]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/commitdiff
Btrfs: fix handling of faults from btrfs_copy_from_user
authorChris Mason <clm@fb.com>
Mon, 16 May 2016 16:21:01 +0000 (09:21 -0700)
committerChris Mason <clm@fb.com>
Thu, 26 May 2016 20:23:59 +0000 (13:23 -0700)
When btrfs_copy_from_user isn't able to copy all of the pages, we need
to adjust our accounting to reflect the work that was actually done.

Commit 2e78c927d79 changed around the decisions a little and we ended up
skipping the accounting adjustments some of the time.  This commit makes
sure that when we don't copy anything at all, we still hop into
the adjustments, and switches to release_bytes instead of write_bytes,
since write_bytes isn't aligned.

The accounting errors led to warnings during btrfs_destroy_inode:

[   70.847532] WARNING: CPU: 10 PID: 514 at fs/btrfs/inode.c:9350 btrfs_destroy_inode+0x2b3/0x2c0
[   70.847536] Modules linked in: i2c_piix4 virtio_net i2c_core input_leds button led_class serio_raw acpi_cpufreq sch_fq_codel autofs4 virtio_blk
[   70.847538] CPU: 10 PID: 514 Comm: umount Tainted: G        W 4.6.0-rc6_00062_g2997da1-dirty #23
[   70.847539] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.9.0-1.fc24 04/01/2014
[   70.847542]  0000000000000000 ffff880ff5cafab8 ffffffff8149d5e9 0000000000000202
[   70.847543]  0000000000000000 0000000000000000 0000000000000000 ffff880ff5cafb08
[   70.847547]  ffffffff8107bdfd ffff880ff5cafaf8 000024868120013d ffff880ff5cafb28
[   70.847547] Call Trace:
[   70.847550]  [<ffffffff8149d5e9>] dump_stack+0x51/0x78
[   70.847551]  [<ffffffff8107bdfd>] __warn+0xfd/0x120
[   70.847553]  [<ffffffff8107be3d>] warn_slowpath_null+0x1d/0x20
[   70.847555]  [<ffffffff8139c9e3>] btrfs_destroy_inode+0x2b3/0x2c0
[   70.847556]  [<ffffffff812003a1>] ? __destroy_inode+0x71/0x140
[   70.847558]  [<ffffffff812004b3>] destroy_inode+0x43/0x70
[   70.847559]  [<ffffffff810b7b5f>] ? wake_up_bit+0x2f/0x40
[   70.847560]  [<ffffffff81200c68>] evict+0x148/0x1d0
[   70.847562]  [<ffffffff81398ade>] ? start_transaction+0x3de/0x460
[   70.847564]  [<ffffffff81200d49>] dispose_list+0x59/0x80
[   70.847565]  [<ffffffff81201ba0>] evict_inodes+0x180/0x190
[   70.847566]  [<ffffffff812191ff>] ? __sync_filesystem+0x3f/0x50
[   70.847568]  [<ffffffff811e95f8>] generic_shutdown_super+0x48/0x100
[   70.847569]  [<ffffffff810b75c0>] ? woken_wake_function+0x20/0x20
[   70.847571]  [<ffffffff811e9796>] kill_anon_super+0x16/0x30
[   70.847573]  [<ffffffff81365cde>] btrfs_kill_super+0x1e/0x130
[   70.847574]  [<ffffffff811e99be>] deactivate_locked_super+0x4e/0x90
[   70.847576]  [<ffffffff811e9e61>] deactivate_super+0x51/0x70
[   70.847577]  [<ffffffff8120536f>] cleanup_mnt+0x3f/0x80
[   70.847579]  [<ffffffff81205402>] __cleanup_mnt+0x12/0x20
[   70.847581]  [<ffffffff81098358>] task_work_run+0x68/0xa0
[   70.847582]  [<ffffffff810022b6>] exit_to_usermode_loop+0xd6/0xe0
[   70.847583]  [<ffffffff81002e1d>] do_syscall_64+0xbd/0x170
[   70.847586]  [<ffffffff817d4dbc>] entry_SYSCALL64_slow_path+0x25/0x25

