[Intel-gfx] [RFC PATCH 1/4] drm/i915: Drop user contexts on driver remove

Thu May 28 14:00:15 UTC 2020

Hi Chris,

On Thu, 2020-05-28 at 14:41 +0100, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2020-05-28 14:34:42)
> > On 28/05/2020 13:10, Janusz Krzysztofik wrote:
> > > Hi Tvrtko,
> > > 
> > > On Thu, 2020-05-28 at 11:14 +0100, Tvrtko Ursulin wrote:
> > > > On 18/05/2020 19:17, Janusz Krzysztofik wrote:
> > > > > Contexts associated with open device file descriptors together with
> > > > > their assigned address spaces are now closed on device file close.  On
> > > > 
> > > > i915_gem_driver_remove looks like module unload to me, not device file
> > > > close. So..
> > > 
> > > Not only module unload ...
> > > 
> > > > > address space closure its associated DMA mappings are revoked.  If the
> > > > > device is removed while being open, subsequent attempts to revoke
> > > > > those mappings while closing the device file descriptor may may be
> > > > > judged by intel-iommu code as a bug and result in kernel panic.
> > > > > 
> > > > > Since user contexts become useless after the device is no longer
> > > > > available, drop them on device removal.
> > > > > 
> > > > > <4> [36.900985] ------------[ cut here ]------------
> > > > > <2> [36.901005] kernel BUG at drivers/iommu/intel-iommu.c:3717!
> > > > > <4> [36.901105] invalid opcode: 0000 [#1] PREEMPT SMP NOPTI
> > > > > <4> [36.901117] CPU: 0 PID: 39 Comm: kworker/u8:1 Tainted: G     U  W         5.7.0-rc5-CI-CI_DRM_8485+ #1
> > > > > <4> [36.901133] Hardware name: Intel Corporation Elkhart Lake Embedded Platform/ElkhartLake LPDDR4x T3 CRB, BIOS EHLSFWI1.R00.1484.A00.1911290833 11/29/2019
> > > > > <4> [36.901250] Workqueue: i915 __i915_vm_release [i915]
> > > > > <4> [36.901264] RIP: 0010:intel_unmap+0x1f5/0x230
> > > > > <4> [36.901274] Code: 01 e8 9f bc a9 ff 85 c0 74 09 80 3d df 60 09 01 00 74 19 65 ff 0d 13 12 97 7e 0f 85 fc fe ff ff e8 82 b0 95 ff e9 f2 fe ff ff <0f> 0b e8 d4 bd a9 ff 85 c0 75 de 48 c7 c2 10 84 2c 82 be 54 00 00
> > > > > <4> [36.901302] RSP: 0018:ffffc900001ebdc0 EFLAGS: 00010246
> > > > > <4> [36.901313] RAX: 0000000000000000 RBX: ffff8882561dd000 RCX: 0000000000000000
> > > > > <4> [36.901324] RDX: 0000000000001000 RSI: 00000000ffd9c000 RDI: ffff888274c94000
> > > > > <4> [36.901336] RBP: ffff888274c940b0 R08: 0000000000000000 R09: 0000000000000001
> > > > > <4> [36.901348] R10: 000000000a25d812 R11: 00000000112af2d4 R12: ffff888252c70200
> > > > > <4> [36.901360] R13: 00000000ffd9c000 R14: 0000000000001000 R15: ffff8882561dd010
> > > > > <4> [36.901372] FS:  0000000000000000(0000) GS:ffff888278000000(0000) knlGS:0000000000000000
> > > > > <4> [36.901386] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > > > <4> [36.901396] CR2: 00007f06def54950 CR3: 0000000255844000 CR4: 0000000000340ef0
> > > > > <4> [36.901408] Call Trace:
> > > > > <4> [36.901418]  ? process_one_work+0x1de/0x600
> > > > > <4> [36.901494]  cleanup_page_dma+0x37/0x70 [i915]
> > > > > <4> [36.901573]  free_pd+0x9/0x20 [i915]
> > > > > <4> [36.901644]  gen8_ppgtt_cleanup+0x59/0xc0 [i915]
> > > > > <4> [36.901721]  __i915_vm_release+0x14/0x30 [i915]
> > > > > <4> [36.901733]  process_one_work+0x268/0x600
> > > > > <4> [36.901744]  ? __schedule+0x307/0x8d0
> > > > > <4> [36.901756]  worker_thread+0x37/0x380
> > > > > <4> [36.901766]  ? process_one_work+0x600/0x600
> > > > > <4> [36.901775]  kthread+0x140/0x160
