xref: /openbmc/linux/drivers/scsi/virtio_scsi.c (revision 8730046c)
1 /*
2  * Virtio SCSI HBA driver
3  *
4  * Copyright IBM Corp. 2010
5  * Copyright Red Hat, Inc. 2011
6  *
7  * Authors:
8  *  Stefan Hajnoczi   <stefanha@linux.vnet.ibm.com>
9  *  Paolo Bonzini   <pbonzini@redhat.com>
10  *
11  * This work is licensed under the terms of the GNU GPL, version 2 or later.
12  * See the COPYING file in the top-level directory.
13  *
14  */
15 
16 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 
18 #include <linux/module.h>
19 #include <linux/slab.h>
20 #include <linux/mempool.h>
21 #include <linux/virtio.h>
22 #include <linux/virtio_ids.h>
23 #include <linux/virtio_config.h>
24 #include <linux/virtio_scsi.h>
25 #include <linux/cpu.h>
26 #include <linux/blkdev.h>
27 #include <scsi/scsi_host.h>
28 #include <scsi/scsi_device.h>
29 #include <scsi/scsi_cmnd.h>
30 #include <scsi/scsi_tcq.h>
31 #include <linux/seqlock.h>
32 
33 #define VIRTIO_SCSI_MEMPOOL_SZ 64
34 #define VIRTIO_SCSI_EVENT_LEN 8
35 #define VIRTIO_SCSI_VQ_BASE 2
36 
37 /* Command queue element */
38 struct virtio_scsi_cmd {
39 	struct scsi_cmnd *sc;
40 	struct completion *comp;
41 	union {
42 		struct virtio_scsi_cmd_req       cmd;
43 		struct virtio_scsi_cmd_req_pi    cmd_pi;
44 		struct virtio_scsi_ctrl_tmf_req  tmf;
45 		struct virtio_scsi_ctrl_an_req   an;
46 	} req;
47 	union {
48 		struct virtio_scsi_cmd_resp      cmd;
49 		struct virtio_scsi_ctrl_tmf_resp tmf;
50 		struct virtio_scsi_ctrl_an_resp  an;
51 		struct virtio_scsi_event         evt;
52 	} resp;
53 } ____cacheline_aligned_in_smp;
54 
55 struct virtio_scsi_event_node {
56 	struct virtio_scsi *vscsi;
57 	struct virtio_scsi_event event;
58 	struct work_struct work;
59 };
60 
61 struct virtio_scsi_vq {
62 	/* Protects vq */
63 	spinlock_t vq_lock;
64 
65 	struct virtqueue *vq;
66 };
67 
68 /*
69  * Per-target queue state.
70  *
71  * This struct holds the data needed by the queue steering policy.  When a
72  * target is sent multiple requests, we need to drive them to the same queue so
73  * that FIFO processing order is kept.  However, if a target was idle, we can
74  * choose a queue arbitrarily.  In this case the queue is chosen according to
75  * the current VCPU, so the driver expects the number of request queues to be
76  * equal to the number of VCPUs.  This makes it easy and fast to select the
77  * queue, and also lets the driver optimize the IRQ affinity for the virtqueues
78  * (each virtqueue's affinity is set to the CPU that "owns" the queue).
79  *
80  * tgt_seq is held to serialize reading and writing req_vq.
81  *
82  * Decrements of reqs are never concurrent with writes of req_vq: before the
83  * decrement reqs will be != 0; after the decrement the virtqueue completion
84  * routine will not use the req_vq so it can be changed by a new request.
85  * Thus they can happen outside the tgt_seq, provided of course we make reqs
86  * an atomic_t.
87  */
88 struct virtio_scsi_target_state {
89 	seqcount_t tgt_seq;
90 
91 	/* Count of outstanding requests. */
92 	atomic_t reqs;
93 
94 	/* Currently active virtqueue for requests sent to this target. */
95 	struct virtio_scsi_vq *req_vq;
96 };
97 
98 /* Driver instance state */
99 struct virtio_scsi {
100 	struct virtio_device *vdev;
101 
102 	/* Get some buffers ready for event vq */
103 	struct virtio_scsi_event_node event_list[VIRTIO_SCSI_EVENT_LEN];
104 
105 	u32 num_queues;
106 
107 	/* If the affinity hint is set for virtqueues */
108 	bool affinity_hint_set;
109 
110 	struct hlist_node node;
111 	struct hlist_node node_dead;
112 
113 	/* Protected by event_vq lock */
114 	bool stop_events;
115 
116 	struct virtio_scsi_vq ctrl_vq;
117 	struct virtio_scsi_vq event_vq;
118 	struct virtio_scsi_vq req_vqs[];
119 };
120 
121 static enum cpuhp_state virtioscsi_online;
122 static struct kmem_cache *virtscsi_cmd_cache;
123 static mempool_t *virtscsi_cmd_pool;
124 
125 static inline struct Scsi_Host *virtio_scsi_host(struct virtio_device *vdev)
126 {
127 	return vdev->priv;
128 }
129 
130 static void virtscsi_compute_resid(struct scsi_cmnd *sc, u32 resid)
131 {
132 	if (!resid)
133 		return;
134 
135 	if (!scsi_bidi_cmnd(sc)) {
136 		scsi_set_resid(sc, resid);
137 		return;
138 	}
139 
140 	scsi_in(sc)->resid = min(resid, scsi_in(sc)->length);
141 	scsi_out(sc)->resid = resid - scsi_in(sc)->resid;
142 }
143 
144 /**
145  * virtscsi_complete_cmd - finish a scsi_cmd and invoke scsi_done
146  *
147  * Called with vq_lock held.
