]>
Commit | Line | Data |
---|---|---|
75818250 | 1 | /* |
a7c8ed36 | 2 | * Copyright Red Hat |
7a5ca864 FB |
3 | * Copyright (C) 2005 Anthony Liguori <anthony@codemonkey.ws> |
4 | * | |
798bfe00 | 5 | * Network Block Device Server Side |
7a5ca864 FB |
6 | * |
7 | * This program is free software; you can redistribute it and/or modify | |
8 | * it under the terms of the GNU General Public License as published by | |
9 | * the Free Software Foundation; under version 2 of the License. | |
10 | * | |
11 | * This program is distributed in the hope that it will be useful, | |
12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | * GNU General Public License for more details. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
8167ee88 | 17 | * along with this program; if not, see <http://www.gnu.org/licenses/>. |
75818250 | 18 | */ |
7a5ca864 | 19 | |
d38ea87a | 20 | #include "qemu/osdep.h" |
56ee8626 | 21 | |
e2c1c34f | 22 | #include "block/block_int.h" |
56ee8626 | 23 | #include "block/export.h" |
e2c1c34f | 24 | #include "block/dirty-bitmap.h" |
da34e65c | 25 | #include "qapi/error.h" |
dc5e9ac7 | 26 | #include "qemu/queue.h" |
9588463e | 27 | #include "trace.h" |
798bfe00 | 28 | #include "nbd-internal.h" |
416e34bd | 29 | #include "qemu/units.h" |
5df022cf | 30 | #include "qemu/memalign.h" |
ca441480 | 31 | |
e7b1948d | 32 | #define NBD_META_ID_BASE_ALLOCATION 0 |
71719cd5 | 33 | #define NBD_META_ID_ALLOCATION_DEPTH 1 |
3b1f244c | 34 | /* Dirty bitmaps use 'NBD_META_ID_DIRTY_BITMAP + i', so keep this id last. */ |
71719cd5 | 35 | #define NBD_META_ID_DIRTY_BITMAP 2 |
3d068aff | 36 | |
416e34bd EB |
37 | /* |
38 | * NBD_MAX_BLOCK_STATUS_EXTENTS: 1 MiB of extents data. An empirical | |
3d068aff VSO |
39 | * constant. If an increase is needed, note that the NBD protocol |
40 | * recommends no larger than 32 mb, so that the client won't consider | |
416e34bd EB |
41 | * the reply as a denial of service attack. |
42 | */ | |
43 | #define NBD_MAX_BLOCK_STATUS_EXTENTS (1 * MiB / 8) | |
e7b1948d | 44 | |
ca441480 PB |
45 | static int system_errno_to_nbd_errno(int err) |
46 | { | |
47 | switch (err) { | |
48 | case 0: | |
49 | return NBD_SUCCESS; | |
50 | case EPERM: | |
c0301fcc | 51 | case EROFS: |
ca441480 PB |
52 | return NBD_EPERM; |
53 | case EIO: | |
54 | return NBD_EIO; | |
55 | case ENOMEM: | |
56 | return NBD_ENOMEM; | |
57 | #ifdef EDQUOT | |
58 | case EDQUOT: | |
59 | #endif | |
60 | case EFBIG: | |
61 | case ENOSPC: | |
62 | return NBD_ENOSPC; | |
bae245d1 EB |
63 | case EOVERFLOW: |
64 | return NBD_EOVERFLOW; | |
0a479545 EB |
65 | case ENOTSUP: |
66 | #if ENOTSUP != EOPNOTSUPP | |
67 | case EOPNOTSUPP: | |
68 | #endif | |
69 | return NBD_ENOTSUP; | |
b6f5d3b5 EB |
70 | case ESHUTDOWN: |
71 | return NBD_ESHUTDOWN; | |
ca441480 PB |
72 | case EINVAL: |
73 | default: | |
74 | return NBD_EINVAL; | |
75 | } | |
76 | } | |
77 | ||
9a304d29 PB |
78 | /* Definitions for opaque data types */ |
79 | ||
315f78ab | 80 | typedef struct NBDRequestData NBDRequestData; |
9a304d29 | 81 | |
315f78ab | 82 | struct NBDRequestData { |
9a304d29 PB |
83 | NBDClient *client; |
84 | uint8_t *data; | |
29b6c3b3 | 85 | bool complete; |
9a304d29 PB |
86 | }; |
87 | ||
88 | struct NBDExport { | |
56ee8626 | 89 | BlockExport common; |
0ddf08db | 90 | |
ee0a19ec | 91 | char *name; |
b1a75b33 | 92 | char *description; |
9d26dfcb | 93 | uint64_t size; |
7423f417 | 94 | uint16_t nbdflags; |
4b9441f6 | 95 | QTAILQ_HEAD(, NBDClient) clients; |
ee0a19ec | 96 | QTAILQ_ENTRY(NBDExport) next; |
958c717d | 97 | |
cd7fca95 | 98 | BlockBackend *eject_notifier_blk; |
741cc431 | 99 | Notifier eject_notifier; |
3d068aff | 100 | |
71719cd5 | 101 | bool allocation_depth; |
3b1f244c EB |
102 | BdrvDirtyBitmap **export_bitmaps; |
103 | size_t nr_export_bitmaps; | |
9a304d29 PB |
104 | }; |
105 | ||
ee0a19ec PB |
106 | static QTAILQ_HEAD(, NBDExport) exports = QTAILQ_HEAD_INITIALIZER(exports); |
107 | ||
fd358d83 EB |
108 | /* |
109 | * NBDMetaContexts represents a list of meta contexts in use, | |
e7b1948d | 110 | * as selected by NBD_OPT_SET_META_CONTEXT. Also used for |
fd358d83 EB |
111 | * NBD_OPT_LIST_META_CONTEXT. |
112 | */ | |
113 | struct NBDMetaContexts { | |
114 | const NBDExport *exp; /* associated export */ | |
47ec485e | 115 | size_t count; /* number of negotiated contexts */ |
e7b1948d | 116 | bool base_allocation; /* export base:allocation context (block status) */ |
71719cd5 | 117 | bool allocation_depth; /* export qemu:allocation-depth */ |
3b1f244c EB |
118 | bool *bitmaps; /* |
119 | * export qemu:dirty-bitmap:<export bitmap name>, | |
120 | * sized by exp->nr_export_bitmaps | |
121 | */ | |
fd358d83 | 122 | }; |
e7b1948d | 123 | |
9a304d29 | 124 | struct NBDClient { |
f816310d | 125 | int refcount; /* atomic */ |
0c9390d9 | 126 | void (*close_fn)(NBDClient *client, bool negotiated); |
9a304d29 | 127 | |
7075d235 SH |
128 | QemuMutex lock; |
129 | ||
9a304d29 | 130 | NBDExport *exp; |
f95910fe | 131 | QCryptoTLSCreds *tlscreds; |
b25e12da | 132 | char *tlsauthz; |
1c778ef7 DB |
133 | QIOChannelSocket *sioc; /* The underlying data channel */ |
134 | QIOChannel *ioc; /* The current I/O channel which may differ (eg TLS) */ | |
9a304d29 | 135 | |
7075d235 | 136 | Coroutine *recv_coroutine; /* protected by lock */ |
9a304d29 PB |
137 | |
138 | CoMutex send_lock; | |
139 | Coroutine *send_coroutine; | |
140 | ||
7075d235 SH |
141 | bool read_yielding; /* protected by lock */ |
142 | bool quiescing; /* protected by lock */ | |
f148ae7d | 143 | |
4b9441f6 | 144 | QTAILQ_ENTRY(NBDClient) next; |
7075d235 SH |
145 | int nb_requests; /* protected by lock */ |
146 | bool closing; /* protected by lock */ | |
5c54e7fa | 147 | |
6e280648 EB |
148 | uint32_t check_align; /* If non-zero, check for aligned client requests */ |
149 | ||
ac132d05 | 150 | NBDMode mode; |
fd358d83 | 151 | NBDMetaContexts contexts; /* Negotiated meta contexts */ |
9a304d29 | 152 | |
0cfae925 VSO |
153 | uint32_t opt; /* Current option being negotiated */ |
154 | uint32_t optlen; /* remaining length of data in ioc for the option being | |
155 | negotiated now */ | |
156 | }; | |
7a5ca864 | 157 | |
ff82911c | 158 | static void nbd_client_receive_next_request(NBDClient *client); |
958c717d | 159 | |
6b8c01e7 | 160 | /* Basic flow for negotiation |
7a5ca864 FB |
161 | |
162 | Server Client | |
7a5ca864 | 163 | Negotiate |
6b8c01e7 PB |
164 | |
165 | or | |
166 | ||
167 | Server Client | |
168 | Negotiate #1 | |
169 | Option | |
170 | Negotiate #2 | |
171 | ||
172 | ---- | |
173 | ||
174 | followed by | |
175 | ||
176 | Server Client | |
7a5ca864 FB |
177 | Request |
178 | Response | |
179 | Request | |
180 | Response | |
181 | ... | |
182 | ... | |
183 | Request (type == 2) | |
6b8c01e7 | 184 | |
7a5ca864 FB |
185 | */ |
186 | ||
1d17922a VSO |
187 | static inline void set_be_option_rep(NBDOptionReply *rep, uint32_t option, |
188 | uint32_t type, uint32_t length) | |
189 | { | |
190 | stq_be_p(&rep->magic, NBD_REP_MAGIC); | |
191 | stl_be_p(&rep->option, option); | |
192 | stl_be_p(&rep->type, type); | |
193 | stl_be_p(&rep->length, length); | |
194 | } | |
195 | ||
526e5c65 EB |
196 | /* Send a reply header, including length, but no payload. |
197 | * Return -errno on error, 0 on success. */ | |
0cfae925 VSO |
198 | static int nbd_negotiate_send_rep_len(NBDClient *client, uint32_t type, |
199 | uint32_t len, Error **errp) | |
6b8c01e7 | 200 | { |
1d17922a | 201 | NBDOptionReply rep; |
6b8c01e7 | 202 | |
1d17922a | 203 | trace_nbd_negotiate_send_rep_len(client->opt, nbd_opt_lookup(client->opt), |
3736cc5b | 204 | type, nbd_rep_lookup(type), len); |
f95910fe | 205 | |
f37708f6 | 206 | assert(len < NBD_MAX_BUFFER_SIZE); |
2fd2c840 | 207 | |
1d17922a VSO |
208 | set_be_option_rep(&rep, client->opt, type, len); |
209 | return nbd_write(client->ioc, &rep, sizeof(rep), errp); | |
f5076b5a | 210 | } |
6b8c01e7 | 211 | |
526e5c65 EB |
212 | /* Send a reply header with default 0 length. |
213 | * Return -errno on error, 0 on success. */ | |
0cfae925 | 214 | static int nbd_negotiate_send_rep(NBDClient *client, uint32_t type, |
2fd2c840 | 215 | Error **errp) |
526e5c65 | 216 | { |
0cfae925 | 217 | return nbd_negotiate_send_rep_len(client, type, 0, errp); |
526e5c65 EB |
218 | } |
219 | ||
36683283 EB |
220 | /* Send an error reply. |
221 | * Return -errno on error, 0 on success. */ | |
9edc6313 | 222 | static int G_GNUC_PRINTF(4, 0) |
41f5dfaf EB |
223 | nbd_negotiate_send_rep_verr(NBDClient *client, uint32_t type, |
224 | Error **errp, const char *fmt, va_list va) | |
36683283 | 225 | { |
795d946d | 226 | ERRP_GUARD(); |
df18c04e | 227 | g_autofree char *msg = NULL; |
36683283 EB |
228 | int ret; |
229 | size_t len; | |
230 | ||
36683283 | 231 | msg = g_strdup_vprintf(fmt, va); |
36683283 | 232 | len = strlen(msg); |
5c4fe018 | 233 | assert(len < NBD_MAX_STRING_SIZE); |
9588463e | 234 | trace_nbd_negotiate_send_rep_err(msg); |
0cfae925 | 235 | ret = nbd_negotiate_send_rep_len(client, type, len, errp); |
36683283 | 236 | if (ret < 0) { |
df18c04e | 237 | return ret; |
36683283 | 238 | } |
0cfae925 | 239 | if (nbd_write(client->ioc, msg, len, errp) < 0) { |
2fd2c840 | 240 | error_prepend(errp, "write failed (error message): "); |
df18c04e | 241 | return -EIO; |
36683283 | 242 | } |
2fd2c840 | 243 | |
df18c04e | 244 | return 0; |
36683283 EB |
245 | } |
246 | ||
5c4fe018 EB |
247 | /* |
248 | * Return a malloc'd copy of @name suitable for use in an error reply. | |
249 | */ | |
250 | static char * | |
251 | nbd_sanitize_name(const char *name) | |
252 | { | |
253 | if (strnlen(name, 80) < 80) { | |
254 | return g_strdup(name); | |
255 | } | |
256 | /* XXX Should we also try to sanitize any control characters? */ | |
257 | return g_strdup_printf("%.80s...", name); | |
258 | } | |
259 | ||
41f5dfaf EB |
260 | /* Send an error reply. |
261 | * Return -errno on error, 0 on success. */ | |
9edc6313 | 262 | static int G_GNUC_PRINTF(4, 5) |
41f5dfaf EB |
263 | nbd_negotiate_send_rep_err(NBDClient *client, uint32_t type, |
264 | Error **errp, const char *fmt, ...) | |
265 | { | |
266 | va_list va; | |
267 | int ret; | |
268 | ||
269 | va_start(va, fmt); | |
270 | ret = nbd_negotiate_send_rep_verr(client, type, errp, fmt, va); | |
271 | va_end(va); | |
272 | return ret; | |
273 | } | |
274 | ||
894e0280 EB |
275 | /* Drop remainder of the current option, and send a reply with the |
276 | * given error type and message. Return -errno on read or write | |
277 | * failure; or 0 if connection is still live. */ | |
9edc6313 | 278 | static int G_GNUC_PRINTF(4, 0) |
2e425fd5 VSO |
279 | nbd_opt_vdrop(NBDClient *client, uint32_t type, Error **errp, |
280 | const char *fmt, va_list va) | |
894e0280 EB |
281 | { |
282 | int ret = nbd_drop(client->ioc, client->optlen, errp); | |
894e0280 EB |
283 | |
284 | client->optlen = 0; | |
285 | if (!ret) { | |
894e0280 | 286 | ret = nbd_negotiate_send_rep_verr(client, type, errp, fmt, va); |
894e0280 EB |
287 | } |
288 | return ret; | |
289 | } | |
290 | ||
9edc6313 | 291 | static int G_GNUC_PRINTF(4, 5) |
2e425fd5 VSO |
292 | nbd_opt_drop(NBDClient *client, uint32_t type, Error **errp, |
293 | const char *fmt, ...) | |
294 | { | |
295 | int ret; | |
296 | va_list va; | |
297 | ||
298 | va_start(va, fmt); | |
299 | ret = nbd_opt_vdrop(client, type, errp, fmt, va); | |
300 | va_end(va); | |
301 | ||
302 | return ret; | |
303 | } | |
304 | ||
9edc6313 | 305 | static int G_GNUC_PRINTF(3, 4) |
2e425fd5 VSO |
306 | nbd_opt_invalid(NBDClient *client, Error **errp, const char *fmt, ...) |
307 | { | |
308 | int ret; | |
309 | va_list va; | |
310 | ||
311 | va_start(va, fmt); | |
312 | ret = nbd_opt_vdrop(client, NBD_REP_ERR_INVALID, errp, fmt, va); | |
313 | va_end(va); | |
314 | ||
315 | return ret; | |
316 | } | |
317 | ||
894e0280 | 318 | /* Read size bytes from the unparsed payload of the current option. |
d1e2c3e7 | 319 | * If @check_nul, require that no NUL bytes appear in buffer. |
894e0280 EB |
320 | * Return -errno on I/O error, 0 if option was completely handled by |
321 | * sending a reply about inconsistent lengths, or 1 on success. */ | |
322 | static int nbd_opt_read(NBDClient *client, void *buffer, size_t size, | |
d1e2c3e7 | 323 | bool check_nul, Error **errp) |
894e0280 EB |
324 | { |
325 | if (size > client->optlen) { | |
2e425fd5 VSO |
326 | return nbd_opt_invalid(client, errp, |
327 | "Inconsistent lengths in option %s", | |
328 | nbd_opt_lookup(client->opt)); | |
894e0280 EB |
329 | } |
330 | client->optlen -= size; | |
d1e2c3e7 EB |
331 | if (qio_channel_read_all(client->ioc, buffer, size, errp) < 0) { |
332 | return -EIO; | |
333 | } | |
334 | ||
335 | if (check_nul && strnlen(buffer, size) != size) { | |
336 | return nbd_opt_invalid(client, errp, | |
337 | "Unexpected embedded NUL in option %s", | |
338 | nbd_opt_lookup(client->opt)); | |
339 | } | |
340 | return 1; | |
894e0280 EB |
341 | } |
342 | ||
e7b1948d VSO |
343 | /* Drop size bytes from the unparsed payload of the current option. |
344 | * Return -errno on I/O error, 0 if option was completely handled by | |
345 | * sending a reply about inconsistent lengths, or 1 on success. */ | |
346 | static int nbd_opt_skip(NBDClient *client, size_t size, Error **errp) | |
347 | { | |
348 | if (size > client->optlen) { | |
349 | return nbd_opt_invalid(client, errp, | |
350 | "Inconsistent lengths in option %s", | |
351 | nbd_opt_lookup(client->opt)); | |
352 | } | |
353 | client->optlen -= size; | |
354 | return nbd_drop(client->ioc, size, errp) < 0 ? -EIO : 1; | |
355 | } | |
356 | ||
12296459 VSO |
357 | /* nbd_opt_read_name |
358 | * | |
359 | * Read a string with the format: | |
93676c88 | 360 | * uint32_t len (<= NBD_MAX_STRING_SIZE) |
12296459 VSO |
361 | * len bytes string (not 0-terminated) |
362 | * | |
9d7ab222 | 363 | * On success, @name will be allocated. |
12296459 VSO |
364 | * If @length is non-null, it will be set to the actual string length. |
365 | * | |
366 | * Return -errno on I/O error, 0 if option was completely handled by | |
367 | * sending a reply about inconsistent lengths, or 1 on success. | |
368 | */ | |
9d7ab222 | 369 | static int nbd_opt_read_name(NBDClient *client, char **name, uint32_t *length, |
12296459 VSO |
370 | Error **errp) |
371 | { | |
372 | int ret; | |
373 | uint32_t len; | |
9d7ab222 | 374 | g_autofree char *local_name = NULL; |
12296459 | 375 | |
9d7ab222 | 376 | *name = NULL; |
d1e2c3e7 | 377 | ret = nbd_opt_read(client, &len, sizeof(len), false, errp); |
12296459 VSO |
378 | if (ret <= 0) { |
379 | return ret; | |
380 | } | |
80c7c2b0 | 381 | len = cpu_to_be32(len); |
12296459 | 382 | |
93676c88 | 383 | if (len > NBD_MAX_STRING_SIZE) { |
12296459 VSO |
384 | return nbd_opt_invalid(client, errp, |
385 | "Invalid name length: %" PRIu32, len); | |
386 | } | |
387 | ||
9d7ab222 | 388 | local_name = g_malloc(len + 1); |
d1e2c3e7 | 389 | ret = nbd_opt_read(client, local_name, len, true, errp); |
12296459 VSO |
390 | if (ret <= 0) { |
391 | return ret; | |
392 | } | |
9d7ab222 | 393 | local_name[len] = '\0'; |
12296459 VSO |
394 | |
395 | if (length) { | |
396 | *length = len; | |
397 | } | |
9d7ab222 | 398 | *name = g_steal_pointer(&local_name); |
12296459 VSO |
399 | |
400 | return 1; | |
401 | } | |
402 | ||
526e5c65 EB |
403 | /* Send a single NBD_REP_SERVER reply to NBD_OPT_LIST, including payload. |
404 | * Return -errno on error, 0 on success. */ | |
0cfae925 | 405 | static int nbd_negotiate_send_rep_list(NBDClient *client, NBDExport *exp, |
2fd2c840 | 406 | Error **errp) |
32d7d2e0 | 407 | { |
795d946d | 408 | ERRP_GUARD(); |
b1a75b33 | 409 | size_t name_len, desc_len; |
526e5c65 | 410 | uint32_t len; |
b1a75b33 EB |
411 | const char *name = exp->name ? exp->name : ""; |
412 | const char *desc = exp->description ? exp->description : ""; | |
0cfae925 | 413 | QIOChannel *ioc = client->ioc; |
2e5c9ad6 | 414 | int ret; |
32d7d2e0 | 415 | |
9588463e | 416 | trace_nbd_negotiate_send_rep_list(name, desc); |
