the current value of 1024 bytes (16 * MFI_FRAME_SIZE) we map is not enough to hold
the maximum number of scatter gather elements we advertise. We actually need a
maximum of 2048 bytes. This is 128 max sg elements * 16 bytes (sizeof (union mfi_sgl)).
Cc: qemu-stable@nongnu.org
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <20190404121015.28634-1-pl@kamp.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
the min_sparse convert parameter can overflow (e.g. -S 1024G)
in the conversion from int64_t to int resulting in a negative
min_sparse parameter. Avoid this by limiting the valid parameters
to sane values. In fact anything exceeding the convert buffer size
is also pointless. While at it also forbid values that are non
multiple of 512 to avoid undesired behaviour. For instance, values
between 1 and 511 were legal, but resulted in full allocation.
Cc: qemu-stable@nongnu.org
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
We currently don't enforce that the sparse segments we detect during convert are
aligned. This leads to unnecessary and costly read-modify-write cycles either
internally in Qemu or in the background on the storage device as nearly all
modern filesystems or hardware have a 4k alignment internally.
This patch modifies is_allocated_sectors so that its *pnum result will always
end at an alignment boundary. This way all requests will end at an alignment
boundary. The start of all requests will also be aligned as long as the results
of get_block_status do not lead to an unaligned offset.
The number of RMW cycles when converting an example image [1] to a raw device that
has 4k sector size is about 4600 4k read requests to perform a total of about 15000
write requests. With this path the additional 4600 read requests are eliminated while
the number of total write requests stays constant.
[1] https://cloud-images.ubuntu.com/releases/16.04/release/ubuntu-16.04-server-cloudimg-amd64-disk1.vmdk
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
the current implementation submits up to 512 I/O requests in parallel
which is much to high especially for a background task.
This patch adds a maximum limit of 16 I/O requests that can
be submitted in parallel to avoid monopolizing the I/O device.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <1520507908-16743-5-git-send-email-pl@kamp.de>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
this actually limits (as the original commit mesage suggests) the
number of I/O buffers that can be allocated and not the number
of parallel (inflight) I/O requests.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <1520507908-16743-4-git-send-email-pl@kamp.de>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
this patch makes the bulk phase of a block migration to take
place before we start transferring ram. As the bulk block migration
can take a long time its pointless to transfer ram during that phase.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <1520507908-16743-2-git-send-email-pl@kamp.de>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
we currently report an "iSCSI Failure" in iscsi_co_generic_cb if the task
hasn't completed with SCSI_STATUS_GOOD. However, we expect a failure in
some cases and handle it gracefully. This is the case for misaligned UNMAPs
and WRITESAME10/16 calls without UNMAP. In this case a failure in the
logs can be quite misleading.
While we are at it improve the logging to reveal which operation failed
at what LBA.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <1512733868-9009-3-git-send-email-pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
auto-converge and block migration currently do not play well together.
During block migration the auto-converge logic detects that ram
migration makes no progress and thus throttles down the vm until
it nearly stalls completely. Avoid this by disabling the throttling
logic during the bulk phase of the block migration.
Cc: qemu-stable@nongnu.org
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <1506421996-12513-1-git-send-email-pl@kamp.de>
Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
img_convert has been around before there was an ImgConvertState or
a block backend, but it has never been modified to directly use
these structs. Change this by parsing parameters directly into
the ImgConvertState and directly use BlockBackend where possible.
Furthermore variable initialization has been reworked and sorted.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
commit 3c80ca15 fixed a deadlock scenarion with nested aio_poll invocations.
However, the rescheduling of the completion BH introcuded unnecessary spinning
in the main-loop. On very fast file backends this can even lead to the
"WARNING: I/O thread spun for 1000 iterations" message popping up.
Callgrind reports about 3-4% less instructions with this patch running
qemu-img bench on a ramdisk based VMDK file.
Fixes: 3c80ca158c
Cc: qemu-stable@nongnu.org
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
the current implementation fails if we try to freeze an
already frozen filesystem. This can happen if a filesystem
is mounted more than once (e.g. with a bind mount).
