xref: /linux/drivers/block/virtio_blk.c (revision 3c4fc7bf4c9e66fe71abcbf93f62f4ddb89b7f15)
1 // SPDX-License-Identifier: GPL-2.0-only
2 //#define DEBUG
3 #include <linux/spinlock.h>
4 #include <linux/slab.h>
5 #include <linux/blkdev.h>
6 #include <linux/hdreg.h>
7 #include <linux/module.h>
8 #include <linux/mutex.h>
9 #include <linux/interrupt.h>
10 #include <linux/virtio.h>
11 #include <linux/virtio_blk.h>
12 #include <linux/scatterlist.h>
13 #include <linux/string_helpers.h>
14 #include <linux/idr.h>
15 #include <linux/blk-mq.h>
16 #include <linux/blk-mq-virtio.h>
17 #include <linux/numa.h>
18 #include <uapi/linux/virtio_ring.h>
19 
20 #define PART_BITS 4
21 #define VQ_NAME_LEN 16
22 #define MAX_DISCARD_SEGMENTS 256u
23 
24 /* The maximum number of sg elements that fit into a virtqueue */
25 #define VIRTIO_BLK_MAX_SG_ELEMS 32768
26 
27 #ifdef CONFIG_ARCH_NO_SG_CHAIN
28 #define VIRTIO_BLK_INLINE_SG_CNT	0
29 #else
30 #define VIRTIO_BLK_INLINE_SG_CNT	2
31 #endif
32 
33 static unsigned int num_request_queues;
34 module_param(num_request_queues, uint, 0644);
35 MODULE_PARM_DESC(num_request_queues,
36 		 "Limit the number of request queues to use for blk device. "
37 		 "0 for no limit. "
38 		 "Values > nr_cpu_ids truncated to nr_cpu_ids.");
39 
40 static unsigned int poll_queues;
41 module_param(poll_queues, uint, 0644);
42 MODULE_PARM_DESC(poll_queues, "The number of dedicated virtqueues for polling I/O");
43 
44 static int major;
45 static DEFINE_IDA(vd_index_ida);
46 
47 static struct workqueue_struct *virtblk_wq;
48 
49 struct virtio_blk_vq {
50 	struct virtqueue *vq;
51 	spinlock_t lock;
52 	char name[VQ_NAME_LEN];
53 } ____cacheline_aligned_in_smp;
54 
55 struct virtio_blk {
56 	/*
57 	 * This mutex must be held by anything that may run after
58 	 * virtblk_remove() sets vblk->vdev to NULL.
59 	 *
60 	 * blk-mq, virtqueue processing, and sysfs attribute code paths are
61 	 * shut down before vblk->vdev is set to NULL and therefore do not need
62 	 * to hold this mutex.
63 	 */
64 	struct mutex vdev_mutex;
65 	struct virtio_device *vdev;
66 
67 	/* The disk structure for the kernel. */
68 	struct gendisk *disk;
69 
70 	/* Block layer tags. */
71 	struct blk_mq_tag_set tag_set;
72 
73 	/* Process context for config space updates */
74 	struct work_struct config_work;
75 
76 	/* Ida index - used to track minor number allocations. */
77 	int index;
78 
79 	/* num of vqs */
80 	int num_vqs;
81 	int io_queues[HCTX_MAX_TYPES];
82 	struct virtio_blk_vq *vqs;
83 };
84 
85 struct virtblk_req {
86 	struct virtio_blk_outhdr out_hdr;
87 	u8 status;
88 	struct sg_table sg_table;
89 	struct scatterlist sg[];
90 };
91 
92 static inline blk_status_t virtblk_result(struct virtblk_req *vbr)
93 {
94 	switch (vbr->status) {
95 	case VIRTIO_BLK_S_OK:
96 		return BLK_STS_OK;
97 	case VIRTIO_BLK_S_UNSUPP:
98 		return BLK_STS_NOTSUPP;
99 	default:
100 		return BLK_STS_IOERR;
101 	}
102 }
103 
104 static inline struct virtio_blk_vq *get_virtio_blk_vq(struct blk_mq_hw_ctx *hctx)
105 {
106 	struct virtio_blk *vblk = hctx->queue->queuedata;
107 	struct virtio_blk_vq *vq = &vblk->vqs[hctx->queue_num];
108 
109 	return vq;
110 }
111 
112 static int virtblk_add_req(struct virtqueue *vq, struct virtblk_req *vbr)
113 {
114 	struct scatterlist hdr, status, *sgs[3];
115 	unsigned int num_out = 0, num_in = 0;
116 
117 	sg_init_one(&hdr, &vbr->out_hdr, sizeof(vbr->out_hdr));
118 	sgs[num_out++] = &hdr;
119 
120 	if (vbr->sg_table.nents) {
121 		if (vbr->out_hdr.type & cpu_to_virtio32(vq->vdev, VIRTIO_BLK_T_OUT))
122 			sgs[num_out++] = vbr->sg_table.sgl;
123 		else
124 			sgs[num_out + num_in++] = vbr->sg_table.sgl;
125 	}
126 
127 	sg_init_one(&status, &vbr->status, sizeof(vbr->status));
128 	sgs[num_out + num_in++] = &status;
129 
130 	return virtqueue_add_sgs(vq, sgs, num_out, num_in, vbr, GFP_ATOMIC);
131 }
132 
133 static int virtblk_setup_discard_write_zeroes(struct request *req, bool unmap)
134 {
135 	unsigned short segments = blk_rq_nr_discard_segments(req);
136 	unsigned short n = 0;
137 	struct virtio_blk_discard_write_zeroes *range;
138 	struct bio *bio;
139 	u32 flags = 0;
140 
141 	if (unmap)
142 		flags |= VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP;
143 
144 	range = kmalloc_array(segments, sizeof(*range), GFP_ATOMIC);
145 	if (!range)
146 		return -ENOMEM;
147 
148 	/*
149 	 * Single max discard segment means multi-range discard isn't
150 	 * supported, and block layer only runs contiguity merge like
151 	 * normal RW request. So we can't reply on bio for retrieving
152 	 * each range info.
