xref: /openbmc/linux/drivers/scsi/virtio_scsi.c (revision 33ac9dba)
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 	struct virtio_scsi_vq ctrl_vq;
114 	struct virtio_scsi_vq event_vq;
115 	struct virtio_scsi_vq req_vqs[];
116 };
117 
118 static struct kmem_cache *virtscsi_cmd_cache;
119 static mempool_t *virtscsi_cmd_pool;
120 
121 static inline struct Scsi_Host *virtio_scsi_host(struct virtio_device *vdev)
122 {
123 	return vdev->priv;
124 }
125 
126 static void virtscsi_compute_resid(struct scsi_cmnd *sc, u32 resid)
127 {
128 	if (!resid)
129 		return;
130 
131 	if (!scsi_bidi_cmnd(sc)) {
132 		scsi_set_resid(sc, resid);
133 		return;
134 	}
135 
136 	scsi_in(sc)->resid = min(resid, scsi_in(sc)->length);
137 	scsi_out(sc)->resid = resid - scsi_in(sc)->resid;
138 }
139 
140 /**
141  * virtscsi_complete_cmd - finish a scsi_cmd and invoke scsi_done
142  *
143  * Called with vq_lock held.
144  */
145 static void virtscsi_complete_cmd(struct virtio_scsi *vscsi, void *buf)
146 {
147 	struct virtio_scsi_cmd *cmd = buf;
148 	struct scsi_cmnd *sc = cmd->sc;
149 	struct virtio_scsi_cmd_resp *resp = &cmd->resp.cmd;
150 	struct virtio_scsi_target_state *tgt =
151 				scsi_target(sc->device)->hostdata;
152 
153 	dev_dbg(&sc->device->sdev_gendev,
154 		"cmd %p response %u status %#02x sense_len %u\n",
155 		sc, resp->response, resp->status, resp->sense_len);
156 
157 	sc->result = resp->status;
158 	virtscsi_compute_resid(sc, resp->resid);
159 	switch (resp->response) {
160 	case VIRTIO_SCSI_S_OK:
161 		set_host_byte(sc, DID_OK);
162 		break;
163 	case VIRTIO_SCSI_S_OVERRUN:
164 		set_host_byte(sc, DID_ERROR);
165 		break;
166 	case VIRTIO_SCSI_S_ABORTED:
167 		set_host_byte(sc, DID_ABORT);
168 		break;
169 	case VIRTIO_SCSI_S_BAD_TARGET:
170 		set_host_byte(sc, DID_BAD_TARGET);
171 		break;
172 	case VIRTIO_SCSI_S_RESET:
173 		set_host_byte(sc, DID_RESET);
174 		break;
175 	case VIRTIO_SCSI_S_BUSY:
176 		set_host_byte(sc, DID_BUS_BUSY);
177 		break;
178 	case VIRTIO_SCSI_S_TRANSPORT_FAILURE:
179 		set_host_byte(sc, DID_TRANSPORT_DISRUPTED);
180 		break;
181 	case VIRTIO_SCSI_S_TARGET_FAILURE:
182 		set_host_byte(sc, DID_TARGET_FAILURE);
183 		break;
184 	case VIRTIO_SCSI_S_NEXUS_FAILURE:
185 		set_host_byte(sc, DID_NEXUS_FAILURE);
186 		break;
187 	default:
188 		scmd_printk(KERN_WARNING, sc, "Unknown response %d",
189 			    resp->response);
190 		/* fall through */
191 	case VIRTIO_SCSI_S_FAILURE:
192 		set_host_byte(sc, DID_ERROR);
193 		break;
194 	}
195 
196 	WARN_ON(resp->sense_len > VIRTIO_SCSI_SENSE_SIZE);
197 	if (sc->sense_buffer) {
198 		memcpy(sc->sense_buffer, resp->sense,
199 		       min_t(u32, resp->sense_len, VIRTIO_SCSI_SENSE_SIZE));
200 		if (resp->sense_len)
201 			set_driver_byte(sc, DRIVER_SENSE);
202 	}
203 
204 	sc->scsi_done(sc);
205 
206 	atomic_dec(&tgt->reqs);
207 }
208 
209 static void virtscsi_vq_done(struct virtio_scsi *vscsi,
210 			     struct virtio_scsi_vq *virtscsi_vq,
211 			     void (*fn)(struct virtio_scsi *vscsi, void *buf))
212 {
213 	void *buf;
214 	unsigned int len;
215 	unsigned long flags;
216 	struct virtqueue *vq = virtscsi_vq->vq;
217 
218 	spin_lock_irqsave(&virtscsi_vq->vq_lock, flags);
219 	do {
220 		virtqueue_disable_cb(vq);
221 		while ((buf = virtqueue_get_buf(vq, &len)) != NULL)
222 			fn(vscsi, buf);
223 
224 		if (unlikely(virtqueue_is_broken(vq)))
225 			break;
226 	} while (!virtqueue_enable_cb(vq));
227 	spin_unlock_irqrestore(&virtscsi_vq->vq_lock, flags);
228 }
229 
230 static void virtscsi_req_done(struct virtqueue *vq)
231 {
232 	struct Scsi_Host *sh = virtio_scsi_host(vq->vdev);
233 	struct virtio_scsi *vscsi = shost_priv(sh);
234 	int index = vq->index - VIRTIO_SCSI_VQ_BASE;
235 	struct virtio_scsi_vq *req_vq = &vscsi->req_vqs[index];
236 
237 	virtscsi_vq_done(vscsi, req_vq, virtscsi_complete_cmd);
238 };
239 
240 static void virtscsi_poll_requests(struct virtio_scsi *vscsi)
241 {
