Allocate uninitialized extent before ext4 buffer write and
convert the extent to initialized after io completes.
The purpose is to make sure an extent can only be marked
initialized after it has been written with new data so
we can safely drop the i_mutex lock in ext4 DIO read without
exposing stale data. This helps to improve multi-thread DIO
read performance on high-speed disks.
Skip the nobh and data=journal mount cases to make things simple for now.
Signed-off-by: Jiaying Zhang <jiayingz@google.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
This commit renames some of the direct I/O's block allocation flags,
variables, and functions introduced in Mingming's "Direct IO for holes
and fallocate" patches so that they can be used by ext4's buffered
write path as well. Also changed the related function comments
accordingly to cover both direct write and buffered write cases.
Signed-off-by: Jiaying Zhang <jiayingz@google.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
fallocate() may potentially instantiate blocks past EOF, depending
on the flags used when it is called.
e2fsck currently has a test for blocks past i_size, and it
sometimes trips up - noticeably on xfstests 013 which runs fsstress.
This patch from Jiayang does fix it up - it (along with
e2fsprogs updates and other patches recently from Aneesh) has
survived many fsstress runs in a row.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Jiaying Zhang <jiayingz@google.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Just a pet peeve of mine; we had a mishash of calls with either __func__
or "function_name" and the latter tends to get out of sync.
I think it's easier to just hide the __func__ in a macro, and it'll
be consistent from then on.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
At several places we modify EXT4_I(inode)->i_state without holding
i_mutex (ext4_release_file, ext4_bmap, ext4_journalled_writepage,
ext4_do_update_inode, ...). These modifications are racy and we can
lose updates to i_state. So convert handling of i_state to use bitops
which are atomic.
Cc: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Add checks to ext4_free_branches() to make sure a block number found
in an indirect block are valid before trying to free it. If a bad
block number is found, stop freeing the indirect block immediately,
since the file system is corrupt and we will need to run fsck anyway.
This also avoids spamming the logs, and specifically avoids
driver-level "attempt to access beyond end of device" errors obscure
what is really going on.
If you get *really*, *really*, *really* unlucky, without this patch, a
supposed indirect block containing garbage might contain a reference
to a primary block group descriptor, in which case
ext4_free_branches() could end up zero'ing out a block group
descriptor block, and if then one of the block bitmaps for a block
group described by that bg descriptor block is not in memory, and is
read in by ext4_read_block_bitmap(). This function calls
ext4_valid_block_bitmap(), which assumes that bg_inode_table() was
validated at mount time and hasn't been modified since. Since this
assumption is no longer valid, it's possible for the value
(ext4_inode_table(sb, desc) - group_first_block) to go negative, which
will cause ext4_find_next_zero_bit() to trigger a kernel GPF.
Addresses-Google-Bug: #2220436
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The "offset" member in ext4_io_end holds bytes, not blocks, so
ext4_lblk_t is wrong - and too small (u32).
This caused the async i/o writes to sparse files beyond 4GB to fail
when they wrapped around to 0.
Also fix up the type of arguments to ext4_convert_unwritten_extents(),
it gets ssize_t from ext4_end_aio_dio_nolock() and
ext4_ext_direct_IO().
Reported-by: Giel de Nijs <giel@vectorwise.com>
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
We should update reserve space if it is delalloc buffer
and that is indicated by EXT4_GET_BLOCKS_DELALLOC_RESERVE flag.
So use EXT4_GET_BLOCKS_DELALLOC_RESERVE in place of
EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
When we fallocate a region of the file which we had recently written,
and which is still in the page cache marked as delayed allocated blocks
we need to make sure we don't do the quota update on writepage path.
This is because the needed quota updated would have already be done
by fallocate.
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
In the past, ext4_calc_metadata_amount(), and its sub-functions
ext4_ext_calc_metadata_amount() and ext4_indirect_calc_metadata_amount()
badly over-estimated the number of metadata blocks that might be
required for delayed allocation blocks. This didn't matter as much
when functions which managed the reserved metadata blocks were more
aggressive about dropping reserved metadata blocks as delayed
allocation blocks were written, but unfortunately they were too
aggressive. This was fixed in commit 0637c6f, but as a result the
over-estimation by ext4_calc_metadata_amount() would lead to reserving
2-3 times the number of pending delayed allocation blocks as
potentially required metadata blocks. So if there are 1 megabytes of
blocks which have been not yet been allocation, up to 3 megabytes of
space would get reserved out of the user's quota and from the file
system free space pool until all of the inode's data blocks have been
allocated.
