xref: /freebsd/tests/sys/fs/fusefs/write.cc (revision 7cc42f6d)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2019 The FreeBSD Foundation
5  *
6  * This software was developed by BFF Storage Systems, LLC under sponsorship
7  * from the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  * $FreeBSD$
31  */
32 
33 extern "C" {
34 #include <sys/param.h>
35 #include <sys/mman.h>
36 #include <sys/resource.h>
37 #include <sys/stat.h>
38 #include <sys/time.h>
39 #include <sys/uio.h>
40 
41 #include <aio.h>
42 #include <fcntl.h>
43 #include <signal.h>
44 #include <unistd.h>
45 }
46 
47 #include "mockfs.hh"
48 #include "utils.hh"
49 
50 using namespace testing;
51 
52 class Write: public FuseTest {
53 
54 public:
55 static sig_atomic_t s_sigxfsz;
56 
57 void SetUp() {
58 	s_sigxfsz = 0;
59 	FuseTest::SetUp();
60 }
61 
62 void TearDown() {
63 	struct sigaction sa;
64 
65 	bzero(&sa, sizeof(sa));
66 	sa.sa_handler = SIG_DFL;
67 	sigaction(SIGXFSZ, &sa, NULL);
68 
69 	FuseTest::TearDown();
70 }
71 
72 void expect_lookup(const char *relpath, uint64_t ino, uint64_t size)
73 {
74 	FuseTest::expect_lookup(relpath, ino, S_IFREG | 0644, size, 1);
75 }
76 
77 void expect_release(uint64_t ino, ProcessMockerT r)
78 {
79 	EXPECT_CALL(*m_mock, process(
80 		ResultOf([=](auto in) {
81 			return (in.header.opcode == FUSE_RELEASE &&
82 				in.header.nodeid == ino);
83 		}, Eq(true)),
84 		_)
85 	).WillRepeatedly(Invoke(r));
86 }
87 
88 void expect_write(uint64_t ino, uint64_t offset, uint64_t isize,
89 	uint64_t osize, const void *contents)
90 {
91 	FuseTest::expect_write(ino, offset, isize, osize, 0, 0, contents);
92 }
93 
94 /* Expect a write that may or may not come, depending on the cache mode */
95 void maybe_expect_write(uint64_t ino, uint64_t offset, uint64_t size,
96 	const void *contents)
97 {
98 	EXPECT_CALL(*m_mock, process(
99 		ResultOf([=](auto in) {
100 			const char *buf = (const char*)in.body.bytes +
101 				sizeof(struct fuse_write_in);
102 
103 			return (in.header.opcode == FUSE_WRITE &&
104 				in.header.nodeid == ino &&
105 				in.body.write.offset == offset  &&
106 				in.body.write.size == size &&
107 				0 == bcmp(buf, contents, size));
108 		}, Eq(true)),
109 		_)
110 	).Times(AtMost(1))
111 	.WillRepeatedly(Invoke(
112 		ReturnImmediate([=](auto in __unused, auto& out) {
113 			SET_OUT_HEADER_LEN(out, write);
114 			out.body.write.size = size;
115 		})
116 	));
117 }
118 
119 };
120 
121 sig_atomic_t Write::s_sigxfsz = 0;
122 
123 class Write_7_8: public FuseTest {
124 
125 public:
126 virtual void SetUp() {
127 	m_kernel_minor_version = 8;
128 	FuseTest::SetUp();
129 }
130 
131 void expect_lookup(const char *relpath, uint64_t ino, uint64_t size)
132 {
133 	FuseTest::expect_lookup_7_8(relpath, ino, S_IFREG | 0644, size, 1);
134 }
135 
136 };
137 
138 class AioWrite: public Write {
139 virtual void SetUp() {
140 	if (!is_unsafe_aio_enabled())
141 		GTEST_SKIP() <<
142 			"vfs.aio.enable_unsafe must be set for this test";
143 	FuseTest::SetUp();
144 }
145 };
146 
147 /* Tests for the writeback cache mode */
148 class WriteBack: public Write {
149 public:
150 virtual void SetUp() {
151 	m_init_flags |= FUSE_WRITEBACK_CACHE;
152 	FuseTest::SetUp();
153 	if (IsSkipped())
154 		return;
155 }
156 
157 void expect_write(uint64_t ino, uint64_t offset, uint64_t isize,
158 	uint64_t osize, const void *contents)
159 {
160 	FuseTest::expect_write(ino, offset, isize, osize, FUSE_WRITE_CACHE, 0,
161 		contents);
162 }
163 };
164 
165 class WriteBackAsync: public WriteBack {
166 public:
167 virtual void SetUp() {
168 	m_async = true;
169 	WriteBack::SetUp();
170 }
171 };
172 
173 class TimeGran: public WriteBackAsync, public WithParamInterface<unsigned> {
174 public:
175 virtual void SetUp() {
176 	m_time_gran = 1 << GetParam();
177 	WriteBackAsync::SetUp();
178 }
179 };
180 
181 /* Tests for clustered writes with WriteBack cacheing */
182 class WriteCluster: public WriteBack {
183 public:
184 virtual void SetUp() {
185 	m_async = true;
186 	m_maxwrite = 1 << 25;	// Anything larger than MAXPHYS will suffice
187 	WriteBack::SetUp();
188 	if (m_maxphys < 2 * DFLTPHYS)
189 		GTEST_SKIP() << "MAXPHYS must be at least twice DFLTPHYS"
190 			<< " for this test";
191 	if (m_maxphys < 2 * m_maxbcachebuf)
192 		GTEST_SKIP() << "MAXPHYS must be at least twice maxbcachebuf"
193 			<< " for this test";
194 }
195 };
196 
197 void sigxfsz_handler(int __unused sig) {
198 	Write::s_sigxfsz = 1;
199 }
200 
201 /* AIO writes need to set the header's pid field correctly */
202 /* https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=236379 */
203 TEST_F(AioWrite, DISABLED_aio_write)
204 {
205 	const char FULLPATH[] = "mountpoint/some_file.txt";
206 	const char RELPATH[] = "some_file.txt";
207 	const char *CONTENTS = "abcdefgh";
208 	uint64_t ino = 42;
209 	uint64_t offset = 4096;
210 	int fd;
211 	ssize_t bufsize = strlen(CONTENTS);
212 	struct aiocb iocb, *piocb;
213 
214 	expect_lookup(RELPATH, ino, 0);
215 	expect_open(ino, 0, 1);
216 	expect_write(ino, offset, bufsize, bufsize, CONTENTS);
217 
218 	fd = open(FULLPATH, O_WRONLY);
219 	EXPECT_LE(0, fd) << strerror(errno);
220 
221 	iocb.aio_nbytes = bufsize;
222 	iocb.aio_fildes = fd;
223 	iocb.aio_buf = __DECONST(void *, CONTENTS);
224 	iocb.aio_offset = offset;
225 	iocb.aio_sigevent.sigev_notify = SIGEV_NONE;
226 	ASSERT_EQ(0, aio_write(&iocb)) << strerror(errno);
227 	ASSERT_EQ(bufsize, aio_waitcomplete(&piocb, NULL)) << strerror(errno);
228 	leak(fd);
229 }
230 
231 /*
232  * When a file is opened with O_APPEND, we should forward that flag to
233  * FUSE_OPEN (tested by Open.o_append) but still attempt to calculate the
234  * offset internally.  That way we'll work both with filesystems that
235  * understand O_APPEND (and ignore the offset) and filesystems that don't (and
236  * simply use the offset).
