On 2018/10/29 下午5:43, fdman...@kernel.org wrote: > From: Filipe Manana <fdman...@suse.com> > > Unless the '-s' option is passed to fssum, it should not detect file holes > and have their existence influence the computed checksum for a file. This > tool was added to test btrfs' send/receive feature, so that it checks for > any metadata and data differences between the original filesystem and the > filesystem that receives send streams. > > For a long time the test btrfs/007, which tests btrfs' send/receive with > fsstress, fails sporadically reporting data differences between files. > However the md5sum/sha1sum from the reported files in the original and > new filesystems are the same. The reason why fssum fails is because even > in normal mode it still accounts for number of holes that exist in the > file and their respective lengths. This is done using the SEEK_DATA mode > of lseek. The btrfs send feature does not preserve holes nor prealloc > extents (not supported by the current protocol), so whenever a hole or > prealloc (unwritten) extent is detected in the source filesystem, it > issues a write command full of zeroes, which will translate to a regular > (written) extent in the destination filesystem. This is why fssum reports > a different checksum. A prealloc extent also counts as hole when using > lseek. > > For example when passing a seed of 1540592967 to fsstress in btrfs/007, > the test fails, as file p0/d0/f7 has a prealloc extent in the original > filesystem (in the incr snapshot). > > Fix this by making fssum just read the hole file and feed its data to the > digest calculation function when option '-s' is not given. If we ever get > btrfs' send/receive to support holes and fallocate, we can just change > the test and pass the '-s' option to all fssum calls.
However this is causing more problem here.
Now since fssum doesn't skip holes, for a super large sparse file, fssum
will try to fill all the holes with zero and spend tons of CPU time
calculating the result.
E.g. when seed = 1550703281.
We will have a PiB level file:
$ ls /mnt/scratch/base/p0/f3 -alh
-rw-rw-rw- 1 1441 2774 666P Feb 13 13:39 /mnt/scratch/base/p0/f3
(well, 666, no wonder this will fail)
But it only takes aroud 1M:
$ du /mnt/scratch/base/p0/f3
1044 /mnt/scratch/base/p0/f3
This is even more annoying than test failure.
Can't we just filter out the fallocate/hole punching part in the test
itself?
Thanks,
Qu
>
> Signed-off-by: Filipe Manana <fdman...@suse.com>
> ---
> src/fssum.c | 65
> +++++--------------------------------------------------------
> 1 file changed, 5 insertions(+), 60 deletions(-)
>
> diff --git a/src/fssum.c b/src/fssum.c
> index 5da39abf..f1da72fb 100644
> --- a/src/fssum.c
> +++ b/src/fssum.c
> @@ -224,71 +224,16 @@ int
> sum_file_data_permissive(int fd, sum_t *dst)
> {
> int ret;
> - off_t pos;
> - off_t old;
> - int i;
> - uint64_t zeros = 0;
> -
> - pos = lseek(fd, 0, SEEK_CUR);
> - if (pos == (off_t)-1)
> - return errno == ENXIO ? 0 : -2;
>
> while (1) {
> - old = pos;
> - pos = lseek(fd, pos, SEEK_DATA);
> - if (pos == (off_t)-1) {
> - if (errno == ENXIO) {
> - ret = 0;
> - pos = lseek(fd, 0, SEEK_END);
> - if (pos != (off_t)-1)
> - zeros += pos - old;
> - } else {
> - ret = -2;
> - }
> - break;
> - }
> ret = read(fd, buf, sizeof(buf));
> - assert(ret); /* eof found by lseek */
> - if (ret <= 0)
> + if (ret < 0)
> + return -errno;
> + sum_add(dst, buf, ret);
> + if (ret < sizeof(buf))
> break;
> - if (old < pos) /* hole */
> - zeros += pos - old;
> - for (i = 0; i < ret; ++i) {
> - for (old = i; buf[i] == 0 && i < ret; ++i)
> - ;
> - if (old < i) /* code like a hole */
> - zeros += i - old;
> - if (i == ret)
> - break;
> - if (zeros) {
> - if (verbose >= 2)
> - fprintf(stderr,
> - "adding %llu zeros to sum\n",
> - (unsigned long long)zeros);
> - sum_add_u64(dst, 0);
> - sum_add_u64(dst, zeros);
> - zeros = 0;
> - }
> - for (old = i; buf[i] != 0 && i < ret; ++i)
> - ;
> - if (verbose >= 2)
> - fprintf(stderr, "adding %u non-zeros to sum\n",
> - i - (int)old);
> - sum_add(dst, buf + old, i - old);
> - }
> - pos += ret;
> }
> -
> - if (zeros) {
> - if (verbose >= 2)
> - fprintf(stderr,
> - "adding %llu zeros to sum (finishing)\n",
> - (unsigned long long)zeros);
> - sum_add_u64(dst, 0);
> - sum_add_u64(dst, zeros);
> - }
> -
> - return ret;
> + return 0;
> }
>
> int
>
signature.asc
Description: OpenPGP digital signature
