xref: /openbmc/linux/drivers/scsi/virtio_scsi.c (revision bc5aa3a0)
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 	/* CPU hotplug notifier */
111 	struct notifier_block nb;
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 struct kmem_cache *virtscsi_cmd_cache;
122 static mempool_t *virtscsi_cmd_pool;
123 
124 static inline struct Scsi_Host *virtio_scsi_host(struct virtio_device *vdev)
125 {
126 	return vdev->priv;
127 }
128 
129 static void virtscsi_compute_resid(struct scsi_cmnd *sc, u32 resid)
130 {
131 	if (!resid)
132 		return;
133 
134 	if (!scsi_bidi_cmnd(sc)) {
135 		scsi_set_resid(sc, resid);
136 		return;
137 	}
138 
139 	scsi_in(sc)->resid = min(resid, scsi_in(sc)->length);
140 	scsi_out(sc)->resid = resid - scsi_in(sc)->resid;
141 }
142 
143 /**
144  * virtscsi_complete_cmd - finish a scsi_cmd and invoke scsi_done
145  *
146  * Called with vq_lock held.
147  */
148 static void virtscsi_complete_cmd(struct virtio_scsi *vscsi, void *buf)
149 {
150 	struct virtio_scsi_cmd *cmd = buf;
151 	struct scsi_cmnd *sc = cmd->sc;
152 	struct virtio_scsi_cmd_resp *resp = &cmd->resp.cmd;
153 	struct virtio_scsi_target_state *tgt =
154 				scsi_target(sc->device)->hostdata;
155 
156 	dev_dbg(&sc->device->sdev_gendev,
157 		"cmd %p response %u status %#02x sense_len %u\n",
158 		sc, resp->response, resp->status, resp->sense_len);
159 
160 	sc->result = resp->status;
161 	virtscsi_compute_resid(sc, virtio32_to_cpu(vscsi->vdev, resp->resid));
162 	switch (resp->response) {
163 	case VIRTIO_SCSI_S_OK:
164 		set_host_byte(sc, DID_OK);
165 		break;
166 	case VIRTIO_SCSI_S_OVERRUN:
167 		set_host_byte(sc, DID_ERROR);
168 		break;
169 	case VIRTIO_SCSI_S_ABORTED:
170 		set_host_byte(sc, DID_ABORT);
171 		break;
172 	case VIRTIO_SCSI_S_BAD_TARGET:
173 		set_host_byte(sc, DID_BAD_TARGET);
174 		break;
175 	case VIRTIO_SCSI_S_RESET:
176 		set_host_byte(sc, DID_RESET);
177 		break;
178 	case VIRTIO_SCSI_S_BUSY:
179 		set_host_byte(sc, DID_BUS_BUSY);
180 		break;
181 	case VIRTIO_SCSI_S_TRANSPORT_FAILURE:
182 		set_host_byte(sc, DID_TRANSPORT_DISRUPTED);
183 		break;
184 	case VIRTIO_SCSI_S_TARGET_FAILURE:
185 		set_host_byte(sc, DID_TARGET_FAILURE);
186 		break;
187 	case VIRTIO_SCSI_S_NEXUS_FAILURE:
188 		set_host_byte(sc, DID_NEXUS_FAILURE);
189 		break;
190 	default:
191 		scmd_printk(KERN_WARNING, sc, "Unknown response %d",
192 			    resp->response);
193 		/* fall through */
194 	case VIRTIO_SCSI_S_FAILURE:
195 		set_host_byte(sc, DID_ERROR);
196 		break;
197 	}
198 
199 	WARN_ON(virtio32_to_cpu(vscsi->vdev, resp->sense_len) >
200 		VIRTIO_SCSI_SENSE_SIZE);
201 	if (sc->sense_buffer) {
202 		memcpy(sc->sense_buffer, resp->sense,
203 		       min_t(u32,
204 			     virtio32_to_cpu(vscsi->vdev, resp->sense_len),
205 			     VIRTIO_SCSI_SENSE_SIZE));
206 		if (resp->sense_len)
207 			set_driver_byte(sc, DRIVER_SENSE);
208 	}
209 
210 	sc->scsi_done(sc);
211 
212 	atomic_dec(&tgt->reqs);
213 }
214 
215 static void virtscsi_vq_done(struct virtio_scsi *vscsi,
216 			     struct virtio_scsi_vq *virtscsi_vq,
217 			     void (*fn)(struct virtio_scsi *vscsi, void *buf))
218 {
219 	void *buf;
220 	unsigned int len;
221 	unsigned long flags;
222 	struct virtqueue *vq = virtscsi_vq->vq;
223 
224 	spin_lock_irqsave(&virtscsi_vq->vq_lock, flags);
225 	do {
226 		virtqueue_disable_cb(vq);
227 		while ((buf = virtqueue_get_buf(vq, &len)) != NULL)
228 			fn(vscsi, buf);
229 
230 		if (unlikely(virtqueue_is_broken(vq)))
231 			break;
232 	} while (!virtqueue_enable_cb(vq));
233 	spin_unlock_irqrestore(&virtscsi_vq->vq_lock, flags);
234 }
235 
236 static void virtscsi_req_done(struct virtqueue *vq)
237 {
238 	struct Scsi_Host *sh = virtio_scsi_host(vq->vdev);
239 	struct virtio_scsi *vscsi = shost_priv(sh);
240 	int index = vq->index - VIRTIO_SCSI_VQ_BASE;
241 	struct virtio_scsi_vq *req_vq = &vscsi->req_vqs[index];
242 
243 	virtscsi_vq_done(vscsi, req_vq, virtscsi_complete_cmd);
244 };
245 
246 static void virtscsi_poll_requests(struct virtio_scsi *vscsi)
247 {
248 	int i, num_vqs;
249 
250 	num_vqs = vscsi->num_queues;
251 	for (i = 0; i < num_vqs; i++)
252 		virtscsi_vq_done(vscsi, &vscsi->req_vqs[i],
253 				 virtscsi_complete_cmd);
254 }
255 
