xref: /openbmc/linux/drivers/nvme/host/pci.c (revision 5ef12cb4a3a78ffb331c03a795a15eea4ae35155)
1 /*
2  * NVM Express device driver
3  * Copyright (c) 2011-2014, Intel Corporation.
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  */
14 
15 #include <linux/aer.h>
16 #include <linux/blkdev.h>
17 #include <linux/blk-mq.h>
18 #include <linux/blk-mq-pci.h>
19 #include <linux/dmi.h>
20 #include <linux/init.h>
21 #include <linux/interrupt.h>
22 #include <linux/io.h>
23 #include <linux/mm.h>
24 #include <linux/module.h>
25 #include <linux/mutex.h>
26 #include <linux/once.h>
27 #include <linux/pci.h>
28 #include <linux/t10-pi.h>
29 #include <linux/types.h>
30 #include <linux/io-64-nonatomic-lo-hi.h>
31 #include <linux/sed-opal.h>
32 
33 #include "nvme.h"
34 
35 #define SQ_SIZE(depth)		(depth * sizeof(struct nvme_command))
36 #define CQ_SIZE(depth)		(depth * sizeof(struct nvme_completion))
37 
38 #define SGES_PER_PAGE	(PAGE_SIZE / sizeof(struct nvme_sgl_desc))
39 
40 static int use_threaded_interrupts;
41 module_param(use_threaded_interrupts, int, 0);
42 
43 static bool use_cmb_sqes = true;
44 module_param(use_cmb_sqes, bool, 0644);
45 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
46 
47 static unsigned int max_host_mem_size_mb = 128;
48 module_param(max_host_mem_size_mb, uint, 0444);
49 MODULE_PARM_DESC(max_host_mem_size_mb,
50 	"Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
51 
52 static unsigned int sgl_threshold = SZ_32K;
53 module_param(sgl_threshold, uint, 0644);
54 MODULE_PARM_DESC(sgl_threshold,
55 		"Use SGLs when average request segment size is larger or equal to "
56 		"this size. Use 0 to disable SGLs.");
57 
58 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
59 static const struct kernel_param_ops io_queue_depth_ops = {
60 	.set = io_queue_depth_set,
61 	.get = param_get_int,
62 };
63 
64 static int io_queue_depth = 1024;
65 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
66 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2");
67 
68 struct nvme_dev;
69 struct nvme_queue;
70 
71 static void nvme_process_cq(struct nvme_queue *nvmeq);
72 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
73 
74 /*
75  * Represents an NVM Express device.  Each nvme_dev is a PCI function.
76  */
77 struct nvme_dev {
78 	struct nvme_queue *queues;
79 	struct blk_mq_tag_set tagset;
80 	struct blk_mq_tag_set admin_tagset;
81 	u32 __iomem *dbs;
82 	struct device *dev;
83 	struct dma_pool *prp_page_pool;
84 	struct dma_pool *prp_small_pool;
85 	unsigned online_queues;
86 	unsigned max_qid;
87 	unsigned int num_vecs;
88 	int q_depth;
89 	u32 db_stride;
90 	void __iomem *bar;
91 	unsigned long bar_mapped_size;
92 	struct work_struct remove_work;
93 	struct mutex shutdown_lock;
94 	bool subsystem;
95 	void __iomem *cmb;
96 	pci_bus_addr_t cmb_bus_addr;
97 	u64 cmb_size;
98 	u32 cmbsz;
99 	u32 cmbloc;
100 	struct nvme_ctrl ctrl;
101 	struct completion ioq_wait;
102 
103 	/* shadow doorbell buffer support: */
104 	u32 *dbbuf_dbs;
105 	dma_addr_t dbbuf_dbs_dma_addr;
106 	u32 *dbbuf_eis;
107 	dma_addr_t dbbuf_eis_dma_addr;
108 
109 	/* host memory buffer support: */
110 	u64 host_mem_size;
111 	u32 nr_host_mem_descs;
112 	dma_addr_t host_mem_descs_dma;
113 	struct nvme_host_mem_buf_desc *host_mem_descs;
114 	void **host_mem_desc_bufs;
115 };
116 
117 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
118 {
119 	int n = 0, ret;
120 
121 	ret = kstrtoint(val, 10, &n);
122 	if (ret != 0 || n < 2)
123 		return -EINVAL;
124 
125 	return param_set_int(val, kp);
126 }
127 
128 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
129 {
130 	return qid * 2 * stride;
131 }
132 
133 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
134 {
135 	return (qid * 2 + 1) * stride;
136 }
137 
138 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
139 {
140 	return container_of(ctrl, struct nvme_dev, ctrl);
141 }
142 
143 /*
144  * An NVM Express queue.  Each device has at least two (one for admin
145  * commands and one for I/O commands).
146  */
147 struct nvme_queue {
148 	struct device *q_dmadev;
149 	struct nvme_dev *dev;
150 	spinlock_t q_lock;
151 	struct nvme_command *sq_cmds;
152 	struct nvme_command __iomem *sq_cmds_io;
153 	volatile struct nvme_completion *cqes;
154 	struct blk_mq_tags **tags;
155 	dma_addr_t sq_dma_addr;
156 	dma_addr_t cq_dma_addr;
157 	u32 __iomem *q_db;
158 	u16 q_depth;
159 	s16 cq_vector;
160 	u16 sq_tail;
161 	u16 cq_head;
162 	u16 qid;
163 	u8 cq_phase;
164 	u8 cqe_seen;
165 	u32 *dbbuf_sq_db;
166 	u32 *dbbuf_cq_db;
167 	u32 *dbbuf_sq_ei;
168 	u32 *dbbuf_cq_ei;
169 };
170 
171 /*
172  * The nvme_iod describes the data in an I/O, including the list of PRP
173  * entries.  You can't see it in this data structure because C doesn't let
174  * me express that.  Use nvme_init_iod to ensure there's enough space
175  * allocated to store the PRP list.
176  */
177 struct nvme_iod {
178 	struct nvme_request req;
179 	struct nvme_queue *nvmeq;
180 	bool use_sgl;
181 	int aborted;
182 	int npages;		/* In the PRP list. 0 means small pool in use */
183 	int nents;		/* Used in scatterlist */
184 	int length;		/* Of data, in bytes */
185 	dma_addr_t first_dma;
186 	struct scatterlist meta_sg; /* metadata requires single contiguous buffer */
187 	struct scatterlist *sg;
188 	struct scatterlist inline_sg[0];
189 };
190 
191 /*
192  * Check we didin't inadvertently grow the command struct
193  */
194 static inline void _nvme_check_size(void)
195 {
196 	BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
197 	BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
198 	BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
199 	BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
200 	BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
201 	BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
202 	BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
203 	BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
204 	BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != NVME_IDENTIFY_DATA_SIZE);
205 	BUILD_BUG_ON(sizeof(struct nvme_id_ns) != NVME_IDENTIFY_DATA_SIZE);
206 	BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
207 	BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
208 	BUILD_BUG_ON(sizeof(struct nvme_dbbuf) != 64);
209 }
210 
211 static inline unsigned int nvme_dbbuf_size(u32 stride)
212 {
213 	return ((num_possible_cpus() + 1) * 8 * stride);
214 }
215 
216 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
217 {
218 	unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
219 
220 	if (dev->dbbuf_dbs)
221 		return 0;
222 
223 	dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
224 					    &dev->dbbuf_dbs_dma_addr,
225 					    GFP_KERNEL);
226 	if (!dev->dbbuf_dbs)
227 		return -ENOMEM;
228 	dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
229 					    &dev->dbbuf_eis_dma_addr,
230 					    GFP_KERNEL);
231 	if (!dev->dbbuf_eis) {
232 		dma_free_coherent(dev->dev, mem_size,
233 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
234 		dev->dbbuf_dbs = NULL;
235 		return -ENOMEM;
236 	}
237 
238 	return 0;
239 }
240 
241 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
242 {
243 	unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
244 
245 	if (dev->dbbuf_dbs) {
246 		dma_free_coherent(dev->dev, mem_size,
247 				  dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
248 		dev->dbbuf_dbs = NULL;
249 	}
250 	if (dev->dbbuf_eis) {
251 		dma_free_coherent(dev->dev, mem_size,
252 				  dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
253 		dev->dbbuf_eis = NULL;
254 	}
255 }
256 
257 static void nvme_dbbuf_init(struct nvme_dev *dev,
258 			    struct nvme_queue *nvmeq, int qid)
259 {
260 	if (!dev->dbbuf_dbs || !qid)
261 		return;
262 
263 	nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
264 	nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
265 	nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
266 	nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
267 }
268 
269 static void nvme_dbbuf_set(struct nvme_dev *dev)
270 {
271 	struct nvme_command c;
272 
273 	if (!dev->dbbuf_dbs)
274 		return;
275 
276 	memset(&c, 0, sizeof(c));
277 	c.dbbuf.opcode = nvme_admin_dbbuf;
278 	c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
279 	c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
280 
281 	if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
282 		dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
283 		/* Free memory and continue on */
284 		nvme_dbbuf_dma_free(dev);
285 	}
286 }
287 
288 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
289 {
290 	return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
291 }
292 
293 /* Update dbbuf and return true if an MMIO is required */
294 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
295 					      volatile u32 *dbbuf_ei)
296 {
297 	if (dbbuf_db) {
298 		u16 old_value;
299 
300 		/*
301 		 * Ensure that the queue is written before updating
302 		 * the doorbell in memory
303 		 */
304 		wmb();
305 
306 		old_value = *dbbuf_db;
307 		*dbbuf_db = value;
308 
309 		if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
310 			return false;
311 	}
312 
313 	return true;
314 }
315 
316 /*
317  * Max size of iod being embedded in the request payload
318  */
319 #define NVME_INT_PAGES		2
320 #define NVME_INT_BYTES(dev)	(NVME_INT_PAGES * (dev)->ctrl.page_size)
321 
322 /*
323  * Will slightly overestimate the number of pages needed.  This is OK
324  * as it only leads to a small amount of wasted memory for the lifetime of
325  * the I/O.
326  */
327 static int nvme_npages(unsigned size, struct nvme_dev *dev)
328 {
329 	unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
330 				      dev->ctrl.page_size);
331 	return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
332 }
333 
334 /*
335  * Calculates the number of pages needed for the SGL segments. For example a 4k
336  * page can accommodate 256 SGL descriptors.