242 	int i, num_vqs;
243 
244 	num_vqs = vscsi->num_queues;
245 	for (i = 0; i < num_vqs; i++)
246 		virtscsi_vq_done(vscsi, &vscsi->req_vqs[i],
247 				 virtscsi_complete_cmd);
248 }
249 
250 static void virtscsi_complete_free(struct virtio_scsi *vscsi, void *buf)
251 {
252 	struct virtio_scsi_cmd *cmd = buf;
253 
254 	if (cmd->comp)
255 		complete_all(cmd->comp);
256 }
257 
258 static void virtscsi_ctrl_done(struct virtqueue *vq)
259 {
260 	struct Scsi_Host *sh = virtio_scsi_host(vq->vdev);
261 	struct virtio_scsi *vscsi = shost_priv(sh);
262 
263 	virtscsi_vq_done(vscsi, &vscsi->ctrl_vq, virtscsi_complete_free);
264 };
265 
266 static void virtscsi_handle_event(struct work_struct *work);
267 
268 static int virtscsi_kick_event(struct virtio_scsi *vscsi,
269 			       struct virtio_scsi_event_node *event_node)
270 {
271 	int err;
272 	struct scatterlist sg;
273 	unsigned long flags;
274 
275 	INIT_WORK(&event_node->work, virtscsi_handle_event);
276 	sg_init_one(&sg, &event_node->event, sizeof(struct virtio_scsi_event));
277 
278 	spin_lock_irqsave(&vscsi->event_vq.vq_lock, flags);
279 
280 	err = virtqueue_add_inbuf(vscsi->event_vq.vq, &sg, 1, event_node,
281 				  GFP_ATOMIC);
282 	if (!err)
283 		virtqueue_kick(vscsi->event_vq.vq);
284 
285 	spin_unlock_irqrestore(&vscsi->event_vq.vq_lock, flags);
286 
287 	return err;
288 }
289 
290 static int virtscsi_kick_event_all(struct virtio_scsi *vscsi)
291 {
292 	int i;
293 
294 	for (i = 0; i < VIRTIO_SCSI_EVENT_LEN; i++) {
295 		vscsi->event_list[i].vscsi = vscsi;
296 		virtscsi_kick_event(vscsi, &vscsi->event_list[i]);
297 	}
298 
299 	return 0;
300 }
301 
302 static void virtscsi_cancel_event_work(struct virtio_scsi *vscsi)
303 {
304 	int i;
305 
306 	for (i = 0; i < VIRTIO_SCSI_EVENT_LEN; i++)
307 		cancel_work_sync(&vscsi->event_list[i].work);
308 }
309 
310 static void virtscsi_handle_transport_reset(struct virtio_scsi *vscsi,
311 					    struct virtio_scsi_event *event)
312 {
313 	struct scsi_device *sdev;
314 	struct Scsi_Host *shost = virtio_scsi_host(vscsi->vdev);
315 	unsigned int target = event->lun[1];
316 	unsigned int lun = (event->lun[2] << 8) | event->lun[3];
317 
318 	switch (event->reason) {
319 	case VIRTIO_SCSI_EVT_RESET_RESCAN:
320 		scsi_add_device(shost, 0, target, lun);
321 		break;
322 	case VIRTIO_SCSI_EVT_RESET_REMOVED:
323 		sdev = scsi_device_lookup(shost, 0, target, lun);
324 		if (sdev) {
325 			scsi_remove_device(sdev);
326 			scsi_device_put(sdev);
327 		} else {
328 			pr_err("SCSI device %d 0 %d %d not found\n",
329 				shost->host_no, target, lun);
330 		}
331 		break;
332 	default:
333 		pr_info("Unsupport virtio scsi event reason %x\n", event->reason);
334 	}
335 }
336 
337 static void virtscsi_handle_param_change(struct virtio_scsi *vscsi,
338 					 struct virtio_scsi_event *event)
339 {
340 	struct scsi_device *sdev;
341 	struct Scsi_Host *shost = virtio_scsi_host(vscsi->vdev);
342 	unsigned int target = event->lun[1];
343 	unsigned int lun = (event->lun[2] << 8) | event->lun[3];
344 	u8 asc = event->reason & 255;
345 	u8 ascq = event->reason >> 8;
346 
347 	sdev = scsi_device_lookup(shost, 0, target, lun);
348 	if (!sdev) {
349 		pr_err("SCSI device %d 0 %d %d not found\n",
350 			shost->host_no, target, lun);
351 		return;
352 	}
353 
354 	/* Handle "Parameters changed", "Mode parameters changed", and
355 	   "Capacity data has changed".  */
356 	if (asc == 0x2a && (ascq == 0x00 || ascq == 0x01 || ascq == 0x09))
357 		scsi_rescan_device(&sdev->sdev_gendev);
358 
359 	scsi_device_put(sdev);
360 }
361 
362 static void virtscsi_handle_event(struct work_struct *work)
363 {
364 	struct virtio_scsi_event_node *event_node =
365 		container_of(work, struct virtio_scsi_event_node, work);
366 	struct virtio_scsi *vscsi = event_node->vscsi;
367 	struct virtio_scsi_event *event = &event_node->event;
368 
369 	if (event->event & VIRTIO_SCSI_T_EVENTS_MISSED) {
370 		event->event &= ~VIRTIO_SCSI_T_EVENTS_MISSED;