256 static void virtscsi_complete_free(struct virtio_scsi *vscsi, void *buf)
257 {
258 	struct virtio_scsi_cmd *cmd = buf;
259 
260 	if (cmd->comp)
261 		complete_all(cmd->comp);
262 }
263 
264 static void virtscsi_ctrl_done(struct virtqueue *vq)
265 {
266 	struct Scsi_Host *sh = virtio_scsi_host(vq->vdev);
267 	struct virtio_scsi *vscsi = shost_priv(sh);
268 
269 	virtscsi_vq_done(vscsi, &vscsi->ctrl_vq, virtscsi_complete_free);
270 };
271 
272 static void virtscsi_handle_event(struct work_struct *work);
273 
274 static int virtscsi_kick_event(struct virtio_scsi *vscsi,
275 			       struct virtio_scsi_event_node *event_node)
276 {
277 	int err;
278 	struct scatterlist sg;
279 	unsigned long flags;
280 
281 	INIT_WORK(&event_node->work, virtscsi_handle_event);
282 	sg_init_one(&sg, &event_node->event, sizeof(struct virtio_scsi_event));
283 
284 	spin_lock_irqsave(&vscsi->event_vq.vq_lock, flags);
285 
286 	err = virtqueue_add_inbuf(vscsi->event_vq.vq, &sg, 1, event_node,
287 				  GFP_ATOMIC);
288 	if (!err)
289 		virtqueue_kick(vscsi->event_vq.vq);
290 
291 	spin_unlock_irqrestore(&vscsi->event_vq.vq_lock, flags);
292 
293 	return err;
294 }
295 
296 static int virtscsi_kick_event_all(struct virtio_scsi *vscsi)
297 {
298 	int i;
299 
300 	for (i = 0; i < VIRTIO_SCSI_EVENT_LEN; i++) {
301 		vscsi->event_list[i].vscsi = vscsi;
302 		virtscsi_kick_event(vscsi, &vscsi->event_list[i]);
303 	}
304 
305 	return 0;
306 }
307 
308 static void virtscsi_cancel_event_work(struct virtio_scsi *vscsi)
309 {
310 	int i;
311 
312 	/* Stop scheduling work before calling cancel_work_sync.  */
313 	spin_lock_irq(&vscsi->event_vq.vq_lock);
314 	vscsi->stop_events = true;
315 	spin_unlock_irq(&vscsi->event_vq.vq_lock);
316 
317 	for (i = 0; i < VIRTIO_SCSI_EVENT_LEN; i++)
318 		cancel_work_sync(&vscsi->event_list[i].work);
319 }
320 
321 static void virtscsi_handle_transport_reset(struct virtio_scsi *vscsi,
322 					    struct virtio_scsi_event *event)
323 {
324 	struct scsi_device *sdev;
325 	struct Scsi_Host *shost = virtio_scsi_host(vscsi->vdev);
326 	unsigned int target = event->lun[1];
327 	unsigned int lun = (event->lun[2] << 8) | event->lun[3];
328 
329 	switch (virtio32_to_cpu(vscsi->vdev, event->reason)) {
330 	case VIRTIO_SCSI_EVT_RESET_RESCAN:
331 		scsi_add_device(shost, 0, target, lun);
332 		break;
333 	case VIRTIO_SCSI_EVT_RESET_REMOVED:
334 		sdev = scsi_device_lookup(shost, 0, target, lun);
335 		if (sdev) {
336 			scsi_remove_device(sdev);
337 			scsi_device_put(sdev);
338 		} else {
339 			pr_err("SCSI device %d 0 %d %d not found\n",
340 				shost->host_no, target, lun);
341 		}
342 		break;
343 	default:
344 		pr_info("Unsupport virtio scsi event reason %x\n", event->reason);
345 	}
346 }
347 
348 static void virtscsi_handle_param_change(struct virtio_scsi *vscsi,
349 					 struct virtio_scsi_event *event)
350 {
351 	struct scsi_device *sdev;
352 	struct Scsi_Host *shost = virtio_scsi_host(vscsi->vdev);
353 	unsigned int target = event->lun[1];
354 	unsigned int lun = (event->lun[2] << 8) | event->lun[3];
355 	u8 asc = virtio32_to_cpu(vscsi->vdev, event->reason) & 255;
356 	u8 ascq = virtio32_to_cpu(vscsi->vdev, event->reason) >> 8;
357 
358 	sdev = scsi_device_lookup(shost, 0, target, lun);
359 	if (!sdev) {
360 		pr_err("SCSI device %d 0 %d %d not found\n",
361 			shost->host_no, target, lun);
362 		return;
363 	}
364 
365 	/* Handle "Parameters changed", "Mode parameters changed", and
366 	   "Capacity data has changed".  */
367 	if (asc == 0x2a && (ascq == 0x00 || ascq == 0x01 || ascq == 0x09))
368 		scsi_rescan_device(&sdev->sdev_gendev);
369 
370 	scsi_device_put(sdev);
371 }
372 
373 static void virtscsi_handle_event(struct work_struct *work)
374 {
375 	struct virtio_scsi_event_node *event_node =
376 		container_of(work, struct virtio_scsi_event_node, work);
377 	struct virtio_scsi *vscsi = event_node->vscsi;
378 	struct virtio_scsi_event *event = &event_node->event;
379 
380 	if (event->event &
381 	    cpu_to_virtio32(vscsi->vdev, VIRTIO_SCSI_T_EVENTS_MISSED)) {
382 		event->event &= ~cpu_to_virtio32(vscsi->vdev,
383 						   VIRTIO_SCSI_T_EVENTS_MISSED);
384 		scsi_scan_host(virtio_scsi_host(vscsi->vdev));
385 	}
386 
387 	switch (virtio32_to_cpu(vscsi->vdev, event->event)) {