148  */
149 static void virtscsi_complete_cmd(struct virtio_scsi *vscsi, void *buf)
150 {
151 	struct virtio_scsi_cmd *cmd = buf;
152 	struct scsi_cmnd *sc = cmd->sc;
153 	struct virtio_scsi_cmd_resp *resp = &cmd->resp.cmd;
154 	struct virtio_scsi_target_state *tgt =
155 				scsi_target(sc->device)->hostdata;
156 
157 	dev_dbg(&sc->device->sdev_gendev,
158 		"cmd %p response %u status %#02x sense_len %u\n",
159 		sc, resp->response, resp->status, resp->sense_len);
160 
161 	sc->result = resp->status;
162 	virtscsi_compute_resid(sc, virtio32_to_cpu(vscsi->vdev, resp->resid));
163 	switch (resp->response) {
164 	case VIRTIO_SCSI_S_OK:
165 		set_host_byte(sc, DID_OK);
166 		break;
167 	case VIRTIO_SCSI_S_OVERRUN:
168 		set_host_byte(sc, DID_ERROR);
169 		break;
170 	case VIRTIO_SCSI_S_ABORTED:
171 		set_host_byte(sc, DID_ABORT);
172 		break;
173 	case VIRTIO_SCSI_S_BAD_TARGET:
174 		set_host_byte(sc, DID_BAD_TARGET);
175 		break;
176 	case VIRTIO_SCSI_S_RESET:
177 		set_host_byte(sc, DID_RESET);
178 		break;
179 	case VIRTIO_SCSI_S_BUSY:
180 		set_host_byte(sc, DID_BUS_BUSY);
181 		break;
182 	case VIRTIO_SCSI_S_TRANSPORT_FAILURE:
183 		set_host_byte(sc, DID_TRANSPORT_DISRUPTED);
184 		break;
185 	case VIRTIO_SCSI_S_TARGET_FAILURE:
186 		set_host_byte(sc, DID_TARGET_FAILURE);
187 		break;
188 	case VIRTIO_SCSI_S_NEXUS_FAILURE:
189 		set_host_byte(sc, DID_NEXUS_FAILURE);
190 		break;
191 	default:
192 		scmd_printk(KERN_WARNING, sc, "Unknown response %d",
193 			    resp->response);
194 		/* fall through */
195 	case VIRTIO_SCSI_S_FAILURE:
196 		set_host_byte(sc, DID_ERROR);
197 		break;
198 	}
199 
200 	WARN_ON(virtio32_to_cpu(vscsi->vdev, resp->sense_len) >
201 		VIRTIO_SCSI_SENSE_SIZE);
202 	if (sc->sense_buffer) {
203 		memcpy(sc->sense_buffer, resp->sense,
204 		       min_t(u32,
205 			     virtio32_to_cpu(vscsi->vdev, resp->sense_len),
206 			     VIRTIO_SCSI_SENSE_SIZE));
207 		if (resp->sense_len)
208 			set_driver_byte(sc, DRIVER_SENSE);
209 	}
210 
211 	sc->scsi_done(sc);
212 
213 	atomic_dec(&tgt->reqs);
214 }
215 
216 static void virtscsi_vq_done(struct virtio_scsi *vscsi,
217 			     struct virtio_scsi_vq *virtscsi_vq,
218 			     void (*fn)(struct virtio_scsi *vscsi, void *buf))
219 {
220 	void *buf;
221 	unsigned int len;
222 	unsigned long flags;
223 	struct virtqueue *vq = virtscsi_vq->vq;
224 
225 	spin_lock_irqsave(&virtscsi_vq->vq_lock, flags);
226 	do {
227 		virtqueue_disable_cb(vq);
228 		while ((buf = virtqueue_get_buf(vq, &len)) != NULL)
229 			fn(vscsi, buf);
230 
231 		if (unlikely(virtqueue_is_broken(vq)))
232 			break;
233 	} while (!virtqueue_enable_cb(vq));
234 	spin_unlock_irqrestore(&virtscsi_vq->vq_lock, flags);
235 }
236 
237 static void virtscsi_req_done(struct virtqueue *vq)
238 {
239 	struct Scsi_Host *sh = virtio_scsi_host(vq->vdev);
240 	struct virtio_scsi *vscsi = shost_priv(sh);
241 	int index = vq->index - VIRTIO_SCSI_VQ_BASE;
242 	struct virtio_scsi_vq *req_vq = &vscsi->req_vqs[index];
243 
244 	virtscsi_vq_done(vscsi, req_vq, virtscsi_complete_cmd);
245 };
246 
247 static void virtscsi_poll_requests(struct virtio_scsi *vscsi)
248 {
249 	int i, num_vqs;
250 
251 	num_vqs = vscsi->num_queues;
252 	for (i = 0; i < num_vqs; i++)
253 		virtscsi_vq_done(vscsi, &vscsi->req_vqs[i],
254 				 virtscsi_complete_cmd);
255 }
256 
257 static void virtscsi_complete_free(struct virtio_scsi *vscsi, void *buf)
258 {
259 	struct virtio_scsi_cmd *cmd = buf;
260 
261 	if (cmd->comp)
262 		complete(cmd->comp);
263 }
264 
265 static void virtscsi_ctrl_done(struct virtqueue *vq)
266 {
267 	struct Scsi_Host *sh = virtio_scsi_host(vq->vdev);
268 	struct virtio_scsi *vscsi = shost_priv(sh);
269 
270 	virtscsi_vq_done(vscsi, &vscsi->ctrl_vq, virtscsi_complete_free);
271 };
272 
273 static void virtscsi_handle_event(struct work_struct *work);
274 
275 static int virtscsi_kick_event(struct virtio_scsi *vscsi,
276 			       struct virtio_scsi_event_node *event_node)
277 {
278 	int err;
279 	struct scatterlist sg;
280 	unsigned long flags;
281 
282 	INIT_WORK(&event_node->work, virtscsi_handle_event);
283 	sg_init_one(&sg, &event_node->event, sizeof(struct virtio_scsi_event));
284 
285 	spin_lock_irqsave(&vscsi->event_vq.vq_lock, flags);
286 
287 	err = virtqueue_add_inbuf(vscsi->event_vq.vq, &sg, 1, event_node,
288 				  GFP_ATOMIC);
289 	if (!err)
290 		virtqueue_kick(vscsi->event_vq.vq);
291 
