]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- |
2 | // vim: ts=8 sw=2 smarttab | |
3 | /* | |
4 | * Ceph - scalable distributed file system | |
5 | * | |
6 | * Copyright (C) 2014 Red Hat | |
7 | * | |
8 | * This is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License version 2.1, as published by the Free Software | |
11 | * Foundation. See file COPYING. | |
12 | * | |
13 | */ | |
14 | ||
15 | #include <unistd.h> | |
16 | #include <stdlib.h> | |
17 | #include <sys/types.h> | |
18 | #include <sys/stat.h> | |
19 | #include <fcntl.h> | |
20 | ||
21 | #include "KernelDevice.h" | |
22 | #include "include/types.h" | |
23 | #include "include/compat.h" | |
24 | #include "include/stringify.h" | |
25 | #include "common/errno.h" | |
26 | #include "common/debug.h" | |
27 | #include "common/blkdev.h" | |
28 | #include "common/align.h" | |
29 | #include "common/blkdev.h" | |
30 | ||
31 | #define dout_context cct | |
32 | #define dout_subsys ceph_subsys_bdev | |
33 | #undef dout_prefix | |
34 | #define dout_prefix *_dout << "bdev(" << this << " " << path << ") " | |
35 | ||
36 | KernelDevice::KernelDevice(CephContext* cct, aio_callback_t cb, void *cbpriv) | |
37 | : BlockDevice(cct), | |
38 | fd_direct(-1), | |
39 | fd_buffered(-1), | |
40 | size(0), block_size(0), | |
41 | fs(NULL), aio(false), dio(false), | |
42 | debug_lock("KernelDevice::debug_lock"), | |
43 | aio_queue(cct->_conf->bdev_aio_max_queue_depth), | |
44 | aio_callback(cb), | |
45 | aio_callback_priv(cbpriv), | |
46 | aio_stop(false), | |
47 | aio_thread(this), | |
48 | injecting_crash(0) | |
49 | { | |
50 | } | |
51 | ||
52 | int KernelDevice::_lock() | |
53 | { | |
54 | struct flock l; | |
55 | memset(&l, 0, sizeof(l)); | |
56 | l.l_type = F_WRLCK; | |
57 | l.l_whence = SEEK_SET; | |
58 | int r = ::fcntl(fd_direct, F_SETLK, &l); | |
59 | if (r < 0) | |
60 | return -errno; | |
61 | return 0; | |
62 | } | |
63 | ||
64 | int KernelDevice::open(const string& p) | |
65 | { | |
66 | path = p; | |
67 | int r = 0; | |
68 | dout(1) << __func__ << " path " << path << dendl; | |
69 | ||
91327a77 | 70 | fd_direct = ::open(path.c_str(), O_RDWR | O_DIRECT | O_CLOEXEC); |
7c673cae FG |
71 | if (fd_direct < 0) { |
72 | r = -errno; | |
73 | derr << __func__ << " open got: " << cpp_strerror(r) << dendl; | |
74 | return r; | |
75 | } | |
91327a77 | 76 | fd_buffered = ::open(path.c_str(), O_RDWR | O_CLOEXEC); |
7c673cae FG |
77 | if (fd_buffered < 0) { |
78 | r = -errno; | |
79 | derr << __func__ << " open got: " << cpp_strerror(r) << dendl; | |
80 | goto out_direct; | |
81 | } | |
82 | dio = true; | |
83 | aio = cct->_conf->bdev_aio; | |
84 | if (!aio) { | |
85 | assert(0 == "non-aio not supported"); | |
86 | } | |
87 | ||
88 | // disable readahead as it will wreak havoc on our mix of | |
89 | // directio/aio and buffered io. | |
90 | r = posix_fadvise(fd_buffered, 0, 0, POSIX_FADV_RANDOM); | |
91 | if (r) { | |
92 | r = -r; | |
93 | derr << __func__ << " open got: " << cpp_strerror(r) << dendl; | |
94 | goto out_fail; | |
95 | } | |
96 | ||
97 | r = _lock(); | |
98 | if (r < 0) { | |
99 | derr << __func__ << " failed to lock " << path << ": " << cpp_strerror(r) | |
100 | << dendl; | |
101 | goto out_fail; | |
102 | } | |
103 | ||
104 | struct stat st; | |
105 | r = ::fstat(fd_direct, &st); | |
106 | if (r < 0) { | |
107 | r = -errno; | |
108 | derr << __func__ << " fstat got " << cpp_strerror(r) << dendl; | |
109 | goto out_fail; | |
110 | } | |
111 | ||
112 | // Operate as though the block size is 4 KB. The backing file | |
113 | // blksize doesn't strictly matter except that some file systems may | |
114 | // require a read/modify/write if we write something smaller than | |
115 | // it. | |
116 | block_size = cct->_conf->bdev_block_size; | |