Suggested-by: Christian Theune <ct@flyingcircus.io>
Cc: qemu-stable@nongnu.org
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Michael Roth <mdroth@linux.vnet.ibm.com>
the convert process is currently completely implemented with sync operations.
That means it reads one buffer and then writes it. No parallelism and each sync
request takes as long as it takes until it is completed.
This can be a big performance hit when the convert process reads and writes
to devices which do not benefit from kernel readahead or pagecache.
In our environment we heavily have the following two use cases when using
qemu-img convert.
a) reading from NFS and writing to iSCSI for deploying templates
b) reading from iSCSI and writing to NFS for backups
In both processes we use libiscsi and libnfs so we have no kernel cache.
This patch changes the convert process to work with parallel running coroutines
which can significantly improve performance for network storage devices:
qemu-img (master)
nfs -> iscsi 22.8 secs
nfs -> ram 11.7 secs
ram -> iscsi 12.3 secs
qemu-img-async (8 coroutines, in-order write disabled)
nfs -> iscsi 11.0 secs
nfs -> ram 10.4 secs
ram -> iscsi 9.0 secs
This patches introduces 2 new cmdline parameters. The -m parameter to specify
the number of coroutines running in parallel (defaults to 8). And the -W parameter to
allow qemu-img to write to the target out of order rather than sequential. This improves
performance as the writes do not have to wait for each other to complete.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
commit 94d6a7a accidentally left the naming of runtime opts and QAPI
scheme inconsistent. As one consequence passing of parameters in the
URI is broken. Sync the naming of the runtime opts to the QAPI
scheme.
Please note that this is technically backwards incompatible with the 2.8
release, but the 2.8 release is the only version that had the wrong naming.
Furthermore release 2.8 suffered from a NULL pointer dereference during
URI parsing.
Fixes: 94d6a7a76e
Cc: qemu-stable@nongnu.org
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-id: 1485942829-10756-3-git-send-email-pl@kamp.de
[mreitz: Fixed commit message]
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
nb_cls_shrunk in iscsi_allocmap_update can become -1 if the
request starts and ends within the same cluster. This results
in passing -1 to bitmap_set and bitmap_clear and they don't
handle negative values properly. In the end this leads to data
corruption.
Fixes: e1123a3b40
Cc: qemu-stable@nongnu.org
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <1484579832-18589-1-git-send-email-pl@kamp.de>
Reviewed-by: Fam Zheng <famz@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
this adds a knob to track the maximum stack usage of stacks
created by qemu_alloc_stack.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
the allocated stack will be adjusted to the minimum supported stack size
by the OS and rounded up to be a multiple of the system pagesize.
Additionally an architecture dependent guard page is added to the stack
to catch stack overflows.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
we only need to allocate enough memory to hold the packet. This might be
less than NET_BUFSIZE. Additionally fail early if the packet is larger
than NET_BUFSIZE.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
until now the allocation map was used only as a hint if a cluster
is allocated or not. If a block was not allocated (or Qemu had
no info about the allocation status) a get_block_status call was
issued to check the allocation status and possibly avoid
a subsequent read of unallocated sectors. If a block known to be
allocated the get_block_status call was omitted. In the other case
a get_block_status call was issued before every read to avoid
the necessity for a consistent allocation map. To avoid the
potential overhead of calling get_block_status for each and
every read request this took only place for the bigger requests.
This patch enhances this mechanism to cache the allocation
status and avoid calling get_block_status for blocks where
the allocation status has been queried before. This allows
for bypassing the read request even for smaller requests and
additionally omits calling get_block_status for known to be
unallocated blocks.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <1468831940-15556-3-git-send-email-pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
when setting clusters as alloacted the boundaries have
to be expanded. As Paolo pointed out the calculation of
the number of clusters is wrong:
Suppose cluster_sectors is 2, sector_num = 1, nb_sectors = 6:
In the "mark allocated" case, you want to set 0..8, i.e.
cluster_num=0, nb_clusters=4.