292 	spin_unlock_irqrestore(&vscsi->event_vq.vq_lock, flags);
293 
294 	return err;
295 }
296 
297 static int virtscsi_kick_event_all(struct virtio_scsi *vscsi)
298 {
299 	int i;
300 
301 	for (i = 0; i < VIRTIO_SCSI_EVENT_LEN; i++) {
302 		vscsi->event_list[i].vscsi = vscsi;
303 		virtscsi_kick_event(vscsi, &vscsi->event_list[i]);
304 	}
305 
306 	return 0;
307 }
308 
309 static void virtscsi_cancel_event_work(struct virtio_scsi *vscsi)
310 {
311 	int i;
312 
313 	/* Stop scheduling work before calling cancel_work_sync.  */
314 	spin_lock_irq(&vscsi->event_vq.vq_lock);
315 	vscsi->stop_events = true;
316 	spin_unlock_irq(&vscsi->event_vq.vq_lock);
317 
318 	for (i = 0; i < VIRTIO_SCSI_EVENT_LEN; i++)
319 		cancel_work_sync(&vscsi->event_list[i].work);
320 }
321 
322 static void virtscsi_handle_transport_reset(struct virtio_scsi *vscsi,
323 					    struct virtio_scsi_event *event)
324 {
325 	struct scsi_device *sdev;
326 	struct Scsi_Host *shost = virtio_scsi_host(vscsi->vdev);
327 	unsigned int target = event->lun[1];
328 	unsigned int lun = (event->lun[2] << 8) | event->lun[3];
329 
330 	switch (virtio32_to_cpu(vscsi->vdev, event->reason)) {
331 	case VIRTIO_SCSI_EVT_RESET_RESCAN:
332 		scsi_add_device(shost, 0, target, lun);
333 		break;
334 	case VIRTIO_SCSI_EVT_RESET_REMOVED:
335 		sdev = scsi_device_lookup(shost, 0, target, lun);
336 		if (sdev) {
337 			scsi_remove_device(sdev);
338 			scsi_device_put(sdev);
339 		} else {
340 			pr_err("SCSI device %d 0 %d %d not found\n",
341 				shost->host_no, target, lun);
342 		}
343 		break;
344 	default:
345 		pr_info("Unsupport virtio scsi event reason %x\n", event->reason);
346 	}
347 }
348 
349 static void virtscsi_handle_param_change(struct virtio_scsi *vscsi,
350 					 struct virtio_scsi_event *event)
351 {
352 	struct scsi_device *sdev;
353 	struct Scsi_Host *shost = virtio_scsi_host(vscsi->vdev);
354 	unsigned int target = event->lun[1];
355 	unsigned int lun = (event->lun[2] << 8) | event->lun[3];
356 	u8 asc = virtio32_to_cpu(vscsi->vdev, event->reason) & 255;
357 	u8 ascq = virtio32_to_cpu(vscsi->vdev, event->reason) >> 8;
358 
359 	sdev = scsi_device_lookup(shost, 0, target, lun);
360 	if (!sdev) {
361 		pr_err("SCSI device %d 0 %d %d not found\n",
362 			shost->host_no, target, lun);
363 		return;
364 	}
365 
366 	/* Handle "Parameters changed", "Mode parameters changed", and
367 	   "Capacity data has changed".  */
368 	if (asc == 0x2a && (ascq == 0x00 || ascq == 0x01 || ascq == 0x09))
369 		scsi_rescan_device(&sdev->sdev_gendev);
370 
371 	scsi_device_put(sdev);
372 }
373 
374 static void virtscsi_handle_event(struct work_struct *work)
375 {
376 	struct virtio_scsi_event_node *event_node =
377 		container_of(work, struct virtio_scsi_event_node, work);
378 	struct virtio_scsi *vscsi = event_node->vscsi;
379 	struct virtio_scsi_event *event = &event_node->event;
380 
381 	if (event->event &
382 	    cpu_to_virtio32(vscsi->vdev, VIRTIO_SCSI_T_EVENTS_MISSED)) {
383 		event->event &= ~cpu_to_virtio32(vscsi->vdev,
384 						   VIRTIO_SCSI_T_EVENTS_MISSED);
385 		scsi_scan_host(virtio_scsi_host(vscsi->vdev));
386 	}
387 
388 	switch (virtio32_to_cpu(vscsi->vdev, event->event)) {
389 	case VIRTIO_SCSI_T_NO_EVENT:
390 		break;
391 	case VIRTIO_SCSI_T_TRANSPORT_RESET:
392 		virtscsi_handle_transport_reset(vscsi, event);
393 		break;
394 	case VIRTIO_SCSI_T_PARAM_CHANGE:
395 		virtscsi_handle_param_change(vscsi, event);
396 		break;
397 	default:
398 		pr_err("Unsupport virtio scsi event %x\n", event->event);
399 	}
400 	virtscsi_kick_event(vscsi, event_node);
401 }
402 
403 static void virtscsi_complete_event(struct virtio_scsi *vscsi, void *buf)
404 {
405 	struct virtio_scsi_event_node *event_node = buf;
406 
407 	if (!vscsi->stop_events)
408 		queue_work(system_freezable_wq, &event_node->work);
409 }
410 
411 static void virtscsi_event_done(struct virtqueue *vq)
412 {
413 	struct Scsi_Host *sh = virtio_scsi_host(vq->vdev);
414 	struct virtio_scsi *vscsi = shost_priv(sh);
415 
416 	virtscsi_vq_done(vscsi, &vscsi->event_vq, virtscsi_complete_event);
417 };
418 
419 /**
420  * virtscsi_add_cmd - add a virtio_scsi_cmd to a virtqueue
421  * @vq		: the struct virtqueue we're talking about
422  * @cmd		: command structure
423  * @req_size	: size of the request buffer
424  * @resp_size	: size of the response buffer
425  */
426 static int virtscsi_add_cmd(struct virtqueue *vq,
427 			    struct virtio_scsi_cmd *cmd,
