]>
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 | ||
70 | fd_direct = ::open(path.c_str(), O_RDWR | O_DIRECT); | |
71 | if (fd_direct < 0) { | |
72 | r = -errno; | |
73 | derr << __func__ << " open got: " << cpp_strerror(r) << dendl; | |
74 | return r; | |
75 | } | |
76 | fd_buffered = ::open(path.c_str(), O_RDWR); | |
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 << ", " | |
166 | << pretty_si_t(size) << "B)" | |
167 | << " block_size " << block_size | |
168 | << " (" << pretty_si_t(block_size) << "B)" | |
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 | ||
341 | void KernelDevice::_aio_thread() | |
342 | { | |
343 | dout(10) << __func__ << " start" << dendl; | |
344 | int inject_crash_count = 0; | |
345 | while (!aio_stop) { | |
346 | dout(40) << __func__ << " polling" << dendl; | |
224ce89b | 347 | int max = cct->_conf->bdev_aio_reap_max; |
7c673cae FG |
348 | aio_t *aio[max]; |
349 | int r = aio_queue.get_next_completed(cct->_conf->bdev_aio_poll_ms, | |
350 | aio, max); | |
351 | if (r < 0) { | |
352 | derr << __func__ << " got " << cpp_strerror(r) << dendl; | |
b32b8144 | 353 | assert(0 == "got unexpected error from io_getevents"); |
7c673cae FG |
354 | } |
355 | if (r > 0) { | |
356 | dout(30) << __func__ << " got " << r << " completed aios" << dendl; | |
357 | for (int i = 0; i < r; ++i) { | |
358 | IOContext *ioc = static_cast<IOContext*>(aio[i]->priv); | |
359 | _aio_log_finish(ioc, aio[i]->offset, aio[i]->length); | |
360 | if (aio[i]->queue_item.is_linked()) { | |
361 | std::lock_guard<std::mutex> l(debug_queue_lock); | |
362 | debug_aio_unlink(*aio[i]); | |
363 | } | |
364 | ||
365 | // set flag indicating new ios have completed. we do this *before* | |
366 | // any completion or notifications so that any user flush() that | |
367 | // follows the observed io completion will include this io. Note | |
368 | // that an earlier, racing flush() could observe and clear this | |
369 | // flag, but that also ensures that the IO will be stable before the | |
370 | // later flush() occurs. | |
371 | io_since_flush.store(true); | |
372 | ||
373 | int r = aio[i]->get_return_value(); | |
b32b8144 FG |
374 | if (r < 0) { |
375 | derr << __func__ << " got " << cpp_strerror(r) << dendl; | |
376 | if (ioc->allow_eio && r == -EIO) { | |
377 | ioc->set_return_value(r); | |
378 | } else { | |
379 | assert(0 == "got unexpected error from io_getevents"); | |
380 | } | |
381 | } else if (aio[i]->length != (uint64_t)r) { | |
382 | derr << "aio to " << aio[i]->offset << "~" << aio[i]->length | |
383 | << " but returned: " << r << dendl; | |
384 | assert(0 == "unexpected aio error"); | |
385 | } | |
386 | ||
387 | dout(10) << __func__ << " finished aio " << aio[i] << " r " << r | |
388 | << " ioc " << ioc | |
389 | << " with " << (ioc->num_running.load() - 1) | |
390 | << " aios left" << dendl; | |
7c673cae FG |
391 | |
392 | // NOTE: once num_running and we either call the callback or | |
393 | // call aio_wake we cannot touch ioc or aio[] as the caller | |
394 | // may free it. | |
395 | if (ioc->priv) { | |
396 | if (--ioc->num_running == 0) { | |
397 | aio_callback(aio_callback_priv, ioc->priv); | |
398 | } | |
399 | } else { | |
31f18b77 | 400 | ioc->try_aio_wake(); |
7c673cae FG |
401 | } |
402 | } | |
403 | } | |
404 | if (cct->_conf->bdev_debug_aio) { | |
405 | utime_t now = ceph_clock_now(); | |
406 | std::lock_guard<std::mutex> l(debug_queue_lock); | |
407 | if (debug_oldest) { | |
408 | if (debug_stall_since == utime_t()) { | |
409 | debug_stall_since = now; | |
410 | } else { | |
411 | utime_t cutoff = now; | |