371 		scsi_scan_host(virtio_scsi_host(vscsi->vdev));
372 	}
373 
374 	switch (event->event) {
375 	case VIRTIO_SCSI_T_NO_EVENT:
376 		break;
377 	case VIRTIO_SCSI_T_TRANSPORT_RESET:
378 		virtscsi_handle_transport_reset(vscsi, event);
379 		break;
380 	case VIRTIO_SCSI_T_PARAM_CHANGE:
381 		virtscsi_handle_param_change(vscsi, event);
382 		break;
383 	default:
384 		pr_err("Unsupport virtio scsi event %x\n", event->event);
385 	}
386 	virtscsi_kick_event(vscsi, event_node);
387 }
388 
389 static void virtscsi_complete_event(struct virtio_scsi *vscsi, void *buf)
390 {
391 	struct virtio_scsi_event_node *event_node = buf;
392 
393 	schedule_work(&event_node->work);
394 }
395 
396 static void virtscsi_event_done(struct virtqueue *vq)
397 {
398 	struct Scsi_Host *sh = virtio_scsi_host(vq->vdev);
399 	struct virtio_scsi *vscsi = shost_priv(sh);
400 
401 	virtscsi_vq_done(vscsi, &vscsi->event_vq, virtscsi_complete_event);
402 };
403 
404 /**
405  * virtscsi_add_cmd - add a virtio_scsi_cmd to a virtqueue
406  * @vq		: the struct virtqueue we're talking about
407  * @cmd		: command structure
408  * @req_size	: size of the request buffer
409  * @resp_size	: size of the response buffer
410  */
411 static int virtscsi_add_cmd(struct virtqueue *vq,
412 			    struct virtio_scsi_cmd *cmd,
413 			    size_t req_size, size_t resp_size)
414 {
415 	struct scsi_cmnd *sc = cmd->sc;
416 	struct scatterlist *sgs[6], req, resp;
417 	struct sg_table *out, *in;
418 	unsigned out_num = 0, in_num = 0;
419 
420 	out = in = NULL;
421 
422 	if (sc && sc->sc_data_direction != DMA_NONE) {
423 		if (sc->sc_data_direction != DMA_FROM_DEVICE)
424 			out = &scsi_out(sc)->table;
425 		if (sc->sc_data_direction != DMA_TO_DEVICE)
426 			in = &scsi_in(sc)->table;
427 	}
428 
429 	/* Request header.  */
430 	sg_init_one(&req, &cmd->req, req_size);
431 	sgs[out_num++] = &req;
432 
433 	/* Data-out buffer.  */
434 	if (out) {
435 		/* Place WRITE protection SGLs before Data OUT payload */
436 		if (scsi_prot_sg_count(sc))
437 			sgs[out_num++] = scsi_prot_sglist(sc);
438 		sgs[out_num++] = out->sgl;
439 	}
440 
441 	/* Response header.  */
442 	sg_init_one(&resp, &cmd->resp, resp_size);
443 	sgs[out_num + in_num++] = &resp;
444 
445 	/* Data-in buffer */
446 	if (in) {
447 		/* Place READ protection SGLs before Data IN payload */
448 		if (scsi_prot_sg_count(sc))
449 			sgs[out_num + in_num++] = scsi_prot_sglist(sc);
450 		sgs[out_num + in_num++] = in->sgl;
451 	}
452 
453 	return virtqueue_add_sgs(vq, sgs, out_num, in_num, cmd, GFP_ATOMIC);
454 }
455 
456 static int virtscsi_kick_cmd(struct virtio_scsi_vq *vq,
457 			     struct virtio_scsi_cmd *cmd,
458 			     size_t req_size, size_t resp_size)
459 {
460 	unsigned long flags;
461 	int err;
462 	bool needs_kick = false;
463 
464 	spin_lock_irqsave(&vq->vq_lock, flags);
465 	err = virtscsi_add_cmd(vq->vq, cmd, req_size, resp_size);
466 	if (!err)
467 		needs_kick = virtqueue_kick_prepare(vq->vq);
468 
469 	spin_unlock_irqrestore(&vq->vq_lock, flags);
470 
471 	if (needs_kick)
472 		virtqueue_notify(vq->vq);
473 	return err;
474 }
475 
476 static void virtio_scsi_init_hdr(struct virtio_scsi_cmd_req *cmd,
477 				 struct scsi_cmnd *sc)
478 {
479 	cmd->lun[0] = 1;
480 	cmd->lun[1] = sc->device->id;
481 	cmd->lun[2] = (sc->device->lun >> 8) | 0x40;
482 	cmd->lun[3] = sc->device->lun & 0xff;
483 	cmd->tag = (unsigned long)sc;
484 	cmd->task_attr = VIRTIO_SCSI_S_SIMPLE;
485 	cmd->prio = 0;
486 	cmd->crn = 0;
487 }
488 
489 static void virtio_scsi_init_hdr_pi(struct virtio_scsi_cmd_req_pi *cmd_pi,
490 				    struct scsi_cmnd *sc)
491 {
492 	struct request *rq = sc->request;
493 	struct blk_integrity *bi;
494 
495 	virtio_scsi_init_hdr((struct virtio_scsi_cmd_req *)cmd_pi, sc);
496 
497 	if (!rq || !scsi_prot_sg_count(sc))
498 		return;
499 
500 	bi = blk_get_integrity(rq->rq_disk);
501 
502 	if (sc->sc_data_direction == DMA_TO_DEVICE)
