- Devel - libvirt List Archives

[libvirt] [PATCH] qemu: Add support for postcopy-requests migration statistics
by Jiri Denemark 05 Dec '18

05 Dec '18

QEMU can report how many times during post-copy migration the domain running on the destination host tried to access a page which has not been migrated yet. Signed-off-by: Jiri Denemark <jdenemar(a)redhat.com> --- include/libvirt/libvirt-domain.h | 10 ++++++++++ src/qemu/qemu_domain.c | 5 ++++- src/qemu/qemu_migration_cookie.c | 5 +++++ src/qemu/qemu_monitor.h | 1 + src/qemu/qemu_monitor_json.c | 2 ++ tools/virsh-domain.c | 8 ++++++++ 6 files changed, 30 insertions(+), 1 deletion(-) diff --git a/include/libvirt/libvirt-domain.h b/include/libvirt/libvirt-domain.h index 71debd92b8..897b06e6d0 100644 --- a/include/libvirt/libvirt-domain.h +++ b/include/libvirt/libvirt-domain.h @@ -3378,6 +3378,16 @@ typedef enum { */ # define VIR_DOMAIN_JOB_MEMORY_ITERATION "memory_iteration" +/** + * VIR_DOMAIN_JOB_MEMORY_POSTCOPY_REQS: + * + * virDomainGetJobStats field: number page requests received from the + * destination host during post-copy migration, as VIR_TYPED_PARAM_ULLONG. + * This counter is incremented whenever the migrated domain tries to access + * a memory page which has not been transferred from the source host yet. + */ +# define VIR_DOMAIN_JOB_MEMORY_POSTCOPY_REQS "memory_postcopy_requests" + /** * VIR_DOMAIN_JOB_DISK_TOTAL: * diff --git a/src/qemu/qemu_domain.c b/src/qemu/qemu_domain.c index fbe63e2e1d..3dbd9d47ca 100644 --- a/src/qemu/qemu_domain.c +++ b/src/qemu/qemu_domain.c @@ -609,7 +609,10 @@ qemuDomainMigrationJobInfoToParams(qemuDomainJobInfoPtr jobInfo, stats->ram_dirty_rate) < 0 || virTypedParamsAddULLong(&par, &npar, &maxpar, VIR_DOMAIN_JOB_MEMORY_ITERATION, - stats->ram_iteration) < 0) + stats->ram_iteration) < 0 || + virTypedParamsAddULLong(&par, &npar, &maxpar, + VIR_DOMAIN_JOB_MEMORY_POSTCOPY_REQS, + stats->ram_postcopy_reqs) < 0) goto error; if (stats->ram_page_size > 0 && diff --git a/src/qemu/qemu_migration_cookie.c b/src/qemu/qemu_migration_cookie.c index 60df449d53..84f0101382 100644 --- a/src/qemu/qemu_migration_cookie.c +++ b/src/qemu/qemu_migration_cookie.c @@ -703,6 +703,9 @@ qemuMigrationCookieStatisticsXMLFormat(virBufferPtr buf, virBufferAsprintf(buf, "<%1$s>%2$llu</%1$s>\n", VIR_DOMAIN_JOB_MEMORY_ITERATION, stats->ram_iteration); + virBufferAsprintf(buf, "<%1$s>%2$llu</%1$s>\n", + VIR_DOMAIN_JOB_MEMORY_POSTCOPY_REQS, + stats->ram_postcopy_reqs); virBufferAsprintf(buf, "<%1$s>%2$llu</%1$s>\n", VIR_DOMAIN_JOB_MEMORY_PAGE_SIZE, @@ -1102,6 +1105,8 @@ qemuMigrationCookieStatisticsXMLParse(xmlXPathContextPtr ctxt) ctxt, &stats->ram_dirty_rate); virXPathULongLong("string(./" VIR_DOMAIN_JOB_MEMORY_ITERATION "[1])", ctxt, &stats->ram_iteration); + virXPathULongLong("string(./" VIR_DOMAIN_JOB_MEMORY_POSTCOPY_REQS "[1])", + ctxt, &stats->ram_postcopy_reqs); virXPathULongLong("string(./" VIR_DOMAIN_JOB_MEMORY_PAGE_SIZE "[1])", ctxt, &stats->ram_page_size); diff --git a/src/qemu/qemu_monitor.h b/src/qemu/qemu_monitor.h index 48b142a4f4..3430426c4b 100644 --- a/src/qemu/qemu_monitor.h +++ b/src/qemu/qemu_monitor.h @@ -706,6 +706,7 @@ struct _qemuMonitorMigrationStats { unsigned long long ram_dirty_rate; unsigned long long ram_page_size; unsigned long long ram_iteration; + unsigned long long ram_postcopy_reqs; unsigned long long disk_transferred; unsigned long long disk_remaining; diff --git a/src/qemu/qemu_monitor_json.c b/src/qemu/qemu_monitor_json.c index 3de298c9e2..4c151dd405 100644 --- a/src/qemu/qemu_monitor_json.c +++ b/src/qemu/qemu_monitor_json.c @@ -3296,6 +3296,8 @@ qemuMonitorJSONGetMigrationStatsReply(virJSONValuePtr reply, &stats->ram_page_size)); ignore_value(virJSONValueObjectGetNumberUlong(ram, "dirty-sync-count", &stats->ram_iteration)); + ignore_value(virJSONValueObjectGetNumberUlong(ram, "postcopy-requests", + &stats->ram_postcopy_reqs)); disk = virJSONValueObjectGetObject(ret, "disk"); if (disk) { diff --git a/tools/virsh-domain.c b/tools/virsh-domain.c index 372bdb95d3..a71cc9d0be 100644 --- a/tools/virsh-domain.c +++ b/tools/virsh-domain.c @@ -6233,6 +6233,14 @@ cmdDomjobinfo(vshControl *ctl, const vshCmd *cmd) } else if (rc) { vshPrint(ctl, "%-17s %-12llu\n", _("Iteration:"), value); } + + if ((rc = virTypedParamsGetULLong(params, nparams, + VIR_DOMAIN_JOB_MEMORY_POSTCOPY_REQS, + &value)) < 0) { + goto save_error; + } else if (rc) { + vshPrint(ctl, "%-17s %-12llu\n", _("Postcopy requests:"), value); + } } if (info.fileTotal || info.fileRemaining || info.fileProcessed) { -- 2.19.1