412 | cutoff -= cct->_conf->bdev_debug_aio_suicide_timeout; | |
413 | if (debug_stall_since < cutoff) { | |
414 | derr << __func__ << " stalled aio " << debug_oldest | |
415 | << " since " << debug_stall_since << ", timeout is " | |
416 | << cct->_conf->bdev_debug_aio_suicide_timeout | |
417 | << "s, suicide" << dendl; | |
418 | assert(0 == "stalled aio... buggy kernel or bad device?"); | |
419 | } | |
420 | } | |
421 | } | |
422 | } | |
423 | reap_ioc(); | |
424 | if (cct->_conf->bdev_inject_crash) { | |
425 | ++inject_crash_count; | |
426 | if (inject_crash_count * cct->_conf->bdev_aio_poll_ms / 1000 > | |
427 | cct->_conf->bdev_inject_crash + cct->_conf->bdev_inject_crash_flush_delay) { | |
428 | derr << __func__ << " bdev_inject_crash trigger from aio thread" | |
429 | << dendl; | |
430 | cct->_log->flush(); | |
431 | _exit(1); | |
432 | } | |
433 | } | |
434 | } | |
435 | reap_ioc(); | |
436 | dout(10) << __func__ << " end" << dendl; | |
437 | } | |
438 | ||
439 | void KernelDevice::_aio_log_start( | |
440 | IOContext *ioc, | |
441 | uint64_t offset, | |
442 | uint64_t length) | |
443 | { | |
444 | dout(20) << __func__ << " 0x" << std::hex << offset << "~" << length | |
445 | << std::dec << dendl; | |
446 | if (cct->_conf->bdev_debug_inflight_ios) { | |
447 | Mutex::Locker l(debug_lock); | |
448 | if (debug_inflight.intersects(offset, length)) { | |
449 | derr << __func__ << " inflight overlap of 0x" | |
450 | << std::hex | |
451 | << offset << "~" << length << std::dec | |
452 | << " with " << debug_inflight << dendl; | |
453 | ceph_abort(); | |
454 | } | |
455 | debug_inflight.insert(offset, length); | |
456 | } | |
457 | } | |
458 | ||
459 | void KernelDevice::debug_aio_link(aio_t& aio) | |
460 | { | |
461 | if (debug_queue.empty()) { | |
462 | debug_oldest = &aio; | |
463 | } | |
464 | debug_queue.push_back(aio); | |
465 | } | |
466 | ||
467 | void KernelDevice::debug_aio_unlink(aio_t& aio) | |
468 | { | |
469 | if (aio.queue_item.is_linked()) { | |
470 | debug_queue.erase(debug_queue.iterator_to(aio)); | |
471 | if (debug_oldest == &aio) { | |
472 | if (debug_queue.empty()) { | |
473 | debug_oldest = nullptr; | |
474 | } else { | |
475 | debug_oldest = &debug_queue.front(); | |
476 | } | |
477 | debug_stall_since = utime_t(); | |
478 | } | |
479 | } | |
480 | } | |
481 | ||
482 | void KernelDevice::_aio_log_finish( | |
483 | IOContext *ioc, | |
484 | uint64_t offset, | |
485 | uint64_t length) | |
486 | { | |
487 | dout(20) << __func__ << " " << aio << " 0x" | |
488 | << std::hex << offset << "~" << length << std::dec << dendl; | |
489 | if (cct->_conf->bdev_debug_inflight_ios) { | |
490 | Mutex::Locker l(debug_lock); | |
491 | debug_inflight.erase(offset, length); | |
492 | } | |
493 | } | |
494 | ||
495 | void KernelDevice::aio_submit(IOContext *ioc) | |
496 | { | |
497 | dout(20) << __func__ << " ioc " << ioc | |
498 | << " pending " << ioc->num_pending.load() | |
499 | << " running " << ioc->num_running.load() | |
500 | << dendl; | |
224ce89b | 501 | |
7c673cae FG |
502 | if (ioc->num_pending.load() == 0) { |
503 | return; | |
504 | } | |
224ce89b | 505 | |
7c673cae FG |
506 | // move these aside, and get our end iterator position now, as the |
507 | // aios might complete as soon as they are submitted and queue more | |
508 | // wal aio's. | |
509 | list<aio_t>::iterator e = ioc->running_aios.begin(); | |
510 | ioc->running_aios.splice(e, ioc->pending_aios); | |
7c673cae FG |
511 | |
512 | int pending = ioc->num_pending.load(); | |
513 | ioc->num_running += pending; | |
514 | ioc->num_pending -= pending; | |
