* src/lisp.h: Include <sys-limits.h>.
(MAX_RW_COUNT): Remove, as it is the same thing as sys-limits.h’s
SYS_BUFSIZE_MAX. All uses replaced by SYS_BUFSIZE_MAX.
* src/fileio.c (Finsert_file_contents): Don’t give up finding the
file’s end merely because SEEK_END fails. Try a small test read
anyway. This might work in non-POSIX file systems where SEEK_END
is not allowed, and trying is easy and simplifies the code.
This fixes some longstanding bugs, along with bugs recently
introduced as part of fixes for Bug#77315.
* src/fileio.c (Finsert_file_contents): Arrange for END_OFFSET
to never be less than BEG_OFFSET, to simplify the code.
Do not go past END_OFFSET when sampling or optimizing reading.
Avoid a few unnecessary lseeks.
This is related to recent fixes for Bug#77315.
* src/fileio.c (Finsert_file_contents): Defend against
a file shrinking to be smaller than its initial offset.
This can happen only on platforms like Solaris where
the initial offset can be positive.
This increases the max read size from 0.25 to 1 MiB, which is a
bit better for Emacs in my measurements. The old value was taken
from coreutils, but was for a different purpose.
* src/fileio.c (INSERT_READ_SIZE_MAX): New constant, replacing
IO_BUFSIZE. All uses changed.
(Finsert_file_contents): Do not use a read buffer that is larger
than INSERT_READ_SIZE_MAX.
This partially reverts 55f41ca3aa,
which was a partial fix for Bug#77315.
* src/fileio.c (Finsert_file_contents): Go back to accepting
partial reads when doing the optimized way when code conversion is
not needed. This avoids an unlikely bug when END happens to lie
at a buffer boundary, causing an infloop. For consistency, do the
same when doing code conversion (though there’s no bug in then),
as there’s not a real need for emacs_full_read there either.
* src/fileio.c (Finsert_file_contents): Be more conservative
when reading from seekable non-regular files when END is given.
The code assumed that in this case, reads would succeed or fail
quickly so there was no need to call internal_condition_case_1.
However, in some cases, e.g., /dev/random, it could take a while.
This fixes a glitch introduced in commit
cb4579ed6b dated 2022-06-11.
This fixes a potential (albeit quite unlikely) bug
introduced during recent fixes to Bug#77315.
* src/fileio.c (union read_non_regular.s.inserted):
Bring back this member.
(read_non_regular, Finsert_file_contents): Work in the unlikely
case when internal_condition_case_1 relocates the buffer before
calling read_non_regular.
This improves on recent fixes to Bug#77315.
When sampling the first 1 KiB and last 3 KiB, do not seek before
BEG if given. Instead, sample starting at BEG, to be consistent
with the non-optimized version.
* src/fileio.c (xlseek): Return POS, for convenience.
(Finsert_file_contents): Sample the first 1 KiB correctly when BEG.
In a CURPOS local, keep track of the input file offset, or for
nonseekable files the number of bytes read, while this value is
important. This lets us avoid some unnecessary seeks. Report an
error earlier if the file is not seekable and BEG is nonzero,
to save work and simplify the code. When sampling, discard less
data, as this is simpler and there’s little point to discarding it.
* src/fileio.c (xlseek): New function.
(Finsert_file_contents): Use it instead of doing things by hand.
(write_region): Match diagnostic wording to xlseek’s.
This improves on recent fixes to Bug#77315.
* src/fileio.c (Finsert_file_contents): When optimizing
do not seek to 0 if the initial file offset is positive,
as that causes the optimized code to behave differently
from the nonoptimized.
This improves on recent fixes to Bug#77315.
* src/fileio.c (Finsert_file_contents):
Do not give up on seeking merely because SEEK_END fails.
On some files (/proc/cpuinfo, say) SEEK_END can fail
even though SEEK_SET and SEEK_CUR succeed, and the
code still works in this case.
* src/fileio.c (Finsert_file_contents): Store st_size,
not file_size_hint, into current_buffer->modtime_size,
since modtime_size is compared to a later st_size, not
to the actual file size. Simplify buffer overflow detection
from file size, since this occurred only on regular files.
Problem reported by Zhengyi Fu (Bug#79031).
* src/fileio.c (Finsert_file_contents): Do not read more than
requested, even when the gap is so small that we use a local buffer.
In commit 61a8ce0280 I mistakenly
focused on overlap in the buffer. The code also needs to check
for overlap in the inserted file data.
Problem reported by Gerd Möllmann (Bug#79020).
* src/fileio.c (Finsert_file_contents): Prevent overlap in both
the buffer head and tail, and in the inserted file’s head and tail.
Also, shrink the file’s head if the file shrank to be smaller
than its head.
This change makes insert-file-contents-literally 30% faster on my
platform, when inserting xdisp.c.
* src/fileio.c (READ_BUF_SIZE): Remove, replacing with ...
(IO_BUFSIZE): ... this new constant from Coreutils.
