posix_fallocate: Emulation fixes and documentation [BZ #15661]

Handle signed integer overflow correctly.  Detect and reject O_APPEND.
Document drawbacks of emulation.

This does not completely address bug 15661, but improves the situation
somewhat.

ChangeLog

@@ -1,3 +1,12 @@
+2015-06-05  Florian Weimer  <fweimer@redhat.com>
+
+	[BZ #15661]
+	* sysdeps/posix/posix_fallocate64.c (__posix_fallocate64_l64):
+	Check for overflow properly.  Check for O_APPEND.  Ignore large
+	file system block sizes.  Add comments about problems.
+	* sysdeps/posix/posix_fallocate.c (posix_fallocate): Likewise.
+	* manual/filesys.texi (Storage Allocation): New node.
+
 2015-06-04  Adhemerval Zanella  <adhemerval.zanella@linaro.org>
 
 	* sysdeps/unix/sysdep.h [SYSCALL_CANCEL]: New macro: define

manual/filesys.texi

@@ -1723,6 +1723,7 @@ modify the attributes of a file.
                                  access a file.
 * File Times::                  About the time attributes of a file.
 * File Size::			Manually changing the size of a file.
+* Storage Allocation::          Allocate backing storage for files.
 @end menu
 
 @node Attribute Meanings
@@ -3233,6 +3234,99 @@ is a requirement of @code{mmap}.  The program has to keep track of the
 real size, and when it has finished a final @code{ftruncate} call should
 set the real size of the file.
 
+@node Storage Allocation
+@subsection Storage Allocation
+@cindex allocating file storage
+@cindex file allocation
+@cindex storage allocating
+
+@cindex file fragmentation
+@cindex fragmentation of files
+@cindex sparse files
+@cindex files, sparse
+Most file systems support allocating large files in a non-contiguous
+fashion: the file is split into @emph{fragments} which are allocated
+sequentially, but the fragments themselves can be scattered across the
+disk.  File systems generally try to avoid such fragmentation because it
+decreases performance, but if a file gradually increases in size, there
+might be no other option than to fragment it.  In addition, many file
+systems support @emph{sparse files} with @emph{holes}: regions of null
+bytes for which no backing storage has been allocated by the file
+system.  When the holes are finally overwritten with data, fragmentation
+can occur as well.
+
+Explicit allocation of storage for yet-unwritten parts of the file can
+help the system to avoid fragmentation.  Additionally, if storage
+pre-allocation fails, it is possible to report the out-of-disk error
+early, often without filling up the entire disk.  However, due to
+deduplication, copy-on-write semantics, and file compression, such
+pre-allocation may not reliably prevent the out-of-disk-space error from
+occurring later.  Checking for write errors is still required, and
+writes to memory-mapped regions created with @code{mmap} can still
+result in @code{SIGBUS}.
+
+@deftypefun int posix_fallocate (int @var{fd}, off_t @var{offset}, off_t @var{length})
+@safety{@prelim{}@mtsafe{}@assafe{}@acsafe{}}
+@c If the file system does not support allocation,
+@c @code{posix_fallocate} has a race with file extension (if
+@c @var{length} is zero) or with concurrent writes of non-NUL bytes (if
+@c @var{length} is positive).
+
+Allocate backing store for the region of @var{length} bytes starting at
+byte @var{offset} in the file for the descriptor @var{fd}.  The file
+length is increased to @samp{@var{length} + @var{offset}} if necessary.
+
+@var{fd} must be a regular file opened for writing, or @code{EBADF} is
+returned.  If there is insufficient disk space to fulfill the allocation
+request, @code{ENOSPC} is returned.
+
+@strong{Note:} If @code{fallocate} is not available (because the file
+system does not support it), @code{posix_fallocate} is emulated, which
+has the following drawbacks:
+
+@itemize @bullet
+@item
+It is very inefficient because all file system blocks in the requested
+range need to be examined (even if they have been allocated before) and
+potentially rewritten.  In contrast, with proper @code{fallocate}
+support (see below), the file system can examine the internal file
+allocation data structures and eliminate holes directly, maybe even
+using unwritten extents (which are pre-allocated but uninitialized on
+disk).
+
+@item
+There is a race condition if another thread or process modifies the
+underlying file in the to-be-allocated area.  Non-null bytes could be
+overwritten with null bytes.
+
+@item
+If @var{fd} has been opened with the @code{O_APPEND} flag, the function
+will fail with an @code{errno} value of @code{EBADF}.
+
+@item
+If @var{length} is zero, @code{ftruncate} is used to increase the file
+size as requested, without allocating file system blocks.  There is a
+race condition which means that @code{ftruncate} can accidentally
+truncate the file if it has been extended concurrently.
+@end itemize
+
+On Linux, if an application does not benefit from emulation or if the
+emulation is harmful due to its inherent race conditions, the
+application can use the Linux-specific @code{fallocate} function, with a
+zero flag argument.  For the @code{fallocate} function, @theglibc{} does
+not perform allocation emulation if the file system does not support
+allocation.  Instead, an @code{EOPNOTSUPP} is returned to the caller.
+
+@end deftypefun
+
+@deftypefun int posix_fallocate64 (int @var{fd}, off64_t @var{length}, off64_t @var{offset})
+@safety{@prelim{}@mtsafe{}@assafe{}@acsafe{}}
+
+This function is a variant of @code{posix_fallocate64} which accepts
+64-bit file offsets on all platforms.
+
+@end deftypefun
+
 @node Making Special Files
 @section Making Special Files
 @cindex creating special files