0--.--2--.--4--.--6--.--8
<--|_________________|--> (<--> = expanded)
Instead you are setting nb_clusters=3, so that 6..8 is not marked.
0--.--2--.--4--.--6--.--8
<--|______________|!!! (! = wrong)
Cc: qemu-stable@nongnu.org
Reported-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <1468831940-15556-2-git-send-email-pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
this is the first step in reducing the brk heap fragmentation
created by the map->nodes memory allocation. Since the introduction
of RCU the freeing of the PhysPageMaps is delayed so that sometimes
several hundred are allocated at the same time.
Even worse the memory for map->nodes is allocated and shortly
afterwards reallocated. Since the number of nodes it grows
to in the end is the same for all PhysPageMaps remember this value
and at least avoid the reallocation.
The large number of simultaneous allocations (about 450 x 70kB in
my configuration) has to be addressed later.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <1468577030-21097-1-git-send-email-pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
commit 095497ff added thread local storage for the color counting
palette. Unfortunately, a VncPalette is about 7kB on a x86_64 system.
This memory is reserved from the stack of every thread and it
exhausted the stack space of a libxenstore thread.
Fix this by allocating memory only for the VNC encoding thread.
Fixes: 095497ffc6
Reported-by: Juergen Gross <jgross@suse.com>
Tested-by: Juergen Gross <jgross@suse.com>
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-id: 1468575911-20656-1-git-send-email-pl@kamp.de
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
currently the color counting palette is allocated from heap, used and destroyed
for each single subrect. Use a static palette per thread for this purpose and
avoid the malloc and free for each update.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Message-id: 1467280846-9674-1-git-send-email-pl@kamp.de
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Commit 94d047a added an assertion the the request alignment check.
This introduced 2 issues:
a) A off-by-one error since a request of BDRV_REQUEST_MAX_SECTORS
is actually allowed.
b) The bdrv_get_block_status call in the read path to check the allocation
status requests up to INT_MAX sectors which triggers the assertion.
Fixes: 94d047a35b
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <1466414680-18383-1-git-send-email-pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
upcoming libnfs will have support for a read cache that can
significantly help to speed up requests since libnfs by design
circumvents the kernel cache.
Example:
qemu -cdrom nfs://127.0.0.1/iso/my.iso?pagecache=1024
The pagecache parameters takes the maximum amount of pages to
cache. A page in libnfs is always the NFS_BLKSIZE which is
4KB.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Jeff Cody <jcody@redhat.com>
Message-id: 1463662083-20814-3-git-send-email-pl@kamp.de
Signed-off-by: Jeff Cody <jcody@redhat.com>
if we open a NFS export with disabled cache we should refuse
the readahead feature as it will cache data inside libnfs.
If a export was opened with readahead enabled it should
futher not be allowed to disable the cache while running.
Cc: qemu-stable@nongnu.org
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Jeff Cody <jcody@redhat.com>
Message-id: 1463662083-20814-2-git-send-email-pl@kamp.de
Signed-off-by: Jeff Cody <jcody@redhat.com>
commit fefe2a78 accidently dropped the code path for injecting
raw packets. This feature is needed for sending gratuitous ARPs
after an incoming migration has completed. The result is increased
network downtime for vservers where the network card is not virtio-net
with the VIRTIO_NET_F_GUEST_ANNOUNCE feature.
Fixes: fefe2a78ab
Cc: qemu-stable@nongnu.org
Cc: hongyang.yang@easystack.cn
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Jason Wang <jasowang@redhat.com>
at least bdrv_co_preadv/pwritev expect this.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
in a read-modify-write cycle a small request might cause
head and tail to fall into the same aligned block. Currently
QEMU reads the same block twice in this case which is
not necessary.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-id: 1464607873-28206-1-git-send-email-pl@kamp.de
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
recent libnfs versions support logging debug messages. Add
support for it in qemu through an URL parameter.
Example:
qemu -cdrom nfs://127.0.0.1/iso/my.iso?debug=2
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Fam Zheng <famz@redhat.com>
Message-id: 1447052973-14513-1-git-send-email-pl@kamp.de
Signed-off-by: Jeff Cody <jcody@redhat.com>
The check for the cleared BSY flag has to be performed
before each data transfer and not just before the
first one.
