[PATCH 00/10] drm/vkms: rework crc worker
Daniel Vetter
daniel at ffwll.ch
Tue Jun 18 08:56:47 UTC 2019
On Mon, Jun 17, 2019 at 11:49:04PM -0300, Rodrigo Siqueira wrote:
> On 06/12, Daniel Vetter wrote:
> > On Wed, Jun 12, 2019 at 10:28:41AM -0300, Rodrigo Siqueira wrote:
> > > Hi Daniel,
> > >
> > > First of all, thank you very much for your patchset.
> > >
> > > I tried to make a detailed review of your series, and you can see my
> > > comments in each patch. You’ll notice that I asked many things related
> > > to the DRM subsystem with the hope that I could learn a little bit
> > > more about DRM from your comments.
> > >
> > > Before you go through the review, I would like to start a discussion
> > > about the vkms race conditions. First, I have to admit that I did not
> > > understand the race conditions that you described before because I
> > > couldn’t reproduce them. Now, I'm suspecting that I could not
> > > experience the problem because I'm using QEMU with KVM; with this idea
> > > in mind, I suppose that we have two scenarios for using vkms in a
> > > virtual machine:
> > >
> > > * Scenario 1: The user has hardware virtualization support; in this
> > > case, it is a little bit harder to experience race conditions with
> > > vkms.
> > >
> > > * Scenario 2: Without hardware virtualization support, it is much
> > > easier to experience race conditions.
> > >
> > > With these two scenarios in mind, I conducted a bunch of experiments
> > > for trying to shed light on this issue. I did:
> > >
> > > 1. Enabled lockdep
> > >
> > > 2. Defined two different environments for testing both using QEMU with
> > > and without kvm. See below the QEMU hardware options:
> > >
> > > * env_kvm: -enable-kvm -daemonize -m 1G -smp cores=2,cpus=2
> > > * env_no_kvm: -daemonize -m 2G -smp cores=4,cpus=4
> > >
> > > 3. My test protocol: I) turn on the vm, II) clean /proc/lock_stat,
> > > III) execute kms_cursor_crc, III) save the lock_stat file, and IV)
> > > turn off the vm.
> > >
> > > 4. From the lockdep_stat, I just highlighted the row related to vkms
> > > and the columns holdtime-total and holdtime-avg
> > >
> > > I would like to highlight that the following data does not have any
> > > statistical approach, and its intention is solely to assist our
> > > discussion. See below the summary of the collected data:
> > >
> > > Summary of the experiment results:
> > >
> > > +----------------+----------------+----------------+
> > > | | env_kvm | env_no_kvm |
> > > + +----------------+----------------+
> > > | Test | Before | After | Before | After |
> > > +----------------+--------+-------+--------+-------+
> > > | kms_cursor_crc | S1 | S2 | M1 | M2 |
> > > +----------------+--------+-------+--------+-------+
> > >
> > > * Before: before apply this patchset
> > > * After: after apply this patchset
> > >
> > > -----------------------------------------+------------------+-----------
> > > S1: Without this patchset and with kvm | holdtime-total | holdtime-avg
> > > -----------------------------------------+----------------+-------------
> > > &(&vkms_out->lock)->rlock: | 21983.52 | 6.21
> > > (work_completion)(&vkms_state->crc_wo: | 20.47 | 20.47
> > > (wq_completion)vkms_crc_workq: | 3999507.87 | 3352.48
> > > &(&vkms_out->state_lock)->rlock: | 378.47 | 0.30
> > > (work_completion)(&vkms_state->crc_#2: | 3999066.30 | 2848.34
> > > -----------------------------------------+----------------+-------------
> > > S2: With this patchset and with kvm | |
> > > -----------------------------------------+----------------+-------------
> > > &(&vkms_out->lock)->rlock: | 23262.83 | 6.34
> > > (work_completion)(&vkms_state->crc_wo: | 8.98 | 8.98
> > > &(&vkms_out->crc_lock)->rlock: | 307.28 | 0.32
> > > (wq_completion)vkms_crc_workq: | 6567727.05 | 12345.35
> > > (work_completion)(&vkms_state->crc_#2: | 6567135.15 | 4488.81
> > > -----------------------------------------+----------------+-------------