2 1

[libvirt] [PATCH 0/5] Various cleanups and fixes (blockdev-add saga)
by Peter Krempa 05 Dec '18

05 Dec '18

Few random patches extracted from my blockdev series. Peter Krempa (5): tests: qemuxml2xml: Add few debug statements for status XML testing util: xml: Always consume args of virXMLFormatElement qemu: domain: Initialize proper element in qemuDomainPrepareStorageSourceBlockdev conf: snapshot: Remove file format check from parser tests: qemuxml2argv: Remove disks from few tests which don't need them src/conf/domain_conf.c | 12 ++---------- src/conf/snapshot_conf.c | 12 +++--------- src/qemu/qemu_domain.c | 9 +++------ src/util/virxml.c | 16 ++++++++++++---- .../channel-unix-guestfwd.x86_64-2.5.0.args | 2 -- .../channel-unix-guestfwd.x86_64-latest.args | 2 -- tests/qemuxml2argvdata/channel-unix-guestfwd.xml | 5 ----- .../console-virtio-unix.x86_64-2.5.0.args | 2 -- .../console-virtio-unix.x86_64-latest.args | 2 -- tests/qemuxml2argvdata/console-virtio-unix.xml | 5 ----- tests/qemuxml2argvdata/net-vhostuser.args | 3 --- .../net-vhostuser.x86_64-2.5.0.args | 2 -- .../net-vhostuser.x86_64-latest.args | 2 -- tests/qemuxml2argvdata/net-vhostuser.xml | 6 ------ .../parallel-unix-chardev.x86_64-2.5.0.args | 2 -- .../parallel-unix-chardev.x86_64-latest.args | 2 -- tests/qemuxml2argvdata/parallel-unix-chardev.xml | 5 ----- tests/qemuxml2xmloutdata/net-vhostuser.xml | 6 ------ tests/qemuxml2xmltest.c | 9 ++++++--- 19 files changed, 26 insertions(+), 78 deletions(-) -- 2.19.2

2 6

[libvirt] [PATCH v2] qemuDomainRemoveRNGDevice: Remove associated chardev too
by Michal Privoznik 05 Dec '18