503 		cmd_pi->pi_bytesout = blk_rq_sectors(rq) * bi->tuple_size;
504 	else if (sc->sc_data_direction == DMA_FROM_DEVICE)
505 		cmd_pi->pi_bytesin = blk_rq_sectors(rq) * bi->tuple_size;
506 }
507 
508 static int virtscsi_queuecommand(struct virtio_scsi *vscsi,
509 				 struct virtio_scsi_vq *req_vq,
510 				 struct scsi_cmnd *sc)
511 {
512 	struct Scsi_Host *shost = virtio_scsi_host(vscsi->vdev);
513 	struct virtio_scsi_cmd *cmd = scsi_cmd_priv(sc);
514 	int req_size;
515 
516 	BUG_ON(scsi_sg_count(sc) > shost->sg_tablesize);
517 
518 	/* TODO: check feature bit and fail if unsupported?  */
519 	BUG_ON(sc->sc_data_direction == DMA_BIDIRECTIONAL);
520 
521 	dev_dbg(&sc->device->sdev_gendev,
522 		"cmd %p CDB: %#02x\n", sc, sc->cmnd[0]);
523 
524 	memset(cmd, 0, sizeof(*cmd));
525 	cmd->sc = sc;
526 
527 	BUG_ON(sc->cmd_len > VIRTIO_SCSI_CDB_SIZE);
528 
529 	if (virtio_has_feature(vscsi->vdev, VIRTIO_SCSI_F_T10_PI)) {
530 		virtio_scsi_init_hdr_pi(&cmd->req.cmd_pi, sc);
531 		memcpy(cmd->req.cmd_pi.cdb, sc->cmnd, sc->cmd_len);
532 		req_size = sizeof(cmd->req.cmd_pi);
533 	} else {
534 		virtio_scsi_init_hdr(&cmd->req.cmd, sc);
535 		memcpy(cmd->req.cmd.cdb, sc->cmnd, sc->cmd_len);
536 		req_size = sizeof(cmd->req.cmd);
537 	}
538 
539 	if (virtscsi_kick_cmd(req_vq, cmd, req_size, sizeof(cmd->resp.cmd)) != 0)
540 		return SCSI_MLQUEUE_HOST_BUSY;
541 	return 0;
542 }
543 
544 static int virtscsi_queuecommand_single(struct Scsi_Host *sh,
545 					struct scsi_cmnd *sc)
546 {
547 	struct virtio_scsi *vscsi = shost_priv(sh);
548 	struct virtio_scsi_target_state *tgt =
549 				scsi_target(sc->device)->hostdata;
550 
551 	atomic_inc(&tgt->reqs);
552 	return virtscsi_queuecommand(vscsi, &vscsi->req_vqs[0], sc);
553 }
554 
555 static struct virtio_scsi_vq *virtscsi_pick_vq(struct virtio_scsi *vscsi,
556 					       struct virtio_scsi_target_state *tgt)
557 {
558 	struct virtio_scsi_vq *vq;
559 	unsigned long flags;
560 	u32 queue_num;
561 
562 	local_irq_save(flags);
563 	if (atomic_inc_return(&tgt->reqs) > 1) {
564 		unsigned long seq;
565 
566 		do {
567 			seq = read_seqcount_begin(&tgt->tgt_seq);
568 			vq = tgt->req_vq;
569 		} while (read_seqcount_retry(&tgt->tgt_seq, seq));
570 	} else {
571 		/* no writes can be concurrent because of atomic_t */
572 		write_seqcount_begin(&tgt->tgt_seq);
573 
574 		/* keep previous req_vq if a reader just arrived */
575 		if (unlikely(atomic_read(&tgt->reqs) > 1)) {
576 			vq = tgt->req_vq;
577 			goto unlock;
578 		}
579 
580 		queue_num = smp_processor_id();
581 		while (unlikely(queue_num >= vscsi->num_queues))
582 			queue_num -= vscsi->num_queues;
583 		tgt->req_vq = vq = &vscsi->req_vqs[queue_num];
584  unlock:
585 		write_seqcount_end(&tgt->tgt_seq);
586 	}
587 	local_irq_restore(flags);
588 
589 	return vq;
590 }
591 
592 static int virtscsi_queuecommand_multi(struct Scsi_Host *sh,
593 				       struct scsi_cmnd *sc)
594 {
595 	struct virtio_scsi *vscsi = shost_priv(sh);
596 	struct virtio_scsi_target_state *tgt =
597 				scsi_target(sc->device)->hostdata;
598 	struct virtio_scsi_vq *req_vq = virtscsi_pick_vq(vscsi, tgt);
599 
600 	return virtscsi_queuecommand(vscsi, req_vq, sc);
601 }
602 
603 static int virtscsi_tmf(struct virtio_scsi *vscsi, struct virtio_scsi_cmd *cmd)
604 {
605 	DECLARE_COMPLETION_ONSTACK(comp);
606 	int ret = FAILED;
607 
608 	cmd->comp = &comp;
609 	if (virtscsi_kick_cmd(&vscsi->ctrl_vq, cmd,
610 			      sizeof cmd->req.tmf, sizeof cmd->resp.tmf) < 0)
611 		goto out;
612 
613 	wait_for_completion(&comp);
614 	if (cmd->resp.tmf.response == VIRTIO_SCSI_S_OK ||
615 	    cmd->resp.tmf.response == VIRTIO_SCSI_S_FUNCTION_SUCCEEDED)
616 		ret = SUCCESS;
617 
618 	/*
619 	 * The spec guarantees that all requests related to the TMF have
620 	 * been completed, but the callback might not have run yet if
621 	 * we're using independent interrupts (e.g. MSI).  Poll the
622 	 * virtqueues once.