153 	 */
154 	if (queue_max_discard_segments(req->q) == 1) {
155 		range[0].flags = cpu_to_le32(flags);
156 		range[0].num_sectors = cpu_to_le32(blk_rq_sectors(req));
157 		range[0].sector = cpu_to_le64(blk_rq_pos(req));
158 		n = 1;
159 	} else {
160 		__rq_for_each_bio(bio, req) {
161 			u64 sector = bio->bi_iter.bi_sector;
162 			u32 num_sectors = bio->bi_iter.bi_size >> SECTOR_SHIFT;
163 
164 			range[n].flags = cpu_to_le32(flags);
165 			range[n].num_sectors = cpu_to_le32(num_sectors);
166 			range[n].sector = cpu_to_le64(sector);
167 			n++;
168 		}
169 	}
170 
171 	WARN_ON_ONCE(n != segments);
172 
173 	req->special_vec.bv_page = virt_to_page(range);
174 	req->special_vec.bv_offset = offset_in_page(range);
175 	req->special_vec.bv_len = sizeof(*range) * segments;
176 	req->rq_flags |= RQF_SPECIAL_PAYLOAD;
177 
178 	return 0;
179 }
180 
181 static void virtblk_unmap_data(struct request *req, struct virtblk_req *vbr)
182 {
183 	if (blk_rq_nr_phys_segments(req))
184 		sg_free_table_chained(&vbr->sg_table,
185 				      VIRTIO_BLK_INLINE_SG_CNT);
186 }
187 
188 static int virtblk_map_data(struct blk_mq_hw_ctx *hctx, struct request *req,
189 		struct virtblk_req *vbr)
190 {
191 	int err;
192 
193 	if (!blk_rq_nr_phys_segments(req))
194 		return 0;
195 
196 	vbr->sg_table.sgl = vbr->sg;
197 	err = sg_alloc_table_chained(&vbr->sg_table,
198 				     blk_rq_nr_phys_segments(req),
199 				     vbr->sg_table.sgl,
200 				     VIRTIO_BLK_INLINE_SG_CNT);
201 	if (unlikely(err))
202 		return -ENOMEM;
203 
204 	return blk_rq_map_sg(hctx->queue, req, vbr->sg_table.sgl);
205 }
206 
207 static void virtblk_cleanup_cmd(struct request *req)
208 {
209 	if (req->rq_flags & RQF_SPECIAL_PAYLOAD)
210 		kfree(bvec_virt(&req->special_vec));
211 }
212 
213 static blk_status_t virtblk_setup_cmd(struct virtio_device *vdev,
214 				      struct request *req,
215 				      struct virtblk_req *vbr)
216 {
217 	bool unmap = false;
218 	u32 type;
219 
220 	vbr->out_hdr.sector = 0;
221 
222 	switch (req_op(req)) {
223 	case REQ_OP_READ:
224 		type = VIRTIO_BLK_T_IN;
225 		vbr->out_hdr.sector = cpu_to_virtio64(vdev,
226 						      blk_rq_pos(req));
227 		break;
228 	case REQ_OP_WRITE:
229 		type = VIRTIO_BLK_T_OUT;
230 		vbr->out_hdr.sector = cpu_to_virtio64(vdev,
231 						      blk_rq_pos(req));
232 		break;
233 	case REQ_OP_FLUSH:
234 		type = VIRTIO_BLK_T_FLUSH;
235 		break;
236 	case REQ_OP_DISCARD:
237 		type = VIRTIO_BLK_T_DISCARD;
238 		break;
239 	case REQ_OP_WRITE_ZEROES:
240 		type = VIRTIO_BLK_T_WRITE_ZEROES;
241 		unmap = !(req->cmd_flags & REQ_NOUNMAP);
242 		break;
243 	case REQ_OP_DRV_IN:
244 		type = VIRTIO_BLK_T_GET_ID;
245 		break;
246 	default:
247 		WARN_ON_ONCE(1);
248 		return BLK_STS_IOERR;
249 	}
250 
251 	vbr->out_hdr.type = cpu_to_virtio32(vdev, type);
252 	vbr->out_hdr.ioprio = cpu_to_virtio32(vdev, req_get_ioprio(req));
253 
254 	if (type == VIRTIO_BLK_T_DISCARD || type == VIRTIO_BLK_T_WRITE_ZEROES) {
255 		if (virtblk_setup_discard_write_zeroes(req, unmap))
256 			return BLK_STS_RESOURCE;
257 	}
258 
259 	return 0;
260 }
261 
262 static inline void virtblk_request_done(struct request *req)
263 {
264 	struct virtblk_req *vbr = blk_mq_rq_to_pdu(req);
265 
266 	virtblk_unmap_data(req, vbr);
267 	virtblk_cleanup_cmd(req);
268 	blk_mq_end_request(req, virtblk_result(vbr));
269 }
270 
271 static void virtblk_done(struct virtqueue *vq)
272 {
273 	struct virtio_blk *vblk = vq->vdev->priv;
274 	bool req_done = false;
275 	int qid = vq->index;
276 	struct virtblk_req *vbr;
277 	unsigned long flags;
278 	unsigned int len;
279 
280 	spin_lock_irqsave(&vblk->vqs[qid].lock, flags);
281 	do {
282 		virtqueue_disable_cb(vq);
283 		while ((vbr = virtqueue_get_buf(vblk->vqs[qid].vq, &len)) != NULL) {
284 			struct request *req = blk_mq_rq_from_pdu(vbr);
285 
286 			if (likely(!blk_should_fake_timeout(req->q)))
287 				blk_mq_complete_request(req);
288 			req_done = true;
289 		}
290 		if (unlikely(virtqueue_is_broken(vq)))
291 			break;
292 	} while (!virtqueue_enable_cb(vq));
293 
294 	/* In case queue is stopped waiting for more buffers. */
295 	if (req_done)
296 		blk_mq_start_stopped_hw_queues(vblk->disk->queue, true);
297 	spin_unlock_irqrestore(&vblk->vqs[qid].lock, flags);
298 }
299 
300 static void virtio_commit_rqs(struct blk_mq_hw_ctx *hctx)
301 {
302 	struct virtio_blk *vblk = hctx->queue->queuedata;
303 	struct virtio_blk_vq *vq = &vblk->vqs[hctx->queue_num];
304 	bool kick;
305 
306 	spin_lock_irq(&vq->lock);
307 	kick = virtqueue_kick_prepare(vq->vq);
308 	spin_unlock_irq(&vq->lock);
309 
310 	if (kick)
311 		virtqueue_notify(vq->vq);
312 }
313 
314 static blk_status_t virtblk_prep_rq(struct blk_mq_hw_ctx *hctx,
315 					struct virtio_blk *vblk,
316 					struct request *req,
317 					struct virtblk_req *vbr)
318 {
319 	blk_status_t status;