05 Dec '18

https://bugzilla.redhat.com/show_bug.cgi?id=1656014 An RNG device can consists of more devices than RND device itself. For instance, in case of EGD there is a chardev that connects to EGD daemon and feeds the qemu with random data. When doing RNG device removal we have to remove the associated chardev as well. Signed-off-by: Michal Privoznik <mprivozn(a)redhat.com> --- diff to v1: - Detach Extension Device too - Reoder the way in which devices are detached (as suggested in v1) src/qemu/qemu_hotplug.c | 16 +++++++++++++--- 1 file changed, 13 insertions(+), 3 deletions(-) diff --git a/src/qemu/qemu_hotplug.c b/src/qemu/qemu_hotplug.c index 5f756b7267..f5f8cec59e 100644 --- a/src/qemu/qemu_hotplug.c +++ b/src/qemu/qemu_hotplug.c @@ -4814,7 +4814,7 @@ qemuDomainRemoveRNGDevice(virQEMUDriverPtr driver, qemuDomainObjPrivatePtr priv = vm->privateData; ssize_t idx; int ret = -1; - int rc; + int rc = 0; VIR_DEBUG("Removing RNG device %s from domain %p %s", rng->info.alias, vm, vm->def->name); @@ -4828,7 +4828,17 @@ qemuDomainRemoveRNGDevice(virQEMUDriverPtr driver, qemuDomainObjEnterMonitor(driver, vm); - rc = qemuMonitorDelObject(priv->mon, objAlias); + if (qemuDomainDetachExtensionDevice(priv->mon, &rng->info) < 0) + rc = -1; + + if (rc == 0 && + qemuMonitorDelObject(priv->mon, objAlias) < 0) + rc = -1; + + if (rng->backend == VIR_DOMAIN_RNG_BACKEND_EGD && + rc == 0 && + qemuMonitorDetachCharDev(priv->mon, charAlias) < 0) + rc = -1; if (qemuDomainObjExitMonitor(driver, vm) < 0) goto cleanup; @@ -4837,7 +4847,7 @@ qemuDomainRemoveRNGDevice(virQEMUDriverPtr driver, rc == 0 && qemuDomainDelChardevTLSObjects(driver, vm, rng->source.chardev, charAlias) < 0) - goto cleanup; + rc = -1; virDomainAuditRNG(vm, rng, NULL, "detach", rc == 0); -- 2.19.2

2 1

[libvirt] [PATCH libvirt-python] virNodeInfo.memory is in KiB
by Philipp Hahn 05 Dec '18

05 Dec '18

but the Python library does an extra left shift of 10 bits returning MiB instead: > # cat y.c > #include <stdlib.h> > #include <stdio.h> > #include <libvirt.h> > int main(void) { > virConnectPtr conn = virConnectOpen("qemu:///system"); > virNodeInfo info; > int rv = virNodeGetInfo(conn, &info); > printf("%ld\n", info.memory); > return rv; > } > # gcc y.c -I/usr/include/libvirt -lvirt > # ./a.out > 4041088 > # python -c 'import libvirt;c=libvirt.open("qemu:///system");print(c.getInfo()[1])' > 3946 Fixes: 197153c6 Signed-off-by: Philipp Hahn <hahn(a)univention.de> --- libvirt-override.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/libvirt-override.c b/libvirt-override.c index f7b2f6b..616fa1c 100644 --- a/libvirt-override.c +++ b/libvirt-override.c @@ -2740,7 +2740,7 @@ libvirt_virNodeGetInfo(PyObject *self ATTRIBUTE_UNUSED, VIR_PY_LIST_SET_GOTO(py_retval, 0, libvirt_constcharPtrWrap(&info.model[0]), error); VIR_PY_LIST_SET_GOTO(py_retval, 1, - libvirt_longWrap((long) info.memory >> 10), error); + libvirt_longWrap((long) info.memory), error); VIR_PY_LIST_SET_GOTO(py_retval, 2, libvirt_intWrap((int) info.cpus), error); VIR_PY_LIST_SET_GOTO(py_retval, 3, libvirt_intWrap((int) info.mhz), error); VIR_PY_LIST_SET_GOTO(py_retval, 4, libvirt_intWrap((int) info.nodes), error); -- 2.11.0

2 2

[libvirt] [PATCH RFC v2] qemu: fix deadlock when waiting in non async jobs
by Nikolay Shirokovskiy 05 Dec '18

