]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Network block device - make block devices work over TCP | |
3 | * | |
4 | * Note that you can not swap over this thing, yet. Seems to work but | |
5 | * deadlocks sometimes - you can not swap over TCP in general. | |
6 | * | |
a2531293 | 7 | * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz> |
1da177e4 LT |
8 | * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com> |
9 | * | |
dbf492d6 | 10 | * This file is released under GPLv2 or later. |
1da177e4 | 11 | * |
dbf492d6 | 12 | * (part of code stolen from loop.c) |
1da177e4 LT |
13 | */ |
14 | ||
15 | #include <linux/major.h> | |
16 | ||
17 | #include <linux/blkdev.h> | |
18 | #include <linux/module.h> | |
19 | #include <linux/init.h> | |
20 | #include <linux/sched.h> | |
21 | #include <linux/fs.h> | |
22 | #include <linux/bio.h> | |
23 | #include <linux/stat.h> | |
24 | #include <linux/errno.h> | |
25 | #include <linux/file.h> | |
26 | #include <linux/ioctl.h> | |
2a48fc0a | 27 | #include <linux/mutex.h> |
4b2f0260 HX |
28 | #include <linux/compiler.h> |
29 | #include <linux/err.h> | |
30 | #include <linux/kernel.h> | |
5a0e3ad6 | 31 | #include <linux/slab.h> |
1da177e4 | 32 | #include <net/sock.h> |
91cf45f0 | 33 | #include <linux/net.h> |
48cf6061 | 34 | #include <linux/kthread.h> |
b9c495bb | 35 | #include <linux/types.h> |
30d53d9c | 36 | #include <linux/debugfs.h> |
fd8383fd | 37 | #include <linux/blk-mq.h> |
1da177e4 | 38 | |
7c0f6ba6 | 39 | #include <linux/uaccess.h> |
1da177e4 LT |
40 | #include <asm/types.h> |
41 | ||
42 | #include <linux/nbd.h> | |
43 | ||
9561a7ad JB |
44 | struct nbd_sock { |
45 | struct socket *sock; | |
46 | struct mutex tx_lock; | |
47 | }; | |
48 | ||
9b4a6ba9 JB |
49 | #define NBD_TIMEDOUT 0 |
50 | #define NBD_DISCONNECT_REQUESTED 1 | |
9561a7ad JB |
51 | #define NBD_DISCONNECTED 2 |
52 | #define NBD_RUNNING 3 | |
9b4a6ba9 | 53 | |
13e71d69 | 54 | struct nbd_device { |
22d109c1 | 55 | u32 flags; |
9b4a6ba9 | 56 | unsigned long runtime_flags; |
9561a7ad | 57 | struct nbd_sock **socks; |
13e71d69 MP |
58 | int magic; |
59 | ||
fd8383fd | 60 | struct blk_mq_tag_set tag_set; |
13e71d69 | 61 | |
9561a7ad | 62 | struct mutex config_lock; |
13e71d69 | 63 | struct gendisk *disk; |
9561a7ad JB |
64 | int num_connections; |
65 | atomic_t recv_threads; | |
66 | wait_queue_head_t recv_wq; | |
ef77b515 | 67 | loff_t blksize; |
b9c495bb | 68 | loff_t bytesize; |
7e2893a1 | 69 | |
7e2893a1 | 70 | struct task_struct *task_recv; |
9561a7ad | 71 | struct task_struct *task_setup; |
30d53d9c MP |
72 | |
73 | #if IS_ENABLED(CONFIG_DEBUG_FS) | |
74 | struct dentry *dbg_dir; | |
75 | #endif | |
13e71d69 MP |
76 | }; |
77 | ||
fd8383fd JB |
78 | struct nbd_cmd { |
79 | struct nbd_device *nbd; | |
9561a7ad | 80 | struct completion send_complete; |
fd8383fd JB |
81 | }; |
82 | ||
30d53d9c MP |
83 | #if IS_ENABLED(CONFIG_DEBUG_FS) |
84 | static struct dentry *nbd_dbg_dir; | |
85 | #endif | |
86 | ||
87 | #define nbd_name(nbd) ((nbd)->disk->disk_name) | |
88 | ||
f4507164 | 89 | #define NBD_MAGIC 0x68797548 |
1da177e4 | 90 | |
9c7a4169 | 91 | static unsigned int nbds_max = 16; |
20a8143e | 92 | static struct nbd_device *nbd_dev; |
d71a6d73 | 93 | static int max_part; |
1da177e4 | 94 | |
d18509f5 | 95 | static inline struct device *nbd_to_dev(struct nbd_device *nbd) |
1da177e4 | 96 | { |
d18509f5 | 97 | return disk_to_dev(nbd->disk); |
1da177e4 LT |
98 | } |
99 | ||
37091fdd MP |
100 | static bool nbd_is_connected(struct nbd_device *nbd) |
101 | { | |
102 | return !!nbd->task_recv; | |
103 | } | |
104 | ||
1da177e4 LT |
105 | static const char *nbdcmd_to_ascii(int cmd) |
106 | { | |
107 | switch (cmd) { | |
108 | case NBD_CMD_READ: return "read"; | |
109 | case NBD_CMD_WRITE: return "write"; | |
110 | case NBD_CMD_DISC: return "disconnect"; | |
75f187ab | 111 | case NBD_CMD_FLUSH: return "flush"; |
a336d298 | 112 | case NBD_CMD_TRIM: return "trim/discard"; |
1da177e4 LT |
113 | } |
114 | return "invalid"; | |
115 | } | |
1da177e4 | 116 | |
37091fdd MP |
117 | static int nbd_size_clear(struct nbd_device *nbd, struct block_device *bdev) |
118 | { | |
119 | bdev->bd_inode->i_size = 0; | |
120 | set_capacity(nbd->disk, 0); | |
121 | kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE); | |
122 | ||
123 | return 0; | |
124 | } | |
125 | ||
126 | static void nbd_size_update(struct nbd_device *nbd, struct block_device *bdev) | |
127 | { | |
128 | if (!nbd_is_connected(nbd)) | |
129 | return; | |
130 | ||
131 | bdev->bd_inode->i_size = nbd->bytesize; | |
132 | set_capacity(nbd->disk, nbd->bytesize >> 9); | |
133 | kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE); | |
134 | } | |
135 | ||
136 | static int nbd_size_set(struct nbd_device *nbd, struct block_device *bdev, | |
ef77b515 | 137 | loff_t blocksize, loff_t nr_blocks) |
37091fdd MP |
138 | { |
139 | int ret; | |
140 | ||
141 | ret = set_blocksize(bdev, blocksize); | |
142 | if (ret) | |
143 | return ret; | |
144 | ||
145 | nbd->blksize = blocksize; | |
ef77b515 | 146 | nbd->bytesize = blocksize * nr_blocks; |
37091fdd MP |
147 | |
148 | nbd_size_update(nbd, bdev); | |
149 | ||
150 | return 0; | |
151 | } | |
152 | ||
fd8383fd | 153 | static void nbd_end_request(struct nbd_cmd *cmd) |
1da177e4 | 154 | { |