Commit 5f81724d revealed this glitch as the BSY flag
was not set in ATAPI PIO transfers before.
While at it fix the descriptions and add a comment before
the nested for loop that transfers the data.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-id: 1448029742-19771-1-git-send-email-pl@kamp.de
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
If the guests canceles a DMA request we can prematurely
invoke all callbacks of buffered requests and flag all them
as orphaned. Ideally this avoids the need for draining all
requests. For CDROM devices this works in 100% of all cases.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Fam Zheng <famz@redhat.com>
Message-id: 1447345846-15624-5-git-send-email-pl@kamp.de
Signed-off-by: John Snow <jsnow@redhat.com>
this patch adds a new aio readv compatible function which copies
all data through a bounce buffer. These buffered requests can be
flagged as orphaned which means that their original callback has
already been invoked and the request has just not been completed
by the backend storage. The bounce buffer guarantees that guest
memory corruption is avoided when such a orphaned request is
completed by the backend at a later stage.
This trick only works for read requests as a write request completed
at a later stage might corrupt data as there is no way to control
if and what data has already been written to the storage.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Fam Zheng <famz@redhat.com>
Message-id: 1447345846-15624-4-git-send-email-pl@kamp.de
Signed-off-by: John Snow <jsnow@redhat.com>
PIO read requests on the ATAPI interface used to be sync blk requests.
This has two significant drawbacks. First the main loop hangs util an
I/O request is completed and secondly if the I/O request does not
complete (e.g. due to an unresponsive storage) Qemu hangs completely.
Note: Due to possible race conditions requests during an ongoing
elementary transfer are still sync.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: John Snow <jsnow@redhat.com>
Message-id: 1447345846-15624-2-git-send-email-pl@kamp.de
Signed-off-by: John Snow <jsnow@redhat.com>
the idea behind this patch is to allow the buffer to shrink, but
make this a seldom operation. The buffers average size is measured
exponentionally smoothed with am alpha of 1/128.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
Message-id: 1446203414-4013-20-git-send-email-kraxel@redhat.com
If the vs->output buffer is empty it will be dropped
by the next qio_buffer_move_empty in vnc_jobs_consume_buffer
anyway. So reuse the allocated buffer from this buffer
in the worker thread where we otherwise would start with
an empty (unallocated buffer).
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
Message-id: 1446203414-4013-17-git-send-email-kraxel@redhat.com
[ added a comment describing the non-obvious optimization ]
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
I have observed that depending on the contents and the encoding it happens
that sending data as RAW sometimes would take less space than the encoded data.
This is especially the case for small updates or areas with high color images.
If sending RAW encoded data is beneficial allow a fall back to RAW encoding
for the framebuffer update.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
If the file is readonly its not expected to grow so
save the blocking call to nfs_fstat_async and use
the value saved at connection time. Also important
the monitor (and thus the main loop) will not hang
if block device info is queried and the NFS share
is unresponsive.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Jeff Cody <jcody@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: 1440671441-7978-1-git-send-email-pl@kamp.de
Signed-off-by: Jeff Cody <jcody@redhat.com>
It has been reported that at least tgtd returns a block size of 0
for LUN 0. To avoid running into divide by zero later on and protect
against other problematic block sizes validate the block size right
at connection time.
Cc: qemu-stable@nongnu.org
Reported-by: Andrey Korolyov <andrey@xdel.ru>
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-Id: <1439552016-8557-1-git-send-email-pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
libiscsi starting with 1.15 will properly support timeout of iscsi
commands. The default will remain no timeout, but this can
be changed via cmdline parameters, e.g.:
qemu -iscsi timeout=30 -drive file=iscsi://...
If a timeout occurs a reconnect is scheduled and the timed out command
will be requeued for processing after a successful reconnect.
The required API call iscsi_set_timeout is present since libiscsi
1.10 which was released in October 2013. However, due to some bugs
in the libiscsi code the use is not recommended before version 1.15.