388 	case VIRTIO_SCSI_T_NO_EVENT:
389 		break;
390 	case VIRTIO_SCSI_T_TRANSPORT_RESET:
391 		virtscsi_handle_transport_reset(vscsi, event);
392 		break;
393 	case VIRTIO_SCSI_T_PARAM_CHANGE:
394 		virtscsi_handle_param_change(vscsi, event);
395 		break;
396 	default:
397 		pr_err("Unsupport virtio scsi event %x\n", event->event);
398 	}
399 	virtscsi_kick_event(vscsi, event_node);
400 }
401 
402 static void virtscsi_complete_event(struct virtio_scsi *vscsi, void *buf)
403 {
404 	struct virtio_scsi_event_node *event_node = buf;
405 
406 	if (!vscsi->stop_events)
407 		queue_work(system_freezable_wq, &event_node->work);
408 }
409 
410 static void virtscsi_event_done(struct virtqueue *vq)
411 {
412 	struct Scsi_Host *sh = virtio_scsi_host(vq->vdev);
413 	struct virtio_scsi *vscsi = shost_priv(sh);
414 
415 	virtscsi_vq_done(vscsi, &vscsi->event_vq, virtscsi_complete_event);
416 };
417 
418 /**
419  * virtscsi_add_cmd - add a virtio_scsi_cmd to a virtqueue
420  * @vq		: the struct virtqueue we're talking about
421  * @cmd		: command structure
422  * @req_size	: size of the request buffer
423  * @resp_size	: size of the response buffer
424  */
425 static int virtscsi_add_cmd(struct virtqueue *vq,
426 			    struct virtio_scsi_cmd *cmd,
427 			    size_t req_size, size_t resp_size)
428 {
429 	struct scsi_cmnd *sc = cmd->sc;
430 	struct scatterlist *sgs[6], req, resp;
431 	struct sg_table *out, *in;
432 	unsigned out_num = 0, in_num = 0;
433 
434 	out = in = NULL;
435 
436 	if (sc && sc->sc_data_direction != DMA_NONE) {
437 		if (sc->sc_data_direction != DMA_FROM_DEVICE)
438 			out = &scsi_out(sc)->table;
439 		if (sc->sc_data_direction != DMA_TO_DEVICE)
440 			in = &scsi_in(sc)->table;
441 	}
442 
443 	/* Request header.  */
444 	sg_init_one(&req, &cmd->req, req_size);
445 	sgs[out_num++] = &req;
446 
447 	/* Data-out buffer.  */
448 	if (out) {
449 		/* Place WRITE protection SGLs before Data OUT payload */
450 		if (scsi_prot_sg_count(sc))
451 			sgs[out_num++] = scsi_prot_sglist(sc);
452 		sgs[out_num++] = out->sgl;
453 	}
454 
455 	/* Response header.  */
456 	sg_init_one(&resp, &cmd->resp, resp_size);
457 	sgs[out_num + in_num++] = &resp;
458 
459 	/* Data-in buffer */
460 	if (in) {
461 		/* Place READ protection SGLs before Data IN payload */
462 		if (scsi_prot_sg_count(sc))
463 			sgs[out_num + in_num++] = scsi_prot_sglist(sc);
464 		sgs[out_num + in_num++] = in->sgl;
465 	}
466 
467 	return virtqueue_add_sgs(vq, sgs, out_num, in_num, cmd, GFP_ATOMIC);
468 }
469 
470 static int virtscsi_kick_cmd(struct virtio_scsi_vq *vq,
471 			     struct virtio_scsi_cmd *cmd,
472 			     size_t req_size, size_t resp_size)
473 {
474 	unsigned long flags;
475 	int err;
476 	bool needs_kick = false;
477 
478 	spin_lock_irqsave(&vq->vq_lock, flags);
479 	err = virtscsi_add_cmd(vq->vq, cmd, req_size, resp_size);
480 	if (!err)
481 		needs_kick = virtqueue_kick_prepare(vq->vq);
482 
483 	spin_unlock_irqrestore(&vq->vq_lock, flags);
484 
485 	if (needs_kick)
486 		virtqueue_notify(vq->vq);
487 	return err;
488 }
489 
490 static void virtio_scsi_init_hdr(struct virtio_device *vdev,
491 				 struct virtio_scsi_cmd_req *cmd,
492 				 struct scsi_cmnd *sc)
493 {
494 	cmd->lun[0] = 1;
495 	cmd->lun[1] = sc->device->id;
496 	cmd->lun[2] = (sc->device->lun >> 8) | 0x40;
497 	cmd->lun[3] = sc->device->lun & 0xff;
498 	cmd->tag = cpu_to_virtio64(vdev, (unsigned long)sc);
499 	cmd->task_attr = VIRTIO_SCSI_S_SIMPLE;
500 	cmd->prio = 0;
501 	cmd->crn = 0;
502 }
503 
504 #ifdef CONFIG_BLK_DEV_INTEGRITY
505 static void virtio_scsi_init_hdr_pi(struct virtio_device *vdev,
506 				    struct virtio_scsi_cmd_req_pi *cmd_pi,
507 				    struct scsi_cmnd *sc)
508 {
509 	struct request *rq = sc->request;
510 	struct blk_integrity *bi;
511 
512 	virtio_scsi_init_hdr(vdev, (struct virtio_scsi_cmd_req *)cmd_pi, sc);
513 
514 	if (!rq || !scsi_prot_sg_count(sc))
515 		return;
516 
517 	bi = blk_get_integrity(rq->rq_disk);
518 
519 	if (sc->sc_data_direction == DMA_TO_DEVICE)
520 		cmd_pi->pi_bytesout = cpu_to_virtio32(vdev,
521 							blk_rq_sectors(rq) *
522 							bi->tuple_size);
523 	else if (sc->sc_data_direction == DMA_FROM_DEVICE)
524 		cmd_pi->pi_bytesin = cpu_to_virtio32(vdev,
525 						       blk_rq_sectors(rq) *
