- 28 Jul, 2021 32 commits
-
-
Matthew Brost authored
Implement a simple static mapping algorithm of the i915 priority levels (int, -1k to 1k exposed to user) to the 4 GuC levels. Mapping is as follows: i915 level < 0 -> GuC low level (3) i915 level == 0 -> GuC normal level (2) i915 level < INT_MAX -> GuC high level (1) i915 level == INT_MAX -> GuC highest level (0) We believe this mapping should cover the UMD use cases (3 distinct user levels + 1 kernel level). In addition to static mapping, a simple counter system is attached to each context tracking the number of requests inflight on the context at each level. This is needed as the GuC levels are per context while in the i915 levels are per request. v2: (Daniele) - Add BUILD_BUG_ON to enforce ordering of priority levels - Add missing lockdep to guc_prio_fini - Check for return before setting context registered flag - Map DISPLAY priority or higher to highest guc prio - Update comment for guc_prio Signed-off-by:
Matthew Brost <matthew.brost@intel.com> Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Reviewed-by:
Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Signed-off-by:
John Harrison <John.C.Harrison@Intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20210727002348.97202-33-matthew.brost@intel.com
-
John Harrison authored
Some testing environments and some heavier tests are slower than previous limits allowed for. For example, it can take multiple seconds for the 'context has been reset' notification handler to reach the 'kill the requests' code in the 'active' version of the 'reset engines' test. During which time the selftest gets bored, gives up waiting and fails the test. There is also an async thread that the selftest uses to pump work through the hardware in parallel to the context that is marked for reset. That also could get bored waiting for completions and kill the test off. Lastly, the flush at the of various test sections can also see timeouts due to the large amount of work backed up. This is also true of the live_hwsp_read test. Signed-off-by:
-
John Harrison authored
When GuC submission is enabled, the GuC controls engine resets. Rather than explicitly triggering a reset, the driver must submit a hanging context to GuC and wait for the reset to occur. Conversely, one of the tests specifically sends hanging batches to the engines but wants them to sit around until a manual reset of the full GT (including GuC itself). That means disabling GuC based engine resets to prevent those from killing the hanging batch too soon. So, add support to the scheduling policy helper for disabling resets as well as making them quicker! In GuC submission mode, the 'is engine idle' test basically turns into 'is engine PM wakelock held'. Independently, there is a heartbeat disable helper function that the tests use. For unexplained reasons, this acquires the engine wakelock before disabling the heartbeat and only releases it when re-enabling the heartbeat. As one of the tests tries to do a wait for idle in the middle of a heartbeat disabled section, it is therefore guaranteed to always fail. Added a 'no_pm' variant of the heartbeat helper that allows the engine to be asleep while also having heartbeats disabled. Signed-off-by:
-
Matthew Brost authored
Requests may take slightly longer with GuC submission, let's increase the timeouts in live_requests. Signed-off-by:
-
Rahul Kumar Singh authored
When GuC submission is enabled, the GuC controls engine resets. Rather than explicitly triggering a reset, the driver must submit a hanging context to GuC and wait for the reset to occur. Signed-off-by:
Rahul Kumar Singh <rahul.kumar.singh@intel.com> Signed-off-by:
-
Rahul Kumar Singh authored
When GuC submission is enabled, the GuC controls engine resets. Rather than explicitly triggering a reset, the driver must submit a hanging context to GuC and wait for the reset to occur. Signed-off-by:
-
John Harrison authored
There are many ways in which the hangcheck selftest can fail. Very few of them actually printed an error message to say what happened. So, fill in the missing messages. Signed-off-by:
-
Matthew Brost authored