> > > > > <4> [36.901783]  ? kthread_park+0x80/0x80
> > > > > <4> [36.901792]  ret_from_fork+0x24/0x50
> > > > > <4> [36.901804] Modules linked in: mei_hdcp i915 x86_pkg_temp_thermal coretemp crct10dif_pclmul crc32_pclmul ghash_clmulni_intel ax88179_178a usbnet mii mei_me mei prime_numbers intel_lpss_pci
> > > > > <4> [36.901857] ---[ end trace 52d1b4d81f8d1ea7 ]---
> > > > > 
> > > > > Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik at linux.intel.com>
> > > > > ---
> > > > >    drivers/gpu/drm/i915/gem/i915_gem_context.c | 38 +++++++++++++++++++++
> > > > >    drivers/gpu/drm/i915/gem/i915_gem_context.h |  1 +
> > > > >    drivers/gpu/drm/i915/i915_gem.c             |  2 ++
> > > > >    3 files changed, 41 insertions(+)
> > > > > 
> > > > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c
> > > > > index 900ea8b7fc8f..0096a69fbfd3 100644
> > > > > --- a/drivers/gpu/drm/i915/gem/i915_gem_context.c
> > > > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c
> > > > > @@ -927,6 +927,44 @@ void i915_gem_driver_release__contexts(struct drm_i915_private *i915)
> > > > >     rcu_barrier(); /* and flush the left over RCU frees */
> > > > >    }
> > > > >    
> > > > > +void i915_gem_driver_remove__contexts(struct drm_i915_private *i915)
> > > > > +{
> > > > > +   struct i915_gem_context *ctx, *cn;
> > > > > +
> > > > > +   list_for_each_entry_safe(ctx, cn, &i915->gem.contexts.list, link) {
> > > > > +           struct drm_i915_file_private *file_priv = ctx->file_priv;
> > > > > +           struct i915_gem_context *entry;
> > > > > +           unsigned long int id;
> > > > > +
> > > > > +           if (i915_gem_context_is_closed(ctx) || IS_ERR(file_priv))
> > > > > +                   continue;
> > > > > +
> > > > > +           xa_for_each(&file_priv->context_xa, id, entry) {
> > > > 
> > > > ... how is driver unload possible with open drm file descriptors, or
> > > > active contexts?
> > > 
> > > ... but also PCI driver unbind or PCI device remove, with the module
> > > still loaded.  That may perfectly happen even if a device file
> > > descriptor is still kept open.
> > 
> > I see. What do we do, or plan to do, with those left open drm fds after 
> > the driver is unbound from the device? Is there a path connected to keep 
> > saying -ENODEV (Or is a different errno standard for this case, like 
> > ENXIO?) from that point onward for everything done with that fd. So 
> > userspace couldn't do anything more with it, attempt to create a new 
> > context etc. Is the DRM core handling this?
> 
> We mark the device as wedged, so any attempt to use the GPU results in
> -EIO. There's a loose edge in the faulthandler, but marking the ggtt as
> closed and SIGBUS on trying to fault in an non-existent page takes care
> of that.
> 
> There's a very expensive srcu in every drm_ioctl for checking unplugged
> that we don't need because we already do the -EIO.

Thanks for help with that explanation.

> 
> In particular this patch is worthless, or worse misdirection. The issue
> is not the contexts or page directories and this is not fixing the
> problem and only a few of the symptoms.

Please note this patch is only a part of a series.  My findings are
that it is needed for the solution the series proposes to be effective.
The idea is to release all DMA mappings still on driver remove.  I
would really appreciate your comments on that idea and maybe on other
patches of the series as well.

Thanks,
Janusz

> -Chris