623 	 *
624 	 * In the abort case, sc->scsi_done will do nothing, because
625 	 * the block layer must have detected a timeout and as a result
626 	 * REQ_ATOM_COMPLETE has been set.
627 	 */
628 	virtscsi_poll_requests(vscsi);
629 
630 out:
631 	mempool_free(cmd, virtscsi_cmd_pool);
632 	return ret;
633 }
634 
635 static int virtscsi_device_reset(struct scsi_cmnd *sc)
636 {
637 	struct virtio_scsi *vscsi = shost_priv(sc->device->host);
638 	struct virtio_scsi_cmd *cmd;
639 
640 	sdev_printk(KERN_INFO, sc->device, "device reset\n");
641 	cmd = mempool_alloc(virtscsi_cmd_pool, GFP_NOIO);
642 	if (!cmd)
643 		return FAILED;
644 
645 	memset(cmd, 0, sizeof(*cmd));
646 	cmd->sc = sc;
647 	cmd->req.tmf = (struct virtio_scsi_ctrl_tmf_req){
648 		.type = VIRTIO_SCSI_T_TMF,
649 		.subtype = VIRTIO_SCSI_T_TMF_LOGICAL_UNIT_RESET,
650 		.lun[0] = 1,
651 		.lun[1] = sc->device->id,
652 		.lun[2] = (sc->device->lun >> 8) | 0x40,
653 		.lun[3] = sc->device->lun & 0xff,
654 	};
655 	return virtscsi_tmf(vscsi, cmd);
656 }
657 
658 /**
659  * virtscsi_change_queue_depth() - Change a virtscsi target's queue depth
660  * @sdev:	Virtscsi target whose queue depth to change
661  * @qdepth:	New queue depth
662  * @reason:	Reason for the queue depth change.
663  */
664 static int virtscsi_change_queue_depth(struct scsi_device *sdev,
665 				       int qdepth,
666 				       int reason)
667 {
668 	struct Scsi_Host *shost = sdev->host;
669 	int max_depth = shost->cmd_per_lun;
670 
671 	switch (reason) {
672 	case SCSI_QDEPTH_QFULL: /* Drop qdepth in response to BUSY state */
673 		scsi_track_queue_full(sdev, qdepth);
674 		break;
675 	case SCSI_QDEPTH_RAMP_UP: /* Raise qdepth after BUSY state resolved */
676 	case SCSI_QDEPTH_DEFAULT: /* Manual change via sysfs */
677 		scsi_adjust_queue_depth(sdev,
678 					scsi_get_tag_type(sdev),
679 					min(max_depth, qdepth));
680 		break;
681 	default:
682 		return -EOPNOTSUPP;
683 	}
684 
685 	return sdev->queue_depth;
686 }
687 
688 static int virtscsi_abort(struct scsi_cmnd *sc)
689 {
690 	struct virtio_scsi *vscsi = shost_priv(sc->device->host);
691 	struct virtio_scsi_cmd *cmd;
692 
693 	scmd_printk(KERN_INFO, sc, "abort\n");
694 	cmd = mempool_alloc(virtscsi_cmd_pool, GFP_NOIO);
695 	if (!cmd)
696 		return FAILED;
697 
698 	memset(cmd, 0, sizeof(*cmd));
699 	cmd->sc = sc;
700 	cmd->req.tmf = (struct virtio_scsi_ctrl_tmf_req){
701 		.type = VIRTIO_SCSI_T_TMF,
702 		.subtype = VIRTIO_SCSI_T_TMF_ABORT_TASK,
703 		.lun[0] = 1,
704 		.lun[1] = sc->device->id,
705 		.lun[2] = (sc->device->lun >> 8) | 0x40,
706 		.lun[3] = sc->device->lun & 0xff,
707 		.tag = (unsigned long)sc,
708 	};
709 	return virtscsi_tmf(vscsi, cmd);
710 }
711 
712 static int virtscsi_target_alloc(struct scsi_target *starget)
713 {
714 	struct Scsi_Host *sh = dev_to_shost(starget->dev.parent);
715 	struct virtio_scsi *vscsi = shost_priv(sh);
716 
717 	struct virtio_scsi_target_state *tgt =
718 				kmalloc(sizeof(*tgt), GFP_KERNEL);
719 	if (!tgt)
720 		return -ENOMEM;
721 
722 	seqcount_init(&tgt->tgt_seq);
723 	atomic_set(&tgt->reqs, 0);
724 	tgt->req_vq = &vscsi->req_vqs[0];
725 
726 	starget->hostdata = tgt;
727 	return 0;
728 }
729 
730 static void virtscsi_target_destroy(struct scsi_target *starget)
731 {
732 	struct virtio_scsi_target_state *tgt = starget->hostdata;
733 	kfree(tgt);
734 }
735 
736 static struct scsi_host_template virtscsi_host_template_single = {
737 	.module = THIS_MODULE,
738 	.name = "Virtio SCSI HBA",
739 	.proc_name = "virtio_scsi",
740 	.this_id = -1,
741 	.cmd_size = sizeof(struct virtio_scsi_cmd),
742 	.queuecommand = virtscsi_queuecommand_single,
743 	.change_queue_depth = virtscsi_change_queue_depth,
744 	.eh_abort_handler = virtscsi_abort,
745 	.eh_device_reset_handler = virtscsi_device_reset,
746 
747 	.can_queue = 1024,
748 	.dma_boundary = UINT_MAX,