This adds GuC backend support for i915_request_cancel(), which in turn makes CONFIG_DRM_I915_REQUEST_TIMEOUT work. This implementation makes use of fence while there are likely simplier options. A fence was chosen because of another feature coming soon which requires a user to block on a context until scheduling is disabled. In that case we return the fence to the user and the user can wait on that fence. v2: (Daniele) - A comment about locking the blocked incr / decr - A comments about the use of the fence - Update commit message explaining why fence - Delete redundant check blocked count in unblock function - Ring buffer implementation - Comment about blocked in submission path - Shorter rpm path v3: (Checkpatch) - Fix typos in commit message (Daniel) - Rework to simplier locking structure in guc_context_block / unblock Signed-off-by:
-
Matthew Brost authored
When using GuC submission, if a context gets banned disable scheduling and mark all inflight requests as complete. Cc: John Harrison <John.C.Harrison@Intel.com> Signed-off-by:
-
John Harrison authored
The media watchdog mechanism involves GuC doing a silent reset and continue of the hung context. This requires the i915 driver provide a golden context to GuC in the ADS. v2: (Matthew Brost): - Fix memory corruption in shmem_read (John H) - Use locals rather than defines for LR_* + SKIP_SIZE Signed-off-by:
-
John Harrison authored
Added the scheduling policy parameters to the 'guc_info' debugfs state dump. Signed-off-by:
-
John Harrison authored
Changing the reset module parameter has no effect on a running GuC. The corresponding entry in the ADS must be updated and then the GuC informed via a Host2GuC message. The new debugfs interface to module parameters allows this to happen. However, connecting the parameter data address back to anything useful is messy. One option would be to pass a new private data structure address through instead of just the parameter pointer. However, that means having a new (and different) data structure for each parameter and a new (and different) write function for each parameter. This method keeps everything generic by instead using a string lookup on the directory entry name. Signed-off-by:
-
John Harrison authored
Use the official driver default scheduling policies for configuring the GuC scheduler rather than a bunch of hardcoded values. v2: (Matthew Brost) - Move I915_ENGINE_WANT_FORCED_PREEMPTION to later patch Signed-off-by:
-
John Harrison authored
In the case of a full GPU reset (e.g. because GuC has died or because GuC's hang detection has been disabled), the driver can't rely on GuC reporting the guilty context. Instead, the driver needs to scan all active contexts and find one that is currently executing, as per the execlist mode behaviour. In GuC mode, this scan is different to execlist mode as the active request list is handled very differently. Similarly, the request state dump in debugfs needs to be handled differently when in GuC submission mode. Also refactured some of the request scanning code to avoid duplication across the multiple code paths that are now replicating it. Signed-off-by:
-
Matthew Brost authored
We receive notification of an engine reset from GuC at its completion. Meaning GuC has potentially cleared any HW state we may have been interested in capturing. GuC resumes scheduling on the engine post-reset, as the resets are meant to be transparent, further muddling our error state. There is ongoing work to define an API for a GuC debug state dump. The suggestion for now is to manually disable FW initiated resets in cases where debug state is needed. Signed-off-by:
-
John Harrison authored
Clear the 'disable resets' flag to allow GuC to reset hung contexts (detected via pre-emption timeout). Signed-off-by:
-
John Harrison authored
It is impossible to seal all race conditions of resets occurring concurrent to other operations. At least, not without introducing excesive mutex locking. Instead, don't complain if it occurs. In particular, don't complain if trying to send a H2G during a reset. Whatever the H2G was about should get redone once the reset is over. Signed-off-by:
-
John Harrison authored
The driver must provide GuC with a list of mmio registers that should be saved/restored during a GuC-based engine reset. Unfortunately, the list must be dynamically allocated as its size is variable. That means the driver must generate the list twice - once to work out the size and a second time to actually save it. v2: (Alan / CI) - GEN7_GT_MODE -> GEN6_GT_MODE to fix WA selftest failure Signed-off-by:
Fernando Pacheco <fernando.pacheco@intel.com> Signed-off-by:
-
Matthew Brost authored
The GuC can implement execution qunatums, detect hung contexts and other such things but it requires the timer expired interrupt to do so. Signed-off-by:
-
Matthew Brost authored