117 | if (block_size != (unsigned)st.st_blksize) { | |
118 | dout(1) << __func__ << " backing device/file reports st_blksize " | |
119 | << st.st_blksize << ", using bdev_block_size " | |
120 | << block_size << " anyway" << dendl; | |
121 | } | |
122 | ||
123 | if (S_ISBLK(st.st_mode)) { | |
124 | int64_t s; | |
125 | r = get_block_device_size(fd_direct, &s); | |
126 | if (r < 0) { | |
127 | goto out_fail; | |
128 | } | |
129 | size = s; | |
130 | } else { | |
131 | size = st.st_size; | |
132 | } | |
3efd9988 FG |
133 | if (cct->_conf->get_val<bool>("bdev_inject_bad_size")) { |
134 | derr << "injecting bad size; actual 0x" << std::hex << size | |
135 | << " but using 0x" << (size & ~block_size) << std::dec << dendl; | |
136 | size &= ~(block_size); | |
137 | } | |
7c673cae FG |
138 | |
139 | { | |
140 | char partition[PATH_MAX], devname[PATH_MAX]; | |
141 | r = get_device_by_fd(fd_buffered, partition, devname, sizeof(devname)); | |
142 | if (r < 0) { | |
143 | derr << "unable to get device name for " << path << ": " | |
144 | << cpp_strerror(r) << dendl; | |
145 | rotational = true; | |
146 | } else { | |
147 | dout(20) << __func__ << " devname " << devname << dendl; | |
148 | rotational = block_device_is_rotational(devname); | |
149 | } | |
150 | } | |
151 | ||
31f18b77 FG |
152 | r = _aio_start(); |
153 | if (r < 0) { | |
154 | goto out_fail; | |
155 | } | |
156 | ||
7c673cae FG |
157 | fs = FS::create_by_fd(fd_direct); |
158 | assert(fs); | |
159 | ||
160 | // round size down to an even block | |
161 | size &= ~(block_size - 1); | |
162 | ||
7c673cae FG |
163 | dout(1) << __func__ |
164 | << " size " << size | |
165 | << " (0x" << std::hex << size << std::dec << ", " | |
1adf2230 | 166 | << byte_u_t(size) << ")" |
7c673cae | 167 | << " block_size " << block_size |
1adf2230 | 168 | << " (" << byte_u_t(block_size) << ")" |
7c673cae FG |
169 | << " " << (rotational ? "rotational" : "non-rotational") |
170 | << dendl; | |
171 | return 0; | |
172 | ||
173 | out_fail: | |
174 | VOID_TEMP_FAILURE_RETRY(::close(fd_buffered)); | |
175 | fd_buffered = -1; | |
176 | out_direct: | |
177 | VOID_TEMP_FAILURE_RETRY(::close(fd_direct)); | |
178 | fd_direct = -1; | |
179 | return r; | |
180 | } | |
181 | ||
182 | void KernelDevice::close() | |
183 | { | |
184 | dout(1) << __func__ << dendl; | |
185 | _aio_stop(); | |
186 | ||
187 | assert(fs); | |
188 | delete fs; | |
189 | fs = NULL; | |
190 | ||
191 | assert(fd_direct >= 0); | |
192 | VOID_TEMP_FAILURE_RETRY(::close(fd_direct)); | |
193 | fd_direct = -1; | |
194 | ||
195 | assert(fd_buffered >= 0); | |
196 | VOID_TEMP_FAILURE_RETRY(::close(fd_buffered)); | |
197 | fd_buffered = -1; | |
198 | ||
199 | path.clear(); | |
200 | } | |
201 | ||
202 | static string get_dev_property(const char *dev, const char *property) | |
203 | { | |
204 | char val[1024] = {0}; | |
205 | get_block_device_string_property(dev, property, val, sizeof(val)); | |
206 | return val; | |
207 | } | |
208 | ||
209 | int KernelDevice::collect_metadata(string prefix, map<string,string> *pm) const | |
210 | { | |
211 | (*pm)[prefix + "rotational"] = stringify((int)(bool)rotational); | |
212 | (*pm)[prefix + "size"] = stringify(get_size()); | |
213 | (*pm)[prefix + "block_size"] = stringify(get_block_size()); | |
214 | (*pm)[prefix + "driver"] = "KernelDevice"; | |
215 | if (rotational) { | |
216 | (*pm)[prefix + "type"] = "hdd"; | |
217 | } else { | |
218 | (*pm)[prefix + "type"] = "ssd"; | |
219 | } | |
220 | ||
221 | struct stat st; | |
222 | int r = ::fstat(fd_buffered, &st); | |
223 | if (r < 0) | |
224 | return -errno; | |
225 | if (S_ISBLK(st.st_mode)) { | |
226 | (*pm)[prefix + "access_mode"] = "blk"; | |
227 | char partition_path[PATH_MAX]; | |