428 			    size_t req_size, size_t resp_size)
429 {
430 	struct scsi_cmnd *sc = cmd->sc;
431 	struct scatterlist *sgs[6], req, resp;
432 	struct sg_table *out, *in;
433 	unsigned out_num = 0, in_num = 0;
434 
435 	out = in = NULL;
436 
437 	if (sc && sc->sc_data_direction != DMA_NONE) {
438 		if (sc->sc_data_direction != DMA_FROM_DEVICE)
439 			out = &scsi_out(sc)->table;
440 		if (sc->sc_data_direction != DMA_TO_DEVICE)
441 			in = &scsi_in(sc)->table;
442 	}
443 
444 	/* Request header.  */
445 	sg_init_one(&req, &cmd->req, req_size);
446 	sgs[out_num++] = &req;
447 
448 	/* Data-out buffer.  */
449 	if (out) {
450 		/* Place WRITE protection SGLs before Data OUT payload */
451 		if (scsi_prot_sg_count(sc))
452 			sgs[out_num++] = scsi_prot_sglist(sc);
453 		sgs[out_num++] = out->sgl;
454 	}
455 
456 	/* Response header.  */
457 	sg_init_one(&resp, &cmd->resp, resp_size);
458 	sgs[out_num + in_num++] = &resp;
459 
460 	/* Data-in buffer */
461 	if (in) {
462 		/* Place READ protection SGLs before Data IN payload */
463 		if (scsi_prot_sg_count(sc))
464 			sgs[out_num + in_num++] = scsi_prot_sglist(sc);
465 		sgs[out_num + in_num++] = in->sgl;
466 	}
467 
468 	return virtqueue_add_sgs(vq, sgs, out_num, in_num, cmd, GFP_ATOMIC);
469 }
470 
471 static int virtscsi_kick_cmd(struct virtio_scsi_vq *vq,
472 			     struct virtio_scsi_cmd *cmd,
473 			     size_t req_size, size_t resp_size)
474 {
475 	unsigned long flags;
476 	int err;
477 	bool needs_kick = false;
478 
479 	spin_lock_irqsave(&vq->vq_lock, flags);
480 	err = virtscsi_add_cmd(vq->vq, cmd, req_size, resp_size);
481 	if (!err)
482 		needs_kick = virtqueue_kick_prepare(vq->vq);
483 
484 	spin_unlock_irqrestore(&vq->vq_lock, flags);
485 
486 	if (needs_kick)
487 		virtqueue_notify(vq->vq);
488 	return err;
489 }
490 
491 static void virtio_scsi_init_hdr(struct virtio_device *vdev,
492 				 struct virtio_scsi_cmd_req *cmd,
493 				 struct scsi_cmnd *sc)
494 {
495 	cmd->lun[0] = 1;
496 	cmd->lun[1] = sc->device->id;
497 	cmd->lun[2] = (sc->device->lun >> 8) | 0x40;
498 	cmd->lun[3] = sc->device->lun & 0xff;
499 	cmd->tag = cpu_to_virtio64(vdev, (unsigned long)sc);
500 	cmd->task_attr = VIRTIO_SCSI_S_SIMPLE;
501 	cmd->prio = 0;
502 	cmd->crn = 0;
503 }
504 
505 #ifdef CONFIG_BLK_DEV_INTEGRITY
506 static void virtio_scsi_init_hdr_pi(struct virtio_device *vdev,
507 				    struct virtio_scsi_cmd_req_pi *cmd_pi,
508 				    struct scsi_cmnd *sc)
509 {
510 	struct request *rq = sc->request;
511 	struct blk_integrity *bi;
512 
513 	virtio_scsi_init_hdr(vdev, (struct virtio_scsi_cmd_req *)cmd_pi, sc);
514 
515 	if (!rq || !scsi_prot_sg_count(sc))
516 		return;
517 
518 	bi = blk_get_integrity(rq->rq_disk);
519 
520 	if (sc->sc_data_direction == DMA_TO_DEVICE)
521 		cmd_pi->pi_bytesout = cpu_to_virtio32(vdev,
522 							blk_rq_sectors(rq) *
523 							bi->tuple_size);
524 	else if (sc->sc_data_direction == DMA_FROM_DEVICE)
525 		cmd_pi->pi_bytesin = cpu_to_virtio32(vdev,
526 						       blk_rq_sectors(rq) *
527 						       bi->tuple_size);
528 }
529 #endif
530 
531 static int virtscsi_queuecommand(struct virtio_scsi *vscsi,
532 				 struct virtio_scsi_vq *req_vq,
533 				 struct scsi_cmnd *sc)
534 {
535 	struct Scsi_Host *shost = virtio_scsi_host(vscsi->vdev);
536 	struct virtio_scsi_cmd *cmd = scsi_cmd_priv(sc);
537 	int req_size;
538 
539 	BUG_ON(scsi_sg_count(sc) > shost->sg_tablesize);
540 
541 	/* TODO: check feature bit and fail if unsupported?  */
542 	BUG_ON(sc->sc_data_direction == DMA_BIDIRECTIONAL);
543 
544 	dev_dbg(&sc->device->sdev_gendev,
545 		"cmd %p CDB: %#02x\n", sc, sc->cmnd[0]);
546 
547 	memset(cmd, 0, sizeof(*cmd));
548 	cmd->sc = sc;
549 
550 	BUG_ON(sc->cmd_len > VIRTIO_SCSI_CDB_SIZE);
551 
552 #ifdef CONFIG_BLK_DEV_INTEGRITY
553 	if (virtio_has_feature(vscsi->vdev, VIRTIO_SCSI_F_T10_PI)) {
554 		virtio_scsi_init_hdr_pi(vscsi->vdev, &cmd->req.cmd_pi, sc);
555 		memcpy(cmd->req.cmd_pi.cdb, sc->cmnd, sc->cmd_len);
556 		req_size = sizeof(cmd->req.cmd_pi);
557 	} else
558 #endif
559 	{
560 		virtio_scsi_init_hdr(vscsi->vdev, &cmd->req.cmd, sc);
561 		memcpy(cmd->req.cmd.cdb, sc->cmnd, sc->cmd_len);
562 		req_size = sizeof(cmd->req.cmd);
563 	}
564 
565 	if (virtscsi_kick_cmd(req_vq, cmd, req_size, sizeof(cmd->resp.cmd)) != 0)
566 		return SCSI_MLQUEUE_HOST_BUSY;
567 	return 0;
568 }
569 
570 static int virtscsi_queuecommand_single(struct Scsi_Host *sh,
571 					struct scsi_cmnd *sc)
