On 3/6/23 16:31, Pádraig Brady wrote:
syntax check now looks good thanks.
You're welcome. I'm getting a failure on tests/du/threshold.sh, though, on Fedora 37 x86-64. Are you seeing that? See attached (compressed) log.
I see this as a more fundamental operational thing than a limit issue. `split -n` needing the size up front alludes to the operation,
To ameliorate a bit, we can document this (done in the attached patch).I don't see this as being significantly more serious than the other utilities that read all their input (possibly from a pipe) before outputting anything. They all need to deal with exhausting temporary space, and 'split' is merely joining the company of 'sort' and 'tac'.
given split is often used with large data, explicit control is often desired over where and how it's persisted.
Anybody needing that control can copy the data themselves to a temp file, before calling 'split'. (Like 'sort' and 'tac'.)
Since the change doesn't invalidate existing uses, it sounds like you're worried that users will want 'split' to work even on enormous pipe inputs, inputs so large that /tmp fills up. I think this unlikely in practice, but if it turns into a real concern I can work around most of the problem by using the first output file as the temporary. However, I'd prefer to avoid doing this unless it's necessary, as it seems overkill.
I installed the attached patch so that 'split' uses a temp file in this situation, so it is no longer limited to the 128 KiB size that it was before the attached patch. Hope this is good enough; if not please let me know.
make-check-log.txt.gz
Description: application/gzip
From bb9dbcbbfd5c3159f975f39336a33319c6c5df04 Mon Sep 17 00:00:00 2001
From: Paul Eggert <egg...@cs.ucla.edu>
Date: Tue, 7 Mar 2023 12:58:12 -0800
Subject: [PATCH] split: support split -n on larger pipe input
* bootstrap.conf (gnulib_modules): Add free-posix, tmpfile.
* src/split.c (copy_to_tmpfile): New function.
(input_file_size): Use it to split larger files when sizes cannot
easily be determined via fstat or lseek. See Bug#61386#235.
* tests/split/l-chunk.sh: Mark tests of /dev/zero as
very expensive since they exhaust /tmp.
---
NEWS | 3 ++
bootstrap.conf | 2 +
doc/coreutils.texi | 6 ++-
src/split.c | 89 +++++++++++++++++++++++++-----------------
tests/split/l-chunk.sh | 15 ++++---
5 files changed, 72 insertions(+), 43 deletions(-)
diff --git a/NEWS b/NEWS
index 5b0dc939c..3df17d3b3 100644
--- a/NEWS
+++ b/NEWS
@@ -145,6 +145,9 @@ GNU coreutils NEWS -*- outline -*-
split now accepts options like '-n SIZE' that exceed machine integer
range, when they can be implemented as if they were infinity.
+ split -n now accepts piped input even when not in round-robin mode,
+ by first copying input to a temporary file to determine its size.
+
wc now accepts the --total={auto,never,always,only} option
to give explicit control over when the total is output.
diff --git a/bootstrap.conf b/bootstrap.conf
index 01430429b..c122354a1 100644
--- a/bootstrap.conf
+++ b/bootstrap.conf
@@ -103,6 +103,7 @@ gnulib_modules="
fnmatch-gnu
fopen-safer
fprintftime
+ free-posix
freopen
freopen-safer
fseeko
@@ -270,6 +271,7 @@ gnulib_modules="
time_rz
timer-time
timespec
+ tmpfile
tzset
uname
unicodeio
diff --git a/doc/coreutils.texi b/doc/coreutils.texi
index f0e46b9ee..7852e9f8a 100644
--- a/doc/coreutils.texi
+++ b/doc/coreutils.texi
@@ -3409,8 +3409,10 @@ are not split even if they overlap a partition, the files written
can be larger or smaller than the partition size, and even empty
if a line/record is so long as to completely overlap the partition.
-For @samp{r} mode, the size of @var{input} is irrelevant,
-and so can be a pipe for example.
+When the input is a pipe or some other special file where the size
+cannot easily be determined, there is no trouble for @samp{r} mode
+because the size of the input is irrelevant. For other modes, such an
+input is first copied to a temporary to determine its size.
@item -a @var{length}
@itemx --suffix-length=@var{length}
diff --git a/src/split.c b/src/split.c
index 95d174a8b..d872ec56a 100644
--- a/src/split.c
+++ b/src/split.c
@@ -275,6 +275,39 @@ CHUNKS may be:\n\
exit (status);
}
+/* Copy the data in FD to a temporary file, then make that file FD.
