1 /*
2  * Sharing QEMU block devices via vhost-user protocal
3  *
4  * Parts of the code based on nbd/server.c.
5  *
6  * Copyright (c) Coiby Xu <coiby.xu@gmail.com>.
7  * Copyright (c) 2020 Red Hat, Inc.
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2 or
10  * later.  See the COPYING file in the top-level directory.
11  */
12 #include "qemu/osdep.h"
13 #include "block/block.h"
14 #include "subprojects/libvhost-user/libvhost-user.h" /* only for the type definitions */
15 #include "standard-headers/linux/virtio_blk.h"
16 #include "qemu/vhost-user-server.h"
17 #include "vhost-user-blk-server.h"
18 #include "qapi/error.h"
19 #include "qom/object_interfaces.h"
20 #include "sysemu/block-backend.h"
21 #include "util/block-helpers.h"
22 
23 /*
24  * Sector units are 512 bytes regardless of the
25  * virtio_blk_config->blk_size value.
26  */
27 #define VIRTIO_BLK_SECTOR_BITS 9
28 #define VIRTIO_BLK_SECTOR_SIZE (1ull << VIRTIO_BLK_SECTOR_BITS)
29 
30 enum {
31     VHOST_USER_BLK_NUM_QUEUES_DEFAULT = 1,
32     VHOST_USER_BLK_MAX_DISCARD_SECTORS = 32768,
33     VHOST_USER_BLK_MAX_WRITE_ZEROES_SECTORS = 32768,
34 };
35 struct virtio_blk_inhdr {
36     unsigned char status;
37 };
38 
39 typedef struct VuBlkReq {
40     VuVirtqElement elem;
41     int64_t sector_num;
42     size_t size;
43     struct virtio_blk_inhdr *in;
44     struct virtio_blk_outhdr out;
45     VuServer *server;
46     struct VuVirtq *vq;
47 } VuBlkReq;
48 
49 /* vhost user block device */
50 typedef struct {
51     BlockExport export;
52     VuServer vu_server;
53     uint32_t blk_size;
54     QIOChannelSocket *sioc;
55     struct virtio_blk_config blkcfg;
56     bool writable;
57 } VuBlkExport;
58 
59 static void vu_blk_req_complete(VuBlkReq *req)
60 {
61     VuDev *vu_dev = &req->server->vu_dev;
62 
63     /* IO size with 1 extra status byte */
64     vu_queue_push(vu_dev, req->vq, &req->elem, req->size + 1);
65     vu_queue_notify(vu_dev, req->vq);
66 
67     free(req);
68 }
69 
70 static bool vu_blk_sect_range_ok(VuBlkExport *vexp, uint64_t sector,
71                                  size_t size)
72 {
73     uint64_t nb_sectors;
74     uint64_t total_sectors;
75 
76     if (size % VIRTIO_BLK_SECTOR_SIZE) {
77         return false;
78     }
79 
80     nb_sectors = size >> VIRTIO_BLK_SECTOR_BITS;
81 
82     QEMU_BUILD_BUG_ON(BDRV_SECTOR_SIZE != VIRTIO_BLK_SECTOR_SIZE);
83     if (nb_sectors > BDRV_REQUEST_MAX_SECTORS) {
84         return false;
85     }
86     if ((sector << VIRTIO_BLK_SECTOR_BITS) % vexp->blk_size) {
87         return false;
88     }
89     blk_get_geometry(vexp->export.blk, &total_sectors);
90     if (sector > total_sectors || nb_sectors > total_sectors - sector) {
91         return false;
92     }
93     return true;
94 }
95 
96 static int coroutine_fn
97 vu_blk_discard_write_zeroes(VuBlkExport *vexp, struct iovec *iov,
98                             uint32_t iovcnt, uint32_t type)
99 {
100     BlockBackend *blk = vexp->export.blk;
101     struct virtio_blk_discard_write_zeroes desc;
102     ssize_t size;
103     uint64_t sector;
104     uint32_t num_sectors;
105     uint32_t max_sectors;
106     uint32_t flags;
107     int bytes;
108 
109     /* Only one desc is currently supported */
110     if (unlikely(iov_size(iov, iovcnt) > sizeof(desc))) {
111         return VIRTIO_BLK_S_UNSUPP;
112     }
113 
114     size = iov_to_buf(iov, iovcnt, 0, &desc, sizeof(desc));
115     if (unlikely(size != sizeof(desc))) {
116         error_report("Invalid size %zd, expected %zu", size, sizeof(desc));
117         return VIRTIO_BLK_S_IOERR;
118     }
119 
120     sector = le64_to_cpu(desc.sector);
121     num_sectors = le32_to_cpu(desc.num_sectors);
122     flags = le32_to_cpu(desc.flags);
123     max_sectors = (type == VIRTIO_BLK_T_WRITE_ZEROES) ?
