cache: sendfile(2) usage results in short writes to pipes

Leah Neukirchen leah at
Sun Dec 18 18:38:56 UTC 2022


in cache.c:print_slot, sendfile is called only once, but not checked
for short writes.  These happen for example when serving cgit via
fcgiwrap, which uses a pipe, so the write only writes 65k:

[pid 31211] openat(AT_FDCWD, "/var/cache/cgit/b9200000", O_RDONLY) = 4
[pid 31211] fstat(4, {st_mode=S_IFREG|0600, st_size=19824215, ...}) = 0
[pid 31211] read(4, "mirror/glibc/snapshot/glibc-3aae"..., 4096) = 4096
[pid 31211] sendfile(1, 4, [77] => [65536], 19824138) = 65459
[pid 31211] close(4 <unfinished ...>

This results in truncated cache responses, and therefore broken
downloads from snapshots (larger than 65k).

I assume this is also the cause for the bug reported in
<0ef89b1e-6543-e726-2cf0-b68e7167fabe at>, I added the author to Cc:.

I recommend running sendfile in a loop until the whole file is sent.
I also recommend detecting sendfile errors and falling back to the
default read/write implementation, as there are some Linux filesystems
where sendfile fails permanently.

Disabling HAVE_LINUX_SENDFILE on build serves as a workaround.

Leah Neukirchen  <leah at>

More information about the CGit mailing list