Commit 0d8d769085 introduced
a regression in virtio-net performance because it looks
into the ring aggressively while we really only care
about a single packet worth of buffers.
Reported as bugzilla 1066055 in launchpad.
To fix, add parameters limiting lookahead, and
use in virtqueue_avail_bytes.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reported-by: Edivaldo de Araujo Pereira <edivaldoapereira@yahoo.com.br>
Tested-by: Edivaldo de Araujo Pereira <edivaldoapereira@yahoo.com.br>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Debug option is available since QEMU-1.2; FP coprocessor and
coprocessor context is available since QEMU-1.3.
Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
We are currently checking for an exact type match. Use QOM dynamic_cast to
check for a compatible type instead.
Cc: Konrad Frederic <fred.konrad@greensocs.com>
Cc: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
---
v1 -> v2:
- also add cast to qbus_find_recursive (Peter)
- simplify by doing object_dynamic_cast instead of messing with classes
The event notifier state is only reset by test_and_clear. But we created
the windows event object with auto-reset, which subtly swallowed events.
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
* bonzini/nbd-next:
nbd-server-add: Fix the default for 'writable'
nbd: fix use of two uninitialized bytes when connecting to a named export
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
* bonzini/scsi-next:
virtio-scsi: Fix subtle (guest) endian bug
virtio-scsi: Fix some endian bugs with virtio-scsi
iscsi: do not assume device is zero initialized
iscsi: fix deadlock during login
iscsi: fix segfault in url parsing
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
* agraf/s390-for-upstream-1.3:
sclp: Fix uninitialized var in handle_write_event_buf().
s390: Fix ram_size updating in machine init
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
* afaerber/qom-cpu-1.3:
target-i386: Enable SSSE3 TCG support
target-i386/cpu: Add missing flags to Haswell CPU model
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
These versions of GCC require insane (>2GB) amounts of memory to compile
translate.o. As a countermeasure, disable the culprit optimization pass.
This should fix the buildbot failure for default_x86_64_fedora16. Anyway
this is a good thing to do because people will try to compile 1.3 with
less than 2GB of memory and complain.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
The virtio-scsi config space is, by specification, in guest endian (which
is ill-defined, but there you go). In virtio_scsi_get_config() we set up
all the fields in there, using stl_raw(). Which is a problem for the
max_channel and max_target fields, which are 16-bit, not 32-bit. For
little-endian targets we get away with it by accident, since the first
two bytes will still be correct, and the extra two bytes written (with
zeroes) will be overwritten correctly by the next store.
But for big-endian guests, this means the max_target field ends up as zero,
which means the guest will only recognize a single disk on the virtio-scsi
bus. This patch fixes the problem.
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Paul 'Rusty' Russell <rusty@rustcorp.com.au>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The virtio-scsi specification does not specify the correct endianness for
fields in the request structure. It's therefore best to assume that it is
"guest native" endian since that's the (stupid and poorly defined) norm in
virtio.
However, the qemu device for virtio-scsi has no byteswaps at all, and so
will break if the guest has different endianness from the host. This patch
fixes it by adding tswap() calls for the sense_len and resid fields in
the request structure. In theory status_qualifier needs swaps as well,
but that field is never actually touched. The tag field is a uint64_t, but
since its value is completely arbitrary, it might as well be uint8_t[8]
and so it does not need swapping.
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Paul 'Rusty' Russell <rusty@rustcorp.com.au>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Without any complex checks we can't assume that an
iscsi target is initialized to zero.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If the connection is interrupted before the first login is successfully
completed qemu-kvm is waiting forever in qemu_aio_wait().
This is fixed by performing an sync login to the target. If the
connection breaks after the first successful login errors are
handled internally by libiscsi.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If an invalid URL is specified iscsi_get_error(iscsi) is called
with iscsi == NULL.