124                   VHOST_USER_BLK_MAX_WRITE_ZEROES_SECTORS :
125                   VHOST_USER_BLK_MAX_DISCARD_SECTORS;
126 
127     /* This check ensures that 'bytes' fits in an int */
128     if (unlikely(num_sectors > max_sectors)) {
129         return VIRTIO_BLK_S_IOERR;
130     }
131 
132     bytes = num_sectors << VIRTIO_BLK_SECTOR_BITS;
133 
134     if (unlikely(!vu_blk_sect_range_ok(vexp, sector, bytes))) {
135         return VIRTIO_BLK_S_IOERR;
136     }
137 
138     /*
139      * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP for discard
140      * and write zeroes commands if any unknown flag is set.
141      */
142     if (unlikely(flags & ~VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP)) {
143         return VIRTIO_BLK_S_UNSUPP;
144     }
145 
146     if (type == VIRTIO_BLK_T_WRITE_ZEROES) {
147         int blk_flags = 0;
148 
149         if (flags & VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP) {
150             blk_flags |= BDRV_REQ_MAY_UNMAP;
151         }
152 
153         if (blk_co_pwrite_zeroes(blk, sector << VIRTIO_BLK_SECTOR_BITS,
154                                  bytes, blk_flags) == 0) {
155             return VIRTIO_BLK_S_OK;
156         }
157     } else if (type == VIRTIO_BLK_T_DISCARD) {
158         /*
159          * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP for
160          * discard commands if the unmap flag is set.
161          */
162         if (unlikely(flags & VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP)) {
163             return VIRTIO_BLK_S_UNSUPP;
164         }
165 
166         if (blk_co_pdiscard(blk, sector << VIRTIO_BLK_SECTOR_BITS,
167                             bytes) == 0) {
168             return VIRTIO_BLK_S_OK;
169         }
170     }
171 
172     return VIRTIO_BLK_S_IOERR;
173 }
174 
175 static void coroutine_fn vu_blk_virtio_process_req(void *opaque)
176 {
177     VuBlkReq *req = opaque;
178     VuServer *server = req->server;
179     VuVirtqElement *elem = &req->elem;
180     uint32_t type;
181 
182     VuBlkExport *vexp = container_of(server, VuBlkExport, vu_server);
183     BlockBackend *blk = vexp->export.blk;
184 
185     struct iovec *in_iov = elem->in_sg;
186     struct iovec *out_iov = elem->out_sg;
187     unsigned in_num = elem->in_num;
188     unsigned out_num = elem->out_num;
189 
190     /* refer to hw/block/virtio_blk.c */
191     if (elem->out_num < 1 || elem->in_num < 1) {
192         error_report("virtio-blk request missing headers");
193         goto err;
194     }
195 
196     if (unlikely(iov_to_buf(out_iov, out_num, 0, &req->out,
197                             sizeof(req->out)) != sizeof(req->out))) {
198         error_report("virtio-blk request outhdr too short");
199         goto err;
200     }
201 
202     iov_discard_front(&out_iov, &out_num, sizeof(req->out));
203 
204     if (in_iov[in_num - 1].iov_len < sizeof(struct virtio_blk_inhdr)) {
205         error_report("virtio-blk request inhdr too short");
206         goto err;
207     }
208 
209     /* We always touch the last byte, so just see how big in_iov is.  */