337  */
338 static int nvme_pci_npages_sgl(unsigned int num_seg)
339 {
340 	return DIV_ROUND_UP(num_seg * sizeof(struct nvme_sgl_desc), PAGE_SIZE);
341 }
342 
343 static unsigned int nvme_pci_iod_alloc_size(struct nvme_dev *dev,
344 		unsigned int size, unsigned int nseg, bool use_sgl)
345 {
346 	size_t alloc_size;
347 
348 	if (use_sgl)
349 		alloc_size = sizeof(__le64 *) * nvme_pci_npages_sgl(nseg);
350 	else
351 		alloc_size = sizeof(__le64 *) * nvme_npages(size, dev);
352 
353 	return alloc_size + sizeof(struct scatterlist) * nseg;
354 }
355 
356 static unsigned int nvme_pci_cmd_size(struct nvme_dev *dev, bool use_sgl)
357 {
358 	unsigned int alloc_size = nvme_pci_iod_alloc_size(dev,
359 				    NVME_INT_BYTES(dev), NVME_INT_PAGES,
360 				    use_sgl);
361 
362 	return sizeof(struct nvme_iod) + alloc_size;
363 }
364 
365 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
366 				unsigned int hctx_idx)
367 {
368 	struct nvme_dev *dev = data;
369 	struct nvme_queue *nvmeq = &dev->queues[0];
370 
371 	WARN_ON(hctx_idx != 0);
372 	WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
373 	WARN_ON(nvmeq->tags);
374 
375 	hctx->driver_data = nvmeq;
376 	nvmeq->tags = &dev->admin_tagset.tags[0];
377 	return 0;
378 }
379 
380 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
381 {
382 	struct nvme_queue *nvmeq = hctx->driver_data;
383 
384 	nvmeq->tags = NULL;
385 }
386 
387 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
388 			  unsigned int hctx_idx)
389 {
390 	struct nvme_dev *dev = data;
391 	struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
392 
393 	if (!nvmeq->tags)
394 		nvmeq->tags = &dev->tagset.tags[hctx_idx];
395 
396 	WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
397 	hctx->driver_data = nvmeq;
398 	return 0;
399 }
400 
401 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
402 		unsigned int hctx_idx, unsigned int numa_node)
403 {
404 	struct nvme_dev *dev = set->driver_data;
405 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
406 	int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
407 	struct nvme_queue *nvmeq = &dev->queues[queue_idx];
408 
409 	BUG_ON(!nvmeq);
410 	iod->nvmeq = nvmeq;
411 	return 0;
412 }
413 
414 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
415 {
416 	struct nvme_dev *dev = set->driver_data;
417 
418 	return blk_mq_pci_map_queues(set, to_pci_dev(dev->dev),
419 			dev->num_vecs > 1 ? 1 /* admin queue */ : 0);
420 }
421 
422 /**
423  * __nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
424  * @nvmeq: The queue to use
425  * @cmd: The command to send
426  *
427  * Safe to use from interrupt context
428  */
429 static void __nvme_submit_cmd(struct nvme_queue *nvmeq,
430 						struct nvme_command *cmd)
431 {
432 	u16 tail = nvmeq->sq_tail;
433 
434 	if (nvmeq->sq_cmds_io)
435 		memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd));
436 	else
437 		memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
438 
439 	if (++tail == nvmeq->q_depth)
440 		tail = 0;
441 	if (nvme_dbbuf_update_and_check_event(tail, nvmeq->dbbuf_sq_db,
442 					      nvmeq->dbbuf_sq_ei))
443 		writel(tail, nvmeq->q_db);
444 	nvmeq->sq_tail = tail;
445 }
446 
447 static void **nvme_pci_iod_list(struct request *req)
448 {
449 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
450 	return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
451 }
452 
453 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req)
454 {
455 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
456 	int nseg = blk_rq_nr_phys_segments(req);
457 	unsigned int avg_seg_size;
458 
459 	if (nseg == 0)
460 		return false;
461 
462 	avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
463 
464 	if (!(dev->ctrl.sgls & ((1 << 0) | (1 << 1))))
465 		return false;
466 	if (!iod->nvmeq->qid)
467 		return false;
468 	if (!sgl_threshold || avg_seg_size < sgl_threshold)
469 		return false;
470 	return true;
471 }
472 
473 static blk_status_t nvme_init_iod(struct request *rq, struct nvme_dev *dev)
474 {
475 	struct nvme_iod *iod = blk_mq_rq_to_pdu(rq);
476 	int nseg = blk_rq_nr_phys_segments(rq);
477 	unsigned int size = blk_rq_payload_bytes(rq);
478 
479 	iod->use_sgl = nvme_pci_use_sgls(dev, rq);
480 
481 	if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) {
482 		size_t alloc_size = nvme_pci_iod_alloc_size(dev, size, nseg,
483 				iod->use_sgl);
484 
485 		iod->sg = kmalloc(alloc_size, GFP_ATOMIC);
486 		if (!iod->sg)
487 			return BLK_STS_RESOURCE;
488 	} else {
489 		iod->sg = iod->inline_sg;
490 	}
491 
492 	iod->aborted = 0;
493 	iod->npages = -1;
494 	iod->nents = 0;
495 	iod->length = size;
496 
497 	return BLK_STS_OK;
498 }
499 
500 static void nvme_free_iod(struct nvme_dev *dev, struct request *req)
501 {
502 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
503 	const int last_prp = dev->ctrl.page_size / sizeof(__le64) - 1;
504 	dma_addr_t dma_addr = iod->first_dma, next_dma_addr;
505 
506 	int i;
507 
508 	if (iod->npages == 0)
509 		dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
510 			dma_addr);
511 
512 	for (i = 0; i < iod->npages; i++) {
513 		void *addr = nvme_pci_iod_list(req)[i];
514 
515 		if (iod->use_sgl) {
516 			struct nvme_sgl_desc *sg_list = addr;
517 
518 			next_dma_addr =
519 			    le64_to_cpu((sg_list[SGES_PER_PAGE - 1]).addr);
520 		} else {
521 			__le64 *prp_list = addr;
522 
523 			next_dma_addr = le64_to_cpu(prp_list[last_prp]);
524 		}
525 
526 		dma_pool_free(dev->prp_page_pool, addr, dma_addr);
527 		dma_addr = next_dma_addr;
528 	}
529 
530 	if (iod->sg != iod->inline_sg)
531 		kfree(iod->sg);
532 }
533 
534 #ifdef CONFIG_BLK_DEV_INTEGRITY
535 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
536 {
537 	if (be32_to_cpu(pi->ref_tag) == v)
538 		pi->ref_tag = cpu_to_be32(p);
539 }
540 
541 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
542 {
543 	if (be32_to_cpu(pi->ref_tag) == p)
544 		pi->ref_tag = cpu_to_be32(v);
545 }
546 
547 /**
548  * nvme_dif_remap - remaps ref tags to bip seed and physical lba
549  *
550  * The virtual start sector is the one that was originally submitted by the
551  * block layer.	Due to partitioning, MD/DM cloning, etc. the actual physical
552  * start sector may be different. Remap protection information to match the
553  * physical LBA on writes, and back to the original seed on reads.
554  *
555  * Type 0 and 3 do not have a ref tag, so no remapping required.
556  */
557 static void nvme_dif_remap(struct request *req,
558 			void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
559 {
560 	struct nvme_ns *ns = req->rq_disk->private_data;
561 	struct bio_integrity_payload *bip;
562 	struct t10_pi_tuple *pi;
563 	void *p, *pmap;
564 	u32 i, nlb, ts, phys, virt;
565 
566 	if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
567 		return;
568 
569 	bip = bio_integrity(req->bio);
570 	if (!bip)
571 		return;
572 
573 	pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
574 
575 	p = pmap;
576 	virt = bip_get_seed(bip);
577 	phys = nvme_block_nr(ns, blk_rq_pos(req));
578 	nlb = (blk_rq_bytes(req) >> ns->lba_shift);
579 	ts = ns->disk->queue->integrity.tuple_size;
580 
581 	for (i = 0; i < nlb; i++, virt++, phys++) {
582 		pi = (struct t10_pi_tuple *)p;
583 		dif_swap(phys, virt, pi);
584 		p += ts;
585 	}
586 	kunmap_atomic(pmap);
587 }
588 #else /* CONFIG_BLK_DEV_INTEGRITY */
589 static void nvme_dif_remap(struct request *req,
590 			void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
591 {
592 }
593 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
594 {
595 }
596 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
597 {
598 }
599 #endif
600 
601 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
602 {
603 	int i;
604 	struct scatterlist *sg;
605 
606 	for_each_sg(sgl, sg, nents, i) {
607 		dma_addr_t phys = sg_phys(sg);
608 		pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
609 			"dma_address:%pad dma_length:%d\n",
610 			i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
611 			sg_dma_len(sg));
612 	}
613 }
614 
615 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
616 		struct request *req, struct nvme_rw_command *cmnd)
617 {
618 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
619 	struct dma_pool *pool;
620 	int length = blk_rq_payload_bytes(req);
621 	struct scatterlist *sg = iod->sg;
622 	int dma_len = sg_dma_len(sg);
623 	u64 dma_addr = sg_dma_address(sg);
624 	u32 page_size = dev->ctrl.page_size;
625 	int offset = dma_addr & (page_size - 1);
626 	__le64 *prp_list;
627 	void **list = nvme_pci_iod_list(req);
628 	dma_addr_t prp_dma;
629 	int nprps, i;
630 
631 	length -= (page_size - offset);
632 	if (length <= 0) {
633 		iod->first_dma = 0;
634 		goto done;
635 	}
636 
637 	dma_len -= (page_size - offset);
638 	if (dma_len) {
639 		dma_addr += (page_size - offset);
640 	} else {
641 		sg = sg_next(sg);
642 		dma_addr = sg_dma_address(sg);
643 		dma_len = sg_dma_len(sg);
644 	}
645 
646 	if (length <= page_size) {
647 		iod->first_dma = dma_addr;
648 		goto done;
649 	}
650 
651 	nprps = DIV_ROUND_UP(length, page_size);
652 	if (nprps <= (256 / 8)) {
653 		pool = dev->prp_small_pool;
654 		iod->npages = 0;
655 	} else {
656 		pool = dev->prp_page_pool;
657 		iod->npages = 1;
658 	}
659 
660 	prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
661 	if (!prp_list) {
662 		iod->first_dma = dma_addr;
663 		iod->npages = -1;
664 		return BLK_STS_RESOURCE;
665 	}
666 	list[0] = prp_list;
667 	iod->first_dma = prp_dma;
668 	i = 0;
669 	for (;;) {
670 		if (i == page_size >> 3) {
671 			__le64 *old_prp_list = prp_list;
672 			prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
673 			if (!prp_list)
674 				return BLK_STS_RESOURCE;
675 			list[iod->npages++] = prp_list;
676 			prp_list[0] = old_prp_list[i - 1];
677 			old_prp_list[i - 1] = cpu_to_le64(prp_dma);
678 			i = 1;
679 		}
680 		prp_list[i++] = cpu_to_le64(dma_addr);
681 		dma_len -= page_size;
682 		dma_addr += page_size;
683 		length -= page_size;
684 		if (length <= 0)
685 			break;
686 		if (dma_len > 0)
687 			continue;
688 		if (unlikely(dma_len < 0))
689 			goto bad_sgl;
690 		sg = sg_next(sg);
691 		dma_addr = sg_dma_address(sg);
692 		dma_len = sg_dma_len(sg);
693 	}
694 
695 done:
696 	cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
697 	cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
698 
699 	return BLK_STS_OK;
700 
701  bad_sgl:
702 	WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
703 			"Invalid SGL for payload:%d nents:%d\n",
704 			blk_rq_payload_bytes(req), iod->nents);
705 	return BLK_STS_IOERR;
706 }
707 
708 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
709 		struct scatterlist *sg)
710 {
711 	sge->addr = cpu_to_le64(sg_dma_address(sg));
712 	sge->length = cpu_to_le32(sg_dma_len(sg));
713 	sge->type = NVME_SGL_FMT_DATA_DESC << 4;
714 }
715 
716 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
717 		dma_addr_t dma_addr, int entries)
718 {
719 	sge->addr = cpu_to_le64(dma_addr);
720 	if (entries < SGES_PER_PAGE) {
721 		sge->length = cpu_to_le32(entries * sizeof(*sge));
722 		sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
723 	} else {
724 		sge->length = cpu_to_le32(PAGE_SIZE);
725 		sge->type = NVME_SGL_FMT_SEG_DESC << 4;
726 	}
727 }
728 
729 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
730 		struct request *req, struct nvme_rw_command *cmd, int entries)
731 {
732 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
733 	struct dma_pool *pool;
734 	struct nvme_sgl_desc *sg_list;
735 	struct scatterlist *sg = iod->sg;
736 	dma_addr_t sgl_dma;
737 	int i = 0;
738 
739 	/* setting the transfer type as SGL */
740 	cmd->flags = NVME_CMD_SGL_METABUF;
741 
742 	if (entries == 1) {
743 		nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
744 		return BLK_STS_OK;
745 	}
746 
747 	if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
748 		pool = dev->prp_small_pool;
749 		iod->npages = 0;
750 	} else {
751 		pool = dev->prp_page_pool;
752 		iod->npages = 1;
753 	}
754 
755 	sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
756 	if (!sg_list) {
757 		iod->npages = -1;
758 		return BLK_STS_RESOURCE;
759 	}
760 
761 	nvme_pci_iod_list(req)[0] = sg_list;
762 	iod->first_dma = sgl_dma;
763 
764 	nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
765 
766 	do {
767 		if (i == SGES_PER_PAGE) {
768 			struct nvme_sgl_desc *old_sg_desc = sg_list;
769 			struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
770 
771 			sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
772 			if (!sg_list)
773 				return BLK_STS_RESOURCE;
774 
775 			i = 0;
776 			nvme_pci_iod_list(req)[iod->npages++] = sg_list;
777 			sg_list[i++] = *link;
778 			nvme_pci_sgl_set_seg(link, sgl_dma, entries);
779 		}
780 
781 		nvme_pci_sgl_set_data(&sg_list[i++], sg);
782 		sg = sg_next(sg);
783 	} while (--entries > 0);
784 
785 	return BLK_STS_OK;
786 }
787 
788 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
789 		struct nvme_command *cmnd)
790 {
791 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
792 	struct request_queue *q = req->q;
793 	enum dma_data_direction dma_dir = rq_data_dir(req) ?
