From: Akash Goel <[email protected]>
This implementation is losely based on the MSM shrinker, and it's
relying on the drm_gpuvm eviction/validation infrastructure.
Right now we only support swapout/eviction, but we could add an extra
flag to specify when buffer content doesn't need to be preserved to
avoid the swapout/swapin dance.
Locking is a bit of a nightmare, but using _trylock() all the way in
the reclaim path seems to make lockdep happy. And yes, we might be
missing opportunities to reclaim when the system is under heavy GPU
load/heavy memory pressure/heavy GPU VM activity, but that's better
than no reclaim at all.
v2:
- Move gpu_mapped_shared next to the mmapped LRU
- Add a bunch of missing is_[vm_bo,vma]_evicted() tests
- Only test mmap_count to check if a BO is mmaped
- Remove stale comment about shrinker not being a thing
- Allow pin_count to be non-zero in panthor_gem_swapin_locked()
- Fix panthor_gem_sync() to check for BO residency before doing the CPU sync
- Fix the value returned by panthor_gem_shrinker_count() in case some
memory has been released
- Check drmm_mutex_init() ret code
- Explicitly mention that PANTHOR_GEM_UNRECLAIMABLE is the initial state
of all BOs
v3:
- Make panthor_gem_try_evict() static
- Collect {A,R}-bs
v4:
- Update the reclaim_state in panthor_gem_mmap()
- Don't reclaim GPU-mapped BOs if can_block() returns false
- Skip evicited vm_bos in panthor_vm_update_bo_reclaim_lru_locked() to
avoid spurious WARN_ON()s
- Explain why we have to do this
select_evicted_vma/repopulate_evicted_vma dance
v5:
- Properly report the reclaimable size in panthor_gem_debugfs_print_bos()
- Check panthor_vm_lock_region() errors in
panthor_vm_evict_bo_mappings_locked()
- Fix lock order inversion (dma_resv_wait_timeout() inside gpuva.lock)
v6:
- Don't remap if the unmapped VMA is evicted.
- Drop a stale comment in panthor_gem_dev_map_get_sgt_locked()
- s/PANTHOR_GEM_GPU_MAPPED_PRIVATE/PANTHOR_GEM_GPU_MAPPED_SINGLE_VM/
- s/PANTHOR_GEM_GPU_MAPPED_SHARED/PANTHOR_GEM_GPU_MAPPED_MULTI_VM/
- Just count the number of vm_bo to determine the reclaim state in
is_gpu_mapped()
- Drop a redundant panthor_gem_backing_get_pages_locked() call in
panthor_gem_swapin_locked()
- Add more comments to panthor_vm_evict_bo_mappings_locked() to
convince Claude it's actually safe.
- Add a comment in panthor_gem_shrinker_count() to mention the race
and hopefully clear Claude's concerns.
- Rework the "is-still-in-list" check we have in
panthor_mmu_reclaim_priv_bos() to address Claude's concerns
- Don't call panthor_vm_unlock_region() if panthor_vm_lock_region()
fails in remap_evicted_vma(). Was harmless, but confusing, as
pointed out by Claude
- Fix can_block() to allow blocking on KSWAPD_RECLAIM-only in a kswapd
context
- Fix a lockdep warning when the last ref on a BO is released in the
shrinker path, where the resv lock can't be acquired other than with
a try_lock()
v7:
- Skip drm_gpuvm_bo_deferred_cleanup() when repopulating VMAs to fix a
deadlock
Signed-off-by: Akash Goel <[email protected]>
Co-developed-by: Boris Brezillon <[email protected]>
Signed-off-by: Boris Brezillon <[email protected]>
Acked-by: Liviu Dudau <[email protected]>
Reviewed-by: Steven Price <[email protected]>
---