b1a75b33 EB |
417 | name_len = strlen(name); |
418 | desc_len = strlen(desc); | |
93676c88 | 419 | assert(name_len <= NBD_MAX_STRING_SIZE && desc_len <= NBD_MAX_STRING_SIZE); |
526e5c65 | 420 | len = name_len + desc_len + sizeof(len); |
0cfae925 | 421 | ret = nbd_negotiate_send_rep_len(client, NBD_REP_SERVER, len, errp); |
2e5c9ad6 VSO |
422 | if (ret < 0) { |
423 | return ret; | |
32d7d2e0 | 424 | } |
526e5c65 | 425 | |
32d7d2e0 | 426 | len = cpu_to_be32(name_len); |
2fd2c840 VSO |
427 | if (nbd_write(ioc, &len, sizeof(len), errp) < 0) { |
428 | error_prepend(errp, "write failed (name length): "); | |
b1a75b33 EB |
429 | return -EINVAL; |
430 | } | |
2fd2c840 VSO |
431 | |
432 | if (nbd_write(ioc, name, name_len, errp) < 0) { | |
433 | error_prepend(errp, "write failed (name buffer): "); | |
32d7d2e0 HB |
434 | return -EINVAL; |
435 | } | |
2fd2c840 VSO |
436 | |
437 | if (nbd_write(ioc, desc, desc_len, errp) < 0) { | |
438 | error_prepend(errp, "write failed (description buffer): "); | |
32d7d2e0 HB |
439 | return -EINVAL; |
440 | } | |
2fd2c840 | 441 | |
32d7d2e0 HB |
442 | return 0; |
443 | } | |
444 | ||
526e5c65 EB |
445 | /* Process the NBD_OPT_LIST command, with a potential series of replies. |
446 | * Return -errno on error, 0 on success. */ | |
e68c35cf | 447 | static int nbd_negotiate_handle_list(NBDClient *client, Error **errp) |
32d7d2e0 | 448 | { |
32d7d2e0 | 449 | NBDExport *exp; |
0cfae925 | 450 | assert(client->opt == NBD_OPT_LIST); |
32d7d2e0 | 451 | |
32d7d2e0 HB |
452 | /* For each export, send a NBD_REP_SERVER reply. */ |
453 | QTAILQ_FOREACH(exp, &exports, next) { | |
0cfae925 | 454 | if (nbd_negotiate_send_rep_list(client, exp, errp)) { |
32d7d2e0 HB |
455 | return -EINVAL; |
456 | } | |
457 | } | |
458 | /* Finish with a NBD_REP_ACK. */ | |
0cfae925 | 459 | return nbd_negotiate_send_rep(client, NBD_REP_ACK, errp); |
32d7d2e0 HB |
460 | } |
461 | ||
fd358d83 | 462 | static void nbd_check_meta_export(NBDClient *client, NBDExport *exp) |
e7b1948d | 463 | { |
fd358d83 EB |
464 | if (exp != client->contexts.exp) { |
465 | client->contexts.count = 0; | |
47ec485e | 466 | } |
e7b1948d VSO |
467 | } |
468 | ||
f37708f6 EB |
469 | /* Send a reply to NBD_OPT_EXPORT_NAME. |
470 | * Return -errno on error, 0 on success. */ | |
dbb38caa | 471 | static int nbd_negotiate_handle_export_name(NBDClient *client, bool no_zeroes, |
2fd2c840 | 472 | Error **errp) |
f5076b5a | 473 | { |
795d946d | 474 | ERRP_GUARD(); |
9d7ab222 | 475 | g_autofree char *name = NULL; |
5f66d060 | 476 | char buf[NBD_REPLY_EXPORT_NAME_SIZE] = ""; |
23e099c3 EB |
477 | size_t len; |
478 | int ret; | |
dbb38caa | 479 | uint16_t myflags; |
6b8c01e7 | 480 | |
f5076b5a HB |
481 | /* Client sends: |
482 | [20 .. xx] export name (length bytes) | |
5f66d060 EB |
483 | Server replies: |
484 | [ 0 .. 7] size | |
485 | [ 8 .. 9] export flags | |
486 | [10 .. 133] reserved (0) [unless no_zeroes] | |
f5076b5a | 487 | */ |
9588463e | 488 | trace_nbd_negotiate_handle_export_name(); |
9c1d2614 EB |
489 | if (client->mode >= NBD_MODE_EXTENDED) { |
490 | error_setg(errp, "Extended headers already negotiated"); | |
491 | return -EINVAL; | |
492 | } | |
93676c88 | 493 | if (client->optlen > NBD_MAX_STRING_SIZE) { |
2fd2c840 | 494 | error_setg(errp, "Bad length received"); |
d9faeed8 | 495 | return -EINVAL; |
6b8c01e7 | 496 | } |
9d7ab222 | 497 | name = g_malloc(client->optlen + 1); |
e6798f06 | 498 | if (nbd_read(client->ioc, name, client->optlen, "export name", errp) < 0) { |
32f158a6 | 499 | return -EIO; |
6b8c01e7 | 500 | } |
0cfae925 VSO |
501 | name[client->optlen] = '\0'; |
502 | client->optlen = 0; | |
6b8c01e7 | 503 | |
9588463e | 504 | trace_nbd_negotiate_handle_export_name_request(name); |
9344e5f5 | 505 | |
6b8c01e7 PB |
506 | client->exp = nbd_export_find(name); |
507 | if (!client->exp) { | |
2fd2c840 | 508 | error_setg(errp, "export not found"); |
d9faeed8 | 509 | return -EINVAL; |
6b8c01e7 | 510 | } |
fd358d83 | 511 | nbd_check_meta_export(client, client->exp); |
6b8c01e7 | 512 | |
dbb38caa | 513 | myflags = client->exp->nbdflags; |
ac132d05 | 514 | if (client->mode >= NBD_MODE_STRUCTURED) { |
dbb38caa EB |
515 | myflags |= NBD_FLAG_SEND_DF; |
516 | } | |
2dcbb11b EB |
517 | if (client->mode >= NBD_MODE_EXTENDED && client->contexts.count) { |
518 | myflags |= NBD_FLAG_BLOCK_STAT_PAYLOAD; | |
519 | } | |
dbb38caa | 520 | trace_nbd_negotiate_new_style_size_flags(client->exp->size, myflags); |
23e099c3 | 521 | stq_be_p(buf, client->exp->size); |
dbb38caa | 522 | stw_be_p(buf + 8, myflags); |
23e099c3 EB |
523 | len = no_zeroes ? 10 : sizeof(buf); |
524 | ret = nbd_write(client->ioc, buf, len, errp); | |
525 | if (ret < 0) { | |
526 | error_prepend(errp, "write failed: "); | |
527 | return ret; | |
528 | } | |
529 | ||
6b8c01e7 | 530 | QTAILQ_INSERT_TAIL(&client->exp->clients, client, next); |
c69de1be | 531 | blk_exp_ref(&client->exp->common); |
d9faeed8 VSO |
532 | |
533 | return 0; | |
6b8c01e7 PB |
534 | } |
535 | ||
f37708f6 EB |
536 | /* Send a single NBD_REP_INFO, with a buffer @buf of @length bytes. |
537 | * The buffer does NOT include the info type prefix. | |
538 | * Return -errno on error, 0 if ready to send more. */ | |
0cfae925 | 539 | static int nbd_negotiate_send_info(NBDClient *client, |
f37708f6 EB |
540 | uint16_t info, uint32_t length, void *buf, |
541 | Error **errp) | |
542 | { | |
543 | int rc; | |
544 | ||
545 | trace_nbd_negotiate_send_info(info, nbd_info_lookup(info), length); | |
0cfae925 | 546 | rc = nbd_negotiate_send_rep_len(client, NBD_REP_INFO, |
f37708f6 EB |
547 | sizeof(info) + length, errp); |
548 | if (rc < 0) { | |
549 | return rc; | |
550 | } | |
80c7c2b0 | 551 | info = cpu_to_be16(info); |
f37708f6 EB |
552 | if (nbd_write(client->ioc, &info, sizeof(info), errp) < 0) { |
553 | return -EIO; | |
554 | } | |
555 | if (nbd_write(client->ioc, buf, length, errp) < 0) { | |
556 | return -EIO; | |
557 | } | |
558 | return 0; | |
559 | } | |
560 | ||
a16a7907 EB |
561 | /* nbd_reject_length: Handle any unexpected payload. |
562 | * @fatal requests that we quit talking to the client, even if we are able | |
563 | * to successfully send an error reply. | |
564 | * Return: | |
565 | * -errno transmission error occurred or @fatal was requested, errp is set | |
566 | * 0 error message successfully sent to client, errp is not set | |
567 | */ | |
0cfae925 | 568 | static int nbd_reject_length(NBDClient *client, bool fatal, Error **errp) |
a16a7907 EB |
569 | { |
570 | int ret; | |
571 | ||
0cfae925 | 572 | assert(client->optlen); |
2e425fd5 VSO |
573 | ret = nbd_opt_invalid(client, errp, "option '%s' has unexpected length", |
574 | nbd_opt_lookup(client->opt)); | |
a16a7907 | 575 | if (fatal && !ret) { |
894e0280 | 576 | error_setg(errp, "option '%s' has unexpected length", |
0cfae925 | 577 | nbd_opt_lookup(client->opt)); |
a16a7907 EB |
578 | return -EINVAL; |
579 | } | |
580 | return ret; | |
581 | } | |
582 | ||
f37708f6 EB |
583 | /* Handle NBD_OPT_INFO and NBD_OPT_GO. |
584 | * Return -errno on error, 0 if ready for next option, and 1 to move | |
585 | * into transmission phase. */ | |
dbb38caa | 586 | static int nbd_negotiate_handle_info(NBDClient *client, Error **errp) |
f37708f6 EB |
587 | { |
588 | int rc; | |
9d7ab222 | 589 | g_autofree char *name = NULL; |
f37708f6 EB |
590 | NBDExport *exp; |
591 | uint16_t requests; | |
592 | uint16_t request; | |
bbc35fc2 | 593 | uint32_t namelen = 0; |
f37708f6 | 594 | bool sendname = false; |
0c1d50bd EB |
595 | bool blocksize = false; |
596 | uint32_t sizes[3]; | |
f37708f6 | 597 | char buf[sizeof(uint64_t) + sizeof(uint16_t)]; |
6e280648 | 598 | uint32_t check_align = 0; |
dbb38caa | 599 | uint16_t myflags; |
f37708f6 EB |
600 | |
601 | /* Client sends: | |
602 | 4 bytes: L, name length (can be 0) | |
603 | L bytes: export name | |
604 | 2 bytes: N, number of requests (can be 0) | |
605 | N * 2 bytes: N requests | |
606 | */ | |
9d7ab222 | 607 | rc = nbd_opt_read_name(client, &name, &namelen, errp); |
894e0280 EB |
608 | if (rc <= 0) { |
609 | return rc; | |
f37708f6 | 610 | } |
f37708f6 EB |
611 | trace_nbd_negotiate_handle_export_name_request(name); |
612 | ||
d1e2c3e7 | 613 | rc = nbd_opt_read(client, &requests, sizeof(requests), false, errp); |
894e0280 EB |
614 | if (rc <= 0) { |
615 | return rc; | |
f37708f6 | 616 | } |
80c7c2b0 | 617 | requests = be16_to_cpu(requests); |
f37708f6 | 618 | trace_nbd_negotiate_handle_info_requests(requests); |
f37708f6 | 619 | while (requests--) { |
d1e2c3e7 | 620 | rc = nbd_opt_read(client, &request, sizeof(request), false, errp); |
894e0280 EB |
621 | if (rc <= 0) { |
622 | return rc; | |
f37708f6 | 623 | } |
80c7c2b0 | 624 | request = be16_to_cpu(request); |
f37708f6 EB |
625 | trace_nbd_negotiate_handle_info_request(request, |
626 | nbd_info_lookup(request)); | |
0c1d50bd EB |
627 | /* We care about NBD_INFO_NAME and NBD_INFO_BLOCK_SIZE; |
628 | * everything else is either a request we don't know or | |
629 | * something we send regardless of request */ | |
630 | switch (request) { | |
631 | case NBD_INFO_NAME: | |
f37708f6 | 632 | sendname = true; |
0c1d50bd EB |
633 | break; |
634 | case NBD_INFO_BLOCK_SIZE: | |
635 | blocksize = true; | |
636 | break; | |
f37708f6 EB |
637 | } |
638 | } | |
894e0280 EB |
639 | if (client->optlen) { |
640 | return nbd_reject_length(client, false, errp); | |
641 | } | |
f37708f6 EB |
642 | |
643 | exp = nbd_export_find(name); | |
644 | if (!exp) { | |
5c4fe018 EB |
645 | g_autofree char *sane_name = nbd_sanitize_name(name); |
646 | ||
0cfae925 VSO |
647 | return nbd_negotiate_send_rep_err(client, NBD_REP_ERR_UNKNOWN, |
648 | errp, "export '%s' not present", | |
5c4fe018 | 649 | sane_name); |
f37708f6 | 650 | } |
fd358d83 EB |
651 | if (client->opt == NBD_OPT_GO) { |
652 | nbd_check_meta_export(client, exp); | |
653 | } | |
f37708f6 EB |
654 | |
655 | /* Don't bother sending NBD_INFO_NAME unless client requested it */ | |
656 | if (sendname) { | |
0cfae925 | 657 | rc = nbd_negotiate_send_info(client, NBD_INFO_NAME, namelen, name, |
f37708f6 EB |
658 | errp); |
659 | if (rc < 0) { | |
660 | return rc; | |
661 | } | |
662 | } | |
663 | ||
664 | /* Send NBD_INFO_DESCRIPTION only if available, regardless of | |
665 | * client request */ | |
666 | if (exp->description) { | |
667 | size_t len = strlen(exp->description); | |
668 | ||
93676c88 | 669 | assert(len <= NBD_MAX_STRING_SIZE); |
0cfae925 | 670 | rc = nbd_negotiate_send_info(client, NBD_INFO_DESCRIPTION, |
f37708f6 EB |
671 | len, exp->description, errp); |
672 | if (rc < 0) { | |
673 | return rc; | |
674 | } | |
675 | } | |
676 | ||
0c1d50bd EB |
677 | /* Send NBD_INFO_BLOCK_SIZE always, but tweak the minimum size |
678 | * according to whether the client requested it, and according to | |
679 | * whether this is OPT_INFO or OPT_GO. */ | |
b0245d64 EB |
680 | /* minimum - 1 for back-compat, or actual if client will obey it. */ |
681 | if (client->opt == NBD_OPT_INFO || blocksize) { | |
37a4f70c | 682 | check_align = sizes[0] = blk_get_request_alignment(exp->common.blk); |
b0245d64 EB |
683 | } else { |
684 | sizes[0] = 1; | |
685 | } | |
686 | assert(sizes[0] <= NBD_MAX_BUFFER_SIZE); | |
0c1d50bd EB |
687 | /* preferred - Hard-code to 4096 for now. |
688 | * TODO: is blk_bs(blk)->bl.opt_transfer appropriate? */ | |
b0245d64 | 689 | sizes[1] = MAX(4096, sizes[0]); |
0c1d50bd | 690 | /* maximum - At most 32M, but smaller as appropriate. */ |
37a4f70c | 691 | sizes[2] = MIN(blk_get_max_transfer(exp->common.blk), NBD_MAX_BUFFER_SIZE); |
0c1d50bd | 692 | trace_nbd_negotiate_handle_info_block_size(sizes[0], sizes[1], sizes[2]); |
80c7c2b0 PM |
693 | sizes[0] = cpu_to_be32(sizes[0]); |
694 | sizes[1] = cpu_to_be32(sizes[1]); | |
695 | sizes[2] = cpu_to_be32(sizes[2]); | |
0cfae925 | 696 | rc = nbd_negotiate_send_info(client, NBD_INFO_BLOCK_SIZE, |
0c1d50bd EB |
697 | sizeof(sizes), sizes, errp); |
698 | if (rc < 0) { | |
699 | return rc; | |
700 | } | |
701 | ||
f37708f6 | 702 | /* Send NBD_INFO_EXPORT always */ |
dbb38caa | 703 | myflags = exp->nbdflags; |
ac132d05 | 704 | if (client->mode >= NBD_MODE_STRUCTURED) { |
dbb38caa EB |
705 | myflags |= NBD_FLAG_SEND_DF; |
706 | } | |
2dcbb11b EB |
707 | if (client->mode >= NBD_MODE_EXTENDED && |
708 | (client->contexts.count || client->opt == NBD_OPT_INFO)) { | |
709 | myflags |= NBD_FLAG_BLOCK_STAT_PAYLOAD; | |
710 | } | |
dbb38caa | 711 | trace_nbd_negotiate_new_style_size_flags(exp->size, myflags); |
f37708f6 | 712 | stq_be_p(buf, exp->size); |
dbb38caa | 713 | stw_be_p(buf + 8, myflags); |
0cfae925 | 714 | rc = nbd_negotiate_send_info(client, NBD_INFO_EXPORT, |
f37708f6 EB |
715 | sizeof(buf), buf, errp); |
716 | if (rc < 0) { | |
717 | return rc; | |
718 | } | |
719 | ||
099fbcd6 EB |
720 | /* |
721 | * If the client is just asking for NBD_OPT_INFO, but forgot to | |
722 | * request block sizes in a situation that would impact | |
723 | * performance, then return an error. But for NBD_OPT_GO, we | |
724 | * tolerate all clients, regardless of alignments. | |
725 | */ | |
726 | if (client->opt == NBD_OPT_INFO && !blocksize && | |
37a4f70c | 727 | blk_get_request_alignment(exp->common.blk) > 1) { |
0cfae925 VSO |
728 | return nbd_negotiate_send_rep_err(client, |
729 | NBD_REP_ERR_BLOCK_SIZE_REQD, | |
0c1d50bd EB |
730 | errp, |
731 | "request NBD_INFO_BLOCK_SIZE to " | |
732 | "use this export"); | |
733 | } | |
734 | ||
f37708f6 | 735 | /* Final reply */ |
0cfae925 | 736 | rc = nbd_negotiate_send_rep(client, NBD_REP_ACK, errp); |
f37708f6 EB |
737 | if (rc < 0) { |
738 | return rc; | |
739 | } | |
740 | ||
0cfae925 | 741 | if (client->opt == NBD_OPT_GO) { |
f37708f6 | 742 | client->exp = exp; |
6e280648 | 743 | client->check_align = check_align; |
f37708f6 | 744 | QTAILQ_INSERT_TAIL(&client->exp->clients, client, next); |
c69de1be | 745 | blk_exp_ref(&client->exp->common); |
f37708f6 EB |
746 | rc = 1; |
747 | } | |
748 | return rc; | |
f37708f6 EB |
749 | } |
750 | ||
751 | ||
36683283 EB |
752 | /* Handle NBD_OPT_STARTTLS. Return NULL to drop connection, or else the |
753 | * new channel for all further (now-encrypted) communication. */ | |
f95910fe | 754 | static QIOChannel *nbd_negotiate_handle_starttls(NBDClient *client, |
2fd2c840 | 755 | Error **errp) |
f95910fe DB |
756 | { |
757 | QIOChannel *ioc; | |
758 | QIOChannelTLS *tioc; | |
759 | struct NBDTLSHandshakeData data = { 0 }; | |
760 | ||
0cfae925 VSO |
761 | assert(client->opt == NBD_OPT_STARTTLS); |
762 | ||
9588463e | 763 | trace_nbd_negotiate_handle_starttls(); |
f95910fe | 764 | ioc = client->ioc; |
f95910fe | 765 | |
0cfae925 | 766 | if (nbd_negotiate_send_rep(client, NBD_REP_ACK, errp) < 0) { |
63d5ef86 EB |
767 | return NULL; |
768 | } | |
f95910fe DB |
769 | |
770 | tioc = qio_channel_tls_new_server(ioc, | |
771 | client->tlscreds, | |
b25e12da | 772 | client->tlsauthz, |
2fd2c840 | 773 | errp); |
f95910fe DB |
774 | if (!tioc) { |
775 | return NULL; | |
776 | } | |
777 | ||
0d73f725 | 778 | qio_channel_set_name(QIO_CHANNEL(tioc), "nbd-server-tls"); |
9588463e | 779 | trace_nbd_negotiate_handle_starttls_handshake(); |
f95910fe DB |
780 | data.loop = g_main_loop_new(g_main_context_default(), FALSE); |
781 | qio_channel_tls_handshake(tioc, | |
782 | nbd_tls_handshake, | |
783 | &data, | |
1939ccda | 784 | NULL, |
f95910fe DB |
785 | NULL); |
786 | ||
787 | if (!data.complete) { | |
788 | g_main_loop_run(data.loop); | |
789 | } | |
790 | g_main_loop_unref(data.loop); | |
791 | if (data.error) { | |
792 | object_unref(OBJECT(tioc)); | |
2fd2c840 | 793 | error_propagate(errp, data.error); |
f95910fe DB |
794 | return NULL; |
795 | } | |
796 | ||
797 | return QIO_CHANNEL(tioc); | |
798 | } | |
799 | ||
e7b1948d VSO |
800 | /* nbd_negotiate_send_meta_context |
801 | * | |
802 | * Send one chunk of reply to NBD_OPT_{LIST,SET}_META_CONTEXT | |
803 | * | |
804 | * For NBD_OPT_LIST_META_CONTEXT @context_id is ignored, 0 is used instead. | |