210     req->in = (void *)in_iov[in_num - 1].iov_base
211               + in_iov[in_num - 1].iov_len
212               - sizeof(struct virtio_blk_inhdr);
213     iov_discard_back(in_iov, &in_num, sizeof(struct virtio_blk_inhdr));
214 
215     type = le32_to_cpu(req->out.type);
216     switch (type & ~VIRTIO_BLK_T_BARRIER) {
217     case VIRTIO_BLK_T_IN:
218     case VIRTIO_BLK_T_OUT: {
219         QEMUIOVector qiov;
220         int64_t offset;
221         ssize_t ret = 0;
222         bool is_write = type & VIRTIO_BLK_T_OUT;
223         req->sector_num = le64_to_cpu(req->out.sector);
224 
225         if (is_write && !vexp->writable) {
226             req->in->status = VIRTIO_BLK_S_IOERR;
227             break;
228         }
229 
230         if (is_write) {
231             qemu_iovec_init_external(&qiov, out_iov, out_num);
232         } else {
233             qemu_iovec_init_external(&qiov, in_iov, in_num);
234         }
235 
236         if (unlikely(!vu_blk_sect_range_ok(vexp,
237                                            req->sector_num,
238                                            qiov.size))) {
239             req->in->status = VIRTIO_BLK_S_IOERR;
240             break;
241         }
242 
243         offset = req->sector_num << VIRTIO_BLK_SECTOR_BITS;
244 
245         if (is_write) {
246             ret = blk_co_pwritev(blk, offset, qiov.size, &qiov, 0);
247         } else {
248             ret = blk_co_preadv(blk, offset, qiov.size, &qiov, 0);
249         }
250         if (ret >= 0) {
251             req->in->status = VIRTIO_BLK_S_OK;
252         } else {
253             req->in->status = VIRTIO_BLK_S_IOERR;
254         }
255         break;
256     }
257     case VIRTIO_BLK_T_FLUSH:
258         if (blk_co_flush(blk) == 0) {
259             req->in->status = VIRTIO_BLK_S_OK;
260         } else {
261             req->in->status = VIRTIO_BLK_S_IOERR;
262         }
263         break;
264     case VIRTIO_BLK_T_GET_ID: {
265         size_t size = MIN(iov_size(&elem->in_sg[0], in_num),
266                           VIRTIO_BLK_ID_BYTES);
267         snprintf(elem->in_sg[0].iov_base, size, "%s", "vhost_user_blk");
268         req->in->status = VIRTIO_BLK_S_OK;
269         req->size = elem->in_sg[0].iov_len;
270         break;
271     }
272     case VIRTIO_BLK_T_DISCARD:
273     case VIRTIO_BLK_T_WRITE_ZEROES: {
274         if (!vexp->writable) {
275             req->in->status = VIRTIO_BLK_S_IOERR;
276             break;
277         }
278 
279         req->in->status = vu_blk_discard_write_zeroes(vexp, out_iov, out_num,
280                                                       type);
281         break;
282     }
283     default:
284         req->in->status = VIRTIO_BLK_S_UNSUPP;
285         break;
286     }
287 
288     vu_blk_req_complete(req);
289     return;
290 
291 err:
292     free(req);
293 }
294 
295 static void vu_blk_process_vq(VuDev *vu_dev, int idx)
296 {
297     VuServer *server = container_of(vu_dev, VuServer, vu_dev);
298     VuVirtq *vq = vu_get_queue(vu_dev, idx);
299 
300     while (1) {
301         VuBlkReq *req;
302 
303         req = vu_queue_pop(vu_dev, vq, sizeof(VuBlkReq));