drivers/gpu/drm/panthor/panthor_device.c | 11 +-
drivers/gpu/drm/panthor/panthor_device.h | 73 ++++
drivers/gpu/drm/panthor/panthor_gem.c | 508 ++++++++++++++++++++++-
drivers/gpu/drm/panthor/panthor_gem.h | 68 +++
drivers/gpu/drm/panthor/panthor_mmu.c | 397 +++++++++++++++++-
drivers/gpu/drm/panthor/panthor_mmu.h | 8 +
6 files changed, 1027 insertions(+), 38 deletions(-)
diff --git a/drivers/gpu/drm/panthor/panthor_device.c
b/drivers/gpu/drm/panthor/panthor_device.c
index 54fbb1aa07c5..bc62a498a8a8 100644
--- a/drivers/gpu/drm/panthor/panthor_device.c
+++ b/drivers/gpu/drm/panthor/panthor_device.c
@@ -2,6 +2,7 @@
/* Copyright 2018 Marty E. Plummer <[email protected]> */
/* Copyright 2019 Linaro, Ltd, Rob Herring <[email protected]> */
/* Copyright 2023 Collabora ltd. */
+/* Copyright 2025 ARM Limited. All rights reserved. */
#include <linux/clk.h>
#include <linux/mm.h>
@@ -122,6 +123,7 @@ void panthor_device_unplug(struct panthor_device *ptdev)
panthor_sched_unplug(ptdev);
panthor_fw_unplug(ptdev);
panthor_mmu_unplug(ptdev);
+ panthor_gem_shrinker_unplug(ptdev);
panthor_gpu_unplug(ptdev);
panthor_pwr_unplug(ptdev);
@@ -291,10 +293,14 @@ int panthor_device_init(struct panthor_device *ptdev)
if (ret)
goto err_unplug_gpu;
- ret = panthor_mmu_init(ptdev);
+ ret = panthor_gem_shrinker_init(ptdev);
if (ret)
goto err_unplug_gpu;
+ ret = panthor_mmu_init(ptdev);
+ if (ret)
+ goto err_unplug_shrinker;
+
ret = panthor_fw_init(ptdev);
if (ret)
goto err_unplug_mmu;
@@ -326,6 +332,9 @@ int panthor_device_init(struct panthor_device *ptdev)
err_unplug_mmu:
panthor_mmu_unplug(ptdev);
+err_unplug_shrinker:
+ panthor_gem_shrinker_unplug(ptdev);
+
err_unplug_gpu:
panthor_gpu_unplug(ptdev);
diff --git a/drivers/gpu/drm/panthor/panthor_device.h
b/drivers/gpu/drm/panthor/panthor_device.h
index b6696f73a536..5cba272f9b4d 100644
--- a/drivers/gpu/drm/panthor/panthor_device.h
+++ b/drivers/gpu/drm/panthor/panthor_device.h
@@ -14,6 +14,7 @@
#include <linux/spinlock.h>
#include <drm/drm_device.h>
+#include <drm/drm_gem.h>
#include <drm/drm_mm.h>
#include <drm/gpu_scheduler.h>
#include <drm/panthor_drm.h>
@@ -178,6 +179,78 @@ struct panthor_device {
/** @devfreq: Device frequency scaling management data. */
struct panthor_devfreq *devfreq;
+ /** @reclaim: Reclaim related stuff */
+ struct {
+ /** @reclaim.shrinker: Shrinker instance */
+ struct shrinker *shrinker;
+
+ /** @reclaim.lock: Lock protecting all LRUs */
+ struct mutex lock;
+
+ /**
+ * @reclaim.unused: BOs with unused pages
+ *
+ * Basically all buffers that got mmapped, vmapped or GPU
mapped and
+ * then unmapped. There should be no contention on these
buffers,
+ * making them ideal to reclaim.
+ */
+ struct drm_gem_lru unused;
+
+ /**
+ * @reclaim.mmapped: mmap()-ed buffers
+ *
+ * Those are relatively easy to reclaim since we don't need user
+ * agreement, we can simply teardown the mapping and let it
fault on
+ * the next access.
+ */
+ struct drm_gem_lru mmapped;
+
+ /**
+ * @reclaim.gpu_mapped_shared: shared BO LRU list
+ *
+ * That's the most tricky BO type to reclaim, because it
involves
+ * tearing down all mappings in all VMs where this BO is mapped,
+ * which increases the risk of contention and thus decreases the
+ * likeliness of success.