805 | */ | |
806 | static int nbd_negotiate_send_meta_context(NBDClient *client, | |
807 | const char *context, | |
808 | uint32_t context_id, | |
809 | Error **errp) | |
810 | { | |
811 | NBDOptionReplyMetaContext opt; | |
812 | struct iovec iov[] = { | |
813 | {.iov_base = &opt, .iov_len = sizeof(opt)}, | |
814 | {.iov_base = (void *)context, .iov_len = strlen(context)} | |
815 | }; | |
816 | ||
93676c88 | 817 | assert(iov[1].iov_len <= NBD_MAX_STRING_SIZE); |
e7b1948d VSO |
818 | if (client->opt == NBD_OPT_LIST_META_CONTEXT) { |
819 | context_id = 0; | |
820 | } | |
821 | ||
2b53af25 | 822 | trace_nbd_negotiate_meta_query_reply(context, context_id); |
e7b1948d VSO |
823 | set_be_option_rep(&opt.h, client->opt, NBD_REP_META_CONTEXT, |
824 | sizeof(opt) - sizeof(opt.h) + iov[1].iov_len); | |
825 | stl_be_p(&opt.context_id, context_id); | |
826 | ||
827 | return qio_channel_writev_all(client->ioc, iov, 2, errp) < 0 ? -EIO : 0; | |
828 | } | |
829 | ||
ebd57062 EB |
830 | /* |
831 | * Return true if @query matches @pattern, or if @query is empty when | |
832 | * the @client is performing _LIST_. | |
dbb8b396 | 833 | */ |
ebd57062 EB |
834 | static bool nbd_meta_empty_or_pattern(NBDClient *client, const char *pattern, |
835 | const char *query) | |
e7b1948d | 836 | { |
ebd57062 EB |
837 | if (!*query) { |
838 | trace_nbd_negotiate_meta_query_parse("empty"); | |
839 | return client->opt == NBD_OPT_LIST_META_CONTEXT; | |
e7b1948d | 840 | } |
ebd57062 | 841 | if (strcmp(query, pattern) == 0) { |
b0769d8f | 842 | trace_nbd_negotiate_meta_query_parse(pattern); |
ebd57062 | 843 | return true; |
e7b1948d | 844 | } |
ebd57062 EB |
845 | trace_nbd_negotiate_meta_query_skip("pattern not matched"); |
846 | return false; | |
e7b1948d VSO |
847 | } |
848 | ||
b0769d8f | 849 | /* |
ebd57062 | 850 | * Return true and adjust @str in place if it begins with @prefix. |
b0769d8f | 851 | */ |
ebd57062 | 852 | static bool nbd_strshift(const char **str, const char *prefix) |
b0769d8f | 853 | { |
ebd57062 | 854 | size_t len = strlen(prefix); |
b0769d8f | 855 | |
ebd57062 EB |
856 | if (strncmp(*str, prefix, len) == 0) { |
857 | *str += len; | |
858 | return true; | |
b0769d8f | 859 | } |
ebd57062 | 860 | return false; |
b0769d8f VSO |
861 | } |
862 | ||
863 | /* nbd_meta_base_query | |
864 | * | |
865 | * Handle queries to 'base' namespace. For now, only the base:allocation | |
ebd57062 | 866 | * context is available. Return true if @query has been handled. |
b0769d8f | 867 | */ |
fd358d83 | 868 | static bool nbd_meta_base_query(NBDClient *client, NBDMetaContexts *meta, |
ebd57062 | 869 | const char *query) |
b0769d8f | 870 | { |
ebd57062 EB |
871 | if (!nbd_strshift(&query, "base:")) { |
872 | return false; | |
873 | } | |
874 | trace_nbd_negotiate_meta_query_parse("base:"); | |
875 | ||
876 | if (nbd_meta_empty_or_pattern(client, "allocation", query)) { | |
877 | meta->base_allocation = true; | |
878 | } | |
879 | return true; | |
b0769d8f VSO |
880 | } |
881 | ||
ebd57062 | 882 | /* nbd_meta_qemu_query |
3d068aff | 883 | * |
ebd57062 | 884 | * Handle queries to 'qemu' namespace. For now, only the qemu:dirty-bitmap: |
71719cd5 EB |
885 | * and qemu:allocation-depth contexts are available. Return true if @query |
886 | * has been handled. | |
ebd57062 | 887 | */ |
fd358d83 | 888 | static bool nbd_meta_qemu_query(NBDClient *client, NBDMetaContexts *meta, |
ebd57062 | 889 | const char *query) |
3d068aff | 890 | { |
3b1f244c EB |
891 | size_t i; |
892 | ||
ebd57062 EB |
893 | if (!nbd_strshift(&query, "qemu:")) { |
894 | return false; | |
3d068aff | 895 | } |
ebd57062 | 896 | trace_nbd_negotiate_meta_query_parse("qemu:"); |
3d068aff | 897 | |
ebd57062 | 898 | if (!*query) { |
3d068aff | 899 | if (client->opt == NBD_OPT_LIST_META_CONTEXT) { |
71719cd5 | 900 | meta->allocation_depth = meta->exp->allocation_depth; |
76df2b8d EB |
901 | if (meta->exp->nr_export_bitmaps) { |
902 | memset(meta->bitmaps, 1, meta->exp->nr_export_bitmaps); | |
903 | } | |
3d068aff VSO |
904 | } |
905 | trace_nbd_negotiate_meta_query_parse("empty"); | |
ebd57062 | 906 | return true; |
3d068aff VSO |
907 | } |
908 | ||
71719cd5 EB |
909 | if (strcmp(query, "allocation-depth") == 0) { |
910 | trace_nbd_negotiate_meta_query_parse("allocation-depth"); | |
911 | meta->allocation_depth = meta->exp->allocation_depth; | |
912 | return true; | |
913 | } | |
914 | ||
ebd57062 EB |
915 | if (nbd_strshift(&query, "dirty-bitmap:")) { |
916 | trace_nbd_negotiate_meta_query_parse("dirty-bitmap:"); | |
3b1f244c | 917 | if (!*query) { |
76df2b8d EB |
918 | if (client->opt == NBD_OPT_LIST_META_CONTEXT && |
919 | meta->exp->nr_export_bitmaps) { | |
3b1f244c EB |
920 | memset(meta->bitmaps, 1, meta->exp->nr_export_bitmaps); |
921 | } | |
922 | trace_nbd_negotiate_meta_query_parse("empty"); | |
ebd57062 EB |
923 | return true; |
924 | } | |
3b1f244c EB |
925 | |
926 | for (i = 0; i < meta->exp->nr_export_bitmaps; i++) { | |
927 | const char *bm_name; | |
928 | ||
929 | bm_name = bdrv_dirty_bitmap_name(meta->exp->export_bitmaps[i]); | |
930 | if (strcmp(bm_name, query) == 0) { | |
931 | meta->bitmaps[i] = true; | |
932 | trace_nbd_negotiate_meta_query_parse(query); | |
933 | return true; | |
934 | } | |
ebd57062 | 935 | } |
3b1f244c | 936 | trace_nbd_negotiate_meta_query_skip("no dirty-bitmap match"); |
ebd57062 | 937 | return true; |
3d068aff VSO |
938 | } |
939 | ||
71719cd5 | 940 | trace_nbd_negotiate_meta_query_skip("unknown qemu context"); |
ebd57062 | 941 | return true; |
3d068aff VSO |
942 | } |
943 | ||
e7b1948d VSO |
944 | /* nbd_negotiate_meta_query |
945 | * | |
946 | * Parse namespace name and call corresponding function to parse body of the | |
947 | * query. | |
948 | * | |
93676c88 | 949 | * The only supported namespaces are 'base' and 'qemu'. |
e7b1948d | 950 | * |
e7b1948d VSO |
951 | * Return -errno on I/O error, 0 if option was completely handled by |
952 | * sending a reply about inconsistent lengths, or 1 on success. */ | |
953 | static int nbd_negotiate_meta_query(NBDClient *client, | |
fd358d83 | 954 | NBDMetaContexts *meta, Error **errp) |
e7b1948d VSO |
955 | { |
956 | int ret; | |
ebd57062 | 957 | g_autofree char *query = NULL; |
e7b1948d VSO |
958 | uint32_t len; |
959 | ||
d1e2c3e7 | 960 | ret = nbd_opt_read(client, &len, sizeof(len), false, errp); |
e7b1948d VSO |
961 | if (ret <= 0) { |
962 | return ret; | |
963 | } | |
80c7c2b0 | 964 | len = cpu_to_be32(len); |
e7b1948d | 965 | |
93676c88 EB |
966 | if (len > NBD_MAX_STRING_SIZE) { |
967 | trace_nbd_negotiate_meta_query_skip("length too long"); | |
968 | return nbd_opt_skip(client, len, errp); | |
969 | } | |
e7b1948d | 970 | |
ebd57062 EB |
971 | query = g_malloc(len + 1); |
972 | ret = nbd_opt_read(client, query, len, true, errp); | |
e7b1948d VSO |
973 | if (ret <= 0) { |
974 | return ret; | |
975 | } | |
ebd57062 | 976 | query[len] = '\0'; |
3d068aff | 977 | |
ebd57062 EB |
978 | if (nbd_meta_base_query(client, meta, query)) { |
979 | return 1; | |
980 | } | |
981 | if (nbd_meta_qemu_query(client, meta, query)) { | |
982 | return 1; | |
e7b1948d VSO |
983 | } |
984 | ||
3d068aff | 985 | trace_nbd_negotiate_meta_query_skip("unknown namespace"); |
ebd57062 | 986 | return 1; |
e7b1948d VSO |
987 | } |
988 | ||
989 | /* nbd_negotiate_meta_queries | |
990 | * Handle NBD_OPT_LIST_META_CONTEXT and NBD_OPT_SET_META_CONTEXT | |
991 | * | |
992 | * Return -errno on I/O error, or 0 if option was completely handled. */ | |
fd358d83 | 993 | static int nbd_negotiate_meta_queries(NBDClient *client, Error **errp) |
e7b1948d VSO |
994 | { |
995 | int ret; | |
9d7ab222 | 996 | g_autofree char *export_name = NULL; |
cd1675f8 RH |
997 | /* Mark unused to work around https://bugs.llvm.org/show_bug.cgi?id=3888 */ |
998 | g_autofree G_GNUC_UNUSED bool *bitmaps = NULL; | |
fd358d83 EB |
999 | NBDMetaContexts local_meta = {0}; |
1000 | NBDMetaContexts *meta; | |
e7b1948d | 1001 | uint32_t nb_queries; |
3b1f244c | 1002 | size_t i; |
47ec485e | 1003 | size_t count = 0; |
e7b1948d | 1004 | |
ac132d05 EB |
1005 | if (client->opt == NBD_OPT_SET_META_CONTEXT && |
1006 | client->mode < NBD_MODE_STRUCTURED) { | |
e7b1948d VSO |
1007 | return nbd_opt_invalid(client, errp, |
1008 | "request option '%s' when structured reply " | |
1009 | "is not negotiated", | |
1010 | nbd_opt_lookup(client->opt)); | |
1011 | } | |
1012 | ||
1013 | if (client->opt == NBD_OPT_LIST_META_CONTEXT) { | |
1014 | /* Only change the caller's meta on SET. */ | |
1015 | meta = &local_meta; | |
fd358d83 EB |
1016 | } else { |
1017 | meta = &client->contexts; | |
e7b1948d VSO |
1018 | } |
1019 | ||
3b1f244c | 1020 | g_free(meta->bitmaps); |
e7b1948d VSO |
1021 | memset(meta, 0, sizeof(*meta)); |
1022 | ||
9d7ab222 | 1023 | ret = nbd_opt_read_name(client, &export_name, NULL, errp); |
e7b1948d VSO |
1024 | if (ret <= 0) { |
1025 | return ret; | |
1026 | } | |
1027 | ||
af736e54 VSO |
1028 | meta->exp = nbd_export_find(export_name); |
1029 | if (meta->exp == NULL) { | |
5c4fe018 EB |
1030 | g_autofree char *sane_name = nbd_sanitize_name(export_name); |
1031 | ||
e7b1948d | 1032 | return nbd_opt_drop(client, NBD_REP_ERR_UNKNOWN, errp, |
5c4fe018 | 1033 | "export '%s' not present", sane_name); |
e7b1948d | 1034 | } |
3b1f244c EB |
1035 | meta->bitmaps = g_new0(bool, meta->exp->nr_export_bitmaps); |
1036 | if (client->opt == NBD_OPT_LIST_META_CONTEXT) { | |
1037 | bitmaps = meta->bitmaps; | |
1038 | } | |
e7b1948d | 1039 | |
d1e2c3e7 | 1040 | ret = nbd_opt_read(client, &nb_queries, sizeof(nb_queries), false, errp); |
e7b1948d VSO |
1041 | if (ret <= 0) { |
1042 | return ret; | |
1043 | } | |
80c7c2b0 | 1044 | nb_queries = cpu_to_be32(nb_queries); |
2b53af25 | 1045 | trace_nbd_negotiate_meta_context(nbd_opt_lookup(client->opt), |
af736e54 | 1046 | export_name, nb_queries); |
e7b1948d VSO |
1047 | |
1048 | if (client->opt == NBD_OPT_LIST_META_CONTEXT && !nb_queries) { | |
1049 | /* enable all known contexts */ | |
1050 | meta->base_allocation = true; | |
71719cd5 | 1051 | meta->allocation_depth = meta->exp->allocation_depth; |
76df2b8d EB |
1052 | if (meta->exp->nr_export_bitmaps) { |
1053 | memset(meta->bitmaps, 1, meta->exp->nr_export_bitmaps); | |
1054 | } | |
e7b1948d VSO |
1055 | } else { |
1056 | for (i = 0; i < nb_queries; ++i) { | |
1057 | ret = nbd_negotiate_meta_query(client, meta, errp); | |
1058 | if (ret <= 0) { | |
1059 | return ret; | |
1060 | } | |
1061 | } | |
1062 | } | |
1063 | ||
1064 | if (meta->base_allocation) { | |
1065 | ret = nbd_negotiate_send_meta_context(client, "base:allocation", | |
1066 | NBD_META_ID_BASE_ALLOCATION, | |
1067 | errp); | |
1068 | if (ret < 0) { | |
1069 | return ret; | |
1070 | } | |
47ec485e | 1071 | count++; |
e7b1948d VSO |
1072 | } |
1073 | ||
71719cd5 EB |
1074 | if (meta->allocation_depth) { |
1075 | ret = nbd_negotiate_send_meta_context(client, "qemu:allocation-depth", | |
1076 | NBD_META_ID_ALLOCATION_DEPTH, | |
1077 | errp); | |
1078 | if (ret < 0) { | |
1079 | return ret; | |
1080 | } | |
1081 | count++; | |
1082 | } | |
1083 | ||
3b1f244c EB |
1084 | for (i = 0; i < meta->exp->nr_export_bitmaps; i++) { |
1085 | const char *bm_name; | |
1086 | g_autofree char *context = NULL; | |
1087 | ||
1088 | if (!meta->bitmaps[i]) { | |
1089 | continue; | |
1090 | } | |
1091 | ||
1092 | bm_name = bdrv_dirty_bitmap_name(meta->exp->export_bitmaps[i]); | |
1093 | context = g_strdup_printf("qemu:dirty-bitmap:%s", bm_name); | |
02e87e3b EB |
1094 | |
1095 | ret = nbd_negotiate_send_meta_context(client, context, | |
3b1f244c | 1096 | NBD_META_ID_DIRTY_BITMAP + i, |
3d068aff VSO |
1097 | errp); |
1098 | if (ret < 0) { | |
1099 | return ret; | |
1100 | } | |
47ec485e | 1101 | count++; |
3d068aff VSO |
1102 | } |
1103 | ||
e7b1948d VSO |
1104 | ret = nbd_negotiate_send_rep(client, NBD_REP_ACK, errp); |
1105 | if (ret == 0) { | |
47ec485e | 1106 | meta->count = count; |
e7b1948d VSO |
1107 | } |
1108 | ||
1109 | return ret; | |
1110 | } | |
1111 | ||
1e120ffe | 1112 | /* nbd_negotiate_options |
f37708f6 EB |
1113 | * Process all NBD_OPT_* client option commands, during fixed newstyle |
1114 | * negotiation. | |
1e120ffe | 1115 | * Return: |
2fd2c840 VSO |
1116 | * -errno on error, errp is set |
1117 | * 0 on successful negotiation, errp is not set | |
1118 | * 1 if client sent NBD_OPT_ABORT, i.e. on valid disconnect, | |
1119 | * errp is not set | |
1e120ffe | 1120 | */ |
dbb38caa | 1121 | static int nbd_negotiate_options(NBDClient *client, Error **errp) |
f5076b5a | 1122 | { |
9c122ada | 1123 | uint32_t flags; |
26afa868 | 1124 | bool fixedNewstyle = false; |
23e099c3 | 1125 | bool no_zeroes = false; |
9c122ada HR |
1126 | |
1127 | /* Client sends: | |
1128 | [ 0 .. 3] client flags | |
1129 | ||
f37708f6 | 1130 | Then we loop until NBD_OPT_EXPORT_NAME or NBD_OPT_GO: |
9c122ada HR |
1131 | [ 0 .. 7] NBD_OPTS_MAGIC |
1132 | [ 8 .. 11] NBD option | |
1133 | [12 .. 15] Data length | |
1134 | ... Rest of request | |
1135 | ||
1136 | [ 0 .. 7] NBD_OPTS_MAGIC | |
1137 | [ 8 .. 11] Second NBD option | |
1138 | [12 .. 15] Data length | |
1139 | ... Rest of request | |
1140 | */ | |
1141 | ||
e6798f06 | 1142 | if (nbd_read32(client->ioc, &flags, "flags", errp) < 0) { |
9c122ada HR |
1143 | return -EIO; |
1144 | } | |
ac132d05 | 1145 | client->mode = NBD_MODE_EXPORT_NAME; |
621c4f4e | 1146 | trace_nbd_negotiate_options_flags(flags); |
26afa868 | 1147 | if (flags & NBD_FLAG_C_FIXED_NEWSTYLE) { |
26afa868 DB |
1148 | fixedNewstyle = true; |
1149 | flags &= ~NBD_FLAG_C_FIXED_NEWSTYLE; | |
ac132d05 | 1150 | client->mode = NBD_MODE_SIMPLE; |
26afa868 | 1151 | } |
c203c59a | 1152 | if (flags & NBD_FLAG_C_NO_ZEROES) { |
23e099c3 | 1153 | no_zeroes = true; |
c203c59a EB |
1154 | flags &= ~NBD_FLAG_C_NO_ZEROES; |
1155 | } | |
26afa868 | 1156 | if (flags != 0) { |
2fd2c840 | 1157 | error_setg(errp, "Unknown client flags 0x%" PRIx32 " received", flags); |
621c4f4e | 1158 | return -EINVAL; |
9c122ada HR |
1159 | } |
1160 | ||
f5076b5a | 1161 | while (1) { |
9c122ada | 1162 | int ret; |
7f9039cd | 1163 | uint32_t option, length; |
f5076b5a HB |
1164 | uint64_t magic; |
1165 | ||
e6798f06 | 1166 | if (nbd_read64(client->ioc, &magic, "opts magic", errp) < 0) { |
f5076b5a HB |
1167 | return -EINVAL; |
1168 | } | |
9588463e VSO |
1169 | trace_nbd_negotiate_options_check_magic(magic); |
1170 | if (magic != NBD_OPTS_MAGIC) { | |
2fd2c840 | 1171 | error_setg(errp, "Bad magic received"); |
f5076b5a HB |
1172 | return -EINVAL; |
1173 | } | |
1174 | ||
e6798f06 | 1175 | if (nbd_read32(client->ioc, &option, "option", errp) < 0) { |
f5076b5a HB |
1176 | return -EINVAL; |
1177 | } | |
0cfae925 | 1178 | client->opt = option; |
f5076b5a | 1179 | |
e6798f06 | 1180 | if (nbd_read32(client->ioc, &length, "option length", errp) < 0) { |
f5076b5a HB |
1181 | return -EINVAL; |
1182 | } | |
894e0280 | 1183 | assert(!client->optlen); |
0cfae925 | 1184 | client->optlen = length; |
f5076b5a | 1185 | |
fdad35ef | 1186 | if (length > NBD_MAX_BUFFER_SIZE) { |
b2578459 | 1187 | error_setg(errp, "len (%" PRIu32 ") is larger than max len (%u)", |
fdad35ef EB |
1188 | length, NBD_MAX_BUFFER_SIZE); |
1189 | return -EINVAL; | |
1190 | } | |
1191 | ||
3736cc5b EB |
1192 | trace_nbd_negotiate_options_check_option(option, |
1193 | nbd_opt_lookup(option)); | |
f95910fe DB |
1194 | if (client->tlscreds && |
1195 | client->ioc == (QIOChannel *)client->sioc) { | |
1196 | QIOChannel *tioc; | |
1197 | if (!fixedNewstyle) { | |
7f9039cd | 1198 | error_setg(errp, "Unsupported option 0x%" PRIx32, option); |
f95910fe DB |
1199 | return -EINVAL; |
1200 | } | |
7f9039cd | 1201 | switch (option) { |
f95910fe | 1202 | case NBD_OPT_STARTTLS: |
e68c35cf EB |
1203 | if (length) { |
1204 | /* Unconditionally drop the connection if the client | |
1205 | * can't start a TLS negotiation correctly */ | |
0cfae925 | 1206 | return nbd_reject_length(client, true, errp); |
e68c35cf EB |
1207 | } |
