btrfs: make fiemap more efficient and accurate reporting extent sharedness

btrfs: make fiemap more efficient and accurate reporting extent sharedness

The current fiemap implementation does not scale very well with the number
of extents a file has. This is both because the main algorithm to find out
the extents has a high algorithmic complexity and because for each extent
we have to check if it's shared. This second part, checking if an extent
is shared, is significantly improved by the two previous patches in this
patchset, while the first part is improved by this specific patch. Every
now and then we get reports from users mentioning fiemap is too slow or
even unusable for files with a very large number of extents, such as the
two recent reports referred to by the Link tags at the bottom of this
change log.

To understand why the part of finding which extents a file has is very
inefficient, consider the example of doing a full ranged fiemap against
a file that has over 100K extents (normal for example for a file with
more than 10G of data and using compression, which limits the extent size
to 128K). When we enter fiemap at extent_fiemap(), the following happens:

1) Before entering the main loop, we call get_extent_skip_holes() to get
   the first extent map. This leads us to btrfs_get_extent_fiemap(), which
   in turn calls btrfs_get_extent(), to find the first extent map that
   covers the file range [0, LLONG_MAX).

   btrfs_get_extent() will first search the inode's extent map tree, to
   see if we have an extent map there that covers the range. If it does
   not find one, then it will search the inode's subvolume b+tree for a
   fitting file extent item. After finding the file extent item, it will
   allocate an extent map, fill it in with information extracted from the
   file extent item, and add it to the inode's extent map tree (which
   requires a search for insertion in the tree).

2) Then we enter the main loop at extent_fiemap(), emit the details of
   the extent, and call again get_extent_skip_holes(), with a start
   offset matching the end of the extent map we previously processed.

   We end up at btrfs_get_extent() again, will search the extent map tree
   and then search the subvolume b+tree for a file extent item if we could
   not find an extent map in the extent tree. We allocate an extent map,
   fill it in with the details in the file extent item, and then insert
   it into the extent map tree (yet another search in this tree).

3) The second step is repeated over and over, until we have processed the
   whole file range. Each iteration ends at btrfs_get_extent(), which
   does a red black tree search on the extent map tree, then searches the
   subvolume b+tree, allocates an extent map and then does another search
   in the extent map tree in order to insert the extent map.

   In the best scenario we have all the extent maps already in the extent
   tree, and so for each extent we do a single search on a red black tree,
   so we have a complexity of O(n log n).

   In the worst scenario we don't have any extent map already loaded in
   the extent map tree, or have very few already there. In this case the
   complexity is much higher since we do:

   - A red black tree search on the extent map tree, which has O(log n)
     complexity, initially very fast since the tree is empty or very
     small, but as we end up allocating extent maps and adding them to
     the tree when we don't find them there, each subsequent search on
     the tree gets slower, since it's getting bigger and bigger after
     each iteration.

   - A search on the subvolume b+tree, also O(log n) complexity, but it
     has items for all inodes in the subvolume, not just items for our
     inode. Plus on a filesystem with concurrent operations on other
     inodes, we can block doing the search due to lock contention on
     b+tree nodes/leaves.

   - Allocate an extent map - this can block, and can also fail if we
     are under serious memory pressure.

   - Do another search on the extent maps red black tree, with the goal
     of inserting the extent map we just allocated. Again, after every
     iteration this tree is getting bigger by 1 element, so after many
     iterations the searches are slower and slower.

   - We will not need the allocated extent map anymore, so it's pointless
     to add it to the extent map tree. It's just wasting time and memory.

   In short we end up searching the extent map tree multiple times, on a
   tree that is growing bigger and bigger after each iteration. And
   besides that we visit the same leaf of the subvolume b+tree many times,
   since a leaf with the default size of 16K can easily have more than 200
   file extent items.

This is very inefficient overall. This patch changes the algorithm to
instead iterate over the subvolume b+tree, visiting each leaf only once,
and only searching in the extent map tree for file ranges that have holes
or prealloc extents, in order to figure out if we have delalloc there.
It will never allocate an extent map and add it to the extent map tree.
This is very similar to what was previously done for the lseek's hole and
data seeking features.

Also, the current implementation relying on extent maps for figuring out
which extents we have is not correct. This is because extent maps can be
merged even if they represent different extents - we do this to minimize
memory utilization and keep extent map trees smaller. For example if we
have two extents that are contiguous on disk, once we load the two extent
maps, they get merged into a single one - however if only one of the
extents is shared, we end up reporting both as shared or both as not
shared, which is incorrect.

