xref: /openbmc/linux/drivers/nvme/host/pci.c (revision b830f94f)
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(depth)		(depth * sizeof(struct nvme_command))
32 #define CQ_SIZE(depth)		(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 write_queues;
72 module_param(write_queues, int, 0644);
73 MODULE_PARM_DESC(write_queues,
74 	"Number of queues to use for writes. If not set, reads and writes "
75 	"will share a queue set.");
76 
77 static int poll_queues;
78 module_param(poll_queues, int, 0644);
79 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
80 
81 struct nvme_dev;
82 struct nvme_queue;
83 
84 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
85 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode);
86 
87 /*
88  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
89  */
90 struct nvme_dev {
91 	struct nvme_queue *queues;
92 	struct blk_mq_tag_set tagset;
93 	struct blk_mq_tag_set admin_tagset;
94 	u32 __iomem *dbs;
95 	struct device *dev;
96 	struct dma_pool *prp_page_pool;
97 	struct dma_pool *prp_small_pool;
98 	unsigned online_queues;
99 	unsigned max_qid;
100 	unsigned io_queues[HCTX_MAX_TYPES];
101 	unsigned int num_vecs;
102 	int q_depth;
103 	u32 db_stride;
104 	void __iomem *bar;
105 	unsigned long bar_mapped_size;
106 	struct work_struct remove_work;
107 	struct mutex shutdown_lock;
108 	bool subsystem;
109 	u64 cmb_size;
110 	bool cmb_use_sqes;
111 	u32 cmbsz;
112 	u32 cmbloc;
113 	struct nvme_ctrl ctrl;
114 	u32 last_ps;
115 
116 	mempool_t *iod_mempool;
117 
118 	/* shadow doorbell buffer support: */
119 	u32 *dbbuf_dbs;
120 	dma_addr_t dbbuf_dbs_dma_addr;
121 	u32 *dbbuf_eis;
122 	dma_addr_t dbbuf_eis_dma_addr;
123 
124 	/* host memory buffer support: */
125 	u64 host_mem_size;
126 	u32 nr_host_mem_descs;
127 	dma_addr_t host_mem_descs_dma;
128 	struct nvme_host_mem_buf_desc *host_mem_descs;
129 	void **host_mem_desc_bufs;
130 };
131 
132 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
133 {
134 	int n = 0, ret;
135 
136 	ret = kstrtoint(val, 10, &n);
137 	if (ret != 0 || n < 2)
138 		return -EINVAL;
139 
140 	return param_set_int(val, kp);
141 }
142 
143 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
144 {
145 	return qid * 2 * stride;
146 }
147 
148 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
149 {
150 	return (qid * 2 + 1) * stride;
151 }
152 
153 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
154 {
155 	return container_of(ctrl, struct nvme_dev, ctrl);
156 }
157 
158 /*
159  * An NVM Express queue.  Each device has at least two (one for admin
160  * commands and one for I/O commands).
161  */
162 struct nvme_queue {
163 	struct nvme_dev *dev;
164 	spinlock_t sq_lock;
165 	struct nvme_command *sq_cmds;
166 	 /* only used for poll queues: */
167 	spinlock_t cq_poll_lock ____cacheline_aligned_in_smp;
168 	volatile struct nvme_completion *cqes;
169 	struct blk_mq_tags **tags;
170 	dma_addr_t sq_dma_addr;
171 	dma_addr_t cq_dma_addr;
172 	u32 __iomem *q_db;
173 	u16 q_depth;
174 	u16 cq_vector;
175 	u16 sq_tail;
176 	u16 last_sq_tail;
177 	u16 cq_head;
178 	u16 last_cq_head;
179 	u16 qid;
180 	u8 cq_phase;
181 	unsigned long flags;
182 #define NVMEQ_ENABLED		0
183 #define NVMEQ_SQ_CMB		1
184 #define NVMEQ_DELETE_ERROR	2
185 #define NVMEQ_POLLED		3
186 	u32 *dbbuf_sq_db;
187 	u32 *dbbuf_cq_db;
188 	u32 *dbbuf_sq_ei;
189 	u32 *dbbuf_cq_ei;
190 	struct completion delete_done;
191 };
192 
193 /*
194  * The nvme_iod describes the data in an I/O.
195  *
196  * The sg pointer contains the list of PRP/SGL chunk allocations in addition
197  * to the actual struct scatterlist.
198  */
199 struct nvme_iod {
200 	struct nvme_request req;
201 	struct nvme_queue *nvmeq;
202 	bool use_sgl;
203 	int aborted;
204 	int npages;		/* In the PRP list. 0 means small pool in use */
205 	int nents;		/* Used in scatterlist */
206 	dma_addr_t first_dma;
207 	unsigned int dma_len;	/* length of single DMA segment mapping */
208 	dma_addr_t meta_dma;
209 	struct scatterlist *sg;
210 };
211 
212 static unsigned int max_io_queues(void)
213 {
214 	return num_possible_cpus() + write_queues + poll_queues;
215 }
216 
217 static unsigned int max_queue_count(void)
218 {
219 	/* IO queues + admin queue */
220 	return 1 + max_io_queues();
221 }
222 
223 static inline unsigned int nvme_dbbuf_size(u32 stride)
224 {
225 	return (max_queue_count() * 8 * stride);
226 }
227 
228 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
229 {
230 	unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
231 
232 	if (dev->dbbuf_dbs)
233 		return 0;
234 
235 	dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
236 					    &dev->dbbuf_dbs_dma_addr,
237 					    GFP_KERNEL);
238 	if (!dev->dbbuf_dbs)
239 		return -ENOMEM;
240 	dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
241 					    &dev->dbbuf_eis_dma_addr,
242 					    GFP_KERNEL);
243 	if (!dev->dbbuf_eis) {
244 		dma_free_coherent(dev->dev, mem_size,
245 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
246 		dev->dbbuf_dbs = NULL;
247 		return -ENOMEM;
248 	}
249 
250 	return 0;
251 }
252 
253 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
254 {
255 	unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
256 
257 	if (dev->dbbuf_dbs) {
258 		dma_free_coherent(dev->dev, mem_size,
259 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
260 		dev->dbbuf_dbs = NULL;
261 	}
262 	if (dev->dbbuf_eis) {
263 		dma_free_coherent(dev->dev, mem_size,
264 				  dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
265 		dev->dbbuf_eis = NULL;
266 	}
267 }
268 
269 static void nvme_dbbuf_init(struct nvme_dev *dev,
270 			    struct nvme_queue *nvmeq, int qid)
271 {
272 	if (!dev->dbbuf_dbs || !qid)
273 		return;
274 
275 	nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
276 	nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
277 	nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
278 	nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
279 }
280 
281 static void nvme_dbbuf_set(struct nvme_dev *dev)
282 {
283 	struct nvme_command c;
284 
285 	if (!dev->dbbuf_dbs)
286 		return;
287 
288 	memset(&c, 0, sizeof(c));
289 	c.dbbuf.opcode = nvme_admin_dbbuf;
290 	c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
291 	c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
292 
293 	if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
294 		dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
295 		/* Free memory and continue on */
296 		nvme_dbbuf_dma_free(dev);
297 	}
298 }
299 
300 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
301 {
302 	return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
303 }
304 
305 /* Update dbbuf and return true if an MMIO is required */
306 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
307 					      volatile u32 *dbbuf_ei)
308 {
309 	if (dbbuf_db) {
310 		u16 old_value;
311 
312 		/*
313 		 * Ensure that the queue is written before updating
314 		 * the doorbell in memory
315 		 */
316 		wmb();
317 
318 		old_value = *dbbuf_db;
319 		*dbbuf_db = value;
320 
321 		/*
322 		 * Ensure that the doorbell is updated before reading the event
323 		 * index from memory.  The controller needs to provide similar
324 		 * ordering to ensure the envent index is updated before reading
325 		 * the doorbell.
326 		 */
327 		mb();
328 
329 		if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
330 			return false;
331 	}
332 
333 	return true;
334 }
335 
336 /*
337  * Will slightly overestimate the number of pages needed.  This is OK
338  * as it only leads to a small amount of wasted memory for the lifetime of
339  * the I/O.
340  */
341 static int nvme_npages(unsigned size, struct nvme_dev *dev)
342 {
343 	unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
344 				      dev->ctrl.page_size);
345 	return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
346 }
347 
348 /*
349  * Calculates the number of pages needed for the SGL segments. For example a 4k
350  * page can accommodate 256 SGL descriptors.
351  */
352 static int nvme_pci_npages_sgl(unsigned int num_seg)
353 {
354 	return DIV_ROUND_UP(num_seg * sizeof(struct nvme_sgl_desc), PAGE_SIZE);
355 }
356 
357 static unsigned int nvme_pci_iod_alloc_size(struct nvme_dev *dev,
358 		unsigned int size, unsigned int nseg, bool use_sgl)
359 {
360 	size_t alloc_size;
361 
362 	if (use_sgl)
363 		alloc_size = sizeof(__le64 *) * nvme_pci_npages_sgl(nseg);
364 	else
365 		alloc_size = sizeof(__le64 *) * nvme_npages(size, dev);
366 
367 	return alloc_size + sizeof(struct scatterlist) * nseg;
368 }
369 
370 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
371 				unsigned int hctx_idx)
372 {
373 	struct nvme_dev *dev = data;
374 	struct nvme_queue *nvmeq = &dev->queues[0];
375 
376 	WARN_ON(hctx_idx != 0);
377 	WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
378 	WARN_ON(nvmeq->tags);
379 
380 	hctx->driver_data = nvmeq;
381 	nvmeq->tags = &dev->admin_tagset.tags[0];
382 	return 0;
383 }
384 
385 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
386 {
387 	struct nvme_queue *nvmeq = hctx->driver_data;
388 
389 	nvmeq->tags = NULL;
390 }
391 
392 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
393 			  unsigned int hctx_idx)
394 {
395 	struct nvme_dev *dev = data;
396 	struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
397 
398 	if (!nvmeq->tags)
399 		nvmeq->tags = &dev->tagset.tags[hctx_idx];
400 
401 	WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
402 	hctx->driver_data = nvmeq;
403 	return 0;
404 }
405 
406 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
407 		unsigned int hctx_idx, unsigned int numa_node)
408 {
409 	struct nvme_dev *dev = set->driver_data;
410 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
411 	int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
412 	struct nvme_queue *nvmeq = &dev->queues[queue_idx];
413 
414 	BUG_ON(!nvmeq);
415 	iod->nvmeq = nvmeq;
416 
417 	nvme_req(req)->ctrl = &dev->ctrl;
418 	return 0;
419 }
420 
421 static int queue_irq_offset(struct nvme_dev *dev)
422 {
423 	/* if we have more than 1 vec, admin queue offsets us by 1 */
424 	if (dev->num_vecs > 1)
425 		return 1;
426 
427 	return 0;
428 }
429 
430 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
431 {
432 	struct nvme_dev *dev = set->driver_data;
433 	int i, qoff, offset;
434 
435 	offset = queue_irq_offset(dev);
436 	for (i = 0, qoff = 0; i < set->nr_maps; i++) {
437 		struct blk_mq_queue_map *map = &set->map[i];
438 
439 		map->nr_queues = dev->io_queues[i];
440 		if (!map->nr_queues) {
441 			BUG_ON(i == HCTX_TYPE_DEFAULT);
442 			continue;
443 		}
444 
445 		/*
446 		 * The poll queue(s) doesn't have an IRQ (and hence IRQ
447 		 * affinity), so use the regular blk-mq cpu mapping
448 		 */
449 		map->queue_offset = qoff;
450 		if (i != HCTX_TYPE_POLL && offset)
451 			blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
452 		else
453 			blk_mq_map_queues(map);
454 		qoff += map->nr_queues;
455 		offset += map->nr_queues;
456 	}
457 
458 	return 0;
459 }
460 
461 /*
462  * Write sq tail if we are asked to, or if the next command would wrap.