1208 | tioc = nbd_negotiate_handle_starttls(client, errp); | |
f95910fe DB |
1209 | if (!tioc) { |
1210 | return -EIO; | |
1211 | } | |
8cbee49e | 1212 | ret = 0; |
f95910fe | 1213 | object_unref(OBJECT(client->ioc)); |
7d5b0d68 | 1214 | client->ioc = tioc; |
f95910fe DB |
1215 | break; |
1216 | ||
d1129a8a EB |
1217 | case NBD_OPT_EXPORT_NAME: |
1218 | /* No way to return an error to client, so drop connection */ | |
2fd2c840 | 1219 | error_setg(errp, "Option 0x%x not permitted before TLS", |
7f9039cd | 1220 | option); |
d1129a8a EB |
1221 | return -EINVAL; |
1222 | ||
f95910fe | 1223 | default: |
3e99ebb9 EB |
1224 | /* Let the client keep trying, unless they asked to |
1225 | * quit. Always try to give an error back to the | |
1226 | * client; but when replying to OPT_ABORT, be aware | |
1227 | * that the client may hang up before receiving the | |
1228 | * error, in which case we are fine ignoring the | |
1229 | * resulting EPIPE. */ | |
1230 | ret = nbd_opt_drop(client, NBD_REP_ERR_TLS_REQD, | |
1231 | option == NBD_OPT_ABORT ? NULL : errp, | |
894e0280 | 1232 | "Option 0x%" PRIx32 |
0b0bb124 | 1233 | " not permitted before TLS", option); |
7f9039cd | 1234 | if (option == NBD_OPT_ABORT) { |
1e120ffe | 1235 | return 1; |
b6f5d3b5 | 1236 | } |
d1129a8a | 1237 | break; |
f95910fe DB |
1238 | } |
1239 | } else if (fixedNewstyle) { | |
7f9039cd | 1240 | switch (option) { |
26afa868 | 1241 | case NBD_OPT_LIST: |
e68c35cf | 1242 | if (length) { |
0cfae925 | 1243 | ret = nbd_reject_length(client, false, errp); |
e68c35cf EB |
1244 | } else { |
1245 | ret = nbd_negotiate_handle_list(client, errp); | |
1246 | } | |
26afa868 DB |
1247 | break; |
1248 | ||
1249 | case NBD_OPT_ABORT: | |
b6f5d3b5 EB |
1250 | /* NBD spec says we must try to reply before |
1251 | * disconnecting, but that we must also tolerate | |
1252 | * guests that don't wait for our reply. */ | |
0cfae925 | 1253 | nbd_negotiate_send_rep(client, NBD_REP_ACK, NULL); |
1e120ffe | 1254 | return 1; |
26afa868 DB |
1255 | |
1256 | case NBD_OPT_EXPORT_NAME: | |
dbb38caa | 1257 | return nbd_negotiate_handle_export_name(client, no_zeroes, |
23e099c3 | 1258 | errp); |
26afa868 | 1259 | |
f37708f6 EB |
1260 | case NBD_OPT_INFO: |
1261 | case NBD_OPT_GO: | |
dbb38caa | 1262 | ret = nbd_negotiate_handle_info(client, errp); |
f37708f6 EB |
1263 | if (ret == 1) { |
1264 | assert(option == NBD_OPT_GO); | |
1265 | return 0; | |
1266 | } | |
f37708f6 EB |
1267 | break; |
1268 | ||
f95910fe | 1269 | case NBD_OPT_STARTTLS: |
e68c35cf | 1270 | if (length) { |
0cfae925 | 1271 | ret = nbd_reject_length(client, false, errp); |
e68c35cf | 1272 | } else if (client->tlscreds) { |
0cfae925 VSO |
1273 | ret = nbd_negotiate_send_rep_err(client, |
1274 | NBD_REP_ERR_INVALID, errp, | |
36683283 | 1275 | "TLS already enabled"); |
f95910fe | 1276 | } else { |
0cfae925 VSO |
1277 | ret = nbd_negotiate_send_rep_err(client, |
1278 | NBD_REP_ERR_POLICY, errp, | |
36683283 | 1279 | "TLS not configured"); |
63d5ef86 | 1280 | } |
d1129a8a | 1281 | break; |
5c54e7fa VSO |
1282 | |
1283 | case NBD_OPT_STRUCTURED_REPLY: | |
1284 | if (length) { | |
0cfae925 | 1285 | ret = nbd_reject_length(client, false, errp); |
9c1d2614 EB |
1286 | } else if (client->mode >= NBD_MODE_EXTENDED) { |
1287 | ret = nbd_negotiate_send_rep_err( | |
1288 | client, NBD_REP_ERR_EXT_HEADER_REQD, errp, | |
1289 | "extended headers already negotiated"); | |
ac132d05 | 1290 | } else if (client->mode >= NBD_MODE_STRUCTURED) { |
5c54e7fa | 1291 | ret = nbd_negotiate_send_rep_err( |
0cfae925 | 1292 | client, NBD_REP_ERR_INVALID, errp, |
5c54e7fa VSO |
1293 | "structured reply already negotiated"); |
1294 | } else { | |
0cfae925 | 1295 | ret = nbd_negotiate_send_rep(client, NBD_REP_ACK, errp); |
ac132d05 | 1296 | client->mode = NBD_MODE_STRUCTURED; |
5c54e7fa VSO |
1297 | } |
1298 | break; | |
1299 | ||
e7b1948d VSO |
1300 | case NBD_OPT_LIST_META_CONTEXT: |
1301 | case NBD_OPT_SET_META_CONTEXT: | |
fd358d83 | 1302 | ret = nbd_negotiate_meta_queries(client, errp); |
e7b1948d VSO |
1303 | break; |
1304 | ||
9c1d2614 EB |
1305 | case NBD_OPT_EXTENDED_HEADERS: |
1306 | if (length) { | |
1307 | ret = nbd_reject_length(client, false, errp); | |
1308 | } else if (client->mode >= NBD_MODE_EXTENDED) { | |
1309 | ret = nbd_negotiate_send_rep_err( | |
1310 | client, NBD_REP_ERR_INVALID, errp, | |
1311 | "extended headers already negotiated"); | |
1312 | } else { | |
1313 | ret = nbd_negotiate_send_rep(client, NBD_REP_ACK, errp); | |
1314 | client->mode = NBD_MODE_EXTENDED; | |
1315 | } | |
1316 | break; | |
1317 | ||
26afa868 | 1318 | default: |
894e0280 | 1319 | ret = nbd_opt_drop(client, NBD_REP_ERR_UNSUP, errp, |
28fb494f | 1320 | "Unsupported option %" PRIu32 " (%s)", |
894e0280 | 1321 | option, nbd_opt_lookup(option)); |
156f6a10 | 1322 | break; |
26afa868 DB |
1323 | } |
1324 | } else { | |
1325 | /* | |
1326 | * If broken new-style we should drop the connection | |
1327 | * for anything except NBD_OPT_EXPORT_NAME | |
1328 | */ | |
7f9039cd | 1329 | switch (option) { |
26afa868 | 1330 | case NBD_OPT_EXPORT_NAME: |
dbb38caa | 1331 | return nbd_negotiate_handle_export_name(client, no_zeroes, |
23e099c3 | 1332 | errp); |
26afa868 DB |
1333 | |
1334 | default: | |
28fb494f | 1335 | error_setg(errp, "Unsupported option %" PRIu32 " (%s)", |
3736cc5b | 1336 | option, nbd_opt_lookup(option)); |
26afa868 | 1337 | return -EINVAL; |
32d7d2e0 | 1338 | } |
f5076b5a | 1339 | } |
8cbee49e EB |
1340 | if (ret < 0) { |
1341 | return ret; | |
1342 | } | |
f5076b5a HB |
1343 | } |
1344 | } | |
1345 | ||
1e120ffe VSO |
1346 | /* nbd_negotiate |
1347 | * Return: | |
2fd2c840 VSO |
1348 | * -errno on error, errp is set |
1349 | * 0 on successful negotiation, errp is not set | |
1350 | * 1 if client sent NBD_OPT_ABORT, i.e. on valid disconnect, | |
1351 | * errp is not set | |
1e120ffe | 1352 | */ |
2fd2c840 | 1353 | static coroutine_fn int nbd_negotiate(NBDClient *client, Error **errp) |
7a5ca864 | 1354 | { |
795d946d | 1355 | ERRP_GUARD(); |
5f66d060 | 1356 | char buf[NBD_OLDSTYLE_NEGOTIATE_SIZE] = ""; |
2e5c9ad6 | 1357 | int ret; |
b2e3d87f | 1358 | |
5f66d060 | 1359 | /* Old style negotiation header, no room for options |
6b8c01e7 PB |
1360 | [ 0 .. 7] passwd ("NBDMAGIC") |
1361 | [ 8 .. 15] magic (NBD_CLIENT_MAGIC) | |
b2e3d87f | 1362 | [16 .. 23] size |
5f66d060 | 1363 | [24 .. 27] export flags (zero-extended) |
6b8c01e7 PB |
1364 | [28 .. 151] reserved (0) |
1365 | ||
5f66d060 | 1366 | New style negotiation header, client can send options |
6b8c01e7 PB |
1367 | [ 0 .. 7] passwd ("NBDMAGIC") |
1368 | [ 8 .. 15] magic (NBD_OPTS_MAGIC) | |
1369 | [16 .. 17] server flags (0) | |
f37708f6 | 1370 | ....options sent, ending in NBD_OPT_EXPORT_NAME or NBD_OPT_GO.... |
b2e3d87f NT |
1371 | */ |
1372 | ||
1c778ef7 | 1373 | qio_channel_set_blocking(client->ioc, false, NULL); |
06e0f098 | 1374 | qio_channel_set_follow_coroutine_ctx(client->ioc, true); |
185b4338 | 1375 | |
9588463e | 1376 | trace_nbd_negotiate_begin(); |
b2e3d87f | 1377 | memcpy(buf, "NBDMAGIC", 8); |
f95910fe | 1378 | |
7f7dfe2a VSO |
1379 | stq_be_p(buf + 8, NBD_OPTS_MAGIC); |
1380 | stw_be_p(buf + 16, NBD_FLAG_FIXED_NEWSTYLE | NBD_FLAG_NO_ZEROES); | |
b2e3d87f | 1381 | |
7f7dfe2a VSO |
1382 | if (nbd_write(client->ioc, buf, 18, errp) < 0) { |
1383 | error_prepend(errp, "write failed: "); | |
1384 | return -EINVAL; | |
1385 | } | |
dbb38caa | 1386 | ret = nbd_negotiate_options(client, errp); |
7f7dfe2a VSO |
1387 | if (ret != 0) { |
1388 | if (ret < 0) { | |
1389 | error_prepend(errp, "option negotiation failed: "); | |
6b8c01e7 | 1390 | } |
7f7dfe2a | 1391 | return ret; |
b2e3d87f NT |
1392 | } |
1393 | ||
0cfae925 | 1394 | assert(!client->optlen); |
9588463e | 1395 | trace_nbd_negotiate_success(); |
d9faeed8 VSO |
1396 | |
1397 | return 0; | |
7a5ca864 FB |
1398 | } |
1399 | ||
f148ae7d SL |
1400 | /* nbd_read_eof |
1401 | * Tries to read @size bytes from @ioc. This is a local implementation of | |
1402 | * qio_channel_readv_all_eof. We have it here because we need it to be | |
1403 | * interruptible and to know when the coroutine is yielding. | |
1404 | * Returns 1 on success | |
1405 | * 0 on eof, when no data was read (errp is not set) | |
1406 | * negative errno on failure (errp is set) | |
1407 | */ | |
1408 | static inline int coroutine_fn | |
1409 | nbd_read_eof(NBDClient *client, void *buffer, size_t size, Error **errp) | |
1410 | { | |
1411 | bool partial = false; | |
1412 | ||
1413 | assert(size); | |
1414 | while (size > 0) { | |
1415 | struct iovec iov = { .iov_base = buffer, .iov_len = size }; | |
1416 | ssize_t len; | |
1417 | ||
1418 | len = qio_channel_readv(client->ioc, &iov, 1, errp); | |
1419 | if (len == QIO_CHANNEL_ERR_BLOCK) { | |
7075d235 SH |
1420 | WITH_QEMU_LOCK_GUARD(&client->lock) { |
1421 | client->read_yielding = true; | |
1422 | ||
1423 | /* Prompt main loop thread to re-run nbd_drained_poll() */ | |
1424 | aio_wait_kick(); | |
1425 | } | |
f148ae7d | 1426 | qio_channel_yield(client->ioc, G_IO_IN); |
7075d235 SH |
1427 | WITH_QEMU_LOCK_GUARD(&client->lock) { |
1428 | client->read_yielding = false; | |
1429 | if (client->quiescing) { | |
1430 | return -EAGAIN; | |
1431 | } | |
f148ae7d SL |
1432 | } |
1433 | continue; | |
1434 | } else if (len < 0) { | |
1435 | return -EIO; | |
1436 | } else if (len == 0) { | |
1437 | if (partial) { | |
1438 | error_setg(errp, | |
1439 | "Unexpected end-of-file before all bytes were read"); | |
1440 | return -EIO; | |
1441 | } else { | |
1442 | return 0; | |
1443 | } | |
1444 | } | |
1445 | ||
1446 | partial = true; | |
1447 | size -= len; | |
1448 | buffer = (uint8_t *) buffer + len; | |
1449 | } | |
1450 | return 1; | |
1451 | } | |
1452 | ||
d2223cdd PB |
1453 | static int coroutine_fn nbd_receive_request(NBDClient *client, NBDRequest *request, |
1454 | Error **errp) | |
75818250 | 1455 | { |
c8720ca0 EB |
1456 | uint8_t buf[NBD_EXTENDED_REQUEST_SIZE]; |
1457 | uint32_t magic, expect; | |
a0dc63a6 | 1458 | int ret; |
c8720ca0 EB |
1459 | size_t size = client->mode >= NBD_MODE_EXTENDED ? |
1460 | NBD_EXTENDED_REQUEST_SIZE : NBD_REQUEST_SIZE; | |
b2e3d87f | 1461 | |
c8720ca0 | 1462 | ret = nbd_read_eof(client, buf, size, errp); |
185b4338 PB |
1463 | if (ret < 0) { |
1464 | return ret; | |
1465 | } | |
1644ccce EB |
1466 | if (ret == 0) { |
1467 | return -EIO; | |
1468 | } | |
185b4338 | 1469 | |
c8720ca0 EB |
1470 | /* |
1471 | * Compact request | |
1472 | * [ 0 .. 3] magic (NBD_REQUEST_MAGIC) | |
1473 | * [ 4 .. 5] flags (NBD_CMD_FLAG_FUA, ...) | |
1474 | * [ 6 .. 7] type (NBD_CMD_READ, ...) | |
1475 | * [ 8 .. 15] cookie | |
1476 | * [16 .. 23] from | |
1477 | * [24 .. 27] len | |
1478 | * Extended request | |
1479 | * [ 0 .. 3] magic (NBD_EXTENDED_REQUEST_MAGIC) | |
1480 | * [ 4 .. 5] flags (NBD_CMD_FLAG_FUA, NBD_CMD_FLAG_PAYLOAD_LEN, ...) | |
1481 | * [ 6 .. 7] type (NBD_CMD_READ, ...) | |
1482 | * [ 8 .. 15] cookie | |
1483 | * [16 .. 23] from | |
1484 | * [24 .. 31] len | |
b2e3d87f NT |
1485 | */ |
1486 | ||
773dce3c | 1487 | magic = ldl_be_p(buf); |
b626b51a EB |
1488 | request->flags = lduw_be_p(buf + 4); |
1489 | request->type = lduw_be_p(buf + 6); | |
22efd811 | 1490 | request->cookie = ldq_be_p(buf + 8); |
773dce3c | 1491 | request->from = ldq_be_p(buf + 16); |
c8720ca0 EB |
1492 | if (client->mode >= NBD_MODE_EXTENDED) { |
1493 | request->len = ldq_be_p(buf + 24); | |
1494 | expect = NBD_EXTENDED_REQUEST_MAGIC; | |
1495 | } else { | |
1496 | request->len = (uint32_t)ldl_be_p(buf + 24); /* widen 32 to 64 bits */ | |
1497 | expect = NBD_REQUEST_MAGIC; | |
1498 | } | |
b2e3d87f | 1499 | |
9588463e VSO |
1500 | trace_nbd_receive_request(magic, request->flags, request->type, |
1501 | request->from, request->len); | |
b2e3d87f | 1502 | |
c8720ca0 EB |
1503 | if (magic != expect) { |
1504 | error_setg(errp, "invalid magic (got 0x%" PRIx32 ", expected 0x%" | |
1505 | PRIx32 ")", magic, expect); | |
185b4338 | 1506 | return -EINVAL; |
b2e3d87f NT |
1507 | } |
1508 | return 0; | |
75818250 TS |
1509 | } |
1510 | ||
41996e38 PB |
1511 | #define MAX_NBD_REQUESTS 16 |
1512 | ||
f816310d | 1513 | /* Runs in export AioContext and main loop thread */ |
ce33967a | 1514 | void nbd_client_get(NBDClient *client) |
1743b515 | 1515 | { |
f816310d | 1516 | qatomic_inc(&client->refcount); |
1743b515 PB |
1517 | } |
1518 | ||
ce33967a | 1519 | void nbd_client_put(NBDClient *client) |
1743b515 | 1520 | { |
f816310d SH |
1521 | assert(qemu_in_main_thread()); |
1522 | ||
1523 | if (qatomic_fetch_dec(&client->refcount) == 1) { | |
ff2b68aa | 1524 | /* The last reference should be dropped by client->close, |
f53a829b | 1525 | * which is called by client_close. |
ff2b68aa PB |
1526 | */ |
1527 | assert(client->closing); | |
1528 | ||
1c778ef7 DB |
1529 | object_unref(OBJECT(client->sioc)); |
1530 | object_unref(OBJECT(client->ioc)); | |
f95910fe DB |
1531 | if (client->tlscreds) { |
1532 | object_unref(OBJECT(client->tlscreds)); | |
1533 | } | |
b25e12da | 1534 | g_free(client->tlsauthz); |
6b8c01e7 PB |
1535 | if (client->exp) { |
1536 | QTAILQ_REMOVE(&client->exp->clients, client, next); | |
c69de1be | 1537 | blk_exp_unref(&client->exp->common); |
6b8c01e7 | 1538 | } |
fd358d83 | 1539 | g_free(client->contexts.bitmaps); |
7075d235 | 1540 | qemu_mutex_destroy(&client->lock); |
1743b515 PB |
1541 | g_free(client); |
1542 | } | |
1543 | } | |
1544 | ||
f816310d SH |
1545 | /* |
1546 | * Tries to release the reference to @client, but only if other references | |
1547 | * remain. This is an optimization for the common case where we want to avoid | |
1548 | * the expense of scheduling nbd_client_put() in the main loop thread. | |
1549 | * | |
1550 | * Returns true upon success or false if the reference was not released because | |
1551 | * it is the last reference. | |
1552 | */ | |
1553 | static bool nbd_client_put_nonzero(NBDClient *client) | |
1554 | { | |
1555 | int old = qatomic_read(&client->refcount); | |
1556 | int expected; | |
1557 | ||
1558 | do { | |
1559 | if (old == 1) { | |
1560 | return false; | |
1561 | } | |
1562 | ||
1563 | expected = old; | |
1564 | old = qatomic_cmpxchg(&client->refcount, expected, expected - 1); | |
1565 | } while (old != expected); | |
1566 | ||
1567 | return true; | |
1568 | } | |
1569 | ||
0c9390d9 | 1570 | static void client_close(NBDClient *client, bool negotiated) |
1743b515 | 1571 | { |
f816310d SH |
1572 | assert(qemu_in_main_thread()); |
1573 | ||
7075d235 SH |
1574 | WITH_QEMU_LOCK_GUARD(&client->lock) { |
1575 | if (client->closing) { | |
1576 | return; | |
1577 | } | |
ff2b68aa | 1578 | |
7075d235 SH |
1579 | client->closing = true; |
1580 | } | |
ff2b68aa PB |
1581 | |
1582 | /* Force requests to finish. They will drop their own references, | |
1583 | * then we'll close the socket and free the NBDClient. | |
1584 | */ | |
1c778ef7 DB |
1585 | qio_channel_shutdown(client->ioc, QIO_CHANNEL_SHUTDOWN_BOTH, |
1586 | NULL); | |
ff2b68aa PB |
1587 | |
1588 | /* Also tell the client, so that they release their reference. */ | |
0c9390d9 EB |
1589 | if (client->close_fn) { |
1590 | client->close_fn(client, negotiated); | |
1743b515 | 1591 | } |
1743b515 PB |
1592 | } |
1593 | ||
7075d235 | 1594 | /* Runs in export AioContext with client->lock held */ |
315f78ab | 1595 | static NBDRequestData *nbd_request_get(NBDClient *client) |
d9a73806 | 1596 | { |
315f78ab | 1597 | NBDRequestData *req; |
72deddc5 | 1598 | |
41996e38 PB |
1599 | assert(client->nb_requests <= MAX_NBD_REQUESTS - 1); |
1600 | client->nb_requests++; | |
1601 | ||
315f78ab | 1602 | req = g_new0(NBDRequestData, 1); |
72deddc5 | 1603 | req->client = client; |
d9a73806 PB |
1604 | return req; |
1605 | } | |
1606 | ||
7075d235 | 1607 | /* Runs in export AioContext with client->lock held */ |
315f78ab | 1608 | static void nbd_request_put(NBDRequestData *req) |
d9a73806 | 1609 | { |
72deddc5 | 1610 | NBDClient *client = req->client; |
e1adb27a | 1611 | |
2d821488 SH |