Signed-off-by: Peter Lieven <pl@kamp.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The documentation to this monitor command tells, that 'writable'
argument is optional and defaults to false. However, the code sets
true as the default. But since some applications may already been
using this, it's safer to fix the code and not documentation which
would break those applications.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This bug occurs when the SET flag of Register B is enabled. When an RTC
data register (i.e. any of the ten time/calender CMOS bytes) is set, the
data is (as expected) correctly stored in the cmos_data array. However,
since the SET flag is enabled, the function rtc_set_time is not invoked.
As a result, the field base_rtc in RTCState remains uninitialized. This
causes a problem on subsequent writes which can end up overwriting data.
To see this, consider writing data to Register A after having written
data to any of the RTC data registers; the following figure illustrates
the call stack for the Register A write operation:
+- cmos_io_port_write
+-- check_update_timer
+---- get_next_alarm
+------ rtc_update_time
In rtc_update_time, get_guest_rtc calculates the wrong time and
overwrites the previously written RTC data register values.
Signed-off-by: Alex Horn <alex.horn@cs.ox.ac.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
g_assert_cmpint is not available on glib 2.12, which is the minimum
version required to build QEMU (we only require 2.16 to run tests,
since that is the first version including GTester). Do not use it
in hardware models, use a normal assertion instead.
This fixes the buildbot failure for default_x86_64_rhel5.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
The cancellation test is failing on the buildbots. While the failure
merits a little more investigation to understand what is going on,
the logs show that the failure is not impacting the coverage
provided by the test. Hence, loosen a bit the assertions in a
way that should let the test proceed and hopefully pass.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Fix typos, whitespace and update comments to match current
implementation.
Signed-off-by: Amit Shah <amit.shah@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Disable the rate-limit timer on device remove (e.g. hot-unplug).
Signed-off-by: Amit Shah <amit.shah@redhat.com>
Reviewed-by: Anthony Liguori <aliguori@us.ibm.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
If we got fewer bytes from the backend than requested, don't poke the
backend for more bytes; the guest will ask for more (or if the guest has
already asked for more, the backend knows about it via handle_input()).
Signed-off-by: Amit Shah <amit.shah@redhat.com>
Reviewed-by: Anthony Liguori <aliguori@us.ibm.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Popping an elem from the vq just to find out its length causes problems
with save/load later on. Use the new virtqueue_get_avail_bytes()
function instead, saves us the complexity in the migration code, as well
as makes the migration endian-safe.
Signed-off-by: Amit Shah <amit.shah@redhat.com>
Reviewed-by: Anthony Liguori <aliguori@us.ibm.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
It used a wrong struct type name since its introduction in
8f04ee0882 (isa: pic: convert to QEMU
Object Model), apparently it is unused so far.
Signed-off-by: Andreas Färber <afaerber@suse.de>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
visit_type_size() requires either visitor->type_size() or
visitor_uint64() to be implemented, otherwise a NULL function pointer is
invoked.
It is possible to trigger this crash as follows:
$ qemu-system-x86_64 -netdev tap,sndbuf=0,id=netdev0 \
-device virtio-blk-pci,netdev=netdev0
The 'sndbuf' option has type "size".
Reviewed-by: Andreas Färber <afaerber@suse.de>
Reviewed-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
When adding the Haswell CPU model, I intended to make it a superset of the
features present on the SandyBridge model, but I have removed the SEP and
RDTSCP features from the feature list by mistake. This patch adds the missing
SEP and RDTSCP features (that are present on SandyBridge) to Haswell.
Reported-by: Martin Kletzander <mkletzan@redhat.com>
Acked-by: Andreas Färber <afaerber@suse.de>
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Commit 094b287f0b accidentally broke the "-usb" command line
option, so it would have no effect if the user had not specified
any machine options at that point. (the return value from
'qemu_opts_find(qemu_find_opts("machine"), 0);' is NULL if there
are no user specified options, so it is only to be used for
looking up an option, not when trying to set one.) Similarly,
would '-usbdevice' no longer cause USB to default to enabled.
