Now that the code is ready to handle L2 slices we can finally add an
option to allow configuring their size.
An L2 slice is the portion of an L2 table that is read by the qcow2
cache. Until now the cache was always reading full L2 tables, and
since the L2 table size is equal to the cluster size this was not very
efficient with large clusters. Here's a more detailed explanation of
why it makes sense to have smaller cache entries in order to load L2
data:
https://lists.gnu.org/archive/html/qemu-block/2017-09/msg00635.html
This patch introduces a new command-line option to the qcow2 driver
named l2-cache-entry-size (cf. l2-cache-size). The cache entry size
has the same restrictions as the cluster size: it must be a power of
two and it has the same range of allowed values, with the additional
requirement that it must not be larger than the cluster size.
The L2 cache entry size (L2 slice size) remains equal to the cluster
size for now by default, so this feature must be explicitly enabled.
Although my tests show that 4KB slices consistently improve
performance and give the best results, let's wait and make more tests
with different cluster sizes before deciding on an optimal default.
Now that the cache entry size is not necessarily equal to the cluster
size we need to reflect that in the MIN_L2_CACHE_SIZE documentation.
That minimum value is a requirement of the COW algorithm: we need to
read two L2 slices (and not two L2 tables) in order to do COW, see
l2_allocate() for the actual code.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: c73e5611ff4a9ec5d20de68a6c289553a13d2354.1517840877.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
This function was only using the BlockDriverState parameter to pass it
to qcow2_cache_get_table_addr(). This is no longer necessary so this
parameter can be removed.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: eb0ed90affcf302e5a954bafb5931b5215483d3a.1517840877.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
This function was only using the BlockDriverState parameter to pass it
to qcow2_cache_get_table_idx() and qcow2_cache_table_release(). This
is no longer necessary so this parameter can be removed.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: 9724f7e38e763ad3be32627c6b7fe8df9edb1476.1517840877.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
This function was only using the BlockDriverState parameter to pass it
to qcow2_cache_table_release(). This is no longer necessary so this
parameter can be removed.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: b74f17591af52f201de0ea3a3b2dd0a81932334d.1517840876.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
This function was never using the BlockDriverState parameter so it can
be safely removed.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: 49c74fe8b3aead9056e61a85b145ce787d06262b.1517840876.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
This function was only using the BlockDriverState parameter to pass it
to qcow2_cache_get_table_idx(). This is no longer necessary so this
parameter can be removed.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: 6f98155489054a457563da77cdad1a66ebb3e896.1517840876.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
This function was only using the BlockDriverState parameter to pass it
to qcow2_cache_get_table_idx(). This is no longer necessary so this
parameter can be removed.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: 5c40516a91782b083c1428b7b6a41bb9e2679bfb.1517840876.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
This function was only using the BlockDriverState parameter to get the
cache table size (since it was equal to the cluster size). This is no
longer necessary so this parameter can be removed.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: 7c1b262344375d52544525f85bbbf0548d5ba575.1517840876.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
This function was only using the BlockDriverState parameter to get the
cache table size (since it was equal to the cluster size). This is no
longer necessary so this parameter can be removed.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: da3575d47c9a181a2cfd4715e53dd84a2c651017.1517840876.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
This function was only using the BlockDriverState parameter to get the
cache table size (since it was equal to the cluster size). This is no
longer necessary so this parameter can be removed.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: e1f943a9e89e1deb876f45de1bb22419ccdb6ad3.1517840876.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
The table size in the qcow2 cache is currently equal to the cluster
size. This doesn't allow us to use the cache memory efficiently,
particularly with large cluster sizes, so we need to be able to have
smaller cache tables that are independent from the cluster size. This
patch adds a new field to Qcow2Cache that we can use instead of the
cluster size.
The current table size is still being initialized to the cluster size,
so there are no semantic changes yet, but this patch will allow us to
prepare the rest of the code and simplify a few function calls.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: 67a1bf9e55f417005c567bead95a018dc34bc687.1517840876.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
@mem_size and @offset are both size_t, thus subtracting them from one
another will just return a big size_t if mem_size < offset -- even more
obvious here because the result is stored in another size_t.
Checking that result to be positive is therefore not sufficient to
exclude the case that offset > mem_size. Thus, we currently sometimes
issue an madvise() over a very large address range.
This is triggered by iotest 163, but with -m64, this does not result in
tangible problems. But with -m32, this test produces three segfaults,
all of which are fixed by this patch.
Signed-off-by: Max Reitz <mreitz@redhat.com>
Message-id: 20171114184127.24238-1-mreitz@redhat.com
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Darren Kenny <darren.kenny@oracle.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
Instead of using an assertion, it is better to emit a corruption event
here. Checking all offsets for correct alignment can be tedious and it
is easily possible to forget to do so. qcow2_cache_do_get() is a
function every L2 and refblock access has to go through, so this is a
good central point to add such a check.