This is the test program I used to force short returns from
btrfs_copy_from_user

void *dontneed(void *arg)
{
char *p = arg;
int ret;

while(1) {
ret = madvise(p, BUFSIZE/4, MADV_DONTNEED);
if (ret) {
perror("madvise");
exit(1);
}
}
}

int main(int ac, char **av) {
int ret;
int fd;
char *filename;
unsigned long offset;
char *buf;
int i;
pthread_t tid;

if (ac != 2) {
fprintf(stderr, "usage: dammitdave filename\n");
exit(1);
}

buf = mmap(NULL, BUFSIZE, PROT_READ|PROT_WRITE,
   MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
if (buf == MAP_FAILED) {
perror("mmap");
exit(1);
}
memset(buf, 'a', BUFSIZE);
filename = av[1];

ret = pthread_create(&tid, NULL, dontneed, buf);
if (ret) {
fprintf(stderr, "error %d from pthread_create\n", ret);
exit(1);
}

ret = pthread_detach(tid);
if (ret) {
fprintf(stderr, "pthread detach failed %d\n", ret);
exit(1);
}

while (1) {
fd = open(filename, O_RDWR | O_CREAT, 0600);
if (fd < 0) {
perror("open");
exit(1);
}

for (i = 0; i < ROUNDS; i++) {
int this_write = BUFSIZE;

offset = rand() % MAXSIZE;
ret = pwrite(fd, buf, this_write, offset);
if (ret < 0) {
perror("pwrite");
exit(1);
} else if (ret != this_write) {
fprintf(stderr, "short write to %s offset %lu ret %d\n",
filename, offset, ret);
exit(1);
}
if (i == ROUNDS - 1) {
ret = sync_file_range(fd, offset, 4096,
    SYNC_FILE_RANGE_WRITE);
if (ret < 0) {
perror("sync_file_range");
exit(1);
}
}
}
ret = ftruncate(fd, 0);
if (ret < 0) {
perror("ftruncate");
exit(1);
}
ret = close(fd);
if (ret) {
perror("close");
exit(1);
}
ret = unlink(filename);
if (ret) {
perror("unlink");
exit(1);
}

}
return 0;
}

Signed-off-by: Chris Mason <clm@fb.com>
Reported-by: Dave Jones <dsj@fb.com>
Fixes: 2e78c927d79333f299a8ac81c2fd2952caeef335
cc: stable@vger.kernel.org # v4.6
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/file.c

index 1933e8b07ed7f386b78f93efb50f5047f4395596..159a93450e269a41de65787a9fc402218342bb79 100644 (file)
@@ -1596,6 +1596,13 @@ again:
 
                copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
 
+               num_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info,
+                                               reserve_bytes);
+               dirty_sectors = round_up(copied + sector_offset,
+                                       root->sectorsize);
+               dirty_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info,
+                                               dirty_sectors);
+
                /*
                 * if we have trouble faulting in the pages, fall
                 * back to one page at a time
@@ -1605,6 +1612,7 @@ again:
 
                if (copied == 0) {
                        force_page_uptodate = true;
+                       dirty_sectors = 0;
                        dirty_pages = 0;
                } else {
                        force_page_uptodate = false;
@@ -1615,20 +1623,19 @@ again:
                /*
                 * If we had a short copy we need to release the excess delaloc
                 * bytes we reserved.  We need to increment outstanding_extents
-                * because btrfs_delalloc_release_space will decrement it, but
+                * because btrfs_delalloc_release_space and
+                * btrfs_delalloc_release_metadata will decrement it, but
                 * we still have an outstanding extent for the chunk we actually
                 * managed to copy.
                 */
-               num_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info,
-                                               reserve_bytes);
-               dirty_sectors = round_up(copied + sector_offset,
-                                       root->sectorsize);
-               dirty_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info,
-                                               dirty_sectors);
-
                if (num_sectors > dirty_sectors) {
-                       release_bytes = (write_bytes - copied)
-                               & ~((u64)root->sectorsize - 1);
+                       /*
+                        * we round down because we don't want to count
+                        * any partial blocks actually sent through the
+                        * IO machines
+                        */
+                       release_bytes = round_down(release_bytes - copied,
+                                     root->sectorsize);
                        if (copied > 0) {
                                spin_lock(&BTRFS_I(inode)->lock);
                                BTRFS_I(inode)->outstanding_extents++;