	switch (state) {
3038 	case pci_channel_io_normal:
3039 		return PCI_ERS_RESULT_CAN_RECOVER;
3040 	case pci_channel_io_frozen:
3041 		dev_warn(dev->ctrl.device,
3042 			"frozen state error detected, reset controller\n");
3043 		nvme_dev_disable(dev, false);
3044 		return PCI_ERS_RESULT_NEED_RESET;
3045 	case pci_channel_io_perm_failure:
3046 		dev_warn(dev->ctrl.device,
3047 			"failure state error detected, request disconnect\n");
3048 		return PCI_ERS_RESULT_DISCONNECT;
3049 	}
3050 	return PCI_ERS_RESULT_NEED_RESET;
3051 }
3052 
3053 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3054 {
3055 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3056 
3057 	dev_info(dev->ctrl.device, "restart after slot reset\n");
3058 	pci_restore_state(pdev);
3059 	nvme_reset_ctrl(&dev->ctrl);
3060 	return PCI_ERS_RESULT_RECOVERED;
3061 }
3062 
3063 static void nvme_error_resume(struct pci_dev *pdev)
3064 {
3065 	struct nvme_dev *dev = pci_get_drvdata(pdev);
3066 
3067 	flush_work(&dev->ctrl.reset_work);
3068 }
3069 
3070 static const struct pci_error_handlers nvme_err_handler = {
3071 	.error_detected	= nvme_error_detected,
3072 	.slot_reset	= nvme_slot_reset,
3073 	.resume		= nvme_error_resume,
3074 	.reset_prepare	= nvme_reset_prepare,
3075 	.reset_done	= nvme_reset_done,
3076 };
3077 
3078 static const struct pci_device_id nvme_id_table[] = {
3079 	{ PCI_VDEVICE(INTEL, 0x0953),
3080 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3081 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3082 	{ PCI_VDEVICE(INTEL, 0x0a53),
3083 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3084 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3085 	{ PCI_VDEVICE(INTEL, 0x0a54),
3086 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3087 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3088 	{ PCI_VDEVICE(INTEL, 0x0a55),
3089 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3090 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3091 	{ PCI_VDEVICE(INTEL, 0xf1a5),	/* Intel 600P/P3100 */
3092 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3093 				NVME_QUIRK_MEDIUM_PRIO_SQ |
3094 				NVME_QUIRK_NO_TEMP_THRESH_CHANGE },
3095 	{ PCI_VDEVICE(INTEL, 0xf1a6),	/* Intel 760p/Pro 7600p */
3096 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3097 	{ PCI_VDEVICE(INTEL, 0x5845),	/* Qemu emulated controller */
3098 		.driver_data = NVME_QUIRK_IDENTIFY_CNS |
3099 				NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3100 	{ PCI_DEVICE(0x1bb1, 0x0100),   /* Seagate Nytro Flash Storage */
3101 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3102 	{ PCI_DEVICE(0x1c58, 0x0003),	/* HGST adapter */
3103 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3104 	{ PCI_DEVICE(0x1c58, 0x0023),	/* WDC SN200 adapter */
3105 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3106 	{ PCI_DEVICE(0x1c5f, 0x0540),	/* Memblaze Pblaze4 adapter */
3107 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3108 	{ PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
3109 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3110 	{ PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
3111 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3112 	{ PCI_DEVICE(0x1d1d, 0x1f1f),	/* LighNVM qemu device */
3113 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3114 	{ PCI_DEVICE(0x1d1d, 0x2807),	/* CNEX WL */
3115 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3116 	{ PCI_DEVICE(0x1d1d, 0x2601),	/* CNEX Granby */
3117 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3118 	{ PCI_DEVICE(0x10ec, 0x5762),   /* ADATA SX6000LNP */
3119 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3120 	{ PCI_DEVICE(0x1cc1, 0x8201),   /* ADATA SX8200PNP 512GB */
3121 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3122 				NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3123 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3124 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001),
3125 		.driver_data = NVME_QUIRK_SINGLE_VECTOR },
3126 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3127 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3128 		.driver_data = NVME_QUIRK_SINGLE_VECTOR |
3129 				NVME_QUIRK_128_BYTES_SQES |
3130 				NVME_QUIRK_SHARED_TAGS },
3131 	{ 0, }
3132 };
3133 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3134 
3135 static struct pci_driver nvme_driver = {
3136 	.name		= "nvme",
3137 	.id_table	= nvme_id_table,
3138 	.probe		= nvme_probe,
3139 	.remove		= nvme_remove,
3140 	.shutdown	= nvme_shutdown,
3141 #ifdef CONFIG_PM_SLEEP
3142 	.driver		= {
3143 		.pm	= &nvme_dev_pm_ops,
3144 	},
3145 #endif
3146 	.sriov_configure = pci_sriov_configure_simple,
3147 	.err_handler	= &nvme_err_handler,
3148 };
3149 
3150 static int __init nvme_init(void)
3151 {
3152 	BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3153 	BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3154 	BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3155 	BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3156 
3157 	return pci_register_driver(&nvme_driver);
3158 }
3159 
3160 static void __exit nvme_exit(void)
3161 {
3162 	pci_unregister_driver(&nvme_driver);
3163 	flush_workqueue(nvme_wq);
3164 }
3165 
3166 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3167 MODULE_LICENSE("GPL");
3168 MODULE_VERSION("1.0");
3169 module_init(nvme_init);
3170 module_exit(nvme_exit);
3171