Please note that this patch bumps the libiscsi requirement to 1.10
to have all function and macros defined. The patch fixes also a
off-by-one error in the NOP timeout calculation which was fixed
while touching these code parts.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-id: 1434455107-19328-1-git-send-email-pl@kamp.de
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
the allocationmap has only a hint character. The driver always
double checks that blocks marked unallocated in the cache are
still unallocated before taking the fast path and return zeroes.
So using the allocationmap is migration safe and can
also be enabled with cache.direct=on.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-id: 1429193313-4263-10-git-send-email-pl@kamp.de
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The idea is that a command is retried in a BUSY condition
up a time of approx. 60 seconds before it is failed. This should
be far higher than any command timeout in the guest.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-id: 1429193313-4263-7-git-send-email-pl@kamp.de
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
SCSI allowes to tell the target to not return from a write command
if the date is not written to the disk. Use this so called FUA
bit if it is supported to optimize WRITE commands if writeback is
not allowed.
In this case qemu always issues a WRITE followed by a FLUSH. This
is 2 round trip times. If we set the FUA bit we can ignore the
following FLUSH.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-id: 1429193313-4263-6-git-send-email-pl@kamp.de
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
newer libiscsi versions may return zero events from iscsi_which_events.
In this case iscsi_service will return immediately without any progress.
To avoid busy waiting for iscsi_which_events to change we deregister all
read and write handlers in this case and schedule a timer to periodically
check iscsi_which_events for changed events.
Next libiscsi version will introduce async reconnects and zero events
are returned while libiscsi is waiting for a reconnect retry.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-id: 1428437295-29577-1-git-send-email-pl@kamp.de
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The code to check the bitmap for the allocation status of each sector
has been "disabled by reason" ever since the vpc driver existed.
The reason might be that we might end up reading sector by sector
in vpc_read if we really used it. This would be a performance desaster.
The current code would furthermore not work if the disabled parts get
reactivated since vpc_read and vpc_write only use get_sector_offset to
check the allocation status of the first sector of a read/write operation.
This might lead to sectors incorrectly treated as zero in vpc_read and
to sectors getting allocated twice in vpc_write.
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-id: 1425379316-19639-6-git-send-email-pl@kamp.de
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
The VHD spec [1] allows for total_sectors of 65535 x 16 x 255 (~127GB)
represented by a CHS geometry. If total_sectors is greater
than 65535 x 16 x 255 this geometry is set as a maximum.
Qemu, Hyper-V and disk2vhd use this special geometry as an indicator
to use the image current size from the footer as disk size.
This patch changes vpc_create to effectively calculate a CxHxS geometry
for the given image size if possible while rounding up if necessary.
If the image size is too big to be represented in CHS we set the maximum
and write the exact requested image size into the footer.
This partly reverts commit 258d2edb, but leaves support for >127G disks
intact.
[1] http://download.microsoft.com/download/f/f/e/ffef50a5-07dd-4cf8-aaa3-442c0673a029/Virtual%20Hard%20Disk%20Format%20Spec_10_18_06.doc
Signed-off-by: Peter Lieven <pl@kamp.de>
Message-id: 1425379316-19639-4-git-send-email-pl@kamp.de
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
in case we send and update for a complete scanline increment
the y offset to avoid running to find_next_bit for that lines
twice.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
we check and adjust request sizes at several places with
sometimes inconsistent checks or default values:
INT_MAX
INT_MAX >> BDRV_SECTOR_BITS
UINT_MAX >> BDRV_SECTOR_BITS
SIZE_MAX >> BDRV_SECTOR_BITS
This patches introdocues a macro for the maximal allowed sectors
per request and uses it at several places.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Denis V. Lunev <den@openvz.org>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
this adds a knob to disable request merging for debugging or benchmarks if dedired.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
this patch finally introduces multiread support to virtio-blk. While
multiwrite support was there for a long time, read support was missing.
The complete merge logic is moved into virtio-blk.c which has
been the only user of request merging ever since. This is required
to be able to merge chunks of requests and immediately invoke callbacks
for those requests. Secondly, this is required to switch to
direct invocation of coroutines which is planned at a later stage.