fd8383fd JB |
155 | struct nbd_device *nbd = cmd->nbd; |
156 | struct request *req = blk_mq_rq_from_pdu(cmd); | |
097c94a4 | 157 | int error = req->errors ? -EIO : 0; |
1da177e4 | 158 | |
fd8383fd | 159 | dev_dbg(nbd_to_dev(nbd), "request %p: %s\n", cmd, |
d18509f5 | 160 | error ? "failed" : "done"); |
1da177e4 | 161 | |
fd8383fd | 162 | blk_mq_complete_request(req, error); |
1da177e4 LT |
163 | } |
164 | ||
e018e757 MP |
165 | /* |
166 | * Forcibly shutdown the socket causing all listeners to error | |
167 | */ | |
36e47bee | 168 | static void sock_shutdown(struct nbd_device *nbd) |
7fdfd406 | 169 | { |
9561a7ad | 170 | int i; |
23272a67 | 171 | |
9561a7ad JB |
172 | if (nbd->num_connections == 0) |
173 | return; | |
174 | if (test_and_set_bit(NBD_DISCONNECTED, &nbd->runtime_flags)) | |
260bbce4 | 175 | return; |
23272a67 | 176 | |
9561a7ad JB |
177 | for (i = 0; i < nbd->num_connections; i++) { |
178 | struct nbd_sock *nsock = nbd->socks[i]; | |
179 | mutex_lock(&nsock->tx_lock); | |
180 | kernel_sock_shutdown(nsock->sock, SHUT_RDWR); | |
181 | mutex_unlock(&nsock->tx_lock); | |
182 | } | |
183 | dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n"); | |
7fdfd406 PC |
184 | } |
185 | ||
0eadf37a JB |
186 | static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req, |
187 | bool reserved) | |
7fdfd406 | 188 | { |
0eadf37a JB |
189 | struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req); |
190 | struct nbd_device *nbd = cmd->nbd; | |
dcc909d9 | 191 | |
9561a7ad | 192 | dev_err(nbd_to_dev(nbd), "Connection timed out, shutting down connection\n"); |
9b4a6ba9 | 193 | set_bit(NBD_TIMEDOUT, &nbd->runtime_flags); |
0eadf37a | 194 | req->errors++; |
9561a7ad | 195 | |
9561a7ad JB |
196 | mutex_lock(&nbd->config_lock); |
197 | sock_shutdown(nbd); | |
198 | mutex_unlock(&nbd->config_lock); | |
0eadf37a | 199 | return BLK_EH_HANDLED; |
7fdfd406 PC |
200 | } |
201 | ||
1da177e4 LT |
202 | /* |
203 | * Send or receive packet. | |
204 | */ | |
9561a7ad JB |
205 | static int sock_xmit(struct nbd_device *nbd, int index, int send, void *buf, |
206 | int size, int msg_flags) | |
1da177e4 | 207 | { |
9561a7ad | 208 | struct socket *sock = nbd->socks[index]->sock; |
1da177e4 LT |
209 | int result; |
210 | struct msghdr msg; | |
211 | struct kvec iov; | |
7f338fe4 | 212 | unsigned long pflags = current->flags; |
1da177e4 | 213 | |
ffc41cf8 | 214 | if (unlikely(!sock)) { |
a897b666 | 215 | dev_err_ratelimited(disk_to_dev(nbd->disk), |
7f1b90f9 WC |
216 | "Attempted %s on closed socket in sock_xmit\n", |
217 | (send ? "send" : "recv")); | |
ffc41cf8 MS |
218 | return -EINVAL; |
219 | } | |
220 | ||
7f338fe4 | 221 | current->flags |= PF_MEMALLOC; |
1da177e4 | 222 | do { |
7f338fe4 | 223 | sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC; |
1da177e4 LT |
224 | iov.iov_base = buf; |
225 | iov.iov_len = size; | |
226 | msg.msg_name = NULL; | |
227 | msg.msg_namelen = 0; | |
228 | msg.msg_control = NULL; | |
229 | msg.msg_controllen = 0; | |
1da177e4 LT |
230 | msg.msg_flags = msg_flags | MSG_NOSIGNAL; |
231 | ||
7e2893a1 | 232 | if (send) |
1da177e4 | 233 | result = kernel_sendmsg(sock, &msg, &iov, 1, size); |
7e2893a1 | 234 | else |
35fbf5bc NK |
235 | result = kernel_recvmsg(sock, &msg, &iov, 1, size, |
236 | msg.msg_flags); | |
1da177e4 | 237 | |
1da177e4 LT |
238 | if (result <= 0) { |
239 | if (result == 0) | |
240 | result = -EPIPE; /* short read */ | |
241 | break; | |
242 | } | |
243 | size -= result; | |
244 | buf += result; | |
245 | } while (size > 0); | |
246 | ||
7f338fe4 | 247 | tsk_restore_flags(current, pflags, PF_MEMALLOC); |
1da177e4 LT |
248 | |
249 | return result; | |
250 | } | |
251 | ||
9561a7ad JB |
252 | static inline int sock_send_bvec(struct nbd_device *nbd, int index, |
253 | struct bio_vec *bvec, int flags) | |
1da177e4 LT |
254 | { |
255 | int result; | |
256 | void *kaddr = kmap(bvec->bv_page); | |
9561a7ad | 257 | result = sock_xmit(nbd, index, 1, kaddr + bvec->bv_offset, |
f4507164 | 258 | bvec->bv_len, flags); |
1da177e4 LT |
259 | kunmap(bvec->bv_page); |
260 | return result; | |
261 | } | |
262 | ||
7fdfd406 | 263 | /* always call with the tx_lock held */ |
9561a7ad | 264 | static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index) |
1da177e4 | 265 | { |
fd8383fd | 266 | struct request *req = blk_mq_rq_from_pdu(cmd); |
d61b7f97 | 267 | int result; |
1da177e4 | 268 | struct nbd_request request; |
1011c1b9 | 269 | unsigned long size = blk_rq_bytes(req); |
429a787b | 270 | struct bio *bio; |
9dc6c806 | 271 | u32 type; |
9561a7ad | 272 | u32 tag = blk_mq_unique_tag(req); |
9dc6c806 | 273 | |
aebf526b CH |
274 | switch (req_op(req)) { |
275 | case REQ_OP_DISCARD: | |
9dc6c806 | 276 | type = NBD_CMD_TRIM; |
aebf526b CH |
277 | break; |
278 | case REQ_OP_FLUSH: | |
9dc6c806 | 279 | type = NBD_CMD_FLUSH; |
aebf526b CH |
280 | break; |
281 | case REQ_OP_WRITE: | |
9dc6c806 | 282 | type = NBD_CMD_WRITE; |
aebf526b CH |
283 | break; |
284 | case REQ_OP_READ: | |
9dc6c806 | 285 | type = NBD_CMD_READ; |
aebf526b CH |
286 | break; |
287 | default: | |
288 | return -EIO; | |
289 | } | |
1da177e4 | 290 | |
09fc54cc CH |
291 | if (rq_data_dir(req) == WRITE && |
292 | (nbd->flags & NBD_FLAG_READ_ONLY)) { | |
293 | dev_err_ratelimited(disk_to_dev(nbd->disk), | |