GuC will notify the driver, via G2H, if it fails to reset an engine. We recover by resorting to a full GPU reset. v2: (John Harrison): - s/drm_dbg/drm_err Signed-off-by:
-
Matthew Brost authored
GuC will issue a reset on detecting an engine hang and will notify the driver via a G2H message. The driver will service the notification by resetting the guilty context to a simple state or banning it completely. v2: (John Harrison) - Move msg[0] lookup after length check v3: (John Harrison) - s/drm_dbg/drm_err Cc: Matthew Brost <matthew.brost@intel.com> Cc: John Harrison <John.C.Harrison@Intel.com> Signed-off-by:
-
Matthew Brost authored
The new GuC interface introduces an MMIO H2G command, INTEL_GUC_ACTION_RESET_CLIENT, which is used to implement suspend. This MMIO tears down any active contexts generating a context reset G2H CTB for each. Once that step completes the GuC tears down the CTB channels. It is safe to suspend once this MMIO H2G command completes and all G2H CTBs have been processed. In practice the i915 will likely never receive a G2H as suspend should only be called after the GPU is idle. Resume is implemented in the same manner as before - simply reload the GuC firmware and reinitialize everything (e.g. CTB channels, contexts, etc..). v2: (Michel / John H) - INTEL_GUC_ACTION_RESET_CLIENT 0x5B01 -> 0x5507 Cc: John Harrison <john.c.harrison@intel.com> Signed-off-by:
Michal Wajdeczko <michal.wajdeczko@intel.com> Reviewed-by:
-
Matthew Brost authored
Add disable GuC interrupts to intel_guc_sanitize(). Part of this requires moving the guc_*_interrupt wrapper function into header file intel_guc.h. Signed-off-by:
-
Matthew Brost authored
If submission is disabled by the backend for any reason, reset the GPU immediately in the heartbeat code as the backend can't be reenabled until the GPU is reset. Signed-off-by:
-
Matthew Brost authored
Reset implementation for new GuC interface. This is the legacy reset implementation which is called when the i915 owns the engine hang check. Future patches will offload the engine hang check to GuC but we will continue to maintain this legacy path as a fallback and this code path is also required if the GuC dies. With the new GuC interface it is not possible to reset individual engines - it is only possible to reset the GPU entirely. This patch forces an entire chip reset if any engine hangs. v2: (Michal) - Check for -EPIPE rather than -EIO (CT deadlock/corrupt check) v3: (John H) - Split into a series of smaller patches v4: (John H) - Fix typo - Add braces around if statements in reset code v5: (Checkpatch) - Fix warnings Cc: John Harrison <john.c.harrison@intel.com> Signed-off-by:
-
Matthew Brost authored
Move active request tracking to a backend vfunc rather than assuming all backends want to do this in the manner. In the of case execlists / ring submission the tracking is on the physical engine while with GuC submission it is on the context. Signed-off-by:
-
Matthew Brost authored
This is required to allow backend specific cleanup v2: (John H) - Rework commit message Signed-off-by:
-
Matthew Brost authored
With GuC virtual engines the physical engine which a request executes and completes on isn't known to the i915. Therefore we can't attach a request to a physical engines breadcrumbs. To work around this we create a single breadcrumbs per engine class when using GuC submission and direct all physical engine interrupts to this breadcrumbs. v2: (John H) - Rework header file structure so intel_engine_mask_t can be in intel_engine_types.h Signed-off-by: -
Matthew Brost authored
Update the bonding extension to return -ENODEV when using GuC submission as this extension fundamentally will not work with the GuC submission interface. Signed-off-by:
-
Matthew Brost authored
Hold a reference to the intel_context over life of an i915_request. Without this an i915_request can exist after the context has been destroyed (e.g. request retired, context closed, but user space holds a reference to the request from an out fence). In the case of GuC submission + virtual engine, the engine that the request references is also destroyed which can trigger bad pointer dref in fence ops (e.g. i915_fence_get_driver_name). We could likely change i915_fence_get_driver_name to avoid touching the engine but let's just be safe and hold the intel_context reference. v2: (John Harrison) - Update comment explaining how GuC mode and execlists mode deal with virtual engines differently Signed-off-by: -
John Harrison authored