This reproducer triggers that bug:

    $ cat fiemap-bug.sh
    #!/bin/bash

    DEV=/dev/sdj
    MNT=/mnt/sdj

    mkfs.btrfs -f $DEV
    mount $DEV $MNT

    # Create a file with two 256K extents.
    # Since there is no other write activity, they will be contiguous,
    # and their extent maps merged, despite having two distinct extents.
    xfs_io -f -c "pwrite -S 0xab 0 256K" \
              -c "fsync" \
              -c "pwrite -S 0xcd 256K 256K" \
              -c "fsync" \
              $MNT/foo

    # Now clone only the second extent into another file.
    xfs_io -f -c "reflink $MNT/foo 256K 0 256K" $MNT/bar

    # Filefrag will report a single 512K extent, and say it's not shared.
    echo
    filefrag -v $MNT/foo

    umount $MNT

Running the reproducer:

    $ ./fiemap-bug.sh
    wrote 262144/262144 bytes at offset 0
    256 KiB, 64 ops; 0.0038 sec (65.479 MiB/sec and 16762.7030 ops/sec)
    wrote 262144/262144 bytes at offset 262144
    256 KiB, 64 ops; 0.0040 sec (61.125 MiB/sec and 15647.9218 ops/sec)
    linked 262144/262144 bytes at offset 0
    256 KiB, 1 ops; 0.0002 sec (1.034 GiB/sec and 4237.2881 ops/sec)

    Filesystem type is: 9123683e
    File size of /mnt/sdj/foo is 524288 (128 blocks of 4096 bytes)
     ext:     logical_offset:        physical_offset: length:   expected: flags:
       0:        0..     127:       3328..      3455:    128:             last,eof
    /mnt/sdj/foo: 1 extent found

We end up reporting that we have a single 512K that is not shared, however
we have two 256K extents, and the second one is shared. Changing the
reproducer to clone instead the first extent into file 'bar', makes us
report a single 512K extent that is shared, which is algo incorrect since
we have two 256K extents and only the first one is shared.

This patch is part of a larger patchset that is comprised of the following
patches:

    btrfs: allow hole and data seeking to be interruptible
    btrfs: make hole and data seeking a lot more efficient
    btrfs: remove check for impossible block start for an extent map at fiemap
    btrfs: remove zero length check when entering fiemap
    btrfs: properly flush delalloc when entering fiemap
    btrfs: allow fiemap to be interruptible
    btrfs: rename btrfs_check_shared() to a more descriptive name
    btrfs: speedup checking for extent sharedness during fiemap
    btrfs: skip unnecessary extent buffer sharedness checks during fiemap
    btrfs: make fiemap more efficient and accurate reporting extent sharedness

The patchset was tested on a machine running a non-debug kernel (Debian's
default config) and compared the tests below on a branch without the
patchset versus the same branch with the whole patchset applied.

The following test for a large compressed file without holes:

    $ cat fiemap-perf-test.sh
    #!/bin/bash

    DEV=/dev/sdi
    MNT=/mnt/sdi

    mkfs.btrfs -f $DEV
    mount -o compress=lzo $DEV $MNT

    # 40G gives 327680 128K file extents (due to compression).
    xfs_io -f -c "pwrite -S 0xab -b 1M 0 20G" $MNT/foobar

    umount $MNT
    mount -o compress=lzo $DEV $MNT

    start=$(date +%s%N)
    filefrag $MNT/foobar
    end=$(date +%s%N)
    dur=$(( (end - start) / 1000000 ))
    echo "fiemap took $dur milliseconds (metadata not cached)"

    start=$(date +%s%N)
    filefrag $MNT/foobar
    end=$(date +%s%N)
    dur=$(( (end - start) / 1000000 ))
    echo "fiemap took $dur milliseconds (metadata cached)"

    umount $MNT

Before patchset:

    $ ./fiemap-perf-test.sh
    (...)
    /mnt/sdi/foobar: 327680 extents found
    fiemap took 3597 milliseconds (metadata not cached)
    /mnt/sdi/foobar: 327680 extents found
    fiemap took 2107 milliseconds (metadata cached)

After patchset:

    $ ./fiemap-perf-test.sh
    (...)
    /mnt/sdi/foobar: 327680 extents found
    fiemap took 1214 milliseconds (metadata not cached)
    /mnt/sdi/foobar: 327680 extents found
    fiemap took 684 milliseconds (metadata cached)

That's a speedup of about 3x for both cases (no metadata cached and all
metadata cached).

The test provided by Pavel (first Link tag at the bottom), which uses
files with a large number of holes, was also used to measure the gains,
and it consists on a small C program and a shell script to invoke it.
The C program is the following:

    $ cat pavels-test.c
    #include <stdio.h>
    #include <unistd.h>
    #include <stdlib.h>
    #include <fcntl.h>

    #include <sys/stat.h>
    #include <sys/time.h>
    #include <sys/ioctl.h>

    #include <linux/fs.h>
    #include <linux/fiemap.h>

    #define FILE_INTERVAL (1<<13) /* 8Kb */

    long long interval(struct timeval t1, struct timeval t2)
    {
        long long val = 0;
        val += (t2.tv_usec - t1.tv_usec);
        val += (t2.tv_sec - t1.tv_sec) * 1000 * 1000;
        return val;
    }

    int main(int argc, char **argv)
    {
        struct fiemap fiemap = {};
        struct timeval t1, t2;
        char data = 'a';
        struct stat st;
        int fd, off, file_size = FILE_INTERVAL;

        if (argc != 3 && argc != 2) {
                printf("usage: %s <path> [size]\n", argv[0]);
                return 1;
        }