sysdeps/posix/posix_fallocate.c

@@ -18,26 +18,36 @@
 #include <errno.h>
 #include <fcntl.h>
 #include <unistd.h>
+#include <stdint.h>
+#include <sys/fcntl.h>
 #include <sys/stat.h>
 #include <sys/statfs.h>
 
-/* Reserve storage for the data of the file associated with FD.  */
+/* Reserve storage for the data of the file associated with FD.  This
+   emulation is far from perfect, but the kernel cannot do not much
+   better for network file systems, either.  */
 
 int
 posix_fallocate (int fd, __off_t offset, __off_t len)
 {
   struct stat64 st;
-  struct statfs f;
 
-  /* `off_t' is a signed type.  Therefore we can determine whether
-     OFFSET + LEN is too large if it is a negative value.  */
   if (offset < 0 || len < 0)
     return EINVAL;
-  if (offset + len < 0)
+
+  /* Perform overflow check.  The outer cast relies on a GCC
+     extension.  */
+  if ((__off_t) ((uint64_t) offset) + ((uint64_t) len) < 0)
     return EFBIG;
 
-  /* First thing we have to make sure is that this is really a regular
+  /* pwrite below will not do the right thing in O_APPEND mode.  */
-     file.  */
+  {
+    int flags = __fcntl (fd, F_GETFL, 0);
+    if (flags < 0 || (flags & O_APPEND) != 0)
+      return EBADF;
+  }
+
+  /* We have to make sure that this is really a regular file.  */
   if (__fxstat64 (_STAT_VER, fd, &st) != 0)
     return EBADF;
   if (S_ISFIFO (st.st_mode))
@@ -47,6 +57,8 @@ posix_fallocate (int fd, __off_t offset, __off_t len)
 
   if (len == 0)
     {
+      /* This is racy, but there is no good way to satisfy a
+	 zero-length allocation request.  */
       if (st.st_size < offset)
 	{
 	  int ret = __ftruncate (fd, offset);
@@ -58,19 +70,36 @@ posix_fallocate (int fd, __off_t offset, __off_t len)
       return 0;
     }
 
-  /* We have to know the block size of the filesystem to get at least some
+  /* Minimize data transfer for network file systems, by issuing
-     sort of performance.  */
+     single-byte write requests spaced by the file system block size.
-  if (__fstatfs (fd, &f) != 0)
+     (Most local file systems have fallocate support, so this fallback
-    return errno;
+     code is not used there.)  */
 
-  /* Try to play safe.  */
+  unsigned increment;
-  if (f.f_bsize == 0)
+  {
-    f.f_bsize = 512;
+    struct statfs64 f;
 