The serial number tracking of engines happens at the backend of request submission and was expecting to only be given physical engines. However, in GuC submission mode, the decomposition of virtual to physical engines does not happen in i915. Instead, requests are submitted to their virtual engine mask all the way through to the hardware (i.e. to GuC). This would mean that the heart beat code thinks the physical engines are idle due to the serial number not incrementing. Which in turns means hangcheck does not work for GuC virtual engines. This patch updates the tracking to decompose virtual engines into their physical constituents and tracks the request against each. This is not entirely accurate as the GuC will only be issuing the request to one physical engine. However, it is the best that i915 can do given that it has no knowledge of the GuC's scheduling decisions. Downside of this is that all physical engines constituting a GuC virtual engine will be periodically unparked (even during just a single context executing) in order to be pinged with a heartbeat request. However the power and performance cost of this is not expected to be measurable (due low frequency of heartbeat pulses) and it is considered an easier option than trying to make changes to GuC firmware. v2: (Tvrtko) - Update commit message - Have default behavior if no vfunc present Signed-off-by:
-
Matthew Brost authored
Implement GuC virtual engines. Rather simple implementation, basically just allocate an engine, setup context enter / exit function to virtual engine specific functions, set all other variables / functions to guc versions, and set the engine mask to that of all the siblings. v2: Update to work with proto-ctx v3: (Daniele) - Drop include, add comment to intel_virtual_engine_has_heartbeat Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Signed-off-by:
-
- 27 Jul, 2021 2 commits
-
-
Matthew Auld authored
EHL and JSL add the 'Bypass LLC' MOCS entry, which should make it possible for userspace to bypass the GTT caching bits set by the kernel, as per the given object cache_level. This is troublesome since the heavy flush we apply when first acquiring the pages is skipped if the kernel thinks the object is coherent with the GPU. As a result it might be possible to bypass the cache and read the contents of the page directly, which could be stale data. If it's just a case of userspace shooting themselves in the foot then so be it, but since i915 takes the stance of always zeroing memory before handing it to userspace, we need to prevent this. v2: this time actually set cache_dirty in put_pages() v3: move to get_pages() which looks simpler BSpec: 34007 References: 04609175 ("Revert "drm/i915/ehl: Update MOCS table for EHL"") Signed-off-by:
Matthew Auld <matthew.auld@intel.com> Cc: Tejas Upadhyay <tejaskumarx.surendrakumar.upadhyay@intel.com> Cc: Francisco Jerez <francisco.jerez.plata@intel.com> Cc: Lucas De Marchi <lucas.demarchi@intel.com> Cc: Jon Bloomfield <jon.bloomfield@intel.com> Cc: Chris Wilson <chris.p.wilson@intel.com> Cc: Matt Roper <matthew.d.roper@intel.com> Cc: Daniel Vetter <daniel@ffwll.ch> Reviewed-by:
Daniel Vetter <daniel.vetter@ffwll.ch> Link: https://patchwork.freedesktop.org/patch/msgid/20210723105045.400841-2-matthew.auld@intel.com
-
Matthew Auld authored
Try to document the object caching related bits, like cache_coherent and cache_dirty. v2(Ville): - As pointed out by Ville, fix the completely incorrect assumptions about the "partial" coherency on shared LLC platforms. v3(Daniel): - Fix nonsense about "dirtying" the cache with reads. v4(Daniel): - Various improvements, including adding some more details for WT. Suggested-by:
-
- 26 Jul, 2021 6 commits
-
-
Thomas Hellström authored
Until we support p2p dma or as a complement to that, migrate data to system memory at dma-buf attach time if possible. v2: - Rebase on dynamic exporter. Update the igt_dmabuf_import_same_driver selftest to migrate if we are LMEM capable. v3: - Migrate also in the pin() callback. v4: - Migrate in attach v5: (jason) - Lock around the migration v6: (jason) - Move the can_migrate check outside the lock - Rework the selftests to test more migration conditions. In particular, SMEM, LMEM, and LMEM+SMEM are all checked. v7: (mauld) - Misc style nits Signed-off-by:
Thomas Hellström <thomas.hellstrom@linux.intel.com> Signed-off-by:
Michael J. Ruhl <michael.j.ruhl@intel.com> Reported-by:
kernel test robot <lkp@intel.com> Signed-off-by:
Jason Ekstrand <jason@jlekstrand.net> Reviewed-by:
-
Thomas Hellström authored