526 						       bi->tuple_size);
527 }
528 #endif
529 
530 static int virtscsi_queuecommand(struct virtio_scsi *vscsi,
531 				 struct virtio_scsi_vq *req_vq,
532 				 struct scsi_cmnd *sc)
533 {
534 	struct Scsi_Host *shost = virtio_scsi_host(vscsi->vdev);
535 	struct virtio_scsi_cmd *cmd = scsi_cmd_priv(sc);
536 	int req_size;
537 
538 	BUG_ON(scsi_sg_count(sc) > shost->sg_tablesize);
539 
540 	/* TODO: check feature bit and fail if unsupported?  */
541 	BUG_ON(sc->sc_data_direction == DMA_BIDIRECTIONAL);
542 
543 	dev_dbg(&sc->device->sdev_gendev,
544 		"cmd %p CDB: %#02x\n", sc, sc->cmnd[0]);
545 
546 	memset(cmd, 0, sizeof(*cmd));
547 	cmd->sc = sc;
548 
549 	BUG_ON(sc->cmd_len > VIRTIO_SCSI_CDB_SIZE);
550 
551 #ifdef CONFIG_BLK_DEV_INTEGRITY
552 	if (virtio_has_feature(vscsi->vdev, VIRTIO_SCSI_F_T10_PI)) {
553 		virtio_scsi_init_hdr_pi(vscsi->vdev, &cmd->req.cmd_pi, sc);
554 		memcpy(cmd->req.cmd_pi.cdb, sc->cmnd, sc->cmd_len);
555 		req_size = sizeof(cmd->req.cmd_pi);
556 	} else
557 #endif
558 	{
559 		virtio_scsi_init_hdr(vscsi->vdev, &cmd->req.cmd, sc);
560 		memcpy(cmd->req.cmd.cdb, sc->cmnd, sc->cmd_len);
561 		req_size = sizeof(cmd->req.cmd);
562 	}
563 
564 	if (virtscsi_kick_cmd(req_vq, cmd, req_size, sizeof(cmd->resp.cmd)) != 0)
565 		return SCSI_MLQUEUE_HOST_BUSY;
566 	return 0;
567 }
568 
569 static int virtscsi_queuecommand_single(struct Scsi_Host *sh,
570 					struct scsi_cmnd *sc)
571 {
572 	struct virtio_scsi *vscsi = shost_priv(sh);
573 	struct virtio_scsi_target_state *tgt =
574 				scsi_target(sc->device)->hostdata;
575 
576 	atomic_inc(&tgt->reqs);
577 	return virtscsi_queuecommand(vscsi, &vscsi->req_vqs[0], sc);
578 }
579 
580 static struct virtio_scsi_vq *virtscsi_pick_vq_mq(struct virtio_scsi *vscsi,
581 						  struct scsi_cmnd *sc)
582 {
583 	u32 tag = blk_mq_unique_tag(sc->request);
584 	u16 hwq = blk_mq_unique_tag_to_hwq(tag);
585 
586 	return &vscsi->req_vqs[hwq];
587 }
588 
589 static struct virtio_scsi_vq *virtscsi_pick_vq(struct virtio_scsi *vscsi,
590 					       struct virtio_scsi_target_state *tgt)
591 {
592 	struct virtio_scsi_vq *vq;
593 	unsigned long flags;
594 	u32 queue_num;
595 
596 	local_irq_save(flags);
597 	if (atomic_inc_return(&tgt->reqs) > 1) {
598 		unsigned long seq;
599 
600 		do {
601 			seq = read_seqcount_begin(&tgt->tgt_seq);
602 			vq = tgt->req_vq;
603 		} while (read_seqcount_retry(&tgt->tgt_seq, seq));
604 	} else {
605 		/* no writes can be concurrent because of atomic_t */
606 		write_seqcount_begin(&tgt->tgt_seq);
607 
608 		/* keep previous req_vq if a reader just arrived */
609 		if (unlikely(atomic_read(&tgt->reqs) > 1)) {
610 			vq = tgt->req_vq;
611 			goto unlock;
612 		}
613 
614 		queue_num = smp_processor_id();
615 		while (unlikely(queue_num >= vscsi->num_queues))
616 			queue_num -= vscsi->num_queues;
617 		tgt->req_vq = vq = &vscsi->req_vqs[queue_num];
618  unlock:
619 		write_seqcount_end(&tgt->tgt_seq);
620 	}
621 	local_irq_restore(flags);
622 
623 	return vq;
624 }
625 
626 static int virtscsi_queuecommand_multi(struct Scsi_Host *sh,
627 				       struct scsi_cmnd *sc)
628 {
629 	struct virtio_scsi *vscsi = shost_priv(sh);
630 	struct virtio_scsi_target_state *tgt =
631 				scsi_target(sc->device)->hostdata;
632 	struct virtio_scsi_vq *req_vq;
633 
634 	if (shost_use_blk_mq(sh))
635 		req_vq = virtscsi_pick_vq_mq(vscsi, sc);
636 	else
637 		req_vq = virtscsi_pick_vq(vscsi, tgt);
638 
639 	return virtscsi_queuecommand(vscsi, req_vq, sc);
640 }
641 
642 static int virtscsi_tmf(struct virtio_scsi *vscsi, struct virtio_scsi_cmd *cmd)
643 {
644 	DECLARE_COMPLETION_ONSTACK(comp);
645 	int ret = FAILED;
646 
647 	cmd->comp = &comp;
648 	if (virtscsi_kick_cmd(&vscsi->ctrl_vq, cmd,
649 			      sizeof cmd->req.tmf, sizeof cmd->resp.tmf) < 0)
650 		goto out;
651 
652 	wait_for_completion(&comp);
653 	if (cmd->resp.tmf.response == VIRTIO_SCSI_S_OK ||
654 	    cmd->resp.tmf.response == VIRTIO_SCSI_S_FUNCTION_SUCCEEDED)
655 		ret = SUCCESS;
656 
657 	/*
658 	 * The spec guarantees that all requests related to the TMF have
659 	 * been completed, but the callback might not have run yet if
660 	 * we're using independent interrupts (e.g. MSI).  Poll the
661 	 * virtqueues once.
