2 * Copyright 2014 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
17 #include <folly/experimental/io/AsyncIO.h>
19 #include <sys/eventfd.h>
26 #include <boost/intrusive/parent_from_member.hpp>
27 #include <glog/logging.h>
29 #include <folly/Exception.h>
30 #include <folly/Format.h>
31 #include <folly/Likely.h>
32 #include <folly/String.h>
36 AsyncIOOp::AsyncIOOp(NotificationCallback cb)
38 state_(State::UNINITIALIZED),
40 memset(&iocb_, 0, sizeof(iocb_));
43 void AsyncIOOp::reset(NotificationCallback cb) {
44 CHECK_NE(state_, State::PENDING);
46 state_ = State::UNINITIALIZED;
48 memset(&iocb_, 0, sizeof(iocb_));
51 AsyncIOOp::~AsyncIOOp() {
52 CHECK_NE(state_, State::PENDING);
55 void AsyncIOOp::start() {
56 DCHECK_EQ(state_, State::INITIALIZED);
57 state_ = State::PENDING;
60 void AsyncIOOp::complete(ssize_t result) {
61 DCHECK_EQ(state_, State::PENDING);
62 state_ = State::COMPLETED;
69 ssize_t AsyncIOOp::result() const {
70 CHECK_EQ(state_, State::COMPLETED);
74 void AsyncIOOp::pread(int fd, void* buf, size_t size, off_t start) {
76 io_prep_pread(&iocb_, fd, buf, size, start);
79 void AsyncIOOp::pread(int fd, Range<unsigned char*> range, off_t start) {
80 pread(fd, range.begin(), range.size(), start);
83 void AsyncIOOp::preadv(int fd, const iovec* iov, int iovcnt, off_t start) {
85 io_prep_preadv(&iocb_, fd, iov, iovcnt, start);
88 void AsyncIOOp::pwrite(int fd, const void* buf, size_t size, off_t start) {
90 io_prep_pwrite(&iocb_, fd, const_cast<void*>(buf), size, start);
93 void AsyncIOOp::pwrite(int fd, Range<const unsigned char*> range, off_t start) {
94 pwrite(fd, range.begin(), range.size(), start);
97 void AsyncIOOp::pwritev(int fd, const iovec* iov, int iovcnt, off_t start) {
99 io_prep_pwritev(&iocb_, fd, iov, iovcnt, start);
102 void AsyncIOOp::init() {
103 CHECK_EQ(state_, State::UNINITIALIZED);
104 state_ = State::INITIALIZED;
107 AsyncIO::AsyncIO(size_t capacity, PollMode pollMode)
114 CHECK_GT(capacity_, 0);
115 completed_.reserve(capacity_);
116 if (pollMode == POLLABLE) {
117 pollFd_ = eventfd(0, EFD_NONBLOCK);
118 checkUnixError(pollFd_, "AsyncIO: eventfd creation failed");
122 AsyncIO::~AsyncIO() {
123 CHECK_EQ(pending_, 0);
125 int rc = io_queue_release(ctx_);
126 CHECK_EQ(rc, 0) << "io_queue_release: " << errnoStr(-rc);
129 CHECK_ERR(close(pollFd_));
133 void AsyncIO::decrementPending() {
134 auto p = pending_.fetch_add(-1, std::memory_order_acq_rel);
138 void AsyncIO::initializeContext() {
139 if (!ctxSet_.load(std::memory_order_acquire)) {
140 std::lock_guard<std::mutex> lock(initMutex_);
141 if (!ctxSet_.load(std::memory_order_relaxed)) {
142 int rc = io_queue_init(capacity_, &ctx_);
143 // returns negative errno
145 long aio_nr, aio_max;
146 std::unique_ptr<FILE, int(*)(FILE*)>
147 fp(fopen("/proc/sys/fs/aio-nr", "r"), fclose);
149 CHECK_EQ(fscanf(fp.get(), "%ld", &aio_nr), 1);
151 std::unique_ptr<FILE, int(*)(FILE*)>
152 aio_max_fp(fopen("/proc/sys/fs/aio-max-nr", "r"), fclose);
154 CHECK_EQ(fscanf(aio_max_fp.get(), "%ld", &aio_max), 1);
156 LOG(ERROR) << "No resources for requested capacity of " << capacity_;
157 LOG(ERROR) << "aio_nr " << aio_nr << ", aio_max_nr " << aio_max;
160 checkKernelError(rc, "AsyncIO: io_queue_init failed");
162 ctxSet_.store(true, std::memory_order_release);
167 void AsyncIO::submit(Op* op) {
168 CHECK_EQ(op->state(), Op::State::INITIALIZED);
169 initializeContext(); // on demand
171 // We can increment past capacity, but we'll clean up after ourselves.
