- 14 Mar, 2017 2 commits
-
-
Daniel Vetter authored
The trouble we have is that we can't really test all the shrinker recursion stuff exhaustively in BAT because any kind of thrashing stress test just takes too long. But that leaves a really big gap open, since shrinker recursions are one of the most annoying bugs. Now lockdep already has support for checking allocation deadlocks: - Direct reclaim paths are marked up with lockdep_set_current_reclaim_state() and lockdep_clear_current_reclaim_state(). - Any allocation paths are marked with lockdep_trace_alloc(). If we simply mark up our debugfs with the reclaim annotations, any code and locks taken in there will automatically complete the picture with any allocation paths we already have, as long as we have a simple testcase in BAT which throws out a few objects using this interface. Not stress test or thrashing needed at all. v2: Need to EXPORT_SYMBOL_GPL to make it compile as a module. v3: Fixup rebase fail (spotted by Chris). Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: linux-kernel@vger.kernel.org Reviewed-by:
Chris Wilson <chris@chris-wilson.co.uk> Acked-by:
Peter Zijlstra (Intel) <peterz@infradead.org> Link: http://patchwork.freedesktop.org/patch/msgid/20170312205340.16202-1-daniel.vetter@ffwll.chSigned-off-by:
Daniel Vetter <daniel.vetter@intel.com>
-
Jani Nikula authored
The main thing are the DDI ports. If there's a VBT that says there are no outputs, we should trust that, and not have semi-random defaults. Unfortunately, the defaults have resulted in some Chromebooks without VBT to rely on this behaviour, so we split out the defaults for the missing VBT case. Reviewed-by:
Manasi Navare <manasi.d.navare@intel.com> Cc: Manasi Navare <manasi.d.navare@intel.com> Cc: Ville Syrjälä <ville.syrjala@linux.intel.com> Signed-off-by:
Jani Nikula <jani.nikula@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/95c26079ff640d43f53b944f17e9fc356b36daec.1489152288.git.jani.nikula@intel.com
-
- 13 Mar, 2017 19 commits
-
-
Ville Syrjälä authored
Use I915_{READ,WRITE}_FW() for updating the DSPARB registers on VLV/CHV. This is less expesive as we can grab the uncore.lock across the entire sequence of reads and writes instead of each register access grabbing it. This also allows us to eliminate the dsparb lock entirely as the uncore.lock now effectively protects the contents of the DSPARB registers. v2: Add a note that interrupts are already disabled (Chris) Cc: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by:Ville Syrjälä <ville.syrjala@linux.intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/20170309154434.29303-6-ville.syrjala@linux.intel.comReviewed-by:
Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
-
Ville Syrjälä authored
Optimize the plane register accesses a little bit by grabbing the uncore lock manually across the entire pile of accesses and using I915_READ_FW(). This helps keep the pipe update vblank evade critical section below our 100 usec deadline, particularly with lockdep enabled. And in general we want to keep that critical section as short as possible as it's executed with interrupts disabled. Not all plane updates currently happen from within the vblank evade critical section, so we must use the irqsave/irqrestore variants of the spinlock functions in the plane hooks. Signed-off-by:
-
Ville Syrjälä authored
Pull all the plane register writes closer together to avoid having a lot of unrelated stuff in between them. This will make things more clear once we'll grab the uncore lock around the entire bunch. Also in the future we might even consider moving more of the register value computation out from the plane update hooks. This should make that easier to do. Signed-off-by:
-
Ville Syrjälä authored
Replace __raw_i915_read32() with I915_READ_FW() in the workaround for the SKL+ scanline counter hardware fail. The two are the same thing but everyone else uses I915_READ_FW() so let's follow suit. Signed-off-by:
Mika Kahola <mika.kahola@intel.com> Reviewed-by:
-
Ville Syrjälä authored
Optimize the multi-register read in i915_get_vblank_counter() a little bit by grabbing the uncore lock manually and using I915_READ_FW(). Signed-off-by:
-
Ville Syrjälä authored
Check that the sink really declared 12bpc support before we enable it. This should not actually never happen since it's mandatory for HDMI sinks to support 12bpc if they support any deep color modes. But reality disagrees with the theory and there are actually sinks in the wild that violate the spec. v2: Fix the output_types check Update commit message to state that these things are in fact real Cc: stable@vger.kernel.org Cc: Nicholas Sielicki <nicholas.sielicki@gmail.com> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=99250Signed-off-by:Shashank Sharma <shashank.sharma@intel.com>
-
Chris Wilson authored
i915_drpc_info missed covering a few register read with the runtime pm wakelock. Be simple and cover the entire function with a single wakelock so that new additions are not similarly missed in future. WARNING: CPU: 2 PID: 1334 at drivers/gpu/drm/i915/intel_drv.h:1743 gen6_read32+0x192/0x1e0 [i915] RPM wakelock ref not held during HW access Modules linked in: rpcsec_gss_krb5 nfsv4 dns_resolver netconsole nfsd auth_rpcgss ipmi_watchdog ipmi_poweroff ipmi_devintf ipmi_msghandler overlay btrfs xor raid6_pq dm_mod sg sd_mod snd_hda_codec_hdmi snd_hda_codec_realtek snd_hda_codec_generic ata_generic pata_acpi intel_rapl x86_pkg_temp_thermal intel_powerclamp coretemp snd_hda_intel kvm_intel snd_hda_codec kvm eeepc_wmi irqbypass snd_hda_core crct10dif_pclmul crc32_pclmul crc32c_intel asus_wmi sparse_keymap ghash_clmulni_intel snd_hwdep i915 rfkill ppdev pcbc aesni_intel ata_piix crypto_simd glue_helper snd_pcm pata_via cryptd pcspkr snd_timer drm_kms_helper syscopyarea snd sysfillrect libata sysimgblt fb_sys_fops soundcore shpchp drm wmi parport_pc parport tpm_infineon video CPU: 2 PID: 1334 Comm: php5 Not tainted 4.10.0-rc8-01615-g1f58c8e7 #1 Hardware name: System manufacturer System Product Name/P8H67-M PRO, BIOS 1002 04/01/2011 Call Trace: dump_stack+0x63/0x8a __warn+0xcb/0xf0 warn_slowpath_fmt+0x4f/0x60 ? seq_vprintf+0x35/0x50 gen6_read32+0x192/0x1e0 [i915] i915_drpc_info+0x55d/0x990 [i915] seq_read+0xf2/0x3b0 full_proxy_read+0x51/0x80 __vfs_read+0x28/0x130 ? security_file_permission+0x9b/0xc0 ? rw_verify_area+0x4e/0xb0 vfs_read+0xa8/0x170 SyS_read+0x46/0xa0 entry_SYSCALL_64_fastpath+0x1a/0xa9 RIP: 0033:0x7fd97bf175a0 RSP: 002b:00007ffdf730db68 EFLAGS: 00000246 ORIG_RAX: 0000000000000000 RAX: ffffffffffffffda RBX: 00007fd978028738 RCX: 00007fd97bf175a0 RDX: 0000000000002000 RSI: 00007fd97740e0d8 RDI: 0000000000000005 RBP: 0000000000000001 R08: 0000000000e97840 R09: 00007fd977ef8d58 R10: 0000000000000027 R11: 0000000000000246 R12: 00007fd977ef8d58 R13: 0000000000000000 R14: 0000000000eb4640 R15: 0000000000000000 Reported-by:
kernel test robot <xiaolong.ye@intel.com> Signed-off-by:
Mika Kuoppala <mika.kuoppala@intel.com>
-
Chris Wilson authored
gen6_sanitize_rps_pm_mask() is small enough that inlining it shrinks the object code. v2: Use const markup Signed-off-by:
-
Chris Wilson authored
The patch 791ff39a: "drm/i915: Live testing for context execution" from Feb 13, 2017, leads to the following static checker warning: drivers/gpu/drm/i915/selftests/i915_gem_context.c:347 igt_ctx_exec() error: 'file' dereferencing possible ERR_PTR() Reported-by:
Dan Carpenter <dan.carpenter@oracle.com> Fixes: 791ff39a ("drm/i915: Live testing for context execution") Signed-off-by:
Tvrtko Ursulin <tvrtko.ursulin@intel.com>
-
Chris Wilson authored
The patch 6e32ab3d: "drm/i915: Fill different pages of the GTT" from Feb 13, 2017, leads to the following static checker warning: drivers/gpu/drm/i915/selftests/i915_gem_gtt.c:583 walk_hole() error: 'vma' dereferencing possible ERR_PTR() Reported-by:
Matthew Auld <matthew.auld@intel.com>
-
Chris Wilson authored
If the object is coherent, we can simply update the cache domain on the whole object rather than calculate the before/after clflushes. The advantage is that we then get correct tracking of ellided flushes when changing coherency later. Testcase: igt/gem_pwrite_snooped Signed-off-by:
Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
-
Chris Wilson authored
The REDIRECT_TO_GUC bit is a strange beast as it is a disable bit - setting the bit in the pm interrupt generation stops the interrupt going to the guc (not sending it to the guc as the name implies). To help the reader rename it to DISABLE_REDIRECT_TO_GUC so that we keep the bspec greppable name without it being as confusing! Signed-off-by:
Sagar Arun Kamble <sagar.a.kamble@intel.com>
-
Maarten Lankhorst authored
Calculating the max pixel rate requires the new state, so use it there. Signed-off-by:
Daniel Vetter <daniel.vetter@ffwll.ch>
-
Maarten Lankhorst authored
Add a big fat warning in __intel_display_resume that the old state is invalid, and use the correct state everywhere. Signed-off-by:
-
Maarten Lankhorst authored
The watermark code needs to look at the new allocations, so use for_each_new_crtc_in_state everywhere. Signed-off-by:
-
Maarten Lankhorst authored
Use for_each_new_plane_in_state, only the new state is needed. Signed-off-by:
-
Maarten Lankhorst authored
Use for_each_new_connector_in_state instead of for_each_connector_in_state. Also make the function static, it's only used inside intel_ddi.c Signed-off-by:
-
Jani Nikula authored
We don't use the error return for anything other than reporting and logging that there is no VBT. We can pull the logging in the function, and remove the error status return. Moreover, if we needed the information for something later on, we'd probably be better off storing the bit in dev_priv, and using it where it's needed, instead of using the error return. While at it, improve the comments. Cc: Manasi Navare <manasi.d.navare@intel.com> Cc: Ville Syrjälä <ville.syrjala@linux.intel.com> Reviewed-by:
-
Daniel Vetter authored
Merge tag 'topic/designware-baytrail-2017-03-02' of git://anongit.freedesktop.org/git/drm-intel into drm-intel-next-queued Baytrail PMIC vs. PMU race fixes from Hans de Goede This time the right version (v4), with the compile fix. Signed-off-by:
-
- 12 Mar, 2017 3 commits
-
-
Sagar Arun Kamble authored
Different state is to be maintained for rps.pm_intrmsk_mbz for GuC and Execlists. Updating it inside guc_interrupts_* routines as in those routines GuC load/submission params are sanitized and it should not be set based on HAS_GUC_SCHED during intel_irq_init. Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Oscar Mateo <oscar.mateo@intel.com> Cc: Radoslaw Szwichtenberg <radoslaw.szwichtenberg@intel.com> Reviewed-by:
-
Sagar Arun Kamble authored
"pm_intr_keep" is not conveying the intent that it is bitmask of interrupts that must be zero(mbz) in GEN6_PMINTRMSK. Name it "pm_intrmsk_mbz". Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Radoslaw Szwichtenberg <radoslaw.szwichtenberg@intel.com> Reviewed-by:
-
Sagar Arun Kamble authored
Like capture of GuC interrupts while enabling GuC submission, release them while disabling GuC submission. Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Oscar Mateo <oscar.mateo@intel.com> Reviewed-by:
-
- 10 Mar, 2017 9 commits
-
-
Chris Wilson authored
To make our adjustments to RPS requires taking a mutex and potentially sleeping for an unknown duration - until we have completed our adjustments further RPS interrupts are immaterial (they are based on stale thresholds) and we can safely ignore them. Signed-off-by:
Mika Kuoppala <mika.kuoppala@linux.intel.com>
-
Chris Wilson authored