And for good measure, let us also add an assertion that the offset is
non-zero. Making this a corruption event is not feasible, because a
zero offset usually means something special (such as the cluster is
unused), so all callers should be checking this anyway. If they do not,
it is their fault, hence the assertion here.
Signed-off-by: Max Reitz <mreitz@redhat.com>
Message-id: 20171110203111.7666-6-mreitz@redhat.com
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
Whenever l2/refcount table clusters are discarded from the file we can
automatically drop unnecessary content of the cache tables. This reduces
the chance of eviction useful cache data and eliminates inconsistent data
in the cache with the data in the file.
Signed-off-by: Pavel Butsykin <pbutsykin@virtuozzo.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
Message-id: 20170918124230.8152-3-pbutsykin@virtuozzo.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
We are using QEMU_MADV_DONTNEED to discard the memory of individual L2
cache tables. The problem with this is that those semantics are
specific to the Linux madvise() system call. Other implementations of
madvise() (including the very Linux implementation of posix_madvise())
don't do that, so we cannot use them for the same purpose.
This patch makes the code Linux-specific and uses madvise() directly
since there's no point in going through qemu_madvise() for this.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
qemu/osdep.h checks whether MAP_ANONYMOUS is defined, but this check
is bogus without a previous inclusion of sys/mman.h. Include it in
sysemu/os-posix.h and remove it from everywhere else.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The problem with excessive flushing was found by a couple of performance
tests:
- parallel directory tree creation (from 2 processes)
- 32 cached writes + fsync at the end in a loop
For the first one results improved from 2.6 loops/sec to 3.5 loops/sec.
Each loop creates 10^3 directories with 10 files in each.
For the second one results improved from ~600 fsync/sec to ~1100
fsync/sec. Though, it was run on SSD so it probably won't show such
performance gain on rotational media.
qcow2_cache_flush() calls bdrv_flush() unconditionally after writing
cache entries of a particular cache. This can lead to as many as
2 additional fdatasyncs inside bdrv_flush.
We can simply skip all fdatasync calls inside qcow2_co_flush_to_os
as bdrv_flush for sure will do the job. These flushes are necessary to
keep the right order of writes to the different caches. Though this is
not necessary in the current code base as this ordering is ensured through
the flush in qcow2_cache_flush_dependency().
Signed-off-by: Denis V. Lunev <den@openvz.org>
CC: Pavel Borzenkov <pborzenkov@virtuozzo.com>
CC: Kevin Wolf <kwolf@redhat.com>
CC: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Clean up includes so that osdep.h is included first and headers
which it implies are not included manually.
This commit was created with scripts/clean-includes.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This patch removes the temporary duplication between bs->file and
bs->file_child by converting everything to BdrvChild.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
BDRVQcowState is already used by qcow1, and gdb is always confused which
one to use. Rename the qcow2 one so they can be distinguished.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Alberto Garcia <berto@igalia.com>
Changing the current ordering saves 8 bytes per cache entry in x86_64.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-id: 0bd55291211df3dfb514d0e7d2031dd5c4f9f807.1438690126.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
This adds a new 'cache-clean-interval' option that cleans all qcow2
cache entries that haven't been used in a certain interval, given in
seconds.
This allows setting a large L2 cache size so it can handle scenarios
with lots of I/O and at the same time use little memory during periods
of inactivity.
This feature currently relies on MADV_DONTNEED to free that memory, so
it is not useful in systems that don't follow that behavior.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: a70d12da60433df9360ada648b3f34b8f6f354ce.1438690126.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
After having emptied the cache, the data in the cache tables is no
longer useful, so we can tell the kernel that we are done with it. In
Linux this frees the resources associated with it.
The effect of this can be seen in the HMP commit operation: it moves
data from the top to the base image (and fills both caches), then it
empties the top image. At this point the data in that cache is no
longer needed so it's just wasting memory.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: 08538b098e1faf6c92496477cf9b47a20e5aacea.1438690126.git.berto@igalia.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
The value of 'i' is guaranteed to be >= 0
Signed-off-by: Alberto Garcia <berto@igalia.com>
Message-id: 1435824371-2660-1-git-send-email-berto@igalia.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Fix pointer declaration to make it consistent with the rest of the
code.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This function never receives an invalid table pointer, so we can make
it void and remove all the error checking code.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The current cache algorithm traverses the array starting always from
the beginning, so the average number of comparisons needed to perform
a lookup is proportional to the size of the array.
By using a hash of the offset as the starting point, lookups are
faster and independent from the array size.
The hash is computed using the cluster number of the table, multiplied
by 4 to make it perform better when there are collisions.