294 | "Write on read-only\n"); | |
295 | return -EIO; | |
296 | } | |
297 | ||
04cfac4e | 298 | memset(&request, 0, sizeof(request)); |
1da177e4 | 299 | request.magic = htonl(NBD_REQUEST_MAGIC); |
9dc6c806 | 300 | request.type = htonl(type); |
9561a7ad | 301 | if (type != NBD_CMD_FLUSH) { |
75f187ab AB |
302 | request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9); |
303 | request.len = htonl(size); | |
304 | } | |
9561a7ad | 305 | memcpy(request.handle, &tag, sizeof(tag)); |
1da177e4 | 306 | |
d18509f5 | 307 | dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n", |
fd8383fd | 308 | cmd, nbdcmd_to_ascii(type), |
d18509f5 | 309 | (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req)); |
9561a7ad | 310 | result = sock_xmit(nbd, index, 1, &request, sizeof(request), |
9dc6c806 | 311 | (type == NBD_CMD_WRITE) ? MSG_MORE : 0); |
1da177e4 | 312 | if (result <= 0) { |
a897b666 | 313 | dev_err_ratelimited(disk_to_dev(nbd->disk), |
7f1b90f9 | 314 | "Send control failed (result %d)\n", result); |
dab5313a | 315 | return -EIO; |
1da177e4 LT |
316 | } |
317 | ||
429a787b JA |
318 | if (type != NBD_CMD_WRITE) |
319 | return 0; | |
320 | ||
429a787b JA |
321 | bio = req->bio; |
322 | while (bio) { | |
323 | struct bio *next = bio->bi_next; | |
324 | struct bvec_iter iter; | |
7988613b | 325 | struct bio_vec bvec; |
429a787b JA |
326 | |
327 | bio_for_each_segment(bvec, bio, iter) { | |
328 | bool is_last = !next && bio_iter_last(bvec, iter); | |
d61b7f97 | 329 | int flags = is_last ? 0 : MSG_MORE; |
429a787b | 330 | |
d18509f5 | 331 | dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n", |
fd8383fd | 332 | cmd, bvec.bv_len); |
9561a7ad | 333 | result = sock_send_bvec(nbd, index, &bvec, flags); |
6c92e699 | 334 | if (result <= 0) { |
f4507164 | 335 | dev_err(disk_to_dev(nbd->disk), |
7f1b90f9 WC |
336 | "Send data failed (result %d)\n", |
337 | result); | |
dab5313a | 338 | return -EIO; |
6c92e699 | 339 | } |
429a787b JA |
340 | /* |
341 | * The completion might already have come in, | |
342 | * so break for the last one instead of letting | |
343 | * the iterator do it. This prevents use-after-free | |
344 | * of the bio. | |
345 | */ | |
346 | if (is_last) | |
347 | break; | |
1da177e4 | 348 | } |
429a787b | 349 | bio = next; |
1da177e4 | 350 | } |
1da177e4 | 351 | return 0; |
1da177e4 LT |
352 | } |
353 | ||
9561a7ad JB |
354 | static inline int sock_recv_bvec(struct nbd_device *nbd, int index, |
355 | struct bio_vec *bvec) | |
1da177e4 LT |
356 | { |
357 | int result; | |
358 | void *kaddr = kmap(bvec->bv_page); | |
9561a7ad JB |
359 | result = sock_xmit(nbd, index, 0, kaddr + bvec->bv_offset, |
360 | bvec->bv_len, MSG_WAITALL); | |
1da177e4 LT |
361 | kunmap(bvec->bv_page); |
362 | return result; | |
363 | } | |
364 | ||
365 | /* NULL returned = something went wrong, inform userspace */ | |
9561a7ad | 366 | static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index) |
1da177e4 LT |
367 | { |
368 | int result; | |
369 | struct nbd_reply reply; | |
fd8383fd JB |
370 | struct nbd_cmd *cmd; |
371 | struct request *req = NULL; | |
372 | u16 hwq; | |
9561a7ad | 373 | u32 tag; |
1da177e4 LT |
374 | |
375 | reply.magic = 0; | |
9561a7ad | 376 | result = sock_xmit(nbd, index, 0, &reply, sizeof(reply), MSG_WAITALL); |
1da177e4 | 377 | if (result <= 0) { |
9561a7ad JB |
378 | if (!test_bit(NBD_DISCONNECTED, &nbd->runtime_flags) && |
379 | !test_bit(NBD_DISCONNECT_REQUESTED, &nbd->runtime_flags)) | |
380 | dev_err(disk_to_dev(nbd->disk), | |
381 | "Receive control failed (result %d)\n", result); | |
19391830 | 382 | return ERR_PTR(result); |
1da177e4 | 383 | } |
e4b57e08 MF |
384 | |
385 | if (ntohl(reply.magic) != NBD_REPLY_MAGIC) { | |
f4507164 | 386 | dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n", |
e4b57e08 | 387 | (unsigned long)ntohl(reply.magic)); |
19391830 | 388 | return ERR_PTR(-EPROTO); |
e4b57e08 MF |
389 | } |
390 | ||
9561a7ad | 391 | memcpy(&tag, reply.handle, sizeof(u32)); |
4b2f0260 | 392 | |
fd8383fd JB |
393 | hwq = blk_mq_unique_tag_to_hwq(tag); |
394 | if (hwq < nbd->tag_set.nr_hw_queues) | |
395 | req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq], | |
396 | blk_mq_unique_tag_to_tag(tag)); | |
397 | if (!req || !blk_mq_request_started(req)) { | |
398 | dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n", | |
399 | tag, req); | |
400 | return ERR_PTR(-ENOENT); | |
1da177e4 | 401 | } |
fd8383fd | 402 | cmd = blk_mq_rq_to_pdu(req); |
1da177e4 | 403 | if (ntohl(reply.error)) { |
f4507164 | 404 | dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n", |
7f1b90f9 | 405 | ntohl(reply.error)); |
1da177e4 | 406 | req->errors++; |
fd8383fd | 407 | return cmd; |
1da177e4 LT |
408 | } |
409 | ||
fd8383fd | 410 | dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", cmd); |
9dc6c806 | 411 | if (rq_data_dir(req) != WRITE) { |
5705f702 | 412 | struct req_iterator iter; |
7988613b | 413 | struct bio_vec bvec; |
5705f702 N |
414 | |
415 | rq_for_each_segment(bvec, req, iter) { | |
9561a7ad | 416 | result = sock_recv_bvec(nbd, index, &bvec); |
6c92e699 | 417 | if (result <= 0) { |
f4507164 | 418 | dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n", |
7f1b90f9 | 419 | result); |
6c92e699 | 420 | req->errors++; |