05 Dec '18

Block job abort operation can not handle properly qemu crashes when waiting for abort/pivot completion. Deadlock scenario is next: - qemuDomainBlockJobAbort waits for pivot/abort completion - qemu crashes, then qemuProcessBeginStopJob broadcasts for VM condition and then waits for job condition (taken by qemuDomainBlockJobAbort) - qemuDomainBlockJobAbort awakes but nothing really changed, VM is still active (vm->def->id != -1) so thread starts waiting for completion again. Now two threads are in deadlock. First let's remove broadcast in qemuProcessBeginStopJob. It is simply wrong because it is not set any condition before broadcast so that awaked threads can not detect any changes. Crashing domain during async job will continue to be handled properly because destroy job can run concurrently with async job and destroy job calls qemuProcessStop which sets vm->def->id to -1 and broadcasts. Second let's introduce flag that EOF is received and broadcast after that. Now non async jobs can check this flag in wait loop. Signed-off-by: Nikolay Shirokovskiy <nshirokovskiy(a)virtuozzo.com> --- Diff from v1: - patches 1 and 2 are already merged - don't bother with reporting monitor EOF reason to user as most of time it is simply "unexpected eof" (this implies dropping patch 3) - drop patch 5 as we now always report "domain is being stopped" in qemuDomainObjWait - don't signal on monitor error for simplicity (otherwise we need to report something more elaborate that "domain is being stopped" as we don't kill domain on monitor errors. On the other hand I guess monitor error is rare case to handle it right now) - keep virDomainObjWait for async jobs It's a bit uneven that for async jobs domain is destroyed concurrently and for non async jobs it will be actually destroyed after job get completed. Also if non async job needs issuing commands to qemu on cleanup then we will send these commands in vain polluting logs etc because qemu process in not running at this moment but typical check (virDomainObjIsActive) will think it is still running. Domain is destroyed (qemuProcessStop) in a job due to patches [1] and [2]. However AFAIU it is not neccessary. If qemuProcessStop does not drop VM lock then we don't need extra job to make qemuProcessStop and main job not interleave. And we can drop the lock now only in qemuDomainObjBeginNestedJob in qemuProcessStop which is introduced in [2]. AFAIU we can fix issues mentioned in [2] the other way for example like it is done for qemu agent - we save agent monitor reference on stack for entering/exiting agent monitor. So I wonder can we instead of this fix remove job for qemuProcessStop and run destroying domain cuncurrently for non async jobs too. [1] commit 8c9ff9960b29d4703a99efdd1cadcf6f48799cc0 Author: Jiri Denemark <jdenemar(a)redhat.com> Date: Thu Feb 11 15:32:48 2016 +0100 qemu: Process monitor EOF in a job [2] commit 81f50cb92d16643bcd749e3ab5b404b8b7cec643 Author: Jiri Denemark <jdenemar(a)redhat.com> Date: Thu Feb 11 11:20:28 2016 +0100 qemu: Avoid calling qemuProcessStop without a job src/qemu/qemu_domain.c | 39 +++++++++++++++++++++++++++++++++++++++ src/qemu/qemu_domain.h | 4 ++++ src/qemu/qemu_driver.c | 2 +- src/qemu/qemu_hotplug.c | 4 ++-- src/qemu/qemu_process.c | 9 +++++---- 5 files changed, 51 insertions(+), 7 deletions(-) diff --git a/src/qemu/qemu_domain.c b/src/qemu/qemu_domain.c index 939b2a3..aead72b 100644 --- a/src/qemu/qemu_domain.c +++ b/src/qemu/qemu_domain.c @@ -13534,3 +13534,42 @@ qemuDomainRunningReasonToResumeEvent(virDomainRunningReason reason) return