1612 | if (req->data) { |
1613 | qemu_vfree(req->data); | |
1614 | } | |
1729404c | 1615 | g_free(req); |
e1adb27a | 1616 | |
958c717d | 1617 | client->nb_requests--; |
fd6afc50 SL |
1618 | |
1619 | if (client->quiescing && client->nb_requests == 0) { | |
1620 | aio_wait_kick(); | |
1621 | } | |
1622 | ||
ff82911c | 1623 | nbd_client_receive_next_request(client); |
d9a73806 PB |
1624 | } |
1625 | ||
aadf99a7 | 1626 | static void blk_aio_attached(AioContext *ctx, void *opaque) |
f2149281 HR |
1627 | { |
1628 | NBDExport *exp = opaque; | |
1629 | NBDClient *client; | |
1630 | ||
7075d235 SH |
1631 | assert(qemu_in_main_thread()); |
1632 | ||
9588463e | 1633 | trace_nbd_blk_aio_attached(exp->name, ctx); |
f2149281 | 1634 | |
8612c686 | 1635 | exp->common.ctx = ctx; |
f2149281 HR |
1636 | |
1637 | QTAILQ_FOREACH(client, &exp->clients, next) { | |
7075d235 SH |
1638 | WITH_QEMU_LOCK_GUARD(&client->lock) { |
1639 | assert(client->nb_requests == 0); | |
1640 | assert(client->recv_coroutine == NULL); | |
1641 | assert(client->send_coroutine == NULL); | |
1642 | } | |
f148ae7d SL |
1643 | } |
1644 | } | |
1645 | ||
fd6afc50 | 1646 | static void blk_aio_detach(void *opaque) |
f148ae7d SL |
1647 | { |
1648 | NBDExport *exp = opaque; | |
f148ae7d | 1649 | |
7075d235 SH |
1650 | assert(qemu_in_main_thread()); |
1651 | ||
fd6afc50 SL |
1652 | trace_nbd_blk_aio_detach(exp->name, exp->common.ctx); |
1653 | ||
fd6afc50 SL |
1654 | exp->common.ctx = NULL; |
1655 | } | |
1656 | ||
1657 | static void nbd_drained_begin(void *opaque) | |
1658 | { | |
1659 | NBDExport *exp = opaque; | |
1660 | NBDClient *client; | |
1661 | ||
7075d235 SH |
1662 | assert(qemu_in_main_thread()); |
1663 | ||
fd6afc50 | 1664 | QTAILQ_FOREACH(client, &exp->clients, next) { |
7075d235 SH |
1665 | WITH_QEMU_LOCK_GUARD(&client->lock) { |
1666 | client->quiescing = true; | |
1667 | } | |
fd6afc50 SL |
1668 | } |
1669 | } | |
f148ae7d | 1670 | |
fd6afc50 SL |
1671 | static void nbd_drained_end(void *opaque) |
1672 | { | |
1673 | NBDExport *exp = opaque; | |
1674 | NBDClient *client; | |
f148ae7d | 1675 | |
7075d235 SH |
1676 | assert(qemu_in_main_thread()); |
1677 | ||
fd6afc50 | 1678 | QTAILQ_FOREACH(client, &exp->clients, next) { |
7075d235 SH |
1679 | WITH_QEMU_LOCK_GUARD(&client->lock) { |
1680 | client->quiescing = false; | |
1681 | nbd_client_receive_next_request(client); | |
1682 | } | |
f2149281 HR |
1683 | } |
1684 | } | |
1685 | ||
7075d235 SH |
1686 | /* Runs in export AioContext */ |
1687 | static void nbd_wake_read_bh(void *opaque) | |
1688 | { | |
1689 | NBDClient *client = opaque; | |
1690 | qio_channel_wake_read(client->ioc); | |
1691 | } | |
1692 | ||
fd6afc50 | 1693 | static bool nbd_drained_poll(void *opaque) |
f2149281 HR |
1694 | { |
1695 | NBDExport *exp = opaque; | |
fd6afc50 | 1696 | NBDClient *client; |
f2149281 | 1697 | |
7075d235 SH |
1698 | assert(qemu_in_main_thread()); |
1699 | ||
fd6afc50 | 1700 | QTAILQ_FOREACH(client, &exp->clients, next) { |
7075d235 SH |
1701 | WITH_QEMU_LOCK_GUARD(&client->lock) { |
1702 | if (client->nb_requests != 0) { | |
1703 | /* | |
1704 | * If there's a coroutine waiting for a request on nbd_read_eof() | |
1705 | * enter it here so we don't depend on the client to wake it up. | |
1706 | * | |
1707 | * Schedule a BH in the export AioContext to avoid missing the | |
1708 | * wake up due to the race between qio_channel_wake_read() and | |
1709 | * qio_channel_yield(). | |
1710 | */ | |
1711 | if (client->recv_coroutine != NULL && client->read_yielding) { | |
1712 | aio_bh_schedule_oneshot(nbd_export_aio_context(client->exp), | |
1713 | nbd_wake_read_bh, client); | |
1714 | } | |
f2149281 | 1715 | |
7075d235 SH |
1716 | return true; |
1717 | } | |
fd6afc50 SL |
1718 | } |
1719 | } | |
f2149281 | 1720 | |
fd6afc50 | 1721 | return false; |
f2149281 HR |
1722 | } |
1723 | ||
741cc431 HR |
1724 | static void nbd_eject_notifier(Notifier *n, void *data) |
1725 | { | |
1726 | NBDExport *exp = container_of(n, NBDExport, eject_notifier); | |
61bc846d | 1727 | |
7075d235 SH |
1728 | assert(qemu_in_main_thread()); |
1729 | ||
bc4ee65b | 1730 | blk_exp_request_shutdown(&exp->common); |
741cc431 HR |
1731 | } |
1732 | ||
9b562c64 KW |
1733 | void nbd_export_set_on_eject_blk(BlockExport *exp, BlockBackend *blk) |
1734 | { | |
1735 | NBDExport *nbd_exp = container_of(exp, NBDExport, common); | |
1736 | assert(exp->drv == &blk_exp_nbd); | |
1737 | assert(nbd_exp->eject_notifier_blk == NULL); | |
1738 | ||
1739 | blk_ref(blk); | |
1740 | nbd_exp->eject_notifier_blk = blk; | |
1741 | nbd_exp->eject_notifier.notify = nbd_eject_notifier; | |
1742 | blk_add_remove_bs_notifier(blk, &nbd_exp->eject_notifier); | |
1743 | } | |
1744 | ||
fd6afc50 SL |
1745 | static const BlockDevOps nbd_block_ops = { |
1746 | .drained_begin = nbd_drained_begin, | |
1747 | .drained_end = nbd_drained_end, | |
1748 | .drained_poll = nbd_drained_poll, | |
1749 | }; | |
1750 | ||
5b1cb497 KW |
1751 | static int nbd_export_create(BlockExport *blk_exp, BlockExportOptions *exp_args, |
1752 | Error **errp) | |
af49bbbe | 1753 | { |
a6ff7989 | 1754 | NBDExport *exp = container_of(blk_exp, NBDExport, common); |
5b1cb497 | 1755 | BlockExportOptionsNbd *arg = &exp_args->u.nbd; |
8461b4d6 | 1756 | const char *name = arg->name ?: exp_args->node_name; |
331170e0 | 1757 | BlockBackend *blk = blk_exp->blk; |
b57e4de0 | 1758 | int64_t size; |
331170e0 | 1759 | uint64_t perm, shared_perm; |
5b1cb497 | 1760 | bool readonly = !exp_args->writable; |
e5fb29d5 | 1761 | BlockDirtyBitmapOrStrList *bitmaps; |
3b1f244c | 1762 | size_t i; |
d7086422 | 1763 | int ret; |
cd7fca95 | 1764 | |
372b69f5 | 1765 | GLOBAL_STATE_CODE(); |
5b1cb497 KW |
1766 | assert(exp_args->type == BLOCK_EXPORT_TYPE_NBD); |
1767 | ||
1768 | if (!nbd_server_is_running()) { | |
1769 | error_setg(errp, "NBD server not running"); | |
1770 | return -EINVAL; | |
1771 | } | |
1772 | ||
8461b4d6 MA |
1773 | if (strlen(name) > NBD_MAX_STRING_SIZE) { |
1774 | error_setg(errp, "export name '%s' too long", name); | |
5b1cb497 KW |
1775 | return -EINVAL; |
1776 | } | |
1777 | ||
1778 | if (arg->description && strlen(arg->description) > NBD_MAX_STRING_SIZE) { | |
1779 | error_setg(errp, "description '%s' too long", arg->description); | |
1780 | return -EINVAL; | |
1781 | } | |
1782 | ||
8461b4d6 MA |
1783 | if (nbd_export_find(name)) { |
1784 | error_setg(errp, "NBD server already has export named '%s'", name); | |
5b1cb497 KW |
1785 | return -EEXIST; |
1786 | } | |
1787 | ||
331170e0 | 1788 | size = blk_getlength(blk); |
b57e4de0 KW |
1789 | if (size < 0) { |
1790 | error_setg_errno(errp, -size, | |
1791 | "Failed to determine the NBD export's length"); | |
a6ff7989 | 1792 | return size; |
b57e4de0 KW |
1793 | } |
1794 | ||
8a7ce4f9 KW |
1795 | /* Don't allow resize while the NBD server is running, otherwise we don't |
1796 | * care what happens with the node. */ | |
331170e0 | 1797 | blk_get_perm(blk, &perm, &shared_perm); |
331170e0 | 1798 | ret = blk_set_perm(blk, perm, shared_perm & ~BLK_PERM_RESIZE, errp); |
d7086422 | 1799 | if (ret < 0) { |
331170e0 | 1800 | return ret; |
d7086422 | 1801 | } |
331170e0 | 1802 | |
4b9441f6 | 1803 | QTAILQ_INIT(&exp->clients); |
8461b4d6 | 1804 | exp->name = g_strdup(name); |
5b1cb497 | 1805 | exp->description = g_strdup(arg->description); |
dbb38caa EB |
1806 | exp->nbdflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_FLUSH | |
1807 | NBD_FLAG_SEND_FUA | NBD_FLAG_SEND_CACHE); | |
58a6fdcc EB |
1808 | |
1809 | if (nbd_server_max_connections() != 1) { | |
1810 | exp->nbdflags |= NBD_FLAG_CAN_MULTI_CONN; | |
1811 | } | |
dbb38caa EB |
1812 | if (readonly) { |
1813 | exp->nbdflags |= NBD_FLAG_READ_ONLY; | |
dbb38caa | 1814 | } else { |
b491dbb7 EB |
1815 | exp->nbdflags |= (NBD_FLAG_SEND_TRIM | NBD_FLAG_SEND_WRITE_ZEROES | |
1816 | NBD_FLAG_SEND_FAST_ZERO); | |
dbb38caa | 1817 | } |
7596bbb3 | 1818 | exp->size = QEMU_ALIGN_DOWN(size, BDRV_SECTOR_SIZE); |
98f44bbe | 1819 | |
372b69f5 KW |
1820 | bdrv_graph_rdlock_main_loop(); |
1821 | ||
cbad81ce | 1822 | for (bitmaps = arg->bitmaps; bitmaps; bitmaps = bitmaps->next) { |
3b1f244c EB |
1823 | exp->nr_export_bitmaps++; |
1824 | } | |
1825 | exp->export_bitmaps = g_new0(BdrvDirtyBitmap *, exp->nr_export_bitmaps); | |
1826 | for (i = 0, bitmaps = arg->bitmaps; bitmaps; | |
e5fb29d5 VSO |
1827 | i++, bitmaps = bitmaps->next) |
1828 | { | |
1829 | const char *bitmap; | |
331170e0 | 1830 | BlockDriverState *bs = blk_bs(blk); |
678ba275 | 1831 | BdrvDirtyBitmap *bm = NULL; |
678ba275 | 1832 | |
e5fb29d5 VSO |
1833 | switch (bitmaps->value->type) { |
1834 | case QTYPE_QSTRING: | |
1835 | bitmap = bitmaps->value->u.local; | |
1836 | while (bs) { | |
1837 | bm = bdrv_find_dirty_bitmap(bs, bitmap); | |
1838 | if (bm != NULL) { | |
1839 | break; | |
1840 | } | |
1841 | ||
1842 | bs = bdrv_filter_or_cow_bs(bs); | |
678ba275 EB |
1843 | } |
1844 | ||
e5fb29d5 VSO |
1845 | if (bm == NULL) { |
1846 | ret = -ENOENT; | |
1847 | error_setg(errp, "Bitmap '%s' is not found", | |
1848 | bitmaps->value->u.local); | |
1849 | goto fail; | |
1850 | } | |
678ba275 | 1851 | |
e5fb29d5 VSO |
1852 | if (readonly && bdrv_is_writable(bs) && |
1853 | bdrv_dirty_bitmap_enabled(bm)) { | |
1854 | ret = -EINVAL; | |
1855 | error_setg(errp, "Enabled bitmap '%s' incompatible with " | |
1856 | "readonly export", bitmap); | |
1857 | goto fail; | |
1858 | } | |
1859 | break; | |
1860 | case QTYPE_QDICT: | |
1861 | bitmap = bitmaps->value->u.external.name; | |
1862 | bm = block_dirty_bitmap_lookup(bitmaps->value->u.external.node, | |
1863 | bitmap, NULL, errp); | |
1864 | if (!bm) { | |
1865 | ret = -ENOENT; | |
1866 | goto fail; | |
1867 | } | |
1868 | break; | |
1869 | default: | |
1870 | abort(); | |
678ba275 EB |
1871 | } |
1872 | ||
e5fb29d5 | 1873 | assert(bm); |
3b78a927 | 1874 | |
e5fb29d5 | 1875 | if (bdrv_dirty_bitmap_check(bm, BDRV_BITMAP_ALLOW_RO, errp)) { |
a6ff7989 | 1876 | ret = -EINVAL; |
678ba275 EB |
1877 | goto fail; |
1878 | } | |
1879 | ||
3b1f244c | 1880 | exp->export_bitmaps[i] = bm; |
cbad81ce | 1881 | assert(strlen(bitmap) <= BDRV_BITMAP_MAX_NAME_SIZE); |
678ba275 EB |
1882 | } |
1883 | ||
3b1f244c EB |
1884 | /* Mark bitmaps busy in a separate loop, to simplify roll-back concerns. */ |
1885 | for (i = 0; i < exp->nr_export_bitmaps; i++) { | |
1886 | bdrv_dirty_bitmap_set_busy(exp->export_bitmaps[i], true); | |
1887 | } | |
1888 | ||
dbc7b014 EB |
1889 | exp->allocation_depth = arg->allocation_depth; |
1890 | ||
fd6afc50 SL |
1891 | /* |
1892 | * We need to inhibit request queuing in the block layer to ensure we can | |
1893 | * be properly quiesced when entering a drained section, as our coroutines | |
1894 | * servicing pending requests might enter blk_pread(). | |
1895 | */ | |
1896 | blk_set_disable_request_queuing(blk, true); | |
1897 | ||
aadf99a7 | 1898 | blk_add_aio_context_notifier(blk, blk_aio_attached, blk_aio_detach, exp); |
741cc431 | 1899 | |
fd6afc50 SL |
1900 | blk_set_dev_ops(blk, &nbd_block_ops, exp); |
1901 | ||
3fa4c765 | 1902 | QTAILQ_INSERT_TAIL(&exports, exp, next); |
c69de1be | 1903 | |
372b69f5 KW |
1904 | bdrv_graph_rdunlock_main_loop(); |
1905 | ||
a6ff7989 | 1906 | return 0; |
98f44bbe HR |
1907 | |
1908 | fail: | |
372b69f5 | 1909 | bdrv_graph_rdunlock_main_loop(); |
3b1f244c | 1910 | g_free(exp->export_bitmaps); |
3fa4c765 EB |
1911 | g_free(exp->name); |
1912 | g_free(exp->description); | |
a6ff7989 | 1913 | return ret; |
af49bbbe PB |
1914 | } |
1915 | ||
ee0a19ec PB |
1916 | NBDExport *nbd_export_find(const char *name) |
1917 | { | |
1918 | NBDExport *exp; | |
1919 | QTAILQ_FOREACH(exp, &exports, next) { | |
1920 | if (strcmp(name, exp->name) == 0) { | |
1921 | return exp; | |
1922 | } | |
1923 | } | |
1924 | ||
1925 | return NULL; | |
1926 | } | |
1927 | ||
61bc846d EB |
1928 | AioContext * |
1929 | nbd_export_aio_context(NBDExport *exp) | |
1930 | { | |
8612c686 | 1931 | return exp->common.ctx; |
61bc846d EB |
1932 | } |
1933 | ||
bc4ee65b | 1934 | static void nbd_export_request_shutdown(BlockExport *blk_exp) |
af49bbbe | 1935 | { |
bc4ee65b | 1936 | NBDExport *exp = container_of(blk_exp, NBDExport, common); |
4b9441f6 | 1937 | NBDClient *client, *next; |
2c8d9f06 | 1938 | |
c69de1be | 1939 | blk_exp_ref(&exp->common); |
3fa4c765 EB |
1940 | /* |
1941 | * TODO: Should we expand QMP NbdServerRemoveNode enum to allow a | |
1942 | * close mode that stops advertising the export to new clients but | |
1943 | * still permits existing clients to run to completion? Because of | |
1944 | * that possibility, nbd_export_close() can be called more than | |
1945 | * once on an export. | |
1946 | */ | |
4b9441f6 | 1947 | QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) { |
0c9390d9 | 1948 | client_close(client, true); |
4b9441f6 | 1949 | } |
3fa4c765 | 1950 | if (exp->name) { |
3fa4c765 EB |
1951 | g_free(exp->name); |
1952 | exp->name = NULL; | |
1953 | QTAILQ_REMOVE(&exports, exp, next); | |
1954 | } | |
c69de1be | 1955 | blk_exp_unref(&exp->common); |
2c8d9f06 PB |
1956 | } |
1957 | ||
c69de1be | 1958 | static void nbd_export_delete(BlockExport *blk_exp) |
2c8d9f06 | 1959 | { |
3b1f244c | 1960 | size_t i; |
c69de1be | 1961 | NBDExport *exp = container_of(blk_exp, NBDExport, common); |
2c8d9f06 | 1962 | |
c69de1be KW |
1963 | assert(exp->name == NULL); |
1964 | assert(QTAILQ_EMPTY(&exp->clients)); | |
d6268348 | 1965 | |
c69de1be KW |
1966 | g_free(exp->description); |
1967 | exp->description = NULL; | |
1968 | ||
dd5b6780 PB |
1969 | if (exp->eject_notifier_blk) { |
1970 | notifier_remove(&exp->eject_notifier); | |
1971 | blk_unref(exp->eject_notifier_blk); | |
c69de1be | 1972 | } |
dd5b6780 PB |
1973 | blk_remove_aio_context_notifier(exp->common.blk, blk_aio_attached, |
1974 | blk_aio_detach, exp); | |
1975 | blk_set_disable_request_queuing(exp->common.blk, false); | |
3d068aff | 1976 | |
3b1f244c EB |
1977 | for (i = 0; i < exp->nr_export_bitmaps; i++) { |
1978 | bdrv_dirty_bitmap_set_busy(exp->export_bitmaps[i], false); | |
2c8d9f06 | 1979 | } |
af49bbbe PB |
1980 | } |
1981 | ||
56ee8626 KW |
1982 | const BlockExportDriver blk_exp_nbd = { |
1983 | .type = BLOCK_EXPORT_TYPE_NBD, | |
a6ff7989 | 1984 | .instance_size = sizeof(NBDExport), |
56ee8626 | 1985 | .create = nbd_export_create, |
c69de1be | 1986 | .delete = nbd_export_delete, |
bc4ee65b | 1987 | .request_shutdown = nbd_export_request_shutdown, |
56ee8626 KW |
1988 | }; |
1989 | ||
de79bfc3 VSO |
1990 | static int coroutine_fn nbd_co_send_iov(NBDClient *client, struct iovec *iov, |
1991 | unsigned niov, Error **errp) | |
1992 | { | |
1993 | int ret; | |
1994 | ||
1995 | g_assert(qemu_in_coroutine()); | |
1996 | qemu_co_mutex_lock(&client->send_lock); | |
1997 | client->send_coroutine = qemu_coroutine_self(); | |
1998 | ||
1999 | ret = qio_channel_writev_all(client->ioc, iov, niov, errp) < 0 ? -EIO : 0; | |
2000 | ||
2001 | client->send_coroutine = NULL; | |
2002 | qemu_co_mutex_unlock(&client->send_lock); | |
2003 | ||
2004 | return ret; | |
2005 | } | |
2006 | ||
caad5384 | 2007 | static inline void set_be_simple_reply(NBDSimpleReply *reply, uint64_t error, |
22efd811 | 2008 | uint64_t cookie) |
caad5384 VSO |
2009 | { |
2010 | stl_be_p(&reply->magic, NBD_SIMPLE_REPLY_MAGIC); | |
2011 | stl_be_p(&reply->error, error); | |
22efd811 | 2012 | stq_be_p(&reply->cookie, cookie); |
caad5384 VSO |
2013 | } |
2014 | ||
d2223cdd | 2015 | static int coroutine_fn nbd_co_send_simple_reply(NBDClient *client, |
66d4f4fe | 2016 | NBDRequest *request, |
d2223cdd PB |
2017 | uint32_t error, |
2018 | void *data, | |
b2578459 | 2019 | uint64_t len, |
d2223cdd | 2020 | Error **errp) |
22045592 | 2021 | { |
de79bfc3 | 2022 | NBDSimpleReply reply; |
14cea41d | 2023 | int nbd_err = system_errno_to_nbd_errno(error); |
de79bfc3 VSO |
2024 | struct iovec iov[] = { |
2025 | {.iov_base = &reply, .iov_len = sizeof(reply)}, | |
2026 | {.iov_base = data, .iov_len = len} | |
2027 | }; | |
6fb2b972 | 2028 | |
