diff --git a/CMakeLists.txt b/CMakeLists.txt
index 7311509a6b3e9a75164c08da5358e766c308c964..cc2352ea100923b28b88557d4e9cded7d40c2428 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -78,6 +78,7 @@ check_function_exists(memmem HAVE_MEMMEM)
check_function_exists(memrchr HAVE_MEMRCHR)
check_function_exists(funopen HAVE_FUNOPEN)
+check_function_exists(fopencookie HAVE_FOPENCOOKIE)
#
# Some versions of GNU libc define non-portable __libc_stack_end
diff --git a/include/fiob.h b/include/fiob.h
index 502b824bc0cc6decb16c61a78905d9b63f312626..9b8de8985d2b8301cf4b173647fbb0734953b158 100644
--- a/include/fiob.h
+++ b/include/fiob.h
@@ -28,8 +28,6 @@
* THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
-#define O_DIRECT_BSIZE ( 4096 * 256 )
-
#include <sys/types.h>
#include <stdio.h>
diff --git a/include/tarantool/config.h.cmake b/include/tarantool/config.h.cmake
index 15a437337e24274097dfb2fc6984603d1e088333..5d8fc06547e35941d8c26b98f27c65c3d3f04710 100644
--- a/include/tarantool/config.h.cmake
+++ b/include/tarantool/config.h.cmake
@@ -67,6 +67,11 @@
*/
#cmakedefine HAVE_FUNOPEN 1
+/*
+ * Defined if this platform has GNU specific fopencookie()
+ */
+#cmakedefine HAVE_FOPENCOOKIE 1
+
/*
* Defined if this platform has GNU specific memmem().
*/
diff --git a/src/fiob.c b/src/fiob.c
index 39779e036164568a24bcec94025e19968afc9b8e..4e8cc794778b4b8609f5d7617195db4ee2806fb5 100644
--- a/src/fiob.c
+++ b/src/fiob.c
@@ -18,6 +18,16 @@
#include <unistd.h>
#include <tarantool/config.h>
+/* Use special implemention if we have O_DIRECT and FOPENCOOKIE or FUNOPEN */
+#if defined(O_DIRECT) && (defined(HAVE_FUNOPEN) || defined(HAVE_FOPENCOOKIE))
+#define FIOB_DIRECT
+#endif
+
+#if defined(FIOB_DIRECT)
+enum {
+ FIOB_ALIGN = 4096,
+ FIOB_BSIZE = FIOB_ALIGN * 256
+};
struct fiob {
int fd;
@@ -37,44 +47,100 @@ struct fiob {
#endif
};
+static inline off_t
+fiob_ceil(off_t off)
+{
+ /* ceil to FIOB_ALIGN */
+ return (off + FIOB_ALIGN - 1) & ~(off_t) (FIOB_ALIGN - 1);
+}
+
+#ifdef HAVE_FUNOPEN
+static int
+fiob_read(void *cookie, char *buf, int len)
+#else
static ssize_t
-fiob_readf(struct fiob *f, char *buf, size_t count)
+fiob_read(void *cookie, char *buf, size_t count)
+#endif
{
+ struct fiob *f = (struct fiob *)cookie;
ssize_t to_read = (ssize_t) count;
+
+ /* The number of starting bytes in f->buf to skip due to alignment */
+ off_t skip = 0;
while (to_read > 0) {
- ssize_t nrd = read(f->fd, buf, to_read);
+ /* Align `to_read' FIOB_ALIGN to be <= size of f->buf */
+ ssize_t to_read_al = MIN(fiob_ceil(to_read), f->bsize);
+ /*
+ * Optimistically try to read aligned size into the aligned
+ * buffer. If the current file position is not aligned then
+ * read(2) returns EINVAL. In this case seek to an aligned
+ * position and try again. This trick saves one extra
+ * syscall for general workflow.
+ */
+ ssize_t nrd = read(f->fd, f->buf, to_read_al);
if (nrd < 0) {
if (errno == EINTR) {
errno = 0;
continue;
+ } else if (errno == EINVAL && skip == 0) {
+ /*
+ * read(2) can return EINVAL only in 3 cases:
+ * 1. read buffer is not aligned - handled in
+ * fiob_open().
+ * 2. read size is not aligned - handled above
+ * 3. current file position is not aligned -
+ * handled here.
