1 /*
2  * Sharing QEMU block devices via vhost-user protocol
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 "qemu/error-report.h"
14 #include "block/block.h"
15 #include "subprojects/libvhost-user/libvhost-user.h" /* only for the type definitions */
16 #include "standard-headers/linux/virtio_blk.h"
17 #include "qemu/vhost-user-server.h"
18 #include "vhost-user-blk-server.h"
19 #include "qapi/error.h"
20 #include "qom/object_interfaces.h"
21 #include "util/block-helpers.h"
22 #include "virtio-blk-handler.h"
23 
24 enum {
25     VHOST_USER_BLK_NUM_QUEUES_DEFAULT = 1,
26 };
27 
28 typedef struct VuBlkReq {
29     VuVirtqElement elem;
30     VuServer *server;
31     struct VuVirtq *vq;
32 } VuBlkReq;
33 
34 /* vhost user block device */
35 typedef struct {
36     BlockExport export;
37     VuServer vu_server;
38     VirtioBlkHandler handler;
39     QIOChannelSocket *sioc;
40     struct virtio_blk_config blkcfg;
41 } VuBlkExport;
42 
43 static void vu_blk_req_complete(VuBlkReq *req, size_t in_len)
44 {
45     VuDev *vu_dev = &req->server->vu_dev;
46 
47     vu_queue_push(vu_dev, req->vq, &req->elem, in_len);
48     vu_queue_notify(vu_dev, req->vq);
49 
50     free(req);
51 }
52 
53 /*
54  * Called with server in_flight counter increased, must decrease before
55  * returning.
56  */
57 static void coroutine_fn vu_blk_virtio_process_req(void *opaque)
58 {
59     VuBlkReq *req = opaque;
60     VuServer *server = req->server;
61     VuVirtqElement *elem = &req->elem;
62     VuBlkExport *vexp = container_of(server, VuBlkExport, vu_server);
63     VirtioBlkHandler *handler = &vexp->handler;
64     struct iovec *in_iov = elem->in_sg;
65     struct iovec *out_iov = elem->out_sg;
66     unsigned in_num = elem->in_num;
67     unsigned out_num = elem->out_num;
68     int in_len;
69 
70     in_len = virtio_blk_process_req(handler, in_iov, out_iov,
71                                     in_num, out_num);
72     if (in_len < 0) {
73         free(req);
74         vhost_user_server_dec_in_flight(server);
75         return;
76     }
77 
78     vu_blk_req_complete(req, in_len);
79     vhost_user_server_dec_in_flight(server);
80 }
81 
82 static void vu_blk_process_vq(VuDev *vu_dev, int idx)
83 {
84     VuServer *server = container_of(vu_dev, VuServer, vu_dev);
85     VuVirtq *vq = vu_get_queue(vu_dev, idx);
86 
87     while (1) {
88         VuBlkReq *req;
89 
90         req = vu_queue_pop(vu_dev, vq, sizeof(VuBlkReq));
91         if (!req) {
92             break;
93         }
94 
95         req->server = server;
96         req->vq = vq;
97 
98         Coroutine *co =
99             qemu_coroutine_create(vu_blk_virtio_process_req, req);
100 
101         vhost_user_server_inc_in_flight(server);
102         qemu_coroutine_enter(co);
103     }
104 }
105 
106 static void vu_blk_queue_set_started(VuDev *vu_dev, int idx, bool started)
107 {
108     VuVirtq *vq;
109 
110     assert(vu_dev);
111 
112     vq = vu_get_queue(vu_dev, idx);
113     vu_set_queue_handler(vu_dev, vq, started ? vu_blk_process_vq : NULL);
114 }
115 
116 static uint64_t vu_blk_get_features(VuDev *dev)
117 {
118     uint64_t features;
119     VuServer *server = container_of(dev, VuServer, vu_dev);
120     VuBlkExport *vexp = container_of(server, VuBlkExport, vu_server);
121     features = 1ull << VIRTIO_BLK_F_SIZE_MAX |
122                1ull << VIRTIO_BLK_F_SEG_MAX |
123                1ull << VIRTIO_BLK_F_TOPOLOGY |
124                1ull << VIRTIO_BLK_F_BLK_SIZE |
125                1ull << VIRTIO_BLK_F_FLUSH |
126                1ull << VIRTIO_BLK_F_DISCARD |
127                1ull << VIRTIO_BLK_F_WRITE_ZEROES |
128                1ull << VIRTIO_BLK_F_CONFIG_WCE |
129                1ull << VIRTIO_BLK_F_MQ |
130                1ull << VIRTIO_F_VERSION_1 |
131                1ull << VIRTIO_RING_F_INDIRECT_DESC |
132                1ull << VIRTIO_RING_F_EVENT_IDX |
133                1ull << VHOST_USER_F_PROTOCOL_FEATURES;
134 
135     if (!vexp->handler.writable) {
136         features |= 1ull << VIRTIO_BLK_F_RO;
137     }
138 
139     return features;
140 }
141 
142 static uint64_t vu_blk_get_protocol_features(VuDev *dev)
143 {
144     return 1ull << VHOST_USER_PROTOCOL_F_CONFIG;
145 }
146 
147 static int
148 vu_blk_get_config(VuDev *vu_dev, uint8_t *config, uint32_t len)
149 {
150     VuServer *server = container_of(vu_dev, VuServer, vu_dev);
151     VuBlkExport *vexp = container_of(server, VuBlkExport, vu_server);
152 
153     if (len > sizeof(struct virtio_blk_config)) {
154         return -1;
155     }
156 
157     memcpy(config, &vexp->blkcfg, len);
158     return 0;
159 }
160 
161 static int
162 vu_blk_set_config(VuDev *vu_dev, const uint8_t *data,