463  */
464 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq)
465 {
466 	if (!write_sq) {
467 		u16 next_tail = nvmeq->sq_tail + 1;
468 
469 		if (next_tail == nvmeq->q_depth)
470 			next_tail = 0;
471 		if (next_tail != nvmeq->last_sq_tail)
472 			return;
473 	}
474 
475 	if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
476 			nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
477 		writel(nvmeq->sq_tail, nvmeq->q_db);
478 	nvmeq->last_sq_tail = nvmeq->sq_tail;
479 }
480 
481 /**
482  * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
483  * @nvmeq: The queue to use
484  * @cmd: The command to send
485  * @write_sq: whether to write to the SQ doorbell
486  */
487 static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
488 			    bool write_sq)
489 {
490 	spin_lock(&nvmeq->sq_lock);
491 	memcpy(&nvmeq->sq_cmds[nvmeq->sq_tail], cmd, sizeof(*cmd));
492 	if (++nvmeq->sq_tail == nvmeq->q_depth)
493 		nvmeq->sq_tail = 0;
494 	nvme_write_sq_db(nvmeq, write_sq);
495 	spin_unlock(&nvmeq->sq_lock);
496 }
497 
498 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
499 {
500 	struct nvme_queue *nvmeq = hctx->driver_data;
501 
502 	spin_lock(&nvmeq->sq_lock);
503 	if (nvmeq->sq_tail != nvmeq->last_sq_tail)
504 		nvme_write_sq_db(nvmeq, true);
505 	spin_unlock(&nvmeq->sq_lock);
506 }
507 
508 static void **nvme_pci_iod_list(struct request *req)
509 {
510 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
511 	return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
512 }
513 
514 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req)
515 {
516 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
517 	int nseg = blk_rq_nr_phys_segments(req);
518 	unsigned int avg_seg_size;
519 
520 	if (nseg == 0)
521 		return false;
522 
523 	avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
524 
525 	if (!(dev->ctrl.sgls & ((1 << 0) | (1 << 1))))
526 		return false;
527 	if (!iod->nvmeq->qid)
528 		return false;
529 	if (!sgl_threshold || avg_seg_size < sgl_threshold)
530 		return false;
531 	return true;
532 }
533 
534 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
535 {
536 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
537 	enum dma_data_direction dma_dir = rq_data_dir(req) ?
538 			DMA_TO_DEVICE : DMA_FROM_DEVICE;
539 	const int last_prp = dev->ctrl.page_size / sizeof(__le64) - 1;
540 	dma_addr_t dma_addr = iod->first_dma, next_dma_addr;
541 	int i;
542 
543 	if (iod->dma_len) {
544 		dma_unmap_page(dev->dev, dma_addr, iod->dma_len, dma_dir);
545 		return;
546 	}
547 
548 	WARN_ON_ONCE(!iod->nents);
549 
550 	/* P2PDMA requests do not need to be unmapped */
551 	if (!is_pci_p2pdma_page(sg_page(iod->sg)))
552 		dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req));
553 
554 
555 	if (iod->npages == 0)
556 		dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
557 			dma_addr);
558 
559 	for (i = 0; i < iod->npages; i++) {
560 		void *addr = nvme_pci_iod_list(req)[i];
561 
562 		if (iod->use_sgl) {
563 			struct nvme_sgl_desc *sg_list = addr;
564 
565 			next_dma_addr =
566 			    le64_to_cpu((sg_list[SGES_PER_PAGE - 1]).addr);
567 		} else {
568 			__le64 *prp_list = addr;
569 
570 			next_dma_addr = le64_to_cpu(prp_list[last_prp]);
571 		}
572 
573 		dma_pool_free(dev->prp_page_pool, addr, dma_addr);
574 		dma_addr = next_dma_addr;
575 	}
576 
577 	mempool_free(iod->sg, dev->iod_mempool);
578 }
579 
580 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
581 {
582 	int i;
583 	struct scatterlist *sg;
584 
585 	for_each_sg(sgl, sg, nents, i) {
586 		dma_addr_t phys = sg_phys(sg);
587 		pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
588 			"dma_address:%pad dma_length:%d\n",
589 			i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
590 			sg_dma_len(sg));
591 	}
592 }
593 
594 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
595 		struct request *req, struct nvme_rw_command *cmnd)
596 {
597 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
598 	struct dma_pool *pool;
599 	int length = blk_rq_payload_bytes(req);
600 	struct scatterlist *sg = iod->sg;
601 	int dma_len = sg_dma_len(sg);
602 	u64 dma_addr = sg_dma_address(sg);
603 	u32 page_size = dev->ctrl.page_size;
604 	int offset = dma_addr & (page_size - 1);
605 	__le64 *prp_list;
606 	void **list = nvme_pci_iod_list(req);
607 	dma_addr_t prp_dma;
608 	int nprps, i;
609 
610 	length -= (page_size - offset);
611 	if (length <= 0) {
612 		iod->first_dma = 0;
613 		goto done;
614 	}
615 
616 	dma_len -= (page_size - offset);
617 	if (dma_len) {
618 		dma_addr += (page_size - offset);
619 	} else {
620 		sg = sg_next(sg);
621 		dma_addr = sg_dma_address(sg);
622 		dma_len = sg_dma_len(sg);
623 	}
624 
625 	if (length <= page_size) {
626 		iod->first_dma = dma_addr;
627 		goto done;
628 	}
629 
630 	nprps = DIV_ROUND_UP(length, page_size);
631 	if (nprps <= (256 / 8)) {
632 		pool = dev->prp_small_pool;
633 		iod->npages = 0;
634 	} else {
635 		pool = dev->prp_page_pool;
636 		iod->npages = 1;
637 	}
638 
639 	prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
640 	if (!prp_list) {
641 		iod->first_dma = dma_addr;
642 		iod->npages = -1;
643 		return BLK_STS_RESOURCE;
644 	}
645 	list[0] = prp_list;
646 	iod->first_dma = prp_dma;
647 	i = 0;
648 	for (;;) {
649 		if (i == page_size >> 3) {
650 			__le64 *old_prp_list = prp_list;
651 			prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
652 			if (!prp_list)
653 				return BLK_STS_RESOURCE;
654 			list[iod->npages++] = prp_list;
655 			prp_list[0] = old_prp_list[i - 1];
656 			old_prp_list[i - 1] = cpu_to_le64(prp_dma);
657 			i = 1;
658 		}
659 		prp_list[i++] = cpu_to_le64(dma_addr);
660 		dma_len -= page_size;
661 		dma_addr += page_size;
662 		length -= page_size;
663 		if (length <= 0)
664 			break;
665 		if (dma_len > 0)
666 			continue;
667 		if (unlikely(dma_len < 0))
668 			goto bad_sgl;
669 		sg = sg_next(sg);
670 		dma_addr = sg_dma_address(sg);
671 		dma_len = sg_dma_len(sg);
672 	}
673 
674 done:
675 	cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
676 	cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
677 
678 	return BLK_STS_OK;
679 
680  bad_sgl:
681 	WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
682 			"Invalid SGL for payload:%d nents:%d\n",
683 			blk_rq_payload_bytes(req), iod->nents);
684 	return BLK_STS_IOERR;
685 }
686 
687 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
688 		struct scatterlist *sg)
689 {
690 	sge->addr = cpu_to_le64(sg_dma_address(sg));
691 	sge->length = cpu_to_le32(sg_dma_len(sg));
692 	sge->type = NVME_SGL_FMT_DATA_DESC << 4;
693 }
694 
695 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
696 		dma_addr_t dma_addr, int entries)
697 {
698 	sge->addr = cpu_to_le64(dma_addr);
699 	if (entries < SGES_PER_PAGE) {
700 		sge->length = cpu_to_le32(entries * sizeof(*sge));
701 		sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
702 	} else {
703 		sge->length = cpu_to_le32(PAGE_SIZE);
704 		sge->type = NVME_SGL_FMT_SEG_DESC << 4;
705 	}
706 }
707 
708 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
709 		struct request *req, struct nvme_rw_command *cmd, int entries)
710 {
711 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
712 	struct dma_pool *pool;
713 	struct nvme_sgl_desc *sg_list;
714 	struct scatterlist *sg = iod->sg;
715 	dma_addr_t sgl_dma;
716 	int i = 0;
717 
718 	/* setting the transfer type as SGL */
719 	cmd->flags = NVME_CMD_SGL_METABUF;
720 
721 	if (entries == 1) {
722 		nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
723 		return BLK_STS_OK;
724 	}
725 
726 	if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
727 		pool = dev->prp_small_pool;
728 		iod->npages = 0;
729 	} else {
730 		pool = dev->prp_page_pool;
731 		iod->npages = 1;
732 	}
733 
734 	sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
735 	if (!sg_list) {
736 		iod->npages = -1;
737 		return BLK_STS_RESOURCE;
738 	}
739 
740 	nvme_pci_iod_list(req)[0] = sg_list;
741 	iod->first_dma = sgl_dma;
742 
743 	nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
744 
745 	do {
746 		if (i == SGES_PER_PAGE) {
747 			struct nvme_sgl_desc *old_sg_desc = sg_list;
748 			struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
749 
750 			sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
751 			if (!sg_list)
752 				return BLK_STS_RESOURCE;
753 
754 			i = 0;
755 			nvme_pci_iod_list(req)[iod->npages++] = sg_list;
756 			sg_list[i++] = *link;
757 			nvme_pci_sgl_set_seg(link, sgl_dma, entries);
758 		}
759 
760 		nvme_pci_sgl_set_data(&sg_list[i++], sg);
761 		sg = sg_next(sg);
762 	} while (--entries > 0);
763 
764 	return BLK_STS_OK;
765 }
766 
767 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
768 		struct request *req, struct nvme_rw_command *cmnd,
769 		struct bio_vec *bv)
770 {
771 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
772 	unsigned int first_prp_len = dev->ctrl.page_size - bv->bv_offset;
773 
774 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
775 	if (dma_mapping_error(dev->dev, iod->first_dma))
776 		return BLK_STS_RESOURCE;
777 	iod->dma_len = bv->bv_len;
778 
779 	cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
780 	if (bv->bv_len > first_prp_len)
781 		cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
782 	return 0;
783 }
784 
785 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
786 		struct request *req, struct nvme_rw_command *cmnd,
787 		struct bio_vec *bv)
788 {
789 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
790 
791 	iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
792 	if (dma_mapping_error(dev->dev, iod->first_dma))
793 		return BLK_STS_RESOURCE;
794 	iod->dma_len = bv->bv_len;
795 
796 	cmnd->flags = NVME_CMD_SGL_METABUF;
797 	cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
798 	cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
799 	cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
800 	return 0;
801 }
802 
803 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
804 		struct nvme_command *cmnd)
805 {
806 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
807 	blk_status_t ret = BLK_STS_RESOURCE;
808 	int nr_mapped;
809 
810 	if (blk_rq_nr_phys_segments(req) == 1) {
811 		struct bio_vec bv = req_bvec(req);
812 
813 		if (!is_pci_p2pdma_page(bv.bv_page)) {
814 			if (bv.bv_offset + bv.bv_len <= dev->ctrl.page_size * 2)
815 				return nvme_setup_prp_simple(dev, req,
816 							     &cmnd->rw, &bv);
817 
818 			if (iod->nvmeq->qid &&
819 			    dev->ctrl.sgls & ((1 << 0) | (1 << 1)))
820 				return nvme_setup_sgl_simple(dev, req,
821 							     &cmnd->rw, &bv);
822 		}
823 	}
824 
825 	iod->dma_len = 0;
826 	iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
827 	if (!iod->sg)
828 		return BLK_STS_RESOURCE;
829 	sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
830 	iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
831 	if (!iod->nents)
832 		goto out;
833 
834 	if (is_pci_p2pdma_page(sg_page(iod->sg)))
835 		nr_mapped = pci_p2pdma_map_sg(dev->dev, iod->sg, iod->nents,
836 					      rq_dma_dir(req));
837 	else
838 		nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
839 					     rq_dma_dir(req), DMA_ATTR_NO_WARN);
840 	if (!nr_mapped)
841 		goto out;
842 
843 	iod->use_sgl = nvme_pci_use_sgls(dev, req);
844 	if (iod->use_sgl)
845 		ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
846 	else
847 		ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
848 out:
849 	if (ret != BLK_STS_OK)
850 		nvme_unmap_data(dev, req);
851 	return ret;
852 }
853 
854 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
855 		struct nvme_command *cmnd)
856 {
857 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
858 
859 	iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
860 			rq_dma_dir(req), 0);
861 	if (dma_mapping_error(dev->dev, iod->meta_dma))
862 		return BLK_STS_IOERR;
863 	cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
864 	return 0;
865 }
866 
867 /*
868  * NOTE: ns is NULL when called on the admin queue.