+ */
+ struct drm_gem_lru gpu_mapped_shared;
+
+ /**
+ * @reclaim.vms: VM LRU list
+ *
+ * VMs that have reclaimable BOs only mapped to a single VM are
placed
+ * in this LRU. Reclaiming such BOs implies waiting for VM
idleness
+ * (no in-flight GPU jobs targeting this VM), meaning we can't
reclaim
+ * those if we're in a context where we can't block/sleep.
+ */
+ struct list_head vms;
+
+ /**
+ * @reclaim.gpu_mapped_count: Global counter of pages that are
GPU mapped
+ *
+ * Allows us to get the number of reclaimable pages without
walking
+ * the vms and gpu_mapped_shared LRUs.
+ */
+ long gpu_mapped_count;
+
+ /**
+ * @reclaim.retry_count: Number of times we ran the shrinker
without being
+ * able to reclaim stuff
+ *
+ * Used to stop scanning GEMs when too many attempts were made
+ * without progress.
+ */
+ atomic_t retry_count;
+
+#ifdef CONFIG_DEBUG_FS
+ /**
+ * @reclaim.nr_pages_reclaimed_on_last_scan: Number of pages
reclaimed on the last
+ * shrinker scan
+ */
+ unsigned long nr_pages_reclaimed_on_last_scan;
+#endif
+ } reclaim;
+
/** @unplug: Device unplug related fields. */
struct {
/** @lock: Lock used to serialize unplug operations. */
diff --git a/drivers/gpu/drm/panthor/panthor_gem.c
b/drivers/gpu/drm/panthor/panthor_gem.c
index eb139759beda..69cef05b6ef7 100644
--- a/drivers/gpu/drm/panthor/panthor_gem.c
+++ b/drivers/gpu/drm/panthor/panthor_gem.c
@@ -2,8 +2,10 @@
/* Copyright 2019 Linaro, Ltd, Rob Herring <[email protected]> */
/* Copyright 2023 Collabora ltd. */
/* Copyright 2025 Amazon.com, Inc. or its affiliates */
+/* Copyright 2025 ARM Limited. All rights reserved. */
#include <linux/cleanup.h>
+#include <linux/debugfs.h>
#include <linux/dma-buf.h>
#include <linux/dma-mapping.h>
#include <linux/err.h>
@@ -12,6 +14,8 @@
#include <drm/drm_debugfs.h>
#include <drm/drm_file.h>
+#include <drm/drm_gpuvm.h>
+#include <drm/drm_managed.h>
#include <drm/drm_prime.h>
#include <drm/drm_print.h>
#include <drm/panthor_drm.h>
@@ -114,10 +118,116 @@ should_map_wc(struct panthor_gem_object *bo)
return true;
}
+static bool is_gpu_mapped(struct panthor_gem_object *bo,
+ enum panthor_gem_reclaim_state *state)
+{
+ struct drm_gpuvm_bo *vm_bo;
+ u32 vm_count = 0;
+
+ drm_gem_for_each_gpuvm_bo(vm_bo, &bo->base) {
+ /* Skip evicted GPU mappings. */
+ if (vm_bo->evicted)
+ continue;
+
+ if (vm_count++) {
+ *state = PANTHOR_GEM_GPU_MAPPED_MULTI_VM;
+ break;
+ }
+
+ *state = PANTHOR_GEM_GPU_MAPPED_SINGLE_VM;
+ }
+
+ return vm_count > 0;
+}
+
+static enum panthor_gem_reclaim_state
+panthor_gem_evaluate_reclaim_state_locked(struct panthor_gem_object *bo)
+{
+ enum panthor_gem_reclaim_state gpu_mapped_state;
+
+ dma_resv_assert_held(bo->base.resv);
+ lockdep_assert_held(&bo->base.gpuva.lock);
+
+ /* If pages have not been allocated, there's nothing to reclaim. */
+ if (!bo->backing.pages)
+ return PANTHOR_GEM_UNRECLAIMABLE;
+
+ /* If memory is pinned, we prevent reclaim. */
+ if (refcount_read(&bo->backing.pin_count))
+ return PANTHOR_GEM_UNRECLAIMABLE;
+
+ if (is_gpu_mapped(bo, &gpu_mapped_state))
