[PATCH v2 4/5] drm/msm: use DRM_GPUVM_RESV_PROTECTED for VM_BIND VMs

[Matthew Brost <matthew.brost@intel.com>]

msm_gem_vm_create() creates every drm_gpuvm without
DRM_GPUVM_RESV_PROTECTED, on the grounds that it makes
drm_gpuvm_bo_evict() lose track of evicted external objects. It does not:
drm_gpuvm_bo_evict() still sets drm_gpuvm_bo::evicted on an extobj, and
drm_gpuvm_prepare_objects() moves any such extobj onto the evicted list
before drm_gpuvm_validate() looks at it. The VM_BIND submit path always
calls the two in that order.

Userspace managed VMs already touch the extobj and evicted lists only
with the VM's resv held: VMAs are created and linked by VM_BIND under
the VM resv, msm_gem_vma_close() asserts it, and every drm_gpuvm_bo_put()
which can drop the last reference of a VM_BIND vm_bo runs with it held,
via msm_gem_lock_vm_and_obj(), with_vm_locks() or the object free path.
The internal spinlocks buy nothing there, so set
DRM_GPUVM_RESV_PROTECTED for those VMs.

Kernel managed VMs are left alone. The legacy submit path holds a vm_bo
reference per BO and drops it in msm_submit_retire() with only the
object's resv held, which could be the last reference once the VMA is
gone.

This is also what two pass locking in drm_gpuvm requires, which a
following patch makes use of.

Cc: Abhinav Kumar <abhinav.kumar@linux.dev>
Cc: Alice Ryhl <aliceryhl@google.com>
Cc: Anna Maniscalco <anna.maniscalco2000@gmail.com>
Cc: Boris Brezillon <boris.brezillon@collabora.com>
Cc: Danilo Krummrich <dakr@kernel.org>
Cc: David Airlie <airlied@gmail.com>
Cc: Dmitry Baryshkov <lumag@kernel.org>
Cc: Jessica Zhang <jesszhan0024@gmail.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Liviu Dudau <liviu.dudau@arm.com>
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Marijn Suijten <marijn.suijten@somainline.org>
Cc: Maxime Ripard <mripard@kernel.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Rob Clark <robin.clark@oss.qualcomm.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Cc: Sean Paul <sean@poorly.run>
Cc: Shuah Khan <skhan@linuxfoundation.org>
Cc: Simona Vetter <simona@ffwll.ch>
Cc: Steven Price <steven.price@arm.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Cc: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Assisted-by: LLM
---
 drivers/gpu/drm/msm/msm_gem_vma.c | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/msm/msm_gem_vma.c b/drivers/gpu/drm/msm/msm_gem_vma.c
index 1badec3caa7b..c7cdd32a798f 100644
--- a/drivers/gpu/drm/msm/msm_gem_vma.c
+++ b/drivers/gpu/drm/msm/msm_gem_vma.c
@@ -818,11 +818,18 @@ msm_gem_vm_create(struct drm_device *drm, struct msm_mmu *mmu, const char *name,
 		  u64 va_start, u64 va_size, bool managed)
 {
 	/*
-	 * We mostly want to use DRM_GPUVM_RESV_PROTECTED, except that
-	 * makes drm_gpuvm_bo_evict() a no-op for extobjs (ie. we loose
-	 * tracking that an extobj is evicted) :facepalm:
+	 * Userspace managed (VM_BIND) VMs only ever touch the gpuvm's extobj
+	 * and evicted lists with the VM's resv held, so use
+	 * DRM_GPUVM_RESV_PROTECTED for those.  drm_gpuvm_bo_evict() cannot
+	 * put an extobj on the evicted list there, but it records the
+	 * eviction and drm_gpuvm_prepare_objects() moves it onto the list
+	 * before drm_gpuvm_validate() runs, so nothing is lost.
+	 *
+	 * Kernel managed VMs keep the internal spinlocks, since the legacy
+	 * submit path can drop the last vm_bo reference with only the
+	 * object's resv held (see msm_submit_retire()).
 	 */
-	enum drm_gpuvm_flags flags = 0;
+	enum drm_gpuvm_flags flags = managed ? 0 : DRM_GPUVM_RESV_PROTECTED;
 	struct msm_gem_vm *vm;
 	struct drm_gem_object *dummy_gem;
 	int ret = 0;
-- 
2.34.1