On Thu, Jul 08, 2021 at 10:48:05AM -0500, Jason Ekstrand wrote:
Overview:
This patch series attempts to clean up some of the IOCTL mess we've created over the last few years. The most egregious bit being context mutability. In summary, this series:
- Drops two never-used context params: RINGSIZE and NO_ZEROMAP
- Drops the entire CONTEXT_CLONE API
- Implements SINGLE_TIMELINE with a syncobj instead of actually sharing intel_timeline between engines.
- Adds a few sanity restrictions to the balancing/bonding API.
- Implements a proto-ctx mechanism so that the engine set and VM can only be set early on in the lifetime of a context, before anything ever executes on it. This effectively makes the VM and engine set immutable.
This series has been tested with IGT as well as the Iris, ANV, and the Intel media driver doing an 8K decode (this uses bonding/balancing). I've also done quite a bit of git archeology to ensure that nothing in here will break anything that's already shipped at some point in history. It's possible I've missed something, but I've dug quite a bit.
Details and motivation:
In very broad strokes, there's an effort going on right now within Intel to try and clean up and simplify i915 anywhere we can. We obviously don't want to break any shipping userspace but, as can be seen by this series, there's a lot i915 theoretically supports which userspace doesn't actually need. Some of this, like the two context params used here, were simply oversights where we went through the usual API review process and merged the i915 bits but the userspace bits never landed for some reason.
Not all are so innocent, however. For instance, there's an entire context cloning API which allows one to create a context with certain parameters "cloned" from some other context. This entire API has never been used by any userspace except IGT and there were never patches to any other userspace to use it. It never should have landed. Also, when we added support for setting explicit engine sets and sharing VMs across contexts, people decided to do so via SET_CONTEXT_PARAM. While this allowed them to re-use existing API, it did so at the cost of making those states mutable which leads to a plethora of potential race conditions. There were even IGT tests merged to cover some of theses:
gem_vm_create@async-destroy and gem_vm_create@destroy-race which test swapping out the VM on a running context.
gem_ctx_persistence@replace* which test whether a client can escape a non-persistent context by submitting a hanging batch and then swapping out the engine set before the hang is detected.
api_intel_bb@bb-with-vm which tests the that intel_bb_assign_vm works properly. This API is never used by any other IGT test.
There is also an entire deferred flush and set state framework in i915_gem_cotnext.c which exists for safely swapping out the VM while there is work in-flight on a context.
So, clearly people knew that this API was inherently racy and difficult to implement but they landed it anyway. Why? The best explanation I've been given is because it makes the API more "unified" or "symmetric" for this stuff to go through SET_CONTEXT_PARAM. It's not because any userspace actually wants to be able to swap out the VM or the set of engines on a running context. That would be utterly insane.
This patch series cleans up this particular mess by introducing the concept of a i915_gem_proto_context data structure which contains context creation information. When you initially call GEM_CONTEXT_CREATE, a proto-context in created instead of an actual context. Then, the first time something is done on the context besides SET_CONTEXT_PARAM, an actual context is created. This allows us to keep the old drivers which use SET_CONTEXT_PARAM to set up the engine set (see also media) while ensuring that, once you have an i915_gem_context, the VM and the engine set are immutable state.
Eventually, there are more clean-ups I'd like to do on top of this which should make working with contexts inside i915 simpler and safer:
Move the GEM handle -> vma LUT from i915_gem_context into either i915_ppgtt or drm_i915_file_private depending on whether or not the hardware has a full PPGTT.
Move the delayed context destruction code into intel_context or a per-engine wrapper struct rather than i915_gem_context.
Get rid of the separation between context close and context destroy
Get rid of the RCU on i915_gem_context
However, these should probably be done as a separate patch series as this one is already starting to get longish, especially if you consider the 89 IGT patches that go along with it.