This commit addresses this problem by much more accurately estimating
the number of metadata blocks that will be required. It will still
somewhat over-estimate the number of blocks needed, since it must make
a worst case estimate not knowing which physical blocks will be
needed, but it is much more accurate than before.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
This patch also fixes write vs chown race condition.
Acked-by: "Theodore Ts'o" <tytso@mit.edu>
Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org>
Signed-off-by: Jan Kara <jack@suse.cz>
We cannot rely on buffer dirty bits during fsync because pdflush can come
before fsync is called and clear dirty bits without forcing a transaction
commit. What we do is that we track which transaction has last changed
the inode and which transaction last changed allocation and force it to
disk on fsync.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Add the facility for ext4_forget() to be called from
ext4_free_blocks(). This simplifies the code in a large number of
places, and centralizes most of the work of calling ext4_forget() into
a single place.
Also fix a bug in the extents migration code; it wasn't calling
ext4_forget() when releasing the indirect blocks during the
conversion. As a result, if the system cashed during or shortly after
the extents migration, and the released indirect blocks get reused as
data blocks, the journal replay would corrupt the data blocks. With
this new patch, fixing this bug was as simple as adding the
EXT4_FREE_BLOCKS_FORGET flags to the call to ext4_free_blocks().
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
ext4_mb_free_blocks() is only called by ext4_free_blocks(), and the
latter function doesn't really do much. So merge the two functions
together, such that ext4_free_blocks() is now found in
fs/ext4/mballoc.c. This saves about 200 bytes of compiled text space.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The ext4_forget() function better belongs in ext4_jbd2.c. This will
allow us to do some cleanup of the ext4_journal_revoke() and
ext4_journal_forget() functions, as well as giving us better error
reporting since we can report the caller of ext4_forget() when things
go wrong.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
It is anticipated that when sb_issue_discard starts doing
real work on trim-capable devices, we may see issues. Make
this mount-time optional, and default it to off until we know
that things are working out OK.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
At the end of direct I/O operation, ext4_ext_direct_IO() always called
ext4_convert_unwritten_extents(), regardless of whether there were any
unwritten extents involved in the I/O or not.
This commit adds a state flag so that ext4_ext_direct_IO() only calls
ext4_convert_unwritten_extents() when necessary.
Signed-off-by: Mingming Cao <cmm@us.ibm.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
"Looking at ext4.h, I think the setting of extra time fields forgets to
mask the epoch bits so the epoch part overwrites nsec part. The second
change is only for coherency (2 -> EXT4_EPOCH_BITS)."
Thanks to Damien Guibouret for pointing out this problem.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The /proc/fs/ext4/<dev>/mb_history was maintained manually, and had a
number of problems: it required a largish amount of memory to be
allocated for each ext4 filesystem, and the s_mb_history_lock
introduced a CPU contention problem.
By ripping out the mb_history code and replacing it with ftrace
tracepoints, and we get more functionality: timestamps, event
filtering, the ability to correlate mballoc history with other ext4
tracepoints, etc.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
For async direct IO that covers holes or fallocate, the end_io
callback function now queued the convertion work on workqueue but
don't flush the work rightaway as it might take too long to afford.
But when fsync is called after all the data is completed, user expects
the metadata also being updated before fsync returns.
Thus we need to flush the conversion work when fsync() is called.
This patch keep track of a listed of completed async direct io that
has a work queued on workqueue. When fsync() is called, it will go
through the list and do the conversion.
Signed-off-by: Mingming Cao <cmm@us.ibm.com>
Currently the DIO VFS code passes create = 0 when writing to the
middle of file. It does this to avoid block allocation for holes, so
as not to expose stale data out when there is a parallel buffered read
(which does not hold the i_mutex lock). Direct I/O writes into holes
falls back to buffered IO for this reason.
Since preallocated extents are treated as holes when doing a
get_block() look up (buffer is not mapped), direct IO over fallocate
also falls back to buffered IO. Thus ext4 actually silently falls
back to buffered IO in above two cases, which is undesirable.
To fix this, this patch creates unitialized extents when a direct I/O
write into holes in sparse files, and registering an end_io callback which
converts the uninitialized extent to an initialized extent after the
I/O is completed.
Singed-Off-By: Mingming Cao <cmm@us.ibm.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
When writing into an unitialized extent via direct I/O, and the direct
I/O doesn't exactly cover the unitialized extent, split the extent
into uninitialized and initialized extents before submitting the I/O.
This avoids needing to deal with an ENOSPC error in the end_io
callback that gets used for direct I/O.
When the IO is complete, the written extent will be marked as initialized.