+ Use BUF, of size BUFSIZE, to copy. Return the number of
+ bytes copied, or -1 (setting errno) on error. */
+static off_t
+copy_to_tmpfile (int fd, char *buf, idx_t bufsize)
+{
+ FILE *tmp = tmpfile ();
+ if (!tmp)
+ return -1;
+ off_t copied = 0;
+ off_t r;
+
+ while (0 < (r = read (fd, buf, bufsize)))
+ {
+ if (fwrite (buf, 1, r, tmp) != r)
+ return -1;
+ if (INT_ADD_WRAPV (copied, r, &copied))
+ {
+ errno = EOVERFLOW;
+ return -1;
+ }
+ }
+
+ if (r < 0)
+ return r;
+ r = dup2 (fileno (tmp), fd);
+ if (r < 0)
+ return r;
+ if (fclose (tmp) < 0)
+ return -1;
+ return copied;
+}
+
/* Return the number of bytes that can be read from FD with status ST.
Store up to the first BUFSIZE bytes of the file's data into BUF,
and advance the file position by the number of bytes read. On
@@ -293,49 +326,35 @@ input_file_size (int fd, struct stat const *st, char *buf, idx_t bufsize)
}
while (size < bufsize);
- off_t cur = lseek (fd, 0, SEEK_CUR);
- if (cur < 0)
- {
- if (errno == ESPIPE)
- errno = 0; /* Suppress confusing seek error. */
- return cur;
- }
-
- off_t end;
- if (usable_st_size (st))
- end = st->st_size;
- else
+ off_t cur, end;
+ if ((usable_st_size (st) && st->st_size < size)
+ || (cur = lseek (fd, 0, SEEK_CUR)) < 0
+ || cur < size /* E.g., /dev/zero on GNU/Linux. */
+ || (end = lseek (fd, 0, SEEK_END)) < 0)
{
- end = lseek (fd, 0, SEEK_END);
+ char *tmpbuf = xmalloc (bufsize);
+ end = copy_to_tmpfile (fd, tmpbuf, bufsize);
+ free (tmpbuf);
if (end < 0)
return end;
- if (end == OFF_T_MAX)
- goto overflow; /* E.g., /dev/zero on GNU/Hurd. */
- if (cur < end)
- {
- off_t cur1 = lseek (fd, cur, SEEK_SET);
- if (cur1 < 0)
- return cur1;
- }
+ cur = 0;
}
- /* Report overflow if we filled the buffer from a file with more
- bytes than stat or lseek reports. This can happen with mutating
- (e.g., /proc) files that are larger than the input block size.
- FIXME: Handle this properly, e.g., by copying the growing file's
- data into the first output file, and then splitting that output
- file (which should not grow) into the other output files. */
- if (end < size)
- goto overflow;
+ if (end == OFF_T_MAX /* E.g., /dev/zero on GNU/Hurd. */
+ || (cur < end && INT_ADD_WRAPV (size, end - cur, &size)))
+ {
+ errno = EOVERFLOW;
+ return -1;
+ }
- if (cur < end && INT_ADD_WRAPV (size, end - cur, &size))
- goto overflow;
+ if (cur < end)
+ {
+ off_t r = lseek (fd, cur, SEEK_SET);
+ if (r < 0)
+ return r;
+ }
return size;
-
- overflow:
- errno = EOVERFLOW;
- return -1;
}
/* Compute the next sequential output file name and store it into the
diff --git a/tests/split/l-chunk.sh b/tests/split/l-chunk.sh
index c94380e87..2c287ee9f 100755
--- a/tests/split/l-chunk.sh
+++ b/tests/split/l-chunk.sh
@@ -37,12 +37,15 @@ rm x??
# 'split' should reject any attempt to create an infinitely
# long output file.
-returns_ 1 split -n l/2 /dev/zero || fail=1
-rm x??
-
-# Repeat the above, but with 1/2, not l/2:
-returns_ 1 split -n 1/2 /dev/zero || fail=1
-rm x??
+# This test is very expensive as it runs out of /tmp space.
+if test "${RUN_VERY_EXPENSIVE_TESTS+set}" = set; then
+ returns_ 1 split -n l/2 /dev/zero || fail=1
+ rm x??
+
+ # Repeat the above, but with 1/2, not l/2:
+ returns_ 1 split -n 1/2 /dev/zero || fail=1
+ rm x??
+fi
# Ensure --elide-empty-files is honored
split -e -n l/10 /dev/null || fail=1
--
2.39.2