237  *
238  * Note that verifying the O_APPEND flag in FUSE_OPEN is done in the
239  * Open.o_append test.
240  */
241 TEST_F(Write, append)
242 {
243 	const ssize_t BUFSIZE = 9;
244 	const char FULLPATH[] = "mountpoint/some_file.txt";
245 	const char RELPATH[] = "some_file.txt";
246 	const char CONTENTS[BUFSIZE] = "abcdefgh";
247 	uint64_t ino = 42;
248 	/*
249 	 * Set offset to a maxbcachebuf boundary so we don't need to RMW when
250 	 * using writeback caching
251 	 */
252 	uint64_t initial_offset = m_maxbcachebuf;
253 	int fd;
254 
255 	expect_lookup(RELPATH, ino, initial_offset);
256 	expect_open(ino, 0, 1);
257 	expect_write(ino, initial_offset, BUFSIZE, BUFSIZE, CONTENTS);
258 
259 	/* Must open O_RDWR or fuse(4) implicitly sets direct_io */
260 	fd = open(FULLPATH, O_RDWR | O_APPEND);
261 	EXPECT_LE(0, fd) << strerror(errno);
262 
263 	ASSERT_EQ(BUFSIZE, write(fd, CONTENTS, BUFSIZE)) << strerror(errno);
264 	leak(fd);
265 }
266 
267 /* If a file is cached, then appending to the end should not cause a read */
268 TEST_F(Write, append_to_cached)
269 {
270 	const ssize_t BUFSIZE = 9;
271 	const char FULLPATH[] = "mountpoint/some_file.txt";
272 	const char RELPATH[] = "some_file.txt";
273 	char *oldcontents, *oldbuf;
274 	const char CONTENTS[BUFSIZE] = "abcdefgh";
275 	uint64_t ino = 42;
276 	/*
277 	 * Set offset in between maxbcachebuf boundary to test buffer handling
278 	 */
279 	uint64_t oldsize = m_maxbcachebuf / 2;
280 	int fd;
281 
282 	oldcontents = (char*)calloc(1, oldsize);
283 	ASSERT_NE(nullptr, oldcontents) << strerror(errno);
284 	oldbuf = (char*)malloc(oldsize);
285 	ASSERT_NE(nullptr, oldbuf) << strerror(errno);
286 
287 	expect_lookup(RELPATH, ino, oldsize);
288 	expect_open(ino, 0, 1);
289 	expect_read(ino, 0, oldsize, oldsize, oldcontents);
290 	maybe_expect_write(ino, oldsize, BUFSIZE, CONTENTS);
291 
292 	/* Must open O_RDWR or fuse(4) implicitly sets direct_io */
293 	fd = open(FULLPATH, O_RDWR | O_APPEND);
294 	EXPECT_LE(0, fd) << strerror(errno);
295 
296 	/* Read the old data into the cache */
297 	ASSERT_EQ((ssize_t)oldsize, read(fd, oldbuf, oldsize))
298 		<< strerror(errno);
299 
300 	/* Write the new data.  There should be no more read operations */
301 	ASSERT_EQ(BUFSIZE, write(fd, CONTENTS, BUFSIZE)) << strerror(errno);
302 	leak(fd);
303 	free(oldbuf);
304 	free(oldcontents);
305 }
306 
307 TEST_F(Write, append_direct_io)
308 {
309 	const ssize_t BUFSIZE = 9;
310 	const char FULLPATH[] = "mountpoint/some_file.txt";
311 	const char RELPATH[] = "some_file.txt";
312 	const char CONTENTS[BUFSIZE] = "abcdefgh";
313 	uint64_t ino = 42;
314 	uint64_t initial_offset = 4096;
315 	int fd;
316 
317 	expect_lookup(RELPATH, ino, initial_offset);
318 	expect_open(ino, FOPEN_DIRECT_IO, 1);
319 	expect_write(ino, initial_offset, BUFSIZE, BUFSIZE, CONTENTS);
320 
321 	fd = open(FULLPATH, O_WRONLY | O_APPEND);
322 	EXPECT_LE(0, fd) << strerror(errno);
323 
324 	ASSERT_EQ(BUFSIZE, write(fd, CONTENTS, BUFSIZE)) << strerror(errno);
325 	leak(fd);
326 }
327 
328 /* A direct write should evict any overlapping cached data */
329 TEST_F(Write, direct_io_evicts_cache)
330 {
331 	const char FULLPATH[] = "mountpoint/some_file.txt";
332 	const char RELPATH[] = "some_file.txt";
333 	const char CONTENTS0[] = "abcdefgh";
334 	const char CONTENTS1[] = "ijklmnop";
335 	uint64_t ino = 42;
336 	int fd;
337 	ssize_t bufsize = strlen(CONTENTS0) + 1;
338 	char readbuf[bufsize];
339 
340 	expect_lookup(RELPATH, ino, bufsize);
341 	expect_open(ino, 0, 1);
342 	expect_read(ino, 0, bufsize, bufsize, CONTENTS0);
343 	expect_write(ino, 0, bufsize, bufsize, CONTENTS1);
344 
345 	fd = open(FULLPATH, O_RDWR);
346 	EXPECT_LE(0, fd) << strerror(errno);
347 
348 	// Prime cache
349 	ASSERT_EQ(bufsize, read(fd, readbuf, bufsize)) << strerror(errno);
350 
351 	// Write directly, evicting cache
352 	ASSERT_EQ(0, fcntl(fd, F_SETFL, O_DIRECT)) << strerror(errno);
353 	ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno);
354 	ASSERT_EQ(bufsize, write(fd, CONTENTS1, bufsize)) << strerror(errno);
355 
356 	// Read again.  Cache should be bypassed
357 	expect_read(ino, 0, bufsize, bufsize, CONTENTS1);
358 	ASSERT_EQ(0, fcntl(fd, F_SETFL, 0)) << strerror(errno);
359 	ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno);
360 	ASSERT_EQ(bufsize, read(fd, readbuf, bufsize)) << strerror(errno);
361 	ASSERT_STREQ(readbuf, CONTENTS1);
362 
363 	leak(fd);
364 }
365 
366 /*
367  * If the server doesn't return FOPEN_DIRECT_IO during FUSE_OPEN, then it's not
368  * allowed to return a short write for that file handle.  However, if it does
369  * then we should still do our darndest to handle it by resending the unwritten
370  * portion.
371  */
372 TEST_F(Write, indirect_io_short_write)
373 {
374 	const char FULLPATH[] = "mountpoint/some_file.txt";
375 	const char RELPATH[] = "some_file.txt";
376 	const char *CONTENTS = "abcdefghijklmnop";
377 	uint64_t ino = 42;
378 	int fd;
379 	ssize_t bufsize = strlen(CONTENTS);
380 	ssize_t bufsize0 = 11;
381 	ssize_t bufsize1 = strlen(CONTENTS) - bufsize0;
382 	const char *contents1 = CONTENTS + bufsize0;
383 
384 	expect_lookup(RELPATH, ino, 0);
385 	expect_open(ino, 0, 1);
386 	expect_write(ino, 0, bufsize, bufsize0, CONTENTS);
387 	expect_write(ino, bufsize0, bufsize1, bufsize1, contents1);
388 
389 	fd = open(FULLPATH, O_WRONLY);
390 	EXPECT_LE(0, fd) << strerror(errno);
391 
392 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
393 	leak(fd);
394 }
395 
396 /*
397  * When the direct_io option is used, filesystems are allowed to write less
398  * data than requested.  We should return the short write to userland.