320 
321 	status = virtblk_setup_cmd(vblk->vdev, req, vbr);
322 	if (unlikely(status))
323 		return status;
324 
325 	blk_mq_start_request(req);
326 
327 	vbr->sg_table.nents = virtblk_map_data(hctx, req, vbr);
328 	if (unlikely(vbr->sg_table.nents < 0)) {
329 		virtblk_cleanup_cmd(req);
330 		return BLK_STS_RESOURCE;
331 	}
332 
333 	return BLK_STS_OK;
334 }
335 
336 static blk_status_t virtio_queue_rq(struct blk_mq_hw_ctx *hctx,
337 			   const struct blk_mq_queue_data *bd)
338 {
339 	struct virtio_blk *vblk = hctx->queue->queuedata;
340 	struct request *req = bd->rq;
341 	struct virtblk_req *vbr = blk_mq_rq_to_pdu(req);
342 	unsigned long flags;
343 	int qid = hctx->queue_num;
344 	bool notify = false;
345 	blk_status_t status;
346 	int err;
347 
348 	status = virtblk_prep_rq(hctx, vblk, req, vbr);
349 	if (unlikely(status))
350 		return status;
351 
352 	spin_lock_irqsave(&vblk->vqs[qid].lock, flags);
353 	err = virtblk_add_req(vblk->vqs[qid].vq, vbr);
354 	if (err) {
355 		virtqueue_kick(vblk->vqs[qid].vq);
356 		/* Don't stop the queue if -ENOMEM: we may have failed to
357 		 * bounce the buffer due to global resource outage.
358 		 */
359 		if (err == -ENOSPC)
360 			blk_mq_stop_hw_queue(hctx);
361 		spin_unlock_irqrestore(&vblk->vqs[qid].lock, flags);
362 		virtblk_unmap_data(req, vbr);
363 		virtblk_cleanup_cmd(req);
364 		switch (err) {
365 		case -ENOSPC:
366 			return BLK_STS_DEV_RESOURCE;
367 		case -ENOMEM:
368 			return BLK_STS_RESOURCE;
369 		default:
370 			return BLK_STS_IOERR;
371 		}
372 	}
373 
374 	if (bd->last && virtqueue_kick_prepare(vblk->vqs[qid].vq))
375 		notify = true;
376 	spin_unlock_irqrestore(&vblk->vqs[qid].lock, flags);
377 
378 	if (notify)
379 		virtqueue_notify(vblk->vqs[qid].vq);
380 	return BLK_STS_OK;
381 }
382 
383 static bool virtblk_prep_rq_batch(struct request *req)
384 {
385 	struct virtio_blk *vblk = req->mq_hctx->queue->queuedata;
386 	struct virtblk_req *vbr = blk_mq_rq_to_pdu(req);
387 
388 	req->mq_hctx->tags->rqs[req->tag] = req;
389 
390 	return virtblk_prep_rq(req->mq_hctx, vblk, req, vbr) == BLK_STS_OK;
391 }
392 
393 static bool virtblk_add_req_batch(struct virtio_blk_vq *vq,
394 					struct request **rqlist,
395 					struct request **requeue_list)
396 {
397 	unsigned long flags;
398 	int err;
399 	bool kick;
400 
401 	spin_lock_irqsave(&vq->lock, flags);
402 
403 	while (!rq_list_empty(*rqlist)) {
404 		struct request *req = rq_list_pop(rqlist);
405 		struct virtblk_req *vbr = blk_mq_rq_to_pdu(req);
406 
407 		err = virtblk_add_req(vq->vq, vbr);
408 		if (err) {
409 			virtblk_unmap_data(req, vbr);
410 			virtblk_cleanup_cmd(req);
411 			rq_list_add(requeue_list, req);
412 		}
413 	}
414 
415 	kick = virtqueue_kick_prepare(vq->vq);
416 	spin_unlock_irqrestore(&vq->lock, flags);
417 
418 	return kick;
419 }
420 
421 static void virtio_queue_rqs(struct request **rqlist)
422 {
423 	struct request *req, *next, *prev = NULL;
424 	struct request *requeue_list = NULL;
425 
426 	rq_list_for_each_safe(rqlist, req, next) {
427 		struct virtio_blk_vq *vq = get_virtio_blk_vq(req->mq_hctx);
428 		bool kick;
429 
430 		if (!virtblk_prep_rq_batch(req)) {
431 			rq_list_move(rqlist, &requeue_list, req, prev);
432 			req = prev;
433 			if (!req)
434 				continue;
435 		}
436 
437 		if (!next || req->mq_hctx != next->mq_hctx) {
438 			req->rq_next = NULL;
439 			kick = virtblk_add_req_batch(vq, rqlist, &requeue_list);
440 			if (kick)
441 				virtqueue_notify(vq->vq);
442 
443 			*rqlist = next;
444 			prev = NULL;
445 		} else
446 			prev = req;
447 	}
448 
449 	*rqlist = requeue_list;
450 }
451 
452 /* return id (s/n) string for *disk to *id_str
453  */
454 static int virtblk_get_id(struct gendisk *disk, char *id_str)
455 {
456 	struct virtio_blk *vblk = disk->private_data;
457 	struct request_queue *q = vblk->disk->queue;
458 	struct request *req;
459 	int err;
460 
461 	req = blk_mq_alloc_request(q, REQ_OP_DRV_IN, 0);
462 	if (IS_ERR(req))
463 		return PTR_ERR(req);
464 
465 	err = blk_rq_map_kern(q, req, id_str, VIRTIO_BLK_ID_BYTES, GFP_KERNEL);
466 	if (err)
467 		goto out;
468 
469 	blk_execute_rq(req, false);
470 	err = blk_status_to_errno(virtblk_result(blk_mq_rq_to_pdu(req)));
471 out:
472 	blk_mq_free_request(req);
473 	return err;
474 }
475 
476 /* We provide getgeo only to please some old bootloader/partitioning tools */
477 static int virtblk_getgeo(struct block_device *bd, struct hd_geometry *geo)
478 {
479 	struct virtio_blk *vblk = bd->bd_disk->private_data;
480 	int ret = 0;
481 
482 	mutex_lock(&vblk->vdev_mutex);
483 
484 	if (!vblk->vdev) {
485 		ret = -ENXIO;
486 		goto out;
487 	}
488 
489 	/* see if the host passed in geometry config */
490 	if (virtio_has_feature(vblk->vdev, VIRTIO_BLK_F_GEOMETRY)) {
491 		virtio_cread(vblk->vdev, struct virtio_blk_config,