228 | char dev_node[PATH_MAX]; | |
229 | int rc = get_device_by_fd(fd_buffered, partition_path, dev_node, PATH_MAX); | |
230 | switch (rc) { | |
231 | case -EOPNOTSUPP: | |
232 | case -EINVAL: | |
233 | (*pm)[prefix + "partition_path"] = "unknown"; | |
234 | (*pm)[prefix + "dev_node"] = "unknown"; | |
235 | break; | |
236 | case -ENODEV: | |
237 | (*pm)[prefix + "partition_path"] = string(partition_path); | |
238 | (*pm)[prefix + "dev_node"] = "unknown"; | |
239 | break; | |
240 | default: | |
241 | { | |
242 | (*pm)[prefix + "partition_path"] = string(partition_path); | |
243 | (*pm)[prefix + "dev_node"] = string(dev_node); | |
244 | (*pm)[prefix + "model"] = get_dev_property(dev_node, "device/model"); | |
245 | (*pm)[prefix + "dev"] = get_dev_property(dev_node, "dev"); | |
246 | ||
247 | // nvme exposes a serial number | |
248 | string serial = get_dev_property(dev_node, "device/serial"); | |
249 | if (serial.length()) { | |
250 | (*pm)[prefix + "serial"] = serial; | |
251 | } | |
252 | ||
253 | // nvme has a device/device/* structure; infer from that. there | |
254 | // is probably a better way? | |
255 | string nvme_vendor = get_dev_property(dev_node, "device/device/vendor"); | |
256 | if (nvme_vendor.length()) { | |
257 | (*pm)[prefix + "type"] = "nvme"; | |
258 | } | |
259 | } | |
260 | } | |
261 | } else { | |
262 | (*pm)[prefix + "access_mode"] = "file"; | |
263 | (*pm)[prefix + "path"] = path; | |
264 | } | |
265 | return 0; | |
266 | } | |
267 | ||
268 | int KernelDevice::flush() | |
269 | { | |
31f18b77 | 270 | // protect flush with a mutex. note that we are not really protecting |
7c673cae FG |
271 | // data here. instead, we're ensuring that if any flush() caller |
272 | // sees that io_since_flush is true, they block any racing callers | |
273 | // until the flush is observed. that allows racing threads to be | |
274 | // calling flush while still ensuring that *any* of them that got an | |
275 | // aio completion notification will not return before that aio is | |
276 | // stable on disk: whichever thread sees the flag first will block | |
277 | // followers until the aio is stable. | |
278 | std::lock_guard<std::mutex> l(flush_mutex); | |
279 | ||
280 | bool expect = true; | |
281 | if (!io_since_flush.compare_exchange_strong(expect, false)) { | |
282 | dout(10) << __func__ << " no-op (no ios since last flush), flag is " | |
283 | << (int)io_since_flush.load() << dendl; | |
284 | return 0; | |
285 | } | |
286 | ||
287 | dout(10) << __func__ << " start" << dendl; | |
288 | if (cct->_conf->bdev_inject_crash) { | |
289 | ++injecting_crash; | |
290 | // sleep for a moment to give other threads a chance to submit or | |
291 | // wait on io that races with a flush. | |
292 | derr << __func__ << " injecting crash. first we sleep..." << dendl; | |
293 | sleep(cct->_conf->bdev_inject_crash_flush_delay); | |
294 | derr << __func__ << " and now we die" << dendl; | |
295 | cct->_log->flush(); | |
296 | _exit(1); | |
297 | } | |
298 | utime_t start = ceph_clock_now(); | |
299 | int r = ::fdatasync(fd_direct); | |
300 | utime_t end = ceph_clock_now(); | |
301 | utime_t dur = end - start; | |
302 | if (r < 0) { | |
303 | r = -errno; | |
304 | derr << __func__ << " fdatasync got: " << cpp_strerror(r) << dendl; | |
305 | ceph_abort(); | |
306 | } | |
307 | dout(5) << __func__ << " in " << dur << dendl;; | |
308 | return r; | |
309 | } | |
310 | ||
311 | int KernelDevice::_aio_start() | |
312 | { | |
313 | if (aio) { | |
314 | dout(10) << __func__ << dendl; | |
315 | int r = aio_queue.init(); | |
316 | if (r < 0) { | |
31f18b77 FG |
317 | if (r == -EAGAIN) { |
318 | derr << __func__ << " io_setup(2) failed with EAGAIN; " | |
319 | << "try increasing /proc/sys/fs/aio-max-nr" << dendl; | |