All uses of READ_BUF_SIZE changed to either MAX_ALLOCA or
IO_BUFSIZE.
* src/fileio.c (union read_non_regular):
New members buf and bufsize replace inserted and trytry.
(read_non_regular): Adjust to this new, simpler interface.
(Finsert_file_contents): If the gap is smaller than read_buf
and we want to read more than the gap, read into read_buf
first, to avoid lots of tiny reads from /proc files.
* src/fileio.c (emacs_full_read): New function.
(Fcopy_file, Finsert_file_contents): Use it.
(Finsert_file_contents): Check for partial reads when scanning
backwards through the file’s tail, as this indicates the file
shrank while we read it. Also, use emacs_full_read in other
situations where it’s simpler and should be a bit faster.
* src/fileio.c (Finsert_file_contents): Use ptrdiff_t, not int, to
store results returned by emacs_fd_read, which return ptrdiff_t.
This doesn’t fix any bugs since the values happen to fit in int,
but is a bit clearer and may simplify future changes.
* src/fileio.c (Finsert_file_contents): Use del_range_byte,
not del_range_1, since we are passing byte counts, not
character counts. Problem reported by Pip Cet (Bug#77315).
* src/fileio.c (Finsert_file_contents): Overlap is impossible, so
omit incorrect overflow adjustments that caused later index
calculations to go off the rails (Bug#77315). Change a ‘ZV_BYTE’
to ‘same_at_end’ for clarity; they have the same value here.
This fix was prompted by Bug#77315.
* src/fileio.c (Finsert_file_contents): Do not rely on st_size for
anything other than a hint about the file size. Trust only the
file size as revealed by a read that returns 0.
* src/fileio.c (Finsert_file_contents):
When counting bytes at file end, don’t trust lseek to seek
to the end of the file. Instead, read a bit after lseeking,
to make sure we get to the end, give up if we don’t get to EOF.
This works around problems on /proc filesystems
where lseek pretends the file is empty.
Update file_size_hint when reaching EOF.
* src/fileio.c (Finsert_file_contents): Change ‘total’ from
ptrdiff_t to off_t since it might not fit in ptrdiff_t.
Check for overflow when estimating the insertion size.
* src/fileio.c (Finsert_file_contents): Remove check for overflow.
The overflow is not possible, and the check was incorrectly
written and would not have caught it anyway.
* src/fileio.c (Finsert_file_contents): Don’t assume that a
newly-opened file is at offset 0, even if it is a regular file.
This might not be true for files like /dev/stdin.
* src/fileio.c (Finsert_file_contents): Check for negative file
sizes earlier, as it’s easy, avoids some unnecessary work, and
will simplify later changes.
* src/fileio.c (Finsert_file_contents): When inserting a file into
a nonempty buffer, improve the heuristic for determining the
file’s coding system by not trusting lseek+SEEK_END, which is
unreliable in /proc or when the file is mutating.
This is part of a fix for Bug#77315,
and improves on the fix for Bug#71258.
* src/fileio.c (Finsert_file_contents): Do not pretend that
directories are regular files. Instead, signal an error when
attempting to read from them in the usual case where the OS
prohibits that; and otherwise read from them. However, when
visiting a directory report an error right away rather than
waiting until later, as this function is documented to not allow
visiting non-regular files. Nest the struct stat into a small
code block, to keep the code simpler and so that the compiler can
see what parts are used and can issue better diagnostics if
uninitialized storage is accessed. Be more skeptical of st_size,
when the file is not regular.
* src/fileio.c (Finsert_file_contents): Do not look at st_size
when scanning the file’s head and tail for a coding system comment.
Instead, just use read and SEEK_END and don’t worry if the latter fails.
st_size and lseek might not work in a /proc file system.
In copy-file, do not trust st_size, since it might change as we run,
or we might be in a /proc system where it is unreliable anyway.
Also, fix some other unlikely copy-file bugs while we’re here.
* src/fileio.c (Fcopy_file): Use O_TRUNC when opening a
destination that already exists. This saves us the trouble
of having to call ftruncate with a possibly-bogus st_size;
the old motivation for using ftruncate is no longer compelling.
Do not assume ptrdiff_t is as wide as ssize_t; although this is
true on all known platforms, it’s easy to not assume it.
Don’t trust st_size. Prefer SSIZE_MAX to TYPE_MAXIMUM (ssize_t).
Always read+write, regardless of whether copy_file_range failed.
* src/fileio.c (Finsert_file_contents): Minor refactoring.
This should help future patches which distinguish read_buf’s size
from other uses of READ_BUF_SIZE which may change.
* src/fileio.c (read_non_regular): Return negation of
errno on failure, instead of -1.
(Finsert_file_contents): Signal with correct errno when a read fails.
* src/fileio.c (Finsert_file_contents): Inhibit ask-supersession
only if we're VISITing in a non-narrowed buffer (bug#78866).
* test/src/fileio-tests.el (ert--tests-dir): New var.
(fileio-tests--insert-file-contents-supersession): New test.