Singed-Off-By: Mingming Cao <cmm@us.ibm.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Work around problems in the writeback code to force out writebacks in
larger chunks than just 4mb, which is just too small. This also works
around limitations in the ext4 block allocator, which can't allocate
more than 2048 blocks at a time. So we need to defeat the round-robin
characteristics of the writeback code and try to write out as many
blocks in one inode before allowing the writeback code to move on to
another inode. We add a a new per-filesystem tunable,
max_writeback_mb_bump, which caps this to a default of 128mb per
inode.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
There's no reason to redefine the maximum allowable offset
in an extent-based file just for defrag;
EXT_MAX_BLOCK already does this.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
EXT4_EXT_MIGRATE is only intended to be used for an in-memory flag,
and the hex value assigned to it collides with FS_DIRECTIO_FL (which
is also stored in i_flags). There's no reason for the
EXT4_EXT_MIGRATE bit to be stored in i_flags, so we switch it to use
i_state instead.
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Today, the ext4 allocator will happily allocate blocks past
2^32 for indirect-block files, which results in the block
numbers getting truncated, and corruption ensues.
This patch limits such allocations to < 2^32, and adds
BUG_ONs if we do get blocks larger than that.
This should address RH Bug 519471, ext4 bitmap allocator
must limit blocks to < 2^32
* ext4_find_goal() is modified to choose a goal < UINT_MAX,
so that our starting point is in an acceptable range.
* ext4_xattr_block_set() is modified such that the goal block
is < UINT_MAX, as above.
* ext4_mb_regular_allocator() is modified so that the group
search does not continue into groups which are too high
* ext4_mb_use_preallocated() has a check that we don't use
preallocated space which is too far out
* ext4_alloc_blocks() and ext4_xattr_block_set() add some BUG_ONs
No attempt has been made to limit inode locations to < 2^32,
so we may wind up with blocks far from their inodes. Doing
this much already will lead to some odd ENOSPC issues when the
"lower 32" gets full, and further restricting inodes could
make that even weirder.
For high inodes, choosing a goal of the original, % UINT_MAX,
may be a bit odd, but then we're in an odd situation anyway,
and I don't know of a better heuristic.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
There's no real cost for the journal checksum feature, and we should
make sure it is enabled all the time.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Add a new tracepoint which shows the pages that will be written using
write_cache_pages() by ext4_da_writepages().
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
unsigned short is potentially too small to track blocks within
a group; today it is safe due to restrictions in e2fsprogs but
we have _lo / _hi bits for group blocks with the intent to go
up to 32 bits, so clean this up now.
There are many more places where we use unsigned/int/unsigned int
to contain a group block but this should at least fix all the
short types.
I added a few comments to the struct ext4_group_info definition
as well.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
ext4_mb_update_group_info is only called in one place, and it's
extremely simple. There's no reason to have it in a separate function
in a separate file as far as I can tell, it just obfuscates what's
really going on.
Perhaps it was intended to keep the grp->bb_* manipulation local to
mballoc.c but we're already accessing other grp-> fields in balloc.c
directly so this seems ok.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
During truncate we are sometimes forced to start a new transaction as
the amount of blocks to be journaled is both quite large and hard to
predict. So far we restarted a transaction while holding i_data_sem
and that violates lock ordering because i_data_sem ranks below a
transaction start (and it can lead to a real deadlock with
ext4_get_blocks() mapping blocks in some page while having a
transaction open).
We fix the problem by dropping the i_data_sem before restarting the
transaction and acquire it afterwards. It's slightly subtle that this
works:
1) By the time ext4_truncate() is called, all the page cache for the
truncated part of the file is dropped so get_block() should not be
called on it (we only have to invalidate extent cache after we
reacquire i_data_sem because some extent from not-truncated part could
extend also into the part we are going to truncate).
2) Writes, migrate or defrag hold i_mutex so they are stopped for all
the time of the truncate.
This bug has been found and analyzed by Theodore Tso <tytso@mit.edu>.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Currently the group preallocation code tries to find a large (512)
free block from which to do per-cpu group allocation for small files.
The problem with this scheme is that it leaves the filesystem horribly
fragmented. In the worst case, if the filesystem is unmounted and
remounted (after a system shutdown, for example) we forget the fact
that wee were using a particular (now-partially filled) 512 block
extent. So the next time we try to allocate space for a small file,
we will find *another* completely free 512 block chunk to allocate
small files. Given that there are 32,768 blocks in a block group,
after 64 iterations of "mount, write one 4k file in a directory,
unmount", the block group will have 64 files, each separated by 511
blocks, and the block group will no longer have any free 512
completely free chunks of blocks for group preallocation space.
