The existing inline data support only works for cases where the entire
file is stored as inline data. For larger files, EROFS stores the
initial blocks separately and the remainder of the file ("file tail")
adjacent to the inode. Generalise inline data to allow reading the
inline file tail. Tails may not cross a page boundary in memory.
We currently have no filesystems that support tails and writing,
so that case is currently disabled (see iomap_write_begin_inline).
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
In iomap_readpage_actor, don't create iop objects for inline inodes.
Otherwise, iomap_read_inline_data will set PageUptodate without setting
iop->uptodate, and iomap_page_release will eventually complain.
To prevent this kind of bug from occurring in the future, make sure the
page doesn't have private data attached in iomap_read_inline_data.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Previously, there is no need to get the full decompressed length since
EROFS supports partial decompression. However for some other cases
such as fiemap, the full decompressed length is necessary for iomap to
make it work properly.
This patch adds a way to get the full decompressed length. Note that
it takes more metadata overhead and it'd be avoided if possible in the
performance sensitive scenario.
Link: https://lore.kernel.org/r/20210818152231.243691-1-hsiangkao@linux.alibaba.com
Reviewed-by: Chao Yu <chao@kernel.org>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Add iomap support for non-tailpacking uncompressed data in order to
support DIO and DAX.
Direct I/O is useful in certain scenarios for uncompressed files.
For example, double pagecache can be avoid by direct I/O when
loop device is used for uncompressed files containing upper layer
compressed filesystem.
This adds iomap DIO support for non-tailpacking cases first and
tail-packing inline files are handled in the follow-up patch.
Link: https://lore.kernel.org/r/20210805003601.183063-2-hsiangkao@linux.alibaba.com
Cc: linux-fsdevel@vger.kernel.org
Reviewed-by: Chao Yu <chao@kernel.org>
Signed-off-by: Huang Jianan <huangjianan@oppo.com>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
If the 1st NONHEAD lcluster of a pcluster isn't CBLKCNT lcluster type
rather than a HEAD or PLAIN type instead, which means its pclustersize
_must_ be 1 lcluster (since its uncompressed size < 2 lclusters),
as illustrated below:
HEAD HEAD / PLAIN lcluster type
____________ ____________
|_:__________|_________:__| file data (uncompressed)
. .
.____________.
|____________| pcluster data (compressed)
Such on-disk case was explained before [1] but missed to be handled
properly in the runtime implementation.
It can be observed if manually generating 1 lcluster-sized pcluster
with 2 lclusters (thus CBLKCNT doesn't exist.) Let's fix it now.
[1] https://lore.kernel.org/r/20210407043927.10623-1-xiang@kernel.org
Link: https://lore.kernel.org/r/20210510064715.29123-1-xiang@kernel.org
Fixes: cec6e93bea ("erofs: support parsing big pcluster compress indexes")
Reviewed-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Gao Xiang <xiang@kernel.org>
Prior to big pcluster, there was only one compressed page so it'd
easy to map this. However, when big pcluster is enabled, more work
needs to be done to handle multiple compressed pages. In detail,
- (maptype 0) if there is only one compressed page + no need
to copy inplace I/O, just map it directly what we did before;
- (maptype 1) if there are more compressed pages + no need to
copy inplace I/O, vmap such compressed pages instead;
- (maptype 2) if inplace I/O needs to be copied, use per-CPU
buffers for decompression then.
Another thing is how to detect inplace decompression is feasable or
not (it's still quite easy for non big pclusters), apart from the
inplace margin calculation, inplace I/O page reusing order is also
needed to be considered for each compressed page. Currently, if the
compressed page is the xth page, it shouldn't be reused as [0 ...
nrpages_out - nrpages_in + x], otherwise a full copy will be triggered.
Although there are some extra optimization ideas for this, I'd like
to make big pcluster work correctly first and obviously it can be
further optimized later since it has nothing with the on-disk format
at all.
Link: https://lore.kernel.org/r/20210407043927.10623-10-xiang@kernel.org
Acked-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Different from non-compact indexes, several lclusters are packed
as the compact form at once and an unique base blkaddr is stored for
each pack, so each lcluster index would take less space on avarage
(e.g. 2 bytes for COMPACT_2B.) btw, that is also why BIG_PCLUSTER
switch should be consistent for compact head0/1.
Prior to big pcluster, the size of all pclusters was 1 lcluster.
Therefore, when a new HEAD lcluster was scanned, blkaddr would be
bumped by 1 lcluster. However, that way doesn't work anymore for
big pcluster since we actually don't know the compressed size of
pclusters in advance (before reading CBLKCNT lcluster).