Fix this regression by using the same style of code for forcing
the usb=on machine option that we use for other aliases such as
'-enable-kvm'.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
It is not used anymore, and there is no need to make it public.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
All conditional deallocation can now be done with object_delete.
Remove the @qom_allocated and @glib_allocated fields; replace the latter
with a direct assignment of the @free function pointer.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Store in the object the freeing function that will be used at deletion
time. This makes it possible to use object_delete on statically-allocated
(embedded) objects. Dually, it makes it possible to use object_unparent
and object_unref without leaking memory, when the lifetime of object
might extend until after the call to object_delete.
Reviewed-by: Andreas Färber <afaerber@suse.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Add an ObjectClass method that is done at object_unparent time. It
should remove any backlinks to the object in the composition tree,
so that object_delete will be able to drop the last reference and
free the object.
Use it for qdev buses.
Reviewed-by: Andreas Färber <afaerber@suse.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
The reference count for embedded objects is always one too low, because
object_initialize_with_type returns with zero references to the object.
This causes premature finalization of the object (or an assertion failure)
after calling object_ref to add an extra reference and object_unref to
remove it.
The fix is to move the initial object_ref call from object_new_with_type
to object_initialize_with_type.
Acked-by: Andreas Färber <afaerber@suse.de>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Some gcc versions rightly complain about a possibly unitialized rc,
so let's move setting it before the QTAILQ_FOREACH().
Signed-off-by: Cornelia Huck <cornelia.huck@de.ibm.com>
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Jens Freimann <jfrei@linux.vnet.ibm.com>
Signed-off-by: Alexander Graf <agraf@suse.de>
The global variable 'ram_size' is hidden by the local variable
declaration in s390_init(). Since we want to update the global
ram size in certain cases we must not use a local ram_size
variable.
- This fixes booting with unusual ram sizes like -m 67001
- This changes behaviour back to the situation before commit
5f072e1f30
(create struct for machine initialization arguments)
Signed-off-by: Heinz Graalfs <graalfs@linux.vnet.ibm.com>
Reviewed-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Alexander Graf <agraf@suse.de>
My QEMU compile failed with the following error:
qemu-pixman.c: In function ‘qemu_pixman_get_type’:
qemu-pixman.c:24: error: ‘PIXMAN_TYPE_BGRA’ undeclared (first use in this function)
qemu-pixman.c:24: error: (Each undeclared identifier is reported only once
qemu-pixman.c:24: error: for each function it appears in.)
Guard the PIXMAN_TYPE_BGRA branch like in the case right above the failing
case, so that compilation is fixed. Functionality on such old pixman is a
different question ;-).
Signed-off-by: Alexander Graf <agraf@suse.de>
Commit 094b287f0b accidentally broke the "-usb" command line
option, so it would have no effect if the user had not specified
any machine options at that point. (the return value from
'qemu_opts_find(qemu_find_opts("machine"), 0);' is NULL if there
are no user specified options, so it is only to be used for
looking up an option, not when trying to set one.) Similarly,
would '-usbdevice' no longer cause USB to default to enabled.
Fix this regression by using the same style of code for forcing
the usb=on machine option that we use for other aliases such as
'-enable-kvm'.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
In one of the recent reworks to the XICS code, a bug was introduced where
we use the wrong sense and allocate level interrupts instead of message
interrupts for PCI MSIs. This patch fixes it.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Alexander Graf <agraf@suse.de>
QEMU allocates a map enough for 4k pages. However the system page size
can be 64K (for example on POWER) and the host kernel uses only a small
part of it as one big stores a dirty flag for 16 pages 4K each,
the hpratio variable stores this ratio and
the kvm_get_dirty_pages_log_range function handles it correctly.
However kvm_get_dirty_pages_log_range still goes beyond the data
provided by the host kernel which is not correct. It does not cause
errors at the moment as the whole bitmap is zeroed before doing KVM ioctl.