399  */
400 TEST_F(Write, direct_io_short_write)
401 {
402 	const char FULLPATH[] = "mountpoint/some_file.txt";
403 	const char RELPATH[] = "some_file.txt";
404 	const char *CONTENTS = "abcdefghijklmnop";
405 	uint64_t ino = 42;
406 	int fd;
407 	ssize_t bufsize = strlen(CONTENTS);
408 	ssize_t halfbufsize = bufsize / 2;
409 
410 	expect_lookup(RELPATH, ino, 0);
411 	expect_open(ino, FOPEN_DIRECT_IO, 1);
412 	expect_write(ino, 0, bufsize, halfbufsize, CONTENTS);
413 
414 	fd = open(FULLPATH, O_WRONLY);
415 	EXPECT_LE(0, fd) << strerror(errno);
416 
417 	ASSERT_EQ(halfbufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
418 	leak(fd);
419 }
420 
421 /*
422  * An insidious edge case: the filesystem returns a short write, and the
423  * difference between what we requested and what it actually wrote crosses an
424  * iov element boundary
425  */
426 TEST_F(Write, direct_io_short_write_iov)
427 {
428 	const char FULLPATH[] = "mountpoint/some_file.txt";
429 	const char RELPATH[] = "some_file.txt";
430 	const char *CONTENTS0 = "abcdefgh";
431 	const char *CONTENTS1 = "ijklmnop";
432 	const char *EXPECTED0 = "abcdefghijklmnop";
433 	uint64_t ino = 42;
434 	int fd;
435 	ssize_t size0 = strlen(CONTENTS0) - 1;
436 	ssize_t size1 = strlen(CONTENTS1) + 1;
437 	ssize_t totalsize = size0 + size1;
438 	struct iovec iov[2];
439 
440 	expect_lookup(RELPATH, ino, 0);
441 	expect_open(ino, FOPEN_DIRECT_IO, 1);
442 	expect_write(ino, 0, totalsize, size0, EXPECTED0);
443 
444 	fd = open(FULLPATH, O_WRONLY);
445 	EXPECT_LE(0, fd) << strerror(errno);
446 
447 	iov[0].iov_base = __DECONST(void*, CONTENTS0);
448 	iov[0].iov_len = strlen(CONTENTS0);
449 	iov[1].iov_base = __DECONST(void*, CONTENTS1);
450 	iov[1].iov_len = strlen(CONTENTS1);
451 	ASSERT_EQ(size0, writev(fd, iov, 2)) << strerror(errno);
452 	leak(fd);
453 }
454 
455 /* fusefs should respect RLIMIT_FSIZE */
456 TEST_F(Write, rlimit_fsize)
457 {
458 	const char FULLPATH[] = "mountpoint/some_file.txt";
459 	const char RELPATH[] = "some_file.txt";
460 	const char *CONTENTS = "abcdefgh";
461 	struct rlimit rl;
462 	ssize_t bufsize = strlen(CONTENTS);
463 	off_t offset = 1'000'000'000;
464 	uint64_t ino = 42;
465 	int fd;
466 
467 	expect_lookup(RELPATH, ino, 0);
468 	expect_open(ino, 0, 1);
469 
470 	rl.rlim_cur = offset;
471 	rl.rlim_max = 10 * offset;
472 	ASSERT_EQ(0, setrlimit(RLIMIT_FSIZE, &rl)) << strerror(errno);
473 	ASSERT_NE(SIG_ERR, signal(SIGXFSZ, sigxfsz_handler)) << strerror(errno);
474 
475 	fd = open(FULLPATH, O_WRONLY);
476 
477 	EXPECT_LE(0, fd) << strerror(errno);
478 
479 	ASSERT_EQ(-1, pwrite(fd, CONTENTS, bufsize, offset));
480 	EXPECT_EQ(EFBIG, errno);
481 	EXPECT_EQ(1, s_sigxfsz);
482 	leak(fd);
483 }
484 
485 /*
486  * A short read indicates EOF.  Test that nothing bad happens if we get EOF
487  * during the R of a RMW operation.
488  */
489 TEST_F(Write, eof_during_rmw)
490 {
491 	const char FULLPATH[] = "mountpoint/some_file.txt";
492 	const char RELPATH[] = "some_file.txt";
493 	const char *CONTENTS = "abcdefgh";
494 	const char *INITIAL   = "XXXXXXXXXX";
495 	uint64_t ino = 42;
496 	uint64_t offset = 1;
497 	ssize_t bufsize = strlen(CONTENTS);
498 	off_t orig_fsize = 10;
499 	off_t truncated_fsize = 5;
500 	off_t final_fsize = bufsize;
501 	int fd;
502 
503 	FuseTest::expect_lookup(RELPATH, ino, S_IFREG | 0644, orig_fsize, 1);
504 	expect_open(ino, 0, 1);
505 	expect_read(ino, 0, orig_fsize, truncated_fsize, INITIAL, O_RDWR);
506 	expect_getattr(ino, truncated_fsize);
507 	expect_read(ino, 0, final_fsize, final_fsize, INITIAL, O_RDWR);
508 	maybe_expect_write(ino, offset, bufsize, CONTENTS);
509 
510 	fd = open(FULLPATH, O_RDWR);
511 	EXPECT_LE(0, fd) << strerror(errno);
512 
513 	ASSERT_EQ(bufsize, pwrite(fd, CONTENTS, bufsize, offset))
514 		<< strerror(errno);
515 	leak(fd);
516 }
517 
518 /*
519  * If the kernel cannot be sure which uid, gid, or pid was responsible for a
520  * write, then it must set the FUSE_WRITE_CACHE bit
521  */
522 /* https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=236378 */
523 TEST_F(Write, mmap)
524 {
525 	const char FULLPATH[] = "mountpoint/some_file.txt";
526 	const char RELPATH[] = "some_file.txt";
527 	const char *CONTENTS = "abcdefgh";
528 	uint64_t ino = 42;
529 	int fd;
530 	ssize_t bufsize = strlen(CONTENTS);
531 	void *p;
532 	uint64_t offset = 10;
533 	size_t len;
534 	void *zeros, *expected;
535 
536 	len = getpagesize();
537 
538 	zeros = calloc(1, len);
539 	ASSERT_NE(nullptr, zeros);
540 	expected = calloc(1, len);
541 	ASSERT_NE(nullptr, expected);
542 	memmove((uint8_t*)expected + offset, CONTENTS, bufsize);
543 
544 	expect_lookup(RELPATH, ino, len);
545 	expect_open(ino, 0, 1);
546 	expect_read(ino, 0, len, len, zeros);
547 	/*
548 	 * Writes from the pager may or may not be associated with the correct
549 	 * pid, so they must set FUSE_WRITE_CACHE.