The following benchmarks show the performance of running fio with
4 worker threads on a local ram disk. The numbers show the average
of 10 test runs after 1 run as warmup phase.
| 4k | 64k | 4k
MB/s | rd seq | rd rand | rd seq | rd rand | wr seq | wr rand
--------------+--------+---------+--------+---------+--------+--------
master | 1221 | 1187 | 4178 | 4114 | 1745 | 1213
multiread | 1829 | 1189 | 4639 | 4110 | 1894 | 1216
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
As it was not obvious (at least for me) where the 32 comes from;
add a constant for it.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
do not trim requests if the driver does not supply a limit
through BlockLimits. For write zeroes we still keep a limit
for the unsupported path to avoid allocating a big bounce buffer.
Suggested-by: Kevin Wolf <kwolf@redhat.com>
Suggested-by: Denis V. Lunev <den@openvz.org>
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Some ancient Linux kernels read from registers 0x09 and 0x3c-3f during
boot. According to the spec these registers are for diag and debug
purposes only. If they are absend qemu aborts on read.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If bs->bl.max_write_zeroes is large and we end up in the unsupported
path we might allocate a lot of memory for the iovector and/or even
generate an oversized requests.
Fix this by limiting the request by the minimum of the reported
maximum transfer size or 16MB (32768 sectors).
Reported-by: Denis V. Lunev <den@openvz.org>
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Denis V. Lunev <den@openvz.org>
Message-id: 1420457389-16332-1-git-send-email-pl@kamp.de
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Placing coroutines on the global pool should be preferrable, because it
can help all threads. But if the global pool is full, we can still
try to save some allocations by stashing completed coroutines on the
local pool. This is quite cheap too, because it does not require
atomic operations, and provides a gain of 15% in the best case.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Message-id: 1417518350-6167-8-git-send-email-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Cancel oversized requests early. They would generate
an iSCSI protocol error anyway; after having transferred
possibly a lot of data over the wire.
Suggested-By: Max Reitz <mreitz@redhat.com>
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
As Max pointed out there is a hidden cast from int64_t to int for all
limits. So use the newly introduced sector_limits_lun2qemu for all
limits received from the target.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Copy the max_xfer_len from the BlockLimits VPD or use the
maximum value fitting in the CDB.
The helper function sector_limits_lun2qemu is introduced to convert
and cap the limits from the VPD to the maximum power of two fitting
in an integer; integer is the range for nb_sectors throughout
the block layer.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
at least in block layer we have the case of limits being defined for a
BlockDriverState. However, in this context often zero (0) has the special
meanining of undefined which means no limit. If two of those limits are
combined and the minimum is needed the minimum function should only return
zero if both parameters are zero.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
In preparation to possible automatic regression and performance
testing for the block layer I found that the iotests don't work
for all protocols anymore.
In commit 1f7bf7d0 I started to change supported protocols from
generic to file for various tests. Unfortunately, some tests
added in the meantime again carry generic protocol altough they
can only work with file because they require local file access.
The other way around for some tests that only support file I added
NFS protocol after confirming they work.
Signed-off-by: Peter Lieven <pl@kamp.de>
Reviewed-by: Benoît Canet <benoit.canet@nodalink.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
this patch extends commit db80fac by not only checking
for unknown flags, but also filtering out unknown flag
combinations.
Suggested-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Juan Quintela <quintela@redhat.com>
regular bitmap_new simply aborts if the memory allocation fails.
bitmap_try_new returns NULL on failure and allows for proper
error handling.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
we currently have the Nagle algorithm enabled for all outgoing VNC updates.
This may delay sensitive updates as mouse movements or typing in the console.
As we currently prepare all data in a buffer and then send as much as we can
disabling the Nagle algorithm should not cause big trouble. Well established
VNC servers like TightVNC set TCP_NODELAY as well.
A regular framebuffer update request generates exactly one framebuffer update
which should be pushed out as fast as possible.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Peter Lieven