a7c8ed36 | 2029 | assert(!len || !nbd_err); |
b2578459 | 2030 | assert(len <= NBD_MAX_BUFFER_SIZE); |
ac132d05 EB |
2031 | assert(client->mode < NBD_MODE_STRUCTURED || |
2032 | (client->mode == NBD_MODE_STRUCTURED && | |
2033 | request->type != NBD_CMD_READ)); | |
22efd811 | 2034 | trace_nbd_co_send_simple_reply(request->cookie, nbd_err, |
66d4f4fe | 2035 | nbd_err_lookup(nbd_err), len); |
22efd811 | 2036 | set_be_simple_reply(&reply, nbd_err, request->cookie); |
262db388 | 2037 | |
a7c8ed36 | 2038 | return nbd_co_send_iov(client, iov, 2, errp); |
22045592 PB |
2039 | } |
2040 | ||
a7c8ed36 EB |
2041 | /* |
2042 | * Prepare the header of a reply chunk for network transmission. | |
2043 | * | |
2044 | * On input, @iov is partially initialized: iov[0].iov_base must point | |
2045 | * to an uninitialized NBDReply, while the remaining @niov elements | |
2046 | * (if any) must be ready for transmission. This function then | |
2047 | * populates iov[0] for transmission. | |
2048 | */ | |
2049 | static inline void set_be_chunk(NBDClient *client, struct iovec *iov, | |
2050 | size_t niov, uint16_t flags, uint16_t type, | |
66d4f4fe | 2051 | NBDRequest *request) |
5c54e7fa | 2052 | { |
a7c8ed36 EB |
2053 | size_t i, length = 0; |
2054 | ||
2055 | for (i = 1; i < niov; i++) { | |
2056 | length += iov[i].iov_len; | |
2057 | } | |
2058 | assert(length <= NBD_MAX_BUFFER_SIZE + sizeof(NBDStructuredReadData)); | |
2059 | ||
11d3355f EB |
2060 | if (client->mode >= NBD_MODE_EXTENDED) { |
2061 | NBDExtendedReplyChunk *chunk = iov->iov_base; | |
2062 | ||
2063 | iov[0].iov_len = sizeof(*chunk); | |
2064 | stl_be_p(&chunk->magic, NBD_EXTENDED_REPLY_MAGIC); | |
2065 | stw_be_p(&chunk->flags, flags); | |
2066 | stw_be_p(&chunk->type, type); | |
2067 | stq_be_p(&chunk->cookie, request->cookie); | |
2068 | stq_be_p(&chunk->offset, request->from); | |
2069 | stq_be_p(&chunk->length, length); | |
2070 | } else { | |
2071 | NBDStructuredReplyChunk *chunk = iov->iov_base; | |
2072 | ||
2073 | iov[0].iov_len = sizeof(*chunk); | |
2074 | stl_be_p(&chunk->magic, NBD_STRUCTURED_REPLY_MAGIC); | |
2075 | stw_be_p(&chunk->flags, flags); | |
2076 | stw_be_p(&chunk->type, type); | |
2077 | stq_be_p(&chunk->cookie, request->cookie); | |
2078 | stl_be_p(&chunk->length, length); | |
2079 | } | |
5c54e7fa VSO |
2080 | } |
2081 | ||
a7c8ed36 | 2082 | static int coroutine_fn nbd_co_send_chunk_done(NBDClient *client, |
66d4f4fe | 2083 | NBDRequest *request, |
a7c8ed36 | 2084 | Error **errp) |
ef8c887e | 2085 | { |
a7c8ed36 | 2086 | NBDReply hdr; |
ef8c887e | 2087 | struct iovec iov[] = { |
a7c8ed36 | 2088 | {.iov_base = &hdr}, |
ef8c887e EB |
2089 | }; |
2090 | ||
22efd811 | 2091 | trace_nbd_co_send_chunk_done(request->cookie); |
a7c8ed36 | 2092 | set_be_chunk(client, iov, 1, NBD_REPLY_FLAG_DONE, |
66d4f4fe | 2093 | NBD_REPLY_TYPE_NONE, request); |
ef8c887e EB |
2094 | return nbd_co_send_iov(client, iov, 1, errp); |
2095 | } | |
2096 | ||
a7c8ed36 | 2097 | static int coroutine_fn nbd_co_send_chunk_read(NBDClient *client, |
66d4f4fe | 2098 | NBDRequest *request, |
a7c8ed36 EB |
2099 | uint64_t offset, |
2100 | void *data, | |
b2578459 | 2101 | uint64_t size, |
a7c8ed36 EB |
2102 | bool final, |
2103 | Error **errp) | |
5c54e7fa | 2104 | { |
a7c8ed36 | 2105 | NBDReply hdr; |
efdc0c10 | 2106 | NBDStructuredReadData chunk; |
5c54e7fa | 2107 | struct iovec iov[] = { |
a7c8ed36 | 2108 | {.iov_base = &hdr}, |
5c54e7fa VSO |
2109 | {.iov_base = &chunk, .iov_len = sizeof(chunk)}, |
2110 | {.iov_base = data, .iov_len = size} | |
2111 | }; | |
2112 | ||
b2578459 | 2113 | assert(size && size <= NBD_MAX_BUFFER_SIZE); |
22efd811 | 2114 | trace_nbd_co_send_chunk_read(request->cookie, offset, data, size); |
a7c8ed36 | 2115 | set_be_chunk(client, iov, 3, final ? NBD_REPLY_FLAG_DONE : 0, |
66d4f4fe | 2116 | NBD_REPLY_TYPE_OFFSET_DATA, request); |
5c54e7fa VSO |
2117 | stq_be_p(&chunk.offset, offset); |
2118 | ||
a7c8ed36 | 2119 | return nbd_co_send_iov(client, iov, 3, errp); |
5c54e7fa | 2120 | } |
ac132d05 | 2121 | |
a7c8ed36 | 2122 | static int coroutine_fn nbd_co_send_chunk_error(NBDClient *client, |
66d4f4fe | 2123 | NBDRequest *request, |
a7c8ed36 EB |
2124 | uint32_t error, |
2125 | const char *msg, | |
2126 | Error **errp) | |
60ace2ba | 2127 | { |
a7c8ed36 | 2128 | NBDReply hdr; |
60ace2ba VSO |
2129 | NBDStructuredError chunk; |
2130 | int nbd_err = system_errno_to_nbd_errno(error); | |
2131 | struct iovec iov[] = { | |
a7c8ed36 | 2132 | {.iov_base = &hdr}, |
60ace2ba VSO |
2133 | {.iov_base = &chunk, .iov_len = sizeof(chunk)}, |
2134 | {.iov_base = (char *)msg, .iov_len = msg ? strlen(msg) : 0}, | |
2135 | }; | |
2136 | ||
2137 | assert(nbd_err); | |
22efd811 | 2138 | trace_nbd_co_send_chunk_error(request->cookie, nbd_err, |
a7c8ed36 EB |
2139 | nbd_err_lookup(nbd_err), msg ? msg : ""); |
2140 | set_be_chunk(client, iov, 3, NBD_REPLY_FLAG_DONE, | |
66d4f4fe | 2141 | NBD_REPLY_TYPE_ERROR, request); |
60ace2ba | 2142 | stl_be_p(&chunk.error, nbd_err); |
a7c8ed36 | 2143 | stw_be_p(&chunk.message_length, iov[2].iov_len); |
60ace2ba | 2144 | |
a7c8ed36 | 2145 | return nbd_co_send_iov(client, iov, 3, errp); |
60ace2ba VSO |
2146 | } |
2147 | ||
37e02aeb | 2148 | /* Do a sparse read and send the structured reply to the client. |
ff7e261b | 2149 | * Returns -errno if sending fails. blk_co_block_status_above() failure is |
37e02aeb VSO |
2150 | * reported to the client, at which point this function succeeds. |
2151 | */ | |
418638d3 | 2152 | static int coroutine_fn nbd_co_send_sparse_read(NBDClient *client, |
66d4f4fe | 2153 | NBDRequest *request, |
418638d3 EB |
2154 | uint64_t offset, |
2155 | uint8_t *data, | |
b2578459 | 2156 | uint64_t size, |
418638d3 EB |
2157 | Error **errp) |
2158 | { | |
2159 | int ret = 0; | |
2160 | NBDExport *exp = client->exp; | |
2161 | size_t progress = 0; | |
2162 | ||
b2578459 | 2163 | assert(size <= NBD_MAX_BUFFER_SIZE); |
418638d3 EB |
2164 | while (progress < size) { |
2165 | int64_t pnum; | |
ff7e261b EGE |
2166 | int status = blk_co_block_status_above(exp->common.blk, NULL, |
2167 | offset + progress, | |
2168 | size - progress, &pnum, NULL, | |
2169 | NULL); | |
e2de3256 | 2170 | bool final; |
418638d3 EB |
2171 | |
2172 | if (status < 0) { | |
37e02aeb VSO |
2173 | char *msg = g_strdup_printf("unable to check for holes: %s", |
2174 | strerror(-status)); | |
2175 | ||
66d4f4fe | 2176 | ret = nbd_co_send_chunk_error(client, request, -status, msg, errp); |
37e02aeb VSO |
2177 | g_free(msg); |
2178 | return ret; | |
418638d3 EB |
2179 | } |
2180 | assert(pnum && pnum <= size - progress); | |
e2de3256 | 2181 | final = progress + pnum == size; |
418638d3 | 2182 | if (status & BDRV_BLOCK_ZERO) { |
a7c8ed36 | 2183 | NBDReply hdr; |
418638d3 EB |
2184 | NBDStructuredReadHole chunk; |
2185 | struct iovec iov[] = { | |
a7c8ed36 | 2186 | {.iov_base = &hdr}, |
418638d3 EB |
2187 | {.iov_base = &chunk, .iov_len = sizeof(chunk)}, |
2188 | }; | |
2189 | ||
22efd811 | 2190 | trace_nbd_co_send_chunk_read_hole(request->cookie, |
66d4f4fe | 2191 | offset + progress, pnum); |
a7c8ed36 EB |
2192 | set_be_chunk(client, iov, 2, |
2193 | final ? NBD_REPLY_FLAG_DONE : 0, | |
66d4f4fe | 2194 | NBD_REPLY_TYPE_OFFSET_HOLE, request); |
418638d3 EB |
2195 | stq_be_p(&chunk.offset, offset + progress); |
2196 | stl_be_p(&chunk.length, pnum); | |
a7c8ed36 | 2197 | ret = nbd_co_send_iov(client, iov, 2, errp); |
418638d3 | 2198 | } else { |
d2223cdd PB |
2199 | ret = blk_co_pread(exp->common.blk, offset + progress, pnum, |
2200 | data + progress, 0); | |
418638d3 EB |
2201 | if (ret < 0) { |
2202 | error_setg_errno(errp, -ret, "reading from file failed"); | |
2203 | break; | |
2204 | } | |
66d4f4fe | 2205 | ret = nbd_co_send_chunk_read(client, request, offset + progress, |
a7c8ed36 | 2206 | data + progress, pnum, final, errp); |
418638d3 EB |
2207 | } |
2208 | ||
2209 | if (ret < 0) { | |
2210 | break; | |
2211 | } | |
2212 | progress += pnum; | |
2213 | } | |
418638d3 EB |
2214 | return ret; |
2215 | } | |
2216 | ||
89cbc7e3 | 2217 | typedef struct NBDExtentArray { |
bcc16cc1 | 2218 | NBDExtent64 *extents; |
89cbc7e3 VSO |
2219 | unsigned int nb_alloc; |
2220 | unsigned int count; | |
2221 | uint64_t total_length; | |
bcc16cc1 | 2222 | bool extended; |
89cbc7e3 VSO |
2223 | bool can_add; |
2224 | bool converted_to_be; | |
2225 | } NBDExtentArray; | |
2226 | ||
bcc16cc1 EB |
2227 | static NBDExtentArray *nbd_extent_array_new(unsigned int nb_alloc, |
2228 | NBDMode mode) | |
89cbc7e3 VSO |
2229 | { |
2230 | NBDExtentArray *ea = g_new0(NBDExtentArray, 1); | |
2231 | ||
bcc16cc1 | 2232 | assert(mode >= NBD_MODE_STRUCTURED); |
89cbc7e3 | 2233 | ea->nb_alloc = nb_alloc; |
bcc16cc1 EB |
2234 | ea->extents = g_new(NBDExtent64, nb_alloc); |
2235 | ea->extended = mode >= NBD_MODE_EXTENDED; | |
89cbc7e3 VSO |
2236 | ea->can_add = true; |
2237 | ||
2238 | return ea; | |
2239 | } | |
2240 | ||
2241 | static void nbd_extent_array_free(NBDExtentArray *ea) | |
2242 | { | |
2243 | g_free(ea->extents); | |
2244 | g_free(ea); | |
2245 | } | |
e0e7fe07 | 2246 | G_DEFINE_AUTOPTR_CLEANUP_FUNC(NBDExtentArray, nbd_extent_array_free) |
89cbc7e3 VSO |
2247 | |
2248 | /* Further modifications of the array after conversion are abandoned */ | |
2249 | static void nbd_extent_array_convert_to_be(NBDExtentArray *ea) | |
2250 | { | |
2251 | int i; | |
2252 | ||
2253 | assert(!ea->converted_to_be); | |
bcc16cc1 | 2254 | assert(ea->extended); |
89cbc7e3 VSO |
2255 | ea->can_add = false; |
2256 | ea->converted_to_be = true; | |
2257 | ||
2258 | for (i = 0; i < ea->count; i++) { | |
bcc16cc1 EB |
2259 | ea->extents[i].length = cpu_to_be64(ea->extents[i].length); |
2260 | ea->extents[i].flags = cpu_to_be64(ea->extents[i].flags); | |
89cbc7e3 VSO |
2261 | } |
2262 | } | |
2263 | ||
bcc16cc1 EB |
2264 | /* Further modifications of the array after conversion are abandoned */ |
2265 | static NBDExtent32 *nbd_extent_array_convert_to_narrow(NBDExtentArray *ea) | |
2266 | { | |
2267 | int i; | |
2268 | NBDExtent32 *extents = g_new(NBDExtent32, ea->count); | |
2269 | ||
2270 | assert(!ea->converted_to_be); | |
2271 | assert(!ea->extended); | |
2272 | ea->can_add = false; | |
2273 | ea->converted_to_be = true; | |
2274 | ||
2275 | for (i = 0; i < ea->count; i++) { | |
2276 | assert((ea->extents[i].length | ea->extents[i].flags) <= UINT32_MAX); | |
2277 | extents[i].length = cpu_to_be32(ea->extents[i].length); | |
2278 | extents[i].flags = cpu_to_be32(ea->extents[i].flags); | |
2279 | } | |
2280 | ||
2281 | return extents; | |
2282 | } | |
2283 | ||
fb7afc79 | 2284 | /* |
89cbc7e3 VSO |
2285 | * Add extent to NBDExtentArray. If extent can't be added (no available space), |
2286 | * return -1. | |
2287 | * For safety, when returning -1 for the first time, .can_add is set to false, | |
314b9026 EB |
2288 | * and further calls to nbd_extent_array_add() will crash. |
2289 | * (this avoids the situation where a caller ignores failure to add one extent, | |
2290 | * where adding another extent that would squash into the last array entry | |
2291 | * would result in an incorrect range reported to the client) | |
fb7afc79 | 2292 | */ |
89cbc7e3 | 2293 | static int nbd_extent_array_add(NBDExtentArray *ea, |
bcc16cc1 | 2294 | uint64_t length, uint32_t flags) |
e7b1948d | 2295 | { |
89cbc7e3 VSO |
2296 | assert(ea->can_add); |
2297 | ||
2298 | if (!length) { | |
2299 | return 0; | |
2300 | } | |
bcc16cc1 EB |
2301 | if (!ea->extended) { |
2302 | assert(length <= UINT32_MAX); | |
2303 | } | |
89cbc7e3 VSO |
2304 | |
2305 | /* Extend previous extent if flags are the same */ | |
2306 | if (ea->count > 0 && flags == ea->extents[ea->count - 1].flags) { | |
bcc16cc1 | 2307 | uint64_t sum = length + ea->extents[ea->count - 1].length; |
89cbc7e3 | 2308 | |
bcc16cc1 EB |
2309 | /* |
2310 | * sum cannot overflow: the block layer bounds image size at | |
2311 | * 2^63, and ea->extents[].length comes from the block layer. | |
2312 | */ | |
2313 | assert(sum >= length); | |
2314 | if (sum <= UINT32_MAX || ea->extended) { | |
89cbc7e3 VSO |
2315 | ea->extents[ea->count - 1].length = sum; |
2316 | ea->total_length += length; | |
2317 | return 0; | |
2318 | } | |
2319 | } | |
2320 | ||
2321 | if (ea->count >= ea->nb_alloc) { | |
2322 | ea->can_add = false; | |
2323 | return -1; | |
2324 | } | |
2325 | ||
2326 | ea->total_length += length; | |
bcc16cc1 | 2327 | ea->extents[ea->count] = (NBDExtent64) {.length = length, .flags = flags}; |
89cbc7e3 | 2328 | ea->count++; |
e7b1948d | 2329 | |
89cbc7e3 VSO |
2330 | return 0; |
2331 | } | |
2332 | ||
ff7e261b | 2333 | static int coroutine_fn blockstatus_to_extents(BlockBackend *blk, |
6f58ac55 EGE |
2334 | uint64_t offset, uint64_t bytes, |
2335 | NBDExtentArray *ea) | |
89cbc7e3 VSO |
2336 | { |
2337 | while (bytes) { | |
e7b1948d VSO |
2338 | uint32_t flags; |
2339 | int64_t num; | |
ff7e261b EGE |
2340 | int ret = blk_co_block_status_above(blk, NULL, offset, bytes, &num, |
2341 | NULL, NULL); | |
fb7afc79 | 2342 | |
e7b1948d VSO |
2343 | if (ret < 0) { |
2344 | return ret; | |
2345 | } | |
2346 | ||
0da98568 NS |
2347 | flags = (ret & BDRV_BLOCK_DATA ? 0 : NBD_STATE_HOLE) | |
2348 | (ret & BDRV_BLOCK_ZERO ? NBD_STATE_ZERO : 0); | |
e7b1948d | 2349 | |
89cbc7e3 VSO |
2350 | if (nbd_extent_array_add(ea, num, flags) < 0) { |
2351 | return 0; | |
e7b1948d | 2352 | } |
fb7afc79 | 2353 | |
89cbc7e3 VSO |
2354 | offset += num; |
2355 | bytes -= num; | |
e7b1948d VSO |
2356 | } |
2357 | ||
e7b1948d VSO |
2358 | return 0; |
2359 | } | |
2360 | ||
ff7e261b | 2361 | static int coroutine_fn blockalloc_to_extents(BlockBackend *blk, |
6f58ac55 EGE |
2362 | uint64_t offset, uint64_t bytes, |
2363 | NBDExtentArray *ea) | |
71719cd5 EB |
2364 | { |
2365 | while (bytes) { | |
2366 | int64_t num; | |
ff7e261b EGE |
2367 | int ret = blk_co_is_allocated_above(blk, NULL, false, offset, bytes, |
2368 | &num); | |
71719cd5 EB |
2369 | |
2370 | if (ret < 0) { | |
2371 | return ret; | |
2372 | } | |
2373 | ||
2374 | if (nbd_extent_array_add(ea, num, ret) < 0) { | |
2375 | return 0; | |
2376 | } | |
2377 | ||
2378 | offset += num; | |
2379 | bytes -= num; | |
2380 | } | |
2381 | ||
2382 | return 0; | |
2383 | } | |
2384 | ||
89cbc7e3 VSO |
2385 | /* |
2386 | * nbd_co_send_extents | |
3d068aff | 2387 | * |
89cbc7e3 VSO |
2388 | * @ea is converted to BE by the function |
2389 | * @last controls whether NBD_REPLY_FLAG_DONE is sent. | |
3d068aff | 2390 | */ |
d2223cdd | 2391 | static int coroutine_fn |
66d4f4fe | 2392 | nbd_co_send_extents(NBDClient *client, NBDRequest *request, NBDExtentArray *ea, |
d2223cdd | 2393 | bool last, uint32_t context_id, Error **errp) |
e7b1948d | 2394 | { |
a7c8ed36 | 2395 | NBDReply hdr; |
bcc16cc1 EB |
2396 | NBDStructuredMeta meta; |
2397 | NBDExtendedMeta meta_ext; | |
2398 | g_autofree NBDExtent32 *extents = NULL; | |
2399 | uint16_t type; | |
2400 | struct iovec iov[] = { {.iov_base = &hdr}, {0}, {0} }; | |
e7b1948d | 2401 | |
bcc16cc1 EB |
2402 | if (client->mode >= NBD_MODE_EXTENDED) { |
2403 | type = NBD_REPLY_TYPE_BLOCK_STATUS_EXT; | |
2404 | ||
2405 | iov[1].iov_base = &meta_ext; | |
2406 | iov[1].iov_len = sizeof(meta_ext); | |
2407 | stl_be_p(&meta_ext.context_id, context_id); | |
2408 | stl_be_p(&meta_ext.count, ea->count); | |
2409 | ||
2410 | nbd_extent_array_convert_to_be(ea); | |
2411 | iov[2].iov_base = ea->extents; | |
2412 | iov[2].iov_len = ea->count * sizeof(ea->extents[0]); | |
2413 | } else { | |
2414 | type = NBD_REPLY_TYPE_BLOCK_STATUS; | |
2415 | ||
2416 | iov[1].iov_base = &meta; | |
2417 | iov[1].iov_len = sizeof(meta); | |
2418 | stl_be_p(&meta.context_id, context_id); | |
2419 | ||
2420 | extents = nbd_extent_array_convert_to_narrow(ea); | |
2421 | iov[2].iov_base = extents; | |
2422 | iov[2].iov_len = ea->count * sizeof(extents[0]); | |
2423 | } | |
89cbc7e3 | 2424 | |
22efd811 | 2425 | trace_nbd_co_send_extents(request->cookie, ea->count, context_id, |