320 | } else { | |
321 | derr << __func__ << " io_setup(2) failed: " << cpp_strerror(r) << dendl; | |
322 | } | |
7c673cae FG |
323 | return r; |
324 | } | |
325 | aio_thread.create("bstore_aio"); | |
326 | } | |
327 | return 0; | |
328 | } | |
329 | ||
330 | void KernelDevice::_aio_stop() | |
331 | { | |
332 | if (aio) { | |
333 | dout(10) << __func__ << dendl; | |
334 | aio_stop = true; | |
335 | aio_thread.join(); | |
336 | aio_stop = false; | |
337 | aio_queue.shutdown(); | |
338 | } | |
339 | } | |
340 | ||
28e407b8 AA |
341 | static bool is_expected_ioerr(const int r) |
342 | { | |
343 | // https://lxr.missinglinkelectronics.com/linux+v4.15/block/blk-core.c#L135 | |
344 | return (r == -EOPNOTSUPP || r == -ETIMEDOUT || r == -ENOSPC || | |
345 | r == -ENOLINK || r == -EREMOTEIO || r == -EBADE || | |
346 | r == -ENODATA || r == -EILSEQ || r == -ENOMEM || | |
347 | r == -EAGAIN || r == -EREMCHG || r == -EIO); | |
348 | } | |
349 | ||
7c673cae FG |
350 | void KernelDevice::_aio_thread() |
351 | { | |
352 | dout(10) << __func__ << " start" << dendl; | |
353 | int inject_crash_count = 0; | |
354 | while (!aio_stop) { | |
355 | dout(40) << __func__ << " polling" << dendl; | |
224ce89b | 356 | int max = cct->_conf->bdev_aio_reap_max; |
7c673cae FG |
357 | aio_t *aio[max]; |
358 | int r = aio_queue.get_next_completed(cct->_conf->bdev_aio_poll_ms, | |
359 | aio, max); | |
360 | if (r < 0) { | |
361 | derr << __func__ << " got " << cpp_strerror(r) << dendl; | |
b32b8144 | 362 | assert(0 == "got unexpected error from io_getevents"); |
7c673cae FG |
363 | } |
364 | if (r > 0) { | |
365 | dout(30) << __func__ << " got " << r << " completed aios" << dendl; | |
366 | for (int i = 0; i < r; ++i) { | |
367 | IOContext *ioc = static_cast<IOContext*>(aio[i]->priv); | |
368 | _aio_log_finish(ioc, aio[i]->offset, aio[i]->length); | |
369 | if (aio[i]->queue_item.is_linked()) { | |
370 | std::lock_guard<std::mutex> l(debug_queue_lock); | |
371 | debug_aio_unlink(*aio[i]); | |
372 | } | |
373 | ||
374 | // set flag indicating new ios have completed. we do this *before* | |
375 | // any completion or notifications so that any user flush() that | |
376 | // follows the observed io completion will include this io. Note | |
377 | // that an earlier, racing flush() could observe and clear this | |
378 | // flag, but that also ensures that the IO will be stable before the | |
379 | // later flush() occurs. | |
380 | io_since_flush.store(true); | |
381 | ||
94b18763 | 382 | long r = aio[i]->get_return_value(); |
b32b8144 | 383 | if (r < 0) { |
28e407b8 AA |
384 | derr << __func__ << " got r=" << r << " (" << cpp_strerror(r) << ")" |
385 | << dendl; | |
386 | if (ioc->allow_eio && is_expected_ioerr(r)) { | |
387 | derr << __func__ << " translating the error to EIO for upper layer" | |
388 | << dendl; | |
389 | ioc->set_return_value(-EIO); | |
b32b8144 | 390 | } else { |
28e407b8 AA |
391 | assert(0 == "got unexpected error from aio_t::get_return_value. " |
392 | "This may suggest HW issue. Please check your dmesg!"); | |
b32b8144 FG |
393 | } |
394 | } else if (aio[i]->length != (uint64_t)r) { | |
395 | derr << "aio to " << aio[i]->offset << "~" << aio[i]->length | |
396 | << " but returned: " << r << dendl; | |
397 | assert(0 == "unexpected aio error"); | |
398 | } | |
399 | ||
400 | dout(10) << __func__ << " finished aio " << aio[i] << " r " << r | |
401 | << " ioc " << ioc | |
402 | << " with " << (ioc->num_running.load() - 1) | |
403 | << " aios left" << dendl; | |
7c673cae FG |
404 | |
405 | // NOTE: once num_running and we either call the callback or | |
406 | // call aio_wake we cannot touch ioc or aio[] as the caller | |
407 | // may free it. | |
408 | if (ioc->priv) { | |
409 | if (--ioc->num_running == 0) { | |