572 {
573 	struct virtio_scsi *vscsi = shost_priv(sh);
574 	struct virtio_scsi_target_state *tgt =
575 				scsi_target(sc->device)->hostdata;
576 
577 	atomic_inc(&tgt->reqs);
578 	return virtscsi_queuecommand(vscsi, &vscsi->req_vqs[0], sc);
579 }
580 
581 static struct virtio_scsi_vq *virtscsi_pick_vq_mq(struct virtio_scsi *vscsi,
582 						  struct scsi_cmnd *sc)
583 {
584 	u32 tag = blk_mq_unique_tag(sc->request);
585 	u16 hwq = blk_mq_unique_tag_to_hwq(tag);
586 
587 	return &vscsi->req_vqs[hwq];
588 }
589 
590 static struct virtio_scsi_vq *virtscsi_pick_vq(struct virtio_scsi *vscsi,
591 					       struct virtio_scsi_target_state *tgt)
592 {
593 	struct virtio_scsi_vq *vq;
594 	unsigned long flags;
595 	u32 queue_num;
596 
597 	local_irq_save(flags);
598 	if (atomic_inc_return(&tgt->reqs) > 1) {
599 		unsigned long seq;
600 
601 		do {
602 			seq = read_seqcount_begin(&tgt->tgt_seq);
603 			vq = tgt->req_vq;
604 		} while (read_seqcount_retry(&tgt->tgt_seq, seq));
605 	} else {
606 		/* no writes can be concurrent because of atomic_t */
607 		write_seqcount_begin(&tgt->tgt_seq);
608 
609 		/* keep previous req_vq if a reader just arrived */
610 		if (unlikely(atomic_read(&tgt->reqs) > 1)) {
611 			vq = tgt->req_vq;
612 			goto unlock;
613 		}
614 
615 		queue_num = smp_processor_id();
616 		while (unlikely(queue_num >= vscsi->num_queues))
617 			queue_num -= vscsi->num_queues;
618 		tgt->req_vq = vq = &vscsi->req_vqs[queue_num];
619  unlock:
620 		write_seqcount_end(&tgt->tgt_seq);
621 	}
622 	local_irq_restore(flags);
623 
624 	return vq;
625 }
626 
627 static int virtscsi_queuecommand_multi(struct Scsi_Host *sh,
628 				       struct scsi_cmnd *sc)
629 {
630 	struct virtio_scsi *vscsi = shost_priv(sh);
631 	struct virtio_scsi_target_state *tgt =
632 				scsi_target(sc->device)->hostdata;
633 	struct virtio_scsi_vq *req_vq;
634 
635 	if (shost_use_blk_mq(sh))
636 		req_vq = virtscsi_pick_vq_mq(vscsi, sc);
637 	else
638 		req_vq = virtscsi_pick_vq(vscsi, tgt);
639 
640 	return virtscsi_queuecommand(vscsi, req_vq, sc);
641 }
642 
643 static int virtscsi_tmf(struct virtio_scsi *vscsi, struct virtio_scsi_cmd *cmd)
644 {
645 	DECLARE_COMPLETION_ONSTACK(comp);
646 	int ret = FAILED;
647 
648 	cmd->comp = &comp;
649 	if (virtscsi_kick_cmd(&vscsi->ctrl_vq, cmd,
650 			      sizeof cmd->req.tmf, sizeof cmd->resp.tmf) < 0)
651 		goto out;
652 
653 	wait_for_completion(&comp);
654 	if (cmd->resp.tmf.response == VIRTIO_SCSI_S_OK ||
655 	    cmd->resp.tmf.response == VIRTIO_SCSI_S_FUNCTION_SUCCEEDED)
656 		ret = SUCCESS;
657 
658 	/*
659 	 * The spec guarantees that all requests related to the TMF have
660 	 * been completed, but the callback might not have run yet if
661 	 * we're using independent interrupts (e.g. MSI).  Poll the
662 	 * virtqueues once.
663 	 *
664 	 * In the abort case, sc->scsi_done will do nothing, because
665 	 * the block layer must have detected a timeout and as a result
666 	 * REQ_ATOM_COMPLETE has been set.
667 	 */
668 	virtscsi_poll_requests(vscsi);
669 
670 out:
671 	mempool_free(cmd, virtscsi_cmd_pool);
672 	return ret;
673 }
674 
675 static int virtscsi_device_reset(struct scsi_cmnd *sc)
676 {
677 	struct virtio_scsi *vscsi = shost_priv(sc->device->host);
678 	struct virtio_scsi_cmd *cmd;
679 
680 	sdev_printk(KERN_INFO, sc->device, "device reset\n");
681 	cmd = mempool_alloc(virtscsi_cmd_pool, GFP_NOIO);
682 	if (!cmd)
683 		return FAILED;
684 
685 	memset(cmd, 0, sizeof(*cmd));
686 	cmd->sc = sc;
687 	cmd->req.tmf = (struct virtio_scsi_ctrl_tmf_req){
688 		.type = VIRTIO_SCSI_T_TMF,
689 		.subtype = cpu_to_virtio32(vscsi->vdev,
690 					     VIRTIO_SCSI_T_TMF_LOGICAL_UNIT_RESET),
691 		.lun[0] = 1,
692 		.lun[1] = sc->device->id,
693 		.lun[2] = (sc->device->lun >> 8) | 0x40,
694 		.lun[3] = sc->device->lun & 0xff,
695 	};
696 	return virtscsi_tmf(vscsi, cmd);
697 }
698 
699 /**
700  * virtscsi_change_queue_depth() - Change a virtscsi target's queue depth
701  * @sdev:	Virtscsi target whose queue depth to change
702  * @qdepth:	New queue depth
703  */
704 static int virtscsi_change_queue_depth(struct scsi_device *sdev, int qdepth)
705 {
706 	struct Scsi_Host *shost = sdev->host;
707 	int max_depth = shost->cmd_per_lun;
708 
709 	return scsi_change_queue_depth(sdev, min(max_depth, qdepth));
710 }
711 
712 static int virtscsi_abort(struct scsi_cmnd *sc)
713 {
714 	struct virtio_scsi *vscsi = shost_priv(sc->device->host);