So, instead, let blkaddr of each pack be the first pcluster blkaddr
with a valid CBLKCNT, in detail,
1) if CBLKCNT starts at the pack, this first valid pcluster is
itself, e.g.
_____________________________________________________________
|_CBLKCNT0_|_NONHEAD_| .. |_HEAD_|_CBLKCNT1_| ... |_HEAD_| ...
^ = blkaddr base ^ += CBLKCNT0 ^ += CBLKCNT1
2) if CBLKCNT doesn't start at the pack, the first valid pcluster
is the next pcluster, e.g.
_________________________________________________________
| NONHEAD_| .. |_HEAD_|_CBLKCNT0_| ... |_HEAD_|_HEAD_| ...
^ = blkaddr base ^ += CBLKCNT0
^ += 1
When a CBLKCNT is found, blkaddr will be increased by CBLKCNT
lclusters, or a new HEAD is found immediately, bump blkaddr by 1
instead (see the picture above.)
Also noted if CBLKCNT is the end of the pack, instead of storing
delta1 (distance of the next HEAD lcluster) as normal NONHEADs,
it still uses the compressed block count (delta0) since delta1
can be calculated indirectly but the block count can't.
Adjust decoding logic to fit big pcluster compact indexes as well.
Link: https://lore.kernel.org/r/20210407043927.10623-9-xiang@kernel.org
Acked-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
When INCOMPAT_BIG_PCLUSTER sb feature is enabled, legacy compress indexes
will also have the same on-disk header compact indexes to keep per-file
configurations instead of leaving it zeroed.
If ADVISE_BIG_PCLUSTER is set for a file, CBLKCNT will be loaded for each
pcluster in this file by parsing 1st non-head lcluster.
Link: https://lore.kernel.org/r/20210407043927.10623-8-xiang@kernel.org
Acked-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Big pcluster indicates the size of compressed data for each physical
pcluster is no longer fixed as block size, but could be more than 1
block (more accurately, 1 logical pcluster)
When big pcluster feature is enabled for head0/1, delta0 of the 1st
non-head lcluster index will keep block count of this pcluster in
lcluster size instead of 1. Or, the compressed size of pcluster
should be 1 lcluster if pcluster has no non-head lcluster index.
Also note that BIG_PCLUSTER feature reuses COMPR_CFGS feature since
it depends on COMPR_CFGS and will be released together.
Link: https://lore.kernel.org/r/20210407043927.10623-6-xiang@kernel.org
Acked-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
When picking up inplace I/O pages, it should be traversed in reverse
order in aligned with the traversal order of file-backed online pages.
Also, index should be updated together when preloading compressed pages.
Previously, only page-sized pclustersize was supported so no problem
at all. Also rename `compressedpages' to `icpage_ptr' to reflect its
functionality.
Link: https://lore.kernel.org/r/20210407043927.10623-5-xiang@kernel.org
Acked-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Since multiple pcluster sizes could be used at once, the number of
compressed pages will become a variable factor. It's necessary to
introduce slab pools rather than a single slab cache now.
This limits the pclustersize to 1M (Z_EROFS_PCLUSTER_MAX_SIZE), and
get rid of the obsolete EROFS_FS_CLUSTER_PAGE_LIMIT, which has no
use now.
Link: https://lore.kernel.org/r/20210407043927.10623-4-xiang@kernel.org
Acked-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
To deal the with the cases which inplace decompression is infeasible
for some inplace I/O. Per-CPU buffers was introduced to get rid of page
allocation latency and thrash for low-latency decompression algorithms
such as lz4.
For the big pcluster feature, introduce multipage per-CPU buffers to
keep such inplace I/O pclusters temporarily as well but note that
per-CPU pages are just consecutive virtually.
When a new big pcluster fs is mounted, its max pclustersize will be
read and per-CPU buffers can be growed if needed. Shrinking adjustable
per-CPU buffers is more complex (because we don't know if such size
is still be used), so currently just release them all when unloading.
Link: https://lore.kernel.org/r/20210409190630.19569-1-xiang@kernel.org
Acked-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Formal big pcluster design is actually more powerful / flexable than
the previous thought whose pclustersize was fixed as power-of-2 blocks,
which was obviously inefficient and space-wasting. Instead, pclustersize
can now be set independently for each pcluster, so various pcluster
sizes can also be used together in one file if mkfs wants (for example,
according to data type and/or compression ratio).
Let's get rid of previous physical_clusterbits[] setting (also notice
that corresponding on-disk fields are still 0 for now). Therefore,
head1/2 can be used for at most 2 different algorithms in one file and
again pclustersize is now independent of these.