66d4f4fe | 2426 | ea->total_length, last); |
bcc16cc1 EB |
2427 | set_be_chunk(client, iov, 3, last ? NBD_REPLY_FLAG_DONE : 0, type, |
2428 | request); | |
e7b1948d | 2429 | |
a7c8ed36 | 2430 | return nbd_co_send_iov(client, iov, 3, errp); |
e7b1948d VSO |
2431 | } |
2432 | ||
2433 | /* Get block status from the exported device and send it to the client */ | |
6f58ac55 | 2434 | static int |
66d4f4fe | 2435 | coroutine_fn nbd_co_send_block_status(NBDClient *client, NBDRequest *request, |
ff7e261b | 2436 | BlockBackend *blk, uint64_t offset, |
bcc16cc1 | 2437 | uint64_t length, bool dont_fragment, |
6f58ac55 EGE |
2438 | bool last, uint32_t context_id, |
2439 | Error **errp) | |
e7b1948d VSO |
2440 | { |
2441 | int ret; | |
416e34bd | 2442 | unsigned int nb_extents = dont_fragment ? 1 : NBD_MAX_BLOCK_STATUS_EXTENTS; |
bcc16cc1 EB |
2443 | g_autoptr(NBDExtentArray) ea = |
2444 | nbd_extent_array_new(nb_extents, client->mode); | |
e7b1948d | 2445 | |
71719cd5 | 2446 | if (context_id == NBD_META_ID_BASE_ALLOCATION) { |
ff7e261b | 2447 | ret = blockstatus_to_extents(blk, offset, length, ea); |
71719cd5 | 2448 | } else { |
ff7e261b | 2449 | ret = blockalloc_to_extents(blk, offset, length, ea); |
71719cd5 | 2450 | } |
e7b1948d | 2451 | if (ret < 0) { |
66d4f4fe | 2452 | return nbd_co_send_chunk_error(client, request, -ret, |
a7c8ed36 | 2453 | "can't get block status", errp); |
e7b1948d VSO |
2454 | } |
2455 | ||
66d4f4fe | 2456 | return nbd_co_send_extents(client, request, ea, last, context_id, errp); |
3d068aff VSO |
2457 | } |
2458 | ||
dacbb6eb | 2459 | /* Populate @ea from a dirty bitmap. */ |
89cbc7e3 VSO |
2460 | static void bitmap_to_extents(BdrvDirtyBitmap *bitmap, |
2461 | uint64_t offset, uint64_t length, | |
dacbb6eb | 2462 | NBDExtentArray *es) |
3d068aff | 2463 | { |
dacbb6eb VSO |
2464 | int64_t start, dirty_start, dirty_count; |
2465 | int64_t end = offset + length; | |
2466 | bool full = false; | |
bcc16cc1 | 2467 | int64_t bound = es->extended ? INT64_MAX : INT32_MAX; |
3d068aff VSO |
2468 | |
2469 | bdrv_dirty_bitmap_lock(bitmap); | |
2470 | ||
dacbb6eb | 2471 | for (start = offset; |
bcc16cc1 | 2472 | bdrv_dirty_bitmap_next_dirty_area(bitmap, start, end, bound, |
dacbb6eb VSO |
2473 | &dirty_start, &dirty_count); |
2474 | start = dirty_start + dirty_count) | |
2475 | { | |
2476 | if ((nbd_extent_array_add(es, dirty_start - start, 0) < 0) || | |
2477 | (nbd_extent_array_add(es, dirty_count, NBD_STATE_DIRTY) < 0)) | |
2478 | { | |
2479 | full = true; | |
89cbc7e3 VSO |
2480 | break; |
2481 | } | |
3d068aff VSO |
2482 | } |
2483 | ||
dacbb6eb | 2484 | if (!full) { |
c0b21f2e EB |
2485 | /* last non dirty extent, nothing to do if array is now full */ |
2486 | (void) nbd_extent_array_add(es, end - start, 0); | |
dacbb6eb | 2487 | } |
3d068aff VSO |
2488 | |
2489 | bdrv_dirty_bitmap_unlock(bitmap); | |
3d068aff VSO |
2490 | } |
2491 | ||
66d4f4fe EB |
2492 | static int coroutine_fn nbd_co_send_bitmap(NBDClient *client, |
2493 | NBDRequest *request, | |
2494 | BdrvDirtyBitmap *bitmap, | |
2495 | uint64_t offset, | |
bcc16cc1 | 2496 | uint64_t length, bool dont_fragment, |
66d4f4fe EB |
2497 | bool last, uint32_t context_id, |
2498 | Error **errp) | |
3d068aff | 2499 | { |
416e34bd | 2500 | unsigned int nb_extents = dont_fragment ? 1 : NBD_MAX_BLOCK_STATUS_EXTENTS; |
bcc16cc1 EB |
2501 | g_autoptr(NBDExtentArray) ea = |
2502 | nbd_extent_array_new(nb_extents, client->mode); | |
3d068aff | 2503 | |
dacbb6eb | 2504 | bitmap_to_extents(bitmap, offset, length, ea); |
3d068aff | 2505 | |
66d4f4fe | 2506 | return nbd_co_send_extents(client, request, ea, last, context_id, errp); |
e7b1948d VSO |
2507 | } |
2508 | ||
2dcbb11b EB |
2509 | /* |
2510 | * nbd_co_block_status_payload_read | |
2511 | * Called when a client wants a subset of negotiated contexts via a | |
2512 | * BLOCK_STATUS payload. Check the payload for valid length and | |
2513 | * contents. On success, return 0 with request updated to effective | |
2514 | * length. If request was invalid but all payload consumed, return 0 | |
2515 | * with request->len and request->contexts->count set to 0 (which will | |
2516 | * trigger an appropriate NBD_EINVAL response later on). Return | |
2517 | * negative errno if the payload was not fully consumed. | |
2518 | */ | |
2519 | static int | |
2520 | nbd_co_block_status_payload_read(NBDClient *client, NBDRequest *request, | |
2521 | Error **errp) | |
2522 | { | |
2523 | uint64_t payload_len = request->len; | |
2524 | g_autofree char *buf = NULL; | |
2525 | size_t count, i, nr_bitmaps; | |
2526 | uint32_t id; | |
2527 | ||
2528 | if (payload_len > NBD_MAX_BUFFER_SIZE) { | |
2529 | error_setg(errp, "len (%" PRIu64 ") is larger than max len (%u)", | |
2530 | request->len, NBD_MAX_BUFFER_SIZE); | |
2531 | return -EINVAL; | |
2532 | } | |
2533 | ||
2534 | assert(client->contexts.exp == client->exp); | |
2535 | nr_bitmaps = client->exp->nr_export_bitmaps; | |
2536 | request->contexts = g_new0(NBDMetaContexts, 1); | |
2537 | request->contexts->exp = client->exp; | |
2538 | ||
2539 | if (payload_len % sizeof(uint32_t) || | |
2540 | payload_len < sizeof(NBDBlockStatusPayload) || | |
2541 | payload_len > (sizeof(NBDBlockStatusPayload) + | |
2542 | sizeof(id) * client->contexts.count)) { | |
2543 | goto skip; | |
2544 | } | |
2545 | ||
2546 | buf = g_malloc(payload_len); | |
2547 | if (nbd_read(client->ioc, buf, payload_len, | |
2548 | "CMD_BLOCK_STATUS data", errp) < 0) { | |
2549 | return -EIO; | |
2550 | } | |
2551 | trace_nbd_co_receive_request_payload_received(request->cookie, | |
2552 | payload_len); | |
2553 | request->contexts->bitmaps = g_new0(bool, nr_bitmaps); | |
2554 | count = (payload_len - sizeof(NBDBlockStatusPayload)) / sizeof(id); | |
2555 | payload_len = 0; | |
2556 | ||
2557 | for (i = 0; i < count; i++) { | |
2558 | id = ldl_be_p(buf + sizeof(NBDBlockStatusPayload) + sizeof(id) * i); | |
2559 | if (id == NBD_META_ID_BASE_ALLOCATION) { | |
2560 | if (!client->contexts.base_allocation || | |
2561 | request->contexts->base_allocation) { | |
2562 | goto skip; | |
2563 | } | |
2564 | request->contexts->base_allocation = true; | |
2565 | } else if (id == NBD_META_ID_ALLOCATION_DEPTH) { | |
2566 | if (!client->contexts.allocation_depth || | |
2567 | request->contexts->allocation_depth) { | |
2568 | goto skip; | |
2569 | } | |
2570 | request->contexts->allocation_depth = true; | |
2571 | } else { | |
2572 | unsigned idx = id - NBD_META_ID_DIRTY_BITMAP; | |
2573 | ||
2574 | if (idx >= nr_bitmaps || !client->contexts.bitmaps[idx] || | |
2575 | request->contexts->bitmaps[idx]) { | |
2576 | goto skip; | |
2577 | } | |
2578 | request->contexts->bitmaps[idx] = true; | |
2579 | } | |
2580 | } | |
2581 | ||
2582 | request->len = ldq_be_p(buf); | |
2583 | request->contexts->count = count; | |
2584 | return 0; | |
2585 | ||
2586 | skip: | |
2587 | trace_nbd_co_receive_block_status_payload_compliance(request->from, | |
2588 | request->len); | |
2589 | request->len = request->contexts->count = 0; | |
2590 | return nbd_drop(client->ioc, payload_len, errp); | |
2591 | } | |
2592 | ||
2a6e128b VSO |
2593 | /* nbd_co_receive_request |
2594 | * Collect a client request. Return 0 if request looks valid, -EIO to drop | |
f148ae7d SL |
2595 | * connection right away, -EAGAIN to indicate we were interrupted and the |
2596 | * channel should be quiesced, and any other negative value to report an error | |
2597 | * to the client (although the caller may still need to disconnect after | |
2598 | * reporting the error). | |
2a6e128b | 2599 | */ |
8db7e2d6 EB |
2600 | static int coroutine_fn nbd_co_receive_request(NBDRequestData *req, |
2601 | NBDRequest *request, | |
d2223cdd | 2602 | Error **errp) |
a030b347 | 2603 | { |
72deddc5 | 2604 | NBDClient *client = req->client; |
009cd866 | 2605 | bool extended_with_payload; |
8db7e2d6 EB |
2606 | bool check_length = false; |
2607 | bool check_rofs = false; | |
2608 | bool allocate_buffer = false; | |
009cd866 EB |
2609 | bool payload_okay = false; |
2610 | uint64_t payload_len = 0; | |
8db7e2d6 | 2611 | int valid_flags = NBD_CMD_FLAG_FUA; |
f148ae7d | 2612 | int ret; |
a030b347 | 2613 | |
1c778ef7 | 2614 | g_assert(qemu_in_coroutine()); |
f148ae7d SL |
2615 | ret = nbd_receive_request(client, request, errp); |
2616 | if (ret < 0) { | |
314b9026 | 2617 | return ret; |
a030b347 PB |
2618 | } |
2619 | ||
22efd811 | 2620 | trace_nbd_co_receive_request_decode_type(request->cookie, request->type, |
3736cc5b | 2621 | nbd_cmd_lookup(request->type)); |
009cd866 EB |
2622 | extended_with_payload = client->mode >= NBD_MODE_EXTENDED && |
2623 | request->flags & NBD_CMD_FLAG_PAYLOAD_LEN; | |
2624 | if (extended_with_payload) { | |
2625 | payload_len = request->len; | |
2626 | check_length = true; | |
2627 | } | |
2628 | ||
8db7e2d6 EB |
2629 | switch (request->type) { |
2630 | case NBD_CMD_DISC: | |
29b6c3b3 EB |
2631 | /* Special case: we're going to disconnect without a reply, |
2632 | * whether or not flags, from, or len are bogus */ | |
8db7e2d6 | 2633 | req->complete = true; |
ee898b87 | 2634 | return -EIO; |
29b6c3b3 | 2635 | |
8db7e2d6 EB |
2636 | case NBD_CMD_READ: |
2637 | if (client->mode >= NBD_MODE_STRUCTURED) { | |
2638 | valid_flags |= NBD_CMD_FLAG_DF; | |
eb38c3b6 | 2639 | } |
8db7e2d6 EB |
2640 | check_length = true; |
2641 | allocate_buffer = true; | |
2642 | break; | |
eb38c3b6 | 2643 | |
8db7e2d6 | 2644 | case NBD_CMD_WRITE: |
009cd866 EB |
2645 | if (client->mode >= NBD_MODE_EXTENDED) { |
2646 | if (!extended_with_payload) { | |
2647 | /* The client is noncompliant. Trace it, but proceed. */ | |
2648 | trace_nbd_co_receive_ext_payload_compliance(request->from, | |
2649 | request->len); | |
2650 | } | |
2651 | valid_flags |= NBD_CMD_FLAG_PAYLOAD_LEN; | |
2652 | } | |
2653 | payload_okay = true; | |
8db7e2d6 EB |
2654 | payload_len = request->len; |
2655 | check_length = true; | |
2656 | allocate_buffer = true; | |
2657 | check_rofs = true; | |
2658 | break; | |
2659 | ||
2660 | case NBD_CMD_FLUSH: | |
2661 | break; | |
2662 | ||
2663 | case NBD_CMD_TRIM: | |
2664 | check_rofs = true; | |
2665 | break; | |
2666 | ||
2667 | case NBD_CMD_CACHE: | |
2668 | check_length = true; | |
2669 | break; | |
2670 | ||
2671 | case NBD_CMD_WRITE_ZEROES: | |
2672 | valid_flags |= NBD_CMD_FLAG_NO_HOLE | NBD_CMD_FLAG_FAST_ZERO; | |
2673 | check_rofs = true; | |
2674 | break; | |
2675 | ||
2676 | case NBD_CMD_BLOCK_STATUS: | |
2dcbb11b EB |
2677 | if (extended_with_payload) { |
2678 | ret = nbd_co_block_status_payload_read(client, request, errp); | |
2679 | if (ret < 0) { | |
2680 | return ret; | |
2681 | } | |
2682 | /* payload now consumed */ | |
2683 | check_length = false; | |
2684 | payload_len = 0; | |
2685 | valid_flags |= NBD_CMD_FLAG_PAYLOAD_LEN; | |
2686 | } else { | |
2687 | request->contexts = &client->contexts; | |
2688 | } | |
8db7e2d6 EB |
2689 | valid_flags |= NBD_CMD_FLAG_REQ_ONE; |
2690 | break; | |
2691 | ||
2692 | default: | |
2693 | /* Unrecognized, will fail later */ | |
2694 | ; | |
2d821488 | 2695 | } |
7fa5c565 | 2696 | |
8db7e2d6 EB |
2697 | /* Payload and buffer handling. */ |
2698 | if (!payload_len) { | |
2699 | req->complete = true; | |
2700 | } | |
2701 | if (check_length && request->len > NBD_MAX_BUFFER_SIZE) { | |
2702 | /* READ, WRITE, CACHE */ | |
2703 | error_setg(errp, "len (%" PRIu64 ") is larger than max len (%u)", | |
2704 | request->len, NBD_MAX_BUFFER_SIZE); | |
2705 | return -EINVAL; | |
2706 | } | |
009cd866 EB |
2707 | if (payload_len && !payload_okay) { |
2708 | /* | |
2709 | * For now, we don't support payloads on other commands; but | |
2710 | * we can keep the connection alive by ignoring the payload. | |
2711 | * We will fail the command later with NBD_EINVAL for the use | |
2712 | * of an unsupported flag (and not for access beyond bounds). | |
2713 | */ | |
2714 | assert(request->type != NBD_CMD_WRITE); | |
2715 | request->len = 0; | |
2716 | } | |
8db7e2d6 EB |
2717 | if (allocate_buffer) { |
2718 | /* READ, WRITE */ | |
2719 | req->data = blk_try_blockalign(client->exp->common.blk, | |
2720 | request->len); | |
2721 | if (req->data == NULL) { | |
2722 | error_setg(errp, "No memory"); | |
2723 | return -ENOMEM; | |
2724 | } | |
2725 | } | |
2726 | if (payload_len) { | |
009cd866 EB |
2727 | if (payload_okay) { |
2728 | /* WRITE */ | |
2729 | assert(req->data); | |
2730 | ret = nbd_read(client->ioc, req->data, payload_len, | |
2731 | "CMD_WRITE data", errp); | |
2732 | } else { | |
2733 | ret = nbd_drop(client->ioc, payload_len, errp); | |
2734 | } | |
8db7e2d6 | 2735 | if (ret < 0) { |
ee898b87 | 2736 | return -EIO; |
a030b347 | 2737 | } |
29b6c3b3 | 2738 | req->complete = true; |
22efd811 | 2739 | trace_nbd_co_receive_request_payload_received(request->cookie, |
8db7e2d6 | 2740 | payload_len); |
a030b347 | 2741 | } |
29b6c3b3 | 2742 | |
fed5f8f8 | 2743 | /* Sanity checks. */ |
8db7e2d6 EB |
2744 | if (client->exp->nbdflags & NBD_FLAG_READ_ONLY && check_rofs) { |
2745 | /* WRITE, TRIM, WRITE_ZEROES */ | |
fed5f8f8 EB |
2746 | error_setg(errp, "Export is read-only"); |
2747 | return -EROFS; | |
2748 | } | |
2749 | if (request->from > client->exp->size || | |
9d26dfcb | 2750 | request->len > client->exp->size - request->from) { |
b2578459 | 2751 | error_setg(errp, "operation past EOF; From: %" PRIu64 ", Len: %" PRIu64 |
2fd2c840 | 2752 | ", Size: %" PRIu64, request->from, request->len, |
9d26dfcb | 2753 | client->exp->size); |
fed5f8f8 EB |
2754 | return (request->type == NBD_CMD_WRITE || |
2755 | request->type == NBD_CMD_WRITE_ZEROES) ? -ENOSPC : -EINVAL; | |
29b6c3b3 | 2756 | } |
6e280648 EB |
2757 | if (client->check_align && !QEMU_IS_ALIGNED(request->from | request->len, |
2758 | client->check_align)) { | |
2759 | /* | |
2760 | * The block layer gracefully handles unaligned requests, but | |
2761 | * it's still worth tracing client non-compliance | |
2762 | */ | |
2763 | trace_nbd_co_receive_align_compliance(nbd_cmd_lookup(request->type), | |
2764 | request->from, | |
2765 | request->len, | |
2766 | client->check_align); | |
2767 | } | |
5c54e7fa VSO |
2768 | if (request->flags & ~valid_flags) { |
2769 | error_setg(errp, "unsupported flags for command %s (got 0x%x)", | |
2770 | nbd_cmd_lookup(request->type), request->flags); | |
ee898b87 | 2771 | return -EINVAL; |
1f4d6d18 | 2772 | } |
29b6c3b3 | 2773 | |
ee898b87 | 2774 | return 0; |
a030b347 PB |
2775 | } |
2776 | ||
6a417599 VSO |
2777 | /* Send simple reply without a payload, or a structured error |
2778 | * @error_msg is ignored if @ret >= 0 | |
2779 | * Returns 0 if connection is still live, -errno on failure to talk to client | |
2780 | */ | |
2781 | static coroutine_fn int nbd_send_generic_reply(NBDClient *client, | |
66d4f4fe | 2782 | NBDRequest *request, |
6a417599 VSO |
2783 | int ret, |
2784 | const char *error_msg, | |
2785 | Error **errp) | |
2786 | { | |
ac132d05 | 2787 | if (client->mode >= NBD_MODE_STRUCTURED && ret < 0) { |
66d4f4fe | 2788 | return nbd_co_send_chunk_error(client, request, -ret, error_msg, errp); |
11d3355f EB |
2789 | } else if (client->mode >= NBD_MODE_EXTENDED) { |
2790 | return nbd_co_send_chunk_done(client, request, errp); | |
6a417599 | 2791 | } else { |
66d4f4fe | 2792 | return nbd_co_send_simple_reply(client, request, ret < 0 ? -ret : 0, |
6a417599 VSO |
2793 | NULL, 0, errp); |
2794 | } | |
2795 | } | |
2796 | ||
2797 | /* Handle NBD_CMD_READ request. | |
2798 | * Return -errno if sending fails. Other errors are reported directly to the | |
2799 | * client as an error reply. */ | |
2800 | static coroutine_fn int nbd_do_cmd_read(NBDClient *client, NBDRequest *request, | |
2801 | uint8_t *data, Error **errp) | |
2802 | { | |
2803 | int ret; | |
2804 | NBDExport *exp = client->exp; | |
2805 | ||
7fa5c565 | 2806 | assert(request->type == NBD_CMD_READ); |
b2578459 | 2807 | assert(request->len <= NBD_MAX_BUFFER_SIZE); |
6a417599 VSO |
2808 | |
2809 | /* XXX: NBD Protocol only documents use of FUA with WRITE */ | |