794 			DMA_TO_DEVICE : DMA_FROM_DEVICE;
795 	blk_status_t ret = BLK_STS_IOERR;
796 	int nr_mapped;
797 
798 	sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
799 	iod->nents = blk_rq_map_sg(q, req, iod->sg);
800 	if (!iod->nents)
801 		goto out;
802 
803 	ret = BLK_STS_RESOURCE;
804 	nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents, dma_dir,
805 			DMA_ATTR_NO_WARN);
806 	if (!nr_mapped)
807 		goto out;
808 
809 	if (iod->use_sgl)
810 		ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
811 	else
812 		ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
813 
814 	if (ret != BLK_STS_OK)
815 		goto out_unmap;
816 
817 	ret = BLK_STS_IOERR;
818 	if (blk_integrity_rq(req)) {
819 		if (blk_rq_count_integrity_sg(q, req->bio) != 1)
820 			goto out_unmap;
821 
822 		sg_init_table(&iod->meta_sg, 1);
823 		if (blk_rq_map_integrity_sg(q, req->bio, &iod->meta_sg) != 1)
824 			goto out_unmap;
825 
826 		if (req_op(req) == REQ_OP_WRITE)
827 			nvme_dif_remap(req, nvme_dif_prep);
828 
829 		if (!dma_map_sg(dev->dev, &iod->meta_sg, 1, dma_dir))
830 			goto out_unmap;
831 	}
832 
833 	if (blk_integrity_rq(req))
834 		cmnd->rw.metadata = cpu_to_le64(sg_dma_address(&iod->meta_sg));
835 	return BLK_STS_OK;
836 
837 out_unmap:
838 	dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
839 out:
840 	return ret;
841 }
842 
843 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
844 {
845 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
846 	enum dma_data_direction dma_dir = rq_data_dir(req) ?
847 			DMA_TO_DEVICE : DMA_FROM_DEVICE;
848 
849 	if (iod->nents) {
850 		dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
851 		if (blk_integrity_rq(req)) {
852 			if (req_op(req) == REQ_OP_READ)
853 				nvme_dif_remap(req, nvme_dif_complete);
854 			dma_unmap_sg(dev->dev, &iod->meta_sg, 1, dma_dir);
855 		}
856 	}
857 
858 	nvme_cleanup_cmd(req);
859 	nvme_free_iod(dev, req);
860 }
861 
862 /*
863  * NOTE: ns is NULL when called on the admin queue.
864  */
865 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
866 			 const struct blk_mq_queue_data *bd)
867 {
868 	struct nvme_ns *ns = hctx->queue->queuedata;
869 	struct nvme_queue *nvmeq = hctx->driver_data;
870 	struct nvme_dev *dev = nvmeq->dev;
871 	struct request *req = bd->rq;
872 	struct nvme_command cmnd;
873 	blk_status_t ret;
874 
875 	ret = nvme_setup_cmd(ns, req, &cmnd);
876 	if (ret)
877 		return ret;
878 
879 	ret = nvme_init_iod(req, dev);
880 	if (ret)
881 		goto out_free_cmd;
882 
883 	if (blk_rq_nr_phys_segments(req)) {
884 		ret = nvme_map_data(dev, req, &cmnd);
885 		if (ret)
886 			goto out_cleanup_iod;
887 	}
888 
889 	blk_mq_start_request(req);
890 
891 	spin_lock_irq(&nvmeq->q_lock);
892 	if (unlikely(nvmeq->cq_vector < 0)) {
893 		ret = BLK_STS_IOERR;
894 		spin_unlock_irq(&nvmeq->q_lock);
895 		goto out_cleanup_iod;
896 	}
897 	__nvme_submit_cmd(nvmeq, &cmnd);
898 	nvme_process_cq(nvmeq);
899 	spin_unlock_irq(&nvmeq->q_lock);
900 	return BLK_STS_OK;
901 out_cleanup_iod:
902 	nvme_free_iod(dev, req);
903 out_free_cmd:
904 	nvme_cleanup_cmd(req);
905 	return ret;
906 }
907 
908 static void nvme_pci_complete_rq(struct request *req)
909 {
910 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
911 
912 	nvme_unmap_data(iod->nvmeq->dev, req);
913 	nvme_complete_rq(req);
914 }
915 
916 /* We read the CQE phase first to check if the rest of the entry is valid */
917 static inline bool nvme_cqe_valid(struct nvme_queue *nvmeq, u16 head,
918 		u16 phase)
919 {
920 	return (le16_to_cpu(nvmeq->cqes[head].status) & 1) == phase;
921 }
922 
923 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
924 {
925 	u16 head = nvmeq->cq_head;
926 
927 	if (likely(nvmeq->cq_vector >= 0)) {
928 		if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
929 						      nvmeq->dbbuf_cq_ei))
930 			writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
931 	}
932 }
933 
934 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq,
935 		struct nvme_completion *cqe)
936 {
937 	struct request *req;
938 
939 	if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
940 		dev_warn(nvmeq->dev->ctrl.device,
941 			"invalid id %d completed on queue %d\n",
942 			cqe->command_id, le16_to_cpu(cqe->sq_id));
943 		return;
944 	}
945 
946 	/*
947 	 * AEN requests are special as they don't time out and can
948 	 * survive any kind of queue freeze and often don't respond to
949 	 * aborts.  We don't even bother to allocate a struct request
950 	 * for them but rather special case them here.
951 	 */
952 	if (unlikely(nvmeq->qid == 0 &&
953 			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) {
954 		nvme_complete_async_event(&nvmeq->dev->ctrl,
955 				cqe->status, &cqe->result);
956 		return;
957 	}
958 
959 	nvmeq->cqe_seen = 1;
960 	req = blk_mq_tag_to_rq(*nvmeq->tags, cqe->command_id);
961 	nvme_end_request(req, cqe->status, cqe->result);
962 }
963 
964 static inline bool nvme_read_cqe(struct nvme_queue *nvmeq,
965 		struct nvme_completion *cqe)
966 {
967 	if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase)) {
968 		*cqe = nvmeq->cqes[nvmeq->cq_head];
969 
970 		if (++nvmeq->cq_head == nvmeq->q_depth) {
971 			nvmeq->cq_head = 0;
972 			nvmeq->cq_phase = !nvmeq->cq_phase;
973 		}
974 		return true;
975 	}
976 	return false;
977 }
978 
979 static void nvme_process_cq(struct nvme_queue *nvmeq)
980 {
981 	struct nvme_completion cqe;
982 	int consumed = 0;
983 
984 	while (nvme_read_cqe(nvmeq, &cqe)) {
985 		nvme_handle_cqe(nvmeq, &cqe);
986 		consumed++;
987 	}
988 
989 	if (consumed)
990 		nvme_ring_cq_doorbell(nvmeq);
991 }
992 
993 static irqreturn_t nvme_irq(int irq, void *data)
994 {
995 	irqreturn_t result;
996 	struct nvme_queue *nvmeq = data;
997 	spin_lock(&nvmeq->q_lock);
998 	nvme_process_cq(nvmeq);
999 	result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
1000 	nvmeq->cqe_seen = 0;
1001 	spin_unlock(&nvmeq->q_lock);
1002 	return result;
1003 }
1004 
1005 static irqreturn_t nvme_irq_check(int irq, void *data)
1006 {
1007 	struct nvme_queue *nvmeq = data;
1008 	if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
1009 		return IRQ_WAKE_THREAD;
1010 	return IRQ_NONE;
1011 }
1012 
1013 static int __nvme_poll(struct nvme_queue *nvmeq, unsigned int tag)
1014 {
1015 	struct nvme_completion cqe;
1016 	int found = 0, consumed = 0;
1017 
1018 	if (!nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
1019 		return 0;
1020 
1021 	spin_lock_irq(&nvmeq->q_lock);
1022 	while (nvme_read_cqe(nvmeq, &cqe)) {
1023 		nvme_handle_cqe(nvmeq, &cqe);
1024 		consumed++;
1025 
1026 		if (tag == cqe.command_id) {
1027 			found = 1;
1028 			break;
1029 		}
1030        }
1031 
1032 	if (consumed)
1033 		nvme_ring_cq_doorbell(nvmeq);
1034 	spin_unlock_irq(&nvmeq->q_lock);
1035 
1036 	return found;
1037 }
1038 
1039 static int nvme_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1040 {
1041 	struct nvme_queue *nvmeq = hctx->driver_data;
1042 
1043 	return __nvme_poll(nvmeq, tag);
1044 }
1045 
1046 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1047 {
1048 	struct nvme_dev *dev = to_nvme_dev(ctrl);
1049 	struct nvme_queue *nvmeq = &dev->queues[0];
1050 	struct nvme_command c;
1051 
1052 	memset(&c, 0, sizeof(c));
1053 	c.common.opcode = nvme_admin_async_event;
1054 	c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1055 
1056 	spin_lock_irq(&nvmeq->q_lock);
1057 	__nvme_submit_cmd(nvmeq, &c);
1058 	spin_unlock_irq(&nvmeq->q_lock);
1059 }
1060 
1061 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1062 {
1063 	struct nvme_command c;
1064 
1065 	memset(&c, 0, sizeof(c));
1066 	c.delete_queue.opcode = opcode;
1067 	c.delete_queue.qid = cpu_to_le16(id);
1068 
1069 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1070 }
1071 
1072 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1073 						struct nvme_queue *nvmeq)
1074 {
1075 	struct nvme_command c;
1076 	int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
1077 
1078 	/*
1079 	 * Note: we (ab)use the fact that the prp fields survive if no data
1080 	 * is attached to the request.