869  */
870 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
871 			 const struct blk_mq_queue_data *bd)
872 {
873 	struct nvme_ns *ns = hctx->queue->queuedata;
874 	struct nvme_queue *nvmeq = hctx->driver_data;
875 	struct nvme_dev *dev = nvmeq->dev;
876 	struct request *req = bd->rq;
877 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
878 	struct nvme_command cmnd;
879 	blk_status_t ret;
880 
881 	iod->aborted = 0;
882 	iod->npages = -1;
883 	iod->nents = 0;
884 
885 	/*
886 	 * We should not need to do this, but we're still using this to
887 	 * ensure we can drain requests on a dying queue.
888 	 */
889 	if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
890 		return BLK_STS_IOERR;
891 
892 	ret = nvme_setup_cmd(ns, req, &cmnd);
893 	if (ret)
894 		return ret;
895 
896 	if (blk_rq_nr_phys_segments(req)) {
897 		ret = nvme_map_data(dev, req, &cmnd);
898 		if (ret)
899 			goto out_free_cmd;
900 	}
901 
902 	if (blk_integrity_rq(req)) {
903 		ret = nvme_map_metadata(dev, req, &cmnd);
904 		if (ret)
905 			goto out_unmap_data;
906 	}
907 
908 	blk_mq_start_request(req);
909 	nvme_submit_cmd(nvmeq, &cmnd, bd->last);
910 	return BLK_STS_OK;
911 out_unmap_data:
912 	nvme_unmap_data(dev, req);
913 out_free_cmd:
914 	nvme_cleanup_cmd(req);
915 	return ret;
916 }
917 
918 static void nvme_pci_complete_rq(struct request *req)
919 {
920 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
921 	struct nvme_dev *dev = iod->nvmeq->dev;
922 
923 	nvme_cleanup_cmd(req);
924 	if (blk_integrity_rq(req))
925 		dma_unmap_page(dev->dev, iod->meta_dma,
926 			       rq_integrity_vec(req)->bv_len, rq_data_dir(req));
927 	if (blk_rq_nr_phys_segments(req))
928 		nvme_unmap_data(dev, req);
929 	nvme_complete_rq(req);
930 }
931 
932 /* We read the CQE phase first to check if the rest of the entry is valid */
933 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
934 {
935 	return (le16_to_cpu(nvmeq->cqes[nvmeq->cq_head].status) & 1) ==
936 			nvmeq->cq_phase;
937 }
938 
939 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
940 {
941 	u16 head = nvmeq->cq_head;
942 
943 	if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
944 					      nvmeq->dbbuf_cq_ei))
945 		writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
946 }
947 
948 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
949 {
950 	volatile struct nvme_completion *cqe = &nvmeq->cqes[idx];
951 	struct request *req;
952 
953 	if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
954 		dev_warn(nvmeq->dev->ctrl.device,
955 			"invalid id %d completed on queue %d\n",
956 			cqe->command_id, le16_to_cpu(cqe->sq_id));
957 		return;
958 	}
959 
960 	/*
961 	 * AEN requests are special as they don't time out and can
962 	 * survive any kind of queue freeze and often don't respond to
963 	 * aborts.  We don't even bother to allocate a struct request
964 	 * for them but rather special case them here.
965 	 */
966 	if (unlikely(nvmeq->qid == 0 &&
967 			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) {
968 		nvme_complete_async_event(&nvmeq->dev->ctrl,
969 				cqe->status, &cqe->result);
970 		return;
971 	}
972 
973 	req = blk_mq_tag_to_rq(*nvmeq->tags, cqe->command_id);
974 	trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
975 	nvme_end_request(req, cqe->status, cqe->result);
976 }
977 
978 static void nvme_complete_cqes(struct nvme_queue *nvmeq, u16 start, u16 end)
979 {
980 	while (start != end) {
981 		nvme_handle_cqe(nvmeq, start);
982 		if (++start == nvmeq->q_depth)
983 			start = 0;
984 	}
985 }
986 
987 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
988 {
989 	if (nvmeq->cq_head == nvmeq->q_depth - 1) {
990 		nvmeq->cq_head = 0;
991 		nvmeq->cq_phase = !nvmeq->cq_phase;
992 	} else {
993 		nvmeq->cq_head++;
994 	}
995 }
996 
997 static inline int nvme_process_cq(struct nvme_queue *nvmeq, u16 *start,
998 				  u16 *end, unsigned int tag)
999 {
1000 	int found = 0;
1001 
1002 	*start = nvmeq->cq_head;
1003 	while (nvme_cqe_pending(nvmeq)) {
1004 		if (tag == -1U || nvmeq->cqes[nvmeq->cq_head].command_id == tag)
1005 			found++;
1006 		nvme_update_cq_head(nvmeq);
1007 	}
1008 	*end = nvmeq->cq_head;
1009 
1010 	if (*start != *end)
1011 		nvme_ring_cq_doorbell(nvmeq);
1012 	return found;
1013 }
1014 
1015 static irqreturn_t nvme_irq(int irq, void *data)
1016 {
1017 	struct nvme_queue *nvmeq = data;
1018 	irqreturn_t ret = IRQ_NONE;
1019 	u16 start, end;
1020 
1021 	/*
1022 	 * The rmb/wmb pair ensures we see all updates from a previous run of
1023 	 * the irq handler, even if that was on another CPU.
1024 	 */
1025 	rmb();
1026 	if (nvmeq->cq_head != nvmeq->last_cq_head)
1027 		ret = IRQ_HANDLED;
1028 	nvme_process_cq(nvmeq, &start, &end, -1);
1029 	nvmeq->last_cq_head = nvmeq->cq_head;
1030 	wmb();
1031 
1032 	if (start != end) {
1033 		nvme_complete_cqes(nvmeq, start, end);
1034 		return IRQ_HANDLED;
1035 	}
1036 
1037 	return ret;
1038 }
1039 
1040 static irqreturn_t nvme_irq_check(int irq, void *data)
1041 {
1042 	struct nvme_queue *nvmeq = data;
1043 	if (nvme_cqe_pending(nvmeq))
1044 		return IRQ_WAKE_THREAD;
1045 	return IRQ_NONE;
1046 }
1047 
1048 /*
1049  * Poll for completions any queue, including those not dedicated to polling.
1050  * Can be called from any context.
1051  */
1052 static int nvme_poll_irqdisable(struct nvme_queue *nvmeq, unsigned int tag)
1053 {
1054 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1055 	u16 start, end;
1056 	int found;
1057 
1058 	/*
1059 	 * For a poll queue we need to protect against the polling thread
1060 	 * using the CQ lock.  For normal interrupt driven threads we have
1061 	 * to disable the interrupt to avoid racing with it.
1062 	 */
1063 	if (test_bit(NVMEQ_POLLED, &nvmeq->flags)) {
1064 		spin_lock(&nvmeq->cq_poll_lock);
1065 		found = nvme_process_cq(nvmeq, &start, &end, tag);
1066 		spin_unlock(&nvmeq->cq_poll_lock);
1067 	} else {
1068 		disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1069 		found = nvme_process_cq(nvmeq, &start, &end, tag);
1070 		enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1071 	}
1072 
1073 	nvme_complete_cqes(nvmeq, start, end);
1074 	return found;
1075 }
1076 
1077 static int nvme_poll(struct blk_mq_hw_ctx *hctx)
1078 {
1079 	struct nvme_queue *nvmeq = hctx->driver_data;
1080 	u16 start, end;
1081 	bool found;
1082 
1083 	if (!nvme_cqe_pending(nvmeq))
1084 		return 0;
1085 
1086 	spin_lock(&nvmeq->cq_poll_lock);
1087 	found = nvme_process_cq(nvmeq, &start, &end, -1);
1088 	spin_unlock(&nvmeq->cq_poll_lock);
1089 
1090 	nvme_complete_cqes(nvmeq, start, end);
1091 	return found;
1092 }
1093 
1094 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1095 {
1096 	struct nvme_dev *dev = to_nvme_dev(ctrl);
1097 	struct nvme_queue *nvmeq = &dev->queues[0];
1098 	struct nvme_command c;
1099 
1100 	memset(&c, 0, sizeof(c));
1101 	c.common.opcode = nvme_admin_async_event;
1102 	c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1103 	nvme_submit_cmd(nvmeq, &c, true);
1104 }
1105 
1106 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1107 {
1108 	struct nvme_command c;
1109 
1110 	memset(&c, 0, sizeof(c));
1111 	c.delete_queue.opcode = opcode;
1112 	c.delete_queue.qid = cpu_to_le16(id);
1113 
1114 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1115 }
1116 
1117 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1118 		struct nvme_queue *nvmeq, s16 vector)
1119 {
1120 	struct nvme_command c;
1121 	int flags = NVME_QUEUE_PHYS_CONTIG;
1122 
1123 	if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1124 		flags |= NVME_CQ_IRQ_ENABLED;
1125 
1126 	/*
1127 	 * Note: we (ab)use the fact that the prp fields survive if no data
1128 	 * is attached to the request.
1129 	 */
1130 	memset(&c, 0, sizeof(c));
1131 	c.create_cq.opcode = nvme_admin_create_cq;
1132 	c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1133 	c.create_cq.cqid = cpu_to_le16(qid);
1134 	c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1135 	c.create_cq.cq_flags = cpu_to_le16(flags);
1136 	c.create_cq.irq_vector = cpu_to_le16(vector);
1137 
1138 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1139 }
1140 
1141 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1142 						struct nvme_queue *nvmeq)
1143 {
1144 	struct nvme_ctrl *ctrl = &dev->ctrl;
1145 	struct nvme_command c;
1146 	int flags = NVME_QUEUE_PHYS_CONTIG;
1147 
1148 	/*
1149 	 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1150 	 * set. Since URGENT priority is zeroes, it makes all queues
1151 	 * URGENT.
1152 	 */
1153 	if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1154 		flags |= NVME_SQ_PRIO_MEDIUM;
1155 
1156 	/*
1157 	 * Note: we (ab)use the fact that the prp fields survive if no data
1158 	 * is attached to the request.
1159 	 */
1160 	memset(&c, 0, sizeof(c));
1161 	c.create_sq.opcode = nvme_admin_create_sq;
1162 	c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1163 	c.create_sq.sqid = cpu_to_le16(qid);
1164 	c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1165 	c.create_sq.sq_flags = cpu_to_le16(flags);
1166 	c.create_sq.cqid = cpu_to_le16(qid);
1167 
1168 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1169 }
1170 
1171 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1172 {
1173 	return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1174 }
1175 
1176 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1177 {
1178 	return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1179 }
1180 
1181 static void abort_endio(struct request *req, blk_status_t error)
1182 {
1183 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1184 	struct nvme_queue *nvmeq = iod->nvmeq;
1185 
1186 	dev_warn(nvmeq->dev->ctrl.device,
1187 		 "Abort status: 0x%x", nvme_req(req)->status);
1188 	atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1189 	blk_mq_free_request(req);
1190 }
1191 
1192 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1193 {
1194 
1195 	/* If true, indicates loss of adapter communication, possibly by a
1196 	 * NVMe Subsystem reset.
1197 	 */
1198 	bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1199 
1200 	/* If there is a reset/reinit ongoing, we shouldn't reset again. */
1201 	switch (dev->ctrl.state) {
1202 	case NVME_CTRL_RESETTING:
1203 	case NVME_CTRL_CONNECTING:
1204 		return false;
1205 	default:
1206 		break;
1207 	}
1208 
1209 	/* We shouldn't reset unless the controller is on fatal error state
1210 	 * _or_ if we lost the communication with it.