749 	.use_clustering = ENABLE_CLUSTERING,
750 	.target_alloc = virtscsi_target_alloc,
751 	.target_destroy = virtscsi_target_destroy,
752 };
753 
754 static struct scsi_host_template virtscsi_host_template_multi = {
755 	.module = THIS_MODULE,
756 	.name = "Virtio SCSI HBA",
757 	.proc_name = "virtio_scsi",
758 	.this_id = -1,
759 	.cmd_size = sizeof(struct virtio_scsi_cmd),
760 	.queuecommand = virtscsi_queuecommand_multi,
761 	.change_queue_depth = virtscsi_change_queue_depth,
762 	.eh_abort_handler = virtscsi_abort,
763 	.eh_device_reset_handler = virtscsi_device_reset,
764 
765 	.can_queue = 1024,
766 	.dma_boundary = UINT_MAX,
767 	.use_clustering = ENABLE_CLUSTERING,
768 	.target_alloc = virtscsi_target_alloc,
769 	.target_destroy = virtscsi_target_destroy,
770 };
771 
772 #define virtscsi_config_get(vdev, fld) \
773 	({ \
774 		typeof(((struct virtio_scsi_config *)0)->fld) __val; \
775 		virtio_cread(vdev, struct virtio_scsi_config, fld, &__val); \
776 		__val; \
777 	})
778 
779 #define virtscsi_config_set(vdev, fld, val) \
780 	do { \
781 		typeof(((struct virtio_scsi_config *)0)->fld) __val = (val); \
782 		virtio_cwrite(vdev, struct virtio_scsi_config, fld, &__val); \
783 	} while(0)
784 
785 static void __virtscsi_set_affinity(struct virtio_scsi *vscsi, bool affinity)
786 {
787 	int i;
788 	int cpu;
789 
790 	/* In multiqueue mode, when the number of cpu is equal
791 	 * to the number of request queues, we let the qeueues
792 	 * to be private to one cpu by setting the affinity hint
793 	 * to eliminate the contention.
794 	 */
795 	if ((vscsi->num_queues == 1 ||
796 	     vscsi->num_queues != num_online_cpus()) && affinity) {
797 		if (vscsi->affinity_hint_set)
798 			affinity = false;
799 		else
800 			return;
801 	}
802 
803 	if (affinity) {
804 		i = 0;
805 		for_each_online_cpu(cpu) {
806 			virtqueue_set_affinity(vscsi->req_vqs[i].vq, cpu);
807 			i++;
808 		}
809 
810 		vscsi->affinity_hint_set = true;
811 	} else {
812 		for (i = 0; i < vscsi->num_queues; i++) {
813 			if (!vscsi->req_vqs[i].vq)
814 				continue;
815 
816 			virtqueue_set_affinity(vscsi->req_vqs[i].vq, -1);
817 		}
818 
819 		vscsi->affinity_hint_set = false;
820 	}
821 }
822 
823 static void virtscsi_set_affinity(struct virtio_scsi *vscsi, bool affinity)
824 {
825 	get_online_cpus();
826 	__virtscsi_set_affinity(vscsi, affinity);
827 	put_online_cpus();
828 }
829 
830 static int virtscsi_cpu_callback(struct notifier_block *nfb,
831 				 unsigned long action, void *hcpu)
832 {
833 	struct virtio_scsi *vscsi = container_of(nfb, struct virtio_scsi, nb);
834 	switch(action) {
835 	case CPU_ONLINE:
836 	case CPU_ONLINE_FROZEN:
837 	case CPU_DEAD:
838 	case CPU_DEAD_FROZEN:
839 		__virtscsi_set_affinity(vscsi, true);
840 		break;
841 	default:
842 		break;
843 	}
844 	return NOTIFY_OK;
845 }
846 
847 static void virtscsi_init_vq(struct virtio_scsi_vq *virtscsi_vq,
848 			     struct virtqueue *vq)
849 {
850 	spin_lock_init(&virtscsi_vq->vq_lock);
851 	virtscsi_vq->vq = vq;
852 }
853 
854 static void virtscsi_scan(struct virtio_device *vdev)
855 {
856 	struct Scsi_Host *shost = (struct Scsi_Host *)vdev->priv;
857 
858 	scsi_scan_host(shost);
859 }
860 
861 static void virtscsi_remove_vqs(struct virtio_device *vdev)
862 {
863 	struct Scsi_Host *sh = virtio_scsi_host(vdev);
864 	struct virtio_scsi *vscsi = shost_priv(sh);
865 
866 	virtscsi_set_affinity(vscsi, false);
867 
868 	/* Stop all the virtqueues. */
869 	vdev->config->reset(vdev);
870 
871 	vdev->config->del_vqs(vdev);
872 }
873 
874 static int virtscsi_init(struct virtio_device *vdev,
875 			 struct virtio_scsi *vscsi)
876 {
877 	int err;
878 	u32 i;
879 	u32 num_vqs;
880 	vq_callback_t **callbacks;
881 	const char **names;
882 	struct virtqueue **vqs;
883 
884 	num_vqs = vscsi->num_queues + VIRTIO_SCSI_VQ_BASE;