492 			     geometry.cylinders, &geo->cylinders);
493 		virtio_cread(vblk->vdev, struct virtio_blk_config,
494 			     geometry.heads, &geo->heads);
495 		virtio_cread(vblk->vdev, struct virtio_blk_config,
496 			     geometry.sectors, &geo->sectors);
497 	} else {
498 		/* some standard values, similar to sd */
499 		geo->heads = 1 << 6;
500 		geo->sectors = 1 << 5;
501 		geo->cylinders = get_capacity(bd->bd_disk) >> 11;
502 	}
503 out:
504 	mutex_unlock(&vblk->vdev_mutex);
505 	return ret;
506 }
507 
508 static void virtblk_free_disk(struct gendisk *disk)
509 {
510 	struct virtio_blk *vblk = disk->private_data;
511 
512 	ida_simple_remove(&vd_index_ida, vblk->index);
513 	mutex_destroy(&vblk->vdev_mutex);
514 	kfree(vblk);
515 }
516 
517 static const struct block_device_operations virtblk_fops = {
518 	.owner  	= THIS_MODULE,
519 	.getgeo		= virtblk_getgeo,
520 	.free_disk	= virtblk_free_disk,
521 };
522 
523 static int index_to_minor(int index)
524 {
525 	return index << PART_BITS;
526 }
527 
528 static int minor_to_index(int minor)
529 {
530 	return minor >> PART_BITS;
531 }
532 
533 static ssize_t serial_show(struct device *dev,
534 			   struct device_attribute *attr, char *buf)
535 {
536 	struct gendisk *disk = dev_to_disk(dev);
537 	int err;
538 
539 	/* sysfs gives us a PAGE_SIZE buffer */
540 	BUILD_BUG_ON(PAGE_SIZE < VIRTIO_BLK_ID_BYTES);
541 
542 	buf[VIRTIO_BLK_ID_BYTES] = '\0';
543 	err = virtblk_get_id(disk, buf);
544 	if (!err)
545 		return strlen(buf);
546 
547 	if (err == -EIO) /* Unsupported? Make it empty. */
548 		return 0;
549 
550 	return err;
551 }
552 
553 static DEVICE_ATTR_RO(serial);
554 
555 /* The queue's logical block size must be set before calling this */
556 static void virtblk_update_capacity(struct virtio_blk *vblk, bool resize)
557 {
558 	struct virtio_device *vdev = vblk->vdev;
559 	struct request_queue *q = vblk->disk->queue;
560 	char cap_str_2[10], cap_str_10[10];
561 	unsigned long long nblocks;
562 	u64 capacity;
563 
564 	/* Host must always specify the capacity. */
565 	virtio_cread(vdev, struct virtio_blk_config, capacity, &capacity);
566 
567 	nblocks = DIV_ROUND_UP_ULL(capacity, queue_logical_block_size(q) >> 9);
568 
569 	string_get_size(nblocks, queue_logical_block_size(q),
570 			STRING_UNITS_2, cap_str_2, sizeof(cap_str_2));
571 	string_get_size(nblocks, queue_logical_block_size(q),
572 			STRING_UNITS_10, cap_str_10, sizeof(cap_str_10));
573 
574 	dev_notice(&vdev->dev,
575 		   "[%s] %s%llu %d-byte logical blocks (%s/%s)\n",
576 		   vblk->disk->disk_name,
577 		   resize ? "new size: " : "",
578 		   nblocks,
579 		   queue_logical_block_size(q),
580 		   cap_str_10,
581 		   cap_str_2);
582 
583 	set_capacity_and_notify(vblk->disk, capacity);
584 }
585 
586 static void virtblk_config_changed_work(struct work_struct *work)
587 {
588 	struct virtio_blk *vblk =
589 		container_of(work, struct virtio_blk, config_work);
590 
591 	virtblk_update_capacity(vblk, true);
592 }
593 
594 static void virtblk_config_changed(struct virtio_device *vdev)
595 {
596 	struct virtio_blk *vblk = vdev->priv;
597 
598 	queue_work(virtblk_wq, &vblk->config_work);
599 }
600 
601 static int init_vq(struct virtio_blk *vblk)
602 {
603 	int err;
604 	int i;
605 	vq_callback_t **callbacks;
606 	const char **names;
607 	struct virtqueue **vqs;
608 	unsigned short num_vqs;
609 	unsigned int num_poll_vqs;
610 	struct virtio_device *vdev = vblk->vdev;
611 	struct irq_affinity desc = { 0, };
612 
613 	err = virtio_cread_feature(vdev, VIRTIO_BLK_F_MQ,
614 				   struct virtio_blk_config, num_queues,
615 				   &num_vqs);
616 	if (err)
617 		num_vqs = 1;
618 
619 	if (!err && !num_vqs) {
620 		dev_err(&vdev->dev, "MQ advertised but zero queues reported\n");
621 		return -EINVAL;
622 	}
623 
624 	num_vqs = min_t(unsigned int,
625 			min_not_zero(num_request_queues, nr_cpu_ids),
626 			num_vqs);
627 
628 	num_poll_vqs = min_t(unsigned int, poll_queues, num_vqs - 1);
629 
630 	vblk->io_queues[HCTX_TYPE_DEFAULT] = num_vqs - num_poll_vqs;
631 	vblk->io_queues[HCTX_TYPE_READ] = 0;
632 	vblk->io_queues[HCTX_TYPE_POLL] = num_poll_vqs;
633 
634 	dev_info(&vdev->dev, "%d/%d/%d default/read/poll queues\n",
635 				vblk->io_queues[HCTX_TYPE_DEFAULT],
636 				vblk->io_queues[HCTX_TYPE_READ],
637 				vblk->io_queues[HCTX_TYPE_POLL]);
638 
639 	vblk->vqs = kmalloc_array(num_vqs, sizeof(*vblk->vqs), GFP_KERNEL);
640 	if (!vblk->vqs)
641 		return -ENOMEM;
642 
643 	names = kmalloc_array(num_vqs, sizeof(*names), GFP_KERNEL);
644 	callbacks = kmalloc_array(num_vqs, sizeof(*callbacks), GFP_KERNEL);
645 	vqs = kmalloc_array(num_vqs, sizeof(*vqs), GFP_KERNEL);
646 	if (!names || !callbacks || !vqs) {
647 		err = -ENOMEM;
648 		goto out;
649 	}
650 
651 	for (i = 0; i < num_vqs - num_poll_vqs; i++) {
652 		callbacks[i] = virtblk_done;