662 	 *
663 	 * In the abort case, sc->scsi_done will do nothing, because
664 	 * the block layer must have detected a timeout and as a result
665 	 * REQ_ATOM_COMPLETE has been set.
666 	 */
667 	virtscsi_poll_requests(vscsi);
668 
669 out:
670 	mempool_free(cmd, virtscsi_cmd_pool);
671 	return ret;
672 }
673 
674 static int virtscsi_device_reset(struct scsi_cmnd *sc)
675 {
676 	struct virtio_scsi *vscsi = shost_priv(sc->device->host);
677 	struct virtio_scsi_cmd *cmd;
678 
679 	sdev_printk(KERN_INFO, sc->device, "device reset\n");
680 	cmd = mempool_alloc(virtscsi_cmd_pool, GFP_NOIO);
681 	if (!cmd)
682 		return FAILED;
683 
684 	memset(cmd, 0, sizeof(*cmd));
685 	cmd->sc = sc;
686 	cmd->req.tmf = (struct virtio_scsi_ctrl_tmf_req){
687 		.type = VIRTIO_SCSI_T_TMF,
688 		.subtype = cpu_to_virtio32(vscsi->vdev,
689 					     VIRTIO_SCSI_T_TMF_LOGICAL_UNIT_RESET),
690 		.lun[0] = 1,
691 		.lun[1] = sc->device->id,
692 		.lun[2] = (sc->device->lun >> 8) | 0x40,
693 		.lun[3] = sc->device->lun & 0xff,
694 	};
695 	return virtscsi_tmf(vscsi, cmd);
696 }
697 
698 /**
699  * virtscsi_change_queue_depth() - Change a virtscsi target's queue depth
700  * @sdev:	Virtscsi target whose queue depth to change
701  * @qdepth:	New queue depth
702  */
703 static int virtscsi_change_queue_depth(struct scsi_device *sdev, int qdepth)
704 {
705 	struct Scsi_Host *shost = sdev->host;
706 	int max_depth = shost->cmd_per_lun;
707 
708 	return scsi_change_queue_depth(sdev, min(max_depth, qdepth));
709 }
710 
711 static int virtscsi_abort(struct scsi_cmnd *sc)
712 {
713 	struct virtio_scsi *vscsi = shost_priv(sc->device->host);
714 	struct virtio_scsi_cmd *cmd;
715 
716 	scmd_printk(KERN_INFO, sc, "abort\n");
717 	cmd = mempool_alloc(virtscsi_cmd_pool, GFP_NOIO);
718 	if (!cmd)
719 		return FAILED;
720 
721 	memset(cmd, 0, sizeof(*cmd));
722 	cmd->sc = sc;
723 	cmd->req.tmf = (struct virtio_scsi_ctrl_tmf_req){
724 		.type = VIRTIO_SCSI_T_TMF,
725 		.subtype = VIRTIO_SCSI_T_TMF_ABORT_TASK,
726 		.lun[0] = 1,
727 		.lun[1] = sc->device->id,
728 		.lun[2] = (sc->device->lun >> 8) | 0x40,
729 		.lun[3] = sc->device->lun & 0xff,
730 		.tag = cpu_to_virtio64(vscsi->vdev, (unsigned long)sc),
731 	};
732 	return virtscsi_tmf(vscsi, cmd);
733 }
734 
735 static int virtscsi_target_alloc(struct scsi_target *starget)
736 {
737 	struct Scsi_Host *sh = dev_to_shost(starget->dev.parent);
738 	struct virtio_scsi *vscsi = shost_priv(sh);
739 
740 	struct virtio_scsi_target_state *tgt =
741 				kmalloc(sizeof(*tgt), GFP_KERNEL);
742 	if (!tgt)
743 		return -ENOMEM;
744 
745 	seqcount_init(&tgt->tgt_seq);
746 	atomic_set(&tgt->reqs, 0);
747 	tgt->req_vq = &vscsi->req_vqs[0];
748 
749 	starget->hostdata = tgt;
750 	return 0;
751 }
752 
753 static void virtscsi_target_destroy(struct scsi_target *starget)
754 {
755 	struct virtio_scsi_target_state *tgt = starget->hostdata;
756 	kfree(tgt);
757 }
758 
759 static struct scsi_host_template virtscsi_host_template_single = {
760 	.module = THIS_MODULE,
761 	.name = "Virtio SCSI HBA",
762 	.proc_name = "virtio_scsi",
763 	.this_id = -1,
764 	.cmd_size = sizeof(struct virtio_scsi_cmd),
765 	.queuecommand = virtscsi_queuecommand_single,
766 	.change_queue_depth = virtscsi_change_queue_depth,
767 	.eh_abort_handler = virtscsi_abort,
768 	.eh_device_reset_handler = virtscsi_device_reset,
769 
770 	.can_queue = 1024,