550 	 */
551 	FuseTest::expect_write(ino, 0, len, len, FUSE_WRITE_CACHE, 0, expected);
552 	expect_flush(ino, 1, ReturnErrno(0));
553 	expect_release(ino, ReturnErrno(0));
554 
555 	fd = open(FULLPATH, O_RDWR);
556 	EXPECT_LE(0, fd) << strerror(errno);
557 
558 	p = mmap(NULL, len, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
559 	ASSERT_NE(MAP_FAILED, p) << strerror(errno);
560 
561 	memmove((uint8_t*)p + offset, CONTENTS, bufsize);
562 
563 	ASSERT_EQ(0, munmap(p, len)) << strerror(errno);
564 	close(fd);	// Write mmap'd data on close
565 
566 	free(expected);
567 	free(zeros);
568 
569 	leak(fd);
570 }
571 
572 TEST_F(Write, pwrite)
573 {
574 	const char FULLPATH[] = "mountpoint/some_file.txt";
575 	const char RELPATH[] = "some_file.txt";
576 	const char *CONTENTS = "abcdefgh";
577 	uint64_t ino = 42;
578 	uint64_t offset = m_maxbcachebuf;
579 	int fd;
580 	ssize_t bufsize = strlen(CONTENTS);
581 
582 	expect_lookup(RELPATH, ino, 0);
583 	expect_open(ino, 0, 1);
584 	expect_write(ino, offset, bufsize, bufsize, CONTENTS);
585 
586 	fd = open(FULLPATH, O_WRONLY);
587 	EXPECT_LE(0, fd) << strerror(errno);
588 
589 	ASSERT_EQ(bufsize, pwrite(fd, CONTENTS, bufsize, offset))
590 		<< strerror(errno);
591 	leak(fd);
592 }
593 
594 /* Writing a file should update its cached mtime and ctime */
595 TEST_F(Write, timestamps)
596 {
597 	const char FULLPATH[] = "mountpoint/some_file.txt";
598 	const char RELPATH[] = "some_file.txt";
599 	const char *CONTENTS = "abcdefgh";
600 	ssize_t bufsize = strlen(CONTENTS);
601 	uint64_t ino = 42;
602 	struct stat sb0, sb1;
603 	int fd;
604 
605 	expect_lookup(RELPATH, ino, 0);
606 	expect_open(ino, 0, 1);
607 	maybe_expect_write(ino, 0, bufsize, CONTENTS);
608 
609 	fd = open(FULLPATH, O_RDWR);
610 	EXPECT_LE(0, fd) << strerror(errno);
611 	ASSERT_EQ(0, fstat(fd, &sb0)) << strerror(errno);
612 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
613 
614 	nap();
615 
616 	ASSERT_EQ(0, fstat(fd, &sb1)) << strerror(errno);
617 
618 	EXPECT_EQ(sb0.st_atime, sb1.st_atime);
619 	EXPECT_NE(sb0.st_mtime, sb1.st_mtime);
620 	EXPECT_NE(sb0.st_ctime, sb1.st_ctime);
621 
622 	leak(fd);
623 }
624 
625 TEST_F(Write, write)
626 {
627 	const char FULLPATH[] = "mountpoint/some_file.txt";
628 	const char RELPATH[] = "some_file.txt";
629 	const char *CONTENTS = "abcdefgh";
630 	uint64_t ino = 42;
631 	int fd;
632 	ssize_t bufsize = strlen(CONTENTS);
633 
634 	expect_lookup(RELPATH, ino, 0);
635 	expect_open(ino, 0, 1);
636 	expect_write(ino, 0, bufsize, bufsize, CONTENTS);
637 
638 	fd = open(FULLPATH, O_WRONLY);
639 	EXPECT_LE(0, fd) << strerror(errno);
640 
641 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
642 	leak(fd);
643 }
644 
645 /* fuse(4) should not issue writes of greater size than the daemon requests */
646 TEST_F(Write, write_large)
647 {
648 	const char FULLPATH[] = "mountpoint/some_file.txt";
649 	const char RELPATH[] = "some_file.txt";
650 	int *contents;
651 	uint64_t ino = 42;
652 	int fd;
653 	ssize_t halfbufsize, bufsize;
654 
655 	halfbufsize = m_mock->m_maxwrite;
656 	bufsize = halfbufsize * 2;
657 	contents = (int*)malloc(bufsize);
658 	ASSERT_NE(nullptr, contents);
659 	for (int i = 0; i < (int)bufsize / (int)sizeof(i); i++) {
660 		contents[i] = i;
661 	}
662 
663 	expect_lookup(RELPATH, ino, 0);
664 	expect_open(ino, 0, 1);
665 	maybe_expect_write(ino, 0, halfbufsize, contents);
666 	maybe_expect_write(ino, halfbufsize, halfbufsize,
667 		&contents[halfbufsize / sizeof(int)]);
668 
669 	fd = open(FULLPATH, O_WRONLY);
670 	EXPECT_LE(0, fd) << strerror(errno);
671 
672 	ASSERT_EQ(bufsize, write(fd, contents, bufsize)) << strerror(errno);
673 	leak(fd);
674 
675 	free(contents);
676 }
677 
678 TEST_F(Write, write_nothing)
679 {
680 	const char FULLPATH[] = "mountpoint/some_file.txt";
681 	const char RELPATH[] = "some_file.txt";
682 	const char *CONTENTS = "";
683 	uint64_t ino = 42;
684 	int fd;
685 	ssize_t bufsize = 0;
686 
687 	expect_lookup(RELPATH, ino, 0);
688 	expect_open(ino, 0, 1);
689 
690 	fd = open(FULLPATH, O_WRONLY);
691 	EXPECT_LE(0, fd) << strerror(errno);
692 
693 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
694 	leak(fd);
695 }
696 
697 TEST_F(Write_7_8, write)
698 {
699 	const char FULLPATH[] = "mountpoint/some_file.txt";
700 	const char RELPATH[] = "some_file.txt";
701 	const char *CONTENTS = "abcdefgh";
702 	uint64_t ino = 42;
703 	int fd;
704 	ssize_t bufsize = strlen(CONTENTS);
705 
706 	expect_lookup(RELPATH, ino, 0);
707 	expect_open(ino, 0, 1);
708 	expect_write_7_8(ino, 0, bufsize, bufsize, CONTENTS);
709 
710 	fd = open(FULLPATH, O_WRONLY);
711 	EXPECT_LE(0, fd) << strerror(errno);
712 
713 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
714 	leak(fd);
715 }
716 
717 /* In writeback mode, dirty data should be written on close */
718 TEST_F(WriteBackAsync, close)
719 {
720 	const char FULLPATH[] = "mountpoint/some_file.txt";
721 	const char RELPATH[] = "some_file.txt";
722 	const char *CONTENTS = "abcdefgh";
723 	uint64_t ino = 42;
724 	int fd;
725 	ssize_t bufsize = strlen(CONTENTS);
726 
727 	expect_lookup(RELPATH, ino, 0);
728 	expect_open(ino, 0, 1);
729 	expect_write(ino, 0, bufsize, bufsize, CONTENTS);
730 	EXPECT_CALL(*m_mock, process(
731 		ResultOf([=](auto in) {
732 			return (in.header.opcode == FUSE_SETATTR);
733 		}, Eq(true)),
734 		_)
735 	).WillRepeatedly(Invoke(ReturnImmediate([=](auto i __unused, auto& out) {
736 		SET_OUT_HEADER_LEN(out, attr);
737 		out.body.attr.attr.ino = ino;	// Must match nodeid
738 	})));
739 	expect_flush(ino, 1, ReturnErrno(0));
740 	expect_release(ino, ReturnErrno(0));
741 
742 	fd = open(FULLPATH, O_RDWR);
743 	ASSERT_LE(0, fd) << strerror(errno);
744 
745 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
746 	close(fd);
747 }
748 
749 /* In writeback mode, adjacent writes will be clustered together */
750 TEST_F(WriteCluster, clustering)
751 {
752 	const char FULLPATH[] = "mountpoint/some_file.txt";
753 	const char RELPATH[] = "some_file.txt";
754 	uint64_t ino = 42;
755 	int i, fd;
756 	void *wbuf, *wbuf2x;
757 	ssize_t bufsize = m_maxbcachebuf;
758 	off_t filesize = 5 * bufsize;
759 
760 	wbuf = malloc(bufsize);
761 	ASSERT_NE(nullptr, wbuf) << strerror(errno);
762 	memset(wbuf, 'X', bufsize);
763 	wbuf2x = malloc(2 * bufsize);
764 	ASSERT_NE(nullptr, wbuf2x) << strerror(errno);
765 	memset(wbuf2x, 'X', 2 * bufsize);
766 
767 	expect_lookup(RELPATH, ino, filesize);
768 	expect_open(ino, 0, 1);
769 	/*
770 	 * Writes of bufsize-bytes each should be clustered into greater sizes.