> > > M1: Without this patchset and without kvm| |
> > > -----------------------------------------+----------------+-------------
> > > &(&vkms_out->state_lock)->rlock: | 4994.72 | 1.61
> > > &(&vkms_out->lock)->rlock: | 247190.04 | 39.39
> > > (work_completion)(&vkms_state->crc_wo: | 31.32 | 31.32
> > > (wq_completion)vkms_crc_workq: | 20991073.78 | 13525.18
> > > (work_completion)(&vkms_state->crc_#2: | 20988347.18 | 11904.90
> > > -----------------------------------------+----------------+-------------
> > > M2: With this patchset and without kvm | |
> > > -----------------------------------------+----------------+-------------
> > > (wq_completion)vkms_crc_workq: | 42819161.68 | 36597.57
> > > &(&vkms_out->lock)->rlock: | 251257.06 | 35.80
> > > (work_completion)(&vkms_state->crc_wo: | 69.37 | 69.37
> > > &(&vkms_out->crc_lock)->rlock: | 3620.92 | 1.54
> > > (work_completion)(&vkms_state->crc_#2: | 42803419.59 | 24306.31
> > >
> > > First, I analyzed the scenarios with KVM (S1 and S2); more
> > > specifically, I focused on these two classes:
> > >
> > > 1. (work_completion)(&vkms_state->crc_wo
> > > 2. (work_completion)(&vkms_state->crc_#2
> > >
> > > After taking a look at the data, it looks like that this patchset
> > > greatly reduces the hold time average for crc_wo. On the other hand,
> > > it increases the hold time average for crc_#2. I didn’t understand
> > > well the reason for the difference. Could you help me here?
> >
> > So there's two real locks here from our code, the ones you can see as
> > spinlocks:
> >
> > &(&vkms_out->state_lock)->rlock: | 4994.72 | 1.61
> > &(&vkms_out->lock)->rlock: | 247190.04 | 39.39
> >
> > All the others are fake locks that the workqueue adds, which only exist in
> > lockdep. They are used to catch special kinds of deadlocks like the below:
> >
> > thread A:
> > 1. mutex_lock(mutex_A)
> > 2. flush_work(work_A)
> >
> > thread B
> > 1. running work_A:
> > 2. mutex_lock(mutex_A)
> >
> > thread B can't continue becuase mutex_A is already held by thread A.
> > thread A can't continue because thread B is blocked and the work never
> > finishes.
> > -> deadlock
> >
> > I haven't checked which is which, but essentially what you measure with
> > the hold times of these fake locks is how long a work execution takes on
> > average.
> >
> > Since my patches are supposed to fix races where the worker can't keep up
> > with the vblank hrtimer, then the average worker will (probably) do more,
> > so that going up is expected. I think.
> >
> > I'm honestly not sure what's going on here, never looking into this in
> > detail.
> >
> > > When I looked for the second set of scenarios (M1 and M2, both without
> > > KVM), the results look much more distant; basically, this patchset
> > > increased the hold time average. Again, could you help me understand a
> > > little bit better this issue?
> > >
> > > Finally, after these tests, I have some questions:
> > >
> > > 1. VKMS is a software-only driver; because of this, how about defining
> > > a minimal system resource for using it?
> >
> > No idea, in reality it's probably "if it fails too often, buy faster cpu".
> > I do think we should make the code robust against a slow cpu, since atm
> > that's needed even on pretty fast machines (because our blending code is
> > really, really slow and inefficient).
> >
> > > 2. During my experiments, I noticed that running tests with a VM that
> > > uses KVM had consistent results. For example, kms_flip never fails
> > > with QEMU+KVM; however, without KVM, two or three tests failed (one of
> > > them looks random). If we use vkms for test DRM stuff, should we
> > > recommend the use of KVM?
> >
> > What do you mean without kvm? In general running without kvm shouldn't be
> > slower, so I'm a bit confused ... I'm running vgem directly on the
> > machine, by booting into new kernels (and controlling the machine over the
> > network).
>
> Sorry, I should have detailed my point.