+ */
+ off_t pos = lseek(f->fd, 0, SEEK_CUR);
+ if (pos < 0) {
+ say_syserror("lseek, [%s]", f->path);
+ return -1;
+ }
+ /* Calculate aligned position */
+ skip = pos % FIOB_ALIGN;
+ pos -= skip;
+ if (skip == 0) {
+ /* Position is aligned. */
+ errno = EINVAL;
+ say_error("read, [%s]", f->path);
+ return -1;
+ }
+ /* Seek to the new position */
+ if (lseek(f->fd, pos, SEEK_SET) != pos) {
+ say_syserror("lseek, [%s]", f->path);
+ return -1;
+ }
+ /* Try to read again. */
+ continue;
}
- if (errno == EAGAIN || errno == EWOULDBLOCK)
- return count != to_read ? count - to_read : -1;
say_syserror("read, [%s]", f->path);
return -1; /* XXX: file position is unspecified */
}
+ /* Ignore starting bytes if the position was aligned. */
+ nrd -= skip;
if (nrd == 0)
break;
+ if (nrd > to_read) {
+ /*
+ * A few more bytes have been read because `to_read'
+ * is not aligned to FIOB_ALIGN. Set the file position
+ * to the expected libc value and ignore extra bytes.
+ */
+ if (lseek(f->fd, to_read - nrd, SEEK_CUR) < 0) {
+ say_syserror("lseek, [%s]", f->path);
+ return -1;
+ }
+ nrd = to_read;
+ }
+ memcpy(buf, f->buf + skip, nrd); /* see nrd -= skip */
+ skip = 0; /* reset alignment offset */
buf += nrd;
to_read -= nrd;
}
return count - to_read;
}
-#ifdef HAVE_FUNOPEN
-static int
-fiob_read(void *cookie, char *buf, int len)
-#else
-static ssize_t
-fiob_read(void *cookie, char *buf, size_t len)
-#endif
-{
- struct fiob *f = (struct fiob *)cookie;
- return fiob_readf(f, buf, len);
-}
-
-
static ssize_t
fiob_writef(struct fiob *f, const char *buf, size_t count)
{
@@ -87,8 +153,6 @@ fiob_writef(struct fiob *f, const char *buf, size_t count)
errno = 0;
continue;
}
- if (errno == EAGAIN || errno == EWOULDBLOCK)
- return count != to_write ? count - to_write : -1;
say_syserror("write, [%s]", f->path);
return -1; /* XXX: file position is unspecified */
}
@@ -101,17 +165,13 @@ fiob_writef(struct fiob *f, const char *buf, size_t count)
return count - to_write;
}
-
static int
fiob_flushb(struct fiob *f)
{
- if (!f->buf || !f->bfill)
+ if (!f->bfill)
return 0;
- size_t tlen = f->bfill / 4096;
- if (f->bfill % 4096)
- tlen++;
- tlen *= 4096;
+ size_t tlen = fiob_ceil(f->bfill);
if (fiob_writef(f, f->buf, tlen) < 0) {
return -1;
@@ -141,9 +201,6 @@ fiob_write(void *cookie, const char *buf, size_t len)
if (len == 0)
return 0;
- if (!f->buf)
- return fiob_writef(f, buf, len);
-
ssize_t bytes_left = len;
ssize_t tocopy;
@@ -223,23 +280,24 @@ fiob_close(void *cookie)
errno = save_errno;
return res;
}
+#endif /* defined(FIOB_DIRECT) */
-/** open file. The same as fiob_open but receives additional open (2) flags */
+/** open file. The same as fopen but receives additional open (2) flags */
FILE *
fiob_open(const char *path, const char *mode)
{
int omode = 0666;
int flags = 0;
int save_errno;
-
- size_t bsize = 0;
- void *buf = NULL;
-
+ int fd = -1;
+ FILE *file = NULL;
+#if defined (FIOB_DIRECT)
+ struct fiob *f = NULL;
+#endif /* defined(FIOB_DIRECT) */
int um = umask(0722);
umask(um);
omode &= ~um;
-
if (strchr(mode, 'r')) {
if (strchr(mode, '+'))
flags |= O_RDWR;
@@ -267,55 +325,50 @@ fiob_open(const char *path, const char *mode)