885 	vqs = kmalloc(num_vqs * sizeof(struct virtqueue *), GFP_KERNEL);
886 	callbacks = kmalloc(num_vqs * sizeof(vq_callback_t *), GFP_KERNEL);
887 	names = kmalloc(num_vqs * sizeof(char *), GFP_KERNEL);
888 
889 	if (!callbacks || !vqs || !names) {
890 		err = -ENOMEM;
891 		goto out;
892 	}
893 
894 	callbacks[0] = virtscsi_ctrl_done;
895 	callbacks[1] = virtscsi_event_done;
896 	names[0] = "control";
897 	names[1] = "event";
898 	for (i = VIRTIO_SCSI_VQ_BASE; i < num_vqs; i++) {
899 		callbacks[i] = virtscsi_req_done;
900 		names[i] = "request";
901 	}
902 
903 	/* Discover virtqueues and write information to configuration.  */
904 	err = vdev->config->find_vqs(vdev, num_vqs, vqs, callbacks, names);
905 	if (err)
906 		goto out;
907 
908 	virtscsi_init_vq(&vscsi->ctrl_vq, vqs[0]);
909 	virtscsi_init_vq(&vscsi->event_vq, vqs[1]);
910 	for (i = VIRTIO_SCSI_VQ_BASE; i < num_vqs; i++)
911 		virtscsi_init_vq(&vscsi->req_vqs[i - VIRTIO_SCSI_VQ_BASE],
912 				 vqs[i]);
913 
914 	virtscsi_set_affinity(vscsi, true);
915 
916 	virtscsi_config_set(vdev, cdb_size, VIRTIO_SCSI_CDB_SIZE);
917 	virtscsi_config_set(vdev, sense_size, VIRTIO_SCSI_SENSE_SIZE);
918 
919 	if (virtio_has_feature(vdev, VIRTIO_SCSI_F_HOTPLUG))
920 		virtscsi_kick_event_all(vscsi);
921 
922 	err = 0;
923 
924 out:
925 	kfree(names);
926 	kfree(callbacks);
927 	kfree(vqs);
928 	if (err)
929 		virtscsi_remove_vqs(vdev);
930 	return err;
931 }
932 
933 static int virtscsi_probe(struct virtio_device *vdev)
934 {
935 	struct Scsi_Host *shost;
936 	struct virtio_scsi *vscsi;
937 	int err, host_prot;
938 	u32 sg_elems, num_targets;
939 	u32 cmd_per_lun;
940 	u32 num_queues;
941 	struct scsi_host_template *hostt;
942 
943 	/* We need to know how many queues before we allocate. */
944 	num_queues = virtscsi_config_get(vdev, num_queues) ? : 1;
945 
946 	num_targets = virtscsi_config_get(vdev, max_target) + 1;
947 
948 	if (num_queues == 1)
949 		hostt = &virtscsi_host_template_single;
950 	else
951 		hostt = &virtscsi_host_template_multi;
952 
953 	shost = scsi_host_alloc(hostt,
954 		sizeof(*vscsi) + sizeof(vscsi->req_vqs[0]) * num_queues);
955 	if (!shost)
956 		return -ENOMEM;
957 
958 	sg_elems = virtscsi_config_get(vdev, seg_max) ?: 1;
959 	shost->sg_tablesize = sg_elems;
960 	vscsi = shost_priv(shost);
961 	vscsi->vdev = vdev;
962 	vscsi->num_queues = num_queues;
963 	vdev->priv = shost;
964 
965 	err = virtscsi_init(vdev, vscsi);
966 	if (err)
967 		goto virtscsi_init_failed;
968 
969 	vscsi->nb.notifier_call = &virtscsi_cpu_callback;
970 	err = register_hotcpu_notifier(&vscsi->nb);
971 	if (err) {
972 		pr_err("registering cpu notifier failed\n");
973 		goto scsi_add_host_failed;
974 	}
975 
976 	cmd_per_lun = virtscsi_config_get(vdev, cmd_per_lun) ?: 1;
977 	shost->cmd_per_lun = min_t(u32, cmd_per_lun, shost->can_queue);
978 	shost->max_sectors = virtscsi_config_get(vdev, max_sectors) ?: 0xFFFF;
979 
980 	/* LUNs > 256 are reported with format 1, so they go in the range
981 	 * 16640-32767.
982 	 */
983 	shost->max_lun = virtscsi_config_get(vdev, max_lun) + 1 + 0x4000;
984 	shost->max_id = num_targets;
985 	shost->max_channel = 0;
986 	shost->max_cmd_len = VIRTIO_SCSI_CDB_SIZE;
987 
988 	if (virtio_has_feature(vdev, VIRTIO_SCSI_F_T10_PI)) {
989 		host_prot = SHOST_DIF_TYPE1_PROTECTION | SHOST_DIF_TYPE2_PROTECTION |
990 			    SHOST_DIF_TYPE3_PROTECTION | SHOST_DIX_TYPE1_PROTECTION |
991 			    SHOST_DIX_TYPE2_PROTECTION | SHOST_DIX_TYPE3_PROTECTION;
992 
993 		scsi_host_set_prot(shost, host_prot);
994 		scsi_host_set_guard(shost, SHOST_DIX_GUARD_CRC);
995 	}
996 
997 	err = scsi_add_host(shost, &vdev->dev);
998 	if (err)
999 		goto scsi_add_host_failed;
1000 	/*
1001 	 * scsi_scan_host() happens in virtscsi_scan() via virtio_driver->scan()
1002 	 * after VIRTIO_CONFIG_S_DRIVER_OK has been set..