fd8383fd | 421 | return cmd; |
6c92e699 | 422 | } |
d18509f5 | 423 | dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n", |
fd8383fd | 424 | cmd, bvec.bv_len); |
1da177e4 | 425 | } |
9561a7ad JB |
426 | } else { |
427 | /* See the comment in nbd_queue_rq. */ | |
428 | wait_for_completion(&cmd->send_complete); | |
1da177e4 | 429 | } |
fd8383fd | 430 | return cmd; |
1da177e4 LT |
431 | } |
432 | ||
edfaa7c3 KS |
433 | static ssize_t pid_show(struct device *dev, |
434 | struct device_attribute *attr, char *buf) | |
6b39bb65 | 435 | { |
edfaa7c3 | 436 | struct gendisk *disk = dev_to_disk(dev); |
6521d39a | 437 | struct nbd_device *nbd = (struct nbd_device *)disk->private_data; |
edfaa7c3 | 438 | |
6521d39a | 439 | return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv)); |
6b39bb65 PC |
440 | } |
441 | ||
edfaa7c3 | 442 | static struct device_attribute pid_attr = { |
01e8ef11 | 443 | .attr = { .name = "pid", .mode = S_IRUGO}, |
6b39bb65 PC |
444 | .show = pid_show, |
445 | }; | |
446 | ||
9561a7ad JB |
447 | struct recv_thread_args { |
448 | struct work_struct work; | |
449 | struct nbd_device *nbd; | |
450 | int index; | |
451 | }; | |
452 | ||
453 | static void recv_work(struct work_struct *work) | |
1da177e4 | 454 | { |
9561a7ad JB |
455 | struct recv_thread_args *args = container_of(work, |
456 | struct recv_thread_args, | |
457 | work); | |
458 | struct nbd_device *nbd = args->nbd; | |
fd8383fd | 459 | struct nbd_cmd *cmd; |
9561a7ad | 460 | int ret = 0; |
1da177e4 | 461 | |
f4507164 | 462 | BUG_ON(nbd->magic != NBD_MAGIC); |
19391830 | 463 | while (1) { |
9561a7ad | 464 | cmd = nbd_read_stat(nbd, args->index); |
fd8383fd JB |
465 | if (IS_ERR(cmd)) { |
466 | ret = PTR_ERR(cmd); | |
19391830 MP |
467 | break; |
468 | } | |
469 | ||
fd8383fd | 470 | nbd_end_request(cmd); |
19391830 | 471 | } |
6b39bb65 | 472 | |
9561a7ad JB |
473 | /* |
474 | * We got an error, shut everybody down if this wasn't the result of a | |
475 | * disconnect request. | |
476 | */ | |
477 | if (ret && !test_bit(NBD_DISCONNECT_REQUESTED, &nbd->runtime_flags)) | |
478 | sock_shutdown(nbd); | |
479 | atomic_dec(&nbd->recv_threads); | |
480 | wake_up(&nbd->recv_wq); | |
1da177e4 LT |
481 | } |
482 | ||
fd8383fd | 483 | static void nbd_clear_req(struct request *req, void *data, bool reserved) |
1da177e4 | 484 | { |
fd8383fd | 485 | struct nbd_cmd *cmd; |
1da177e4 | 486 | |
fd8383fd JB |
487 | if (!blk_mq_request_started(req)) |
488 | return; | |
489 | cmd = blk_mq_rq_to_pdu(req); | |
490 | req->errors++; | |
491 | nbd_end_request(cmd); | |
492 | } | |
493 | ||
494 | static void nbd_clear_que(struct nbd_device *nbd) | |
495 | { | |
f4507164 | 496 | BUG_ON(nbd->magic != NBD_MAGIC); |
1da177e4 | 497 | |
fd8383fd | 498 | blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL); |
e78273c8 | 499 | dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n"); |
1da177e4 LT |
500 | } |
501 | ||
7fdfd406 | 502 | |
9561a7ad | 503 | static void nbd_handle_cmd(struct nbd_cmd *cmd, int index) |
48cf6061 | 504 | { |
fd8383fd JB |
505 | struct request *req = blk_mq_rq_from_pdu(cmd); |
506 | struct nbd_device *nbd = cmd->nbd; | |
9561a7ad | 507 | struct nbd_sock *nsock; |
fd8383fd | 508 | |
9561a7ad | 509 | if (index >= nbd->num_connections) { |
a897b666 JB |
510 | dev_err_ratelimited(disk_to_dev(nbd->disk), |
511 | "Attempted send on invalid socket\n"); | |
48cf6061 | 512 | goto error_out; |
9561a7ad | 513 | } |
48cf6061 | 514 | |
9561a7ad | 515 | if (test_bit(NBD_DISCONNECTED, &nbd->runtime_flags)) { |
a897b666 JB |
516 | dev_err_ratelimited(disk_to_dev(nbd->disk), |
517 | "Attempted send on closed socket\n"); | |
9561a7ad JB |
518 | goto error_out; |
519 | } | |
520 | ||
48cf6061 LV |
521 | req->errors = 0; |
522 | ||
9561a7ad JB |
523 | nsock = nbd->socks[index]; |
524 | mutex_lock(&nsock->tx_lock); | |
525 | if (unlikely(!nsock->sock)) { | |
526 | mutex_unlock(&nsock->tx_lock); | |
a897b666 JB |
527 | dev_err_ratelimited(disk_to_dev(nbd->disk), |
528 | "Attempted send on closed socket\n"); | |
15746fca | 529 | goto error_out; |
48cf6061 LV |
530 | } |
531 | ||
9561a7ad | 532 | if (nbd_send_cmd(nbd, cmd, index) != 0) { |
a897b666 JB |
533 | dev_err_ratelimited(disk_to_dev(nbd->disk), |
534 | "Request send failed\n"); | |
48cf6061 | 535 | req->errors++; |
fd8383fd | 536 | nbd_end_request(cmd); |
48cf6061 LV |
537 | } |
538 | ||
9561a7ad | 539 | mutex_unlock(&nsock->tx_lock); |
48cf6061 LV |
540 | |
541 | return; | |
542 | ||
543 | error_out: | |
544 | req->errors++; | |
fd8383fd | 545 | nbd_end_request(cmd); |
48cf6061 LV |
546 | } |
547 | ||
fd8383fd JB |
548 | static int nbd_queue_rq(struct blk_mq_hw_ctx *hctx, |
549 | const struct blk_mq_queue_data *bd) | |
1da177e4 | 550 | { |
fd8383fd | 551 | struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq); |
1da177e4 | 552 | |
9561a7ad JB |
553 | /* |
554 | * Since we look at the bio's to send the request over the network we | |
555 | * need to make sure the completion work doesn't mark this request done | |
556 | * before we are done doing our send. This keeps us from dereferencing | |
557 | * freed data if we have particularly fast completions (ie we get the | |
558 | * completion before we exit sock_xmit on the last bvec) or in the case | |
