Bug 997666
| Summary: | [RFE] 100 createVolume calls on a block sd with 500 LVs takes too much time | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Product: | Red Hat Enterprise Virtualization Manager | Reporter: | Marina Kalinin <mkalinin> | ||||||||
| Component: | RFEs | Assignee: | Allon Mureinik <amureini> | ||||||||
| Status: | CLOSED ERRATA | QA Contact: | Yuri Obshansky <yobshans> | ||||||||
| Severity: | high | Docs Contact: | |||||||||
| Priority: | high | ||||||||||
| Version: | 3.2.0 | CC: | acanan, amureini, bazulay, bgraveno, cww, iheim, lbopf, lnatapov, lpeer, lyarwood, mkalinin, nsoffer, rbalakri, scohen, tnisan, yeylon, ylavi, yobshans | ||||||||
| Target Milestone: | ovirt-3.6.0-rc | Keywords: | FutureFeature, Improvement | ||||||||
| Target Release: | 3.6.0 | Flags: | sherold:
Triaged+
|
||||||||
| Hardware: | Unspecified | ||||||||||
| OS: | Unspecified | ||||||||||
| Whiteboard: | |||||||||||
| Fixed In Version: | Doc Type: | Enhancement | |||||||||
| Doc Text: | Story Points: | --- | |||||||||
| Clone Of: | Environment: | ||||||||||
| Last Closed: | 2016-03-09 20:31:59 UTC | Type: | Bug | ||||||||
| Regression: | --- | Mount Type: | --- | ||||||||
| Documentation: | --- | CRM: | |||||||||
| Verified Versions: | Category: | --- | |||||||||
| oVirt Team: | Storage | RHEL 7.3 requirements from Atomic Host: | |||||||||
| Cloudforms Team: | --- | Target Upstream Version: | |||||||||
| Embargoed: | |||||||||||
| Bug Depends On: | |||||||||||
| Bug Blocks: | 1018916 | ||||||||||
| Attachments: |
|
||||||||||
Created attachment 788715 [details]
patched adjusted to current downstream
Created attachment 788718 [details]
the working patch
Implemented the change in the attacked working patch. It works - no original errors as on the bug description. But - it takes longer time to process the request (2.5 times longer - 25 second compared to 7-9 sec before). So, probably performing the "release" command per each volume not that efficient. I will wait for your advise, From all the testing I am doing, I do not see when we actually call teardown for the actual new volume. volume.clone creates a new volume from itself(i.e. the template lv), and then calls teardown on itself again. which would never deactivate the newly created lv. Do we really need this method? Or it is not used properly? As for the part that takes more time now, I suspect it is the activate command we added to blockVolume.py and we call in volume.prepare:
elif self.__class__.__name__ == "BlockVolume": 567
self.activate()
blockVolume.activate would perform lvchange -ay. Before applying the patch, volume.prepare would set 'r' attribute to the lv. I don't see if we activate it as well. With the patch, we create a new race condition/redundancy - by calling lvchange -ay for the template lv all the time we clone from this template. But this lv should be active all the time anyway. So we should check if it is a shared volume and skip the activation part, probably. Created attachment 789010 [details]
patch without calling activate lv for cloning from template
(In reply to Marina from comment #11) > blockVolume.activate would perform lvchange -ay. > > Before applying the patch, volume.prepare would set 'r' attribute to the lv. > I don't see if we activate it as well. > With the patch, we create a new race condition/redundancy - by calling > lvchange -ay for the template lv all the time we clone from this template. > > But this lv should be active all the time anyway. > So we should check if it is a shared volume and skip the activation part, > probably. Calling lvm.activateLVs (the unique path that we should take) on an already active LV has no cost since the lvm command will be not performed. Checking if it is a shared volume by reading MD will slow it more. We problably should avoid the clone path at all. Starting to read info. The issue here is not deactivation of volume, it's contention on lvm calls