VIR_DOMAIN_EVENT_RESUMED_UNPAUSED; } + + +/** + * Waits for domain condition to be triggered for a specific period of time. + * if @until is 0 then waits indefinetely. + * + * Returns: + * -1 on error + * 0 on success + * 1 on timeout + */ +int +qemuDomainObjWait(virDomainObjPtr vm, unsigned long long until) +{ + qemuDomainObjPrivatePtr priv = vm->privateData; + int rc; + + if (until) + rc = virCondWaitUntil(&vm->cond, &vm->parent.lock, until); + else + rc = virCondWait(&vm->cond, &vm->parent.lock); + + if (rc < 0) { + if (until && errno == ETIMEDOUT) + return 1; + + virReportSystemError(errno, "%s", + _("failed to wait for domain condition")); + return -1; + } + + if (priv->monEOF) { + virReportError(VIR_ERR_OPERATION_FAILED, "%s", + _("domain is being stopped")); + return -1; + } + + return 0; +} diff --git a/src/qemu/qemu_domain.h b/src/qemu/qemu_domain.h index 2f8a1bf..36ab294 100644 --- a/src/qemu/qemu_domain.h +++ b/src/qemu/qemu_domain.h @@ -281,6 +281,7 @@ struct _qemuDomainObjPrivate { virDomainChrSourceDefPtr monConfig; bool monJSON; bool monError; + bool monEOF; unsigned long long monStart; qemuAgentPtr agent; @@ -1085,4 +1086,7 @@ void qemuDomainStorageIdReset(qemuDomainObjPrivatePtr priv); virDomainEventResumedDetailType qemuDomainRunningReasonToResumeEvent(virDomainRunningReason reason); +int +qemuDomainObjWait(virDomainObjPtr vm, unsigned long long until); + #endif /* __QEMU_DOMAIN_H__ */ diff --git a/src/qemu/qemu_driver.c b/src/qemu/qemu_driver.c index b238309..f4250da 100644 --- a/src/qemu/qemu_driver.c +++ b/src/qemu/qemu_driver.c @@ -17142,7 +17142,7 @@ qemuDomainBlockJobAbort(virDomainPtr dom, qemuDomainDiskPrivatePtr diskPriv = QEMU_DOMAIN_DISK_PRIVATE(disk); qemuBlockJobUpdate(vm, QEMU_ASYNC_JOB_NONE, disk, NULL); while (diskPriv->blockjob) { - if (virDomainObjWait(vm) < 0) { + if (qemuDomainObjWait(vm, 0) < 0) { ret = -1; goto endjob; } diff --git a/src/qemu/qemu_hotplug.c b/src/qemu/qemu_hotplug.c index 4558a3c..8189629 100644 --- a/src/qemu/qemu_hotplug.c +++ b/src/qemu/qemu_hotplug.c @@ -165,7 +165,7 @@ qemuHotplugWaitForTrayEject(virDomainObjPtr vm, return -1; while (disk->tray_status != VIR_DOMAIN_DISK_TRAY_OPEN) { - if ((rc = virDomainObjWaitUntil(vm, now + CHANGE_MEDIA_TIMEOUT)) < 0) + if ((rc = qemuDomainObjWait(vm, now + CHANGE_MEDIA_TIMEOUT)) < 0) return -1; if (rc > 0) { @@ -5002,7 +5002,7 @@ qemuDomainWaitForDeviceRemoval(virDomainObjPtr vm) until += qemuDomainRemoveDeviceWaitTime; while (priv->unplug.alias) { - if ((rc = virDomainObjWaitUntil(vm, until)) == 1) + if ((rc = qemuDomainObjWait(vm, until)) == 1) return 0; if (rc < 0) { diff --git a/src/qemu/qemu_process.c b/src/qemu/qemu_process.c index 29b0ba1..dd03269 100644 --- a/src/qemu/qemu_process.c +++ b/src/qemu/qemu_process.c @@ -290,9 +290,12 @@ qemuProcessHandleMonitorEOF(qemuMonitorPtr mon, virObjectLock(vm); + priv = vm->privateData; + priv->monEOF = true; + virDomainObjBroadcast(vm); + VIR_DEBUG("Received EOF on %p '%s'", vm, vm->def->name); - priv = vm->privateData; if (priv->beingDestroyed) { VIR_DEBUG("Domain is being destroyed, EOF is expected"); goto cleanup; @@ -5996,6 +5999,7 @@ qemuProcessPrepareDomain(virQEMUDriverPtr driver, priv->monJSON = true; priv->monError = false; + priv->monEOF = false; priv->monStart = 0; priv->gotShutdown = false; priv->runningReason = VIR_DOMAIN_RUNNING_UNKNOWN; @@ -6965,9 +6969,6 @@ qemuProcessBeginStopJob(virQEMUDriverPtr driver, if (qemuProcessKill(vm, killFlags) < 0) goto cleanup; - /* Wake up anything waiting on domain condition */ - virDomainObjBroadcast(vm); - if (qemuDomainObjBeginJob(driver, vm, job) < 0) goto cleanup; -- 1.8.3.1