771 	 * The amount of clustering is adaptive, so the first write actually
772 	 * issued will be 2x bufsize and subsequent writes may be larger
773 	 */
774 	expect_write(ino, 0, 2 * bufsize, 2 * bufsize, wbuf2x);
775 	expect_write(ino, 2 * bufsize, 2 * bufsize, 2 * bufsize, wbuf2x);
776 	expect_flush(ino, 1, ReturnErrno(0));
777 	expect_release(ino, ReturnErrno(0));
778 
779 	fd = open(FULLPATH, O_RDWR);
780 	ASSERT_LE(0, fd) << strerror(errno);
781 
782 	for (i = 0; i < 4; i++) {
783 		ASSERT_EQ(bufsize, write(fd, wbuf, bufsize))
784 			<< strerror(errno);
785 	}
786 	close(fd);
787 	free(wbuf2x);
788 	free(wbuf);
789 }
790 
791 /*
792  * When clustering writes, an I/O error to any of the cluster's children should
793  * not panic the system on unmount
794  */
795 /*
796  * Disabled because it panics.
797  * https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=238565
798  */
799 TEST_F(WriteCluster, DISABLED_cluster_write_err)
800 {
801 	const char FULLPATH[] = "mountpoint/some_file.txt";
802 	const char RELPATH[] = "some_file.txt";
803 	uint64_t ino = 42;
804 	int i, fd;
805 	void *wbuf;
806 	ssize_t bufsize = m_maxbcachebuf;
807 	off_t filesize = 4 * bufsize;
808 
809 	wbuf = malloc(bufsize);
810 	ASSERT_NE(nullptr, wbuf) << strerror(errno);
811 	memset(wbuf, 'X', bufsize);
812 
813 	expect_lookup(RELPATH, ino, filesize);
814 	expect_open(ino, 0, 1);
815 	EXPECT_CALL(*m_mock, process(
816 		ResultOf([=](auto in) {
817 			return (in.header.opcode == FUSE_WRITE);
818 		}, Eq(true)),
819 		_)
820 	).WillRepeatedly(Invoke(ReturnErrno(EIO)));
821 	expect_flush(ino, 1, ReturnErrno(0));
822 	expect_release(ino, ReturnErrno(0));
823 
824 	fd = open(FULLPATH, O_RDWR);
825 	ASSERT_LE(0, fd) << strerror(errno);
826 
827 	for (i = 0; i < 3; i++) {
828 		ASSERT_EQ(bufsize, write(fd, wbuf, bufsize))
829 			<< strerror(errno);
830 	}
831 	close(fd);
832 	free(wbuf);
833 }
834 
835 /*
836  * In writeback mode, writes to an O_WRONLY file could trigger reads from the
837  * server.  The FUSE protocol explicitly allows that.
838  */
839 TEST_F(WriteBack, rmw)
840 {
841 	const char FULLPATH[] = "mountpoint/some_file.txt";
842 	const char RELPATH[] = "some_file.txt";
843 	const char *CONTENTS = "abcdefgh";
844 	const char *INITIAL   = "XXXXXXXXXX";
845 	uint64_t ino = 42;
846 	uint64_t offset = 1;
847 	off_t fsize = 10;
848 	int fd;
849 	ssize_t bufsize = strlen(CONTENTS);
850 
851 	FuseTest::expect_lookup(RELPATH, ino, S_IFREG | 0644, fsize, 1);
852 	expect_open(ino, 0, 1);
853 	expect_read(ino, 0, fsize, fsize, INITIAL, O_WRONLY);
854 	maybe_expect_write(ino, offset, bufsize, CONTENTS);
855 
856 	fd = open(FULLPATH, O_WRONLY);
857 	EXPECT_LE(0, fd) << strerror(errno);
858 
859 	ASSERT_EQ(bufsize, pwrite(fd, CONTENTS, bufsize, offset))
860 		<< strerror(errno);
861 	leak(fd);
862 }
863 
864 /*
865  * Without direct_io, writes should be committed to cache
866  */
867 TEST_F(WriteBack, cache)
868 {
869 	const char FULLPATH[] = "mountpoint/some_file.txt";
870 	const char RELPATH[] = "some_file.txt";
871 	const char *CONTENTS = "abcdefgh";
872 	uint64_t ino = 42;
873 	int fd;
874 	ssize_t bufsize = strlen(CONTENTS);
875 	uint8_t readbuf[bufsize];
876 
877 	expect_lookup(RELPATH, ino, 0);
878 	expect_open(ino, 0, 1);
879 	expect_write(ino, 0, bufsize, bufsize, CONTENTS);
880 
881 	fd = open(FULLPATH, O_RDWR);
882 	EXPECT_LE(0, fd) << strerror(errno);
883 
884 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
885 	/*
886 	 * A subsequent read should be serviced by cache, without querying the
887 	 * filesystem daemon
888 	 */
889 	ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno);
890 	ASSERT_EQ(bufsize, read(fd, readbuf, bufsize)) << strerror(errno);
891 	leak(fd);
892 }
893 
894 /*
895  * With O_DIRECT, writes should be not committed to cache.  Admittedly this is
896  * an odd test, because it would be unusual to use O_DIRECT for writes but not
897  * reads.