1081 	 */
1082 	memset(&c, 0, sizeof(c));
1083 	c.create_cq.opcode = nvme_admin_create_cq;
1084 	c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1085 	c.create_cq.cqid = cpu_to_le16(qid);
1086 	c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1087 	c.create_cq.cq_flags = cpu_to_le16(flags);
1088 	c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
1089 
1090 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1091 }
1092 
1093 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1094 						struct nvme_queue *nvmeq)
1095 {
1096 	struct nvme_ctrl *ctrl = &dev->ctrl;
1097 	struct nvme_command c;
1098 	int flags = NVME_QUEUE_PHYS_CONTIG;
1099 
1100 	/*
1101 	 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1102 	 * set. Since URGENT priority is zeroes, it makes all queues
1103 	 * URGENT.
1104 	 */
1105 	if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1106 		flags |= NVME_SQ_PRIO_MEDIUM;
1107 
1108 	/*
1109 	 * Note: we (ab)use the fact that the prp fields survive if no data
1110 	 * is attached to the request.
1111 	 */
1112 	memset(&c, 0, sizeof(c));
1113 	c.create_sq.opcode = nvme_admin_create_sq;
1114 	c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1115 	c.create_sq.sqid = cpu_to_le16(qid);
1116 	c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1117 	c.create_sq.sq_flags = cpu_to_le16(flags);
1118 	c.create_sq.cqid = cpu_to_le16(qid);
1119 
1120 	return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1121 }
1122 
1123 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1124 {
1125 	return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1126 }
1127 
1128 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1129 {
1130 	return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1131 }
1132 
1133 static void abort_endio(struct request *req, blk_status_t error)
1134 {
1135 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1136 	struct nvme_queue *nvmeq = iod->nvmeq;
1137 
1138 	dev_warn(nvmeq->dev->ctrl.device,
1139 		 "Abort status: 0x%x", nvme_req(req)->status);
1140 	atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1141 	blk_mq_free_request(req);
1142 }
1143 
1144 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1145 {
1146 
1147 	/* If true, indicates loss of adapter communication, possibly by a
1148 	 * NVMe Subsystem reset.
1149 	 */
1150 	bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1151 
1152 	/* If there is a reset/reinit ongoing, we shouldn't reset again. */
1153 	switch (dev->ctrl.state) {
1154 	case NVME_CTRL_RESETTING:
1155 	case NVME_CTRL_CONNECTING:
1156 		return false;
1157 	default:
1158 		break;
1159 	}
1160 
1161 	/* We shouldn't reset unless the controller is on fatal error state
1162 	 * _or_ if we lost the communication with it.
1163 	 */
1164 	if (!(csts & NVME_CSTS_CFS) && !nssro)
1165 		return false;
1166 
1167 	return true;
1168 }
1169 
1170 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1171 {
1172 	/* Read a config register to help see what died. */
1173 	u16 pci_status;
1174 	int result;
1175 
1176 	result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1177 				      &pci_status);
1178 	if (result == PCIBIOS_SUCCESSFUL)
1179 		dev_warn(dev->ctrl.device,
1180 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1181 			 csts, pci_status);
1182 	else
1183 		dev_warn(dev->ctrl.device,
1184 			 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1185 			 csts, result);
1186 }
1187 
1188 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1189 {
1190 	struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1191 	struct nvme_queue *nvmeq = iod->nvmeq;
1192 	struct nvme_dev *dev = nvmeq->dev;
1193 	struct request *abort_req;
1194 	struct nvme_command cmd;
1195 	u32 csts = readl(dev->bar + NVME_REG_CSTS);
1196 
1197 	/* If PCI error recovery process is happening, we cannot reset or
1198 	 * the recovery mechanism will surely fail.
1199 	 */
1200 	mb();
1201 	if (pci_channel_offline(to_pci_dev(dev->dev)))
1202 		return BLK_EH_RESET_TIMER;
1203 
1204 	/*
1205 	 * Reset immediately if the controller is failed
1206 	 */
1207 	if (nvme_should_reset(dev, csts)) {
1208 		nvme_warn_reset(dev, csts);
1209 		nvme_dev_disable(dev, false);
1210 		nvme_reset_ctrl(&dev->ctrl);
1211 		return BLK_EH_HANDLED;
1212 	}
1213 
1214 	/*
1215 	 * Did we miss an interrupt?
1216 	 */
1217 	if (__nvme_poll(nvmeq, req->tag)) {
1218 		dev_warn(dev->ctrl.device,
1219 			 "I/O %d QID %d timeout, completion polled\n",
1220 			 req->tag, nvmeq->qid);
1221 		return BLK_EH_HANDLED;
1222 	}
1223 
1224 	/*
1225 	 * Shutdown immediately if controller times out while starting. The
1226 	 * reset work will see the pci device disabled when it gets the forced
1227 	 * cancellation error. All outstanding requests are completed on
1228 	 * shutdown, so we return BLK_EH_HANDLED.
1229 	 */
1230 	switch (dev->ctrl.state) {
1231 	case NVME_CTRL_CONNECTING:
1232 	case NVME_CTRL_RESETTING:
1233 		dev_warn(dev->ctrl.device,
1234 			 "I/O %d QID %d timeout, disable controller\n",
1235 			 req->tag, nvmeq->qid);
1236 		nvme_dev_disable(dev, false);
1237 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1238 		return BLK_EH_HANDLED;
1239 	default:
1240 		break;
1241 	}
1242 
1243 	/*
1244  	 * Shutdown the controller immediately and schedule a reset if the
1245  	 * command was already aborted once before and still hasn't been
1246  	 * returned to the driver, or if this is the admin queue.
1247 	 */
1248 	if (!nvmeq->qid || iod->aborted) {
1249 		dev_warn(dev->ctrl.device,
1250 			 "I/O %d QID %d timeout, reset controller\n",
1251 			 req->tag, nvmeq->qid);
1252 		nvme_dev_disable(dev, false);
1253 		nvme_reset_ctrl(&dev->ctrl);
1254 
1255 		/*
1256 		 * Mark the request as handled, since the inline shutdown
1257 		 * forces all outstanding requests to complete.
1258 		 */
1259 		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1260 		return BLK_EH_HANDLED;
1261 	}
1262 
1263 	if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1264 		atomic_inc(&dev->ctrl.abort_limit);
1265 		return BLK_EH_RESET_TIMER;
1266 	}
1267 	iod->aborted = 1;
1268 
1269 	memset(&cmd, 0, sizeof(cmd));
1270 	cmd.abort.opcode = nvme_admin_abort_cmd;
1271 	cmd.abort.cid = req->tag;
1272 	cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1273 
1274 	dev_warn(nvmeq->dev->ctrl.device,
1275 		"I/O %d QID %d timeout, aborting\n",
1276 		 req->tag, nvmeq->qid);
1277 
1278 	abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1279 			BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1280 	if (IS_ERR(abort_req)) {
1281 		atomic_inc(&dev->ctrl.abort_limit);
1282 		return BLK_EH_RESET_TIMER;
1283 	}
1284 
1285 	abort_req->timeout = ADMIN_TIMEOUT;
1286 	abort_req->end_io_data = NULL;
1287 	blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1288 
1289 	/*
1290 	 * The aborted req will be completed on receiving the abort req.
1291 	 * We enable the timer again. If hit twice, it'll cause a device reset,
1292 	 * as the device then is in a faulty state.
1293 	 */
1294 	return BLK_EH_RESET_TIMER;
1295 }
1296 
1297 static void nvme_free_queue(struct nvme_queue *nvmeq)
1298 {
1299 	dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1300 				(void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1301 	if (nvmeq->sq_cmds)
1302 		dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1303 					nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1304 }
1305 
1306 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1307 {
1308 	int i;
1309 
1310 	for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1311 		dev->ctrl.queue_count--;
1312 		nvme_free_queue(&dev->queues[i]);
1313 	}
1314 }
1315 
1316 /**
1317  * nvme_suspend_queue - put queue into suspended state
1318  * @nvmeq - queue to suspend
1319  */
1320 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1321 {
1322 	int vector;
1323 
1324 	spin_lock_irq(&nvmeq->q_lock);
1325 	if (nvmeq->cq_vector == -1) {
1326 		spin_unlock_irq(&nvmeq->q_lock);
1327 		return 1;
1328 	}
1329 	vector = nvmeq->cq_vector;
1330 	nvmeq->dev->online_queues--;
1331 	nvmeq->cq_vector = -1;
1332 	spin_unlock_irq(&nvmeq->q_lock);
1333 
1334 	if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1335 		blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1336 
1337 	pci_free_irq(to_pci_dev(nvmeq->dev->dev), vector, nvmeq);
1338 
1339 	return 0;
1340 }
1341 
1342 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1343 {
1344 	struct nvme_queue *nvmeq = &dev->queues[0];
1345 
1346 	if (shutdown)
1347 		nvme_shutdown_ctrl(&dev->ctrl);
1348 	else
1349 		nvme_disable_ctrl(&dev->ctrl, dev->ctrl.cap);
1350 
1351 	spin_lock_irq(&nvmeq->q_lock);
1352 	nvme_process_cq(nvmeq);
1353 	spin_unlock_irq(&nvmeq->q_lock);
1354 }
1355 
1356 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1357 				int entry_size)
1358 {
1359 	int q_depth = dev->q_depth;
1360 	unsigned q_size_aligned = roundup(q_depth * entry_size,
1361 					  dev->ctrl.page_size);
1362 
1363 	if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1364 		u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1365 		mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1366 		q_depth = div_u64(mem_per_q, entry_size);
1367 
1368 		/*
1369 		 * Ensure the reduced q_depth is above some threshold where it
1370 		 * would be better to map queues in system memory with the
1371 		 * original depth
1372 		 */
1373 		if (q_depth < 64)
1374 			return -ENOMEM;
1375 	}
1376 
1377 	return q_depth;
1378 }
1379 
1380 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1381 				int qid, int depth)
1382 {
1383 	/* CMB SQEs will be mapped before creation */
1384 	if (qid && dev->cmb && use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS))
1385 		return 0;
1386 
1387 	nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1388 					    &nvmeq->sq_dma_addr, GFP_KERNEL);
1389 	if (!nvmeq->sq_cmds)
1390 		return -ENOMEM;
1391 	return 0;
1392 }
1393 
1394 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1395 {
1396 	struct nvme_queue *nvmeq = &dev->queues[qid];
1397 
1398 	if (dev->ctrl.queue_count > qid)
1399 		return 0;
1400 
1401 	nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1402 					  &nvmeq->cq_dma_addr, GFP_KERNEL);
1403 	if (!nvmeq->cqes)
1404 		goto free_nvmeq;
1405 
1406 	if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
1407 		goto free_cqdma;
1408 
1409 	nvmeq->q_dmadev = dev->dev;
1410 	nvmeq->dev = dev;
1411 	spin_lock_init(&nvmeq->q_lock);
1412 	nvmeq->cq_head = 0;
1413 	nvmeq->cq_phase = 1;
1414 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1415 	nvmeq->q_depth = depth;
1416 	nvmeq->qid = qid;
1417 	nvmeq->cq_vector = -1;
1418 	dev->ctrl.queue_count++;
1419 
1420 	return 0;
1421 
1422  free_cqdma:
1423 	dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1424 							nvmeq->cq_dma_addr);
1425  free_nvmeq:
1426 	return -ENOMEM;
1427 }
1428 
1429 static int queue_request_irq(struct nvme_queue *nvmeq)
1430 {
1431 	struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1432 	int nr = nvmeq->dev->ctrl.instance;
1433 
1434 	if (use_threaded_interrupts) {
1435 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1436 				nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1437 	} else {
1438 		return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1439 				NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1440 	}
1441 }
1442 
1443 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1444 {
1445 	struct nvme_dev *dev = nvmeq->dev;
1446 
1447 	spin_lock_irq(&nvmeq->q_lock);
1448 	nvmeq->sq_tail = 0;
1449 	nvmeq->cq_head = 0;
1450 	nvmeq->cq_phase = 1;
1451 	nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1452 	memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1453 	nvme_dbbuf_init(dev, nvmeq, qid);
1454 	dev->online_queues++;
1455 	spin_unlock_irq(&nvmeq->q_lock);
1456 }
1457 
1458 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1459 {
1460 	struct nvme_dev *dev = nvmeq->dev;
1461 	int result;
1462 
1463 	if (dev->cmb && use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1464 		unsigned offset = (qid - 1) * roundup(SQ_SIZE(nvmeq->q_depth),
1465 						      dev->ctrl.page_size);
1466 		nvmeq->sq_dma_addr = dev->cmb_bus_addr + offset;
1467 		nvmeq->sq_cmds_io = dev->cmb + offset;
1468 	}
1469 
1470 	/*
1471 	 * A queue's vector matches the queue identifier unless the controller
1472 	 * has only one vector available.