1211 	 */
1212 	if (!(csts & NVME_CSTS_CFS) && !nssro)
1213 		return false;
1214 
1215 	return true;
1216 }
1217 
1218 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1219 {
1220 	/* Read a config register to help see what died. */
1221 	u16 pci_status;
1222 	int result;
1223 
1224 	result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1225 				      &pci_status);
1226 	if (result == PCIBIOS_SUCCESSFUL)
1227 		dev_warn(dev->ctrl.device,
1228 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1229 			 csts, pci_status);
1230 	else
1231 		dev_warn(dev->ctrl.device,
1232 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1233 			 csts, result);
1234 }
1235 
1236 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1237 {
1238 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1239 	struct nvme_queue *nvmeq = iod->nvmeq;
1240 	struct nvme_dev *dev = nvmeq->dev;
1241 	struct request *abort_req;
1242 	struct nvme_command cmd;
1243 	u32 csts = readl(dev->bar + NVME_REG_CSTS);
1244 
1245 	/* If PCI error recovery process is happening, we cannot reset or
1246 	 * the recovery mechanism will surely fail.
1247 	 */
1248 	mb();
1249 	if (pci_channel_offline(to_pci_dev(dev->dev)))
1250 		return BLK_EH_RESET_TIMER;
1251 
1252 	/*
1253 	 * Reset immediately if the controller is failed
1254 	 */
1255 	if (nvme_should_reset(dev, csts)) {
1256 		nvme_warn_reset(dev, csts);
1257 		nvme_dev_disable(dev, false);
1258 		nvme_reset_ctrl(&dev->ctrl);
1259 		return BLK_EH_DONE;
1260 	}
1261 
1262 	/*
1263 	 * Did we miss an interrupt?
1264 	 */
1265 	if (nvme_poll_irqdisable(nvmeq, req->tag)) {
1266 		dev_warn(dev->ctrl.device,
1267 			 "I/O %d QID %d timeout, completion polled\n",
1268 			 req->tag, nvmeq->qid);
1269 		return BLK_EH_DONE;
1270 	}
1271 
1272 	/*
1273 	 * Shutdown immediately if controller times out while starting. The
1274 	 * reset work will see the pci device disabled when it gets the forced
1275 	 * cancellation error. All outstanding requests are completed on
1276 	 * shutdown, so we return BLK_EH_DONE.
1277 	 */
1278 	switch (dev->ctrl.state) {
1279 	case NVME_CTRL_CONNECTING:
1280 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1281 		/* fall through */
1282 	case NVME_CTRL_DELETING:
1283 		dev_warn_ratelimited(dev->ctrl.device,
1284 			 "I/O %d QID %d timeout, disable controller\n",
1285 			 req->tag, nvmeq->qid);
1286 		nvme_dev_disable(dev, true);
1287 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1288 		return BLK_EH_DONE;
1289 	case NVME_CTRL_RESETTING:
1290 		return BLK_EH_RESET_TIMER;
1291 	default:
1292 		break;
1293 	}
1294 
1295 	/*
1296  	 * Shutdown the controller immediately and schedule a reset if the
1297  	 * command was already aborted once before and still hasn't been
1298  	 * returned to the driver, or if this is the admin queue.
1299 	 */
1300 	if (!nvmeq->qid || iod->aborted) {
1301 		dev_warn(dev->ctrl.device,
1302 			 "I/O %d QID %d timeout, reset controller\n",
1303 			 req->tag, nvmeq->qid);
1304 		nvme_dev_disable(dev, false);
1305 		nvme_reset_ctrl(&dev->ctrl);
1306 
1307 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1308 		return BLK_EH_DONE;
1309 	}
1310 
1311 	if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1312 		atomic_inc(&dev->ctrl.abort_limit);
1313 		return BLK_EH_RESET_TIMER;
1314 	}
1315 	iod->aborted = 1;
1316 
1317 	memset(&cmd, 0, sizeof(cmd));
1318 	cmd.abort.opcode = nvme_admin_abort_cmd;
1319 	cmd.abort.cid = req->tag;
1320 	cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1321 
1322 	dev_warn(nvmeq->dev->ctrl.device,
1323 		"I/O %d QID %d timeout, aborting\n",
1324 		 req->tag, nvmeq->qid);
1325 
1326 	abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1327 			BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1328 	if (IS_ERR(abort_req)) {
1329 		atomic_inc(&dev->ctrl.abort_limit);
1330 		return BLK_EH_RESET_TIMER;
1331 	}
1332 
1333 	abort_req->timeout = ADMIN_TIMEOUT;
1334 	abort_req->end_io_data = NULL;
1335 	blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1336 
1337 	/*
1338 	 * The aborted req will be completed on receiving the abort req.
1339 	 * We enable the timer again. If hit twice, it'll cause a device reset,
1340 	 * as the device then is in a faulty state.
1341 	 */
1342 	return BLK_EH_RESET_TIMER;
1343 }
1344 
1345 static void nvme_free_queue(struct nvme_queue *nvmeq)
1346 {
1347 	dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq->q_depth),
1348 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1349 	if (!nvmeq->sq_cmds)
1350 		return;
1351 
1352 	if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1353 		pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1354 				nvmeq->sq_cmds, SQ_SIZE(nvmeq->q_depth));
1355 	} else {
1356 		dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq->q_depth),
1357 				nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1358 	}
1359 }
1360 
1361 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1362 {
1363 	int i;
1364 
1365 	for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1366 		dev->ctrl.queue_count--;
1367 		nvme_free_queue(&dev->queues[i]);
1368 	}
1369 }
1370 
1371 /**
1372  * nvme_suspend_queue - put queue into suspended state
1373  * @nvmeq: queue to suspend
1374  */
1375 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1376 {
1377 	if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1378 		return 1;
1379 
1380 	/* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1381 	mb();
1382 
1383 	nvmeq->dev->online_queues--;
1384 	if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1385 		blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1386 	if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1387 		pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1388 	return 0;
1389 }
1390 
1391 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1392 {
1393 	int i;
1394 
1395 	for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1396 		nvme_suspend_queue(&dev->queues[i]);
1397 }
1398 
1399 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1400 {
1401 	struct nvme_queue *nvmeq = &dev->queues[0];
1402 
1403 	if (shutdown)
1404 		nvme_shutdown_ctrl(&dev->ctrl);
1405 	else
1406 		nvme_disable_ctrl(&dev->ctrl, dev->ctrl.cap);
1407 
1408 	nvme_poll_irqdisable(nvmeq, -1);
1409 }
1410 
1411 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1412 				int entry_size)
1413 {
1414 	int q_depth = dev->q_depth;
1415 	unsigned q_size_aligned = roundup(q_depth * entry_size,
1416 					  dev->ctrl.page_size);
1417 
1418 	if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1419 		u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1420 		mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1421 		q_depth = div_u64(mem_per_q, entry_size);
1422 
1423 		/*
1424 		 * Ensure the reduced q_depth is above some threshold where it
1425 		 * would be better to map queues in system memory with the
1426 		 * original depth
1427 		 */
1428 		if (q_depth < 64)
1429 			return -ENOMEM;
1430 	}
1431 
1432 	return q_depth;
1433 }
1434 
1435 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1436 				int qid, int depth)
1437 {
1438 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1439 
1440 	if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1441 		nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(depth));
1442 		if (nvmeq->sq_cmds) {
1443 			nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1444 							nvmeq->sq_cmds);
1445 			if (nvmeq->sq_dma_addr) {
1446 				set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1447 				return 0;
1448 			}
1449 
1450 			pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(depth));
1451 		}
1452 	}
1453 
1454 	nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1455 				&nvmeq->sq_dma_addr, GFP_KERNEL);
1456 	if (!nvmeq->sq_cmds)
1457 		return -ENOMEM;
1458 	return 0;
1459 }
1460 
1461 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1462 {
1463 	struct nvme_queue *nvmeq = &dev->queues[qid];
1464 
1465 	if (dev->ctrl.queue_count > qid)
1466 		return 0;
1467 
1468 	nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(depth),
1469 					 &nvmeq->cq_dma_addr, GFP_KERNEL);
1470 	if (!nvmeq->cqes)
1471 		goto free_nvmeq;
1472 
1473 	if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
1474 		goto free_cqdma;
1475 
1476 	nvmeq->dev = dev;
1477 	spin_lock_init(&nvmeq->sq_lock);
1478 	spin_lock_init(&nvmeq->cq_poll_lock);
1479 	nvmeq->cq_head = 0;
1480 	nvmeq->cq_phase = 1;
1481 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1482 	nvmeq->q_depth = depth;
1483 	nvmeq->qid = qid;
1484 	dev->ctrl.queue_count++;
1485 
1486 	return 0;
1487 
1488  free_cqdma:
1489 	dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1490 							nvmeq->cq_dma_addr);
1491  free_nvmeq:
1492 	return -ENOMEM;
1493 }
1494 
1495 static int queue_request_irq(struct nvme_queue *nvmeq)
1496 {
1497 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1498 	int nr = nvmeq->dev->ctrl.instance;
1499 
1500 	if (use_threaded_interrupts) {
1501 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1502 				nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1503 	} else {
1504 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1505 				NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1506 	}
1507 }
1508 
1509 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1510 {
1511 	struct nvme_dev *dev = nvmeq->dev;
1512 
1513 	nvmeq->sq_tail = 0;
1514 	nvmeq->last_sq_tail = 0;
1515 	nvmeq->cq_head = 0;
1516 	nvmeq->cq_phase = 1;
1517 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1518 	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1519 	nvme_dbbuf_init(dev, nvmeq, qid);
1520 	dev->online_queues++;
1521 	wmb(); /* ensure the first interrupt sees the initialization */
1522 }
1523 
1524 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1525 {
1526 	struct nvme_dev *dev = nvmeq->dev;
1527 	int result;
1528 	u16 vector = 0;
1529 
1530 	clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1531 
1532 	/*
1533 	 * A queue's vector matches the queue identifier unless the controller
1534 	 * has only one vector available.
1535 	 */
1536 	if (!polled)
1537 		vector = dev->num_vecs == 1 ? 0 : qid;
1538 	else
1539 		set_bit(NVMEQ_POLLED, &nvmeq->flags);
1540 
1541 	result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1542 	if (result)
1543 		return result;
1544 
1545 	result = adapter_alloc_sq(dev, qid, nvmeq);
1546 	if (result < 0)
1547 		return result;
1548 	else if (result)
1549 		goto release_cq;
1550 
1551 	nvmeq->cq_vector = vector;
1552 	nvme_init_queue(nvmeq, qid);
1553 
1554 	if (!polled) {
1555 		nvmeq->cq_vector = vector;
1556 		result = queue_request_irq(nvmeq);
1557 		if (result < 0)
1558 			goto release_sq;
1559 	}
1560 
1561 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1562 	return result;
1563 
1564 release_sq:
1565 	dev->online_queues--;
1566 	adapter_delete_sq(dev, qid);
1567 release_cq:
1568 	adapter_delete_cq(dev, qid);
1569 	return result;
1570 }
1571 
1572 static const struct blk_mq_ops nvme_mq_admin_ops = {
1573 	.queue_rq	= nvme_queue_rq,
1574 	.complete	= nvme_pci_complete_rq,
1575 	.init_hctx	= nvme_admin_init_hctx,
1576 	.exit_hctx      = nvme_admin_exit_hctx,
1577 	.init_request	= nvme_init_request,
1578 	.timeout	= nvme_timeout,
1579 };
1580 
1581 static const struct blk_mq_ops nvme_mq_ops = {
1582 	.queue_rq	= nvme_queue_rq,
1583 	.complete	= nvme_pci_complete_rq,
1584 	.commit_rqs	= nvme_commit_rqs,
1585 	.init_hctx	= nvme_init_hctx,
1586 	.init_request	= nvme_init_request,
1587 	.map_queues	= nvme_pci_map_queues,
1588 	.timeout	= nvme_timeout,
1589 	.poll		= nvme_poll,
1590 };
1591 
1592 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1593 {
1594 	if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1595 		/*
1596 		 * If the controller was reset during removal, it's possible
1597 		 * user requests may be waiting on a stopped queue. Start the
1598 		 * queue to flush these to completion.