653 		snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req.%d", i);
654 		names[i] = vblk->vqs[i].name;
655 	}
656 
657 	for (; i < num_vqs; i++) {
658 		callbacks[i] = NULL;
659 		snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req_poll.%d", i);
660 		names[i] = vblk->vqs[i].name;
661 	}
662 
663 	/* Discover virtqueues and write information to configuration.  */
664 	err = virtio_find_vqs(vdev, num_vqs, vqs, callbacks, names, &desc);
665 	if (err)
666 		goto out;
667 
668 	for (i = 0; i < num_vqs; i++) {
669 		spin_lock_init(&vblk->vqs[i].lock);
670 		vblk->vqs[i].vq = vqs[i];
671 	}
672 	vblk->num_vqs = num_vqs;
673 
674 out:
675 	kfree(vqs);
676 	kfree(callbacks);
677 	kfree(names);
678 	if (err)
679 		kfree(vblk->vqs);
680 	return err;
681 }
682 
683 /*
684  * Legacy naming scheme used for virtio devices.  We are stuck with it for
685  * virtio blk but don't ever use it for any new driver.
686  */
687 static int virtblk_name_format(char *prefix, int index, char *buf, int buflen)
688 {
689 	const int base = 'z' - 'a' + 1;
690 	char *begin = buf + strlen(prefix);
691 	char *end = buf + buflen;
692 	char *p;
693 	int unit;
694 
695 	p = end - 1;
696 	*p = '\0';
697 	unit = base;
698 	do {
699 		if (p == begin)
700 			return -EINVAL;
701 		*--p = 'a' + (index % unit);
702 		index = (index / unit) - 1;
703 	} while (index >= 0);
704 
705 	memmove(begin, p, end - p);
706 	memcpy(buf, prefix, strlen(prefix));
707 
708 	return 0;
709 }
710 
711 static int virtblk_get_cache_mode(struct virtio_device *vdev)
712 {
713 	u8 writeback;
714 	int err;
715 
716 	err = virtio_cread_feature(vdev, VIRTIO_BLK_F_CONFIG_WCE,
717 				   struct virtio_blk_config, wce,
718 				   &writeback);
719 
720 	/*
721 	 * If WCE is not configurable and flush is not available,
722 	 * assume no writeback cache is in use.
723 	 */
724 	if (err)
725 		writeback = virtio_has_feature(vdev, VIRTIO_BLK_F_FLUSH);
726 
727 	return writeback;
728 }
729 
730 static void virtblk_update_cache_mode(struct virtio_device *vdev)
731 {
732 	u8 writeback = virtblk_get_cache_mode(vdev);
733 	struct virtio_blk *vblk = vdev->priv;
734 
735 	blk_queue_write_cache(vblk->disk->queue, writeback, false);
736 }
737 
738 static const char *const virtblk_cache_types[] = {
739 	"write through", "write back"
740 };
741 
742 static ssize_t
743 cache_type_store(struct device *dev, struct device_attribute *attr,
744 		 const char *buf, size_t count)
745 {
746 	struct gendisk *disk = dev_to_disk(dev);
747 	struct virtio_blk *vblk = disk->private_data;
748 	struct virtio_device *vdev = vblk->vdev;
749 	int i;
750 
751 	BUG_ON(!virtio_has_feature(vblk->vdev, VIRTIO_BLK_F_CONFIG_WCE));
752 	i = sysfs_match_string(virtblk_cache_types, buf);
753 	if (i < 0)
754 		return i;
755 
756 	virtio_cwrite8(vdev, offsetof(struct virtio_blk_config, wce), i);
757 	virtblk_update_cache_mode(vdev);
758 	return count;
759 }
760 
761 static ssize_t
762 cache_type_show(struct device *dev, struct device_attribute *attr, char *buf)
763 {
764 	struct gendisk *disk = dev_to_disk(dev);
765 	struct virtio_blk *vblk = disk->private_data;
766 	u8 writeback = virtblk_get_cache_mode(vblk->vdev);
767 
768 	BUG_ON(writeback >= ARRAY_SIZE(virtblk_cache_types));
769 	return sysfs_emit(buf, "%s\n", virtblk_cache_types[writeback]);
770 }
771 
772 static DEVICE_ATTR_RW(cache_type);
773 
774 static struct attribute *virtblk_attrs[] = {
775 	&dev_attr_serial.attr,
776 	&dev_attr_cache_type.attr,
777 	NULL,
778 };
779 
780 static umode_t virtblk_attrs_are_visible(struct kobject *kobj,
781 		struct attribute *a, int n)
782 {
783 	struct device *dev = kobj_to_dev(kobj);
784 	struct gendisk *disk = dev_to_disk(dev);
785 	struct virtio_blk *vblk = disk->private_data;
786 	struct virtio_device *vdev = vblk->vdev;
787 
788 	if (a == &dev_attr_cache_type.attr &&
789 	    !virtio_has_feature(vdev, VIRTIO_BLK_F_CONFIG_WCE))
790 		return S_IRUGO;
791 
792 	return a->mode;
793 }
794 
795 static const struct attribute_group virtblk_attr_group = {
796 	.attrs = virtblk_attrs,
797 	.is_visible = virtblk_attrs_are_visible,
798 };
799 
800 static const struct attribute_group *virtblk_attr_groups[] = {
801 	&virtblk_attr_group,
802 	NULL,
803 };
804 
805 static int virtblk_map_queues(struct blk_mq_tag_set *set)
806 {
807 	struct virtio_blk *vblk = set->driver_data;
808 	int i, qoff;
809 
810 	for (i = 0, qoff = 0; i < set->nr_maps; i++) {
811 		struct blk_mq_queue_map *map = &set->map[i];
812 
813 		map->nr_queues = vblk->io_queues[i];
814 		map->queue_offset = qoff;
815 		qoff += map->nr_queues;
816 
817 		if (map->nr_queues == 0)
818 			continue;
819 
820 		/*
821 		 * Regular queues have interrupts and hence CPU affinity is
822 		 * defined by the core virtio code, but polling queues have
823 		 * no interrupts so we let the block layer assign CPU affinity.