304         if (!req) {
305             break;
306         }
307 
308         req->server = server;
309         req->vq = vq;
310 
311         Coroutine *co =
312             qemu_coroutine_create(vu_blk_virtio_process_req, req);
313         qemu_coroutine_enter(co);
314     }
315 }
316 
317 static void vu_blk_queue_set_started(VuDev *vu_dev, int idx, bool started)
318 {
319     VuVirtq *vq;
320 
321     assert(vu_dev);
322 
323     vq = vu_get_queue(vu_dev, idx);
324     vu_set_queue_handler(vu_dev, vq, started ? vu_blk_process_vq : NULL);
325 }
326 
327 static uint64_t vu_blk_get_features(VuDev *dev)
328 {
329     uint64_t features;
330     VuServer *server = container_of(dev, VuServer, vu_dev);
331     VuBlkExport *vexp = container_of(server, VuBlkExport, vu_server);
332     features = 1ull << VIRTIO_BLK_F_SIZE_MAX |
333                1ull << VIRTIO_BLK_F_SEG_MAX |
334                1ull << VIRTIO_BLK_F_TOPOLOGY |
335                1ull << VIRTIO_BLK_F_BLK_SIZE |
336                1ull << VIRTIO_BLK_F_FLUSH |
337                1ull << VIRTIO_BLK_F_DISCARD |
338                1ull << VIRTIO_BLK_F_WRITE_ZEROES |
339                1ull << VIRTIO_BLK_F_CONFIG_WCE |
340                1ull << VIRTIO_BLK_F_MQ |
341                1ull << VIRTIO_F_VERSION_1 |
342                1ull << VIRTIO_RING_F_INDIRECT_DESC |
343                1ull << VIRTIO_RING_F_EVENT_IDX |
344                1ull << VHOST_USER_F_PROTOCOL_FEATURES;
345 
346     if (!vexp->writable) {
347         features |= 1ull << VIRTIO_BLK_F_RO;
348     }
349 
350     return features;
351 }
352 
353 static uint64_t vu_blk_get_protocol_features(VuDev *dev)
354 {
355     return 1ull << VHOST_USER_PROTOCOL_F_CONFIG;
356 }
357 
358 static int
359 vu_blk_get_config(VuDev *vu_dev, uint8_t *config, uint32_t len)
360 {
361     VuServer *server = container_of(vu_dev, VuServer, vu_dev);
362     VuBlkExport *vexp = container_of(server, VuBlkExport, vu_server);
363 
364     if (len > sizeof(struct virtio_blk_config)) {
365         return -1;
366     }
367 
368     memcpy(config, &vexp->blkcfg, len);
369     return 0;
370 }
371 
372 static int
373 vu_blk_set_config(VuDev *vu_dev, const uint8_t *data,
374                     uint32_t offset, uint32_t size, uint32_t flags)
375 {
376     VuServer *server = container_of(vu_dev, VuServer, vu_dev);
377     VuBlkExport *vexp = container_of(server, VuBlkExport, vu_server);
378     uint8_t wce;
379 
380     /* don't support live migration */
381     if (flags != VHOST_SET_CONFIG_TYPE_MASTER) {
382         return -EINVAL;
383     }
384 
385     if (offset != offsetof(struct virtio_blk_config, wce) ||
386         size != 1) {
387         return -EINVAL;
388     }
389 
390     wce = *data;
391     vexp->blkcfg.wce = wce;
392     blk_set_enable_write_cache(vexp->export.blk, wce);
393     return 0;
394 }
395 
396 /*
397  * When the client disconnects, it sends a VHOST_USER_NONE request
398  * and vu_process_message will simple call exit which cause the VM
399  * to exit abruptly.
400  * To avoid this issue,  process VHOST_USER_NONE request ahead
401  * of vu_process_message.