1473 	 */
1474 	nvmeq->cq_vector = dev->num_vecs == 1 ? 0 : qid;
1475 	result = adapter_alloc_cq(dev, qid, nvmeq);
1476 	if (result < 0)
1477 		goto release_vector;
1478 
1479 	result = adapter_alloc_sq(dev, qid, nvmeq);
1480 	if (result < 0)
1481 		goto release_cq;
1482 
1483 	nvme_init_queue(nvmeq, qid);
1484 	result = queue_request_irq(nvmeq);
1485 	if (result < 0)
1486 		goto release_sq;
1487 
1488 	return result;
1489 
1490  release_sq:
1491 	dev->online_queues--;
1492 	adapter_delete_sq(dev, qid);
1493  release_cq:
1494 	adapter_delete_cq(dev, qid);
1495  release_vector:
1496 	nvmeq->cq_vector = -1;
1497 	return result;
1498 }
1499 
1500 static const struct blk_mq_ops nvme_mq_admin_ops = {
1501 	.queue_rq	= nvme_queue_rq,
1502 	.complete	= nvme_pci_complete_rq,
1503 	.init_hctx	= nvme_admin_init_hctx,
1504 	.exit_hctx      = nvme_admin_exit_hctx,
1505 	.init_request	= nvme_init_request,
1506 	.timeout	= nvme_timeout,
1507 };
1508 
1509 static const struct blk_mq_ops nvme_mq_ops = {
1510 	.queue_rq	= nvme_queue_rq,
1511 	.complete	= nvme_pci_complete_rq,
1512 	.init_hctx	= nvme_init_hctx,
1513 	.init_request	= nvme_init_request,
1514 	.map_queues	= nvme_pci_map_queues,
1515 	.timeout	= nvme_timeout,
1516 	.poll		= nvme_poll,
1517 };
1518 
1519 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1520 {
1521 	if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1522 		/*
1523 		 * If the controller was reset during removal, it's possible
1524 		 * user requests may be waiting on a stopped queue. Start the
1525 		 * queue to flush these to completion.
1526 		 */
1527 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1528 		blk_cleanup_queue(dev->ctrl.admin_q);
1529 		blk_mq_free_tag_set(&dev->admin_tagset);
1530 	}
1531 }
1532 
1533 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1534 {
1535 	if (!dev->ctrl.admin_q) {
1536 		dev->admin_tagset.ops = &nvme_mq_admin_ops;
1537 		dev->admin_tagset.nr_hw_queues = 1;
1538 
1539 		dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1540 		dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1541 		dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1542 		dev->admin_tagset.cmd_size = nvme_pci_cmd_size(dev, false);
1543 		dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1544 		dev->admin_tagset.driver_data = dev;
1545 
1546 		if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1547 			return -ENOMEM;
1548 		dev->ctrl.admin_tagset = &dev->admin_tagset;
1549 
1550 		dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1551 		if (IS_ERR(dev->ctrl.admin_q)) {
1552 			blk_mq_free_tag_set(&dev->admin_tagset);
1553 			return -ENOMEM;
1554 		}
1555 		if (!blk_get_queue(dev->ctrl.admin_q)) {
1556 			nvme_dev_remove_admin(dev);
1557 			dev->ctrl.admin_q = NULL;
1558 			return -ENODEV;
1559 		}
1560 	} else
1561 		blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1562 
1563 	return 0;
1564 }
1565 
1566 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1567 {
1568 	return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1569 }
1570 
1571 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1572 {
1573 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1574 
1575 	if (size <= dev->bar_mapped_size)
1576 		return 0;
1577 	if (size > pci_resource_len(pdev, 0))
1578 		return -ENOMEM;
1579 	if (dev->bar)
1580 		iounmap(dev->bar);
1581 	dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1582 	if (!dev->bar) {
1583 		dev->bar_mapped_size = 0;
1584 		return -ENOMEM;
1585 	}
1586 	dev->bar_mapped_size = size;
1587 	dev->dbs = dev->bar + NVME_REG_DBS;
1588 
1589 	return 0;
1590 }
1591 
1592 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1593 {
1594 	int result;
1595 	u32 aqa;
1596 	struct nvme_queue *nvmeq;
1597 
1598 	result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1599 	if (result < 0)
1600 		return result;
1601 
1602 	dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1603 				NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1604 
1605 	if (dev->subsystem &&
1606 	    (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1607 		writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1608 
1609 	result = nvme_disable_ctrl(&dev->ctrl, dev->ctrl.cap);
1610 	if (result < 0)
1611 		return result;
1612 
1613 	result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1614 	if (result)
1615 		return result;
1616 
1617 	nvmeq = &dev->queues[0];
1618 	aqa = nvmeq->q_depth - 1;
1619 	aqa |= aqa << 16;
1620 
1621 	writel(aqa, dev->bar + NVME_REG_AQA);
1622 	lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1623 	lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1624 
1625 	result = nvme_enable_ctrl(&dev->ctrl, dev->ctrl.cap);
1626 	if (result)
1627 		return result;
1628 
1629 	nvmeq->cq_vector = 0;
1630 	nvme_init_queue(nvmeq, 0);
1631 	result = queue_request_irq(nvmeq);
1632 	if (result) {
1633 		nvmeq->cq_vector = -1;
1634 		return result;
1635 	}
1636 
1637 	return result;
1638 }
1639 
1640 static int nvme_create_io_queues(struct nvme_dev *dev)
1641 {
1642 	unsigned i, max;
1643 	int ret = 0;
1644 
1645 	for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1646 		if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1647 			ret = -ENOMEM;
1648 			break;
1649 		}
1650 	}
1651 
1652 	max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1653 	for (i = dev->online_queues; i <= max; i++) {
1654 		ret = nvme_create_queue(&dev->queues[i], i);
1655 		if (ret)
1656 			break;
1657 	}
1658 
1659 	/*
1660 	 * Ignore failing Create SQ/CQ commands, we can continue with less
1661 	 * than the desired amount of queues, and even a controller without
1662 	 * I/O queues can still be used to issue admin commands.  This might
1663 	 * be useful to upgrade a buggy firmware for example.
1664 	 */
1665 	return ret >= 0 ? 0 : ret;
1666 }
1667 
1668 static ssize_t nvme_cmb_show(struct device *dev,
1669 			     struct device_attribute *attr,
1670 			     char *buf)
1671 {
1672 	struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1673 
1674 	return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz  : x%08x\n",
1675 		       ndev->cmbloc, ndev->cmbsz);
1676 }
1677 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1678 
1679 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1680 {
1681 	u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1682 
1683 	return 1ULL << (12 + 4 * szu);
1684 }
1685 
1686 static u32 nvme_cmb_size(struct nvme_dev *dev)
1687 {
1688 	return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1689 }
1690 
1691 static void nvme_map_cmb(struct nvme_dev *dev)
1692 {
1693 	u64 size, offset;
1694 	resource_size_t bar_size;
1695 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1696 	int bar;
1697 
1698 	dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1699 	if (!dev->cmbsz)
1700 		return;
1701 	dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1702 
1703 	if (!use_cmb_sqes)
1704 		return;
1705 
1706 	size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1707 	offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1708 	bar = NVME_CMB_BIR(dev->cmbloc);
1709 	bar_size = pci_resource_len(pdev, bar);
1710 
1711 	if (offset > bar_size)
1712 		return;
1713 
1714 	/*
1715 	 * Controllers may support a CMB size larger than their BAR,
1716 	 * for example, due to being behind a bridge. Reduce the CMB to
1717 	 * the reported size of the BAR
1718 	 */
1719 	if (size > bar_size - offset)
1720 		size = bar_size - offset;
1721 
1722 	dev->cmb = ioremap_wc(pci_resource_start(pdev, bar) + offset, size);
1723 	if (!dev->cmb)
1724 		return;
1725 	dev->cmb_bus_addr = pci_bus_address(pdev, bar) + offset;
1726 	dev->cmb_size = size;
1727 
1728 	if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1729 				    &dev_attr_cmb.attr, NULL))
1730 		dev_warn(dev->ctrl.device,
1731 			 "failed to add sysfs attribute for CMB\n");
1732 }
1733 
1734 static inline void nvme_release_cmb(struct nvme_dev *dev)
1735 {
1736 	if (dev->cmb) {
1737 		iounmap(dev->cmb);
1738 		dev->cmb = NULL;
1739 		sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1740 					     &dev_attr_cmb.attr, NULL);
1741 		dev->cmbsz = 0;
1742 	}
1743 }
1744 
1745 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1746 {
1747 	u64 dma_addr = dev->host_mem_descs_dma;
1748 	struct nvme_command c;
1749 	int ret;
1750 
1751 	memset(&c, 0, sizeof(c));
1752 	c.features.opcode	= nvme_admin_set_features;
1753 	c.features.fid		= cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1754 	c.features.dword11	= cpu_to_le32(bits);
1755 	c.features.dword12	= cpu_to_le32(dev->host_mem_size >>
1756 					      ilog2(dev->ctrl.page_size));
1757 	c.features.dword13	= cpu_to_le32(lower_32_bits(dma_addr));
1758 	c.features.dword14	= cpu_to_le32(upper_32_bits(dma_addr));
1759 	c.features.dword15	= cpu_to_le32(dev->nr_host_mem_descs);
1760 
1761 	ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1762 	if (ret) {
1763 		dev_warn(dev->ctrl.device,
1764 			 "failed to set host mem (err %d, flags %#x).\n",
1765 			 ret, bits);
1766 	}
1767 	return ret;
1768 }
1769 
1770 static void nvme_free_host_mem(struct nvme_dev *dev)
1771 {
1772 	int i;
1773 
1774 	for (i = 0; i < dev->nr_host_mem_descs; i++) {
1775 		struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1776 		size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1777 
1778 		dma_free_coherent(dev->dev, size, dev->host_mem_desc_bufs[i],
1779 				le64_to_cpu(desc->addr));
1780 	}
1781 
1782 	kfree(dev->host_mem_desc_bufs);
1783 	dev->host_mem_desc_bufs = NULL;
1784 	dma_free_coherent(dev->dev,
1785 			dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1786 			dev->host_mem_descs, dev->host_mem_descs_dma);
1787 	dev->host_mem_descs = NULL;
1788 	dev->nr_host_mem_descs = 0;
1789 }
1790 
1791 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1792 		u32 chunk_size)
1793 {
1794 	struct nvme_host_mem_buf_desc *descs;