559 | * that the server is misbehaving (or there was an error) before we're | |
560 | * done sending everything over the wire. | |
561 | */ | |
562 | init_completion(&cmd->send_complete); | |
fd8383fd | 563 | blk_mq_start_request(bd->rq); |
9561a7ad JB |
564 | nbd_handle_cmd(cmd, hctx->queue_num); |
565 | complete(&cmd->send_complete); | |
566 | ||
fd8383fd | 567 | return BLK_MQ_RQ_QUEUE_OK; |
1da177e4 LT |
568 | } |
569 | ||
9561a7ad | 570 | static int nbd_add_socket(struct nbd_device *nbd, struct socket *sock) |
23272a67 | 571 | { |
9561a7ad JB |
572 | struct nbd_sock **socks; |
573 | struct nbd_sock *nsock; | |
23272a67 | 574 | |
9561a7ad JB |
575 | if (!nbd->task_setup) |
576 | nbd->task_setup = current; | |
577 | if (nbd->task_setup != current) { | |
578 | dev_err(disk_to_dev(nbd->disk), | |
579 | "Device being setup by another task"); | |
580 | return -EINVAL; | |
23272a67 MP |
581 | } |
582 | ||
9561a7ad JB |
583 | socks = krealloc(nbd->socks, (nbd->num_connections + 1) * |
584 | sizeof(struct nbd_sock *), GFP_KERNEL); | |
585 | if (!socks) | |
586 | return -ENOMEM; | |
587 | nsock = kzalloc(sizeof(struct nbd_sock), GFP_KERNEL); | |
588 | if (!nsock) | |
589 | return -ENOMEM; | |
590 | ||
591 | nbd->socks = socks; | |
23272a67 | 592 | |
9561a7ad JB |
593 | mutex_init(&nsock->tx_lock); |
594 | nsock->sock = sock; | |
595 | socks[nbd->num_connections++] = nsock; | |
23272a67 | 596 | |
9561a7ad | 597 | return 0; |
23272a67 MP |
598 | } |
599 | ||
0e4f0f6f MP |
600 | /* Reset all properties of an NBD device */ |
601 | static void nbd_reset(struct nbd_device *nbd) | |
602 | { | |
9561a7ad JB |
603 | int i; |
604 | ||
605 | for (i = 0; i < nbd->num_connections; i++) | |
606 | kfree(nbd->socks[i]); | |
607 | kfree(nbd->socks); | |
608 | nbd->socks = NULL; | |
9b4a6ba9 | 609 | nbd->runtime_flags = 0; |
0e4f0f6f MP |
610 | nbd->blksize = 1024; |
611 | nbd->bytesize = 0; | |
612 | set_capacity(nbd->disk, 0); | |
613 | nbd->flags = 0; | |
0eadf37a | 614 | nbd->tag_set.timeout = 0; |
9561a7ad JB |
615 | nbd->num_connections = 0; |
616 | nbd->task_setup = NULL; | |
0e4f0f6f | 617 | queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue); |
0e4f0f6f MP |
618 | } |
619 | ||
620 | static void nbd_bdev_reset(struct block_device *bdev) | |
621 | { | |
622 | set_device_ro(bdev, false); | |
623 | bdev->bd_inode->i_size = 0; | |
624 | if (max_part > 0) { | |
625 | blkdev_reread_part(bdev); | |
626 | bdev->bd_invalidated = 1; | |
627 | } | |
628 | } | |
629 | ||
d02cf531 MP |
630 | static void nbd_parse_flags(struct nbd_device *nbd, struct block_device *bdev) |
631 | { | |
632 | if (nbd->flags & NBD_FLAG_READ_ONLY) | |
633 | set_device_ro(bdev, true); | |
634 | if (nbd->flags & NBD_FLAG_SEND_TRIM) | |
635 | queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue); | |
636 | if (nbd->flags & NBD_FLAG_SEND_FLUSH) | |
aafb1eec | 637 | blk_queue_write_cache(nbd->disk->queue, true, false); |
d02cf531 | 638 | else |
aafb1eec | 639 | blk_queue_write_cache(nbd->disk->queue, false, false); |
d02cf531 MP |
640 | } |
641 | ||
9561a7ad JB |
642 | static void send_disconnects(struct nbd_device *nbd) |
643 | { | |
644 | struct nbd_request request = {}; | |
645 | int i, ret; | |
646 | ||
647 | request.magic = htonl(NBD_REQUEST_MAGIC); | |
648 | request.type = htonl(NBD_CMD_DISC); | |
649 | ||
650 | for (i = 0; i < nbd->num_connections; i++) { | |
651 | ret = sock_xmit(nbd, i, 1, &request, sizeof(request), 0); | |
652 | if (ret <= 0) | |
653 | dev_err(disk_to_dev(nbd->disk), | |
654 | "Send disconnect failed %d\n", ret); | |
655 | } | |
656 | } | |
657 | ||
30d53d9c MP |
658 | static int nbd_dev_dbg_init(struct nbd_device *nbd); |
659 | static void nbd_dev_dbg_close(struct nbd_device *nbd); | |
660 | ||
9561a7ad | 661 | /* Must be called with config_lock held */ |
f4507164 | 662 | static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd, |
1a2ad211 PM |
663 | unsigned int cmd, unsigned long arg) |
664 | { | |
1da177e4 | 665 | switch (cmd) { |
1a2ad211 | 666 | case NBD_DISCONNECT: { |
f4507164 | 667 | dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n"); |
9561a7ad | 668 | if (!nbd->socks) |
3a2d63f8 | 669 | return -EINVAL; |
1a2ad211 | 670 | |
9561a7ad | 671 | mutex_unlock(&nbd->config_lock); |
3a2d63f8 | 672 | fsync_bdev(bdev); |
9561a7ad | 673 | mutex_lock(&nbd->config_lock); |
3a2d63f8 PB |
674 | |
675 | /* Check again after getting mutex back. */ | |
9561a7ad | 676 | if (!nbd->socks) |
1da177e4 | 677 | return -EINVAL; |
c378f70a | 678 | |
9561a7ad JB |
679 | if (!test_and_set_bit(NBD_DISCONNECT_REQUESTED, |
680 | &nbd->runtime_flags)) | |
681 | send_disconnects(nbd); | |
c378f70a | 682 | return 0; |
1a2ad211 | 683 | } |
9561a7ad | 684 | |
23272a67 MP |
685 | case NBD_CLEAR_SOCK: |
686 | sock_shutdown(nbd); | |
f4507164 | 687 | nbd_clear_que(nbd); |
3a2d63f8 | 688 | kill_bdev(bdev); |
9561a7ad JB |
689 | nbd_bdev_reset(bdev); |
690 | /* | |
691 | * We want to give the run thread a chance to wait for everybody | |
692 | * to clean up and then do it's own cleanup. | |
693 | */ | |
694 | if (!test_bit(NBD_RUNNING, &nbd->runtime_flags)) { | |
695 | int i; | |
696 | ||
697 | for (i = 0; i < nbd->num_connections; i++) | |
698 | kfree(nbd->socks[i]); | |
699 | kfree(nbd->socks); | |
700 | nbd->socks = NULL; | |