771 	.dma_boundary = UINT_MAX,
772 	.use_clustering = ENABLE_CLUSTERING,
773 	.target_alloc = virtscsi_target_alloc,
774 	.target_destroy = virtscsi_target_destroy,
775 	.track_queue_depth = 1,
776 };
777 
778 static struct scsi_host_template virtscsi_host_template_multi = {
779 	.module = THIS_MODULE,
780 	.name = "Virtio SCSI HBA",
781 	.proc_name = "virtio_scsi",
782 	.this_id = -1,
783 	.cmd_size = sizeof(struct virtio_scsi_cmd),
784 	.queuecommand = virtscsi_queuecommand_multi,
785 	.change_queue_depth = virtscsi_change_queue_depth,
786 	.eh_abort_handler = virtscsi_abort,
787 	.eh_device_reset_handler = virtscsi_device_reset,
788 
789 	.can_queue = 1024,
790 	.dma_boundary = UINT_MAX,
791 	.use_clustering = ENABLE_CLUSTERING,
792 	.target_alloc = virtscsi_target_alloc,
793 	.target_destroy = virtscsi_target_destroy,
794 	.track_queue_depth = 1,
795 };
796 
797 #define virtscsi_config_get(vdev, fld) \
798 	({ \
799 		typeof(((struct virtio_scsi_config *)0)->fld) __val; \
800 		virtio_cread(vdev, struct virtio_scsi_config, fld, &__val); \
801 		__val; \
802 	})
803 
804 #define virtscsi_config_set(vdev, fld, val) \
805 	do { \
806 		typeof(((struct virtio_scsi_config *)0)->fld) __val = (val); \
807 		virtio_cwrite(vdev, struct virtio_scsi_config, fld, &__val); \
808 	} while(0)
809 
810 static void __virtscsi_set_affinity(struct virtio_scsi *vscsi, bool affinity)
811 {
812 	int i;
813 	int cpu;
814 
815 	/* In multiqueue mode, when the number of cpu is equal
816 	 * to the number of request queues, we let the qeueues
817 	 * to be private to one cpu by setting the affinity hint
818 	 * to eliminate the contention.
819 	 */
820 	if ((vscsi->num_queues == 1 ||
821 	     vscsi->num_queues != num_online_cpus()) && affinity) {
822 		if (vscsi->affinity_hint_set)
823 			affinity = false;
824 		else
825 			return;
826 	}
827 
828 	if (affinity) {
829 		i = 0;
830 		for_each_online_cpu(cpu) {
831 			virtqueue_set_affinity(vscsi->req_vqs[i].vq, cpu);
832 			i++;
833 		}
834 
835 		vscsi->affinity_hint_set = true;
836 	} else {
837 		for (i = 0; i < vscsi->num_queues; i++) {
838 			if (!vscsi->req_vqs[i].vq)
839 				continue;
840 
841 			virtqueue_set_affinity(vscsi->req_vqs[i].vq, -1);
842 		}
843 
844 		vscsi->affinity_hint_set = false;
845 	}
846 }
847 
848 static void virtscsi_set_affinity(struct virtio_scsi *vscsi, bool affinity)
849 {
850 	get_online_cpus();
851 	__virtscsi_set_affinity(vscsi, affinity);
852 	put_online_cpus();
853 }
854 
855 static int virtscsi_cpu_callback(struct notifier_block *nfb,
856 				 unsigned long action, void *hcpu)
857 {
858 	struct virtio_scsi *vscsi = container_of(nfb, struct virtio_scsi, nb);
859 	switch(action) {
860 	case CPU_ONLINE:
861 	case CPU_ONLINE_FROZEN:
862 	case CPU_DEAD:
863 	case CPU_DEAD_FROZEN:
864 		__virtscsi_set_affinity(vscsi, true);
865 		break;
866 	default:
867 		break;
868 	}
869 	return NOTIFY_OK;
870 }
871 
872 static void virtscsi_init_vq(struct virtio_scsi_vq *virtscsi_vq,
873 			     struct virtqueue *vq)
874 {
875 	spin_lock_init(&virtscsi_vq->vq_lock);
876 	virtscsi_vq->vq = vq;
877 }
878 
879 static void virtscsi_remove_vqs(struct virtio_device *vdev)
880 {
881 	struct Scsi_Host *sh = virtio_scsi_host(vdev);
882 	struct virtio_scsi *vscsi = shost_priv(sh);
883 
884 	virtscsi_set_affinity(vscsi, false);
885 
886 	/* Stop all the virtqueues. */
887 	vdev->config->reset(vdev);