>
> Basically, I do all my testing with vkms in QEMU VM. In that sense, I
> did some experiments which I enabled and disabled the KVM (i.e., flag
> '-enable-kvm') to check the vkms behaviour in these two scenarios. I
> noticed that the tests are consistent when I use the '-enable-kvm' flag,
> in that context it seemed a good idea to recommend the use of KVM for
> those users who want to test vkms with igt. Anyway, don't worry about
> that I'll try to add more documentation for vkms in the future and in
> that time we discuss about this again.
Ah, qemu without kvm is going to use software emulation for a lot of the
kernel stuff. That's going to be terribly slow indeed.
> Anyway, from my side, I think we should merge this series. Do you want
> to prepare a V2 with the fixes and maybe update the commit messages by
> using some of your explanations? Or, if you want, I can fix the tiny
> details and merge it.
I'm deeply burried in my prime cleanup/doc series right now, so will take
a few days until I get around to this again. If you want, please go ahead
with merging.
btw if you edit a patch when merging, please add a comment about that to
the commit message. And ime it's best to only augment the commit message
and maybe delete an unused variable or so that got forgotten. For
everything more better to do the edits and resubmit.
Thanks, Daniel
>
> > -Daniel
> >
> > > Best Regards
> > >
> > > On Thu, Jun 6, 2019 at 7:28 PM Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
> > > >
> > > > Hi all,
> > > >
> > > > This here is the first part of a rework for the vkms crc worker. I think
> > > > this should fix all the locking/races/use-after-free issues I spotted in
> > > > the code. There's more work we can do in the future as a follow-up:
> > > >
> > > > - directly access vkms_plane_state->base in the crc worker, with this
> > > > approach in this series here that should be safe now. Follow-up patches
> > > > could switch and remove all the separate crc_data infrastructure.
> > > >
> > > > - I think some kerneldoc for vkms structures would be nice. Documentation
> > > > the various functions is probably overkill.
> > > >
> > > > - Implementing a more generic blending engine, as prep for adding more
> > > > pixel formats, more planes, and more features in general.
> > > >
> > > > Tested with kms_pipe_crc, kms_flip and kms_cursor_crc. Seems to not make
> > > > things worse, but I didn't do a full run.
> > > >
> > > > Cheers, Daniel
> > > >
> > > > Daniel Vetter (10):
> > > > drm/vkms: Fix crc worker races
> > > > drm/vkms: Use spin_lock_irq in process context
> > > > drm/vkms: Rename vkms_output.state_lock to crc_lock
> > > > drm/vkms: Move format arrays to vkms_plane.c
> > > > drm/vkms: Add our own commit_tail
> > > > drm/vkms: flush crc workers earlier in commit flow
> > > > drm/vkms: Dont flush crc worker when we change crc status
> > > > drm/vkms: No _irqsave within spin_lock_irq needed
> > > > drm/vkms: totally reworked crc data tracking
> > > > drm/vkms: No need for ->pages_lock in crc work anymore
> > > >
> > > > drivers/gpu/drm/vkms/vkms_crc.c | 74 +++++++++-------------------
> > > > drivers/gpu/drm/vkms/vkms_crtc.c | 80 ++++++++++++++++++++++++++-----
> > > > drivers/gpu/drm/vkms/vkms_drv.c | 35 ++++++++++++++
> > > > drivers/gpu/drm/vkms/vkms_drv.h | 24 +++++-----
> > > > drivers/gpu/drm/vkms/vkms_plane.c | 8 ++++
> > > > 5 files changed, 146 insertions(+), 75 deletions(-)
> > > >
> > > > --
> > > > 2.20.1
> > > >
> > > > _______________________________________________
> > > > dri-devel mailing list
> > > > dri-devel at lists.freedesktop.org
> > > > https://lists.freedesktop.org/mailman/listinfo/dri-devel
> > >
> > >
> > >
> > > --
> > >
> > > Rodrigo Siqueira
> > > https://siqueira.tech
> >
> > --
> > Daniel Vetter
> > Software Engineer, Intel Corporation
> > http://blog.ffwll.ch
>
> --
> Rodrigo Siqueira
> https://siqueira.tech
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
More information about the dri-devel
mailing list