In my tests, using a cache with 2048 entries, this reduces the average
number of comparisons per lookup from 430 to 2.5.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
A cache miss means that the whole array was traversed and the entry
we were looking for was not found, so there's no need to traverse it
again in order to select an entry to replace.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The current algorithm to evict entries from the cache gives always
preference to those in the lowest positions. As the size of the cache
increases, the chances of the later elements of being removed decrease
exponentially.
In a scenario with random I/O and lots of cache misses, entries in
positions 8 and higher are rarely (if ever) evicted. This can be seen
even with the default cache size, but with larger caches the problem
becomes more obvious.
Using an LRU algorithm makes the chances of being removed from the
cache independent from the position.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Since all tables are now stored together, it is possible to obtain
the position of a particular table directly from its address, so the
operation becomes O(1).
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The qcow2 L2/refcount cache contains one separate table for each cache
entry. Doing one allocation per table adds unnecessary overhead and it
also requires us to store the address of each table separately.
Since the size of the cache is constant during its lifetime, it's
better to have an array that contains all the tables using one single
allocation.
In my tests measuring freshly created caches with sizes 128MB (L2) and
32MB (refcount) this uses around 10MB of RAM less.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The current algorithm to replace entries from the L2 cache gives
priority to newer hits by dividing the hit count of all existing
entries by two everytime there is a cache miss.
However, if there are several cache misses the hit count of the
existing entries can easily go down to 0. This will result in those
entries being replaced even when there are others that have never been
used.
This problem is more noticeable with larger disk images and cache
sizes, since the chances of having several misses before the cache is
full are higher.
If we make sure that the hit count can never go down to 0 again,
unused entries will always have priority.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
With a variable cache size, the number given to qcow2_cache_create() may
be huge. Therefore, use g_try_new0().
While at it, use g_new0() instead of g_malloc0() for allocating the
Qcow2Cache object.
Signed-off-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
g_new(T, n) is safer than g_malloc(sizeof(*v) * n) for two reasons.
One, it catches multiplication overflowing size_t. Two, it returns
T * rather than void *, which lets the compiler catch more type
errors.
Perhaps a conversion to g_malloc_n() would be neater in places, but
that's merely four years old, and we can't use such newfangled stuff.
This commit only touches allocations with size arguments of the form
sizeof(T), plus two that use 4 instead of sizeof(uint32_t). We can
make the others safe by converting to g_malloc_n() when it becomes
available to us in a couple of years.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Jeff Cody <jcody@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Some code in the block layer makes potentially huge allocations. Failure
is not completely unexpected there, so avoid aborting qemu and handle
out-of-memory situations gracefully.
This patch addresses the allocations in the qcow2 block driver.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
In qcow2_check_metadata_overlap and qcow2_pre_write_overlap_check,
change the parameter signifying the checks to perform from its current
positive form to a negative one, i.e., it will no longer explicitly
specify every check to perform but rather a mask of checks not to
perform.
Signed-off-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Add a function for emptying a cache, i.e., flushing it and marking all
elements invalid.
Signed-off-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The pre-write overlap check function is now called before most of the
qcow2 writes (aborting it on collision or other error).
Signed-off-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Writethrough does not need special-casing anymore in the qcow2 caches.
The block layer adds flushes after every guest-initiated data write,
and these will also flush the qcow2 caches to the OS.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
In snapshotting there is no guest involved, so we can safely use a writeback
mode and do the flushes in the right place (i.e. at the very end). This
improves the time that creating/restoring an internal snapshot takes with an
image in writethrough mode.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Move size after the two pointers in struct Qcow2Cache to get better
packing of struct elements on 64 bit architectures.
Signed-off-by: Jes Sorensen <Jes.Sorensen@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
qcow2 calls bdrv_flush() after performing COW in order to ensure that the
L2 table change is never written before the copy is safe on disk. Now that the
L2 table is cached, we can wait with flushing until we write out the next L2
table.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This adds some new cache functions to qcow2 which can be used for caching
refcount blocks and L2 tables. When used with cache=writethrough they work
like the old caching code which is spread all over qcow2, so for this case we
have merely a cleanup.
The interesting case is with writeback caching (this includes cache=none) where
data isn't written to disk immediately but only kept in cache initially. This
leads to some form of metadata write batching which avoids the current "write
to refcount block, flush, write to L2 table" pattern for each single request
when a lot of cluster allocations happen. Instead, cache entries are only
written out if its required to maintain the right order. In the pure cluster
allocation case this means that all metadata updates for requests are done in
memory initially and on sync, first the refcount blocks are written to disk,
then fsync, then L2 tables.
This improves performance of scenarios with lots of cluster allocations
noticably (e.g. installation or after taking a snapshot).
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Alberto Garcia