We should kill free_some_buffers() to stop reclaiming extent buffers or
we will hit a problem described below.
As of commit 53ee1bccf9, we are not
counting a reference for tree->lru anymore. However free_some_buffers()
is still left and is reclaiming extent buffers whose @refs == 1. This
cause extent buffers to be reclaimed unintentionally. Thus the following
steps could happen:
1. A buffer at address A is reclaimed by free_some_buffers()
(address A is also free()ed)
2. Some code call alloc_extent_buffer()
3. Address A is assigned to newly allocated buffer
4. You see a buffer pointed by A suddenly changed its content
This problem is also pointed out here and it has a reproducer:
https://www.mail-archive.com/linux-btrfs@vger.kernel.org/msg36703.html
This commit drop free_some_buffers() and related variables, and also it
modify extent_io_tree_cleanup() to catch non-free'ed buffers properly.
Signed-off-by: Naohiro Aota <naota@elisp.net>
Signed-off-by: David Sterba <dsterba@suse.cz>
free_some_buffer() should not free dirty extent buffers. They are left
to be committed.
Signed-off-by: Naohiro Aota <naota@elisp.net>
Signed-off-by: David Sterba <dsterba@suse.cz>
Now we set @refs to 2 on creating a new extent buffer, meanwhile we
allocate the needed free space, but we don't give enough free_extent_buffer()
to reduce the eb's references to zero so that the eb can finally be freed,
so the problem is we has decrease the referene count of backrefs to zero, which
ends up releasing the space occupied by the eb, and this space can be allocated
again for something else(another eb or disk), usually a crash(core dump) will
occur, I've hit a crash in rb_insert() because another eb re-use the space while
the original one is floating around.
We should do the same thing as the kernel code does, it's necessary to initialize
@refs to 1 instead of 2, this helps us get rid of the above problem.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
A user was reporting an issue with bad transid errors on his blocks. The thing
is that btrfs-progs will ignore transid failures for things like restore and
fsck so we can do a best effort to fix a users file system. So fsck can put
together a coherent view of the file system with stale blocks. So if everything
else is ok in the mind of fsck then we can recow these blocks to fix the
generation and the user can get their file system back. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
The 'prealloc' extent_state structure is leaked for the case when the 'desired
range' encapsulates/covers the 'extent range'.
Signed-off-by: chandan <chandan@linux.vnet.ibm.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
In files copied from the kernel, mark many functions as static,
and remove any resulting dead code.
Some functions are left unmarked if they aren't static in the
kernel tree.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
It should be 'clear', not 'set'.
Signed-off-by: Wang Sheng-Hui <shhuiw@gmail.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
This adds a 'btrfs-image -m' option, which let us restore an image that
is built from a btrfs of multiple disks onto several disks altogether.
This aims to address the following case,
$ mkfs.btrfs -m raid0 sda sdb
$ btrfs-image sda image.file
$ btrfs-image -r image.file sdc
---------
so we can only restore metadata onto sdc, and another thing is we can
only mount sdc with degraded mode as we don't provide informations of
another disk. And, it's built as RAID0 and we have only one disk,
so after mount sdc we'll get into readonly mode.
This is just annoying for people(like me) who're trying to restore image
but turn to find they cannot make it work.
So this'll make your life easier, just tap
$ btrfs-image -m image.file sdc sdd
---------
then you get everything about metadata done, the same offset with that of
the originals(of course, you need offer enough disk size, at least the disk
size of the original disks).
Besides, this also works with raid5 and raid6 metadata image.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
As we know, btrfs can manage several devices in the same fs, so [offset, size]
is not sufficient for unique identification of an device extent, we need the
device id to identify the device extents which have the same offset and size,
but are not in the same device. So, we added a member variant named objectid
into the extent cache, and introduced some functions to make the extent cache
be suitable to manage the device extent.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
In fact, the code of many rb-tree insert/search/delete functions is similar,
so we can abstract them, and implement common functions for rb-tree, and then
simplify them.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
In trying to track down a weird tree log problem I wanted to make sure that the
free space cache was actually valid, which we currently have no way of doing.
So this patch adds a bunch of support for the free space cache code and then a
checker to fsck. Basically we go through and if we can actually load the free
space cache then we will walk the extent tree and verify that the free space
cache exactly matches what is in the extent tree. Hopefully this will always be
correct, the only time it wouldn't is if the extent tree is corrupt or we have
some sort of awful bug in the free space cache. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Please find attached a patch to make the new libbtrfs usable from
C++ (at least for the parts snapper will likely need).