701 | nbd->num_connections = 0; | |
20032ec3 | 702 | nbd->task_setup = NULL; |
9561a7ad | 703 | } |
1a2ad211 | 704 | return 0; |
1a2ad211 PM |
705 | |
706 | case NBD_SET_SOCK: { | |
e2511578 | 707 | int err; |
23272a67 MP |
708 | struct socket *sock = sockfd_lookup(arg, &err); |
709 | ||
710 | if (!sock) | |
711 | return err; | |
712 | ||
9561a7ad | 713 | err = nbd_add_socket(nbd, sock); |
23272a67 MP |
714 | if (!err && max_part) |
715 | bdev->bd_invalidated = 1; | |
716 | ||
717 | return err; | |
1a2ad211 PM |
718 | } |
719 | ||
37091fdd | 720 | case NBD_SET_BLKSIZE: { |
5e454c67 | 721 | loff_t bsize = div_s64(nbd->bytesize, arg); |
37091fdd MP |
722 | |
723 | return nbd_size_set(nbd, bdev, arg, bsize); | |
724 | } | |
1a2ad211 | 725 | |
1da177e4 | 726 | case NBD_SET_SIZE: |
37091fdd | 727 | return nbd_size_set(nbd, bdev, nbd->blksize, |
e88f72cb | 728 | div_s64(arg, nbd->blksize)); |
37091fdd MP |
729 | |
730 | case NBD_SET_SIZE_BLOCKS: | |
731 | return nbd_size_set(nbd, bdev, nbd->blksize, arg); | |
1a2ad211 | 732 | |
7fdfd406 | 733 | case NBD_SET_TIMEOUT: |
0eadf37a | 734 | nbd->tag_set.timeout = arg * HZ; |
7fdfd406 | 735 | return 0; |
1a2ad211 | 736 | |
2f012508 PC |
737 | case NBD_SET_FLAGS: |
738 | nbd->flags = arg; | |
739 | return 0; | |
740 | ||
1a2ad211 | 741 | case NBD_DO_IT: { |
9561a7ad JB |
742 | struct recv_thread_args *args; |
743 | int num_connections = nbd->num_connections; | |
feffa5cc | 744 | int error = 0, i; |
1a2ad211 | 745 | |
6521d39a | 746 | if (nbd->task_recv) |
c91192d6 | 747 | return -EBUSY; |
9561a7ad | 748 | if (!nbd->socks) |
1da177e4 | 749 | return -EINVAL; |
9561a7ad JB |
750 | if (num_connections > 1 && |
751 | !(nbd->flags & NBD_FLAG_CAN_MULTI_CONN)) { | |
752 | dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n"); | |
feffa5cc | 753 | error = -EINVAL; |
9561a7ad JB |
754 | goto out_err; |
755 | } | |
1a2ad211 | 756 | |
9561a7ad JB |
757 | set_bit(NBD_RUNNING, &nbd->runtime_flags); |
758 | blk_mq_update_nr_hw_queues(&nbd->tag_set, nbd->num_connections); | |
759 | args = kcalloc(num_connections, sizeof(*args), GFP_KERNEL); | |
feffa5cc JA |
760 | if (!args) { |
761 | error = -ENOMEM; | |
9561a7ad | 762 | goto out_err; |
feffa5cc | 763 | } |
97240963 | 764 | nbd->task_recv = current; |
9561a7ad | 765 | mutex_unlock(&nbd->config_lock); |
1a2ad211 | 766 | |
d02cf531 | 767 | nbd_parse_flags(nbd, bdev); |
a336d298 | 768 | |
9561a7ad JB |
769 | error = device_create_file(disk_to_dev(nbd->disk), &pid_attr); |
770 | if (error) { | |
771 | dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n"); | |
772 | goto out_recv; | |
773 | } | |
774 | ||
775 | nbd_size_update(nbd, bdev); | |
776 | ||
30d53d9c | 777 | nbd_dev_dbg_init(nbd); |
9561a7ad JB |
778 | for (i = 0; i < num_connections; i++) { |
779 | sk_set_memalloc(nbd->socks[i]->sock->sk); | |
780 | atomic_inc(&nbd->recv_threads); | |
781 | INIT_WORK(&args[i].work, recv_work); | |
782 | args[i].nbd = nbd; | |
783 | args[i].index = i; | |
784 | queue_work(system_long_wq, &args[i].work); | |
785 | } | |
786 | wait_event_interruptible(nbd->recv_wq, | |
787 | atomic_read(&nbd->recv_threads) == 0); | |
788 | for (i = 0; i < num_connections; i++) | |
789 | flush_work(&args[i].work); | |
30d53d9c | 790 | nbd_dev_dbg_close(nbd); |
9561a7ad JB |
791 | nbd_size_clear(nbd, bdev); |
792 | device_remove_file(disk_to_dev(nbd->disk), &pid_attr); | |
793 | out_recv: | |
794 | mutex_lock(&nbd->config_lock); | |
97240963 | 795 | nbd->task_recv = NULL; |
9561a7ad | 796 | out_err: |
36e47bee | 797 | sock_shutdown(nbd); |
f4507164 | 798 | nbd_clear_que(nbd); |
3a2d63f8 | 799 | kill_bdev(bdev); |
0e4f0f6f MP |
800 | nbd_bdev_reset(bdev); |
801 | ||
9b4a6ba9 JB |
802 | /* user requested, ignore socket errors */ |
803 | if (test_bit(NBD_DISCONNECT_REQUESTED, &nbd->runtime_flags)) | |
1f7b5cf1 | 804 | error = 0; |
9b4a6ba9 | 805 | if (test_bit(NBD_TIMEDOUT, &nbd->runtime_flags)) |
1f7b5cf1 MP |
806 | error = -ETIMEDOUT; |
807 | ||
0e4f0f6f | 808 | nbd_reset(nbd); |
19391830 | 809 | return error; |
1a2ad211 PM |
810 | } |
811 | ||
1da177e4 | 812 | case NBD_CLEAR_QUE: |
4b2f0260 HX |
813 | /* |
814 | * This is for compatibility only. The queue is always cleared | |
815 | * by NBD_DO_IT or NBD_CLEAR_SOCK. | |
816 | */ | |
1da177e4 | 817 | return 0; |
1a2ad211 | 818 | |
1da177e4 | 819 | case NBD_PRINT_DEBUG: |
fd8383fd JB |
820 | /* |
821 | * For compatibility only, we no longer keep a list of | |
822 | * outstanding requests. | |
823 | */ | |
1da177e4 LT |
824 | return 0; |
825 | } | |
1a2ad211 PM |
826 | return -ENOTTY; |
827 | } | |
828 | ||
829 | static int nbd_ioctl(struct block_device *bdev, fmode_t mode, | |
830 | unsigned int cmd, unsigned long arg) | |
831 | { | |
f4507164 | 832 | struct nbd_device *nbd = bdev->bd_disk->private_data; |
1a2ad211 PM |
833 | int error; |
834 | ||
835 | if (!capable(CAP_SYS_ADMIN)) | |
836 | return -EPERM; | |
837 | ||
f4507164 | 838 | BUG_ON(nbd->magic != NBD_MAGIC); |
1a2ad211 | 839 | |
9561a7ad | 840 | mutex_lock(&nbd->config_lock); |
f4507164 | 841 | error = __nbd_ioctl(bdev, nbd, cmd, arg); |
9561a7ad | 842 | mutex_unlock(&nbd->config_lock); |
1a2ad211 PM |
843 | |
844 | return error; | |
1da177e4 LT |
845 | } |
846 | ||
83d5cde4 | 847 | static const struct block_device_operations nbd_fops = |