402  *
403  */
404 static int vu_blk_process_msg(VuDev *dev, VhostUserMsg *vmsg, int *do_reply)
405 {
406     if (vmsg->request == VHOST_USER_NONE) {
407         dev->panic(dev, "disconnect");
408         return true;
409     }
410     return false;
411 }
412 
413 static const VuDevIface vu_blk_iface = {
414     .get_features          = vu_blk_get_features,
415     .queue_set_started     = vu_blk_queue_set_started,
416     .get_protocol_features = vu_blk_get_protocol_features,
417     .get_config            = vu_blk_get_config,
418     .set_config            = vu_blk_set_config,
419     .process_msg           = vu_blk_process_msg,
420 };
421 
422 static void blk_aio_attached(AioContext *ctx, void *opaque)
423 {
424     VuBlkExport *vexp = opaque;
425 
426     vexp->export.ctx = ctx;
427     vhost_user_server_attach_aio_context(&vexp->vu_server, ctx);
428 }
429 
430 static void blk_aio_detach(void *opaque)
431 {
432     VuBlkExport *vexp = opaque;
433 
434     vhost_user_server_detach_aio_context(&vexp->vu_server);
435     vexp->export.ctx = NULL;
436 }
437 
438 static void
439 vu_blk_initialize_config(BlockDriverState *bs,
440                          struct virtio_blk_config *config,
441                          uint32_t blk_size,
442                          uint16_t num_queues)
443 {
444     config->capacity =
445         cpu_to_le64(bdrv_getlength(bs) >> VIRTIO_BLK_SECTOR_BITS);
446     config->blk_size = cpu_to_le32(blk_size);
447     config->size_max = cpu_to_le32(0);
448     config->seg_max = cpu_to_le32(128 - 2);
449     config->min_io_size = cpu_to_le16(1);
450     config->opt_io_size = cpu_to_le32(1);
451     config->num_queues = cpu_to_le16(num_queues);
452     config->max_discard_sectors =
453         cpu_to_le32(VHOST_USER_BLK_MAX_DISCARD_SECTORS);
454     config->max_discard_seg = cpu_to_le32(1);
455     config->discard_sector_alignment =
456         cpu_to_le32(blk_size >> VIRTIO_BLK_SECTOR_BITS);
457     config->max_write_zeroes_sectors
458         = cpu_to_le32(VHOST_USER_BLK_MAX_WRITE_ZEROES_SECTORS);
459     config->max_write_zeroes_seg = cpu_to_le32(1);
460 }
461 
462 static void vu_blk_exp_request_shutdown(BlockExport *exp)
463 {
464     VuBlkExport *vexp = container_of(exp, VuBlkExport, export);
465 
466     vhost_user_server_stop(&vexp->vu_server);
467 }
468 
469 static int vu_blk_exp_create(BlockExport *exp, BlockExportOptions *opts,
470                              Error **errp)
471 {
472     VuBlkExport *vexp = container_of(exp, VuBlkExport, export);
473     BlockExportOptionsVhostUserBlk *vu_opts = &opts->u.vhost_user_blk;
474     Error *local_err = NULL;
475     uint64_t logical_block_size;
476     uint16_t num_queues = VHOST_USER_BLK_NUM_QUEUES_DEFAULT;
477 
478     vexp->writable = opts->writable;
479     vexp->blkcfg.wce = 0;
480 
481     if (vu_opts->has_logical_block_size) {
482         logical_block_size = vu_opts->logical_block_size;
483     } else {
484         logical_block_size = VIRTIO_BLK_SECTOR_SIZE;
485     }
486     check_block_size(exp->id, "logical-block-size", logical_block_size,
487                      &local_err);
488     if (local_err) {
489         error_propagate(errp, local_err);
490         return -EINVAL;
491     }
492     vexp->blk_size = logical_block_size;
493     blk_set_guest_block_size(exp->blk, logical_block_size);
494 
495     if (vu_opts->has_num_queues) {
496         num_queues = vu_opts->num_queues;
497     }
498     if (num_queues == 0) {
499         error_setg(errp, "num-queues must be greater than 0");
500         return -EINVAL;
501     }
502 
503     vu_blk_initialize_config(blk_bs(exp->blk), &vexp->blkcfg,
504                              logical_block_size, num_queues);
505 
506     blk_add_aio_context_notifier(exp->blk, blk_aio_attached, blk_aio_detach,
507                                  vexp);
508 
509     if (!vhost_user_server_start(&vexp->vu_server, vu_opts->addr, exp->ctx,
510                                  num_queues, &vu_blk_iface, errp)) {
511         blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
512                                         blk_aio_detach, vexp);
513         return -EADDRNOTAVAIL;
514     }
515 
516     return 0;
517 }
518 
519 static void vu_blk_exp_delete(BlockExport *exp)
520 {
521     VuBlkExport *vexp = container_of(exp, VuBlkExport, export);
522 
523     blk_remove_aio_context_notifier(exp->blk, blk_aio_attached, blk_aio_detach,
524                                     vexp);
525 }
526 
527 const BlockExportDriver blk_exp_vhost_user_blk = {
528     .type               = BLOCK_EXPORT_TYPE_VHOST_USER_BLK,
529     .instance_size      = sizeof(VuBlkExport),
530     .create             = vu_blk_exp_create,
531     .delete             = vu_blk_exp_delete,
532     .request_shutdown   = vu_blk_exp_request_shutdown,
533 };
534