which take too long. createVolume can be rewritten to avoid a lot of the calls hence significantly reducing the total run time (In reply to Ayal Baron from comment #14) > The issue here is not deactivation of volume, it's contention on lvm calls > which take too long. > createVolume can be rewritten to avoid a lot of the calls hence > significantly reducing the total run time Ayal, this bz comes to address the error: FAILED: <err> = ' device-mapper: remove ioctl on failed: Device or resource busy\n which occurs when VDSM tries to perform teardown on the template volume after creating a vm from that template (see volume.clone). Thus, it sounds to me that the original bug title should be set back: "vdsm should not deactivate template lv after cloning a vm from this template". Please correct me if my understanding is wrong. (In reply to Marina from comment #15) > (In reply to Ayal Baron from comment #14) > > The issue here is not deactivation of volume, it's contention on lvm calls > > which take too long. > > createVolume can be rewritten to avoid a lot of the calls hence > > significantly reducing the total run time > > Ayal, this bz comes to address the error: > FAILED: <err> = ' device-mapper: remove ioctl on failed: Device or > resource busy\n > > which occurs when VDSM tries to perform teardown on the template volume > after creating a vm from that template (see volume.clone). Thus, it sounds > to me that the original bug title should be set back: "vdsm should not > deactivate template lv after cloning a vm from this template". Please > correct me if my understanding is wrong. deactivation of the template lv is not the part which is consuming the time, hence the title was wrong. What you want is for concurrent provisioning of many thin volumes derived from the same template to be fast. How we solve this issue is an implementation detail. Initial patch here: http://gerrit.ovirt.org/#/c/18793/ (In reply to Ayal Baron from comment #16) > (In reply to Marina from comment #15) > > (In reply to Ayal Baron from comment #14) > > > The issue here is not deactivation of volume, it's contention on lvm calls > > > which take too long. > > > createVolume can be rewritten to avoid a lot of the calls hence > > > significantly reducing the total run time > > > > Ayal, this bz comes to address the error: > > FAILED: <err> = ' device-mapper: remove ioctl on failed: Device or > > resource busy\n > > > > which occurs when VDSM tries to perform teardown on the template volume > > after creating a vm from that template (see volume.clone). Thus, it sounds > > to me that the original bug title should be set back: "vdsm should not > > deactivate template lv after cloning a vm from this template". Please > > correct me if my understanding is wrong. > > deactivation of the template lv is not the part which is consuming the time, > hence the title was wrong. > What you want is for concurrent provisioning of many thin volumes derived > from the same template to be fast. I want this, very much, but not in this bug. In this bug I ask the question why do we deactivate template's lv. Which causes device mapper error. > How we solve this issue is an implementation detail. > Initial patch here: http://gerrit.ovirt.org/#/c/18793/ http://gerrit.ovirt.org/#/c/18793/ will address the activation and deactivation problems by just avoiding both (not needed at all in this flow, which should also speed things up a bit). That is being done for 3.3. We're keeping this bug to track performance of the entire operation (hence the title). The errors you got (device-mapper: remove ioctl on failed: Device or resource busy) may be related to concurrency issue with lvm, and should be solved by http://gerrit.ovirt.org/#/c/20890. Please test again with 3.3. a no comments Bug verified on RHEV-M 3.5.1.1-0.1 Setup: 1 DC, 1 Cluster, 1 Host and 1 iSCSi SD (1T) + 10 VM Templates (2 CPU, 2G RAM, Disk 10G) Test scenario: 50 concurrent