824 		 */
825 		if (i == HCTX_TYPE_POLL)
826 			blk_mq_map_queues(&set->map[i]);
827 		else
828 			blk_mq_virtio_map_queues(&set->map[i], vblk->vdev, 0);
829 	}
830 
831 	return 0;
832 }
833 
834 static void virtblk_complete_batch(struct io_comp_batch *iob)
835 {
836 	struct request *req;
837 
838 	rq_list_for_each(&iob->req_list, req) {
839 		virtblk_unmap_data(req, blk_mq_rq_to_pdu(req));
840 		virtblk_cleanup_cmd(req);
841 	}
842 	blk_mq_end_request_batch(iob);
843 }
844 
845 static int virtblk_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
846 {
847 	struct virtio_blk *vblk = hctx->queue->queuedata;
848 	struct virtio_blk_vq *vq = get_virtio_blk_vq(hctx);
849 	struct virtblk_req *vbr;
850 	unsigned long flags;
851 	unsigned int len;
852 	int found = 0;
853 
854 	spin_lock_irqsave(&vq->lock, flags);
855 
856 	while ((vbr = virtqueue_get_buf(vq->vq, &len)) != NULL) {
857 		struct request *req = blk_mq_rq_from_pdu(vbr);
858 
859 		found++;
860 		if (!blk_mq_add_to_batch(req, iob, vbr->status,
861 						virtblk_complete_batch))
862 			blk_mq_complete_request(req);
863 	}
864 
865 	if (found)
866 		blk_mq_start_stopped_hw_queues(vblk->disk->queue, true);
867 
868 	spin_unlock_irqrestore(&vq->lock, flags);
869 
870 	return found;
871 }
872 
873 static const struct blk_mq_ops virtio_mq_ops = {
874 	.queue_rq	= virtio_queue_rq,
875 	.queue_rqs	= virtio_queue_rqs,
876 	.commit_rqs	= virtio_commit_rqs,
877 	.complete	= virtblk_request_done,
878 	.map_queues	= virtblk_map_queues,
879 	.poll		= virtblk_poll,
880 };
881 
882 static unsigned int virtblk_queue_depth;
883 module_param_named(queue_depth, virtblk_queue_depth, uint, 0444);
884 
885 static int virtblk_probe(struct virtio_device *vdev)
886 {
887 	struct virtio_blk *vblk;
888 	struct request_queue *q;
889 	int err, index;
890 
891 	u32 v, blk_size, max_size, sg_elems, opt_io_size;
892 	u16 min_io_size;
893 	u8 physical_block_exp, alignment_offset;
894 	unsigned int queue_depth;
895 
896 	if (!vdev->config->get) {
897 		dev_err(&vdev->dev, "%s failure: config access disabled\n",
898 			__func__);
899 		return -EINVAL;
900 	}
901 
902 	err = ida_simple_get(&vd_index_ida, 0, minor_to_index(1 << MINORBITS),
903 			     GFP_KERNEL);
904 	if (err < 0)
905 		goto out;
906 	index = err;
907 
908 	/* We need to know how many segments before we allocate. */
909 	err = virtio_cread_feature(vdev, VIRTIO_BLK_F_SEG_MAX,
910 				   struct virtio_blk_config, seg_max,
911 				   &sg_elems);
912 
913 	/* We need at least one SG element, whatever they say. */
914 	if (err || !sg_elems)
915 		sg_elems = 1;
916 
917 	/* Prevent integer overflows and honor max vq size */
918 	sg_elems = min_t(u32, sg_elems, VIRTIO_BLK_MAX_SG_ELEMS - 2);
919 
920 	vdev->priv = vblk = kmalloc(sizeof(*vblk), GFP_KERNEL);
921 	if (!vblk) {
922 		err = -ENOMEM;
923 		goto out_free_index;
924 	}
925 
926 	mutex_init(&vblk->vdev_mutex);
927 
928 	vblk->vdev = vdev;
929 
930 	INIT_WORK(&vblk->config_work, virtblk_config_changed_work);
931 
932 	err = init_vq(vblk);
933 	if (err)
934 		goto out_free_vblk;
935 
936 	/* Default queue sizing is to fill the ring. */
937 	if (!virtblk_queue_depth) {
938 		queue_depth = vblk->vqs[0].vq->num_free;
939 		/* ... but without indirect descs, we use 2 descs per req */
940 		if (!virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC))
941 			queue_depth /= 2;
942 	} else {
943 		queue_depth = virtblk_queue_depth;
944 	}
945 
946 	memset(&vblk->tag_set, 0, sizeof(vblk->tag_set));
947 	vblk->tag_set.ops = &virtio_mq_ops;
948 	vblk->tag_set.queue_depth = queue_depth;
949 	vblk->tag_set.numa_node = NUMA_NO_NODE;
950 	vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
951 	vblk->tag_set.cmd_size =
952 		sizeof(struct virtblk_req) +
953 		sizeof(struct scatterlist) * VIRTIO_BLK_INLINE_SG_CNT;