Currently, we sum the render and media cycles (on different engines) to compute a percentage - but we fail to factor in the duplication into the threshold calculations. This makes us very eager to upclock! If we just consider the maximum busy cycles of either counter, we should have an accurate reflection on whether there are cycles to spare to handle the workload at this frequency. Signed-off-by:
-
Chris Wilson authored
On Baytrail, we manually calculate busyness over the evaluation interval to avoid issues with miscaluations with RC6 enabled. However, it turns out that the DOWN_EI interrupt generator is completely bust - it operates in two modes, continuous or never. Neither of which are conducive to good behaviour. Stop unmask the DOWN_EI interrupt and just compute everything from the UP_EI which does seem to correspond to the desired interval. v2: Fixup gen6_rps_pm_mask() as well v3: Inline vlv_c0_above() to combine the now identical elapsed calculation for up/down and simplify the threshold testing Fixes: 43cf3bf0 ("drm/i915: Improved w/a for rps on Baytrail") Signed-off-by:
-
Chris Wilson authored
Sometimes we want to explicitly page out all available objects from igt, i.e. call i915_gem_shrink_all() and check that subsequent operations succeed. This adds DROP_SHRINK_ALL [0x8] to the set of flags for debugfs/i915_drop_caches for that purpose. Signed-off-by:
-
Chris Wilson authored
We only need to clflush those cachelines that we have validated to be read by the GPU. Userspace typically fills the batch length in correctly, the exceptions tend to be explicit tests within igt. v2: Use ptr_mask_bits() to make Mika happy v3: cmd is not advanced on MI_BBE, so make sure to include an extra dword in the clflush. Signed-off-by:
-
Tvrtko Ursulin authored
In commit 003342a5 ("drm/i915: Keep track of active forcewake domains in a bitmask") I forgot to adjust the newly introduce fw_domains_active state across reset. This caused the assert_forcewakes_inactive to trigger during suspend and resume if there were user held forcewakes. v2: Bitmask checks are required since vfuncs are not always present. v3: Move bitmask tracking to get/put vfunc for simplicity. (Chris Wilson) Signed-off-by:
-
Shashank Sharma authored
In I915 driver, there are many places where variable name for intel_encoder object is given as 'intel_encoder' whereas it would make more sense to call it just 'encoder' when possible. This patch does this cleanup in file intel_ddi.c. PS: There are few functions where both drm_encoder and intel_encoder are present. For such functions, this patch does nothing. Suggested-by:
Ander Conselvan De Oliveira <conselvan2@gmail.com> Signed-off-by:
Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1489067021-4709-1-git-send-email-shashank.sharma@intel.com
-
Andrew Morton authored
gcc-4.4.4 has issues with anonymous union initializers. In file included from drivers/gpu/drm/i915/selftests/i915_selftest.c:68: drivers/gpu/drm/i915/selftests/i915_mock_selftests.h:11: error: unknown field 'mock' specified in initializer drivers/gpu/drm/i915/selftests/i915_mock_selftests.h:11: warning: missing braces around initializer drivers/gpu/drm/i915/selftests/i915_mock_selftests.h:11: warning: (near initialization for 'mock_selftests[0].<anonymous>') drivers/gpu/drm/i915/selftests/i915_mock_selftests.h:12: error: unknown field 'mock' specified in initializer drivers/gpu/drm/i915/selftests/i915_mock_selftests.h:13: error: unknown field 'm ... Work around this. Fixes: 953c7f82 ("drm/i915: Provide a hook for selftests") Cc: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Reviewed-by:
-
Tvrtko Ursulin authored
In order to ensure no missed interrupts we must first re-direct the interrupts to GuC, and only then re-submit the requests to be replayed after a GPU reset. Otherwise context switch can fire before GuC has been set up to receive it triggering more hangs. v2: Rebase. Signed-off-by:
-
- 09 Mar, 2017 7 commits
-
-
Chris Wilson authored
There is no easily digestible single self-refresh status bit, so don't report one for debugfs/i915_sr_status on gen9+. For the moment this avoids a read of the non-existent WM1_LP_ILK register. Signed-off-by:
-
Sagar Arun Kamble authored
Driver needs to ensure that it doesn't mask the PM interrupts, which are unmasked/needed by GuC firmware. For that, Driver maintains a bitmask of interrupts to be kept unmasked, pm_intr_keep. pm_intr_keep was determined across GuC load. GuC gets loaded in different scenarios and it is not going to change the pm_intr_keep so this patch moves its setup to intel_irq_init. This patch fixes incorrect RPS masking leading to UP interrupts triggered even when at cur_freq=max and inversly for Down interrupts. Cc: Radoslaw Szwichtenberg <radoslaw.szwichtenberg@intel.com> Cc: Arkadiusz Hiler <arkadiusz.hiler@intel.com> Cc: Michal Wajdeczko <michal.wajdeczko@intel.com> Cc: Michal Winiarski <michal.winiarski@intel.com> Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Signed-off-by:
-
Maarten Lankhorst authored
printks are slow so we should not be doing them from the vblank evade critical section. These could explain why we sometimes seem to blow past our 100 usec deadline. The problem has been there ever since commit c331879c ("drm/i915: skylake sprite plane scaling using shared scalers.") but it may not have been readily visible until commit e1edbd44 ("drm/i915: Complain if we take too long under vblank evasion.") increased our chances of noticing it. Signed-off-by:
-
Matthew Auld authored
Add a selftest to exercise evicting neighbouring nodes that conflict due to page colouring in the GTT. v2: add a peppering of comments Signed-off-by:
-
Matthew Auld authored
It looks like we were incorrectly comparing vma->node against itself instead of the target node, when evicting for a node on systems where we need guard pages between regions with different cache domains. As a consequence we can end up trying to needlessly evict neighbouring nodes, even if they have the same cache domain, and if they were pinned we would fail the eviction. Fixes: 625d988a ("drm/i915: Extract reserving space in the GTT to a helper") Signed-off-by:
-
Matthew Auld authored
For our fake dma objects we can leak the underlying gem object if we fail to pin our "backing storage". [ 39.952618] ============================================================================= [ 39.952625] BUG mock_object (Tainted: G U ): Objects remaining in mock_object on __kmem_cache_shutdown() [ 39.952629] ----------------------------------------------------------------------------- [ 39.952633] Disabling lock debugging due to kernel taint [ 39.952635] INFO: Slab 0xffffea00086c6a00 objects=21 used=1 fp=0xffff88021b1abc00 flags=0x5fff8000008100 [ 39.952640] CPU: 1 PID: 1258 Comm: drv_selftest Tainted: G BU 4.10.0+ #46 [ 39.952641] Hardware name: Apple Inc. MacBookPro11,1/Mac-189A3D4F975D5FFC, BIOS MBP111.88Z.0138.B17.1602221718 02/22/2016 [ 39.952642] Call Trace: [ 39.952648] dump_stack+0x4d/0x6f [ 39.952651] slab_err+0x9d/0xb0 [ 39.952654] ? ksm_migrate_page+0xe0/0xe0 [ 39.952657] ? on_each_cpu_cond+0x9a/0xc0 [ 39.952658] ? __kmalloc+0x1af/0x1c0 [ 39.952660] ? __kmem_cache_shutdown+0x173/0x3e0 [ 39.952661] __kmem_cache_shutdown+0x196/0x3e0 [ 39.952664] kmem_cache_destroy+0xa0/0x150 [ 39.952708] mock_device_release+0x113/0x140 [i915] [ 39.952726] drm_dev_release+0x20/0x40 [drm] [ 39.952735] drm_dev_unref+0x23/0x30 [drm] [ 39.952768] i915_gem_gtt_mock_selftests+0x55/0x70 [i915] [ 39.952803] __run_selftests+0x169/0x1c0 [i915] [ 39.952805] ? 0xffffffffa0151000 [ 39.952840] i915_mock_selftests+0x30/0x60 [i915] [ 39.952869] i915_init+0xc/0x78 [i915] [ 39.952870] ? 0xffffffffa0151000 [ 39.952872] do_one_initcall+0x43/0x170 [ 39.952874] ? __vunmap+0x81/0xd0 [ 39.952875] ? kmem_cache_alloc_trace+0x37/0x170 [ 39.952877] ? do_init_module+0x27/0x1f8 [ 39.952879] do_init_module+0x5f/0x1f8 [ 39.952881] load_module+0x2423/0x29b0 [ 39.952882] ? __symbol_put+0x40/0x40 [ 39.952885] ? kernel_read_file+0x1a3/0x1c0 [ 39.952887] SYSC_finit_module+0xbc/0xf0 [ 39.952889] SyS_finit_module+0xe/0x10 [ 39.952892] entry_SYSCALL_64_fastpath+0x13/0x94 v2: use onion teardown and favour i915_gem_object_put Fixes: 8d28ba45 ("drm/i915: Exercise filling the top/bottom portions of the ppgtt") Signed-off-by:
-
Chris Wilson authored
If we allow the user to convert a GTT mmap address into a userptr, we may end up in recursion hell, where currently we hit a mutex deadlock but other possibilities include use-after-free during the unbind/cancel_userptr. [ 143.203989] gem_userptr_bli D 0 902 898 0x00000000 [ 143.204054] Call Trace: [ 143.204137] __schedule+0x511/0x1180 [ 143.204195] ? pci_mmcfg_check_reserved+0xc0/0xc0 [ 143.204274] schedule+0x57/0xe0 [ 143.204327] schedule_timeout+0x383/0x670 [ 143.204374] ? trace_hardirqs_on_caller+0x187/0x280 [ 143.204457] ? trace_hardirqs_on_thunk+0x1a/0x1c [ 143.204507] ? usleep_range+0x110/0x110 [ 143.204657] ? irq_exit+0x89/0x100 [ 143.204710] ? retint_kernel+0x2d/0x2d [ 143.204794] ? trace_hardirqs_on_caller+0x187/0x280 [ 143.204857] ? _raw_spin_unlock_irq+0x33/0x60 [ 143.204944] wait_for_common+0x1f0/0x2f0 [ 143.205006] ? out_of_line_wait_on_atomic_t+0x170/0x170 [ 143.205103] ? wake_up_q+0xa0/0xa0 [ 143.205159] ? flush_workqueue_prep_pwqs+0x15a/0x2c0 [ 143.205237] wait_for_completion+0x1d/0x20 [ 143.205292] flush_workqueue+0x2e9/0xbb0 [ 143.205339] ? flush_workqueue+0x163/0xbb0 [ 143.205418] ? __schedule+0x533/0x1180 [ 143.205498] ? check_flush_dependency+0x1a0/0x1a0 [ 143.205681] i915_gem_userptr_mn_invalidate_range_start+0x1c7/0x270 [i915] [ 143.205865] ? i915_gem_userptr_dmabuf_export+0x40/0x40 [i915] [ 143.205955] __mmu_notifier_invalidate_range_start+0xc6/0x120 [ 143.206044] ? __mmu_notifier_invalidate_range_start+0x51/0x120 [ 143.206123] zap_page_range_single+0x1c7/0x1f0 [ 143.206171] ? unmap_single_vma+0x160/0x160 [ 143.206260] ? unmap_mapping_range+0xa9/0x1b0 [ 143.206308] ? vma_interval_tree_subtree_search+0x75/0xd0 [ 143.206397] unmap_mapping_range+0x18f/0x1b0 [ 143.206444] ? zap_vma_ptes+0x70/0x70 [ 143.206524] ? __pm_runtime_resume+0x67/0xa0 [ 143.206723] i915_gem_release_mmap+0x1ba/0x1c0 [i915] [ 143.206846] i915_vma_unbind+0x5c2/0x690 [i915] [ 143.206925] ? __lock_is_held+0x52/0x100 [ 143.207076] i915_gem_object_set_tiling+0x1db/0x650 [i915] [ 143.207236] i915_gem_set_tiling_ioctl+0x1d3/0x3b0 [i915] [ 143.207377] ? i915_gem_set_tiling_ioctl+0x5/0x3b0 [i915] [ 143.207457] drm_ioctl+0x36c/0x670 [ 143.207535] ? debug_lockdep_rcu_enabled.part.0+0x1a/0x30 [ 143.207730] ? i915_gem_object_set_tiling+0x650/0x650 [i915] [ 143.207793] ? drm_getunique+0x120/0x120 [ 143.207875] ? __handle_mm_fault+0x996/0x14a0 [ 143.207939] ? vm_insert_page+0x340/0x340 [ 143.208028] ? up_write+0x28/0x50 [ 143.208086] ? vm_mmap_pgoff+0x160/0x190 [ 143.208163] do_vfs_ioctl+0x12c/0xa60 [ 143.208218] ? debug_lockdep_rcu_enabled+0x35/0x40 [ 143.208267] ? ioctl_preallocate+0x150/0x150 [ 143.208353] ? __do_page_fault+0x36a/0x6e0 [ 143.208400] ? mark_held_locks+0x23/0xc0 [ 143.208479] ? up_read+0x1f/0x40 [ 143.208526] ? entry_SYSCALL_64_fastpath+0x5/0xc6 [ 143.208669] ? __fget_light+0xa7/0xc0 [ 143.208747] SyS_ioctl+0x41/0x70 To prevent the possibility of a deadlock, we defer scheduling the worker until after we have proven that given the current mm, the userptr range does not overlap a GGTT mmaping. If another thread tries to remap the GGTT over the userptr before the worker is scheduled, it will be stopped by its invalidate-range flushing the current work, before the deadlock can occur. v2: Improve discussion of how we end up in the deadlock. v3: Don't forget to mark the userptr as active after a successful gup_fast. Rename overlaps_ggtt to noncontiguous_or_overlaps_ggtt. v4: Fix test ordering between invalid GTT mmaping and range completion (Tvrtko) Reported-by:
Michał Winiarski <michal.winiarski@intel.com> Testcase: igt/gem_userptr_blits/map-fixed-invalidate-gup Signed-off-by:
-