1da177e4 LT |
848 | { |
849 | .owner = THIS_MODULE, | |
8a6cfeb6 | 850 | .ioctl = nbd_ioctl, |
263a3df1 | 851 | .compat_ioctl = nbd_ioctl, |
1da177e4 LT |
852 | }; |
853 | ||
30d53d9c MP |
854 | #if IS_ENABLED(CONFIG_DEBUG_FS) |
855 | ||
856 | static int nbd_dbg_tasks_show(struct seq_file *s, void *unused) | |
857 | { | |
858 | struct nbd_device *nbd = s->private; | |
859 | ||
860 | if (nbd->task_recv) | |
861 | seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv)); | |
30d53d9c MP |
862 | |
863 | return 0; | |
864 | } | |
865 | ||
866 | static int nbd_dbg_tasks_open(struct inode *inode, struct file *file) | |
867 | { | |
868 | return single_open(file, nbd_dbg_tasks_show, inode->i_private); | |
869 | } | |
870 | ||
871 | static const struct file_operations nbd_dbg_tasks_ops = { | |
872 | .open = nbd_dbg_tasks_open, | |
873 | .read = seq_read, | |
874 | .llseek = seq_lseek, | |
875 | .release = single_release, | |
876 | }; | |
877 | ||
878 | static int nbd_dbg_flags_show(struct seq_file *s, void *unused) | |
879 | { | |
880 | struct nbd_device *nbd = s->private; | |
881 | u32 flags = nbd->flags; | |
882 | ||
883 | seq_printf(s, "Hex: 0x%08x\n\n", flags); | |
884 | ||
885 | seq_puts(s, "Known flags:\n"); | |
886 | ||
887 | if (flags & NBD_FLAG_HAS_FLAGS) | |
888 | seq_puts(s, "NBD_FLAG_HAS_FLAGS\n"); | |
889 | if (flags & NBD_FLAG_READ_ONLY) | |
890 | seq_puts(s, "NBD_FLAG_READ_ONLY\n"); | |
891 | if (flags & NBD_FLAG_SEND_FLUSH) | |
892 | seq_puts(s, "NBD_FLAG_SEND_FLUSH\n"); | |
893 | if (flags & NBD_FLAG_SEND_TRIM) | |
894 | seq_puts(s, "NBD_FLAG_SEND_TRIM\n"); | |
895 | ||
896 | return 0; | |
897 | } | |
898 | ||
899 | static int nbd_dbg_flags_open(struct inode *inode, struct file *file) | |
900 | { | |
901 | return single_open(file, nbd_dbg_flags_show, inode->i_private); | |
902 | } | |
903 | ||
904 | static const struct file_operations nbd_dbg_flags_ops = { | |
905 | .open = nbd_dbg_flags_open, | |
906 | .read = seq_read, | |
907 | .llseek = seq_lseek, | |
908 | .release = single_release, | |
909 | }; | |
910 | ||
911 | static int nbd_dev_dbg_init(struct nbd_device *nbd) | |
912 | { | |
913 | struct dentry *dir; | |
27ea43fe MP |
914 | |
915 | if (!nbd_dbg_dir) | |
916 | return -EIO; | |
30d53d9c MP |
917 | |
918 | dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir); | |
27ea43fe MP |
919 | if (!dir) { |
920 | dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n", | |
921 | nbd_name(nbd)); | |
922 | return -EIO; | |
30d53d9c MP |
923 | } |
924 | nbd->dbg_dir = dir; | |
925 | ||
27ea43fe MP |
926 | debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops); |
927 | debugfs_create_u64("size_bytes", 0444, dir, &nbd->bytesize); | |
0eadf37a | 928 | debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout); |
ef77b515 | 929 | debugfs_create_u64("blocksize", 0444, dir, &nbd->blksize); |
d366a0ff | 930 | debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops); |
30d53d9c MP |
931 | |
932 | return 0; | |
933 | } | |
934 | ||
935 | static void nbd_dev_dbg_close(struct nbd_device *nbd) | |
936 | { | |
937 | debugfs_remove_recursive(nbd->dbg_dir); | |
938 | } | |
939 | ||
940 | static int nbd_dbg_init(void) | |
941 | { | |
942 | struct dentry *dbg_dir; | |
943 | ||
944 | dbg_dir = debugfs_create_dir("nbd", NULL); | |
27ea43fe MP |
945 | if (!dbg_dir) |
946 | return -EIO; | |
30d53d9c MP |
947 | |
948 | nbd_dbg_dir = dbg_dir; | |
949 | ||
950 | return 0; | |
951 | } | |
952 | ||
953 | static void nbd_dbg_close(void) | |
954 | { | |
955 | debugfs_remove_recursive(nbd_dbg_dir); | |
956 | } | |
957 | ||
958 | #else /* IS_ENABLED(CONFIG_DEBUG_FS) */ | |
959 | ||
960 | static int nbd_dev_dbg_init(struct nbd_device *nbd) | |
961 | { | |
962 | return 0; | |
963 | } | |
964 | ||
965 | static void nbd_dev_dbg_close(struct nbd_device *nbd) | |
966 | { | |
967 | } | |
968 | ||
969 | static int nbd_dbg_init(void) | |
970 | { | |
971 | return 0; | |
972 | } | |
973 | ||
974 | static void nbd_dbg_close(void) | |
975 | { | |
976 | } | |
977 | ||
978 | #endif | |
979 | ||
fd8383fd JB |
980 | static int nbd_init_request(void *data, struct request *rq, |
981 | unsigned int hctx_idx, unsigned int request_idx, | |
982 | unsigned int numa_node) | |
983 | { | |
984 | struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq); | |
fd8383fd | 985 | cmd->nbd = data; |
fd8383fd JB |
986 | return 0; |
987 | } | |
988 | ||
989 | static struct blk_mq_ops nbd_mq_ops = { | |
990 | .queue_rq = nbd_queue_rq, | |
fd8383fd | 991 | .init_request = nbd_init_request, |
0eadf37a | 992 | .timeout = nbd_xmit_timeout, |
fd8383fd JB |
993 | }; |
994 | ||
1da177e4 LT |
995 | /* |
996 | * And here should be modules and kernel interface | |
997 | * (Just smiley confuses emacs :-) | |
998 | */ | |
999 | ||
1000 | static int __init nbd_init(void) | |
1001 | { | |
1002 | int err = -ENOMEM; | |
1003 | int i; | |
d71a6d73 | 1004 | int part_shift; |
1da177e4 | 1005 | |
5b7b18cc | 1006 | BUILD_BUG_ON(sizeof(struct nbd_request) != 28); |
1da177e4 | 1007 | |
d71a6d73 | 1008 | if (max_part < 0) { |
7742ce4a | 1009 | printk(KERN_ERR "nbd: max_part must be >= 0\n"); |
d71a6d73 LV |
1010 | return -EINVAL; |
1011 | } | |
1012 | ||
1013 | part_shift = 0; | |
5988ce23 | 1014 | if (max_part > 0) { |
d71a6d73 LV |