So if we try to allocate blocks for a file that has been closed, such
that we know the final size of the file, and the filesystem is not
busy, avoid using group preallocation.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The logic around sbi->s_mb_last_group and sbi->s_mb_last_start was all
screwed up. These fields were getting unconditionally all the time,
set even when stream allocation had not taken place, and if they were
being used when the file was smaller than s_mb_stream_request, which
is when the allocation should _not_ be doing stream allocation.
Fix this by determining whether or not we stream allocation should
take place once, in ext4_mb_group_or_file(), and setting a flag which
gets used in ext4_mb_regular_allocator() and ext4_mb_use_best_found().
This simplifies the code and assures that we are consistently using
(or not using) the stream allocation logic.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The function ext4_mb_free_blocks() was using an "unsigned long" to
pass a block number; this will cause 64-bit block numbers to get
truncated on x86 and other 32-bit platforms.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Enhance the inode allocator to take a goal inode number as a
paremeter; if it is specified, it takes precedence over Orlov or
parent directory inode allocation algorithms.
The extents migration function uses the goal inode number so that the
extent trees allocated the migration function use the correct flex_bg.
In the future, the goal inode functionality will also be used to
allocate an adjacent inode for the extended attributes.
Also, for testing purposes the goal inode number can be specified via
/sys/fs/{dev}/inode_goal. This can be useful for testing inode
allocation beyond 2^32 blocks on very large filesystems.
Signed-off-by: Andreas Dilger <adilger@sun.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Instead of using a random number to determine the goal parent grop for
the Orlov top directories, use a hash of the directory name. This
allows for repeatable results when trying to benchmark filesystem
layout algorithms.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
We're running out of space in the mount options word, and
EXT4_MOUNT_ABORT isn't really a mount option, but a run-time flag. So
move it to become EXT4_MF_FS_ABORTED in s_mount_flags.
Also remove bogus ext2_fs.h / ext4.h simultaneous #include protection,
which can never happen.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
This field can be very helpful when a system administrator is trying
to sort through large numbers of block devices or filesystem images.
What is stored in this field can be ambiguous if multiple filesystem
namespaces are in play; what we store in practice is the mountpoint
interpreted by the process's namespace which first opens a file in the
filesystem.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
We can only fit 32 options in s_mount_opt because an unsigned long is
32-bits on a x86 machine. So use an unsigned int to save space on
64-bit platforms.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The EXT4_IOC_MOVE_EXT exchanges the blocks between orig_fd and donor_fd,
and then write the file data of orig_fd to donor_fd.
ext4_mext_move_extent() is the main fucntion of ext4 online defrag,
and this patch includes all functions related to ext4 online defrag.
Signed-off-by: Akira Fujita <a-fujita@rs.jp.nec.com>
Signed-off-by: Takashi Sato <t-sato@yk.jp.nec.com>
Signed-off-by: Kazuya Mio <k-mio@sx.jp.nec.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
This patch changes ext4 super.c to include the device name with all
warning/error messages, by using a new utility function ext4_msg.
It's a rather large patch, but very mechanic. I left debug printks
alone.
This is a straightforward port of a patch which Andi Kleen did for
ext3.
Cc: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Get rid of EXTEND_DISKSIZE flag of ext4_get_blocks_handle(). This
seems to be a relict from some old days and setting disksize in this
function does not make much sense. Currently it was set only by
ext4_getblk(). Since the parameter has some effect only if create ==
1, it is easy to check by grepping through the sources that the three
callers which end up calling ext4_getblk() with create == 1
(ext4_append, ext4_quota_write, ext4_mkdir) do the right thing and set
disksize themselves.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
To catch filesystem bugs or corruption which could lead to the
filesystem getting severly damaged, this patch adds a facility for
tracking all of the filesystem metadata blocks by contiguous regions
in a red-black tree. This allows quick searching of the tree to
locate extents which might overlap with filesystem metadata blocks.
This facility is also used by the multi-block allocator to assure that
it is not allocating blocks out of the system zone, as well as by the
routines used when reading indirect blocks and extents information
from disk to make sure their contents are valid.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The ext4_get_blocks() function was depending on the value of
bh_result->b_state as an input parameter to decide whether or not
update the delalloc accounting statistics by calling
ext4_da_update_reserve_space(). We now use a separate flag,
EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE, to requests this update, so that
all callers of ext4_get_blocks() can clear map_bh.b_state before
calling ext4_get_blocks() without worrying about any consistency
issues.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
The functions ext4_get_blocks(), ext4_ext_get_blocks(), and
ext4_ind_get_blocks() used an ad-hoc set of integer variables used as
boolean flags passed in as arguments. Use a single flags parameter
and a setandard set of bitfield flags instead. This saves space on
the call stack, and it also makes the code a bit more understandable.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>