threads (with 1 min ramp up) create VMs from template (using Clone/Thin provisioning). JMeter script measured time from Create VM REST API call till VM is created and has status - Down. This bug has 2 parts functional and performance 1. Functional part verified and errors didn't happened during scenario execution 2. Performance part verified by performing 2 scenarios: VM creation with Clone and Thin disk provisioning Below Create VM operation response time: 2.1. Clone Transaction Average 90% Line Min Max Create VM form template (msec) 3624490 5058803 988476 5123454 Create VM form template (min) 60 84 16 85 2.2. Thin Transaction Average 90% Line Min Max Create VM from template msec 33679 89997 119 119196 Create VM from template min 1 1 0 2 Since the problem described in this bug report should be resolved in a recent advisory, it has been closed with a resolution of ERRATA. For information on the advisory, and where to find the updated files, follow the link below. If the solution does not work for you, open a new bug report. https://rhn.redhat.com/errata/RHEA-2016-0376.html |
Description of problem: VDSM will perform teardown on the template volume after creating a vm from that template (see volume.clone). A race condition will occure, when multiple vms are created at the same time, this lv will get 'o' attribute, since being in use; and then, when vdsm will try to deactivate it and will fail with the error: "device-mapper: remove ioctl on failed: Device or resource busy\n Unable to deactivate f48e93bb--e3bb--4291--bd2f--8d622ef8019d-a351bc36--d741--4cf7--aa89--735c82b888e5 (253:508)\n'; <rc> = 5" (you cannot run `lvchange -an` on an open/used lv) Why do we need to deactivate template lv? Template's nature is to be in use by other VMs and when this is happening, they are considered to be active. Suggested fix: Check inside volume.clone, if we are creating a vm from a template, then call teardown with justme=True: self.teardown(self.sdUUID, self.volUUID, 'True') Version-Release number of selected component (if applicable): 3.2.0 How reproducible: Reproducable when trying to deploy multiple vms (~50) from the same template at the same time. Steps to Reproduce: 1. Create a template. 2. Issue 50(or more) simultaneous request of creating vms from this template. 3. Watch vdsm.log for this error: "device-mapper: remove ioctl on failed: Device or resource busy\n Unable to deactivate". Actual results: This will slow down significantly the time required to create those vms. Only after creating all the vms, we will be able to remove 'o' attribute from the lv and deactivate it, and terminate successfully all the createVmFromTemplate flows. Expected results: Do not deactivate template lv after cloning a vm from it. Additional info: This affect one of big customer deployments. I will attach ticket number. Here is a partial flow from the vdsm.log: 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:44,553::volume::263::Storage.Volume::(clone) Volume.clone: /rhev/data-center/89ced3c8-7615-11e2-87d3-0014227472ce/f48e93bb-e3bb-4291-bd2f-8d622ef8019 d/images/a460e3a9-4fd7-4fc4-bdeb-839fc85130df/a351bc36-d741-4cf7-aa89-735c82b888e5 to /rhev/data-center/89ced3c8-7615-11e2-87d3-0014227472ce/f48e93bb-e3bb-4291-bd2f-8d622ef8019d/images/a460e3a9-4fd7-4fc4-bdeb-83 9fc85130df/25fb3d9d-dcd1-417c-b5b2-72b27e9a35c4 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:44,579::misc::83::Storage.Misc.excCmd::(<lambda>) '/bin/dd iflag=direct skip=304 bs=512 if=/dev/f48e93bb-e3bb-4291-bd2f-8d622ef8019d/metadata count=1 ' (cwd None) 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:44,731::misc::83::Storage.Misc.excCmd::(<lambda>) SUCCESS: <err> = '1+0 records in\n1+0 records out\n512 bytes (512 B) copied, 0.000354389 s, 1.4 MB/ s\n'; <rc> = 0 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:44,741::misc::344::Storage.Misc::(validateDDBytes) err: ['1+0 records in', '1+0 records out', '512 bytes (512 B) copied, 0.000354389 s, 1.4 MB/s'], s ize: 512 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:44,765::misc::83::Storage.Misc.excCmd::(<lambda>) '/bin/dd iflag=direct skip=304 bs=512 if=/dev/f48e93bb-e3bb-4291-bd2f-8d622ef8019d/metadata count=1 ' (cwd None) 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:44,934::misc::83::Storage.Misc.excCmd::(<lambda>) SUCCESS: <err> = '1+0 records in\n1+0 records out\n512 bytes (512 B) copied, 0.000243984 s, 2.1 MB/ s\n'; <rc> = 0 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:44,945::misc::344::Storage.Misc::(validateDDBytes) err: ['1+0 records in', '1+0 records out', '512 bytes (512 B) copied, 0.000243984 s, 2.1 MB/s'], s ize: 512 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:44,993::misc::83::Storage.Misc.excCmd::(<lambda>) '/usr/bin/qemu-img create -f qcow2 -F qcow2 -b ../a460e3a9-4fd7-4fc4-bdeb-839fc85130df/a351bc36-d74 1-4cf7-aa89-735c82b888e5 /rhev/data-center/89ced3c8-7615-11e2-87d3-0014227472ce/f48e93bb-e3bb-4291-bd2f-8d622ef8019d/images/a460e3a9-4fd7-4fc4-bdeb-839fc85130df/25fb3d9d-dcd1-417c-b5b2-72b27e9a35c4' (cwd /rhev/d ata-center/89ced3c8-7615-11e2-87d3-0014227472ce/f48e93bb-e3bb-4291-bd2f-8d622ef8019d/images/a460e3a9-4fd7-4fc4-bdeb-839fc85130df) 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:45,089::misc::83::Storage.Misc.excCmd::(<lambda>) SUCCESS: <err> = ''; <rc> = 0 [mku: let's see here why we are doing this ] 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::INFO::2013-08-15 14:04:45,099::blockVolume::374::Storage.Volume::(teardown) Tearing down volume f48e93bb-e3bb-4291-bd2f-8d622ef8019d/a351bc36-d741-4cf7-aa89-735c82b888e5 jus tme False 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:45,107::resourceManager::538::ResourceManager::(releaseResource) Trying to release resource 'f48e93bb-e3bb-4291-bd2f-8d622ef8019d_lvmActivationNS.a35 1bc36-d741-4cf7-aa89-735c82b888e5' 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:45,122::resourceManager::553::ResourceManager::(releaseResource) Released resource 'f48e93bb-e3bb-4291-bd2f-8d622ef8019d_lvmActivationNS.a351bc36-d74 1-4cf7-aa89-735c82b888e5' (0 active users) 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:45,137::resourceManager::558::ResourceManager::(releaseResource) Resource 'f48e93bb-e3bb-4291-bd2f-8d622ef8019d_lvmActivationNS.a351bc36-d741-4cf7-aa 89-735c82b888e5' is free, finding out if anyone is waiting for it. 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:45,230::misc::83::Storage.Misc.excCmd::(<lambda>) '/usr/bin/sudo -n /sbin/lvm lvchange --config " devices { preferred_names = [\\"^/dev/mapper/\\"] i gnore_suspended_devices=1 write_cache_state=0 disable_after_error_count=3 filter = [ \'a|/dev/mapper/1Dell_Internal_Dual_SD_0123456789AB|/dev/mapper/3600601605ab02c0028c71db4c076e211|/dev/mapper/3600601605ab02c0 05a0df4e048a0e211|/dev/mapper/3600601605ab02c005c0df4e048a0e211|/dev/mapper/3600601605ab02c005e0df4e048a0e211|/dev/mapper/3600601605ab02c00600df4e048a0e211|/dev/mapper/3600601605ab02c00620df4e048a0e211|/dev/mapp er/3600601605ab02c00640df4e048a0e211|/dev/mapper/36848f690eb9aa00018d3d7c418bfabf6|\', \'r|.*|\' ] } global { locking_type=1 prioritise_write_locks=1 wait_for_locks=1 } backup { retain_min = 50 retain_day s = 0 } " --autobackup n --available n f48e93bb-e3bb-4291-bd2f-8d622ef8019d/a351bc36-d741-4cf7-aa89-735c82b888e5' (cwd None) 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:51,110::misc::83::Storage.Misc.excCmd::(<lambda>) FAILED: <err> = ' device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n device-mapper: remove ioctl on failed: Device or resource busy\n Unable to deactivate f48e93bb--e3bb--4291--bd2f--8d622ef8019d-a351bc36--d741--4cf7--aa89--735c82b888e5 (253:508)\n'; <rc> = 5 85e6363c-1a1c-4e72-b1e7-0afd0daa67b9::DEBUG::2013-08-15 14:04:52,960::lvm::500::OperationMutex::(_invalidatelvs) Operation 'lvm reload operation' is holding the operation mutex, waiting...