954 	vblk->tag_set.driver_data = vblk;
955 	vblk->tag_set.nr_hw_queues = vblk->num_vqs;
956 	vblk->tag_set.nr_maps = 1;
957 	if (vblk->io_queues[HCTX_TYPE_POLL])
958 		vblk->tag_set.nr_maps = 3;
959 
960 	err = blk_mq_alloc_tag_set(&vblk->tag_set);
961 	if (err)
962 		goto out_free_vq;
963 
964 	vblk->disk = blk_mq_alloc_disk(&vblk->tag_set, vblk);
965 	if (IS_ERR(vblk->disk)) {
966 		err = PTR_ERR(vblk->disk);
967 		goto out_free_tags;
968 	}
969 	q = vblk->disk->queue;
970 
971 	virtblk_name_format("vd", index, vblk->disk->disk_name, DISK_NAME_LEN);
972 
973 	vblk->disk->major = major;
974 	vblk->disk->first_minor = index_to_minor(index);
975 	vblk->disk->minors = 1 << PART_BITS;
976 	vblk->disk->private_data = vblk;
977 	vblk->disk->fops = &virtblk_fops;
978 	vblk->index = index;
979 
980 	/* configure queue flush support */
981 	virtblk_update_cache_mode(vdev);
982 
983 	/* If disk is read-only in the host, the guest should obey */
984 	if (virtio_has_feature(vdev, VIRTIO_BLK_F_RO))
985 		set_disk_ro(vblk->disk, 1);
986 
987 	/* We can handle whatever the host told us to handle. */
988 	blk_queue_max_segments(q, sg_elems);
989 
990 	/* No real sector limit. */
991 	blk_queue_max_hw_sectors(q, -1U);
992 
993 	max_size = virtio_max_dma_size(vdev);
994 
995 	/* Host can optionally specify maximum segment size and number of
996 	 * segments. */
997 	err = virtio_cread_feature(vdev, VIRTIO_BLK_F_SIZE_MAX,
998 				   struct virtio_blk_config, size_max, &v);
999 	if (!err)
1000 		max_size = min(max_size, v);
1001 
1002 	blk_queue_max_segment_size(q, max_size);
1003 
1004 	/* Host can optionally specify the block size of the device */
1005 	err = virtio_cread_feature(vdev, VIRTIO_BLK_F_BLK_SIZE,
1006 				   struct virtio_blk_config, blk_size,
1007 				   &blk_size);
1008 	if (!err) {
1009 		err = blk_validate_block_size(blk_size);
1010 		if (err) {
1011 			dev_err(&vdev->dev,
1012 				"virtio_blk: invalid block size: 0x%x\n",
1013 				blk_size);
1014 			goto out_cleanup_disk;
1015 		}
1016 
1017 		blk_queue_logical_block_size(q, blk_size);
1018 	} else
1019 		blk_size = queue_logical_block_size(q);
1020 
1021 	/* Use topology information if available */
1022 	err = virtio_cread_feature(vdev, VIRTIO_BLK_F_TOPOLOGY,
1023 				   struct virtio_blk_config, physical_block_exp,
1024 				   &physical_block_exp);
1025 	if (!err && physical_block_exp)
1026 		blk_queue_physical_block_size(q,
1027 				blk_size * (1 << physical_block_exp));
1028 
1029 	err = virtio_cread_feature(vdev, VIRTIO_BLK_F_TOPOLOGY,
1030 				   struct virtio_blk_config, alignment_offset,
1031 				   &alignment_offset);
1032 	if (!err && alignment_offset)
1033 		blk_queue_alignment_offset(q, blk_size * alignment_offset);
1034 
1035 	err = virtio_cread_feature(vdev, VIRTIO_BLK_F_TOPOLOGY,
1036 				   struct virtio_blk_config, min_io_size,
1037 				   &min_io_size);
1038 	if (!err && min_io_size)
1039 		blk_queue_io_min(q, blk_size * min_io_size);
1040 
1041 	err = virtio_cread_feature(vdev, VIRTIO_BLK_F_TOPOLOGY,
1042 				   struct virtio_blk_config, opt_io_size,
1043 				   &opt_io_size);
1044 	if (!err && opt_io_size)
1045 		blk_queue_io_opt(q, blk_size * opt_io_size);
1046 
1047 	if (virtio_has_feature(vdev, VIRTIO_BLK_F_DISCARD)) {
1048 		virtio_cread(vdev, struct virtio_blk_config,
1049 			     discard_sector_alignment, &v);
1050 		if (v)
1051 			q->limits.discard_granularity = v << SECTOR_SHIFT;
1052 		else
1053 			q->limits.discard_granularity = blk_size;
1054 
1055 		virtio_cread(vdev, struct virtio_blk_config,
1056 			     max_discard_sectors, &v);
1057 		blk_queue_max_discard_sectors(q, v ? v : UINT_MAX);
1058 
1059 		virtio_cread(vdev, struct virtio_blk_config, max_discard_seg,
1060 			     &v);
1061 
1062 		/*
1063 		 * max_discard_seg == 0 is out of spec but we always
1064 		 * handled it.