898  */
899 TEST_F(WriteBack, o_direct)
900 {
901 	const char FULLPATH[] = "mountpoint/some_file.txt";
902 	const char RELPATH[] = "some_file.txt";
903 	const char *CONTENTS = "abcdefgh";
904 	uint64_t ino = 42;
905 	int fd;
906 	ssize_t bufsize = strlen(CONTENTS);
907 	uint8_t readbuf[bufsize];
908 
909 	expect_lookup(RELPATH, ino, 0);
910 	expect_open(ino, 0, 1);
911 	FuseTest::expect_write(ino, 0, bufsize, bufsize, 0, FUSE_WRITE_CACHE,
912 		CONTENTS);
913 	expect_read(ino, 0, bufsize, bufsize, CONTENTS);
914 
915 	fd = open(FULLPATH, O_RDWR | O_DIRECT);
916 	EXPECT_LE(0, fd) << strerror(errno);
917 
918 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
919 	/* A subsequent read must query the daemon because cache is empty */
920 	ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno);
921 	ASSERT_EQ(0, fcntl(fd, F_SETFL, 0)) << strerror(errno);
922 	ASSERT_EQ(bufsize, read(fd, readbuf, bufsize)) << strerror(errno);
923 	leak(fd);
924 }
925 
926 TEST_F(WriteBack, direct_io)
927 {
928 	const char FULLPATH[] = "mountpoint/some_file.txt";
929 	const char RELPATH[] = "some_file.txt";
930 	const char *CONTENTS = "abcdefgh";
931 	uint64_t ino = 42;
932 	int fd;
933 	ssize_t bufsize = strlen(CONTENTS);
934 	uint8_t readbuf[bufsize];
935 
936 	expect_lookup(RELPATH, ino, 0);
937 	expect_open(ino, FOPEN_DIRECT_IO, 1);
938 	FuseTest::expect_write(ino, 0, bufsize, bufsize, 0, FUSE_WRITE_CACHE,
939 		CONTENTS);
940 	expect_read(ino, 0, bufsize, bufsize, CONTENTS);
941 
942 	fd = open(FULLPATH, O_RDWR);
943 	EXPECT_LE(0, fd) << strerror(errno);
944 
945 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
946 	/* A subsequent read must query the daemon because cache is empty */
947 	ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno);
948 	ASSERT_EQ(0, fcntl(fd, F_SETFL, 0)) << strerror(errno);
949 	ASSERT_EQ(bufsize, read(fd, readbuf, bufsize)) << strerror(errno);
950 	leak(fd);
951 }
952 
953 /*
954  * mmap should still be possible even if the server used direct_io.  Mmap will
955  * still use the cache, though.
956  *
957  * Regression test for bug 247276
958  * https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=247276
959  */
960 TEST_F(WriteBack, mmap_direct_io)
961 {
962 	const char FULLPATH[] = "mountpoint/some_file.txt";
963 	const char RELPATH[] = "some_file.txt";
964 	const char *CONTENTS = "abcdefgh";
965 	uint64_t ino = 42;
966 	int fd;
967 	size_t len;
968 	ssize_t bufsize = strlen(CONTENTS);
969 	void *p, *zeros;
970 
971 	len = getpagesize();
972 	zeros = calloc(1, len);
973 	ASSERT_NE(nullptr, zeros);
974 
975 	expect_lookup(RELPATH, ino, len);
976 	expect_open(ino, FOPEN_DIRECT_IO, 1);
977 	expect_read(ino, 0, len, len, zeros);
978 	expect_flush(ino, 1, ReturnErrno(0));
979 	FuseTest::expect_write(ino, 0, len, len, FUSE_WRITE_CACHE, 0, zeros);
980 	expect_release(ino, ReturnErrno(0));
981 
982 	fd = open(FULLPATH, O_RDWR);
983 	EXPECT_LE(0, fd) << strerror(errno);
984 
985 	p = mmap(NULL, len, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
986 	ASSERT_NE(MAP_FAILED, p) << strerror(errno);
987 
988 	memmove((uint8_t*)p, CONTENTS, bufsize);
989 
990 	ASSERT_EQ(0, munmap(p, len)) << strerror(errno);
991 	close(fd);	// Write mmap'd data on close
992 
993 	free(zeros);
994 }
995 
996 /*
997  * When mounted with -o async, the writeback cache mode should delay writes
998  */
999 TEST_F(WriteBackAsync, delay)
1000 {
1001 	const char FULLPATH[] = "mountpoint/some_file.txt";
1002 	const char RELPATH[] = "some_file.txt";
1003 	const char *CONTENTS = "abcdefgh";
1004 	uint64_t ino = 42;
1005 	int fd;
1006 	ssize_t bufsize = strlen(CONTENTS);
1007 
1008 	expect_lookup(RELPATH, ino, 0);
1009 	expect_open(ino, 0, 1);
1010 	/* Write should be cached, but FUSE_WRITE shouldn't be sent */
1011 	EXPECT_CALL(*m_mock, process(
1012 		ResultOf([=](auto in) {
1013 			return (in.header.opcode == FUSE_WRITE);
1014 		}, Eq(true)),
1015 		_)
1016 	).Times(0);
1017 
1018 	fd = open(FULLPATH, O_RDWR);
1019 	EXPECT_LE(0, fd) << strerror(errno);
1020 
1021 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
1022 
1023 	/* Don't close the file because that would flush the cache */
1024 	leak(fd);
1025 }
1026 
1027 /*
1028  * A direct write should not evict dirty cached data from outside of its own
1029  * byte range.
1030  */
1031 TEST_F(WriteBackAsync, direct_io_ignores_unrelated_cached)
1032 {
1033 	const char FULLPATH[] = "mountpoint/some_file.txt";
1034 	const char RELPATH[] = "some_file.txt";
1035 	const char CONTENTS0[] = "abcdefgh";
1036 	const char CONTENTS1[] = "ijklmnop";
1037 	uint64_t ino = 42;
1038 	int fd;
1039 	ssize_t bufsize = strlen(CONTENTS0) + 1;
1040 	ssize_t fsize = 2 * m_maxbcachebuf;
1041 	char readbuf[bufsize];
1042 	void *zeros;
1043 
1044 	zeros = calloc(1, m_maxbcachebuf);
1045 	ASSERT_NE(nullptr, zeros);
1046 
1047 	expect_lookup(RELPATH, ino, fsize);
1048 	expect_open(ino, 0, 1);
1049 	expect_read(ino, 0, m_maxbcachebuf, m_maxbcachebuf, zeros);
1050 	FuseTest::expect_write(ino, m_maxbcachebuf, bufsize, bufsize, 0, 0,
1051 		CONTENTS1);
1052 
1053 	fd = open(FULLPATH, O_RDWR);
1054 	EXPECT_LE(0, fd) << strerror(errno);
1055 
1056 	// Cache first block with dirty data.  This will entail first reading
1057 	// the existing data.
1058 	ASSERT_EQ(bufsize, pwrite(fd, CONTENTS0, bufsize, 0))
1059 		<< strerror(errno);
1060 
1061 	// Write directly to second block
1062 	ASSERT_EQ(0, fcntl(fd, F_SETFL, O_DIRECT)) << strerror(errno);
1063 	ASSERT_EQ(bufsize, pwrite(fd, CONTENTS1, bufsize, m_maxbcachebuf))
1064 		<< strerror(errno);
1065 
1066 	// Read from the first block again.  Should be serviced by cache.
1067 	ASSERT_EQ(0, fcntl(fd, F_SETFL, 0)) << strerror(errno);
1068 	ASSERT_EQ(bufsize, pread(fd, readbuf, bufsize, 0)) << strerror(errno);
1069 	ASSERT_STREQ(readbuf, CONTENTS0);
1070 
1071 	leak(fd);
1072 	free(zeros);
1073 }
1074 
1075 /*
1076  * If a direct io write partially overlaps one or two blocks of dirty cached
1077  * data, No dirty data should be lost.  Admittedly this is a weird test,
1078  * because it would be unusual to use O_DIRECT and the writeback cache.