2 5

[libvirt] [PATCH] Fix minor typos in messages and docs
by Yuri Chornoivan 05 Dec '18

05 Dec '18

Hi, Please find the attached patch with minor fixes. Many thanks in advance for reviewing it. Best regards, Yuri

2 1

[libvirt] [PATCH 0/2] qemuDomainRemoveRNGDevice: Remove associated chardev too
by Michal Privoznik 04 Dec '18

04 Dec '18

*** BLURB HERE *** Michal Prívozník (2): qemuBuildRNGBackendChrdevStr: Fix formatting qemuDomainRemoveRNGDevice: Remove associated chardev too src/qemu/qemu_command.c | 6 ++++-- src/qemu/qemu_hotplug.c | 23 +++++++++++++++++------ 2 files changed, 21 insertions(+), 8 deletions(-) -- 2.19.2

2 4

[libvirt] [PATCH] util: Fix the build on MinGW because of missing DT_CHR dirent type
by Erik Skultety 04 Dec '18

04 Dec '18

Caused by commit 39480969 Signed-off-by: Erik Skultety <eskultet(a)redhat.com> --- Pushed under the build breaker rule. src/util/virutil.c | 3 --- 1 file changed, 3 deletions(-) diff --git a/src/util/virutil.c b/src/util/virutil.c index da12a11e04..279e6aedc0 100644 --- a/src/util/virutil.c +++ b/src/util/virutil.c @@ -2172,9 +2172,6 @@ virHostGetDRMRenderNode(void) return NULL; while ((dirErr = virDirRead(driDir, &ent, driPath)) > 0) { - if (ent->d_type != DT_CHR) - continue; - if (STRPREFIX(ent->d_name, "renderD")) { have_rendernode = true; break; -- 2.19.2

1 0

[libvirt] [PATCH 0/2] syncNicRxFilterMultiMode: Trivial fixes
by Michal Privoznik 03 Dec '18

03 Dec '18

Pushed under trivial rule. Michal Prívozník (2): syncNicRxFilterMultiMode: Check for helper's retval properly syncNicRxFilterMultiMode: Fix indentation src/qemu/qemu_driver.c | 80 +++++++++++++++++++++--------------------- 1 file changed, 40 insertions(+), 40 deletions(-) -- 2.19.2

1 2

Re: [libvirt] [PATCH] qemu: Qemu process unexpectedly killed in repeated reboot
by Wangjing (King) 03 Dec '18