715 	struct virtio_scsi_cmd *cmd;
716 
717 	scmd_printk(KERN_INFO, sc, "abort\n");
718 	cmd = mempool_alloc(virtscsi_cmd_pool, GFP_NOIO);
719 	if (!cmd)
720 		return FAILED;
721 
722 	memset(cmd, 0, sizeof(*cmd));
723 	cmd->sc = sc;
724 	cmd->req.tmf = (struct virtio_scsi_ctrl_tmf_req){
725 		.type = VIRTIO_SCSI_T_TMF,
726 		.subtype = VIRTIO_SCSI_T_TMF_ABORT_TASK,
727 		.lun[0] = 1,
728 		.lun[1] = sc->device->id,
729 		.lun[2] = (sc->device->lun >> 8) | 0x40,
730 		.lun[3] = sc->device->lun & 0xff,
731 		.tag = cpu_to_virtio64(vscsi->vdev, (unsigned long)sc),
732 	};
733 	return virtscsi_tmf(vscsi, cmd);
734 }
735 
736 static int virtscsi_target_alloc(struct scsi_target *starget)
737 {
738 	struct Scsi_Host *sh = dev_to_shost(starget->dev.parent);
739 	struct virtio_scsi *vscsi = shost_priv(sh);
740 
741 	struct virtio_scsi_target_state *tgt =
742 				kmalloc(sizeof(*tgt), GFP_KERNEL);
743 	if (!tgt)
744 		return -ENOMEM;
745 
746 	seqcount_init(&tgt->tgt_seq);
747 	atomic_set(&tgt->reqs, 0);
748 	tgt->req_vq = &vscsi->req_vqs[0];
749 
750 	starget->hostdata = tgt;
751 	return 0;
752 }
753 
754 static void virtscsi_target_destroy(struct scsi_target *starget)
755 {
756 	struct virtio_scsi_target_state *tgt = starget->hostdata;
757 	kfree(tgt);
758 }
759 
760 static struct scsi_host_template virtscsi_host_template_single = {
761 	.module = THIS_MODULE,
762 	.name = "Virtio SCSI HBA",
763 	.proc_name = "virtio_scsi",
764 	.this_id = -1,
765 	.cmd_size = sizeof(struct virtio_scsi_cmd),
766 	.queuecommand = virtscsi_queuecommand_single,
767 	.change_queue_depth = virtscsi_change_queue_depth,
768 	.eh_abort_handler = virtscsi_abort,
769 	.eh_device_reset_handler = virtscsi_device_reset,
770 
771 	.can_queue = 1024,
772 	.dma_boundary = UINT_MAX,
773 	.use_clustering = ENABLE_CLUSTERING,
774 	.target_alloc = virtscsi_target_alloc,
775 	.target_destroy = virtscsi_target_destroy,
776 	.track_queue_depth = 1,
777 };
778 
779 static struct scsi_host_template virtscsi_host_template_multi = {
780 	.module = THIS_MODULE,
781 	.name = "Virtio SCSI HBA",
782 	.proc_name = "virtio_scsi",
783 	.this_id = -1,
784 	.cmd_size = sizeof(struct virtio_scsi_cmd),
785 	.queuecommand = virtscsi_queuecommand_multi,
786 	.change_queue_depth = virtscsi_change_queue_depth,
787 	.eh_abort_handler = virtscsi_abort,
788 	.eh_device_reset_handler = virtscsi_device_reset,
789 
790 	.can_queue = 1024,
791 	.dma_boundary = UINT_MAX,
792 	.use_clustering = ENABLE_CLUSTERING,
793 	.target_alloc = virtscsi_target_alloc,
794 	.target_destroy = virtscsi_target_destroy,
795 	.track_queue_depth = 1,
796 };
797 
798 #define virtscsi_config_get(vdev, fld) \
799 	({ \
800 		typeof(((struct virtio_scsi_config *)0)->fld) __val; \
801 		virtio_cread(vdev, struct virtio_scsi_config, fld, &__val); \
802 		__val; \
803 	})
804 
805 #define virtscsi_config_set(vdev, fld, val) \
806 	do { \
807 		typeof(((struct virtio_scsi_config *)0)->fld) __val = (val); \
808 		virtio_cwrite(vdev, struct virtio_scsi_config, fld, &__val); \
809 	} while(0)
810 
811 static void __virtscsi_set_affinity(struct virtio_scsi *vscsi, bool affinity)
812 {
813 	int i;
814 	int cpu;
815 
816 	/* In multiqueue mode, when the number of cpu is equal
817 	 * to the number of request queues, we let the qeueues
818 	 * to be private to one cpu by setting the affinity hint
819 	 * to eliminate the contention.
820 	 */
821 	if ((vscsi->num_queues == 1 ||
822 	     vscsi->num_queues != num_online_cpus()) && affinity) {
823 		if (vscsi->affinity_hint_set)
824 			affinity = false;
825 		else
826 			return;
827 	}
828 
829 	if (affinity) {
830 		i = 0;
831 		for_each_online_cpu(cpu) {
832 			virtqueue_set_affinity(vscsi->req_vqs[i].vq, cpu);
833 			i++;
834 		}
835 
836 		vscsi->affinity_hint_set = true;
837 	} else {
838 		for (i = 0; i < vscsi->num_queues; i++) {
839 			if (!vscsi->req_vqs[i].vq)
840 				continue;
841 
842 			virtqueue_set_affinity(vscsi->req_vqs[i].vq, -1);
843 		}
844 
845 		vscsi->affinity_hint_set = false;
846 	}
847 }
848 
849 static void virtscsi_set_affinity(struct virtio_scsi *vscsi, bool affinity)
850 {
851 	get_online_cpus();
852 	__virtscsi_set_affinity(vscsi, affinity);
853 	put_online_cpus();
854 }
855 
856 static int virtscsi_cpu_online(unsigned int cpu, struct hlist_node *node)
857 {
858 	struct virtio_scsi *vscsi = hlist_entry_safe(node, struct virtio_scsi,
