|
| From: | Pádraig Brady |
| Subject: | Re: RFC: posix_fadvise( "to end of file" ) semantics |
| Date: | Fri, 31 Oct 2025 15:56:54 +0000 |
| User-agent: | Mozilla Thunderbird Beta |
On 30/10/2025 02:18, Collin Funk wrote:
Pádraig Brady <P@draigBrady.com> writes:I was debugging a change in coreutils 9.8 and noticed that fadvise was being passed OFF_T_MAX to indicate to apply advice to end of file. This struck me as a bit unusual and coreutils <= 9.7 used 0 to do the same thing. It seems to me like 0 may be treated specially or more efficiently in more cases? How about the following to go back to 0 meaning, to end of file? cheers, Padraig diff --git a/src/copy-file-data.c b/src/copy-file-data.c index 1eefd3071..c3abe41e2 100644 --- a/src/copy-file-data.c +++ b/src/copy-file-data.c @@ -538,7 +538,7 @@ copy_file_data (int ifd, struct stat const *ist, off_t ipos, char const *iname, /* Don't bother calling fadvise for small copies, as it is not likely to help performance and might even hurt it. */ if (IO_BUFSIZE < ibytes) - fdadvise (ifd, ipos, ibytes <= OFF_T_MAX - ipos ? ibytes : 0, + fdadvise (ifd, ipos, ibytes < OFF_T_MAX - ipos ? ibytes : 0, FADVISE_SEQUENTIAL); /* If not making a sparse file, try to use a more-efficientYour change seems to match the expectations of POSIX [1]: If len is zero, all data from offset to the largest possible value of the file offset for that file shall be specified. And the Linux man pages [2]: or until the end of the file if size is 0 Whether or not it causes issues, I am not sure. Collin [1] https://pubs.opengroup.org/onlinepubs/9799919799/functions/posix_fadvise.html [2] https://www.man7.org/linux/man-pages/man2/posix_fadvise.2.html
This patch is in fact required with OpenZFS as my hunch was correct that a zero length would be treated specially (even perhaps because cp has used 0 for so long). With a specific length OpenZFS 2.2.2 synchronously (decompresses and) populates the cache, which defeats the subsequent copy offload. Details at: https://github.com/coreutils/coreutils/issues/122 I'll apply the above, but also the attached more defensive patch, to avoid fdadvise unless we really need it. cheers, Padraig
cp-defensive-fadvise.patch
Description: Text Data
| [Prev in Thread] | Current Thread | [Next in Thread] |