410 | aio_callback(aio_callback_priv, ioc->priv); | |
411 | } | |
412 | } else { | |
31f18b77 | 413 | ioc->try_aio_wake(); |
7c673cae FG |
414 | } |
415 | } | |
416 | } | |
417 | if (cct->_conf->bdev_debug_aio) { | |
418 | utime_t now = ceph_clock_now(); | |
419 | std::lock_guard<std::mutex> l(debug_queue_lock); | |
420 | if (debug_oldest) { | |
421 | if (debug_stall_since == utime_t()) { | |
422 | debug_stall_since = now; | |
423 | } else { | |
424 | utime_t cutoff = now; | |
425 | cutoff -= cct->_conf->bdev_debug_aio_suicide_timeout; | |
426 | if (debug_stall_since < cutoff) { | |
427 | derr << __func__ << " stalled aio " << debug_oldest | |
428 | << " since " << debug_stall_since << ", timeout is " | |
429 | << cct->_conf->bdev_debug_aio_suicide_timeout | |
430 | << "s, suicide" << dendl; | |
431 | assert(0 == "stalled aio... buggy kernel or bad device?"); | |
432 | } | |
433 | } | |
434 | } | |
435 | } | |
436 | reap_ioc(); | |
437 | if (cct->_conf->bdev_inject_crash) { | |
438 | ++inject_crash_count; | |
439 | if (inject_crash_count * cct->_conf->bdev_aio_poll_ms / 1000 > | |
440 | cct->_conf->bdev_inject_crash + cct->_conf->bdev_inject_crash_flush_delay) { | |
441 | derr << __func__ << " bdev_inject_crash trigger from aio thread" | |
442 | << dendl; | |
443 | cct->_log->flush(); | |
444 | _exit(1); | |
445 | } | |
446 | } | |
447 | } | |
448 | reap_ioc(); | |
449 | dout(10) << __func__ << " end" << dendl; | |
450 | } | |
451 | ||
452 | void KernelDevice::_aio_log_start( | |
453 | IOContext *ioc, | |
454 | uint64_t offset, | |
455 | uint64_t length) | |
456 | { | |
457 | dout(20) << __func__ << " 0x" << std::hex << offset << "~" << length | |
458 | << std::dec << dendl; | |
459 | if (cct->_conf->bdev_debug_inflight_ios) { | |
460 | Mutex::Locker l(debug_lock); | |
461 | if (debug_inflight.intersects(offset, length)) { | |
462 | derr << __func__ << " inflight overlap of 0x" | |
463 | << std::hex | |
464 | << offset << "~" << length << std::dec | |
465 | << " with " << debug_inflight << dendl; | |
466 | ceph_abort(); | |
467 | } | |
468 | debug_inflight.insert(offset, length); | |
469 | } | |
470 | } | |
471 | ||
472 | void KernelDevice::debug_aio_link(aio_t& aio) | |
473 | { | |
474 | if (debug_queue.empty()) { | |
475 | debug_oldest = &aio; | |
476 | } | |
477 | debug_queue.push_back(aio); | |
478 | } | |
479 | ||
480 | void KernelDevice::debug_aio_unlink(aio_t& aio) | |
481 | { | |
482 | if (aio.queue_item.is_linked()) { | |
483 | debug_queue.erase(debug_queue.iterator_to(aio)); | |
484 | if (debug_oldest == &aio) { | |
485 | if (debug_queue.empty()) { | |
486 | debug_oldest = nullptr; | |
487 | } else { | |
488 | debug_oldest = &debug_queue.front(); | |
489 | } | |
490 | debug_stall_since = utime_t(); | |
491 | } | |
492 | } | |
493 | } | |
494 | ||
495 | void KernelDevice::_aio_log_finish( | |
496 | IOContext *ioc, | |
497 | uint64_t offset, | |
498 | uint64_t length) | |
499 | { | |
500 | dout(20) << __func__ << " " << aio << " 0x" | |
501 | << std::hex << offset << "~" << length << std::dec << dendl; | |
502 | if (cct->_conf->bdev_debug_inflight_ios) { | |
503 | Mutex::Locker l(debug_lock); | |
504 | debug_inflight.erase(offset, length); | |
505 | } | |
506 | } | |
507 | ||
508 | void KernelDevice::aio_submit(IOContext *ioc) | |
509 | { | |
510 | dout(20) << __func__ << " ioc " << ioc | |
511 | << " pending " << ioc->num_pending.load() | |
512 | << " running " << ioc->num_running.load() | |
513 | << dendl; | |
224ce89b | 514 | |
7c673cae FG |
515 | if (ioc->num_pending.load() == 0) { |
516 | return; | |
517 | } | |
224ce89b | 518 | |
7c673cae FG |
519 | // move these aside, and get our end iterator position now, as the |
520 | // aios might complete as soon as they are submitted and queue more | |