If our exported dma-bufs are imported by another instance of our driver, that instance will typically have the imported dma-bufs locked during dma_buf_map_attachment(). But the exporter also locks the same reservation object in the map_dma_buf() callback, which leads to recursive locking. So taking the lock inside _pin_pages_unlocked() is incorrect. Additionally, the current pinning code path is contrary to the defined way that pinning should occur. Remove the explicit pin/unpin from the map/umap functions and move them to the attach/detach allowing correct locking to occur, and to match the static dma-buf drm_prime pattern. Add a live selftest to exercise both dynamic and non-dynamic exports. v2: - Extend the selftest with a fake dynamic importer. - Provide real pin and unpin callbacks to not abuse the interface. v3: (ruhl) - Remove the dynamic export support and move the pinning into the attach/detach path. v4: (ruhl) - Put pages does not need to assert on the dma-resv v5: (jason) - Lock around dma_buf_unmap_attachment() when emulating a dynamic importer in the subtests. - Use pin_pages_unlocked v6: (jason) - Use dma_buf_attach instead of dma_buf_attach_dynamic in the selftests v7: (mauld) - Use __i915_gem_object_get_pages (2 __underscores) instead of the 4 ____underscore version in the selftests v8: (mauld) - Drop the kernel doc from the static i915_gem_dmabuf_attach function - Add missing "err = PTR_ERR()" to a bunch of selftest error cases Reported-by:
-
Jason Ekstrand authored
Without TTM, we have no such hook so we exit early but this is fine because we use TTM on all LMEM platforms and, on integrated platforms, there is no real migration. If we do have the hook, it's better to just let TTM handle the migration because it knows where things are actually placed. This fixes a bug where i915_gem_object_migrate fails to migrate newly created LMEM objects. In that scenario, the object has obj->mm.region set to LMEM but TTM has it in SMEM because that's where all new objects are placed there prior to getting actual pages. When we invoke i915_gem_object_migrate, it exits early because, from the point of view of the GEM object, it's already in LMEM and no migration is needed. Then, when we try to pin the pages, __i915_ttm_get_pages is called which, unaware of our failed attempt at a migration, places the object in SMEM. This only happens on newly created objects because they have this weird state where TTM thinks they're in SMEM, GEM thinks they're in LMEM, and the reality is that they don't exist at all. It's better if GEM just always calls into TTM and let's TTM handle things. That way the lies stay better contained. Once the migration is complete, the object will have pages, obj->mm.region will be correct, and we're done lying. Signed-off-by:
-
Jason Ekstrand authored
__i915_ttm_get_pages does two things. First, it calls ttm_bo_validate() to check the given placement and migrate the BO if needed. Then, it updates the GEM object to match, in case the object was migrated. If no migration occured, however, we might still have pages on the GEM object in which case we don't need to fetch them from TTM and call __i915_gem_object_set_pages. This hasn't been a problem before because the primary user of __i915_ttm_get_pages is __i915_gem_object_get_pages which only calls it if the GEM object doesn't have pages. However, i915_ttm_migrate also uses __i915_ttm_get_pages to do the migration so this meant it was unsafe to call on an already populated object. This patch checks i915_gem_object_has_pages() before trying to __i915_gem_object_set_pages so i915_ttm_migrate is safe to call, even on populated objects. Signed-off-by:
-
Jason Ekstrand authored
Instead of hand-rolling the same three calls in each function, pull them into an i915_gem_object_create_user helper. Apart from re-ordering of the placements array ENOMEM check, there should be no functional change. v2 (Matthew Auld): - Add the call to i915_gem_flush_free_objects() from i915_gem_dumb_create() in a separate patch - Move i915_gem_object_alloc() below the simple error checks v3 (Matthew Auld): - Add __ to i915_gem_object_create_user and kerneldoc which warns the caller that it's not validating anything. Signed-off-by:
-
Jason Ekstrand authored
This doesn't really fix anything serious since the chances of a client creating and destroying a mass of dumb BOs is pretty low. However, it is called by the other two create IOCTLs to garbage collect old objects. Call it here too for consistency. Signed-off-by:
-