888 
889 	vdev->config->del_vqs(vdev);
890 }
891 
892 static int virtscsi_init(struct virtio_device *vdev,
893 			 struct virtio_scsi *vscsi)
894 {
895 	int err;
896 	u32 i;
897 	u32 num_vqs;
898 	vq_callback_t **callbacks;
899 	const char **names;
900 	struct virtqueue **vqs;
901 
902 	num_vqs = vscsi->num_queues + VIRTIO_SCSI_VQ_BASE;
903 	vqs = kmalloc(num_vqs * sizeof(struct virtqueue *), GFP_KERNEL);
904 	callbacks = kmalloc(num_vqs * sizeof(vq_callback_t *), GFP_KERNEL);
905 	names = kmalloc(num_vqs * sizeof(char *), GFP_KERNEL);
906 
907 	if (!callbacks || !vqs || !names) {
908 		err = -ENOMEM;
909 		goto out;
910 	}
911 
912 	callbacks[0] = virtscsi_ctrl_done;
913 	callbacks[1] = virtscsi_event_done;
914 	names[0] = "control";
915 	names[1] = "event";
916 	for (i = VIRTIO_SCSI_VQ_BASE; i < num_vqs; i++) {
917 		callbacks[i] = virtscsi_req_done;
918 		names[i] = "request";
919 	}
920 
921 	/* Discover virtqueues and write information to configuration.  */
922 	err = vdev->config->find_vqs(vdev, num_vqs, vqs, callbacks, names);
923 	if (err)
924 		goto out;
925 
926 	virtscsi_init_vq(&vscsi->ctrl_vq, vqs[0]);
927 	virtscsi_init_vq(&vscsi->event_vq, vqs[1]);
928 	for (i = VIRTIO_SCSI_VQ_BASE; i < num_vqs; i++)
929 		virtscsi_init_vq(&vscsi->req_vqs[i - VIRTIO_SCSI_VQ_BASE],
930 				 vqs[i]);
931 
932 	virtscsi_set_affinity(vscsi, true);
933 
934 	virtscsi_config_set(vdev, cdb_size, VIRTIO_SCSI_CDB_SIZE);
935 	virtscsi_config_set(vdev, sense_size, VIRTIO_SCSI_SENSE_SIZE);
936 
937 	err = 0;
938 
939 out:
940 	kfree(names);
941 	kfree(callbacks);
942 	kfree(vqs);
943 	if (err)
944 		virtscsi_remove_vqs(vdev);
945 	return err;
946 }
947 
948 static int virtscsi_probe(struct virtio_device *vdev)
949 {
950 	struct Scsi_Host *shost;
951 	struct virtio_scsi *vscsi;
952 	int err;
953 	u32 sg_elems, num_targets;
954 	u32 cmd_per_lun;
955 	u32 num_queues;
956 	struct scsi_host_template *hostt;
957 
958 	if (!vdev->config->get) {
959 		dev_err(&vdev->dev, "%s failure: config access disabled\n",
960 			__func__);
961 		return -EINVAL;
962 	}
963 
964 	/* We need to know how many queues before we allocate. */
965 	num_queues = virtscsi_config_get(vdev, num_queues) ? : 1;
966 
967 	num_targets = virtscsi_config_get(vdev, max_target) + 1;
968 
969 	if (num_queues == 1)
970 		hostt = &virtscsi_host_template_single;
971 	else
972 		hostt = &virtscsi_host_template_multi;
973 
974 	shost = scsi_host_alloc(hostt,
975 		sizeof(*vscsi) + sizeof(vscsi->req_vqs[0]) * num_queues);
976 	if (!shost)
977 		return -ENOMEM;
978 
979 	sg_elems = virtscsi_config_get(vdev, seg_max) ?: 1;
980 	shost->sg_tablesize = sg_elems;
981 	vscsi = shost_priv(shost);
982 	vscsi->vdev = vdev;
983 	vscsi->num_queues = num_queues;
984 	vdev->priv = shost;
985 
986 	err = virtscsi_init(vdev, vscsi);
987 	if (err)
988 		goto virtscsi_init_failed;
989 
990 	vscsi->nb.notifier_call = &virtscsi_cpu_callback;
991 	err = register_hotcpu_notifier(&vscsi->nb);
992 	if (err) {
993 		pr_err("registering cpu notifier failed\n");
994 		goto scsi_add_host_failed;
995 	}
996 
997 	cmd_per_lun = virtscsi_config_get(vdev, cmd_per_lun) ?: 1;
998 	shost->cmd_per_lun = min_t(u32, cmd_per_lun, shost->can_queue);
999 	shost->max_sectors = virtscsi_config_get(vdev, max_sectors) ?: 0xFFFF;
1000 
1001 	/* LUNs > 256 are reported with format 1, so they go in the range
1002 	 * 16640-32767.