-  /* Write something to every block.  */
+    if (__fstatfs64 (fd, &f) != 0)
-  for (offset += (len - 1) % f.f_bsize; len > 0; offset += f.f_bsize)
+      return errno;
+    if (f.f_bsize == 0)
+      increment = 512;
+    else if (f.f_bsize < 4096)
+      increment = f.f_bsize;
+    else
+      /* NFS does not propagate the block size of the underlying
+	 storage and may report a much larger value which would still
+	 leave holes after the loop below, so we cap the increment at
+	 4096.  */
+      increment = 4096;
+  }
+
+  /* Write a null byte to every block.  This is racy; we currently
+     lack a better option.  Compare-and-swap against a file mapping
+     might additional local races, but requires interposition of a
+     signal handler to catch SIGBUS.  */
+  for (offset += (len - 1) % increment; len > 0; offset += increment)
     {
-      len -= f.f_bsize;
+      len -= increment;
 
       if (offset < st.st_size)
 	{

sysdeps/posix/posix_fallocate64.c

@@ -18,26 +18,36 @@
 #include <errno.h>
 #include <fcntl.h>
 #include <unistd.h>
+#include <stdint.h>
+#include <sys/fcntl.h>
 #include <sys/stat.h>
 #include <sys/statfs.h>
 
-/* Reserve storage for the data of the file associated with FD.  */
+/* Reserve storage for the data of the file associated with FD.  This
+   emulation is far from perfect, but the kernel cannot do not much
+   better for network file systems, either.  */
 
 int
 __posix_fallocate64_l64 (int fd, __off64_t offset, __off64_t len)
 {
   struct stat64 st;
-  struct statfs64 f;
 
-  /* `off64_t' is a signed type.  Therefore we can determine whether
-     OFFSET + LEN is too large if it is a negative value.  */
   if (offset < 0 || len < 0)
     return EINVAL;
-  if (offset + len < 0)
+
+  /* Perform overflow check.  The outer cast relies on a GCC
+     extension.  */
+  if ((__off64_t) ((uint64_t) offset) + ((uint64_t) len) < 0)
     return EFBIG;
 
-  /* First thing we have to make sure is that this is really a regular
+  /* pwrite64 below will not do the right thing in O_APPEND mode.  */
-     file.  */
+  {
+    int flags = __fcntl (fd, F_GETFL, 0);
+    if (flags < 0 || (flags & O_APPEND) != 0)
+      return EBADF;
+  }
+
+  /* We have to make sure that this is really a regular file.  */
   if (__fxstat64 (_STAT_VER, fd, &st) != 0)
     return EBADF;
   if (S_ISFIFO (st.st_mode))
@@ -47,6 +57,8 @@ __posix_fallocate64_l64 (int fd, __off64_t offset, __off64_t len)
 
   if (len == 0)
     {
+      /* This is racy, but there is no good way to satisfy a
+	 zero-length allocation request.  */
       if (st.st_size < offset)
 	{
 	  int ret = __ftruncate64 (fd, offset);
@@ -58,19 +70,36 @@ __posix_fallocate64_l64 (int fd, __off64_t offset, __off64_t len)
       return 0;
     }
 
-  /* We have to know the block size of the filesystem to get at least some
+  /* Minimize data transfer for network file systems, by issuing
-     sort of performance.  */
+     single-byte write requests spaced by the file system block size.
-  if (__fstatfs64 (fd, &f) != 0)
+     (Most local file systems have fallocate support, so this fallback
-    return errno;
+     code is not used there.)  */
 
-  /* Try to play safe.  */
+  unsigned increment;
-  if (f.f_bsize == 0)
+  {
-    f.f_bsize = 512;
+    struct statfs64 f;
 
-  /* Write something to every block.  */
+    if (__fstatfs64 (fd, &f) != 0)
-  for (offset += (len - 1) % f.f_bsize; len > 0; offset += f.f_bsize)
+      return errno;
+    if (f.f_bsize == 0)
+      increment = 512;
+    else if (f.f_bsize < 4096)
+      increment = f.f_bsize;
+    else
+      /* NFS clients do not propagate the block size of the underlying
+	 storage and may report a much larger value which would still
+	 leave holes after the loop below, so we cap the increment at
+	 4096.  */
+      increment = 4096;
+  }
+
+  /* Write a null byte to every block.  This is racy; we currently
+     lack a better option.  Compare-and-swap against a file mapping
+     might address local races, but requires interposition of a signal
+     handler to catch SIGBUS.  */
+  for (offset += (len - 1) % increment; len > 0; offset += increment)
     {
-      len -= f.f_bsize;
+      len -= increment;
 
       if (offset < st.st_size)
 	{