1079  */
1080 TEST_F(WriteBackAsync, direct_io_partially_overlaps_cached_block)
1081 {
1082 	const char FULLPATH[] = "mountpoint/some_file.txt";
1083 	const char RELPATH[] = "some_file.txt";
1084 	uint64_t ino = 42;
1085 	int fd;
1086 	off_t bs = m_maxbcachebuf;
1087 	ssize_t fsize = 3 * bs;
1088 	void *readbuf, *zeros, *ones, *zeroones, *onezeros;
1089 
1090 	readbuf = malloc(bs);
1091 	ASSERT_NE(nullptr, readbuf) << strerror(errno);
1092 	zeros = calloc(1, 3 * bs);
1093 	ASSERT_NE(nullptr, zeros);
1094 	ones = calloc(1, 2 * bs);
1095 	ASSERT_NE(nullptr, ones);
1096 	memset(ones, 1, 2 * bs);
1097 	zeroones = calloc(1, bs);
1098 	ASSERT_NE(nullptr, zeroones);
1099 	memset((uint8_t*)zeroones + bs / 2, 1, bs / 2);
1100 	onezeros = calloc(1, bs);
1101 	ASSERT_NE(nullptr, onezeros);
1102 	memset(onezeros, 1, bs / 2);
1103 
1104 	expect_lookup(RELPATH, ino, fsize);
1105 	expect_open(ino, 0, 1);
1106 
1107 	fd = open(FULLPATH, O_RDWR);
1108 	EXPECT_LE(0, fd) << strerror(errno);
1109 
1110 	/* Cache first and third blocks with dirty data.  */
1111 	ASSERT_EQ(3 * bs, pwrite(fd, zeros, 3 * bs, 0)) << strerror(errno);
1112 
1113 	/*
1114 	 * Write directly to all three blocks.  The partially written blocks
1115 	 * will be flushed because they're dirty.
1116 	 */
1117 	FuseTest::expect_write(ino, 0, bs, bs, 0, 0, zeros);
1118 	FuseTest::expect_write(ino, 2 * bs, bs, bs, 0, 0, zeros);
1119 	/* The direct write is split in two because of the m_maxwrite value */
1120 	FuseTest::expect_write(ino,     bs / 2, bs, bs, 0, 0, ones);
1121 	FuseTest::expect_write(ino, 3 * bs / 2, bs, bs, 0, 0, ones);
1122 	ASSERT_EQ(0, fcntl(fd, F_SETFL, O_DIRECT)) << strerror(errno);
1123 	ASSERT_EQ(2 * bs, pwrite(fd, ones, 2 * bs, bs / 2)) << strerror(errno);
1124 
1125 	/*
1126 	 * Read from both the valid and invalid portions of the first and third
1127 	 * blocks again.  This will entail FUSE_READ operations because these
1128 	 * blocks were invalidated by the direct write.
1129 	 */
1130 	expect_read(ino, 0, bs, bs, zeroones);
1131 	expect_read(ino, 2 * bs, bs, bs, onezeros);
1132 	ASSERT_EQ(0, fcntl(fd, F_SETFL, 0)) << strerror(errno);
1133 	ASSERT_EQ(bs / 2, pread(fd, readbuf, bs / 2, 0)) << strerror(errno);
1134 	EXPECT_EQ(0, memcmp(zeros, readbuf, bs / 2));
1135 	ASSERT_EQ(bs / 2, pread(fd, readbuf, bs / 2, 5 * bs / 2))
1136 		<< strerror(errno);
1137 	EXPECT_EQ(0, memcmp(zeros, readbuf, bs / 2));
1138 	ASSERT_EQ(bs / 2, pread(fd, readbuf, bs / 2, bs / 2))
1139 		<< strerror(errno);
1140 	EXPECT_EQ(0, memcmp(ones, readbuf, bs / 2));
1141 	ASSERT_EQ(bs / 2, pread(fd, readbuf, bs / 2, 2 * bs))
1142 		<< strerror(errno);
1143 	EXPECT_EQ(0, memcmp(ones, readbuf, bs / 2));
1144 
1145 	leak(fd);
1146 	free(zeroones);
1147 	free(onezeros);
1148 	free(ones);
1149 	free(zeros);
1150 	free(readbuf);
1151 }
1152 
1153 /*
1154  * In WriteBack mode, writes may be cached beyond what the server thinks is the
1155  * EOF.  In this case, a short read at EOF should _not_ cause fusefs to update
1156  * the file's size.
1157  */
1158 TEST_F(WriteBackAsync, eof)
1159 {
1160 	const char FULLPATH[] = "mountpoint/some_file.txt";
1161 	const char RELPATH[] = "some_file.txt";
1162 	const char *CONTENTS0 = "abcdefgh";
1163 	const char *CONTENTS1 = "ijklmnop";
1164 	uint64_t ino = 42;
1165 	int fd;
1166 	off_t offset = m_maxbcachebuf;
1167 	ssize_t wbufsize = strlen(CONTENTS1);
1168 	off_t old_filesize = (off_t)strlen(CONTENTS0);
1169 	ssize_t rbufsize = 2 * old_filesize;
1170 	char readbuf[rbufsize];
1171 	size_t holesize = rbufsize - old_filesize;
1172 	char hole[holesize];
1173 	struct stat sb;
1174 	ssize_t r;
1175 
1176 	expect_lookup(RELPATH, ino, 0);
1177 	expect_open(ino, 0, 1);
1178 	expect_read(ino, 0, m_maxbcachebuf, old_filesize, CONTENTS0);
1179 
1180 	fd = open(FULLPATH, O_RDWR);
1181 	EXPECT_LE(0, fd) << strerror(errno);
1182 
1183 	/* Write and cache data beyond EOF */
1184 	ASSERT_EQ(wbufsize, pwrite(fd, CONTENTS1, wbufsize, offset))
1185 		<< strerror(errno);
1186 
1187 	/* Read from the old EOF */
1188 	r = pread(fd, readbuf, rbufsize, 0);
1189 	ASSERT_LE(0, r) << strerror(errno);
1190 	EXPECT_EQ(rbufsize, r) << "read should've synthesized a hole";
1191 	EXPECT_EQ(0, memcmp(CONTENTS0, readbuf, old_filesize));
1192 	bzero(hole, holesize);
1193 	EXPECT_EQ(0, memcmp(hole, readbuf + old_filesize, holesize));
1194 
1195 	/* The file's size should still be what was established by pwrite */
1196 	ASSERT_EQ(0, fstat(fd, &sb)) << strerror(errno);
1197 	EXPECT_EQ(offset + wbufsize, sb.st_size);
1198 	leak(fd);
1199 }
1200 
1201 /*
1202  * When a file has dirty writes that haven't been flushed, the server's notion
1203  * of its mtime and ctime will be wrong.  The kernel should ignore those if it
1204  * gets them from a FUSE_GETATTR before flushing.