1795 	u32 max_entries, len;
1796 	dma_addr_t descs_dma;
1797 	int i = 0;
1798 	void **bufs;
1799 	u64 size, tmp;
1800 
1801 	tmp = (preferred + chunk_size - 1);
1802 	do_div(tmp, chunk_size);
1803 	max_entries = tmp;
1804 
1805 	if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1806 		max_entries = dev->ctrl.hmmaxd;
1807 
1808 	descs = dma_zalloc_coherent(dev->dev, max_entries * sizeof(*descs),
1809 			&descs_dma, GFP_KERNEL);
1810 	if (!descs)
1811 		goto out;
1812 
1813 	bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1814 	if (!bufs)
1815 		goto out_free_descs;
1816 
1817 	for (size = 0; size < preferred && i < max_entries; size += len) {
1818 		dma_addr_t dma_addr;
1819 
1820 		len = min_t(u64, chunk_size, preferred - size);
1821 		bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1822 				DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1823 		if (!bufs[i])
1824 			break;
1825 
1826 		descs[i].addr = cpu_to_le64(dma_addr);
1827 		descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1828 		i++;
1829 	}
1830 
1831 	if (!size)
1832 		goto out_free_bufs;
1833 
1834 	dev->nr_host_mem_descs = i;
1835 	dev->host_mem_size = size;
1836 	dev->host_mem_descs = descs;
1837 	dev->host_mem_descs_dma = descs_dma;
1838 	dev->host_mem_desc_bufs = bufs;
1839 	return 0;
1840 
1841 out_free_bufs:
1842 	while (--i >= 0) {
1843 		size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1844 
1845 		dma_free_coherent(dev->dev, size, bufs[i],
1846 				le64_to_cpu(descs[i].addr));
1847 	}
1848 
1849 	kfree(bufs);
1850 out_free_descs:
1851 	dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
1852 			descs_dma);
1853 out:
1854 	dev->host_mem_descs = NULL;
1855 	return -ENOMEM;
1856 }
1857 
1858 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1859 {
1860 	u32 chunk_size;
1861 
1862 	/* start big and work our way down */
1863 	for (chunk_size = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
1864 	     chunk_size >= max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
1865 	     chunk_size /= 2) {
1866 		if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
1867 			if (!min || dev->host_mem_size >= min)
1868 				return 0;
1869 			nvme_free_host_mem(dev);
1870 		}
1871 	}
1872 
1873 	return -ENOMEM;
1874 }
1875 
1876 static int nvme_setup_host_mem(struct nvme_dev *dev)
1877 {
1878 	u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1879 	u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1880 	u64 min = (u64)dev->ctrl.hmmin * 4096;
1881 	u32 enable_bits = NVME_HOST_MEM_ENABLE;
1882 	int ret;
1883 
1884 	preferred = min(preferred, max);
1885 	if (min > max) {
1886 		dev_warn(dev->ctrl.device,
1887 			"min host memory (%lld MiB) above limit (%d MiB).\n",
1888 			min >> ilog2(SZ_1M), max_host_mem_size_mb);
1889 		nvme_free_host_mem(dev);
1890 		return 0;
1891 	}
1892 
1893 	/*
1894 	 * If we already have a buffer allocated check if we can reuse it.
1895 	 */
1896 	if (dev->host_mem_descs) {
1897 		if (dev->host_mem_size >= min)
1898 			enable_bits |= NVME_HOST_MEM_RETURN;
1899 		else
1900 			nvme_free_host_mem(dev);
1901 	}
1902 
1903 	if (!dev->host_mem_descs) {
1904 		if (nvme_alloc_host_mem(dev, min, preferred)) {
1905 			dev_warn(dev->ctrl.device,
1906 				"failed to allocate host memory buffer.\n");
1907 			return 0; /* controller must work without HMB */
1908 		}
1909 
1910 		dev_info(dev->ctrl.device,
1911 			"allocated %lld MiB host memory buffer.\n",
1912 			dev->host_mem_size >> ilog2(SZ_1M));
1913 	}
1914 
1915 	ret = nvme_set_host_mem(dev, enable_bits);
1916 	if (ret)
1917 		nvme_free_host_mem(dev);
1918 	return ret;
1919 }
1920 
1921 static int nvme_setup_io_queues(struct nvme_dev *dev)
1922 {
1923 	struct nvme_queue *adminq = &dev->queues[0];
1924 	struct pci_dev *pdev = to_pci_dev(dev->dev);
1925 	int result, nr_io_queues;
1926 	unsigned long size;
1927 
1928 	struct irq_affinity affd = {
1929 		.pre_vectors = 1
1930 	};
1931 
1932 	nr_io_queues = num_possible_cpus();
1933 	result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
1934 	if (result < 0)
1935 		return result;
1936 
1937 	if (nr_io_queues == 0)
1938 		return 0;
1939 
1940 	if (dev->cmb && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1941 		result = nvme_cmb_qdepth(dev, nr_io_queues,
1942 				sizeof(struct nvme_command));
1943 		if (result > 0)
1944 			dev->q_depth = result;
1945 		else
1946 			nvme_release_cmb(dev);
1947 	}
1948 
1949 	do {
1950 		size = db_bar_size(dev, nr_io_queues);
1951 		result = nvme_remap_bar(dev, size);
1952 		if (!result)
1953 			break;
1954 		if (!--nr_io_queues)
1955 			return -ENOMEM;
1956 	} while (1);
1957 	adminq->q_db = dev->dbs;
1958 
1959 	/* Deregister the admin queue's interrupt */
1960 	pci_free_irq(pdev, 0, adminq);
1961 
1962 	/*
1963 	 * If we enable msix early due to not intx, disable it again before
1964 	 * setting up the full range we need.
1965 	 */
1966 	pci_free_irq_vectors(pdev);
1967 	result = pci_alloc_irq_vectors_affinity(pdev, 1, nr_io_queues + 1,
1968 			PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
1969 	if (result <= 0)
1970 		return -EIO;
1971 	dev->num_vecs = result;
1972 	dev->max_qid = max(result - 1, 1);
1973 
1974 	/*
1975 	 * Should investigate if there's a performance win from allocating
1976 	 * more queues than interrupt vectors; it might allow the submission
1977 	 * path to scale better, even if the receive path is limited by the
1978 	 * number of interrupts.
1979 	 */
1980 
1981 	result = queue_request_irq(adminq);
1982 	if (result) {
1983 		adminq->cq_vector = -1;
1984 		return result;
1985 	}
1986 	return nvme_create_io_queues(dev);
1987 }
1988 
1989 static void nvme_del_queue_end(struct request *req, blk_status_t error)
1990 {
1991 	struct nvme_queue *nvmeq = req->end_io_data;
1992 
1993 	blk_mq_free_request(req);
1994 	complete(&nvmeq->dev->ioq_wait);
1995 }
1996 
1997 static void nvme_del_cq_end(struct request *req, blk_status_t error)
1998 {
1999 	struct nvme_queue *nvmeq = req->end_io_data;
2000 
2001 	if (!error) {
2002 		unsigned long flags;
2003 
2004 		/*
2005 		 * We might be called with the AQ q_lock held
2006 		 * and the I/O queue q_lock should always
2007 		 * nest inside the AQ one.
2008 		 */
2009 		spin_lock_irqsave_nested(&nvmeq->q_lock, flags,
2010 					SINGLE_DEPTH_NESTING);
2011 		nvme_process_cq(nvmeq);
2012 		spin_unlock_irqrestore(&nvmeq->q_lock, flags);
2013 	}
2014 
2015 	nvme_del_queue_end(req, error);
2016 }
2017 
2018 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2019 {
2020 	struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2021 	struct request *req;
2022 	struct nvme_command cmd;
2023 
2024 	memset(&cmd, 0, sizeof(cmd));
2025 	cmd.delete_queue.opcode = opcode;
2026 	cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2027 
2028 	req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
2029 	if (IS_ERR(req))
2030 		return PTR_ERR(req);
2031 
2032 	req->timeout = ADMIN_TIMEOUT;
2033 	req->end_io_data = nvmeq;
2034 
2035 	blk_execute_rq_nowait(q, NULL, req, false,
2036 			opcode == nvme_admin_delete_cq ?
2037 				nvme_del_cq_end : nvme_del_queue_end);
2038 	return 0;
2039 }
2040 
2041 static void nvme_disable_io_queues(struct nvme_dev *dev)
2042 {
2043 	int pass, queues = dev->online_queues - 1;
2044 	unsigned long timeout;
2045 	u8 opcode = nvme_admin_delete_sq;
2046 
2047 	for (pass = 0; pass < 2; pass++) {
2048 		int sent = 0, i = queues;
2049 
2050 		reinit_completion(&dev->ioq_wait);
2051  retry:
2052 		timeout = ADMIN_TIMEOUT;
2053 		for (; i > 0; i--, sent++)
2054 			if (nvme_delete_queue(&dev->queues[i], opcode))
2055 				break;
2056 
2057 		while (sent--) {
2058 			timeout = wait_for_completion_io_timeout(&dev->ioq_wait, timeout);
2059 			if (timeout == 0)
2060 				return;
2061 			if (i)
2062 				goto retry;
2063 		}
2064 		opcode = nvme_admin_delete_cq;
2065 	}
2066 }
2067 
2068 /*
2069  * return error value only when tagset allocation failed
2070  */
2071 static int nvme_dev_add(struct nvme_dev *dev)
2072 {
2073 	int ret;
2074 
2075 	if (!dev->ctrl.tagset) {
2076 		dev->tagset.ops = &nvme_mq_ops;
2077 		dev->tagset.nr_hw_queues = dev->online_queues - 1;
2078 		dev->tagset.timeout = NVME_IO_TIMEOUT;
2079 		dev->tagset.numa_node = dev_to_node(dev->dev);
2080 		dev->tagset.queue_depth =
2081 				min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
2082 		dev->tagset.cmd_size = nvme_pci_cmd_size(dev, false);
2083 		if ((dev->ctrl.sgls & ((1 << 0) | (1 << 1))) && sgl_threshold) {
2084 			dev->tagset.cmd_size = max(dev->tagset.cmd_size,
2085 					nvme_pci_cmd_size(dev, true));
2086 		}
2087 		dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2088 		dev->tagset.driver_data = dev;
2089 
2090 		ret = blk_mq_alloc_tag_set(&dev->tagset);
2091 		if (ret) {
2092 			dev_warn(dev->ctrl.device,
2093 				"IO queues tagset allocation failed %d\n", ret);
2094 			return ret;
2095 		}
2096 		dev->ctrl.tagset = &dev->tagset;
2097 
2098 		nvme_dbbuf_set(dev);
2099 	} else {
2100 		blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2101 
2102 		/* Free previously allocated queues that are no longer usable */
2103 		nvme_free_queues(dev, dev->online_queues);
2104 	}
2105 
2106 	return 0;
2107 }
2108 
2109 static int nvme_pci_enable(struct nvme_dev *dev)
2110 {
2111 	int result = -ENOMEM;
2112 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2113 
2114 	if (pci_enable_device_mem(pdev))
2115 		return result;
2116 
2117 	pci_set_master(pdev);
2118 
2119 	if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
2120 	    dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
2121 		goto disable;
2122 
2123 	if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2124 		result = -ENODEV;
2125 		goto disable;
2126 	}
2127 
2128 	/*
2129 	 * Some devices and/or platforms don't advertise or work with INTx
2130 	 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2131 	 * adjust this later.