Link: https://lore.kernel.org/r/20210407043927.10623-2-xiang@kernel.org
Acked-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Add a bitmap for available compression algorithms and a variable-sized
on-disk table for compression options in preparation for upcoming big
pcluster and LZMA algorithm, which follows the end of super block.
To parse the compression options, the bitmap is scanned one by one.
For each available algorithm, there is data followed by 2-byte `length'
correspondingly (it's enough for most cases, or entire fs blocks should
be used.)
With such available algorithm bitmap, kernel itself can also refuse to
mount such filesystem if any unsupported compression algorithm exists.
Note that COMPR_CFGS feature will be enabled with BIG_PCLUSTER.
Link: https://lore.kernel.org/r/20210329100012.12980-1-hsiangkao@aol.com
Reviewed-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
lz4 uses LZ4_DISTANCE_MAX to record history preservation. When
using rolling decompression, a block with a higher compression
ratio will cause a larger memory allocation (up to 64k). It may
cause a large resource burden in extreme cases on devices with
small memory and a large number of concurrent IOs. So appropriately
reducing this value can improve performance.
Decreasing this value will reduce the compression ratio (except
when input_size <LZ4_DISTANCE_MAX). But considering that erofs
currently only supports 4k output, reducing this value will not
significantly reduce the compression benefits.
The maximum value of LZ4_DISTANCE_MAX defined by lz4 is 64k, and
we can only reduce this value. For the old kernel, it just can't
reduce the memory allocation during rolling decompression without
affecting the decompression result.
Link: https://lore.kernel.org/r/20210329012308.28743-3-hsiangkao@aol.com
Reviewed-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Huang Jianan <huangjianan@oppo.com>
Signed-off-by: Guo Weichao <guoweichao@oppo.com>
[ Gao Xiang: introduce struct erofs_sb_lz4_info for configurations. ]
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Add a missing case which could cause unnecessary page allocation but
not directly use inplace I/O instead, which increases runtime extra
memory footprint.
The detail is, considering an online file-backed page, the right half
of the page is chosen to be cached (e.g. the end page of a readahead
request) and some of its data doesn't exist in managed cache, so the
pcluster will be definitely kept in the submission chain. (IOWs, it
cannot be decompressed without I/O, e.g., due to the bypass queue).
Currently, DELAYEDALLOC/TRYALLOC cases can be downgraded as NOINPLACE,
and stop online pages from inplace I/O. After this patch, unneeded page
allocations won't be observed in pickup_page_for_submission() then.
Link: https://lore.kernel.org/r/20210321183227.5182-1-hsiangkao@aol.com
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Martin reported an issue that directory read could be hung on the
latest -rc kernel with some certain image. The root cause is that
commit baa2c7c971 ("block: set .bi_max_vecs as actual allocated
vector number") changes .bi_max_vecs behavior. bio->bi_max_vecs
is set as actual allocated vector number rather than the requested
number now.
Let's avoid using .bi_max_vecs completely instead.
Link: https://lore.kernel.org/r/20210306040438.8084-1-hsiangkao@aol.com
Reported-by: Martin DEVERA <devik@eaxlabs.cz>
Reviewed-by: Chao Yu <yuchao0@huawei.com>
[ Gao Xiang: note that <= 5.11 kernels are not impacted. ]
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
It's often inconvenient to use BIO_MAX_PAGES due to min() requiring the
sign to be the same. Introduce bio_max_segs() and change BIO_MAX_PAGES to
be unsigned to make it easier for the users.
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, although set_bit() & test_bit() pairs are used as a fast-
path for initialized configurations. However, these atomic ops are
actually relaxed forms. Instead, load-acquire & store-release form is
needed to make sure uninitialized fields won't be observed in advance
here (yet no such corresponding bitops so use full barriers instead.)
Link: https://lore.kernel.org/r/20210209130618.15838-1-hsiangkao@aol.com
Fixes: 62dc45979f ("staging: erofs: fix race of initializing xattrs of a inode at the same time")
Fixes: 152a333a58 ("staging: erofs: add compacted compression indexes support")
Cc: <stable@vger.kernel.org> # 5.3+
Reported-by: Huang Jianan <huangjianan@oppo.com>
Reviewed-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
Previously, we played around with magical page->mapping for short-lived
temporary pages since we need to identify different types of pages in
the same pcluster but both invalidated and short-lived temporary pages
can have page->mapping == NULL. It was considered as safe because that
temporary pages are all non-LRU / non-movable pages.
This patch tends to use specific page->private to identify short-lived
pages instead so it won't rely on page->mapping anymore. Details are
described in "compress.h" as well.
Link: https://lore.kernel.org/r/20201208095834.3133565-1-hsiangkao@redhat.com
Reviewed-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Gao Xiang <hsiangkao@redhat.com>