1003 	 */
1004 	shost->max_lun = virtscsi_config_get(vdev, max_lun) + 1 + 0x4000;
1005 	shost->max_id = num_targets;
1006 	shost->max_channel = 0;
1007 	shost->max_cmd_len = VIRTIO_SCSI_CDB_SIZE;
1008 	shost->nr_hw_queues = num_queues;
1009 
1010 #ifdef CONFIG_BLK_DEV_INTEGRITY
1011 	if (virtio_has_feature(vdev, VIRTIO_SCSI_F_T10_PI)) {
1012 		int host_prot;
1013 
1014 		host_prot = SHOST_DIF_TYPE1_PROTECTION | SHOST_DIF_TYPE2_PROTECTION |
1015 			    SHOST_DIF_TYPE3_PROTECTION | SHOST_DIX_TYPE1_PROTECTION |
1016 			    SHOST_DIX_TYPE2_PROTECTION | SHOST_DIX_TYPE3_PROTECTION;
1017 
1018 		scsi_host_set_prot(shost, host_prot);
1019 		scsi_host_set_guard(shost, SHOST_DIX_GUARD_CRC);
1020 	}
1021 #endif
1022 
1023 	err = scsi_add_host(shost, &vdev->dev);
1024 	if (err)
1025 		goto scsi_add_host_failed;
1026 
1027 	virtio_device_ready(vdev);
1028 
1029 	if (virtio_has_feature(vdev, VIRTIO_SCSI_F_HOTPLUG))
1030 		virtscsi_kick_event_all(vscsi);
1031 
1032 	scsi_scan_host(shost);
1033 	return 0;
1034 
1035 scsi_add_host_failed:
1036 	vdev->config->del_vqs(vdev);
1037 virtscsi_init_failed:
1038 	scsi_host_put(shost);
1039 	return err;
1040 }
1041 
1042 static void virtscsi_remove(struct virtio_device *vdev)
1043 {
1044 	struct Scsi_Host *shost = virtio_scsi_host(vdev);
1045 	struct virtio_scsi *vscsi = shost_priv(shost);
1046 
1047 	if (virtio_has_feature(vdev, VIRTIO_SCSI_F_HOTPLUG))
1048 		virtscsi_cancel_event_work(vscsi);
1049 
1050 	scsi_remove_host(shost);
1051 
1052 	unregister_hotcpu_notifier(&vscsi->nb);
1053 
1054 	virtscsi_remove_vqs(vdev);
1055 	scsi_host_put(shost);
1056 }
1057 
1058 #ifdef CONFIG_PM_SLEEP
1059 static int virtscsi_freeze(struct virtio_device *vdev)
1060 {
1061 	struct Scsi_Host *sh = virtio_scsi_host(vdev);
1062 	struct virtio_scsi *vscsi = shost_priv(sh);
1063 
1064 	unregister_hotcpu_notifier(&vscsi->nb);
1065 	virtscsi_remove_vqs(vdev);
1066 	return 0;
1067 }
1068 
1069 static int virtscsi_restore(struct virtio_device *vdev)
1070 {
1071 	struct Scsi_Host *sh = virtio_scsi_host(vdev);
1072 	struct virtio_scsi *vscsi = shost_priv(sh);
1073 	int err;
1074 
1075 	err = virtscsi_init(vdev, vscsi);
1076 	if (err)
1077 		return err;
1078 
1079 	err = register_hotcpu_notifier(&vscsi->nb);
1080 	if (err) {
1081 		vdev->config->del_vqs(vdev);
1082 		return err;
1083 	}
1084 
1085 	virtio_device_ready(vdev);
1086 
1087 	if (virtio_has_feature(vdev, VIRTIO_SCSI_F_HOTPLUG))
1088 		virtscsi_kick_event_all(vscsi);
1089 
1090 	return err;
1091 }
1092 #endif
1093 
1094 static struct virtio_device_id id_table[] = {
1095 	{ VIRTIO_ID_SCSI, VIRTIO_DEV_ANY_ID },
1096 	{ 0 },
1097 };
1098 
1099 static unsigned int features[] = {
1100 	VIRTIO_SCSI_F_HOTPLUG,
1101 	VIRTIO_SCSI_F_CHANGE,
1102 #ifdef CONFIG_BLK_DEV_INTEGRITY
1103 	VIRTIO_SCSI_F_T10_PI,
1104 #endif
1105 };
1106 
1107 static struct virtio_driver virtio_scsi_driver = {
1108 	.feature_table = features,
1109 	.feature_table_size = ARRAY_SIZE(features),
1110 	.driver.name = KBUILD_MODNAME,
1111 	.driver.owner = THIS_MODULE,
1112 	.id_table = id_table,
1113 	.probe = virtscsi_probe,
1114 #ifdef CONFIG_PM_SLEEP
1115 	.freeze = virtscsi_freeze,
1116 	.restore = virtscsi_restore,
1117 #endif
1118 	.remove = virtscsi_remove,
1119 };
1120 
1121 static int __init init(void)
1122 {
1123 	int ret = -ENOMEM;
1124 
1125 	virtscsi_cmd_cache = KMEM_CACHE(virtio_scsi_cmd, 0);
1126 	if (!virtscsi_cmd_cache) {
1127 		pr_err("kmem_cache_create() for virtscsi_cmd_cache failed\n");
1128 		goto error;
1129 	}
1130 
1131 
1132 	virtscsi_cmd_pool =
1133 		mempool_create_slab_pool(VIRTIO_SCSI_MEMPOOL_SZ,
1134 					 virtscsi_cmd_cache);
1135 	if (!virtscsi_cmd_pool) {
1136 		pr_err("mempool_create() for virtscsi_cmd_pool failed\n");
1137 		goto error;
1138 	}
1139 	ret = register_virtio_driver(&virtio_scsi_driver);
1140 	if (ret < 0)
1141 		goto error;
1142 
1143 	return 0;
1144 
1145 error:
1146 	if (virtscsi_cmd_pool) {
1147 		mempool_destroy(virtscsi_cmd_pool);
1148 		virtscsi_cmd_pool = NULL;
1149 	}
1150 	if (virtscsi_cmd_cache) {
1151 		kmem_cache_destroy(virtscsi_cmd_cache);
1152 		virtscsi_cmd_cache = NULL;
1153 	}
1154 	return ret;
1155 }
1156 
1157 static void __exit fini(void)
1158 {
1159 	unregister_virtio_driver(&virtio_scsi_driver);
1160 	mempool_destroy(virtscsi_cmd_pool);
1161 	kmem_cache_destroy(virtscsi_cmd_cache);
1162 }
1163 module_init(init);
1164 module_exit(fini);
1165 
1166 MODULE_DEVICE_TABLE(virtio, id_table);
1167 MODULE_DESCRIPTION("Virtio SCSI HBA driver");
1168 MODULE_LICENSE("GPL");
1169