515 | assert(ioc->num_pending.load() == 0); // we should be only thread doing this | |
224ce89b WB |
516 | assert(ioc->pending_aios.size() == 0); |
517 | ||
518 | if (cct->_conf->bdev_debug_aio) { | |
519 | list<aio_t>::iterator p = ioc->running_aios.begin(); | |
520 | while (p != e) { | |
521 | for (auto& io : p->iov) | |
522 | dout(30) << __func__ << " iov " << (void*)io.iov_base | |
523 | << " len " << io.iov_len << dendl; | |
7c673cae | 524 | |
7c673cae | 525 | std::lock_guard<std::mutex> l(debug_queue_lock); |
224ce89b | 526 | debug_aio_link(*p++); |
7c673cae FG |
527 | } |
528 | } | |
224ce89b WB |
529 | |
530 | void *priv = static_cast<void*>(ioc); | |
531 | int r, retries = 0; | |
532 | r = aio_queue.submit_batch(ioc->running_aios.begin(), e, | |
533 | ioc->num_running.load(), priv, &retries); | |
534 | ||
535 | if (retries) | |
536 | derr << __func__ << " retries " << retries << dendl; | |
537 | if (r < 0) { | |
538 | derr << " aio submit got " << cpp_strerror(r) << dendl; | |
539 | assert(r == 0); | |
540 | } | |
7c673cae FG |
541 | } |
542 | ||
543 | int KernelDevice::_sync_write(uint64_t off, bufferlist &bl, bool buffered) | |
544 | { | |
545 | uint64_t len = bl.length(); | |
546 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len | |
547 | << std::dec << " buffered" << dendl; | |
548 | if (cct->_conf->bdev_inject_crash && | |
549 | rand() % cct->_conf->bdev_inject_crash == 0) { | |
550 | derr << __func__ << " bdev_inject_crash: dropping io 0x" << std::hex | |
551 | << off << "~" << len << std::dec << dendl; | |
552 | ++injecting_crash; | |
553 | return 0; | |
554 | } | |
555 | vector<iovec> iov; | |
556 | bl.prepare_iov(&iov); | |
557 | int r = ::pwritev(buffered ? fd_buffered : fd_direct, | |
558 | &iov[0], iov.size(), off); | |
559 | ||
560 | if (r < 0) { | |
561 | r = -errno; | |
562 | derr << __func__ << " pwritev error: " << cpp_strerror(r) << dendl; | |
563 | return r; | |
564 | } | |
565 | if (buffered) { | |
566 | // initiate IO (but do not wait) | |
567 | r = ::sync_file_range(fd_buffered, off, len, SYNC_FILE_RANGE_WRITE); | |
568 | if (r < 0) { | |
569 | r = -errno; | |
570 | derr << __func__ << " sync_file_range error: " << cpp_strerror(r) << dendl; | |
571 | return r; | |
572 | } | |
573 | } | |
31f18b77 FG |
574 | |
575 | io_since_flush.store(true); | |
576 | ||
7c673cae FG |
577 | return 0; |
578 | } | |
579 | ||
580 | int KernelDevice::write( | |
581 | uint64_t off, | |
582 | bufferlist &bl, | |
583 | bool buffered) | |
584 | { | |
585 | uint64_t len = bl.length(); | |
586 | dout(20) << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
587 | << (buffered ? " (buffered)" : " (direct)") | |
588 | << dendl; | |
589 | assert(off % block_size == 0); | |
590 | assert(len % block_size == 0); | |
591 | assert(len > 0); | |
592 | assert(off < size); | |
593 | assert(off + len <= size); | |
594 | ||
595 | if ((!buffered || bl.get_num_buffers() >= IOV_MAX) && | |
b32b8144 | 596 | bl.rebuild_aligned_size_and_memory(block_size, block_size, IOV_MAX)) { |
7c673cae FG |
597 | dout(20) << __func__ << " rebuilding buffer to be aligned" << dendl; |
598 | } | |
599 | dout(40) << "data: "; | |
600 | bl.hexdump(*_dout); | |
601 | *_dout << dendl; | |
602 | ||
603 | return _sync_write(off, bl, buffered); | |
604 | } | |
605 | ||
606 | int KernelDevice::aio_write( | |
607 | uint64_t off, | |
608 | bufferlist &bl, | |
609 | IOContext *ioc, | |
610 | bool buffered) | |
611 | { | |
612 | uint64_t len = bl.length(); | |
613 | dout(20) << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
614 | << (buffered ? " (buffered)" : " (direct)") | |