521 | // wal aio's. | |
522 | list<aio_t>::iterator e = ioc->running_aios.begin(); | |
523 | ioc->running_aios.splice(e, ioc->pending_aios); | |
7c673cae FG |
524 | |
525 | int pending = ioc->num_pending.load(); | |
526 | ioc->num_running += pending; | |
527 | ioc->num_pending -= pending; | |
528 | assert(ioc->num_pending.load() == 0); // we should be only thread doing this | |
224ce89b WB |
529 | assert(ioc->pending_aios.size() == 0); |
530 | ||
531 | if (cct->_conf->bdev_debug_aio) { | |
532 | list<aio_t>::iterator p = ioc->running_aios.begin(); | |
533 | while (p != e) { | |
534 | for (auto& io : p->iov) | |
535 | dout(30) << __func__ << " iov " << (void*)io.iov_base | |
536 | << " len " << io.iov_len << dendl; | |
7c673cae | 537 | |
7c673cae | 538 | std::lock_guard<std::mutex> l(debug_queue_lock); |
224ce89b | 539 | debug_aio_link(*p++); |
7c673cae FG |
540 | } |
541 | } | |
224ce89b WB |
542 | |
543 | void *priv = static_cast<void*>(ioc); | |
544 | int r, retries = 0; | |
545 | r = aio_queue.submit_batch(ioc->running_aios.begin(), e, | |
546 | ioc->num_running.load(), priv, &retries); | |
547 | ||
548 | if (retries) | |
549 | derr << __func__ << " retries " << retries << dendl; | |
550 | if (r < 0) { | |
551 | derr << " aio submit got " << cpp_strerror(r) << dendl; | |
552 | assert(r == 0); | |
553 | } | |
7c673cae FG |
554 | } |
555 | ||
556 | int KernelDevice::_sync_write(uint64_t off, bufferlist &bl, bool buffered) | |
557 | { | |
558 | uint64_t len = bl.length(); | |
559 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len | |
560 | << std::dec << " buffered" << dendl; | |
561 | if (cct->_conf->bdev_inject_crash && | |
562 | rand() % cct->_conf->bdev_inject_crash == 0) { | |
563 | derr << __func__ << " bdev_inject_crash: dropping io 0x" << std::hex | |
564 | << off << "~" << len << std::dec << dendl; | |
565 | ++injecting_crash; | |
566 | return 0; | |
567 | } | |
568 | vector<iovec> iov; | |
569 | bl.prepare_iov(&iov); | |
570 | int r = ::pwritev(buffered ? fd_buffered : fd_direct, | |
571 | &iov[0], iov.size(), off); | |
572 | ||
573 | if (r < 0) { | |
574 | r = -errno; | |
575 | derr << __func__ << " pwritev error: " << cpp_strerror(r) << dendl; | |
576 | return r; | |
577 | } | |
578 | if (buffered) { | |
579 | // initiate IO (but do not wait) | |
580 | r = ::sync_file_range(fd_buffered, off, len, SYNC_FILE_RANGE_WRITE); | |
581 | if (r < 0) { | |
582 | r = -errno; | |
583 | derr << __func__ << " sync_file_range error: " << cpp_strerror(r) << dendl; | |
584 | return r; | |
585 | } | |
586 | } | |
31f18b77 FG |
587 | |
588 | io_since_flush.store(true); | |
589 | ||
7c673cae FG |
590 | return 0; |
591 | } | |
592 | ||
593 | int KernelDevice::write( | |
594 | uint64_t off, | |
595 | bufferlist &bl, | |
596 | bool buffered) | |
597 | { | |
598 | uint64_t len = bl.length(); | |
599 | dout(20) << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
600 | << (buffered ? " (buffered)" : " (direct)") | |
601 | << dendl; | |
602 | assert(off % block_size == 0); | |
603 | assert(len % block_size == 0); | |
604 | assert(len > 0); | |
605 | assert(off < size); | |
606 | assert(off + len <= size); | |
607 | ||
608 | if ((!buffered || bl.get_num_buffers() >= IOV_MAX) && | |
b32b8144 | 609 | bl.rebuild_aligned_size_and_memory(block_size, block_size, IOV_MAX)) { |
7c673cae FG |
610 | dout(20) << __func__ << " rebuilding buffer to be aligned" << dendl; |
611 | } | |
612 | dout(40) << "data: "; | |
613 | bl.hexdump(*_dout); | |
614 | *_dout << dendl; | |
615 | ||
616 | return _sync_write(off, bl, buffered); | |
617 | } | |
618 | ||
619 | int KernelDevice::aio_write( | |
620 | uint64_t off, | |
621 | bufferlist &bl, | |
622 | IOContext *ioc, | |