172 auto p = pending_.fetch_add(1, std::memory_order_acq_rel);
173 if (p >= capacity_) {
175 throw std::range_error("AsyncIO: too many pending requests");
177 iocb* cb = &op->iocb_;
178 cb->data = nullptr; // unused
180 io_set_eventfd(cb, pollFd_);
182 int rc = io_submit(ctx_, 1, &cb);
185 throwSystemErrorExplicit(-rc, "AsyncIO: io_submit failed");
192 Range<AsyncIO::Op**> AsyncIO::wait(size_t minRequests) {
194 CHECK_EQ(pollFd_, -1) << "wait() only allowed on non-pollable object";
195 auto p = pending_.load(std::memory_order_acquire);
196 CHECK_LE(minRequests, p);
197 return doWait(minRequests, p);
200 Range<AsyncIO::Op**> AsyncIO::pollCompleted() {
202 CHECK_NE(pollFd_, -1) << "pollCompleted() only allowed on pollable object";
204 // This sets the eventFd counter to 0, see
205 // http://www.kernel.org/doc/man-pages/online/pages/man2/eventfd.2.html
208 rc = ::read(pollFd_, &numEvents, 8);
209 } while (rc == -1 && errno == EINTR);
210 if (UNLIKELY(rc == -1 && errno == EAGAIN)) {
211 return Range<Op**>(); // nothing completed
213 checkUnixError(rc, "AsyncIO: read from event fd failed");
216 DCHECK_GT(numEvents, 0);
217 DCHECK_LE(numEvents, pending_);
219 // Don't reap more than numEvents, as we've just reset the counter to 0.
220 return doWait(numEvents, numEvents);
223 Range<AsyncIO::Op**> AsyncIO::doWait(size_t minRequests, size_t maxRequests) {
224 io_event events[maxRequests];
230 // GOTCHA: io_getevents() may returns less than min_nr results if
231 // interrupted after some events have been read (if before, -EINTR
233 ret = io_getevents(ctx_,
237 /* timeout */ nullptr); // wait forever
238 } while (ret == -EINTR);
239 // Check as may not be able to recover without leaking events.
241 << "AsyncIO: io_getevents failed with error " << errnoStr(-ret);
243 } while (count < minRequests);
244 DCHECK_LE(count, maxRequests);
248 return folly::Range<Op**>();
251 for (size_t i = 0; i < count; ++i) {
252 DCHECK(events[i].obj);
253 Op* op = boost::intrusive::get_parent_from_member(
254 events[i].obj, &AsyncIOOp::iocb_);
256 op->complete(events[i].res);
257 completed_.push_back(op);
260 return folly::Range<Op**>(&completed_.front(), count);
263 AsyncIOQueue::AsyncIOQueue(AsyncIO* asyncIO)
264 : asyncIO_(asyncIO) {
267 AsyncIOQueue::~AsyncIOQueue() {
268 CHECK_EQ(asyncIO_->pending(), 0);
271 void AsyncIOQueue::submit(AsyncIOOp* op) {
272 submit([op]() { return op; });
275 void AsyncIOQueue::submit(OpFactory op) {
276 queue_.push_back(op);
280 void AsyncIOQueue::onCompleted(AsyncIOOp* op) {
284 void AsyncIOQueue::maybeDequeue() {
285 while (!queue_.empty() && asyncIO_->pending() < asyncIO_->capacity()) {
286 auto& opFactory = queue_.front();
287 auto op = opFactory();
290 // Interpose our completion callback
291 auto& nextCb = op->notificationCallback();
292 op->setNotificationCallback([this, nextCb](AsyncIOOp* op) {
293 this->onCompleted(op);
294 if (nextCb) nextCb(op);
297 asyncIO_->submit(op);
301 // debugging helpers:
305 #define X(c) case c: return #c
307 const char* asyncIoOpStateToString(AsyncIOOp::State state) {
309 X(AsyncIOOp::State::UNINITIALIZED);
310 X(AsyncIOOp::State::INITIALIZED);
311 X(AsyncIOOp::State::PENDING);
312 X(AsyncIOOp::State::COMPLETED);
314 return "<INVALID AsyncIOOp::State>";
317 const char* iocbCmdToString(short int cmd_short) {
318 io_iocb_cmd cmd = static_cast<io_iocb_cmd>(cmd_short);
329 return "<INVALID io_iocb_cmd>";
334 std::string fd2name(int fd) {
335 std::string path = folly::to<std::string>("/proc/self/fd/", fd);
337 const ssize_t length =
338 std::max<ssize_t>(readlink(path.c_str(), link, PATH_MAX), 0);
339 return path.assign(link, length);
342 std::ostream& operator<<(std::ostream& os, const iocb& cb) {
344 "data={}, key={}, opcode={}, reqprio={}, fd={}, f={}, ",
345 cb.data, cb.key, iocbCmdToString(cb.aio_lio_opcode),
346 cb.aio_reqprio, cb.aio_fildes, fd2name(cb.aio_fildes));
348 switch (cb.aio_lio_opcode) {
351 os << folly::format("buf={}, off={}, size={}, ",
352 cb.u.c.buf, cb.u.c.nbytes, cb.u.c.offset);
354 os << "[TODO: write debug string for "
355 << iocbCmdToString(cb.aio_lio_opcode) << "] ";
361 } // anonymous namespace
363 std::ostream& operator<<(std::ostream& os, const AsyncIOOp& op) {
364 os << "{" << op.state_ << ", ";
366 if (op.state_ != AsyncIOOp::State::UNINITIALIZED) {
370 if (op.state_ == AsyncIOOp::State::COMPLETED) {
371 os << "result=" << op.result_ << ", ";
377 std::ostream& operator<<(std::ostream& os, AsyncIOOp::State state) {
378 return os << asyncIoOpStateToString(state);