1003 	 */
1004 	return 0;
1005 
1006 scsi_add_host_failed:
1007 	vdev->config->del_vqs(vdev);
1008 virtscsi_init_failed:
1009 	scsi_host_put(shost);
1010 	return err;
1011 }
1012 
1013 static void virtscsi_remove(struct virtio_device *vdev)
1014 {
1015 	struct Scsi_Host *shost = virtio_scsi_host(vdev);
1016 	struct virtio_scsi *vscsi = shost_priv(shost);
1017 
1018 	if (virtio_has_feature(vdev, VIRTIO_SCSI_F_HOTPLUG))
1019 		virtscsi_cancel_event_work(vscsi);
1020 
1021 	scsi_remove_host(shost);
1022 
1023 	unregister_hotcpu_notifier(&vscsi->nb);
1024 
1025 	virtscsi_remove_vqs(vdev);
1026 	scsi_host_put(shost);
1027 }
1028 
1029 #ifdef CONFIG_PM_SLEEP
1030 static int virtscsi_freeze(struct virtio_device *vdev)
1031 {
1032 	struct Scsi_Host *sh = virtio_scsi_host(vdev);
1033 	struct virtio_scsi *vscsi = shost_priv(sh);
1034 
1035 	unregister_hotcpu_notifier(&vscsi->nb);
1036 	virtscsi_remove_vqs(vdev);
1037 	return 0;
1038 }
1039 
1040 static int virtscsi_restore(struct virtio_device *vdev)
1041 {
1042 	struct Scsi_Host *sh = virtio_scsi_host(vdev);
1043 	struct virtio_scsi *vscsi = shost_priv(sh);
1044 	int err;
1045 
1046 	err = virtscsi_init(vdev, vscsi);
1047 	if (err)
1048 		return err;
1049 
1050 	err = register_hotcpu_notifier(&vscsi->nb);
1051 	if (err)
1052 		vdev->config->del_vqs(vdev);
1053 
1054 	return err;
1055 }
1056 #endif
1057 
1058 static struct virtio_device_id id_table[] = {
1059 	{ VIRTIO_ID_SCSI, VIRTIO_DEV_ANY_ID },
1060 	{ 0 },
1061 };
1062 
1063 static unsigned int features[] = {
1064 	VIRTIO_SCSI_F_HOTPLUG,
1065 	VIRTIO_SCSI_F_CHANGE,
1066 	VIRTIO_SCSI_F_T10_PI,
1067 };
1068 
1069 static struct virtio_driver virtio_scsi_driver = {
1070 	.feature_table = features,
1071 	.feature_table_size = ARRAY_SIZE(features),
1072 	.driver.name = KBUILD_MODNAME,
1073 	.driver.owner = THIS_MODULE,
1074 	.id_table = id_table,
1075 	.probe = virtscsi_probe,
1076 	.scan = virtscsi_scan,
1077 #ifdef CONFIG_PM_SLEEP
1078 	.freeze = virtscsi_freeze,
1079 	.restore = virtscsi_restore,
1080 #endif
1081 	.remove = virtscsi_remove,
1082 };
1083 
1084 static int __init init(void)
1085 {
1086 	int ret = -ENOMEM;
1087 
1088 	virtscsi_cmd_cache = KMEM_CACHE(virtio_scsi_cmd, 0);
1089 	if (!virtscsi_cmd_cache) {
1090 		pr_err("kmem_cache_create() for virtscsi_cmd_cache failed\n");
1091 		goto error;
1092 	}
1093 
1094 
1095 	virtscsi_cmd_pool =
1096 		mempool_create_slab_pool(VIRTIO_SCSI_MEMPOOL_SZ,
1097 					 virtscsi_cmd_cache);
1098 	if (!virtscsi_cmd_pool) {
1099 		pr_err("mempool_create() for virtscsi_cmd_pool failed\n");
1100 		goto error;
1101 	}
1102 	ret = register_virtio_driver(&virtio_scsi_driver);
1103 	if (ret < 0)
1104 		goto error;
1105 
1106 	return 0;
1107 
1108 error:
1109 	if (virtscsi_cmd_pool) {
1110 		mempool_destroy(virtscsi_cmd_pool);
1111 		virtscsi_cmd_pool = NULL;
1112 	}
1113 	if (virtscsi_cmd_cache) {
1114 		kmem_cache_destroy(virtscsi_cmd_cache);
1115 		virtscsi_cmd_cache = NULL;
1116 	}
1117 	return ret;
1118 }
1119 
1120 static void __exit fini(void)
1121 {
1122 	unregister_virtio_driver(&virtio_scsi_driver);
1123 	mempool_destroy(virtscsi_cmd_pool);
1124 	kmem_cache_destroy(virtscsi_cmd_cache);
1125 }
1126 module_init(init);
1127 module_exit(fini);
1128 
1129 MODULE_DEVICE_TABLE(virtio, id_table);
1130 MODULE_DESCRIPTION("Virtio SCSI HBA driver");
1131 MODULE_LICENSE("GPL");
1132