1599 		 */
1600 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1601 		blk_cleanup_queue(dev->ctrl.admin_q);
1602 		blk_mq_free_tag_set(&dev->admin_tagset);
1603 	}
1604 }
1605 
1606 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1607 {
1608 	if (!dev->ctrl.admin_q) {
1609 		dev->admin_tagset.ops = &nvme_mq_admin_ops;
1610 		dev->admin_tagset.nr_hw_queues = 1;
1611 
1612 		dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1613 		dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1614 		dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1615 		dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1616 		dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1617 		dev->admin_tagset.driver_data = dev;
1618 
1619 		if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1620 			return -ENOMEM;
1621 		dev->ctrl.admin_tagset = &dev->admin_tagset;
1622 
1623 		dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1624 		if (IS_ERR(dev->ctrl.admin_q)) {
1625 			blk_mq_free_tag_set(&dev->admin_tagset);
1626 			return -ENOMEM;
1627 		}
1628 		if (!blk_get_queue(dev->ctrl.admin_q)) {
1629 			nvme_dev_remove_admin(dev);
1630 			dev->ctrl.admin_q = NULL;
1631 			return -ENODEV;
1632 		}
1633 	} else
1634 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1635 
1636 	return 0;
1637 }
1638 
1639 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1640 {
1641 	return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1642 }
1643 
1644 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1645 {
1646 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1647 
1648 	if (size <= dev->bar_mapped_size)
1649 		return 0;
1650 	if (size > pci_resource_len(pdev, 0))
1651 		return -ENOMEM;
1652 	if (dev->bar)
1653 		iounmap(dev->bar);
1654 	dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1655 	if (!dev->bar) {
1656 		dev->bar_mapped_size = 0;
1657 		return -ENOMEM;
1658 	}
1659 	dev->bar_mapped_size = size;
1660 	dev->dbs = dev->bar + NVME_REG_DBS;
1661 
1662 	return 0;
1663 }
1664 
1665 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1666 {
1667 	int result;
1668 	u32 aqa;
1669 	struct nvme_queue *nvmeq;
1670 
1671 	result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1672 	if (result < 0)
1673 		return result;
1674 
1675 	dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1676 				NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1677 
1678 	if (dev->subsystem &&
1679 	    (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1680 		writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1681 
1682 	result = nvme_disable_ctrl(&dev->ctrl, dev->ctrl.cap);
1683 	if (result < 0)
1684 		return result;
1685 
1686 	result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1687 	if (result)
1688 		return result;
1689 
1690 	nvmeq = &dev->queues[0];
1691 	aqa = nvmeq->q_depth - 1;
1692 	aqa |= aqa << 16;
1693 
1694 	writel(aqa, dev->bar + NVME_REG_AQA);
1695 	lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1696 	lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1697 
1698 	result = nvme_enable_ctrl(&dev->ctrl, dev->ctrl.cap);
1699 	if (result)
1700 		return result;
1701 
1702 	nvmeq->cq_vector = 0;
1703 	nvme_init_queue(nvmeq, 0);
1704 	result = queue_request_irq(nvmeq);
1705 	if (result) {
1706 		dev->online_queues--;
1707 		return result;
1708 	}
1709 
1710 	set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1711 	return result;
1712 }
1713 
1714 static int nvme_create_io_queues(struct nvme_dev *dev)
1715 {
1716 	unsigned i, max, rw_queues;
1717 	int ret = 0;
1718 
1719 	for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1720 		if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1721 			ret = -ENOMEM;
1722 			break;
1723 		}
1724 	}
1725 
1726 	max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1727 	if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1728 		rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1729 				dev->io_queues[HCTX_TYPE_READ];
1730 	} else {
1731 		rw_queues = max;
1732 	}
1733 
1734 	for (i = dev->online_queues; i <= max; i++) {
1735 		bool polled = i > rw_queues;
1736 
1737 		ret = nvme_create_queue(&dev->queues[i], i, polled);
1738 		if (ret)
1739 			break;
1740 	}
1741 
1742 	/*
1743 	 * Ignore failing Create SQ/CQ commands, we can continue with less
1744 	 * than the desired amount of queues, and even a controller without
1745 	 * I/O queues can still be used to issue admin commands.  This might
1746 	 * be useful to upgrade a buggy firmware for example.
1747 	 */
1748 	return ret >= 0 ? 0 : ret;
1749 }
1750 
1751 static ssize_t nvme_cmb_show(struct device *dev,
1752 			     struct device_attribute *attr,
1753 			     char *buf)
1754 {
1755 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1756 
1757 	return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz  : x%08x\n",
1758 		       ndev->cmbloc, ndev->cmbsz);
1759 }
1760 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1761 
1762 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1763 {
1764 	u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1765 
1766 	return 1ULL << (12 + 4 * szu);
1767 }
1768 
1769 static u32 nvme_cmb_size(struct nvme_dev *dev)
1770 {
1771 	return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1772 }
1773 
1774 static void nvme_map_cmb(struct nvme_dev *dev)
1775 {
1776 	u64 size, offset;
1777 	resource_size_t bar_size;
1778 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1779 	int bar;
1780 
1781 	if (dev->cmb_size)
1782 		return;
1783 
1784 	dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1785 	if (!dev->cmbsz)
1786 		return;
1787 	dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1788 
1789 	size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1790 	offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1791 	bar = NVME_CMB_BIR(dev->cmbloc);
1792 	bar_size = pci_resource_len(pdev, bar);
1793 
1794 	if (offset > bar_size)
1795 		return;
1796 
1797 	/*
1798 	 * Controllers may support a CMB size larger than their BAR,
1799 	 * for example, due to being behind a bridge. Reduce the CMB to
1800 	 * the reported size of the BAR
1801 	 */
1802 	if (size > bar_size - offset)
1803 		size = bar_size - offset;
1804 
1805 	if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1806 		dev_warn(dev->ctrl.device,
1807 			 "failed to register the CMB\n");
1808 		return;
1809 	}
1810 
1811 	dev->cmb_size = size;
1812 	dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1813 
1814 	if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1815 			(NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1816 		pci_p2pmem_publish(pdev, true);
1817 
1818 	if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1819 				    &dev_attr_cmb.attr, NULL))
1820 		dev_warn(dev->ctrl.device,
1821 			 "failed to add sysfs attribute for CMB\n");
1822 }
1823 
1824 static inline void nvme_release_cmb(struct nvme_dev *dev)
1825 {
1826 	if (dev->cmb_size) {
1827 		sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1828 					     &dev_attr_cmb.attr, NULL);
1829 		dev->cmb_size = 0;
1830 	}
1831 }
1832 
1833 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1834 {
1835 	u64 dma_addr = dev->host_mem_descs_dma;
1836 	struct nvme_command c;
1837 	int ret;
1838 
1839 	memset(&c, 0, sizeof(c));
1840 	c.features.opcode	= nvme_admin_set_features;
1841 	c.features.fid		= cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1842 	c.features.dword11	= cpu_to_le32(bits);
1843 	c.features.dword12	= cpu_to_le32(dev->host_mem_size >>
1844 					      ilog2(dev->ctrl.page_size));
1845 	c.features.dword13	= cpu_to_le32(lower_32_bits(dma_addr));
1846 	c.features.dword14	= cpu_to_le32(upper_32_bits(dma_addr));
1847 	c.features.dword15	= cpu_to_le32(dev->nr_host_mem_descs);
1848 
1849 	ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1850 	if (ret) {
1851 		dev_warn(dev->ctrl.device,
1852 			 "failed to set host mem (err %d, flags %#x).\n",
1853 			 ret, bits);
1854 	}
1855 	return ret;
1856 }
1857 
1858 static void nvme_free_host_mem(struct nvme_dev *dev)
1859 {
1860 	int i;
1861 
1862 	for (i = 0; i < dev->nr_host_mem_descs; i++) {
1863 		struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1864 		size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1865 
1866 		dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
1867 			       le64_to_cpu(desc->addr),
1868 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1869 	}
1870 
1871 	kfree(dev->host_mem_desc_bufs);
1872 	dev->host_mem_desc_bufs = NULL;
1873 	dma_free_coherent(dev->dev,
1874 			dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1875 			dev->host_mem_descs, dev->host_mem_descs_dma);
1876 	dev->host_mem_descs = NULL;
1877 	dev->nr_host_mem_descs = 0;
1878 }
1879 
1880 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1881 		u32 chunk_size)
1882 {
1883 	struct nvme_host_mem_buf_desc *descs;
1884 	u32 max_entries, len;
1885 	dma_addr_t descs_dma;
1886 	int i = 0;
1887 	void **bufs;
1888 	u64 size, tmp;
1889 
1890 	tmp = (preferred + chunk_size - 1);
1891 	do_div(tmp, chunk_size);
1892 	max_entries = tmp;
1893 
1894 	if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1895 		max_entries = dev->ctrl.hmmaxd;
1896 
1897 	descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
1898 				   &descs_dma, GFP_KERNEL);
1899 	if (!descs)
1900 		goto out;
1901 
1902 	bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1903 	if (!bufs)
1904 		goto out_free_descs;
1905 
1906 	for (size = 0; size < preferred && i < max_entries; size += len) {
1907 		dma_addr_t dma_addr;
1908 
1909 		len = min_t(u64, chunk_size, preferred - size);
1910 		bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1911 				DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1912 		if (!bufs[i])
1913 			break;
1914 
1915 		descs[i].addr = cpu_to_le64(dma_addr);
1916 		descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1917 		i++;
1918 	}
1919 
1920 	if (!size)
1921 		goto out_free_bufs;
1922 
1923 	dev->nr_host_mem_descs = i;
1924 	dev->host_mem_size = size;
1925 	dev->host_mem_descs = descs;
1926 	dev->host_mem_descs_dma = descs_dma;
1927 	dev->host_mem_desc_bufs = bufs;
1928 	return 0;
1929 
1930 out_free_bufs:
1931 	while (--i >= 0) {
1932 		size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1933 
1934 		dma_free_attrs(dev->dev, size, bufs[i],
1935 			       le64_to_cpu(descs[i].addr),
1936 			       DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1937 	}
1938 
1939 	kfree(bufs);
1940 out_free_descs:
1941 	dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
1942 			descs_dma);
1943 out:
1944 	dev->host_mem_descs = NULL;
1945 	return -ENOMEM;
1946 }
1947 
1948 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1949 {
1950 	u32 chunk_size;
1951 
1952 	/* start big and work our way down */
1953 	for (chunk_size = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
1954 	     chunk_size >= max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
1955 	     chunk_size /= 2) {
1956 		if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
1957 			if (!min || dev->host_mem_size >= min)
1958 				return 0;
1959 			nvme_free_host_mem(dev);
1960 		}
1961 	}
1962 
1963 	return -ENOMEM;
1964 }
1965 
1966 static int nvme_setup_host_mem(struct nvme_dev *dev)
1967 {
1968 	u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1969 	u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1970 	u64 min = (u64)dev->ctrl.hmmin * 4096;
1971 	u32 enable_bits = NVME_HOST_MEM_ENABLE;
1972 	int ret;
1973 
1974 	preferred = min(preferred, max);
1975 	if (min > max) {
1976 		dev_warn(dev->ctrl.device,
1977 			"min host memory (%lld MiB) above limit (%d MiB).\n",
1978 			min >> ilog2(SZ_1M), max_host_mem_size_mb);
1979 		nvme_free_host_mem(dev);
1980 		return 0;
1981 	}
1982 
1983 	/*
1984 	 * If we already have a buffer allocated check if we can reuse it.
1985 	 */
1986 	if (dev->host_mem_descs) {
1987 		if (dev->host_mem_size >= min)
1988 			enable_bits |= NVME_HOST_MEM_RETURN;
1989 		else
1990 			nvme_free_host_mem(dev);
1991 	}
1992 
1993 	if (!dev->host_mem_descs) {
1994 		if (nvme_alloc_host_mem(dev, min, preferred)) {
1995 			dev_warn(dev->ctrl.device,
1996 				"failed to allocate host memory buffer.\n");
1997 			return 0; /* controller must work without HMB */
1998 		}
1999 
2000 		dev_info(dev->ctrl.device,
2001 			"allocated %lld MiB host memory buffer.\n",
2002 			dev->host_mem_size >> ilog2(SZ_1M));
2003 	}
2004 
2005 	ret = nvme_set_host_mem(dev, enable_bits);
2006 	if (ret)
2007 		nvme_free_host_mem(dev);
2008 	return ret;
2009 }
2010 
2011 /*
2012  * nirqs is the number of interrupts available for write and read
2013  * queues. The core already reserved an interrupt for the admin queue.