623 | bool buffered) | |
624 | { | |
625 | uint64_t len = bl.length(); | |
626 | dout(20) << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
627 | << (buffered ? " (buffered)" : " (direct)") | |
628 | << dendl; | |
629 | assert(off % block_size == 0); | |
630 | assert(len % block_size == 0); | |
631 | assert(len > 0); | |
632 | assert(off < size); | |
633 | assert(off + len <= size); | |
634 | ||
635 | if ((!buffered || bl.get_num_buffers() >= IOV_MAX) && | |
b32b8144 | 636 | bl.rebuild_aligned_size_and_memory(block_size, block_size, IOV_MAX)) { |
7c673cae FG |
637 | dout(20) << __func__ << " rebuilding buffer to be aligned" << dendl; |
638 | } | |
639 | dout(40) << "data: "; | |
640 | bl.hexdump(*_dout); | |
641 | *_dout << dendl; | |
642 | ||
643 | _aio_log_start(ioc, off, len); | |
644 | ||
645 | #ifdef HAVE_LIBAIO | |
646 | if (aio && dio && !buffered) { | |
647 | ioc->pending_aios.push_back(aio_t(ioc, fd_direct)); | |
648 | ++ioc->num_pending; | |
649 | aio_t& aio = ioc->pending_aios.back(); | |
650 | if (cct->_conf->bdev_inject_crash && | |
651 | rand() % cct->_conf->bdev_inject_crash == 0) { | |
652 | derr << __func__ << " bdev_inject_crash: dropping io 0x" << std::hex | |
653 | << off << "~" << len << std::dec | |
654 | << dendl; | |
655 | // generate a real io so that aio_wait behaves properly, but make it | |
656 | // a read instead of write, and toss the result. | |
657 | aio.pread(off, len); | |
658 | ++injecting_crash; | |
659 | } else { | |
660 | bl.prepare_iov(&aio.iov); | |
661 | for (unsigned i=0; i<aio.iov.size(); ++i) { | |
662 | dout(30) << "aio " << i << " " << aio.iov[i].iov_base | |
663 | << " " << aio.iov[i].iov_len << dendl; | |
664 | } | |
665 | aio.bl.claim_append(bl); | |
666 | aio.pwritev(off, len); | |
667 | } | |
668 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len | |
669 | << std::dec << " aio " << &aio << dendl; | |
670 | } else | |
671 | #endif | |
672 | { | |
673 | int r = _sync_write(off, bl, buffered); | |
674 | _aio_log_finish(ioc, off, len); | |
675 | if (r < 0) | |
676 | return r; | |
677 | } | |
678 | return 0; | |
679 | } | |
680 | ||
681 | int KernelDevice::read(uint64_t off, uint64_t len, bufferlist *pbl, | |
682 | IOContext *ioc, | |
683 | bool buffered) | |
684 | { | |
685 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
686 | << (buffered ? " (buffered)" : " (direct)") | |
687 | << dendl; | |
688 | assert(off % block_size == 0); | |
689 | assert(len % block_size == 0); | |
690 | assert(len > 0); | |
691 | assert(off < size); | |
692 | assert(off + len <= size); | |
693 | ||
694 | _aio_log_start(ioc, off, len); | |
695 | ||
696 | bufferptr p = buffer::create_page_aligned(len); | |
697 | int r = ::pread(buffered ? fd_buffered : fd_direct, | |
698 | p.c_str(), len, off); | |
699 | if (r < 0) { | |
a8e16298 TL |
700 | if (ioc->allow_eio && is_expected_ioerr(r)) { |
701 | r = -EIO; | |
702 | } else { | |
703 | r = -errno; | |
704 | } | |
7c673cae FG |
705 | goto out; |
706 | } | |
707 | assert((uint64_t)r == len); | |
708 | pbl->push_back(std::move(p)); | |
709 | ||
710 | dout(40) << "data: "; | |
711 | pbl->hexdump(*_dout); | |
712 | *_dout << dendl; | |
713 | ||
714 | out: | |
715 | _aio_log_finish(ioc, off, len); | |
716 | return r < 0 ? r : 0; | |
717 | } | |
718 | ||
719 | int KernelDevice::aio_read( | |
720 | uint64_t off, | |
721 | uint64_t len, | |
722 | bufferlist *pbl, | |
723 | IOContext *ioc) | |
724 | { | |
725 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
726 | << dendl; | |
727 | ||
728 | int r = 0; | |
729 | #ifdef HAVE_LIBAIO | |
730 | if (aio && dio) { | |
731 | _aio_log_start(ioc, off, len); | |
732 | ioc->pending_aios.push_back(aio_t(ioc, fd_direct)); | |
733 | ++ioc->num_pending; | |