Signed-off-by: Arvin Schnell <aschnell@suse.de>
Signed-off-by: Mark Fasheh <mfasheh@suse.de>
It looks possible to hit the search_again label without using the
prealloc. A new prealloc is allocated, leaking the current one.
Every use of prealloc sets it to null so let's just allocate a new
prealloc when we don't already have one.
Signed-off-by: Zach Brown <zab@redhat.com>
David Woodhouse originally contributed this code, and Chris Mason
changed it around to reflect the current design goals for raid56.
The original code expected all metadata and data writes to be full
stripes. This meant metadata block size == stripe size, and had a few
other restrictions.
This version allows metadata blocks smaller than the stripe size. It
implements both raid5 and raid6, although it does not have code to
rebuild from parity if one of the drives is missing or incorrect.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
This changes free_some_buffers (called each time we allocate an extent
buffer) to allow a higher hard limit on the number of extent buffers
in use.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
fsck needs to be able to open a damaged FS, which means open_ctree needs
to be able to return a damaged FS.
This adds a new open_ctree_fs_info which can be used to open any and all
roots that are valid. btrfs-debug-tree is changed to use it.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Found by valgrind:
==8968== Use of uninitialised value of size 8
==8968== at 0x41CE7D: crc32c_le (crc32c.c:98)
==8968== by 0x40A1D0: csum_tree_block_size (disk-io.c:82)
==8968== by 0x40A2D4: csum_tree_block (disk-io.c:105)
==8968== by 0x40A7D6: write_tree_block (disk-io.c:241)
==8968== by 0x40ACEE: __commit_transaction (disk-io.c:354)
==8968== by 0x40AE9E: btrfs_commit_transaction (disk-io.c:385)
==8968== by 0x42CF66: make_image (mkfs.c:1061)
==8968== by 0x42DE63: main (mkfs.c:1410)
==8968== Uninitialised value was created by a stack allocation
==8968== at 0x42B5FB: add_inode_items (mkfs.c:493)
1. On-disk inode format has reserved (and thus, random at alloc time) fields:
btrfs_inode_item: __le64 reserved[4]
2. Sometimes extents are created on disk without writing data there.
(Or at least not all data is written there). Kernel code always had
it kzalloc'ed.
Zero them all.
Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org>
Signed-off-by: Hugo Mills <hugo@carfax.org.uk>
gcc-4.6 (as shipped in Fedora) turns on -Wunused-but-set-variable by
default, which breaks the build when combined with -Wall, e.g.:
debug-tree.c: In function ‘print_extent_leaf’:
debug-tree.c:45:13: error: variable ‘last_len’ set but not used [-Werror=unused-but-set-variable]
debug-tree.c:44:13: error: variable ‘last’ set but not used [-Werror=unused-but-set-variable]
debug-tree.c:41:21: error: variable ‘item’ set but not used [-Werror=unused-but-set-variable]
cc1: all warnings being treated as errors
This patch fixes the errors by removing the unused variables.
Signed-off-by: Chris Ball <cjb@laptop.org>
Signed-off-by: Hugo Mills <hugo@carfax.org.uk>
This commit introduces a new kind of back reference for btrfs metadata.
Once a filesystem has been mounted with this commit, IT WILL NO LONGER
BE MOUNTABLE BY OLDER KERNELS.
The new back ref provides information about pointer's key, level and in which
tree the pointer lives. This information allow us to find the pointer by
searching the tree. The shortcoming of the new back ref is that it only works
for pointers in tree blocks referenced by their owner trees.
This is mostly a problem for snapshots, where resolving one of these fuzzy back
references would be O(number_of_snapshots) and quite slow. The solution used
here is to use the fuzzy back references in the common case where a given tree
block is only referenced by one root, and use the full back references when
multiple roots have a reference
The main changes in this patch are adding chunk handing and data relocation
ability. In the last step of conversion, the converter relocates data in system
chunk and move chunk tree into system chunk. In the rollback process, the
converter remove chunk tree from system chunk and copy data back.
Regards
YZ
---
We get lots of warnings of the flavor:
utils.c:441: warning: format '%Lu' expects type 'long long unsigned int' but argument 2 has type 'u64'
And thanks to -Werror, the build fails. Clean up these printfs
by properly casting the arg to the format specified.
Signed-off-by: Alex Chiang <achiang@hp.com>