Test-with: 20210707210215.351483-1-jason@jlekstrand.net
Jason Ekstrand (30): drm/i915: Drop I915_CONTEXT_PARAM_RINGSIZE drm/i915: Stop storing the ring size in the ring pointer (v3) drm/i915: Drop I915_CONTEXT_PARAM_NO_ZEROMAP drm/i915/gem: Set the watchdog timeout directly in intel_context_set_gem (v2) drm/i915/gem: Return void from context_apply_all drm/i915: Drop the CONTEXT_CLONE API (v2) drm/i915: Implement SINGLE_TIMELINE with a syncobj (v4) drm/i915: Drop getparam support for I915_CONTEXT_PARAM_ENGINES drm/i915/gem: Disallow bonding of virtual engines (v3) drm/i915/gem: Remove engine auto-magic with FENCE_SUBMIT (v2) drm/i915/request: Remove the hook from await_execution drm/i915/gem: Disallow creating contexts with too many engines drm/i915: Stop manually RCU banging in reset_stats_ioctl (v2) drm/i915/gem: Add a separate validate_priority helper drm/i915: Add gem/i915_gem_context.h to the docs drm/i915/gem: Add an intermediate proto_context struct (v5) drm/i915/gem: Rework error handling in default_engines drm/i915/gem: Optionally set SSEU in intel_context_set_gem drm/i915: Add an i915_gem_vm_lookup helper drm/i915/gem: Make an alignment check more sensible drm/i915/gem: Use the proto-context to handle create parameters (v5) drm/i915/gem: Return an error ptr from context_lookup drm/i915/gt: Drop i915_address_space::file (v2) drm/i915/gem: Delay context creation (v3) drm/i915/gem: Don't allow changing the VM on running contexts (v4) drm/i915/gem: Don't allow changing the engine set on running contexts (v3) drm/i915/selftests: Take a VM in kernel_context() i915/gem/selftests: Assign the VM at context creation in igt_shared_ctx_exec drm/i915/gem: Roll all of context creation together drm/i915: Finalize contexts in GEM_CONTEXT_CREATE on version 13+
Applied everything, thanks a lot.
Now, could we bake all this hard-earned knowledge about how the uapi actually works into some kerneldoc for the uapi header?
Thanks, Daniel
Documentation/gpu/i915.rst | 2 + drivers/gpu/drm/i915/Makefile | 1 - drivers/gpu/drm/i915/gem/i915_gem_context.c | 2926 ++++++++--------- drivers/gpu/drm/i915/gem/i915_gem_context.h | 3 + .../gpu/drm/i915/gem/i915_gem_context_types.h | 196 +- .../gpu/drm/i915/gem/i915_gem_execbuffer.c | 31 +- .../drm/i915/gem/selftests/i915_gem_context.c | 127 +- .../gpu/drm/i915/gem/selftests/mock_context.c | 67 +- .../gpu/drm/i915/gem/selftests/mock_context.h | 4 +- drivers/gpu/drm/i915/gt/intel_context.c | 3 +- drivers/gpu/drm/i915/gt/intel_context.h | 5 - drivers/gpu/drm/i915/gt/intel_context_param.c | 63 - drivers/gpu/drm/i915/gt/intel_context_param.h | 6 +- drivers/gpu/drm/i915/gt/intel_context_types.h | 1 + drivers/gpu/drm/i915/gt/intel_engine_cs.c | 3 +- drivers/gpu/drm/i915/gt/intel_engine_types.h | 7 - .../drm/i915/gt/intel_execlists_submission.c | 114 - .../drm/i915/gt/intel_execlists_submission.h | 8 +- drivers/gpu/drm/i915/gt/intel_gtt.h | 11 - drivers/gpu/drm/i915/gt/intel_lrc.c | 2 +- drivers/gpu/drm/i915/gt/intel_migrate.c | 3 +- drivers/gpu/drm/i915/gt/selftest_execlists.c | 251 +- drivers/gpu/drm/i915/gt/selftest_hangcheck.c | 2 +- drivers/gpu/drm/i915/gt/selftest_mocs.c | 2 +- drivers/gpu/drm/i915/gt/selftest_timeline.c | 2 +- drivers/gpu/drm/i915/gvt/scheduler.c | 7 +- drivers/gpu/drm/i915/i915_drv.h | 82 +- drivers/gpu/drm/i915/i915_perf.c | 4 +- drivers/gpu/drm/i915/i915_request.c | 42 +- drivers/gpu/drm/i915/i915_request.h | 4 +- .../drm/i915/selftests/i915_mock_selftests.h | 1 - drivers/gpu/drm/i915/selftests/mock_gtt.c | 1 - include/uapi/drm/i915_drm.h | 40 +- 33 files changed, 1681 insertions(+), 2340 deletions(-) delete mode 100644 drivers/gpu/drm/i915/gt/intel_context_param.c
-- 2.31.1