2014  */
2015 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2016 {
2017 	struct nvme_dev *dev = affd->priv;
2018 	unsigned int nr_read_queues;
2019 
2020 	/*
2021 	 * If there is no interupt available for queues, ensure that
2022 	 * the default queue is set to 1. The affinity set size is
2023 	 * also set to one, but the irq core ignores it for this case.
2024 	 *
2025 	 * If only one interrupt is available or 'write_queue' == 0, combine
2026 	 * write and read queues.
2027 	 *
2028 	 * If 'write_queues' > 0, ensure it leaves room for at least one read
2029 	 * queue.
2030 	 */
2031 	if (!nrirqs) {
2032 		nrirqs = 1;
2033 		nr_read_queues = 0;
2034 	} else if (nrirqs == 1 || !write_queues) {
2035 		nr_read_queues = 0;
2036 	} else if (write_queues >= nrirqs) {
2037 		nr_read_queues = 1;
2038 	} else {
2039 		nr_read_queues = nrirqs - write_queues;
2040 	}
2041 
2042 	dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2043 	affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2044 	dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2045 	affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2046 	affd->nr_sets = nr_read_queues ? 2 : 1;
2047 }
2048 
2049 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2050 {
2051 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2052 	struct irq_affinity affd = {
2053 		.pre_vectors	= 1,
2054 		.calc_sets	= nvme_calc_irq_sets,
2055 		.priv		= dev,
2056 	};
2057 	unsigned int irq_queues, this_p_queues;
2058 	unsigned int nr_cpus = num_possible_cpus();
2059 
2060 	/*
2061 	 * Poll queues don't need interrupts, but we need at least one IO
2062 	 * queue left over for non-polled IO.
2063 	 */
2064 	this_p_queues = poll_queues;
2065 	if (this_p_queues >= nr_io_queues) {
2066 		this_p_queues = nr_io_queues - 1;
2067 		irq_queues = 1;
2068 	} else {
2069 		if (nr_cpus < nr_io_queues - this_p_queues)
2070 			irq_queues = nr_cpus + 1;
2071 		else
2072 			irq_queues = nr_io_queues - this_p_queues + 1;
2073 	}
2074 	dev->io_queues[HCTX_TYPE_POLL] = this_p_queues;
2075 
2076 	/* Initialize for the single interrupt case */
2077 	dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2078 	dev->io_queues[HCTX_TYPE_READ] = 0;
2079 
2080 	return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2081 			      PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2082 }
2083 
2084 static void nvme_disable_io_queues(struct nvme_dev *dev)
2085 {
2086 	if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2087 		__nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2088 }
2089 
2090 static int nvme_setup_io_queues(struct nvme_dev *dev)
2091 {
2092 	struct nvme_queue *adminq = &dev->queues[0];
2093 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2094 	int result, nr_io_queues;
2095 	unsigned long size;
2096 
2097 	nr_io_queues = max_io_queues();
2098 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2099 	if (result < 0)
2100 		return result;
2101 
2102 	if (nr_io_queues == 0)
2103 		return 0;
2104 
2105 	clear_bit(NVMEQ_ENABLED, &adminq->flags);
2106 
2107 	if (dev->cmb_use_sqes) {
2108 		result = nvme_cmb_qdepth(dev, nr_io_queues,
2109 				sizeof(struct nvme_command));
2110 		if (result > 0)
2111 			dev->q_depth = result;
2112 		else
2113 			dev->cmb_use_sqes = false;
2114 	}
2115 
2116 	do {
2117 		size = db_bar_size(dev, nr_io_queues);
2118 		result = nvme_remap_bar(dev, size);
2119 		if (!result)
2120 			break;
2121 		if (!--nr_io_queues)
2122 			return -ENOMEM;
2123 	} while (1);
2124 	adminq->q_db = dev->dbs;
2125 
2126  retry:
2127 	/* Deregister the admin queue's interrupt */
2128 	pci_free_irq(pdev, 0, adminq);
2129 
2130 	/*
2131 	 * If we enable msix early due to not intx, disable it again before
2132 	 * setting up the full range we need.
2133 	 */
2134 	pci_free_irq_vectors(pdev);
2135 
2136 	result = nvme_setup_irqs(dev, nr_io_queues);
2137 	if (result <= 0)
2138 		return -EIO;
2139 
2140 	dev->num_vecs = result;
2141 	result = max(result - 1, 1);
2142 	dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2143 
2144 	/*
2145 	 * Should investigate if there's a performance win from allocating
2146 	 * more queues than interrupt vectors; it might allow the submission
2147 	 * path to scale better, even if the receive path is limited by the
2148 	 * number of interrupts.
2149 	 */
2150 	result = queue_request_irq(adminq);
2151 	if (result)
2152 		return result;
2153 	set_bit(NVMEQ_ENABLED, &adminq->flags);
2154 
2155 	result = nvme_create_io_queues(dev);
2156 	if (result || dev->online_queues < 2)
2157 		return result;
2158 
2159 	if (dev->online_queues - 1 < dev->max_qid) {
2160 		nr_io_queues = dev->online_queues - 1;
2161 		nvme_disable_io_queues(dev);
2162 		nvme_suspend_io_queues(dev);
2163 		goto retry;
2164 	}
2165 	dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2166 					dev->io_queues[HCTX_TYPE_DEFAULT],
2167 					dev->io_queues[HCTX_TYPE_READ],
2168 					dev->io_queues[HCTX_TYPE_POLL]);
2169 	return 0;
2170 }
2171 
2172 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2173 {
2174 	struct nvme_queue *nvmeq = req->end_io_data;
2175 
2176 	blk_mq_free_request(req);
2177 	complete(&nvmeq->delete_done);
2178 }
2179 
2180 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2181 {
2182 	struct nvme_queue *nvmeq = req->end_io_data;
2183 
2184 	if (error)
2185 		set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2186 
2187 	nvme_del_queue_end(req, error);
2188 }
2189 
2190 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2191 {
2192 	struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2193 	struct request *req;
2194 	struct nvme_command cmd;
2195 
2196 	memset(&cmd, 0, sizeof(cmd));
2197 	cmd.delete_queue.opcode = opcode;
2198 	cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2199 
2200 	req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
2201 	if (IS_ERR(req))
2202 		return PTR_ERR(req);
2203 
2204 	req->timeout = ADMIN_TIMEOUT;
2205 	req->end_io_data = nvmeq;
2206 
2207 	init_completion(&nvmeq->delete_done);
2208 	blk_execute_rq_nowait(q, NULL, req, false,
2209 			opcode == nvme_admin_delete_cq ?
2210 				nvme_del_cq_end : nvme_del_queue_end);
2211 	return 0;
2212 }
2213 
2214 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2215 {
2216 	int nr_queues = dev->online_queues - 1, sent = 0;
2217 	unsigned long timeout;
2218 
2219  retry:
2220 	timeout = ADMIN_TIMEOUT;
2221 	while (nr_queues > 0) {
2222 		if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2223 			break;
2224 		nr_queues--;
2225 		sent++;
2226 	}
2227 	while (sent) {
2228 		struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2229 
2230 		timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2231 				timeout);
2232 		if (timeout == 0)
2233 			return false;
2234 
2235 		/* handle any remaining CQEs */
2236 		if (opcode == nvme_admin_delete_cq &&
2237 		    !test_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags))
2238 			nvme_poll_irqdisable(nvmeq, -1);
2239 
2240 		sent--;
2241 		if (nr_queues)
2242 			goto retry;
2243 	}
2244 	return true;
2245 }
2246 
2247 /*
2248  * return error value only when tagset allocation failed
2249  */
2250 static int nvme_dev_add(struct nvme_dev *dev)
2251 {
2252 	int ret;
2253 
2254 	if (!dev->ctrl.tagset) {
2255 		dev->tagset.ops = &nvme_mq_ops;
2256 		dev->tagset.nr_hw_queues = dev->online_queues - 1;
2257 		dev->tagset.nr_maps = 1; /* default */
2258 		if (dev->io_queues[HCTX_TYPE_READ])
2259 			dev->tagset.nr_maps++;
2260 		if (dev->io_queues[HCTX_TYPE_POLL])
2261 			dev->tagset.nr_maps++;
2262 		dev->tagset.timeout = NVME_IO_TIMEOUT;
2263 		dev->tagset.numa_node = dev_to_node(dev->dev);
2264 		dev->tagset.queue_depth =
2265 				min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2266 		dev->tagset.cmd_size = sizeof(struct nvme_iod);
2267 		dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2268 		dev->tagset.driver_data = dev;
2269 
2270 		ret = blk_mq_alloc_tag_set(&dev->tagset);
2271 		if (ret) {
2272 			dev_warn(dev->ctrl.device,
2273 				"IO queues tagset allocation failed %d\n", ret);
2274 			return ret;
2275 		}
2276 		dev->ctrl.tagset = &dev->tagset;
2277 	} else {
2278 		blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2279 
2280 		/* Free previously allocated queues that are no longer usable */
2281 		nvme_free_queues(dev, dev->online_queues);
2282 	}
2283 
2284 	nvme_dbbuf_set(dev);
2285 	return 0;
2286 }
2287 
2288 static int nvme_pci_enable(struct nvme_dev *dev)
2289 {
2290 	int result = -ENOMEM;
2291 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2292 
2293 	if (pci_enable_device_mem(pdev))
2294 		return result;
2295 
2296 	pci_set_master(pdev);
2297 
2298 	if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)))
2299 		goto disable;
2300 
2301 	if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2302 		result = -ENODEV;
2303 		goto disable;
2304 	}
2305 
2306 	/*
2307 	 * Some devices and/or platforms don't advertise or work with INTx
2308 	 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2309 	 * adjust this later.
2310 	 */
2311 	result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2312 	if (result < 0)
2313 		return result;
2314 
2315 	dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2316 
2317 	dev->q_depth = min_t(int, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2318 				io_queue_depth);
2319 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2320 	dev->dbs = dev->bar + 4096;
2321 
2322 	/*
2323 	 * Temporary fix for the Apple controller found in the MacBook8,1 and
2324 	 * some MacBook7,1 to avoid controller resets and data loss.
2325 	 */
2326 	if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2327 		dev->q_depth = 2;
2328 		dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2329 			"set queue depth=%u to work around controller resets\n",
2330 			dev->q_depth);
2331 	} else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2332 		   (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2333 		   NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2334 		dev->q_depth = 64;
2335 		dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2336                         "set queue depth=%u\n", dev->q_depth);
2337 	}
2338 
2339 	nvme_map_cmb(dev);
2340 
2341 	pci_enable_pcie_error_reporting(pdev);
2342 	pci_save_state(pdev);
2343 	return 0;
2344 
2345  disable:
2346 	pci_disable_device(pdev);
2347 	return result;
2348 }
2349 
2350 static void nvme_dev_unmap(struct nvme_dev *dev)
2351 {
2352 	if (dev->bar)
2353 		iounmap(dev->bar);
2354 	pci_release_mem_regions(to_pci_dev(dev->dev));
2355 }
2356 
2357 static void nvme_pci_disable(struct nvme_dev *dev)
2358 {
2359 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2360 
2361 	pci_free_irq_vectors(pdev);
2362 
2363 	if (pci_is_enabled(pdev)) {
2364 		pci_disable_pcie_error_reporting(pdev);
2365 		pci_disable_device(pdev);
2366 	}
2367 }
2368 
2369 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2370 {
2371 	bool dead = true, freeze = false;
2372 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2373 
2374 	mutex_lock(&dev->shutdown_lock);
2375 	if (pci_is_enabled(pdev)) {
2376 		u32 csts = readl(dev->bar + NVME_REG_CSTS);
2377 
2378 		if (dev->ctrl.state == NVME_CTRL_LIVE ||
2379 		    dev->ctrl.state == NVME_CTRL_RESETTING) {
2380 			freeze = true;
2381 			nvme_start_freeze(&dev->ctrl);
2382 		}
2383 		dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2384 			pdev->error_state  != pci_channel_io_normal);
2385 	}
2386 
2387 	/*
2388 	 * Give the controller a chance to complete all entered requests if
2389 	 * doing a safe shutdown.