if (strchr(mode, 'x'))
flags |= O_EXCL;
#endif
-
- /* O_DIRECT */
- if (strchr(mode, 'd')) {
-#ifdef O_DIRECT
- flags |= O_DIRECT;
-#endif
- bsize = O_DIRECT_BSIZE;
- posix_memalign(&buf, 4096, bsize);
- if (!buf) {
- errno = ENOMEM;
- return NULL;
- }
- /* for valgrind */
- memset(buf, 0, bsize);
- }
-
/* O_SYNC */
if (strchr(mode, 's')) {
flags |= WAL_SYNC_FLAG;
}
- struct fiob *f = (struct fiob *)calloc(1, sizeof(struct fiob));
- if (!f) {
- free(buf);
- errno = ENOMEM;
- return NULL;
+#if defined(FIOB_DIRECT)
+ if (strchr(mode, 'd')) {
+ /* Try to open file with O_DIRECT */
+ fd = open(path, flags | O_DIRECT, omode);
}
-
- f->path = strdup(path);
- if (!f->path) {
- errno = ENOMEM;
- goto error;
+ if (fd < 0) {
+#endif /* defined(FIOB_DIRECT) */
+ /* Fallback to libc implementation */
+ fd = open(path, flags, omode);
+ if (fd < 0)
+ goto error;
+ file = fdopen(fd, mode);
+ if (!file)
+ goto error;
+ return file;
+#if defined(FIOB_DIRECT)
}
- f->buf = buf;
- f->bsize = bsize;
+ f = (struct fiob *)calloc(1, sizeof(struct fiob));
+ if (!f)
+ goto error;
- f->fd = open(path, flags, omode);
- if (f->fd < 0)
+ f->fd = fd;
+ f->bsize = FIOB_BSIZE;
+ if (posix_memalign(&f->buf, FIOB_ALIGN, f->bsize))
goto error;
+ /* for valgrind */
+ memset(f->buf, 0, f->bsize);
+ f->path = strdup(path);
+ if (!f->path)
+ goto error;
f->io.read = fiob_read;
f->io.write = fiob_write;
f->io.seek = fiob_seek;
f->io.close = fiob_close;
- FILE *file;
#ifdef HAVE_FUNOPEN
file = funopen(f,
f->io.read, f->io.write, f->io.seek, f->io.close);
@@ -333,16 +386,21 @@ fiob_open(const char *path, const char *mode)
#endif
return file;
+#endif /* defined(FIOB_DIRECT) */
error:
save_errno = errno;
say_syserror("Can't open '%s'", path);
- if (f->fd > 0)
- close(f->fd);
-
- free(f->buf);
- free(f->path);
- free(f);
+ if (fd >= 0)
+ close(fd);
+
+#if defined(FIOB_DIRECT)
+ if (f) {
+ free(f->buf);
+ free(f->path);
+ free(f);
+ }
+#endif /* FIOB_DIRECT */
errno = save_errno;
return NULL;
diff --git a/test/unit/fiob.c b/test/unit/fiob.c
index a26570f61f2150a0b57e1b4dfe16558e99057123..dc61f07ee2109b881378a9e928c5037d5514a936 100644
--- a/test/unit/fiob.c
+++ b/test/unit/fiob.c
@@ -22,7 +22,7 @@
-#define PLAN 47
+#define PLAN 65
#define ITEMS 7
@@ -98,6 +98,29 @@ main(void)
is(done, 12, "Hello world is read (%zu bytes)", done);
is(memcmp(buf, "Hello, world", 12), 0, "data");
+ is(fseek(f, 7L, SEEK_SET), 0, "set odd position");
+ is(ftell(f), 7L, "check odd position");
+ is(fread(buf, 1, 4096, f), 5, "read from odd position (size)");
+ is(memcmp(buf, "world", 5), 0, "read from odd position (data)");
+
+ is(fseek(f, 0L, SEEK_SET), 0, "set start position");
+ is(ftell(f), 0L, "check start position");
+
+ is(fseek(f, 0L, SEEK_END), 0, "set eof position");
+ is(ftell(f), 12L, "check eof position");
+ is(fread(buf, 1, 4096, f), 0, "read from eof position (size)");
+ ok(feof(f), "feof");
+ is(fread(buf, 1, 4096, f), 0, "read from eof position (size)");
+ ok(feof(f), "feof");
+
+ is(fseek(f, -1L, SEEK_END), 0, "set -1 position");
+ is(ftell(f), 11L, "check -1 position");
+ is(fread(buf, 1, 4096, f), 1, "read from -1 position (size)");