2132 	 */
2133 	result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2134 	if (result < 0)
2135 		return result;
2136 
2137 	dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2138 
2139 	dev->q_depth = min_t(int, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2140 				io_queue_depth);
2141 	dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2142 	dev->dbs = dev->bar + 4096;
2143 
2144 	/*
2145 	 * Temporary fix for the Apple controller found in the MacBook8,1 and
2146 	 * some MacBook7,1 to avoid controller resets and data loss.
2147 	 */
2148 	if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2149 		dev->q_depth = 2;
2150 		dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2151 			"set queue depth=%u to work around controller resets\n",
2152 			dev->q_depth);
2153 	} else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2154 		   (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2155 		   NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2156 		dev->q_depth = 64;
2157 		dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2158                         "set queue depth=%u\n", dev->q_depth);
2159 	}
2160 
2161 	nvme_map_cmb(dev);
2162 
2163 	pci_enable_pcie_error_reporting(pdev);
2164 	pci_save_state(pdev);
2165 	return 0;
2166 
2167  disable:
2168 	pci_disable_device(pdev);
2169 	return result;
2170 }
2171 
2172 static void nvme_dev_unmap(struct nvme_dev *dev)
2173 {
2174 	if (dev->bar)
2175 		iounmap(dev->bar);
2176 	pci_release_mem_regions(to_pci_dev(dev->dev));
2177 }
2178 
2179 static void nvme_pci_disable(struct nvme_dev *dev)
2180 {
2181 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2182 
2183 	nvme_release_cmb(dev);
2184 	pci_free_irq_vectors(pdev);
2185 
2186 	if (pci_is_enabled(pdev)) {
2187 		pci_disable_pcie_error_reporting(pdev);
2188 		pci_disable_device(pdev);
2189 	}
2190 }
2191 
2192 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2193 {
2194 	int i;
2195 	bool dead = true;
2196 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2197 
2198 	mutex_lock(&dev->shutdown_lock);
2199 	if (pci_is_enabled(pdev)) {
2200 		u32 csts = readl(dev->bar + NVME_REG_CSTS);
2201 
2202 		if (dev->ctrl.state == NVME_CTRL_LIVE ||
2203 		    dev->ctrl.state == NVME_CTRL_RESETTING)
2204 			nvme_start_freeze(&dev->ctrl);
2205 		dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2206 			pdev->error_state  != pci_channel_io_normal);
2207 	}
2208 
2209 	/*
2210 	 * Give the controller a chance to complete all entered requests if
2211 	 * doing a safe shutdown.
2212 	 */
2213 	if (!dead) {
2214 		if (shutdown)
2215 			nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2216 	}
2217 
2218 	nvme_stop_queues(&dev->ctrl);
2219 
2220 	if (!dead && dev->ctrl.queue_count > 0) {
2221 		/*
2222 		 * If the controller is still alive tell it to stop using the
2223 		 * host memory buffer.  In theory the shutdown / reset should
2224 		 * make sure that it doesn't access the host memoery anymore,
2225 		 * but I'd rather be safe than sorry..
2226 		 */
2227 		if (dev->host_mem_descs)
2228 			nvme_set_host_mem(dev, 0);
2229 		nvme_disable_io_queues(dev);
2230 		nvme_disable_admin_queue(dev, shutdown);
2231 	}
2232 	for (i = dev->ctrl.queue_count - 1; i >= 0; i--)
2233 		nvme_suspend_queue(&dev->queues[i]);
2234 
2235 	nvme_pci_disable(dev);
2236 
2237 	blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2238 	blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2239 
2240 	/*
2241 	 * The driver will not be starting up queues again if shutting down so
2242 	 * must flush all entered requests to their failed completion to avoid
2243 	 * deadlocking blk-mq hot-cpu notifier.
2244 	 */
2245 	if (shutdown)
2246 		nvme_start_queues(&dev->ctrl);
2247 	mutex_unlock(&dev->shutdown_lock);
2248 }
2249 
2250 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2251 {
2252 	dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2253 						PAGE_SIZE, PAGE_SIZE, 0);
2254 	if (!dev->prp_page_pool)
2255 		return -ENOMEM;
2256 
2257 	/* Optimisation for I/Os between 4k and 128k */
2258 	dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2259 						256, 256, 0);
2260 	if (!dev->prp_small_pool) {
2261 		dma_pool_destroy(dev->prp_page_pool);
2262 		return -ENOMEM;
2263 	}
2264 	return 0;
2265 }
2266 
2267 static void nvme_release_prp_pools(struct nvme_dev *dev)
2268 {
2269 	dma_pool_destroy(dev->prp_page_pool);
2270 	dma_pool_destroy(dev->prp_small_pool);
2271 }
2272 
2273 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2274 {
2275 	struct nvme_dev *dev = to_nvme_dev(ctrl);
2276 
2277 	nvme_dbbuf_dma_free(dev);
2278 	put_device(dev->dev);
2279 	if (dev->tagset.tags)
2280 		blk_mq_free_tag_set(&dev->tagset);
2281 	if (dev->ctrl.admin_q)
2282 		blk_put_queue(dev->ctrl.admin_q);
2283 	kfree(dev->queues);
2284 	free_opal_dev(dev->ctrl.opal_dev);
2285 	kfree(dev);
2286 }
2287 
2288 static void nvme_remove_dead_ctrl(struct nvme_dev *dev, int status)
2289 {
2290 	dev_warn(dev->ctrl.device, "Removing after probe failure status: %d\n", status);
2291 
2292 	nvme_get_ctrl(&dev->ctrl);
2293 	nvme_dev_disable(dev, false);
2294 	if (!queue_work(nvme_wq, &dev->remove_work))
2295 		nvme_put_ctrl(&dev->ctrl);
2296 }
2297 
2298 static void nvme_reset_work(struct work_struct *work)
2299 {
2300 	struct nvme_dev *dev =
2301 		container_of(work, struct nvme_dev, ctrl.reset_work);
2302 	bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2303 	int result = -ENODEV;
2304 	enum nvme_ctrl_state new_state = NVME_CTRL_LIVE;
2305 
2306 	if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING))
2307 		goto out;
2308 
2309 	/*
2310 	 * If we're called to reset a live controller first shut it down before
2311 	 * moving on.
2312 	 */
2313 	if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2314 		nvme_dev_disable(dev, false);
2315 
2316 	/*
2317 	 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2318 	 * initializing procedure here.
2319 	 */
2320 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2321 		dev_warn(dev->ctrl.device,
2322 			"failed to mark controller CONNECTING\n");
2323 		goto out;
2324 	}
2325 
2326 	result = nvme_pci_enable(dev);
2327 	if (result)
2328 		goto out;
2329 
2330 	result = nvme_pci_configure_admin_queue(dev);
2331 	if (result)
2332 		goto out;
2333 
2334 	result = nvme_alloc_admin_tags(dev);
2335 	if (result)
2336 		goto out;
2337 
2338 	result = nvme_init_identify(&dev->ctrl);
2339 	if (result)
2340 		goto out;
2341 
2342 	if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2343 		if (!dev->ctrl.opal_dev)
2344 			dev->ctrl.opal_dev =
2345 				init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2346 		else if (was_suspend)
2347 			opal_unlock_from_suspend(dev->ctrl.opal_dev);
2348 	} else {
2349 		free_opal_dev(dev->ctrl.opal_dev);
2350 		dev->ctrl.opal_dev = NULL;
2351 	}
2352 
2353 	if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2354 		result = nvme_dbbuf_dma_alloc(dev);
2355 		if (result)
2356 			dev_warn(dev->dev,
2357 				 "unable to allocate dma for dbbuf\n");
2358 	}
2359 
2360 	if (dev->ctrl.hmpre) {
2361 		result = nvme_setup_host_mem(dev);
2362 		if (result < 0)
2363 			goto out;
2364 	}
2365 
2366 	result = nvme_setup_io_queues(dev);
2367 	if (result)
2368 		goto out;
2369 
2370 	/*
2371 	 * Keep the controller around but remove all namespaces if we don't have
2372 	 * any working I/O queue.
2373 	 */
2374 	if (dev->online_queues < 2) {
2375 		dev_warn(dev->ctrl.device, "IO queues not created\n");
2376 		nvme_kill_queues(&dev->ctrl);
2377 		nvme_remove_namespaces(&dev->ctrl);
2378 		new_state = NVME_CTRL_ADMIN_ONLY;
2379 	} else {
2380 		nvme_start_queues(&dev->ctrl);
2381 		nvme_wait_freeze(&dev->ctrl);
2382 		/* hit this only when allocate tagset fails */
2383 		if (nvme_dev_add(dev))
2384 			new_state = NVME_CTRL_ADMIN_ONLY;
2385 		nvme_unfreeze(&dev->ctrl);
2386 	}
2387 
2388 	/*
2389 	 * If only admin queue live, keep it to do further investigation or
2390 	 * recovery.
2391 	 */
2392 	if (!nvme_change_ctrl_state(&dev->ctrl, new_state)) {
2393 		dev_warn(dev->ctrl.device,
2394 			"failed to mark controller state %d\n", new_state);
2395 		goto out;
2396 	}
2397 
2398 	nvme_start_ctrl(&dev->ctrl);
2399 	return;
2400 
2401  out:
2402 	nvme_remove_dead_ctrl(dev, result);
2403 }
2404 
2405 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2406 {
2407 	struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2408 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2409 
2410 	nvme_kill_queues(&dev->ctrl);
2411 	if (pci_get_drvdata(pdev))
2412 		device_release_driver(&pdev->dev);
2413 	nvme_put_ctrl(&dev->ctrl);
2414 }
2415 
2416 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2417 {
2418 	*val = readl(to_nvme_dev(ctrl)->bar + off);
2419 	return 0;
2420 }
2421 
2422 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2423 {
2424 	writel(val, to_nvme_dev(ctrl)->bar + off);
2425 	return 0;
2426 }
2427 
2428 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2429 {
2430 	*val = readq(to_nvme_dev(ctrl)->bar + off);
2431 	return 0;
2432 }
2433 
2434 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2435 {
2436 	struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2437 
2438 	return snprintf(buf, size, "%s", dev_name(&pdev->dev));
2439 }
2440 
2441 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2442 	.name			= "pcie",
2443 	.module			= THIS_MODULE,
2444 	.flags			= NVME_F_METADATA_SUPPORTED,
2445 	.reg_read32		= nvme_pci_reg_read32,
2446 	.reg_write32		= nvme_pci_reg_write32,
2447 	.reg_read64		= nvme_pci_reg_read64,
2448 	.free_ctrl		= nvme_pci_free_ctrl,
2449 	.submit_async_event	= nvme_pci_submit_async_event,
2450 	.get_address		= nvme_pci_get_address,
2451 };
2452 
2453 static int nvme_dev_map(struct nvme_dev *dev)
2454 {
2455 	struct pci_dev *pdev = to_pci_dev(dev->dev);
2456 
2457 	if (pci_request_mem_regions(pdev, "nvme"))
2458 		return -ENODEV;
2459 
2460 	if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2461 		goto release;
2462 
2463 	return 0;
2464   release:
2465 	pci_release_mem_regions(pdev);
2466 	return -ENODEV;
2467 }
2468 
2469 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2470 {
2471 	if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2472 		/*
2473 		 * Several Samsung devices seem to drop off the PCIe bus
2474 		 * randomly when APST is on and uses the deepest sleep state.