2390 	 */
2391 	if (!dead && shutdown && freeze)
2392 		nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2393 
2394 	nvme_stop_queues(&dev->ctrl);
2395 
2396 	if (!dead && dev->ctrl.queue_count > 0) {
2397 		nvme_disable_io_queues(dev);
2398 		nvme_disable_admin_queue(dev, shutdown);
2399 	}
2400 	nvme_suspend_io_queues(dev);
2401 	nvme_suspend_queue(&dev->queues[0]);
2402 	nvme_pci_disable(dev);
2403 
2404 	blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2405 	blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2406 
2407 	/*
2408 	 * The driver will not be starting up queues again if shutting down so
2409 	 * must flush all entered requests to their failed completion to avoid
2410 	 * deadlocking blk-mq hot-cpu notifier.
2411 	 */
2412 	if (shutdown) {
2413 		nvme_start_queues(&dev->ctrl);
2414 		if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2415 			blk_mq_unquiesce_queue(dev->ctrl.admin_q);
2416 	}
2417 	mutex_unlock(&dev->shutdown_lock);
2418 }
2419 
2420 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2421 {
2422 	dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2423 						PAGE_SIZE, PAGE_SIZE, 0);
2424 	if (!dev->prp_page_pool)
2425 		return -ENOMEM;
2426 
2427 	/* Optimisation for I/Os between 4k and 128k */
2428 	dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2429 						256, 256, 0);
2430 	if (!dev->prp_small_pool) {
2431 		dma_pool_destroy(dev->prp_page_pool);
2432 		return -ENOMEM;
2433 	}
2434 	return 0;
2435 }
2436 
2437 static void nvme_release_prp_pools(struct nvme_dev *dev)
2438 {
2439 	dma_pool_destroy(dev->prp_page_pool);
2440 	dma_pool_destroy(dev->prp_small_pool);
2441 }
2442 
2443 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2444 {
2445 	struct nvme_dev *dev = to_nvme_dev(ctrl);
2446 
2447 	nvme_dbbuf_dma_free(dev);
2448 	put_device(dev->dev);
2449 	if (dev->tagset.tags)
2450 		blk_mq_free_tag_set(&dev->tagset);
2451 	if (dev->ctrl.admin_q)
2452 		blk_put_queue(dev->ctrl.admin_q);
2453 	kfree(dev->queues);
2454 	free_opal_dev(dev->ctrl.opal_dev);
2455 	mempool_destroy(dev->iod_mempool);
2456 	kfree(dev);
2457 }
2458 
2459 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2460 {
2461 	nvme_get_ctrl(&dev->ctrl);
2462 	nvme_dev_disable(dev, false);
2463 	nvme_kill_queues(&dev->ctrl);
2464 	if (!queue_work(nvme_wq, &dev->remove_work))
2465 		nvme_put_ctrl(&dev->ctrl);
2466 }
2467 
2468 static void nvme_reset_work(struct work_struct *work)
2469 {
2470 	struct nvme_dev *dev =
2471 		container_of(work, struct nvme_dev, ctrl.reset_work);
2472 	bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2473 	int result;
2474 	enum nvme_ctrl_state new_state = NVME_CTRL_LIVE;
2475 
2476 	if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING)) {
2477 		result = -ENODEV;
2478 		goto out;
2479 	}
2480 
2481 	/*
2482 	 * If we're called to reset a live controller first shut it down before
2483 	 * moving on.
2484 	 */
2485 	if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2486 		nvme_dev_disable(dev, false);
2487 	nvme_sync_queues(&dev->ctrl);
2488 
2489 	mutex_lock(&dev->shutdown_lock);
2490 	result = nvme_pci_enable(dev);
2491 	if (result)
2492 		goto out_unlock;
2493 
2494 	result = nvme_pci_configure_admin_queue(dev);
2495 	if (result)
2496 		goto out_unlock;
2497 
2498 	result = nvme_alloc_admin_tags(dev);
2499 	if (result)
2500 		goto out_unlock;
2501 
2502 	/*
2503 	 * Limit the max command size to prevent iod->sg allocations going
2504 	 * over a single page.
2505 	 */
2506 	dev->ctrl.max_hw_sectors = min_t(u32,
2507 		NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2508 	dev->ctrl.max_segments = NVME_MAX_SEGS;
2509 
2510 	/*
2511 	 * Don't limit the IOMMU merged segment size.
2512 	 */
2513 	dma_set_max_seg_size(dev->dev, 0xffffffff);
2514 
2515 	mutex_unlock(&dev->shutdown_lock);
2516 
2517 	/*
2518 	 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2519 	 * initializing procedure here.
2520 	 */
2521 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2522 		dev_warn(dev->ctrl.device,
2523 			"failed to mark controller CONNECTING\n");
2524 		result = -EBUSY;
2525 		goto out;
2526 	}
2527 
2528 	result = nvme_init_identify(&dev->ctrl);
2529 	if (result)
2530 		goto out;
2531 
2532 	if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2533 		if (!dev->ctrl.opal_dev)
2534 			dev->ctrl.opal_dev =
2535 				init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2536 		else if (was_suspend)
2537 			opal_unlock_from_suspend(dev->ctrl.opal_dev);
2538 	} else {
2539 		free_opal_dev(dev->ctrl.opal_dev);
2540 		dev->ctrl.opal_dev = NULL;
2541 	}
2542 
2543 	if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2544 		result = nvme_dbbuf_dma_alloc(dev);
2545 		if (result)
2546 			dev_warn(dev->dev,
2547 				 "unable to allocate dma for dbbuf\n");
2548 	}
2549 
2550 	if (dev->ctrl.hmpre) {
2551 		result = nvme_setup_host_mem(dev);
2552 		if (result < 0)
2553 			goto out;
2554 	}
2555 
2556 	result = nvme_setup_io_queues(dev);
2557 	if (result)
2558 		goto out;
2559 
2560 	/*
2561 	 * Keep the controller around but remove all namespaces if we don't have
2562 	 * any working I/O queue.
2563 	 */
2564 	if (dev->online_queues < 2) {
2565 		dev_warn(dev->ctrl.device, "IO queues not created\n");
2566 		nvme_kill_queues(&dev->ctrl);
2567 		nvme_remove_namespaces(&dev->ctrl);
2568 		new_state = NVME_CTRL_ADMIN_ONLY;
2569 	} else {
2570 		nvme_start_queues(&dev->ctrl);
2571 		nvme_wait_freeze(&dev->ctrl);
2572 		/* hit this only when allocate tagset fails */
2573 		if (nvme_dev_add(dev))
2574 			new_state = NVME_CTRL_ADMIN_ONLY;
2575 		nvme_unfreeze(&dev->ctrl);
2576 	}
2577 
2578 	/*
2579 	 * If only admin queue live, keep it to do further investigation or
2580 	 * recovery.
2581 	 */
2582 	if (!nvme_change_ctrl_state(&dev->ctrl, new_state)) {
2583 		dev_warn(dev->ctrl.device,
2584 			"failed to mark controller state %d\n", new_state);
2585 		result = -ENODEV;
2586 		goto out;
2587 	}
2588 
2589 	nvme_start_ctrl(&dev->ctrl);
2590 	return;
2591 
2592  out_unlock:
2593 	mutex_unlock(&dev->shutdown_lock);
2594  out:
2595 	if (result)
2596 		dev_warn(dev->ctrl.device,
2597 			 "Removing after probe failure status: %d\n", result);
2598 	nvme_remove_dead_ctrl(dev);
2599 }
2600 
2601 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2602 {
2603 	struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2604 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2605 
2606 	if (pci_get_drvdata(pdev))
2607 		device_release_driver(&pdev->dev);
2608 	nvme_put_ctrl(&dev->ctrl);
2609 }
2610 
2611 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2612 {
2613 	*val = readl(to_nvme_dev(ctrl)->bar + off);
2614 	return 0;
2615 }
2616 
2617 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2618 {
2619 	writel(val, to_nvme_dev(ctrl)->bar + off);
2620 	return 0;
2621 }
2622 
2623 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2624 {
2625 	*val = readq(to_nvme_dev(ctrl)->bar + off);
2626 	return 0;
2627 }
2628 
2629 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2630 {
2631 	struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2632 
2633 	return snprintf(buf, size, "%s", dev_name(&pdev->dev));
2634 }
2635 
2636 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2637 	.name			= "pcie",
2638 	.module			= THIS_MODULE,
2639 	.flags			= NVME_F_METADATA_SUPPORTED |
2640 				  NVME_F_PCI_P2PDMA,
2641 	.reg_read32		= nvme_pci_reg_read32,
2642 	.reg_write32		= nvme_pci_reg_write32,
2643 	.reg_read64		= nvme_pci_reg_read64,
2644 	.free_ctrl		= nvme_pci_free_ctrl,
2645 	.submit_async_event	= nvme_pci_submit_async_event,
2646 	.get_address		= nvme_pci_get_address,
2647 };
2648 
2649 static int nvme_dev_map(struct nvme_dev *dev)
2650 {
2651 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2652 
2653 	if (pci_request_mem_regions(pdev, "nvme"))
2654 		return -ENODEV;
2655 
2656 	if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2657 		goto release;
2658 
2659 	return 0;
2660   release:
2661 	pci_release_mem_regions(pdev);
2662 	return -ENODEV;
2663 }
2664 
2665 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2666 {
2667 	if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2668 		/*
2669 		 * Several Samsung devices seem to drop off the PCIe bus
2670 		 * randomly when APST is on and uses the deepest sleep state.
2671 		 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2672 		 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2673 		 * 950 PRO 256GB", but it seems to be restricted to two Dell
2674 		 * laptops.
2675 		 */
2676 		if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2677 		    (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2678 		     dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2679 			return NVME_QUIRK_NO_DEEPEST_PS;
2680 	} else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2681 		/*
2682 		 * Samsung SSD 960 EVO drops off the PCIe bus after system
2683 		 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2684 		 * within few minutes after bootup on a Coffee Lake board -
2685 		 * ASUS PRIME Z370-A
2686 		 */
2687 		if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2688 		    (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2689 		     dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2690 			return NVME_QUIRK_NO_APST;
2691 	}
2692 
2693 	return 0;
2694 }
2695 
2696 static void nvme_async_probe(void *data, async_cookie_t cookie)
2697 {
2698 	struct nvme_dev *dev = data;
2699 
2700 	nvme_reset_ctrl_sync(&dev->ctrl);
2701 	flush_work(&dev->ctrl.scan_work);
2702 	nvme_put_ctrl(&dev->ctrl);
2703 }
2704 
2705 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2706 {
2707 	int node, result = -ENOMEM;
2708 	struct nvme_dev *dev;
2709 	unsigned long quirks = id->driver_data;
2710 	size_t alloc_size;
2711 
2712 	node = dev_to_node(&pdev->dev);
2713 	if (node == NUMA_NO_NODE)
2714 		set_dev_node(&pdev->dev, first_memory_node);
2715 
2716 	dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2717 	if (!dev)
2718 		return -ENOMEM;
2719 
2720 	dev->queues = kcalloc_node(max_queue_count(), sizeof(struct nvme_queue),
2721 					GFP_KERNEL, node);
2722 	if (!dev->queues)
2723 		goto free;
2724 
2725 	dev->dev = get_device(&pdev->dev);
2726 	pci_set_drvdata(pdev, dev);
2727 
2728 	result = nvme_dev_map(dev);
2729 	if (result)
2730 		goto put_pci;
2731 
2732 	INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2733 	INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2734 	mutex_init(&dev->shutdown_lock);
2735 
2736 	result = nvme_setup_prp_pools(dev);
2737 	if (result)
2738 		goto unmap;
2739 
2740 	quirks |= check_vendor_combination_bug(pdev);
2741 
2742 	/*
2743 	 * Double check that our mempool alloc size will cover the biggest
2744 	 * command we support.