615 | << dendl; | |
616 | assert(off % block_size == 0); | |
617 | assert(len % block_size == 0); | |
618 | assert(len > 0); | |
619 | assert(off < size); | |
620 | assert(off + len <= size); | |
621 | ||
622 | if ((!buffered || bl.get_num_buffers() >= IOV_MAX) && | |
b32b8144 | 623 | bl.rebuild_aligned_size_and_memory(block_size, block_size, IOV_MAX)) { |
7c673cae FG |
624 | dout(20) << __func__ << " rebuilding buffer to be aligned" << dendl; |
625 | } | |
626 | dout(40) << "data: "; | |
627 | bl.hexdump(*_dout); | |
628 | *_dout << dendl; | |
629 | ||
630 | _aio_log_start(ioc, off, len); | |
631 | ||
632 | #ifdef HAVE_LIBAIO | |
633 | if (aio && dio && !buffered) { | |
634 | ioc->pending_aios.push_back(aio_t(ioc, fd_direct)); | |
635 | ++ioc->num_pending; | |
636 | aio_t& aio = ioc->pending_aios.back(); | |
637 | if (cct->_conf->bdev_inject_crash && | |
638 | rand() % cct->_conf->bdev_inject_crash == 0) { | |
639 | derr << __func__ << " bdev_inject_crash: dropping io 0x" << std::hex | |
640 | << off << "~" << len << std::dec | |
641 | << dendl; | |
642 | // generate a real io so that aio_wait behaves properly, but make it | |
643 | // a read instead of write, and toss the result. | |
644 | aio.pread(off, len); | |
645 | ++injecting_crash; | |
646 | } else { | |
647 | bl.prepare_iov(&aio.iov); | |
648 | for (unsigned i=0; i<aio.iov.size(); ++i) { | |
649 | dout(30) << "aio " << i << " " << aio.iov[i].iov_base | |
650 | << " " << aio.iov[i].iov_len << dendl; | |
651 | } | |
652 | aio.bl.claim_append(bl); | |
653 | aio.pwritev(off, len); | |
654 | } | |
655 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len | |
656 | << std::dec << " aio " << &aio << dendl; | |
657 | } else | |
658 | #endif | |
659 | { | |
660 | int r = _sync_write(off, bl, buffered); | |
661 | _aio_log_finish(ioc, off, len); | |
662 | if (r < 0) | |
663 | return r; | |
664 | } | |
665 | return 0; | |
666 | } | |
667 | ||
668 | int KernelDevice::read(uint64_t off, uint64_t len, bufferlist *pbl, | |
669 | IOContext *ioc, | |
670 | bool buffered) | |
671 | { | |
672 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
673 | << (buffered ? " (buffered)" : " (direct)") | |
674 | << dendl; | |
675 | assert(off % block_size == 0); | |
676 | assert(len % block_size == 0); | |
677 | assert(len > 0); | |
678 | assert(off < size); | |
679 | assert(off + len <= size); | |
680 | ||
681 | _aio_log_start(ioc, off, len); | |
682 | ||
683 | bufferptr p = buffer::create_page_aligned(len); | |
684 | int r = ::pread(buffered ? fd_buffered : fd_direct, | |
685 | p.c_str(), len, off); | |
686 | if (r < 0) { | |
687 | r = -errno; | |
688 | goto out; | |
689 | } | |
690 | assert((uint64_t)r == len); | |
691 | pbl->push_back(std::move(p)); | |
692 | ||
693 | dout(40) << "data: "; | |
694 | pbl->hexdump(*_dout); | |
695 | *_dout << dendl; | |
696 | ||
697 | out: | |
698 | _aio_log_finish(ioc, off, len); | |
699 | return r < 0 ? r : 0; | |
700 | } | |
701 | ||
702 | int KernelDevice::aio_read( | |
703 | uint64_t off, | |
704 | uint64_t len, | |
705 | bufferlist *pbl, | |
706 | IOContext *ioc) | |
707 | { | |
708 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
709 | << dendl; | |
710 | ||
711 | int r = 0; | |
712 | #ifdef HAVE_LIBAIO | |
713 | if (aio && dio) { | |
714 | _aio_log_start(ioc, off, len); | |
715 | ioc->pending_aios.push_back(aio_t(ioc, fd_direct)); | |
716 | ++ioc->num_pending; | |
717 | aio_t& aio = ioc->pending_aios.back(); | |
718 | aio.pread(off, len); | |
719 | for (unsigned i=0; i<aio.iov.size(); ++i) { | |
720 | dout(30) << "aio " << i << " " << aio.iov[i].iov_base | |