1065 		 */
1066 		if (!v)
1067 			v = sg_elems;
1068 		blk_queue_max_discard_segments(q,
1069 					       min(v, MAX_DISCARD_SEGMENTS));
1070 	}
1071 
1072 	if (virtio_has_feature(vdev, VIRTIO_BLK_F_WRITE_ZEROES)) {
1073 		virtio_cread(vdev, struct virtio_blk_config,
1074 			     max_write_zeroes_sectors, &v);
1075 		blk_queue_max_write_zeroes_sectors(q, v ? v : UINT_MAX);
1076 	}
1077 
1078 	virtblk_update_capacity(vblk, false);
1079 	virtio_device_ready(vdev);
1080 
1081 	err = device_add_disk(&vdev->dev, vblk->disk, virtblk_attr_groups);
1082 	if (err)
1083 		goto out_cleanup_disk;
1084 
1085 	return 0;
1086 
1087 out_cleanup_disk:
1088 	put_disk(vblk->disk);
1089 out_free_tags:
1090 	blk_mq_free_tag_set(&vblk->tag_set);
1091 out_free_vq:
1092 	vdev->config->del_vqs(vdev);
1093 	kfree(vblk->vqs);
1094 out_free_vblk:
1095 	kfree(vblk);
1096 out_free_index:
1097 	ida_simple_remove(&vd_index_ida, index);
1098 out:
1099 	return err;
1100 }
1101 
1102 static void virtblk_remove(struct virtio_device *vdev)
1103 {
1104 	struct virtio_blk *vblk = vdev->priv;
1105 
1106 	/* Make sure no work handler is accessing the device. */
1107 	flush_work(&vblk->config_work);
1108 
1109 	del_gendisk(vblk->disk);
1110 	blk_mq_free_tag_set(&vblk->tag_set);
1111 
1112 	mutex_lock(&vblk->vdev_mutex);
1113 
1114 	/* Stop all the virtqueues. */
1115 	virtio_reset_device(vdev);
1116 
1117 	/* Virtqueues are stopped, nothing can use vblk->vdev anymore. */
1118 	vblk->vdev = NULL;
1119 
1120 	vdev->config->del_vqs(vdev);
1121 	kfree(vblk->vqs);
1122 
1123 	mutex_unlock(&vblk->vdev_mutex);
1124 
1125 	put_disk(vblk->disk);
1126 }
1127 
1128 #ifdef CONFIG_PM_SLEEP
1129 static int virtblk_freeze(struct virtio_device *vdev)
1130 {
1131 	struct virtio_blk *vblk = vdev->priv;
1132 
1133 	/* Ensure we don't receive any more interrupts */
1134 	virtio_reset_device(vdev);
1135 
1136 	/* Make sure no work handler is accessing the device. */
1137 	flush_work(&vblk->config_work);
1138 
1139 	blk_mq_quiesce_queue(vblk->disk->queue);
1140 
1141 	vdev->config->del_vqs(vdev);
1142 	kfree(vblk->vqs);
1143 
1144 	return 0;
1145 }
1146 
1147 static int virtblk_restore(struct virtio_device *vdev)
1148 {
1149 	struct virtio_blk *vblk = vdev->priv;
1150 	int ret;
1151 
1152 	ret = init_vq(vdev->priv);
1153 	if (ret)
1154 		return ret;
1155 
1156 	virtio_device_ready(vdev);
1157 
1158 	blk_mq_unquiesce_queue(vblk->disk->queue);
1159 	return 0;
1160 }
1161 #endif
1162 
1163 static const struct virtio_device_id id_table[] = {
1164 	{ VIRTIO_ID_BLOCK, VIRTIO_DEV_ANY_ID },
1165 	{ 0 },
1166 };
1167 
1168 static unsigned int features_legacy[] = {
1169 	VIRTIO_BLK_F_SEG_MAX, VIRTIO_BLK_F_SIZE_MAX, VIRTIO_BLK_F_GEOMETRY,
1170 	VIRTIO_BLK_F_RO, VIRTIO_BLK_F_BLK_SIZE,
1171 	VIRTIO_BLK_F_FLUSH, VIRTIO_BLK_F_TOPOLOGY, VIRTIO_BLK_F_CONFIG_WCE,
1172 	VIRTIO_BLK_F_MQ, VIRTIO_BLK_F_DISCARD, VIRTIO_BLK_F_WRITE_ZEROES,
1173 }
1174 ;
1175 static unsigned int features[] = {
1176 	VIRTIO_BLK_F_SEG_MAX, VIRTIO_BLK_F_SIZE_MAX, VIRTIO_BLK_F_GEOMETRY,
1177 	VIRTIO_BLK_F_RO, VIRTIO_BLK_F_BLK_SIZE,
1178 	VIRTIO_BLK_F_FLUSH, VIRTIO_BLK_F_TOPOLOGY, VIRTIO_BLK_F_CONFIG_WCE,
1179 	VIRTIO_BLK_F_MQ, VIRTIO_BLK_F_DISCARD, VIRTIO_BLK_F_WRITE_ZEROES,
1180 };
1181 
1182 static struct virtio_driver virtio_blk = {
1183 	.feature_table			= features,
1184 	.feature_table_size		= ARRAY_SIZE(features),
1185 	.feature_table_legacy		= features_legacy,
1186 	.feature_table_size_legacy	= ARRAY_SIZE(features_legacy),
1187 	.driver.name			= KBUILD_MODNAME,
1188 	.driver.owner			= THIS_MODULE,
1189 	.id_table			= id_table,
1190 	.probe				= virtblk_probe,
1191 	.remove				= virtblk_remove,
1192 	.config_changed			= virtblk_config_changed,
1193 #ifdef CONFIG_PM_SLEEP
1194 	.freeze				= virtblk_freeze,
1195 	.restore			= virtblk_restore,
1196 #endif
1197 };
1198 
1199 static int __init virtio_blk_init(void)
1200 {
1201 	int error;
1202 
1203 	virtblk_wq = alloc_workqueue("virtio-blk", 0, 0);
1204 	if (!virtblk_wq)
1205 		return -ENOMEM;
1206 
1207 	major = register_blkdev(0, "virtblk");
1208 	if (major < 0) {
1209 		error = major;
1210 		goto out_destroy_workqueue;
1211 	}
1212 
1213 	error = register_virtio_driver(&virtio_blk);
1214 	if (error)
1215 		goto out_unregister_blkdev;
1216 	return 0;
1217 
1218 out_unregister_blkdev:
1219 	unregister_blkdev(major, "virtblk");
1220 out_destroy_workqueue:
1221 	destroy_workqueue(virtblk_wq);
1222 	return error;
1223 }
1224 
1225 static void __exit virtio_blk_fini(void)
1226 {
1227 	unregister_virtio_driver(&virtio_blk);
1228 	unregister_blkdev(major, "virtblk");
1229 	destroy_workqueue(virtblk_wq);
1230 }
1231 module_init(virtio_blk_init);
1232 module_exit(virtio_blk_fini);
1233 
1234 MODULE_DEVICE_TABLE(virtio, id_table);
1235 MODULE_DESCRIPTION("Virtio block driver");
1236 MODULE_LICENSE("GPL");
1237