        if (argc == 3)
                file_size = atoi(argv[2]);
        if (file_size < FILE_INTERVAL)
                file_size = FILE_INTERVAL;
        file_size -= file_size % FILE_INTERVAL;

        fd = open(argv[1], O_RDWR | O_CREAT | O_TRUNC, 0644);
        if (fd < 0) {
            perror("open");
            return 1;
        }

        for (off = 0; off < file_size; off += FILE_INTERVAL) {
            if (pwrite(fd, &data, 1, off) != 1) {
                perror("pwrite");
                close(fd);
                return 1;
            }
        }

        if (ftruncate(fd, file_size)) {
            perror("ftruncate");
            close(fd);
            return 1;
        }

        if (fstat(fd, &st) < 0) {
            perror("fstat");
            close(fd);
            return 1;
        }

        printf("size: %ld\n", st.st_size);
        printf("actual size: %ld\n", st.st_blocks * 512);

        fiemap.fm_length = FIEMAP_MAX_OFFSET;
        gettimeofday(&t1, NULL);
        if (ioctl(fd, FS_IOC_FIEMAP, &fiemap) < 0) {
            perror("fiemap");
            close(fd);
            return 1;
        }
        gettimeofday(&t2, NULL);

        printf("fiemap: fm_mapped_extents = %d\n",
               fiemap.fm_mapped_extents);
        printf("time = %lld us\n", interval(t1, t2));

        close(fd);
        return 0;
    }

    $ gcc -o pavels_test pavels_test.c

And the wrapper shell script:

    $ cat fiemap-pavels-test.sh

    #!/bin/bash

    DEV=/dev/sdi
    MNT=/mnt/sdi

    mkfs.btrfs -f -O no-holes $DEV
    mount $DEV $MNT

    echo
    echo "*********** 256M ***********"
    echo

    ./pavels-test $MNT/testfile $((1 << 28))
    echo
    ./pavels-test $MNT/testfile $((1 << 28))

    echo
    echo "*********** 512M ***********"
    echo

    ./pavels-test $MNT/testfile $((1 << 29))
    echo
    ./pavels-test $MNT/testfile $((1 << 29))

    echo
    echo "*********** 1G ***********"
    echo

    ./pavels-test $MNT/testfile $((1 << 30))
    echo
    ./pavels-test $MNT/testfile $((1 << 30))

    umount $MNT

Running his reproducer before applying the patchset:

    *********** 256M ***********

    size: 268435456
    actual size: 134217728
    fiemap: fm_mapped_extents = 32768
    time = 4003133 us

    size: 268435456
    actual size: 134217728
    fiemap: fm_mapped_extents = 32768
    time = 4895330 us

    *********** 512M ***********

    size: 536870912
    actual size: 268435456
    fiemap: fm_mapped_extents = 65536
    time = 30123675 us

    size: 536870912
    actual size: 268435456
    fiemap: fm_mapped_extents = 65536
    time = 33450934 us

    *********** 1G ***********

    size: 1073741824
    actual size: 536870912
    fiemap: fm_mapped_extents = 131072
    time = 224924074 us

    size: 1073741824
    actual size: 536870912
    fiemap: fm_mapped_extents = 131072
    time = 217239242 us

Running it after applying the patchset:

    *********** 256M ***********

    size: 268435456
    actual size: 134217728
    fiemap: fm_mapped_extents = 32768
    time = 29475 us

    size: 268435456
    actual size: 134217728
    fiemap: fm_mapped_extents = 32768
    time = 29307 us

    *********** 512M ***********

    size: 536870912
    actual size: 268435456
    fiemap: fm_mapped_extents = 65536
    time = 58996 us

    size: 536870912
    actual size: 268435456
    fiemap: fm_mapped_extents = 65536
    time = 59115 us

    *********** 1G ***********

    size: 1073741824
    actual size: 536870912
    fiemap: fm_mapped_extents = 116251
    time = 124141 us

    size: 1073741824
    actual size: 536870912
    fiemap: fm_mapped_extents = 131072
    time = 119387 us

The speedup is massive, both on the first fiemap call and on the second
one as well, as his test creates files with many holes and small extents
(every extent follows a hole and precedes another hole).

For the 256M file we go from 4 seconds down to 29 milliseconds in the
first run, and then from 4.9 seconds down to 29 milliseconds again in the
second run, a speedup of 138x and 169x, respectively.

For the 512M file we go from 30.1 seconds down to 59 milliseconds in the
first run, and then from 33.5 seconds down to 59 milliseconds again in the
second run, a speedup of 510x and 568x, respectively.

For the 1G file, we go from 225 seconds down to 124 milliseconds in the
first run, and then from 217 seconds down to 119 milliseconds in the
second run, a speedup of 1815x and 1824x, respectively.

Reported-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
Link: https://lore.kernel.org/linux-btrfs/21dd32c6-f1f9-f44a-466a-e18fdc6788a7@virtuozzo.com/
Reported-by: Dominique MARTINET <dominique.martinet@atmark-techno.com>
Link: https://lore.kernel.org/linux-btrfs/Ysace25wh5BbLd5f@atmark-techno.com/
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>