2810 | if (request->flags & NBD_CMD_FLAG_FUA) { | |
37a4f70c | 2811 | ret = blk_co_flush(exp->common.blk); |
6a417599 | 2812 | if (ret < 0) { |
66d4f4fe | 2813 | return nbd_send_generic_reply(client, request, ret, |
6a417599 VSO |
2814 | "flush failed", errp); |
2815 | } | |
2816 | } | |
2817 | ||
ac132d05 EB |
2818 | if (client->mode >= NBD_MODE_STRUCTURED && |
2819 | !(request->flags & NBD_CMD_FLAG_DF) && request->len) | |
2f454def | 2820 | { |
66d4f4fe | 2821 | return nbd_co_send_sparse_read(client, request, request->from, |
6a417599 VSO |
2822 | data, request->len, errp); |
2823 | } | |
2824 | ||
d2223cdd | 2825 | ret = blk_co_pread(exp->common.blk, request->from, request->len, data, 0); |
7fa5c565 | 2826 | if (ret < 0) { |
66d4f4fe | 2827 | return nbd_send_generic_reply(client, request, ret, |
6a417599 VSO |
2828 | "reading from file failed", errp); |
2829 | } | |
2830 | ||
ac132d05 | 2831 | if (client->mode >= NBD_MODE_STRUCTURED) { |
6a417599 | 2832 | if (request->len) { |
66d4f4fe | 2833 | return nbd_co_send_chunk_read(client, request, request->from, data, |
a7c8ed36 | 2834 | request->len, true, errp); |
6a417599 | 2835 | } else { |
66d4f4fe | 2836 | return nbd_co_send_chunk_done(client, request, errp); |
6a417599 VSO |
2837 | } |
2838 | } else { | |
66d4f4fe | 2839 | return nbd_co_send_simple_reply(client, request, 0, |
6a417599 VSO |
2840 | data, request->len, errp); |
2841 | } | |
2842 | } | |
2843 | ||
7fa5c565 VSO |
2844 | /* |
2845 | * nbd_do_cmd_cache | |
2846 | * | |
2847 | * Handle NBD_CMD_CACHE request. | |
2848 | * Return -errno if sending fails. Other errors are reported directly to the | |
2849 | * client as an error reply. | |
2850 | */ | |
2851 | static coroutine_fn int nbd_do_cmd_cache(NBDClient *client, NBDRequest *request, | |
2852 | Error **errp) | |
2853 | { | |
2854 | int ret; | |
2855 | NBDExport *exp = client->exp; | |
2856 | ||
2857 | assert(request->type == NBD_CMD_CACHE); | |
b2578459 | 2858 | assert(request->len <= NBD_MAX_BUFFER_SIZE); |
7fa5c565 | 2859 | |
37a4f70c | 2860 | ret = blk_co_preadv(exp->common.blk, request->from, request->len, |
7fa5c565 VSO |
2861 | NULL, BDRV_REQ_COPY_ON_READ | BDRV_REQ_PREFETCH); |
2862 | ||
66d4f4fe | 2863 | return nbd_send_generic_reply(client, request, ret, |
7fa5c565 VSO |
2864 | "caching data failed", errp); |
2865 | } | |
2866 | ||
6f302e60 VSO |
2867 | /* Handle NBD request. |
2868 | * Return -errno if sending fails. Other errors are reported directly to the | |
2869 | * client as an error reply. */ | |
2870 | static coroutine_fn int nbd_handle_request(NBDClient *client, | |
2871 | NBDRequest *request, | |
2872 | uint8_t *data, Error **errp) | |
2873 | { | |
2874 | int ret; | |
2875 | int flags; | |
2876 | NBDExport *exp = client->exp; | |
2877 | char *msg; | |
3b1f244c | 2878 | size_t i; |
6f302e60 VSO |
2879 | |
2880 | switch (request->type) { | |
bc37b06a | 2881 | case NBD_CMD_CACHE: |
7fa5c565 VSO |
2882 | return nbd_do_cmd_cache(client, request, errp); |
2883 | ||
2884 | case NBD_CMD_READ: | |
6f302e60 VSO |
2885 | return nbd_do_cmd_read(client, request, data, errp); |
2886 | ||
2887 | case NBD_CMD_WRITE: | |
2888 | flags = 0; | |
2889 | if (request->flags & NBD_CMD_FLAG_FUA) { | |
2890 | flags |= BDRV_REQ_FUA; | |
2891 | } | |
b2578459 | 2892 | assert(request->len <= NBD_MAX_BUFFER_SIZE); |
d2223cdd PB |
2893 | ret = blk_co_pwrite(exp->common.blk, request->from, request->len, data, |
2894 | flags); | |
66d4f4fe | 2895 | return nbd_send_generic_reply(client, request, ret, |
6f302e60 VSO |
2896 | "writing to file failed", errp); |
2897 | ||
2898 | case NBD_CMD_WRITE_ZEROES: | |
2899 | flags = 0; | |
2900 | if (request->flags & NBD_CMD_FLAG_FUA) { | |
2901 | flags |= BDRV_REQ_FUA; | |
2902 | } | |
2903 | if (!(request->flags & NBD_CMD_FLAG_NO_HOLE)) { | |
2904 | flags |= BDRV_REQ_MAY_UNMAP; | |
2905 | } | |
b491dbb7 EB |
2906 | if (request->flags & NBD_CMD_FLAG_FAST_ZERO) { |
2907 | flags |= BDRV_REQ_NO_FALLBACK; | |
2908 | } | |
d2223cdd PB |
2909 | ret = blk_co_pwrite_zeroes(exp->common.blk, request->from, request->len, |
2910 | flags); | |
66d4f4fe | 2911 | return nbd_send_generic_reply(client, request, ret, |
6f302e60 VSO |
2912 | "writing to file failed", errp); |
2913 | ||
2914 | case NBD_CMD_DISC: | |
2915 | /* unreachable, thanks to special case in nbd_co_receive_request() */ | |
2916 | abort(); | |
2917 | ||
2918 | case NBD_CMD_FLUSH: | |
37a4f70c | 2919 | ret = blk_co_flush(exp->common.blk); |
66d4f4fe | 2920 | return nbd_send_generic_reply(client, request, ret, |
6f302e60 VSO |
2921 | "flush failed", errp); |
2922 | ||
2923 | case NBD_CMD_TRIM: | |
e3557422 | 2924 | ret = blk_co_pdiscard(exp->common.blk, request->from, request->len); |
890cbccb | 2925 | if (ret >= 0 && request->flags & NBD_CMD_FLAG_FUA) { |
37a4f70c | 2926 | ret = blk_co_flush(exp->common.blk); |
65529782 | 2927 | } |
66d4f4fe | 2928 | return nbd_send_generic_reply(client, request, ret, |
6f302e60 VSO |
2929 | "discard failed", errp); |
2930 | ||
e7b1948d | 2931 | case NBD_CMD_BLOCK_STATUS: |
1dec4643 | 2932 | assert(request->contexts); |
bcc16cc1 EB |
2933 | assert(client->mode >= NBD_MODE_EXTENDED || |
2934 | request->len <= UINT32_MAX); | |
1dec4643 | 2935 | if (request->contexts->count) { |
fb7afc79 | 2936 | bool dont_fragment = request->flags & NBD_CMD_FLAG_REQ_ONE; |
1dec4643 | 2937 | int contexts_remaining = request->contexts->count; |
fb7afc79 | 2938 | |
2dcbb11b EB |
2939 | if (!request->len) { |
2940 | return nbd_send_generic_reply(client, request, -EINVAL, | |
2941 | "need non-zero length", errp); | |
2942 | } | |
1dec4643 | 2943 | if (request->contexts->base_allocation) { |
66d4f4fe | 2944 | ret = nbd_co_send_block_status(client, request, |
ff7e261b | 2945 | exp->common.blk, |
37a4f70c | 2946 | request->from, |
fb7afc79 | 2947 | request->len, dont_fragment, |
47ec485e | 2948 | !--contexts_remaining, |
3d068aff VSO |
2949 | NBD_META_ID_BASE_ALLOCATION, |
2950 | errp); | |
73e064cc EB |
2951 | if (ret < 0) { |
2952 | return ret; | |
2953 | } | |
2954 | } | |
2955 | ||
1dec4643 | 2956 | if (request->contexts->allocation_depth) { |
66d4f4fe | 2957 | ret = nbd_co_send_block_status(client, request, |
ff7e261b | 2958 | exp->common.blk, |
71719cd5 EB |
2959 | request->from, request->len, |
2960 | dont_fragment, | |
2961 | !--contexts_remaining, | |
2962 | NBD_META_ID_ALLOCATION_DEPTH, | |
2963 | errp); | |
2964 | if (ret < 0) { | |
2965 | return ret; | |
2966 | } | |
2967 | } | |
2968 | ||
1dec4643 | 2969 | assert(request->contexts->exp == client->exp); |
3b1f244c | 2970 | for (i = 0; i < client->exp->nr_export_bitmaps; i++) { |
1dec4643 | 2971 | if (!request->contexts->bitmaps[i]) { |
3b1f244c EB |
2972 | continue; |
2973 | } | |
66d4f4fe | 2974 | ret = nbd_co_send_bitmap(client, request, |
3b1f244c | 2975 | client->exp->export_bitmaps[i], |
3d068aff | 2976 | request->from, request->len, |
47ec485e | 2977 | dont_fragment, !--contexts_remaining, |
3b1f244c | 2978 | NBD_META_ID_DIRTY_BITMAP + i, errp); |
73e064cc EB |
2979 | if (ret < 0) { |
2980 | return ret; | |
2981 | } | |
3d068aff VSO |
2982 | } |
2983 | ||
47ec485e EB |
2984 | assert(!contexts_remaining); |
2985 | ||
73e064cc | 2986 | return 0; |
1dec4643 EB |
2987 | } else if (client->contexts.count) { |
2988 | return nbd_send_generic_reply(client, request, -EINVAL, | |
2989 | "CMD_BLOCK_STATUS payload not valid", | |
2990 | errp); | |
e7b1948d | 2991 | } else { |
66d4f4fe | 2992 | return nbd_send_generic_reply(client, request, -EINVAL, |
e7b1948d VSO |
2993 | "CMD_BLOCK_STATUS not negotiated", |
2994 | errp); | |
2995 | } | |
2996 | ||
6f302e60 VSO |
2997 | default: |
2998 | msg = g_strdup_printf("invalid request type (%" PRIu32 ") received", | |
2999 | request->type); | |
66d4f4fe | 3000 | ret = nbd_send_generic_reply(client, request, -EINVAL, msg, |
6f302e60 VSO |
3001 | errp); |
3002 | g_free(msg); | |
3003 | return ret; | |
3004 | } | |
3005 | } | |
3006 | ||
ff82911c PB |
3007 | /* Owns a reference to the NBDClient passed as opaque. */ |
3008 | static coroutine_fn void nbd_trip(void *opaque) | |
75818250 | 3009 | { |
262db388 | 3010 | NBDClient *client = opaque; |
f816310d | 3011 | NBDRequestData *req = NULL; |
ff82911c | 3012 | NBDRequest request = { 0 }; /* GCC thinks it can be used uninitialized */ |
a0dc63a6 | 3013 | int ret; |
2fd2c840 | 3014 | Error *local_err = NULL; |
b2e3d87f | 3015 | |
f816310d SH |
3016 | /* |
3017 | * Note that nbd_client_put() and client_close() must be called from the | |
3018 | * main loop thread. Use aio_co_reschedule_self() to switch AioContext | |
3019 | * before calling these functions. | |
3020 | */ | |
3021 | ||
9588463e | 3022 | trace_nbd_trip(); |
7075d235 SH |
3023 | |
3024 | qemu_mutex_lock(&client->lock); | |
3025 | ||
ff2b68aa | 3026 | if (client->closing) { |
f816310d | 3027 | goto done; |
ff2b68aa | 3028 | } |
b2e3d87f | 3029 | |
f148ae7d SL |
3030 | if (client->quiescing) { |
3031 | /* | |
3032 | * We're switching between AIO contexts. Don't attempt to receive a new | |
3033 | * request and kick the main context which may be waiting for us. | |
3034 | */ | |
f148ae7d SL |
3035 | client->recv_coroutine = NULL; |
3036 | aio_wait_kick(); | |
f816310d | 3037 | goto done; |
f148ae7d SL |
3038 | } |
3039 | ||
ff2b68aa | 3040 | req = nbd_request_get(client); |
7075d235 SH |
3041 | |
3042 | /* | |
3043 | * nbd_co_receive_request() returns -EAGAIN when nbd_drained_begin() has | |
3044 | * set client->quiescing but by the time we get back nbd_drained_end() may | |
3045 | * have already cleared client->quiescing. In that case we try again | |
3046 | * because nothing else will spawn an nbd_trip() coroutine until we set | |
3047 | * client->recv_coroutine = NULL further down. | |
3048 | */ | |
3049 | do { | |
3050 | assert(client->recv_coroutine == qemu_coroutine_self()); | |
3051 | qemu_mutex_unlock(&client->lock); | |
3052 | ret = nbd_co_receive_request(req, &request, &local_err); | |
3053 | qemu_mutex_lock(&client->lock); | |
3054 | } while (ret == -EAGAIN && !client->quiescing); | |
3055 | ||
ee898b87 | 3056 | client->recv_coroutine = NULL; |
b2e3d87f | 3057 | |
d6268348 WC |
3058 | if (client->closing) { |
3059 | /* | |
3060 | * The client may be closed when we are blocked in | |
3061 | * nbd_co_receive_request() | |
3062 | */ | |
3063 | goto done; | |
3064 | } | |
3065 | ||
f148ae7d | 3066 | if (ret == -EAGAIN) { |
f148ae7d SL |
3067 | goto done; |
3068 | } | |
3069 | ||
a0d7ce20 | 3070 | nbd_client_receive_next_request(client); |
7075d235 | 3071 | |
a0d7ce20 VSO |
3072 | if (ret == -EIO) { |
3073 | goto disconnect; | |
3074 | } | |
3075 | ||
7075d235 | 3076 | qemu_mutex_unlock(&client->lock); |
bd2cd4a4 FW |
3077 | qio_channel_set_cork(client->ioc, true); |
3078 | ||
a0d7ce20 | 3079 | if (ret < 0) { |
314b9026 | 3080 | /* It wasn't -EIO, so, according to nbd_co_receive_request() |
6a417599 VSO |
3081 | * semantics, we should return the error to the client. */ |
3082 | Error *export_err = local_err; | |
3083 | ||
3084 | local_err = NULL; | |
66d4f4fe | 3085 | ret = nbd_send_generic_reply(client, &request, -EINVAL, |
6a417599 VSO |
3086 | error_get_pretty(export_err), &local_err); |
3087 | error_free(export_err); | |
6f302e60 VSO |
3088 | } else { |
3089 | ret = nbd_handle_request(client, &request, req->data, &local_err); | |
5c54e7fa | 3090 | } |
1dec4643 EB |
3091 | if (request.contexts && request.contexts != &client->contexts) { |
3092 | assert(request.type == NBD_CMD_BLOCK_STATUS); | |
3093 | g_free(request.contexts->bitmaps); | |
3094 | g_free(request.contexts); | |
3095 | } | |
7075d235 SH |
3096 | |
3097 | qio_channel_set_cork(client->ioc, false); | |
3098 | qemu_mutex_lock(&client->lock); | |
3099 | ||
5c54e7fa | 3100 | if (ret < 0) { |
c7b97282 | 3101 | error_prepend(&local_err, "Failed to send reply: "); |
2fd2c840 VSO |
3102 | goto disconnect; |
3103 | } | |
3104 | ||
2dcbb11b EB |
3105 | /* |
3106 | * We must disconnect after NBD_CMD_WRITE or BLOCK_STATUS with | |
3107 | * payload if we did not read the payload. | |
8c372a02 | 3108 | */ |
2fd2c840 VSO |
3109 | if (!req->complete) { |
3110 | error_setg(&local_err, "Request handling failed in intermediate state"); | |
8c372a02 | 3111 | goto disconnect; |
b2e3d87f NT |
3112 | } |
3113 | ||
7fe7b68b | 3114 | done: |
f816310d SH |
3115 | if (req) { |
3116 | nbd_request_put(req); | |
3117 | } | |
7075d235 SH |
3118 | |
3119 | qemu_mutex_unlock(&client->lock); | |
3120 | ||
f816310d SH |
3121 | if (!nbd_client_put_nonzero(client)) { |
3122 | aio_co_reschedule_self(qemu_get_aio_context()); | |
3123 | nbd_client_put(client); | |
3124 | } | |
262db388 PB |
3125 | return; |
3126 | ||
8c372a02 | 3127 | disconnect: |
2fd2c840 VSO |
3128 | if (local_err) { |
3129 | error_reportf_err(local_err, "Disconnect client, due to: "); | |
3130 | } | |
7075d235 | 3131 | |
72deddc5 | 3132 | nbd_request_put(req); |
7075d235 | 3133 | qemu_mutex_unlock(&client->lock); |
f816310d SH |
3134 | |
3135 | aio_co_reschedule_self(qemu_get_aio_context()); | |
0c9390d9 | 3136 | client_close(client, true); |
ff82911c | 3137 | nbd_client_put(client); |
7a5ca864 | 3138 | } |
af49bbbe | 3139 | |
7075d235 SH |
3140 | /* |
3141 | * Runs in export AioContext and main loop thread. Caller must hold | |
3142 | * client->lock. | |
3143 | */ | |
ff82911c | 3144 | static void nbd_client_receive_next_request(NBDClient *client) |
958c717d | 3145 | { |
f148ae7d SL |
3146 | if (!client->recv_coroutine && client->nb_requests < MAX_NBD_REQUESTS && |
3147 | !client->quiescing) { | |
ff82911c PB |
3148 | nbd_client_get(client); |
3149 | client->recv_coroutine = qemu_coroutine_create(nbd_trip, client); | |
8612c686 | 3150 | aio_co_schedule(client->exp->common.ctx, client->recv_coroutine); |
958c717d HR |
3151 | } |
3152 | } | |
3153 | ||
1a6245a5 FZ |
3154 | static coroutine_fn void nbd_co_client_start(void *opaque) |
3155 | { | |
c84087f2 | 3156 | NBDClient *client = opaque; |
2fd2c840 | 3157 | Error *local_err = NULL; |
1a6245a5 | 3158 | |
df8ad9f1 EB |
3159 | qemu_co_mutex_init(&client->send_lock); |
3160 | ||
2fd2c840 VSO |
3161 | if (nbd_negotiate(client, &local_err)) { |
3162 | if (local_err) { | |
3163 | error_report_err(local_err); | |
3164 | } | |
0c9390d9 | 3165 | client_close(client, false); |
c84087f2 | 3166 | return; |
1a6245a5 | 3167 | } |
ff82911c | 3168 | |
7075d235 SH |
3169 | WITH_QEMU_LOCK_GUARD(&client->lock) { |
3170 | nbd_client_receive_next_request(client); | |
3171 | } | |
1a6245a5 FZ |
3172 | } |
3173 | ||
0c9390d9 | 3174 | /* |
7f7dfe2a VSO |
3175 | * Create a new client listener using the given channel @sioc. |
3176 | * Begin servicing it in a coroutine. When the connection closes, call | |
3177 | * @close_fn with an indication of whether the client completed negotiation. | |
0c9390d9 | 3178 | */ |
7f7dfe2a | 3179 | void nbd_client_new(QIOChannelSocket *sioc, |
f95910fe | 3180 | QCryptoTLSCreds *tlscreds, |
b25e12da | 3181 | const char *tlsauthz, |
0c9390d9 | 3182 | void (*close_fn)(NBDClient *, bool)) |
af49bbbe | 3183 | { |
1743b515 | 3184 | NBDClient *client; |
c84087f2 | 3185 | Coroutine *co; |
1a6245a5 | 3186 | |
e8d3eb74 | 3187 | client = g_new0(NBDClient, 1); |
7075d235 | 3188 | qemu_mutex_init(&client->lock); |
1743b515 | 3189 | client->refcount = 1; |
f95910fe DB |
3190 | client->tlscreds = tlscreds; |
3191 | if (tlscreds) { | |
3192 | object_ref(OBJECT(client->tlscreds)); | |
3193 | } | |
b25e12da | 3194 | client->tlsauthz = g_strdup(tlsauthz); |
1c778ef7 | 3195 | client->sioc = sioc; |
f1426881 | 3196 | qio_channel_set_delay(QIO_CHANNEL(sioc), false); |
1c778ef7 DB |
3197 | object_ref(OBJECT(client->sioc)); |
3198 | client->ioc = QIO_CHANNEL(sioc); | |
3199 | object_ref(OBJECT(client->ioc)); | |
0c9390d9 | 3200 | client->close_fn = close_fn; |
2c8d9f06 | 3201 | |
c84087f2 VSO |
3202 | co = qemu_coroutine_create(nbd_co_client_start, client); |
3203 | qemu_coroutine_enter(co); | |
af49bbbe | 3204 | } |