859 						     node);
860 	__virtscsi_set_affinity(vscsi, true);
861 	return 0;
862 }
863 
864 static int virtscsi_cpu_notif_add(struct virtio_scsi *vi)
865 {
866 	int ret;
867 
868 	ret = cpuhp_state_add_instance(virtioscsi_online, &vi->node);
869 	if (ret)
870 		return ret;
871 
872 	ret = cpuhp_state_add_instance(CPUHP_VIRT_SCSI_DEAD, &vi->node_dead);
873 	if (ret)
874 		cpuhp_state_remove_instance(virtioscsi_online, &vi->node);
875 	return ret;
876 }
877 
878 static void virtscsi_cpu_notif_remove(struct virtio_scsi *vi)
879 {
880 	cpuhp_state_remove_instance_nocalls(virtioscsi_online, &vi->node);
881 	cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_SCSI_DEAD,
882 					    &vi->node_dead);
883 }
884 
885 static void virtscsi_init_vq(struct virtio_scsi_vq *virtscsi_vq,
886 			     struct virtqueue *vq)
887 {
888 	spin_lock_init(&virtscsi_vq->vq_lock);
889 	virtscsi_vq->vq = vq;
890 }
891 
892 static void virtscsi_remove_vqs(struct virtio_device *vdev)
893 {
894 	struct Scsi_Host *sh = virtio_scsi_host(vdev);
895 	struct virtio_scsi *vscsi = shost_priv(sh);
896 
897 	virtscsi_set_affinity(vscsi, false);
898 
899 	/* Stop all the virtqueues. */
900 	vdev->config->reset(vdev);
901 
902 	vdev->config->del_vqs(vdev);
903 }
904 
905 static int virtscsi_init(struct virtio_device *vdev,
906 			 struct virtio_scsi *vscsi)
907 {
908 	int err;
909 	u32 i;
910 	u32 num_vqs;
911 	vq_callback_t **callbacks;
912 	const char **names;
913 	struct virtqueue **vqs;
914 
915 	num_vqs = vscsi->num_queues + VIRTIO_SCSI_VQ_BASE;
916 	vqs = kmalloc(num_vqs * sizeof(struct virtqueue *), GFP_KERNEL);
917 	callbacks = kmalloc(num_vqs * sizeof(vq_callback_t *), GFP_KERNEL);
918 	names = kmalloc(num_vqs * sizeof(char *), GFP_KERNEL);
919 
920 	if (!callbacks || !vqs || !names) {
921 		err = -ENOMEM;
922 		goto out;
923 	}
924 
925 	callbacks[0] = virtscsi_ctrl_done;
926 	callbacks[1] = virtscsi_event_done;
927 	names[0] = "control";
928 	names[1] = "event";
929 	for (i = VIRTIO_SCSI_VQ_BASE; i < num_vqs; i++) {
930 		callbacks[i] = virtscsi_req_done;
931 		names[i] = "request";
932 	}
933 
934 	/* Discover virtqueues and write information to configuration.  */
935 	err = vdev->config->find_vqs(vdev, num_vqs, vqs, callbacks, names);
936 	if (err)
937 		goto out;
938 
939 	virtscsi_init_vq(&vscsi->ctrl_vq, vqs[0]);
940 	virtscsi_init_vq(&vscsi->event_vq, vqs[1]);
941 	for (i = VIRTIO_SCSI_VQ_BASE; i < num_vqs; i++)
942 		virtscsi_init_vq(&vscsi->req_vqs[i - VIRTIO_SCSI_VQ_BASE],
943 				 vqs[i]);
944 
945 	virtscsi_config_set(vdev, cdb_size, VIRTIO_SCSI_CDB_SIZE);
946 	virtscsi_config_set(vdev, sense_size, VIRTIO_SCSI_SENSE_SIZE);
947 
948 	err = 0;
949 
950 out:
951 	kfree(names);
952 	kfree(callbacks);
953 	kfree(vqs);
954 	if (err)
955 		virtscsi_remove_vqs(vdev);
956 	return err;
957 }
958 
959 static int virtscsi_probe(struct virtio_device *vdev)
960 {
961 	struct Scsi_Host *shost;
962 	struct virtio_scsi *vscsi;
963 	int err;
964 	u32 sg_elems, num_targets;
965 	u32 cmd_per_lun;
966 	u32 num_queues;
967 	struct scsi_host_template *hostt;
968 
969 	if (!vdev->config->get) {
970 		dev_err(&vdev->dev, "%s failure: config access disabled\n",
971 			__func__);
972 		return -EINVAL;
973 	}
974 
975 	/* We need to know how many queues before we allocate. */
976 	num_queues = virtscsi_config_get(vdev, num_queues) ? : 1;
977 
978 	num_targets = virtscsi_config_get(vdev, max_target) + 1;
979 
980 	if (num_queues == 1)
981 		hostt = &virtscsi_host_template_single;
982 	else
983 		hostt = &virtscsi_host_template_multi;
984 
985 	shost = scsi_host_alloc(hostt,
986 		sizeof(*vscsi) + sizeof(vscsi->req_vqs[0]) * num_queues);
987 	if (!shost)
988 		return -ENOMEM;
989 
990 	sg_elems = virtscsi_config_get(vdev, seg_max) ?: 1;
991 	shost->sg_tablesize = sg_elems;
992 	vscsi = shost_priv(shost);
993 	vscsi->vdev = vdev;
994 	vscsi->num_queues = num_queues;
995 	vdev->priv = shost;
996 
997 	err = virtscsi_init(vdev, vscsi);
998 	if (err)
999 		goto virtscsi_init_failed;
1000 
1001 	err = virtscsi_cpu_notif_add(vscsi);
1002 	if (err)
1003 		goto scsi_add_host_failed;
1004 
1005 	cmd_per_lun = virtscsi_config_get(vdev, cmd_per_lun) ?: 1;
1006 	shost->cmd_per_lun = min_t(u32, cmd_per_lun, shost->can_queue);
1007 	shost->max_sectors = virtscsi_config_get(vdev, max_sectors) ?: 0xFFFF;
1008 
1009 	/* LUNs > 256 are reported with format 1, so they go in the range
1010 	 * 16640-32767.