2745 	 */
2746 	alloc_size = nvme_pci_iod_alloc_size(dev, NVME_MAX_KB_SZ,
2747 						NVME_MAX_SEGS, true);
2748 	WARN_ON_ONCE(alloc_size > PAGE_SIZE);
2749 
2750 	dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
2751 						mempool_kfree,
2752 						(void *) alloc_size,
2753 						GFP_KERNEL, node);
2754 	if (!dev->iod_mempool) {
2755 		result = -ENOMEM;
2756 		goto release_pools;
2757 	}
2758 
2759 	result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2760 			quirks);
2761 	if (result)
2762 		goto release_mempool;
2763 
2764 	dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2765 
2766 	nvme_get_ctrl(&dev->ctrl);
2767 	async_schedule(nvme_async_probe, dev);
2768 
2769 	return 0;
2770 
2771  release_mempool:
2772 	mempool_destroy(dev->iod_mempool);
2773  release_pools:
2774 	nvme_release_prp_pools(dev);
2775  unmap:
2776 	nvme_dev_unmap(dev);
2777  put_pci:
2778 	put_device(dev->dev);
2779  free:
2780 	kfree(dev->queues);
2781 	kfree(dev);
2782 	return result;
2783 }
2784 
2785 static void nvme_reset_prepare(struct pci_dev *pdev)
2786 {
2787 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2788 	nvme_dev_disable(dev, false);
2789 }
2790 
2791 static void nvme_reset_done(struct pci_dev *pdev)
2792 {
2793 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2794 	nvme_reset_ctrl_sync(&dev->ctrl);
2795 }
2796 
2797 static void nvme_shutdown(struct pci_dev *pdev)
2798 {
2799 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2800 	nvme_dev_disable(dev, true);
2801 }
2802 
2803 /*
2804  * The driver's remove may be called on a device in a partially initialized
2805  * state. This function must not have any dependencies on the device state in
2806  * order to proceed.
2807  */
2808 static void nvme_remove(struct pci_dev *pdev)
2809 {
2810 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2811 
2812 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2813 	pci_set_drvdata(pdev, NULL);
2814 
2815 	if (!pci_device_is_present(pdev)) {
2816 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2817 		nvme_dev_disable(dev, true);
2818 		nvme_dev_remove_admin(dev);
2819 	}
2820 
2821 	flush_work(&dev->ctrl.reset_work);
2822 	nvme_stop_ctrl(&dev->ctrl);
2823 	nvme_remove_namespaces(&dev->ctrl);
2824 	nvme_dev_disable(dev, true);
2825 	nvme_release_cmb(dev);
2826 	nvme_free_host_mem(dev);
2827 	nvme_dev_remove_admin(dev);
2828 	nvme_free_queues(dev, 0);
2829 	nvme_uninit_ctrl(&dev->ctrl);
2830 	nvme_release_prp_pools(dev);
2831 	nvme_dev_unmap(dev);
2832 	nvme_put_ctrl(&dev->ctrl);
2833 }
2834 
2835 #ifdef CONFIG_PM_SLEEP
2836 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
2837 {
2838 	return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
2839 }
2840 
2841 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
2842 {
2843 	return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
2844 }
2845 
2846 static int nvme_resume(struct device *dev)
2847 {
2848 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
2849 	struct nvme_ctrl *ctrl = &ndev->ctrl;
2850 
2851 	if (pm_resume_via_firmware() || !ctrl->npss ||
2852 	    nvme_set_power_state(ctrl, ndev->last_ps) != 0)
2853 		nvme_reset_ctrl(ctrl);
2854 	return 0;
2855 }
2856 
2857 static int nvme_suspend(struct device *dev)
2858 {
2859 	struct pci_dev *pdev = to_pci_dev(dev);
2860 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2861 	struct nvme_ctrl *ctrl = &ndev->ctrl;
2862 	int ret = -EBUSY;
2863 
2864 	/*
2865 	 * The platform does not remove power for a kernel managed suspend so
2866 	 * use host managed nvme power settings for lowest idle power if
2867 	 * possible. This should have quicker resume latency than a full device
2868 	 * shutdown.  But if the firmware is involved after the suspend or the
2869 	 * device does not support any non-default power states, shut down the
2870 	 * device fully.
2871 	 */
2872 	if (pm_suspend_via_firmware() || !ctrl->npss) {
2873 		nvme_dev_disable(ndev, true);
2874 		return 0;
2875 	}
2876 
2877 	nvme_start_freeze(ctrl);
2878 	nvme_wait_freeze(ctrl);
2879 	nvme_sync_queues(ctrl);
2880 
2881 	if (ctrl->state != NVME_CTRL_LIVE &&
2882 	    ctrl->state != NVME_CTRL_ADMIN_ONLY)
2883 		goto unfreeze;
2884 
2885 	ndev->last_ps = 0;
2886 	ret = nvme_get_power_state(ctrl, &ndev->last_ps);
2887 	if (ret < 0)
2888 		goto unfreeze;
2889 
2890 	ret = nvme_set_power_state(ctrl, ctrl->npss);
2891 	if (ret < 0)
2892 		goto unfreeze;
2893 
2894 	if (ret) {
2895 		/*
2896 		 * Clearing npss forces a controller reset on resume. The
2897 		 * correct value will be resdicovered then.
2898 		 */
2899 		nvme_dev_disable(ndev, true);
2900 		ctrl->npss = 0;
2901 		ret = 0;
2902 		goto unfreeze;
2903 	}
2904 	/*
2905 	 * A saved state prevents pci pm from generically controlling the
2906 	 * device's power. If we're using protocol specific settings, we don't
2907 	 * want pci interfering.
2908 	 */
2909 	pci_save_state(pdev);
2910 unfreeze:
2911 	nvme_unfreeze(ctrl);
2912 	return ret;
2913 }
2914 
2915 static int nvme_simple_suspend(struct device *dev)
2916 {
2917 	struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
2918 
2919 	nvme_dev_disable(ndev, true);
2920 	return 0;
2921 }
2922 
2923 static int nvme_simple_resume(struct device *dev)
2924 {
2925 	struct pci_dev *pdev = to_pci_dev(dev);
2926 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2927 
2928 	nvme_reset_ctrl(&ndev->ctrl);
2929 	return 0;
2930 }
2931 
2932 static const struct dev_pm_ops nvme_dev_pm_ops = {
2933 	.suspend	= nvme_suspend,
2934 	.resume		= nvme_resume,
2935 	.freeze		= nvme_simple_suspend,
2936 	.thaw		= nvme_simple_resume,
2937 	.poweroff	= nvme_simple_suspend,
2938 	.restore	= nvme_simple_resume,
2939 };
2940 #endif /* CONFIG_PM_SLEEP */
2941 
2942 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
2943 						pci_channel_state_t state)
2944 {
2945 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2946 
2947 	/*
2948 	 * A frozen channel requires a reset. When detected, this method will
2949 	 * shutdown the controller to quiesce. The controller will be restarted
2950 	 * after the slot reset through driver's slot_reset callback.
2951 	 */
2952 	switch (state) {
2953 	case pci_channel_io_normal:
2954 		return PCI_ERS_RESULT_CAN_RECOVER;
2955 	case pci_channel_io_frozen:
2956 		dev_warn(dev->ctrl.device,
2957 			"frozen state error detected, reset controller\n");
2958 		nvme_dev_disable(dev, false);
2959 		return PCI_ERS_RESULT_NEED_RESET;
2960 	case pci_channel_io_perm_failure:
2961 		dev_warn(dev->ctrl.device,
2962 			"failure state error detected, request disconnect\n");
2963 		return PCI_ERS_RESULT_DISCONNECT;
2964 	}
2965 	return PCI_ERS_RESULT_NEED_RESET;
2966 }
2967 
2968 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
2969 {
2970 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2971 
2972 	dev_info(dev->ctrl.device, "restart after slot reset\n");
2973 	pci_restore_state(pdev);
2974 	nvme_reset_ctrl(&dev->ctrl);
2975 	return PCI_ERS_RESULT_RECOVERED;
2976 }
2977 
2978 static void nvme_error_resume(struct pci_dev *pdev)
2979 {
2980 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2981 
2982 	flush_work(&dev->ctrl.reset_work);
2983 }
2984 
2985 static const struct pci_error_handlers nvme_err_handler = {
2986 	.error_detected	= nvme_error_detected,
2987 	.slot_reset	= nvme_slot_reset,
2988 	.resume		= nvme_error_resume,
2989 	.reset_prepare	= nvme_reset_prepare,
2990 	.reset_done	= nvme_reset_done,
2991 };
2992 
2993 static const struct pci_device_id nvme_id_table[] = {
2994 	{ PCI_VDEVICE(INTEL, 0x0953),
2995 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2996 				NVME_QUIRK_DEALLOCATE_ZEROES, },
2997 	{ PCI_VDEVICE(INTEL, 0x0a53),
2998 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2999 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3000 	{ PCI_VDEVICE(INTEL, 0x0a54),
3001 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3002 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3003 	{ PCI_VDEVICE(INTEL, 0x0a55),
3004 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
3005 				NVME_QUIRK_DEALLOCATE_ZEROES, },
3006 	{ PCI_VDEVICE(INTEL, 0xf1a5),	/* Intel 600P/P3100 */
3007 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3008 				NVME_QUIRK_MEDIUM_PRIO_SQ },
3009 	{ PCI_VDEVICE(INTEL, 0xf1a6),	/* Intel 760p/Pro 7600p */
3010 		.driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3011 	{ PCI_VDEVICE(INTEL, 0x5845),	/* Qemu emulated controller */
3012 		.driver_data = NVME_QUIRK_IDENTIFY_CNS |
3013 				NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3014 	{ PCI_DEVICE(0x1bb1, 0x0100),   /* Seagate Nytro Flash Storage */
3015 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3016 	{ PCI_DEVICE(0x1c58, 0x0003),	/* HGST adapter */
3017 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3018 	{ PCI_DEVICE(0x1c58, 0x0023),	/* WDC SN200 adapter */
3019 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3020 	{ PCI_DEVICE(0x1c5f, 0x0540),	/* Memblaze Pblaze4 adapter */
3021 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3022 	{ PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
3023 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3024 	{ PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
3025 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3026 	{ PCI_DEVICE(0x1d1d, 0x1f1f),	/* LighNVM qemu device */
3027 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3028 	{ PCI_DEVICE(0x1d1d, 0x2807),	/* CNEX WL */
3029 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3030 	{ PCI_DEVICE(0x1d1d, 0x2601),	/* CNEX Granby */
3031 		.driver_data = NVME_QUIRK_LIGHTNVM, },
3032 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3033 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
3034 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3035 	{ 0, }
3036 };
3037 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3038 
3039 static struct pci_driver nvme_driver = {
3040 	.name		= "nvme",
3041 	.id_table	= nvme_id_table,
3042 	.probe		= nvme_probe,
3043 	.remove		= nvme_remove,
3044 	.shutdown	= nvme_shutdown,
3045 #ifdef CONFIG_PM_SLEEP
3046 	.driver		= {
3047 		.pm	= &nvme_dev_pm_ops,
3048 	},
3049 #endif
3050 	.sriov_configure = pci_sriov_configure_simple,
3051 	.err_handler	= &nvme_err_handler,
3052 };
3053 
3054 static int __init nvme_init(void)
3055 {
3056 	BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3057 	BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3058 	BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3059 	BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3060 	return pci_register_driver(&nvme_driver);
3061 }
3062 
3063 static void __exit nvme_exit(void)
3064 {
3065 	pci_unregister_driver(&nvme_driver);
3066 	flush_workqueue(nvme_wq);
3067 }
3068 
3069 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3070 MODULE_LICENSE("GPL");
3071 MODULE_VERSION("1.0");
3072 module_init(nvme_init);
3073 module_exit(nvme_exit);
3074