The patch reduces number of iterations over the map.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Alexander Graf <agraf@suse.de>
This patch fixes bug 1031698 :
https://bugs.launchpad.net/qemu/+bug/1031698
If we look at the (truncated) translation of the conditional branch
instruction in the test submitted in the bug post, the call to the
exception helper is missing in the "bne-false" chunk of translated
code :
IN:
bne- 0x1800278
OUT:
0xb544236d: jne 0xb5442396
0xb5442373: mov %ebp,(%esp)
0xb5442376: mov $0x44,%ebx
0xb544237b: mov %ebx,0x4(%esp)
0xb544237f: mov $0x1800278,%ebx
0xb5442384: mov %ebx,0x25c(%ebp)
0xb544238a: call 0x827475a
^^^^^^^^^^^^^^^^^^
0xb5442396: mov %ebp,(%esp)
0xb5442399: mov $0x44,%ebx
0xb544239e: mov %ebx,0x4(%esp)
0xb54423a2: mov $0x1800270,%ebx
0xb54423a7: mov %ebx,0x25c(%ebp)
Indeed, gen_exception(ctx, excp) called by gen_goto_tb (called by
gen_bcond) changes ctx->exception's value to excp's :
gen_bcond()
{
gen_goto_tb(ctx, 0, ctx->nip + li - 4);
/* ctx->exception value is POWERPC_EXCP_BRANCH */
gen_goto_tb(ctx, 1, ctx->nip);
/* ctx->exception now value is POWERPC_EXCP_TRACE */
}
Making the following gen_goto_tb()'s test false during the second call :
if ((ctx->singlestep_enabled &
(CPU_BRANCH_STEP | CPU_SINGLE_STEP)) &&
ctx->exception == POWERPC_EXCP_BRANCH /* false...*/) {
target_ulong tmp = ctx->nip;
ctx->nip = dest;
/* ... and this is the missing call */
gen_exception(ctx, POWERPC_EXCP_TRACE);
ctx->nip = tmp;
}
So the patch simply adds the missing matching case, fixing our problem.
Signed-off-by: Julio Guerra <guerr@julio.in>
Signed-off-by: Alexander Graf <agraf@suse.de>
Pass qemu_sglist_init the global dma_context_memory rather than a NULL
pointer; this fixes a segfault in dma_memory_map() when the guest
starts using DMA.
Reported-by: Amadeusz Sławiński <amade@asmblr.net>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
SSSE3 support has been added to TCG more than 4 years ago in commit
4242b1bd8a. It has been disabled by
mistake in commit 551a2dec8f.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Signed-off-by: Andreas Färber <afaerber@suse.de>
When adding the Haswell CPU model, I intended to make it a superset of the
features present on the SandyBridge model, but I have removed the SEP and
RDTSCP features from the feature list by mistake. This patch adds the
missing SEP and RDTSCP features (that are present on SandyBridge) to
Haswell.
Reported-by: Martin Kletzander <mkletzan@redhat.com>
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Andreas Färber <afaerber@suse.de>
Commit 0d93692 (qdev: Convert busses to QEMU Object Model, 2012-05-02)
removed a check on the type of the bus where a SCSI disk is hotplugged.
However, hot-plugging to the wrong kind of device now causes a crash
due to either a NULL pointer dereference (avoided by the previous patch)
or a failed QOM cast.
Instead, in this case we need to use object_dynamic_cast and check for
the result, similar to what was done before that commit.
Reported-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Trying to cast a NULL value will cause a crash. Returning
NULL is also sensible, and it is also what the type-unsafe
DO_UPCAST macro does.
Reported-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
The repeated calls to WaitForMultipleObjects may cause a livelock in aio_poll,
where no progress is made on bottom halves. This patch matches the behavior
of the POSIX code.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Jason Baron <jbaron@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>