163                     uint32_t offset, uint32_t size, uint32_t flags)
164 {
165     VuServer *server = container_of(vu_dev, VuServer, vu_dev);
166     VuBlkExport *vexp = container_of(server, VuBlkExport, vu_server);
167     uint8_t wce;
168 
169     /* don't support live migration */
170     if (flags != VHOST_SET_CONFIG_TYPE_FRONTEND) {
171         return -EINVAL;
172     }
173 
174     if (offset != offsetof(struct virtio_blk_config, wce) ||
175         size != 1) {
176         return -EINVAL;
177     }
178 
179     wce = *data;
180     vexp->blkcfg.wce = wce;
181     blk_set_enable_write_cache(vexp->export.blk, wce);
182     return 0;
183 }
184 
185 /*
186  * When the client disconnects, it sends a VHOST_USER_NONE request
187  * and vu_process_message will simple call exit which cause the VM
188  * to exit abruptly.
189  * To avoid this issue,  process VHOST_USER_NONE request ahead
190  * of vu_process_message.
191  *
192  */
193 static int vu_blk_process_msg(VuDev *dev, VhostUserMsg *vmsg, int *do_reply)
194 {
195     if (vmsg->request == VHOST_USER_NONE) {
196         dev->panic(dev, "disconnect");
197         return true;
198     }
199     return false;
200 }
201 
202 static const VuDevIface vu_blk_iface = {
203     .get_features          = vu_blk_get_features,
204     .queue_set_started     = vu_blk_queue_set_started,
205     .get_protocol_features = vu_blk_get_protocol_features,
206     .get_config            = vu_blk_get_config,
207     .set_config            = vu_blk_set_config,
208     .process_msg           = vu_blk_process_msg,
209 };
210 
211 static void blk_aio_attached(AioContext *ctx, void *opaque)
212 {
213     VuBlkExport *vexp = opaque;
214 
215     /*
216      * The actual attach will happen in vu_blk_drained_end() and we just
217      * restore ctx here.
218      */
219     vexp->export.ctx = ctx;
220 }
221 
222 static void blk_aio_detach(void *opaque)
223 {
224     VuBlkExport *vexp = opaque;
225 
226     /*
227      * The actual detach already happened in vu_blk_drained_begin() but from
228      * this point on we must not access ctx anymore.
229      */
230     vexp->export.ctx = NULL;
231 }
232 
233 static void
234 vu_blk_initialize_config(BlockDriverState *bs,
235                          struct virtio_blk_config *config,
236                          uint32_t blk_size,
237                          uint16_t num_queues)
238 {
239     config->capacity =
240         cpu_to_le64(bdrv_getlength(bs) >> VIRTIO_BLK_SECTOR_BITS);
241     config->blk_size = cpu_to_le32(blk_size);
242     config->size_max = cpu_to_le32(0);
243     config->seg_max = cpu_to_le32(128 - 2);
244     config->min_io_size = cpu_to_le16(1);
245     config->opt_io_size = cpu_to_le32(1);
246     config->num_queues = cpu_to_le16(num_queues);
247     config->max_discard_sectors =
248         cpu_to_le32(VIRTIO_BLK_MAX_DISCARD_SECTORS);
249     config->max_discard_seg = cpu_to_le32(1);
250     config->discard_sector_alignment =
251         cpu_to_le32(blk_size >> VIRTIO_BLK_SECTOR_BITS);
252     config->max_write_zeroes_sectors
253         = cpu_to_le32(VIRTIO_BLK_MAX_WRITE_ZEROES_SECTORS);
254     config->max_write_zeroes_seg = cpu_to_le32(1);
255 }
256 
257 static void vu_blk_exp_request_shutdown(BlockExport *exp)
258 {
259     VuBlkExport *vexp = container_of(exp, VuBlkExport, export);
260 
261     vhost_user_server_stop(&vexp->vu_server);
262 }
263 
264 static void vu_blk_exp_resize(void *opaque)
265 {
266     VuBlkExport *vexp = opaque;
267     BlockDriverState *bs = blk_bs(vexp->handler.blk);
268     int64_t new_size = bdrv_getlength(bs);
269 
270     if (new_size < 0) {
271         error_printf("Failed to get length of block node '%s'",
272                      bdrv_get_node_name(bs));
273         return;
274     }
275 
276     vexp->blkcfg.capacity = cpu_to_le64(new_size >> VIRTIO_BLK_SECTOR_BITS);
277 
278     vu_config_change_msg(&vexp->vu_server.vu_dev);
279 }
280 
281 /* Called with vexp->export.ctx acquired */
282 static void vu_blk_drained_begin(void *opaque)
283 {
284     VuBlkExport *vexp = opaque;
285 
286     vexp->vu_server.quiescing = true;
287     vhost_user_server_detach_aio_context(&vexp->vu_server);
288 }
289 
290 /* Called with vexp->export.blk AioContext acquired */
291 static void vu_blk_drained_end(void *opaque)
292 {
293     VuBlkExport *vexp = opaque;
294 
295     vexp->vu_server.quiescing = false;
296     vhost_user_server_attach_aio_context(&vexp->vu_server, vexp->export.ctx);
297 }
298 
299 /*
300  * Ensures that bdrv_drained_begin() waits until in-flight requests complete
301  * and the server->co_trip coroutine has terminated. It will be restarted in
302  * vhost_user_server_attach_aio_context().