721 | << " " << aio.iov[i].iov_len << dendl; | |
722 | } | |
723 | pbl->append(aio.bl); | |
724 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len | |
725 | << std::dec << " aio " << &aio << dendl; | |
726 | } else | |
727 | #endif | |
728 | { | |
729 | r = read(off, len, pbl, ioc, false); | |
730 | } | |
731 | ||
732 | return r; | |
733 | } | |
734 | ||
735 | int KernelDevice::direct_read_unaligned(uint64_t off, uint64_t len, char *buf) | |
736 | { | |
737 | uint64_t aligned_off = align_down(off, block_size); | |
738 | uint64_t aligned_len = align_up(off+len, block_size) - aligned_off; | |
739 | bufferptr p = buffer::create_page_aligned(aligned_len); | |
740 | int r = 0; | |
741 | ||
742 | r = ::pread(fd_direct, p.c_str(), aligned_len, aligned_off); | |
743 | if (r < 0) { | |
744 | r = -errno; | |
745 | derr << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
746 | << " error: " << cpp_strerror(r) << dendl; | |
747 | goto out; | |
748 | } | |
749 | assert((uint64_t)r == aligned_len); | |
750 | memcpy(buf, p.c_str() + (off - aligned_off), len); | |
751 | ||
752 | dout(40) << __func__ << " data: "; | |
753 | bufferlist bl; | |
754 | bl.append(buf, len); | |
755 | bl.hexdump(*_dout); | |
756 | *_dout << dendl; | |
757 | ||
758 | out: | |
759 | return r < 0 ? r : 0; | |
760 | } | |
761 | ||
762 | int KernelDevice::read_random(uint64_t off, uint64_t len, char *buf, | |
763 | bool buffered) | |
764 | { | |
765 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
766 | << dendl; | |
767 | assert(len > 0); | |
768 | assert(off < size); | |
769 | assert(off + len <= size); | |
770 | int r = 0; | |
771 | ||
772 | //if it's direct io and unaligned, we have to use a internal buffer | |
773 | if (!buffered && ((off % block_size != 0) | |
774 | || (len % block_size != 0) | |
775 | || (uintptr_t(buf) % CEPH_PAGE_SIZE != 0))) | |
776 | return direct_read_unaligned(off, len, buf); | |
777 | ||
778 | if (buffered) { | |
779 | //buffered read | |
780 | char *t = buf; | |
781 | uint64_t left = len; | |
782 | while (left > 0) { | |
783 | r = ::pread(fd_buffered, t, left, off); | |
784 | if (r < 0) { | |
785 | r = -errno; | |
786 | derr << __func__ << " 0x" << std::hex << off << "~" << left | |
787 | << std::dec << " error: " << cpp_strerror(r) << dendl; | |
788 | goto out; | |
789 | } | |
790 | off += r; | |
791 | t += r; | |
792 | left -= r; | |
793 | } | |
794 | } else { | |
795 | //direct and aligned read | |
796 | r = ::pread(fd_direct, buf, len, off); | |
797 | if (r < 0) { | |
798 | r = -errno; | |
799 | derr << __func__ << " direct_aligned_read" << " 0x" << std::hex | |
800 | << off << "~" << left << std::dec << " error: " << cpp_strerror(r) | |
801 | << dendl; | |
802 | goto out; | |
803 | } | |
804 | assert((uint64_t)r == len); | |
805 | } | |
806 | ||
807 | dout(40) << __func__ << " data: "; | |
808 | bufferlist bl; | |
809 | bl.append(buf, len); | |
810 | bl.hexdump(*_dout); | |
811 | *_dout << dendl; | |
812 | ||
813 | out: | |
814 | return r < 0 ? r : 0; | |
815 | } | |
816 | ||
817 | int KernelDevice::invalidate_cache(uint64_t off, uint64_t len) | |
818 | { | |
819 | dout(5) << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
820 | << dendl; | |
821 | assert(off % block_size == 0); | |
822 | assert(len % block_size == 0); | |
823 | int r = posix_fadvise(fd_buffered, off, len, POSIX_FADV_DONTNEED); | |
824 | if (r) { | |
825 | r = -r; | |
826 | derr << __func__ << " 0x" << std::hex << off << "~" << len << std::dec | |
827 | << " error: " << cpp_strerror(r) << dendl; | |
828 | } | |
829 | return r; | |
830 | } | |
831 |