734 | aio_t& aio = ioc->pending_aios.back(); | |
735 | aio.pread(off, len); | |
736 | for (unsigned i=0; i<aio.iov.size(); ++i) { | |
737 | dout(30) << "aio " << i << " " << aio.iov[i].iov_base | |
738 | << " " << aio.iov[i].iov_len << dendl; | |
739 | } | |
740 | pbl->append(aio.bl); | |
741 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len | |
742 | << std::dec << " aio " << &aio << dendl; | |
743 | } else | |
744 | #endif | |
745 | { | |
746 | r = read(off, len, pbl, ioc, false); | |
747 | } | |
748 | ||
749 | return r; | |
750 | } | |
751 | ||
752 | int KernelDevice::direct_read_unaligned(uint64_t off, uint64_t len, char *buf) | |
753 | { | |
754 | uint64_t aligned_off = align_down(off, block_size); | |
755 | uint64_t aligned_len = align_up(off+len, block_size) - aligned_off; | |
756 | bufferptr p = buffer::create_page_aligned(aligned_len); | |
757 | int r = 0; | |
758 | ||
759 | r = ::pread(fd_direct, p.c_str(), aligned_len, aligned_off); | |
760 | if (r < 0) { | |
761 | r = -errno; | |
762 | derr << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
763 | << " error: " << cpp_strerror(r) << dendl; | |
764 | goto out; | |
765 | } | |
766 | assert((uint64_t)r == aligned_len); | |
767 | memcpy(buf, p.c_str() + (off - aligned_off), len); | |
768 | ||
769 | dout(40) << __func__ << " data: "; | |
770 | bufferlist bl; | |
771 | bl.append(buf, len); | |
772 | bl.hexdump(*_dout); | |
773 | *_dout << dendl; | |
774 | ||
775 | out: | |
776 | return r < 0 ? r : 0; | |
777 | } | |
778 | ||
779 | int KernelDevice::read_random(uint64_t off, uint64_t len, char *buf, | |
780 | bool buffered) | |
781 | { | |
782 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
783 | << dendl; | |
784 | assert(len > 0); | |
785 | assert(off < size); | |
786 | assert(off + len <= size); | |
787 | int r = 0; | |
788 | ||
789 | //if it's direct io and unaligned, we have to use a internal buffer | |
790 | if (!buffered && ((off % block_size != 0) | |
791 | || (len % block_size != 0) | |
792 | || (uintptr_t(buf) % CEPH_PAGE_SIZE != 0))) | |
793 | return direct_read_unaligned(off, len, buf); | |
794 | ||
795 | if (buffered) { | |
796 | //buffered read | |
797 | char *t = buf; | |
798 | uint64_t left = len; | |
799 | while (left > 0) { | |
800 | r = ::pread(fd_buffered, t, left, off); | |
801 | if (r < 0) { | |
802 | r = -errno; | |
803 | derr << __func__ << " 0x" << std::hex << off << "~" << left | |
804 | << std::dec << " error: " << cpp_strerror(r) << dendl; | |
805 | goto out; | |
806 | } | |
807 | off += r; | |
808 | t += r; | |
809 | left -= r; | |
810 | } | |
811 | } else { | |
812 | //direct and aligned read | |
813 | r = ::pread(fd_direct, buf, len, off); | |
814 | if (r < 0) { | |
815 | r = -errno; | |
816 | derr << __func__ << " direct_aligned_read" << " 0x" << std::hex | |
817 | << off << "~" << left << std::dec << " error: " << cpp_strerror(r) | |
818 | << dendl; | |
819 | goto out; | |
820 | } | |
821 | assert((uint64_t)r == len); | |
822 | } | |
823 | ||
824 | dout(40) << __func__ << " data: "; | |
825 | bufferlist bl; | |
826 | bl.append(buf, len); | |
827 | bl.hexdump(*_dout); | |
828 | *_dout << dendl; | |
829 | ||
830 | out: | |
831 | return r < 0 ? r : 0; | |
832 | } | |
833 | ||
834 | int KernelDevice::invalidate_cache(uint64_t off, uint64_t len) | |
835 | { | |
836 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
837 | << dendl; | |
838 | assert(off % block_size == 0); | |
839 | assert(len % block_size == 0); | |
840 | int r = posix_fadvise(fd_buffered, off, len, POSIX_FADV_DONTNEED); | |
841 | if (r) { | |
842 | r = -r; | |
843 | derr << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
844 | << " error: " << cpp_strerror(r) << dendl; | |
845 | } | |
846 | return r; | |
847 | } | |
848 |