1015 | part_shift = fls(max_part); |
1016 | ||
5988ce23 NK |
1017 | /* |
1018 | * Adjust max_part according to part_shift as it is exported | |
1019 | * to user space so that user can know the max number of | |
1020 | * partition kernel should be able to manage. | |
1021 | * | |
1022 | * Note that -1 is required because partition 0 is reserved | |
1023 | * for the whole disk. | |
1024 | */ | |
1025 | max_part = (1UL << part_shift) - 1; | |
1026 | } | |
1027 | ||
3b271082 NK |
1028 | if ((1UL << part_shift) > DISK_MAX_PARTS) |
1029 | return -EINVAL; | |
1030 | ||
1031 | if (nbds_max > 1UL << (MINORBITS - part_shift)) | |
1032 | return -EINVAL; | |
1033 | ||
ff6b8090 SM |
1034 | nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL); |
1035 | if (!nbd_dev) | |
1036 | return -ENOMEM; | |
1037 | ||
40be0c28 | 1038 | for (i = 0; i < nbds_max; i++) { |
25b4acfc | 1039 | struct request_queue *q; |
d71a6d73 | 1040 | struct gendisk *disk = alloc_disk(1 << part_shift); |
1da177e4 LT |
1041 | if (!disk) |
1042 | goto out; | |
1043 | nbd_dev[i].disk = disk; | |
fd8383fd JB |
1044 | |
1045 | nbd_dev[i].tag_set.ops = &nbd_mq_ops; | |
1046 | nbd_dev[i].tag_set.nr_hw_queues = 1; | |
1047 | nbd_dev[i].tag_set.queue_depth = 128; | |
1048 | nbd_dev[i].tag_set.numa_node = NUMA_NO_NODE; | |
1049 | nbd_dev[i].tag_set.cmd_size = sizeof(struct nbd_cmd); | |
1050 | nbd_dev[i].tag_set.flags = BLK_MQ_F_SHOULD_MERGE | | |
005043ac | 1051 | BLK_MQ_F_SG_MERGE | BLK_MQ_F_BLOCKING; |
fd8383fd JB |
1052 | nbd_dev[i].tag_set.driver_data = &nbd_dev[i]; |
1053 | ||
1054 | err = blk_mq_alloc_tag_set(&nbd_dev[i].tag_set); | |
1055 | if (err) { | |
1056 | put_disk(disk); | |
1057 | goto out; | |
1058 | } | |
1059 | ||
1da177e4 LT |
1060 | /* |
1061 | * The new linux 2.5 block layer implementation requires | |
1062 | * every gendisk to have its very own request_queue struct. | |
1063 | * These structs are big so we dynamically allocate them. | |
1064 | */ | |
25b4acfc JM |
1065 | q = blk_mq_init_queue(&nbd_dev[i].tag_set); |
1066 | if (IS_ERR(q)) { | |
fd8383fd | 1067 | blk_mq_free_tag_set(&nbd_dev[i].tag_set); |
1da177e4 LT |
1068 | put_disk(disk); |
1069 | goto out; | |
1070 | } | |
25b4acfc | 1071 | disk->queue = q; |
fd8383fd | 1072 | |
31dcfab0 JA |
1073 | /* |
1074 | * Tell the block layer that we are not a rotational device | |
1075 | */ | |
1076 | queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue); | |
b277da0a | 1077 | queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, disk->queue); |
a336d298 | 1078 | disk->queue->limits.discard_granularity = 512; |
2bb4cd5c | 1079 | blk_queue_max_discard_sectors(disk->queue, UINT_MAX); |
a336d298 | 1080 | disk->queue->limits.discard_zeroes_data = 0; |
078be02b MB |
1081 | blk_queue_max_hw_sectors(disk->queue, 65536); |
1082 | disk->queue->limits.max_sectors = 256; | |
1da177e4 LT |
1083 | } |
1084 | ||
1085 | if (register_blkdev(NBD_MAJOR, "nbd")) { | |
1086 | err = -EIO; | |
1087 | goto out; | |
1088 | } | |
1089 | ||
1090 | printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR); | |
1da177e4 | 1091 | |
30d53d9c MP |
1092 | nbd_dbg_init(); |
1093 | ||
40be0c28 | 1094 | for (i = 0; i < nbds_max; i++) { |
1da177e4 | 1095 | struct gendisk *disk = nbd_dev[i].disk; |
f4507164 | 1096 | nbd_dev[i].magic = NBD_MAGIC; |
9561a7ad | 1097 | mutex_init(&nbd_dev[i].config_lock); |
1da177e4 | 1098 | disk->major = NBD_MAJOR; |
d71a6d73 | 1099 | disk->first_minor = i << part_shift; |
1da177e4 LT |
1100 | disk->fops = &nbd_fops; |
1101 | disk->private_data = &nbd_dev[i]; | |
1da177e4 | 1102 | sprintf(disk->disk_name, "nbd%d", i); |
9561a7ad | 1103 | init_waitqueue_head(&nbd_dev[i].recv_wq); |
0e4f0f6f | 1104 | nbd_reset(&nbd_dev[i]); |
1da177e4 LT |
1105 | add_disk(disk); |
1106 | } | |
1107 | ||
1108 | return 0; | |
1109 | out: | |
1110 | while (i--) { | |
fd8383fd | 1111 | blk_mq_free_tag_set(&nbd_dev[i].tag_set); |
1da177e4 LT |
1112 | blk_cleanup_queue(nbd_dev[i].disk->queue); |
1113 | put_disk(nbd_dev[i].disk); | |
1114 | } | |
f3944d61 | 1115 | kfree(nbd_dev); |
1da177e4 LT |
1116 | return err; |
1117 | } | |
1118 | ||
1119 | static void __exit nbd_cleanup(void) | |
1120 | { | |
1121 | int i; | |
30d53d9c MP |
1122 | |
1123 | nbd_dbg_close(); | |
1124 | ||
40be0c28 | 1125 | for (i = 0; i < nbds_max; i++) { |
1da177e4 | 1126 | struct gendisk *disk = nbd_dev[i].disk; |
40be0c28 | 1127 | nbd_dev[i].magic = 0; |
1da177e4 LT |
1128 | if (disk) { |
1129 | del_gendisk(disk); | |
1130 | blk_cleanup_queue(disk->queue); | |
fd8383fd | 1131 | blk_mq_free_tag_set(&nbd_dev[i].tag_set); |
1da177e4 LT |
1132 | put_disk(disk); |
1133 | } | |
1134 | } | |
1da177e4 | 1135 | unregister_blkdev(NBD_MAJOR, "nbd"); |
f3944d61 | 1136 | kfree(nbd_dev); |
1da177e4 LT |
1137 | printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR); |
1138 | } | |
1139 | ||
1140 | module_init(nbd_init); | |
1141 | module_exit(nbd_cleanup); | |
1142 | ||
1143 | MODULE_DESCRIPTION("Network Block Device"); | |
1144 | MODULE_LICENSE("GPL"); | |
1145 | ||
40be0c28 | 1146 | module_param(nbds_max, int, 0444); |
d71a6d73 LV |
1147 | MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)"); |
1148 | module_param(max_part, int, 0444); | |
1149 | MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)"); |