303  *
304  * Called with vexp->export.ctx acquired.
305  */
306 static bool vu_blk_drained_poll(void *opaque)
307 {
308     VuBlkExport *vexp = opaque;
309     VuServer *server = &vexp->vu_server;
310 
311     return server->co_trip || vhost_user_server_has_in_flight(server);
312 }
313 
314 static const BlockDevOps vu_blk_dev_ops = {
315     .drained_begin = vu_blk_drained_begin,
316     .drained_end   = vu_blk_drained_end,
317     .drained_poll  = vu_blk_drained_poll,
318     .resize_cb = vu_blk_exp_resize,
319 };
320 
321 static int vu_blk_exp_create(BlockExport *exp, BlockExportOptions *opts,
322                              Error **errp)
323 {
324     VuBlkExport *vexp = container_of(exp, VuBlkExport, export);
325     BlockExportOptionsVhostUserBlk *vu_opts = &opts->u.vhost_user_blk;
326     Error *local_err = NULL;
327     uint64_t logical_block_size;
328     uint16_t num_queues = VHOST_USER_BLK_NUM_QUEUES_DEFAULT;
329 
330     vexp->blkcfg.wce = 0;
331 
332     if (vu_opts->has_logical_block_size) {
333         logical_block_size = vu_opts->logical_block_size;
334     } else {
335         logical_block_size = VIRTIO_BLK_SECTOR_SIZE;
336     }
337     check_block_size(exp->id, "logical-block-size", logical_block_size,
338                      &local_err);
339     if (local_err) {
340         error_propagate(errp, local_err);
341         return -EINVAL;
342     }
343 
344     if (vu_opts->has_num_queues) {
345         num_queues = vu_opts->num_queues;
346     }
347     if (num_queues == 0) {
348         error_setg(errp, "num-queues must be greater than 0");
349         return -EINVAL;
350     }
351     vexp->handler.blk = exp->blk;
352     vexp->handler.serial = g_strdup("vhost_user_blk");
353     vexp->handler.logical_block_size = logical_block_size;
354     vexp->handler.writable = opts->writable;
355 
356     vu_blk_initialize_config(blk_bs(exp->blk), &vexp->blkcfg,
357                              logical_block_size, num_queues);
358 
359     blk_add_aio_context_notifier(exp->blk, blk_aio_attached, blk_aio_detach,
360                                  vexp);
361 
362     blk_set_dev_ops(exp->blk, &vu_blk_dev_ops, vexp);
363 
364     if (!vhost_user_server_start(&vexp->vu_server, vu_opts->addr, exp->ctx,
365                                  num_queues, &vu_blk_iface, errp)) {
366         blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
367                                         blk_aio_detach, vexp);
368         g_free(vexp->handler.serial);
369         return -EADDRNOTAVAIL;
370     }
371 
372     return 0;
373 }
374 
375 static void vu_blk_exp_delete(BlockExport *exp)
376 {
377     VuBlkExport *vexp = container_of(exp, VuBlkExport, export);
378 
379     blk_remove_aio_context_notifier(exp->blk, blk_aio_attached, blk_aio_detach,
380                                     vexp);
381     g_free(vexp->handler.serial);
382 }
383 
384 const BlockExportDriver blk_exp_vhost_user_blk = {
385     .type               = BLOCK_EXPORT_TYPE_VHOST_USER_BLK,
386     .instance_size      = sizeof(VuBlkExport),
387     .create             = vu_blk_exp_create,
388     .delete             = vu_blk_exp_delete,
389     .request_shutdown   = vu_blk_exp_request_shutdown,
390 };
391