03 Dec '18

On 11/30/18 9:53 AM, Wang King wrote: >> The issue occurs when I make repeated calls to virDomainReboot with >> VIR_DOMAIN_REBOOT_DEFAULT flag. In the first call to reboot domain, >> the qemu driver chose ACPI path, and set priv->fakeReboot to true. >> Then in a second call, qemu driver chose agent to reboot which set >> fakeReboot to false. But because the guest already responded to ACPI >> shut down, libvirtd daemon will process a SHUTDOWN event in >> qemuProcessShutdownOrReboot and checks priv->fakeReboot. Since the >> fakeReboot flag is now false, qemu process is unexpectedly killed. >> > >This sounds fishy. Looking at the code libvirt decides whether to use agent or ACPI based on: > >a) flags (but since you're passing 0 this is out of the picture), >b) guest agent being available, > >This means that agent must have connected between two virDomainReboot() calls. Otherwise libvirt would make the same choice. > >> I have no idea how to fix it. > >Well, the qemuDomainSetFakeReboot(false) call was added in b0c144c5792 which points to: > > https://www.redhat.com/archives/libvir-list/2015-April/msg00732.html > >I think the patch proposed there is actually right and not the one that was merged. > >Michal The problem: Consider this scenario: firstly, call virDomainReboot() with mode ACPI to a domain, then immediately call virDomainReboot() again with mode “agent”. Supposedly, this sequence will reboot the domain (hereinafter referred to as DOM1). Instead, however, the current version of libvirt code will *KILL* the QEMU process associated with the domain. How to reproduce this problem: Simply make a ACPI-mode reboot call to the domain, and immediately make an agent-mode reboot call. You will see that the domain is killed straight away. Root cause: The unexpected SIGKILL comes from Libvirt seeing fakeReboot flag set to *FALSE* when it received an SHUTDOWN event as a result of calling ACPI virDomainReboot(). This happens when DOM1 is already rebooting, yet its guest agent is still available, thus the second agent-mode virDomainReboot() was able to pass the check and set fakeReboot flag to *FALSE* (its “guest-shutdown” command is also sent, which is out of the scope of this email). Finally, DOM1 shut down as part of the reboot process, its QEMU process sent SHUTDOWN event to Libvirt. Libvirt checked fakeReboot flag and saw it being *FALSE*, forcefully kills the QEMU process as a result. Analysis: Say that we make two calls to shutdown/reboot DOM1. DOM1 will eventually respond to only one of the two calls. For another call, only two scenarios happen: DOM1 accepts the second call but do not execute it (meaning only fakeReboot flag is modified on Libvirt side), or DOM1 accepts the second call and overrides the first one. In the second scenario, it is the same situation as first scenario but reversed (DOM1 responds to first call and sets flag with second call vs DOM1 responds to second call and sets flag with first call). Therefore, we hereby simplify the problem to that DOM1 always responds to first call, and the second call is also accepted, but it only serves to change the fakeReboot flag. As you will see, there is what we expect to happen versus what actually happens. The questionable result is marked with tilde(~) and the faulted result is marked with emphasis(*). If you see the chart below, you will see that how the domain should react to shutdown/reboot calls in different order: ----------------------------------------------------------------------------------------------------------- |First Command |Second Command(set flags only) |Expected Result |Actual Result |After Fix | ----------------------------------------------------------------------------------------------------------- |ACPI SHUTDOWN |ACPI SHUTDOWN |SHUTDOWN |SHUTDOWN |SHUTDOWN | |ACPI REBOOT |ACPI SHUTDOWN |REBOOT |~SHUTDOWN~ |~SHUTDOWN~ | |AGENT SHUTDOWN |ACPI SHUTDOWN |SHUTDOWN |SHUTDOWN |SHUTDOWN | |AGENT REBOOT |ACPI SHUTDOWN |REBOOT |REBOOT |REBOOT | |ACPI SHUTDOWN |ACPI REBOOT |SHUTDOWN |~REBOOT~ |~REBOOT~ | |ACPI REBOOT |ACPI REBOOT |REBOOT |REBOOT |REBOOT | |AGENT SHUTDOWN |ACPI REBOOT |SHUTDOWN |~REBOOT~ |~REBOOT~ | |AGENT REBOOT |ACPI REBOOT |REBOOT |REBOOT |REBOOT | |ACPI SHUTDOWN |AGENT SHUTDOWN |SHUTDOWN |SHUTDOWN |SHUTDOWN | |ACPI REBOOT |AGENT SHUTDOWN |REBOOT |~SHUTDOWN~ |~SHUTDOWN~ | |AGENT SHUTDOWN |AGENT SHUTDOWN |SHUTDOWN |SHUTDOWN |SHUTDOWN | |AGENT REBOOT |AGENT SHUTDOWN |REBOOT |REBOOT |REBOOT | |ACPI SHUTDOWN |AGENT REBOOT |SHUTDOWN |SHUTDOWN |SHUTDOWN | |ACPI REBOOT |AGENT REBOOT |REBOOT |**SHUTDOWN** |**REBOOT** | |AGENT SHUTDOWN |AGENT REBOOT |SHUTDOWN |SHUTDOWN |SHUTDOWN | |AGENT REBOOT |AGENT REBOOT |REBOOT |REBOOT |REBOOT | ---------------------------------------------------------------------------------------------------------- As you can see, when DOM1 reboots with ACPI and immediately receives agent-mode reboot. It will *SHUTDOWN* instead. (Certainly, there are questionable results in other combos, but we think they cause less confusions in real-life situations so we are not discussing them here) Fix: The fix is to revert the changes added by b0c144c5792 commit, which introduced setting fakeReboot to False in agent-mode reboot. This should, at least, avoid killing QEMU process during successive reboot calls. Nonetheless, shutdown/reboot process still needs some optimizations to be done. I will leave it to the analysis above. What do you think about this? Should we revert the commit first or find an all-fix solution to this problem?

1 0