+ is(memcmp(buf, "d", 1), 0, "read from -1 position (data)");
+ ok(feof(f), "feof");
+
+ is(fread(buf, 1, 4096, f), 0, "read from eof position (size)");
+
is(fseek(f, 0L, SEEK_SET), 0, "set new position");
done = fread(buf + 1, 1, 12, f);
is(done, 12, "Hello world is read (%zu bytes)", done);
diff --git a/test/unit/fiob.result b/test/unit/fiob.result
index 2cedf473e2d4c16595a053e2d2d5699b28e9e91e..fe92099582ff021b02aed06e2bbec22e318ddc33 100644
--- a/test/unit/fiob.result
+++ b/test/unit/fiob.result
@@ -1,4 +1,4 @@
-1..47
+1..65
ok 1 - tempdir is created
ok 2 - common open
ok 3 - Hello world is written (12 bytes)
@@ -7,42 +7,60 @@ ok 5 - set new position
ok 6 - current position 0
ok 7 - Hello world is read (12 bytes)
ok 8 - data
-ok 9 - set new position
-ok 10 - Hello world is read (12 bytes)
-ok 11 - data
-ok 12 - set new position
-ok 13 - set new position
-ok 14 - data is read
-ok 15 - data is read
-ok 16 - fclose
-ok 17 - reopened file
-ok 18 - move pos at finish
-ok 19 - file size
-ok 20 - fclose
-ok 21 - common open: O_EXCL
-ok 22 - common open
-ok 23 - Hello world is written (12 bytes)
-ok 24 - move pos
-ok 25 - Hello world is written (12 bytes)
-ok 26 - move pos
-ok 27 - read 11 bytes
-ok 28 - content was read
-ok 29 - fclose
-ok 30 - open big file
-ok 31 - Hello world is written (13 bytes)
-ok 32 - all bytes were written
-ok 33 - fclose
-ok 34 - reopen file for reading
-ok 35 - all records were written properly
-ok 36 - eof
-ok 37 - feof
+ok 9 - set odd position
+ok 10 - check odd position
+ok 11 - read from odd position (size)
+ok 12 - read from odd position (data)
+ok 13 - set start position
+ok 14 - check start position
+ok 15 - set eof position
+ok 16 - check eof position
+ok 17 - read from eof position (size)
+ok 18 - feof
+ok 19 - read from eof position (size)
+ok 20 - feof
+ok 21 - set -1 position
+ok 22 - check -1 position
+ok 23 - read from -1 position (size)
+ok 24 - read from -1 position (data)
+ok 25 - feof
+ok 26 - read from eof position (size)
+ok 27 - set new position
+ok 28 - Hello world is read (12 bytes)
+ok 29 - data
+ok 30 - set new position
+ok 31 - set new position
+ok 32 - data is read
+ok 33 - data is read
+ok 34 - fclose
+ok 35 - reopened file
+ok 36 - move pos at finish
+ok 37 - file size
ok 38 - fclose
-ok 39 - open big file
-ok 40 - Hello world is written (13 bytes)
-ok 41 - all bytes were written
-ok 42 - fclose
-ok 43 - reopen file for reading
-ok 44 - all records were written properly
-ok 45 - eof
-ok 46 - feof
+ok 39 - common open: O_EXCL
+ok 40 - common open
+ok 41 - Hello world is written (12 bytes)
+ok 42 - move pos
+ok 43 - Hello world is written (12 bytes)
+ok 44 - move pos
+ok 45 - read 11 bytes
+ok 46 - content was read
ok 47 - fclose
+ok 48 - open big file
+ok 49 - Hello world is written (13 bytes)
+ok 50 - all bytes were written
+ok 51 - fclose
+ok 52 - reopen file for reading
+ok 53 - all records were written properly
+ok 54 - eof
+ok 55 - feof
+ok 56 - fclose
+ok 57 - open big file
+ok 58 - Hello world is written (13 bytes)
+ok 59 - all bytes were written
+ok 60 - fclose
+ok 61 - reopen file for reading
+ok 62 - all records were written properly
+ok 63 - eof
+ok 64 - feof
+ok 65 - fclose