1205  */
1206 TEST_F(WriteBackAsync, timestamps)
1207 {
1208 	const char FULLPATH[] = "mountpoint/some_file.txt";
1209 	const char RELPATH[] = "some_file.txt";
1210 	const char *CONTENTS = "abcdefgh";
1211 	ssize_t bufsize = strlen(CONTENTS);
1212 	uint64_t ino = 42;
1213 	uint64_t attr_valid = 0;
1214 	uint64_t attr_valid_nsec = 0;
1215 	uint64_t server_time = 12345;
1216 	mode_t mode = S_IFREG | 0644;
1217 	int fd;
1218 
1219 	struct stat sb;
1220 
1221 	EXPECT_LOOKUP(FUSE_ROOT_ID, RELPATH)
1222 	.WillRepeatedly(Invoke(
1223 		ReturnImmediate([=](auto in __unused, auto& out) {
1224 		SET_OUT_HEADER_LEN(out, entry);
1225 		out.body.entry.attr.mode = mode;
1226 		out.body.entry.nodeid = ino;
1227 		out.body.entry.attr.nlink = 1;
1228 		out.body.entry.attr_valid = attr_valid;
1229 		out.body.entry.attr_valid_nsec = attr_valid_nsec;
1230 	})));
1231 	expect_open(ino, 0, 1);
1232 	EXPECT_CALL(*m_mock, process(
1233 		ResultOf([=](auto in) {
1234 			return (in.header.opcode == FUSE_GETATTR &&
1235 				in.header.nodeid == ino);
1236 		}, Eq(true)),
1237 		_)
1238 	).WillRepeatedly(Invoke(
1239 	ReturnImmediate([=](auto i __unused, auto& out) {
1240 		SET_OUT_HEADER_LEN(out, attr);
1241 		out.body.attr.attr.ino = ino;
1242 		out.body.attr.attr.mode = mode;
1243 		out.body.attr.attr_valid = attr_valid;
1244 		out.body.attr.attr_valid_nsec = attr_valid_nsec;
1245 		out.body.attr.attr.atime = server_time;
1246 		out.body.attr.attr.mtime = server_time;
1247 		out.body.attr.attr.ctime = server_time;
1248 	})));
1249 
1250 	fd = open(FULLPATH, O_RDWR);
1251 	EXPECT_LE(0, fd) << strerror(errno);
1252 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
1253 
1254 	ASSERT_EQ(0, fstat(fd, &sb)) << strerror(errno);
1255 	EXPECT_EQ((time_t)server_time, sb.st_atime);
1256 	EXPECT_NE((time_t)server_time, sb.st_mtime);
1257 	EXPECT_NE((time_t)server_time, sb.st_ctime);
1258 
1259 	leak(fd);
1260 }
1261 
1262 /* Any dirty timestamp fields should be flushed during a SETATTR */
1263 TEST_F(WriteBackAsync, timestamps_during_setattr)
1264 {
1265 	const char FULLPATH[] = "mountpoint/some_file.txt";
1266 	const char RELPATH[] = "some_file.txt";
1267 	const char *CONTENTS = "abcdefgh";
1268 	ssize_t bufsize = strlen(CONTENTS);
1269 	uint64_t ino = 42;
1270 	const mode_t newmode = 0755;
1271 	int fd;
1272 
1273 	expect_lookup(RELPATH, ino, 0);
1274 	expect_open(ino, 0, 1);
1275 	EXPECT_CALL(*m_mock, process(
1276 		ResultOf([=](auto in) {
1277 			uint32_t valid = FATTR_MODE | FATTR_MTIME | FATTR_CTIME;
1278 			return (in.header.opcode == FUSE_SETATTR &&
1279 				in.header.nodeid == ino &&
1280 				in.body.setattr.valid == valid);
1281 		}, Eq(true)),
1282 		_)
1283 	).WillOnce(Invoke(ReturnImmediate([=](auto in __unused, auto& out) {
1284 		SET_OUT_HEADER_LEN(out, attr);
1285 		out.body.attr.attr.ino = ino;
1286 		out.body.attr.attr.mode = S_IFREG | newmode;
1287 	})));
1288 
1289 	fd = open(FULLPATH, O_RDWR);
1290 	EXPECT_LE(0, fd) << strerror(errno);
1291 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
1292 	ASSERT_EQ(0, fchmod(fd, newmode)) << strerror(errno);
1293 
1294 	leak(fd);
1295 }
1296 
1297 /* fuse_init_out.time_gran controls the granularity of timestamps */
1298 TEST_P(TimeGran, timestamps_during_setattr)
1299 {
1300 	const char FULLPATH[] = "mountpoint/some_file.txt";
1301 	const char RELPATH[] = "some_file.txt";
1302 	const char *CONTENTS = "abcdefgh";
1303 	ssize_t bufsize = strlen(CONTENTS);
1304 	uint64_t ino = 42;
1305 	const mode_t newmode = 0755;
1306 	int fd;
1307 
1308 	expect_lookup(RELPATH, ino, 0);
1309 	expect_open(ino, 0, 1);
1310 	EXPECT_CALL(*m_mock, process(
1311 		ResultOf([=](auto in) {
1312 			uint32_t valid = FATTR_MODE | FATTR_MTIME | FATTR_CTIME;
1313 			return (in.header.opcode == FUSE_SETATTR &&
1314 				in.header.nodeid == ino &&
1315 				in.body.setattr.valid == valid &&
1316 				in.body.setattr.mtimensec % m_time_gran == 0 &&
1317 				in.body.setattr.ctimensec % m_time_gran == 0);
1318 		}, Eq(true)),
1319 		_)
1320 	).WillOnce(Invoke(ReturnImmediate([=](auto in __unused, auto& out) {
1321 		SET_OUT_HEADER_LEN(out, attr);
1322 		out.body.attr.attr.ino = ino;
1323 		out.body.attr.attr.mode = S_IFREG | newmode;
1324 	})));
1325 
1326 	fd = open(FULLPATH, O_RDWR);
1327 	EXPECT_LE(0, fd) << strerror(errno);
1328 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
1329 	ASSERT_EQ(0, fchmod(fd, newmode)) << strerror(errno);
1330 
1331 	leak(fd);
1332 }
1333 
1334 INSTANTIATE_TEST_CASE_P(RA, TimeGran, Range(0u, 10u));
1335 
1336 /*
1337  * Without direct_io, writes should be committed to cache
1338  */
1339 TEST_F(Write, writethrough)
1340 {
1341 	const char FULLPATH[] = "mountpoint/some_file.txt";
1342 	const char RELPATH[] = "some_file.txt";
1343 	const char *CONTENTS = "abcdefgh";
1344 	uint64_t ino = 42;
1345 	int fd;
1346 	ssize_t bufsize = strlen(CONTENTS);
1347 	uint8_t readbuf[bufsize];
1348 
1349 	expect_lookup(RELPATH, ino, 0);
1350 	expect_open(ino, 0, 1);
1351 	expect_write(ino, 0, bufsize, bufsize, CONTENTS);
1352 
1353 	fd = open(FULLPATH, O_RDWR);
1354 	EXPECT_LE(0, fd) << strerror(errno);
1355 
1356 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
1357 	/*
1358 	 * A subsequent read should be serviced by cache, without querying the
1359 	 * filesystem daemon
1360 	 */
1361 	ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno);
1362 	ASSERT_EQ(bufsize, read(fd, readbuf, bufsize)) << strerror(errno);
1363 	leak(fd);
1364 }
1365 
1366 /* Writes that extend a file should update the cached file size */
1367 TEST_F(Write, update_file_size)
1368 {
1369 	const char FULLPATH[] = "mountpoint/some_file.txt";
1370 	const char RELPATH[] = "some_file.txt";
1371 	const char *CONTENTS = "abcdefgh";
1372 	struct stat sb;
1373 	uint64_t ino = 42;
1374 	int fd;
1375 	ssize_t bufsize = strlen(CONTENTS);
1376 
1377 	expect_lookup(RELPATH, ino, 0);
1378 	expect_open(ino, 0, 1);
1379 	expect_write(ino, 0, bufsize, bufsize, CONTENTS);
1380 
1381 	fd = open(FULLPATH, O_RDWR);
1382 	EXPECT_LE(0, fd) << strerror(errno);
1383 
1384 	ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno);
1385 	/* Get cached attributes */
1386 	ASSERT_EQ(0, fstat(fd, &sb)) << strerror(errno);
1387 	ASSERT_EQ(bufsize, sb.st_size);
1388 	leak(fd);
1389 }
1390