2475 		 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2476 		 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2477 		 * 950 PRO 256GB", but it seems to be restricted to two Dell
2478 		 * laptops.
2479 		 */
2480 		if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2481 		    (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2482 		     dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2483 			return NVME_QUIRK_NO_DEEPEST_PS;
2484 	} else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2485 		/*
2486 		 * Samsung SSD 960 EVO drops off the PCIe bus after system
2487 		 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2488 		 * within few minutes after bootup on a Coffee Lake board -
2489 		 * ASUS PRIME Z370-A
2490 		 */
2491 		if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2492 		    (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2493 		     dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2494 			return NVME_QUIRK_NO_APST;
2495 	}
2496 
2497 	return 0;
2498 }
2499 
2500 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2501 {
2502 	int node, result = -ENOMEM;
2503 	struct nvme_dev *dev;
2504 	unsigned long quirks = id->driver_data;
2505 
2506 	node = dev_to_node(&pdev->dev);
2507 	if (node == NUMA_NO_NODE)
2508 		set_dev_node(&pdev->dev, first_memory_node);
2509 
2510 	dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2511 	if (!dev)
2512 		return -ENOMEM;
2513 
2514 	dev->queues = kcalloc_node(num_possible_cpus() + 1,
2515 			sizeof(struct nvme_queue), GFP_KERNEL, node);
2516 	if (!dev->queues)
2517 		goto free;
2518 
2519 	dev->dev = get_device(&pdev->dev);
2520 	pci_set_drvdata(pdev, dev);
2521 
2522 	result = nvme_dev_map(dev);
2523 	if (result)
2524 		goto put_pci;
2525 
2526 	INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2527 	INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2528 	mutex_init(&dev->shutdown_lock);
2529 	init_completion(&dev->ioq_wait);
2530 
2531 	result = nvme_setup_prp_pools(dev);
2532 	if (result)
2533 		goto unmap;
2534 
2535 	quirks |= check_vendor_combination_bug(pdev);
2536 
2537 	result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2538 			quirks);
2539 	if (result)
2540 		goto release_pools;
2541 
2542 	dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2543 
2544 	nvme_reset_ctrl(&dev->ctrl);
2545 
2546 	return 0;
2547 
2548  release_pools:
2549 	nvme_release_prp_pools(dev);
2550  unmap:
2551 	nvme_dev_unmap(dev);
2552  put_pci:
2553 	put_device(dev->dev);
2554  free:
2555 	kfree(dev->queues);
2556 	kfree(dev);
2557 	return result;
2558 }
2559 
2560 static void nvme_reset_prepare(struct pci_dev *pdev)
2561 {
2562 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2563 	nvme_dev_disable(dev, false);
2564 }
2565 
2566 static void nvme_reset_done(struct pci_dev *pdev)
2567 {
2568 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2569 	nvme_reset_ctrl_sync(&dev->ctrl);
2570 }
2571 
2572 static void nvme_shutdown(struct pci_dev *pdev)
2573 {
2574 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2575 	nvme_dev_disable(dev, true);
2576 }
2577 
2578 /*
2579  * The driver's remove may be called on a device in a partially initialized
2580  * state. This function must not have any dependencies on the device state in
2581  * order to proceed.
2582  */
2583 static void nvme_remove(struct pci_dev *pdev)
2584 {
2585 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2586 
2587 	nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2588 
2589 	cancel_work_sync(&dev->ctrl.reset_work);
2590 	pci_set_drvdata(pdev, NULL);
2591 
2592 	if (!pci_device_is_present(pdev)) {
2593 		nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2594 		nvme_dev_disable(dev, false);
2595 	}
2596 
2597 	flush_work(&dev->ctrl.reset_work);
2598 	nvme_stop_ctrl(&dev->ctrl);
2599 	nvme_remove_namespaces(&dev->ctrl);
2600 	nvme_dev_disable(dev, true);
2601 	nvme_free_host_mem(dev);
2602 	nvme_dev_remove_admin(dev);
2603 	nvme_free_queues(dev, 0);
2604 	nvme_uninit_ctrl(&dev->ctrl);
2605 	nvme_release_prp_pools(dev);
2606 	nvme_dev_unmap(dev);
2607 	nvme_put_ctrl(&dev->ctrl);
2608 }
2609 
2610 static int nvme_pci_sriov_configure(struct pci_dev *pdev, int numvfs)
2611 {
2612 	int ret = 0;
2613 
2614 	if (numvfs == 0) {
2615 		if (pci_vfs_assigned(pdev)) {
2616 			dev_warn(&pdev->dev,
2617 				"Cannot disable SR-IOV VFs while assigned\n");
2618 			return -EPERM;
2619 		}
2620 		pci_disable_sriov(pdev);
2621 		return 0;
2622 	}
2623 
2624 	ret = pci_enable_sriov(pdev, numvfs);
2625 	return ret ? ret : numvfs;
2626 }
2627 
2628 #ifdef CONFIG_PM_SLEEP
2629 static int nvme_suspend(struct device *dev)
2630 {
2631 	struct pci_dev *pdev = to_pci_dev(dev);
2632 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2633 
2634 	nvme_dev_disable(ndev, true);
2635 	return 0;
2636 }
2637 
2638 static int nvme_resume(struct device *dev)
2639 {
2640 	struct pci_dev *pdev = to_pci_dev(dev);
2641 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
2642 
2643 	nvme_reset_ctrl(&ndev->ctrl);
2644 	return 0;
2645 }
2646 #endif
2647 
2648 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2649 
2650 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
2651 						pci_channel_state_t state)
2652 {
2653 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2654 
2655 	/*
2656 	 * A frozen channel requires a reset. When detected, this method will
2657 	 * shutdown the controller to quiesce. The controller will be restarted
2658 	 * after the slot reset through driver's slot_reset callback.
2659 	 */
2660 	switch (state) {
2661 	case pci_channel_io_normal:
2662 		return PCI_ERS_RESULT_CAN_RECOVER;
2663 	case pci_channel_io_frozen:
2664 		dev_warn(dev->ctrl.device,
2665 			"frozen state error detected, reset controller\n");
2666 		nvme_dev_disable(dev, false);
2667 		return PCI_ERS_RESULT_NEED_RESET;
2668 	case pci_channel_io_perm_failure:
2669 		dev_warn(dev->ctrl.device,
2670 			"failure state error detected, request disconnect\n");
2671 		return PCI_ERS_RESULT_DISCONNECT;
2672 	}
2673 	return PCI_ERS_RESULT_NEED_RESET;
2674 }
2675 
2676 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
2677 {
2678 	struct nvme_dev *dev = pci_get_drvdata(pdev);
2679 
2680 	dev_info(dev->ctrl.device, "restart after slot reset\n");
2681 	pci_restore_state(pdev);
2682 	nvme_reset_ctrl(&dev->ctrl);
2683 	return PCI_ERS_RESULT_RECOVERED;
2684 }
2685 
2686 static void nvme_error_resume(struct pci_dev *pdev)
2687 {
2688 	pci_cleanup_aer_uncorrect_error_status(pdev);
2689 }
2690 
2691 static const struct pci_error_handlers nvme_err_handler = {
2692 	.error_detected	= nvme_error_detected,
2693 	.slot_reset	= nvme_slot_reset,
2694 	.resume		= nvme_error_resume,
2695 	.reset_prepare	= nvme_reset_prepare,
2696 	.reset_done	= nvme_reset_done,
2697 };
2698 
2699 static const struct pci_device_id nvme_id_table[] = {
2700 	{ PCI_VDEVICE(INTEL, 0x0953),
2701 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2702 				NVME_QUIRK_DEALLOCATE_ZEROES, },
2703 	{ PCI_VDEVICE(INTEL, 0x0a53),
2704 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2705 				NVME_QUIRK_DEALLOCATE_ZEROES, },
2706 	{ PCI_VDEVICE(INTEL, 0x0a54),
2707 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2708 				NVME_QUIRK_DEALLOCATE_ZEROES, },
2709 	{ PCI_VDEVICE(INTEL, 0x0a55),
2710 		.driver_data = NVME_QUIRK_STRIPE_SIZE |
2711 				NVME_QUIRK_DEALLOCATE_ZEROES, },
2712 	{ PCI_VDEVICE(INTEL, 0xf1a5),	/* Intel 600P/P3100 */
2713 		.driver_data = NVME_QUIRK_NO_DEEPEST_PS |
2714 				NVME_QUIRK_MEDIUM_PRIO_SQ },
2715 	{ PCI_VDEVICE(INTEL, 0x5845),	/* Qemu emulated controller */
2716 		.driver_data = NVME_QUIRK_IDENTIFY_CNS, },
2717 	{ PCI_DEVICE(0x1c58, 0x0003),	/* HGST adapter */
2718 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2719 	{ PCI_DEVICE(0x1c58, 0x0023),	/* WDC SN200 adapter */
2720 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2721 	{ PCI_DEVICE(0x1c5f, 0x0540),	/* Memblaze Pblaze4 adapter */
2722 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2723 	{ PCI_DEVICE(0x144d, 0xa821),   /* Samsung PM1725 */
2724 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2725 	{ PCI_DEVICE(0x144d, 0xa822),   /* Samsung PM1725a */
2726 		.driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2727 	{ PCI_DEVICE(0x1d1d, 0x1f1f),	/* LighNVM qemu device */
2728 		.driver_data = NVME_QUIRK_LIGHTNVM, },
2729 	{ PCI_DEVICE(0x1d1d, 0x2807),	/* CNEX WL */
2730 		.driver_data = NVME_QUIRK_LIGHTNVM, },
2731 	{ PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2732 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
2733 	{ PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
2734 	{ 0, }
2735 };
2736 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2737 
2738 static struct pci_driver nvme_driver = {
2739 	.name		= "nvme",
2740 	.id_table	= nvme_id_table,
2741 	.probe		= nvme_probe,
2742 	.remove		= nvme_remove,
2743 	.shutdown	= nvme_shutdown,
2744 	.driver		= {
2745 		.pm	= &nvme_dev_pm_ops,
2746 	},
2747 	.sriov_configure = nvme_pci_sriov_configure,
2748 	.err_handler	= &nvme_err_handler,
2749 };
2750 
2751 static int __init nvme_init(void)
2752 {
2753 	return pci_register_driver(&nvme_driver);
2754 }
2755 
2756 static void __exit nvme_exit(void)
2757 {
2758 	pci_unregister_driver(&nvme_driver);
2759 	flush_workqueue(nvme_wq);
2760 	_nvme_check_size();
2761 }
2762 
2763 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2764 MODULE_LICENSE("GPL");
2765 MODULE_VERSION("1.0");
2766 module_init(nvme_init);
2767 module_exit(nvme_exit);
2768