1011 	 */
1012 	shost->max_lun = virtscsi_config_get(vdev, max_lun) + 1 + 0x4000;
1013 	shost->max_id = num_targets;
1014 	shost->max_channel = 0;
1015 	shost->max_cmd_len = VIRTIO_SCSI_CDB_SIZE;
1016 	shost->nr_hw_queues = num_queues;
1017 
1018 #ifdef CONFIG_BLK_DEV_INTEGRITY
1019 	if (virtio_has_feature(vdev, VIRTIO_SCSI_F_T10_PI)) {
1020 		int host_prot;
1021 
1022 		host_prot = SHOST_DIF_TYPE1_PROTECTION | SHOST_DIF_TYPE2_PROTECTION |
1023 			    SHOST_DIF_TYPE3_PROTECTION | SHOST_DIX_TYPE1_PROTECTION |
1024 			    SHOST_DIX_TYPE2_PROTECTION | SHOST_DIX_TYPE3_PROTECTION;
1025 
1026 		scsi_host_set_prot(shost, host_prot);
1027 		scsi_host_set_guard(shost, SHOST_DIX_GUARD_CRC);
1028 	}
1029 #endif
1030 
1031 	err = scsi_add_host(shost, &vdev->dev);
1032 	if (err)
1033 		goto scsi_add_host_failed;
1034 
1035 	virtio_device_ready(vdev);
1036 
1037 	if (virtio_has_feature(vdev, VIRTIO_SCSI_F_HOTPLUG))
1038 		virtscsi_kick_event_all(vscsi);
1039 
1040 	scsi_scan_host(shost);
1041 	return 0;
1042 
1043 scsi_add_host_failed:
1044 	vdev->config->del_vqs(vdev);
1045 virtscsi_init_failed:
1046 	scsi_host_put(shost);
1047 	return err;
1048 }
1049 
1050 static void virtscsi_remove(struct virtio_device *vdev)
1051 {
1052 	struct Scsi_Host *shost = virtio_scsi_host(vdev);
1053 	struct virtio_scsi *vscsi = shost_priv(shost);
1054 
1055 	if (virtio_has_feature(vdev, VIRTIO_SCSI_F_HOTPLUG))
1056 		virtscsi_cancel_event_work(vscsi);
1057 
1058 	scsi_remove_host(shost);
1059 
1060 	virtscsi_cpu_notif_remove(vscsi);
1061 
1062 	virtscsi_remove_vqs(vdev);
1063 	scsi_host_put(shost);
1064 }
1065 
1066 #ifdef CONFIG_PM_SLEEP
1067 static int virtscsi_freeze(struct virtio_device *vdev)
1068 {
1069 	struct Scsi_Host *sh = virtio_scsi_host(vdev);
1070 	struct virtio_scsi *vscsi = shost_priv(sh);
1071 
1072 	virtscsi_cpu_notif_remove(vscsi);
1073 	virtscsi_remove_vqs(vdev);
1074 	return 0;
1075 }
1076 
1077 static int virtscsi_restore(struct virtio_device *vdev)
1078 {
1079 	struct Scsi_Host *sh = virtio_scsi_host(vdev);
1080 	struct virtio_scsi *vscsi = shost_priv(sh);
1081 	int err;
1082 
1083 	err = virtscsi_init(vdev, vscsi);
1084 	if (err)
1085 		return err;
1086 
1087 	err = virtscsi_cpu_notif_add(vscsi);
1088 	if (err) {
1089 		vdev->config->del_vqs(vdev);
1090 		return err;
1091 	}
1092 	virtio_device_ready(vdev);
1093 
1094 	if (virtio_has_feature(vdev, VIRTIO_SCSI_F_HOTPLUG))
1095 		virtscsi_kick_event_all(vscsi);
1096 
1097 	return err;
1098 }
1099 #endif
1100 
1101 static struct virtio_device_id id_table[] = {
1102 	{ VIRTIO_ID_SCSI, VIRTIO_DEV_ANY_ID },
1103 	{ 0 },
1104 };
1105 
1106 static unsigned int features[] = {
1107 	VIRTIO_SCSI_F_HOTPLUG,
1108 	VIRTIO_SCSI_F_CHANGE,
1109 #ifdef CONFIG_BLK_DEV_INTEGRITY
1110 	VIRTIO_SCSI_F_T10_PI,
1111 #endif
1112 };
1113 
1114 static struct virtio_driver virtio_scsi_driver = {
1115 	.feature_table = features,
1116 	.feature_table_size = ARRAY_SIZE(features),
1117 	.driver.name = KBUILD_MODNAME,
1118 	.driver.owner = THIS_MODULE,
1119 	.id_table = id_table,
1120 	.probe = virtscsi_probe,
1121 #ifdef CONFIG_PM_SLEEP
1122 	.freeze = virtscsi_freeze,
1123 	.restore = virtscsi_restore,
1124 #endif
1125 	.remove = virtscsi_remove,
1126 };
1127 
1128 static int __init init(void)
1129 {
1130 	int ret = -ENOMEM;
1131 
1132 	virtscsi_cmd_cache = KMEM_CACHE(virtio_scsi_cmd, 0);
1133 	if (!virtscsi_cmd_cache) {
1134 		pr_err("kmem_cache_create() for virtscsi_cmd_cache failed\n");
1135 		goto error;
1136 	}
1137 
1138 
1139 	virtscsi_cmd_pool =
1140 		mempool_create_slab_pool(VIRTIO_SCSI_MEMPOOL_SZ,
1141 					 virtscsi_cmd_cache);
1142 	if (!virtscsi_cmd_pool) {
1143 		pr_err("mempool_create() for virtscsi_cmd_pool failed\n");
1144 		goto error;
1145 	}
1146 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
1147 				      "scsi/virtio:online",
1148 				      virtscsi_cpu_online, NULL);
1149 	if (ret < 0)
1150 		goto error;
1151 	virtioscsi_online = ret;
1152 	ret = cpuhp_setup_state_multi(CPUHP_VIRT_SCSI_DEAD, "scsi/virtio:dead",
1153 				      NULL, virtscsi_cpu_online);
1154 	if (ret)
1155 		goto error;
1156 	ret = register_virtio_driver(&virtio_scsi_driver);
1157 	if (ret < 0)
1158 		goto error;
1159 
1160 	return 0;
1161 
1162 error:
1163 	if (virtscsi_cmd_pool) {
1164 		mempool_destroy(virtscsi_cmd_pool);
1165 		virtscsi_cmd_pool = NULL;
1166 	}
1167 	if (virtscsi_cmd_cache) {
1168 		kmem_cache_destroy(virtscsi_cmd_cache);
1169 		virtscsi_cmd_cache = NULL;
1170 	}
1171 	if (virtioscsi_online)
1172 		cpuhp_remove_multi_state(virtioscsi_online);
1173 	cpuhp_remove_multi_state(CPUHP_VIRT_SCSI_DEAD);
1174 	return ret;
1175 }
1176 
1177 static void __exit fini(void)
1178 {
1179 	unregister_virtio_driver(&virtio_scsi_driver);
1180 	cpuhp_remove_multi_state(virtioscsi_online);
1181 	cpuhp_remove_multi_state(CPUHP_VIRT_SCSI_DEAD);
1182 	mempool_destroy(virtscsi_cmd_pool);
1183 	kmem_cache_destroy(virtscsi_cmd_cache);
1184 }
1185 module_init(init);
1186 module_exit(fini);
1187 
1188 MODULE_DEVICE_TABLE(virtio, id_table);
1189 MODULE_DESCRIPTION("Virtio SCSI HBA driver");
1190 MODULE_LICENSE("GPL");
1191