+ return gpu_mapped_state;
+
+ if (refcount_read(&bo->cmap.mmap_count))
+ return PANTHOR_GEM_MMAPPED;
+
+ return PANTHOR_GEM_UNUSED;
+}
+
+void panthor_gem_update_reclaim_state_locked(struct panthor_gem_object *bo,
+ enum panthor_gem_reclaim_state
*old_statep)
+{
+ struct panthor_device *ptdev = container_of(bo->base.dev, struct
panthor_device, base);
+ enum panthor_gem_reclaim_state old_state = bo->reclaim_state;
+ enum panthor_gem_reclaim_state new_state;
+ bool was_gpu_mapped, is_gpu_mapped;
+
+ if (old_statep)
+ *old_statep = old_state;
+
+ new_state = panthor_gem_evaluate_reclaim_state_locked(bo);
+ if (new_state == old_state)
+ return;
+
+ was_gpu_mapped = old_state == PANTHOR_GEM_GPU_MAPPED_MULTI_VM ||
+ old_state == PANTHOR_GEM_GPU_MAPPED_SINGLE_VM;
+ is_gpu_mapped = new_state == PANTHOR_GEM_GPU_MAPPED_MULTI_VM ||
+ new_state == PANTHOR_GEM_GPU_MAPPED_SINGLE_VM;
+
+ if (is_gpu_mapped && !was_gpu_mapped)
+ ptdev->reclaim.gpu_mapped_count += bo->base.size >> PAGE_SHIFT;
+ else if (!is_gpu_mapped && was_gpu_mapped)
+ ptdev->reclaim.gpu_mapped_count -= bo->base.size >> PAGE_SHIFT;
+
+ switch (new_state) {
+ case PANTHOR_GEM_UNUSED:
+ drm_gem_lru_move_tail(&ptdev->reclaim.unused, &bo->base);
+ break;
+ case PANTHOR_GEM_MMAPPED:
+ drm_gem_lru_move_tail(&ptdev->reclaim.mmapped, &bo->base);
+ break;
+ case PANTHOR_GEM_GPU_MAPPED_SINGLE_VM:
+ panthor_vm_update_bo_reclaim_lru_locked(bo);
+ break;
+ case PANTHOR_GEM_GPU_MAPPED_MULTI_VM:
+ drm_gem_lru_move_tail(&ptdev->reclaim.gpu_mapped_shared,
&bo->base);
+ break;
+ case PANTHOR_GEM_UNRECLAIMABLE:
+ drm_gem_lru_remove(&bo->base);
+ break;
+ default:
+ drm_WARN(&ptdev->base, true, "invalid GEM reclaim state
(%d)\n", new_state);
+ break;
+ }
+
+ bo->reclaim_state = new_state;
+}
+
+static void
+bo_assert_locked_or_gone(struct panthor_gem_object *bo)
+{
+ /* If the refcount is zero, the BO is being freed, and we
+ * allow the lock to not be held in that particular case.
+ */
+ if (kref_read(&bo->base.refcount))
+ dma_resv_assert_held(bo->base.resv);
+}
+
static void
panthor_gem_backing_cleanup_locked(struct panthor_gem_object *bo)
{
- dma_resv_assert_held(bo->base.resv);
+ bo_assert_locked_or_gone(bo);
if (!bo->backing.pages)
return;
@@ -157,28 +267,35 @@ static int panthor_gem_backing_pin_locked(struct
panthor_gem_object *bo)
return 0;
ret = panthor_gem_backing_get_pages_locked(bo);
- if (!ret)
+ if (!ret) {
refcount_set(&bo->backing.pin_count, 1);
+ mutex_lock(&bo->base.gpuva.lock);
+ panthor_gem_update_reclaim_state_locked(bo, NULL);
+ mutex_unlock(&bo->base.gpuva.lock);
+ }
return ret;
}
static void panthor_gem_backing_unpin_locked(struct panthor_gem_object *bo)
{
- dma_resv_assert_held(bo->base.resv);
+ bo_assert_locked_or_gone(bo);
drm_WARN_ON_ONCE(bo->base.dev, drm_gem_is_imported(&bo->base));
if (refcount_dec_and_test(&bo->backing.pin_count)) {
/* We don't release anything when pin_count drops to zero.
* Pages stay there until an explicit cleanup is requested.
*/
+ mutex_lock(&bo->base.gpuva.lock);
+ panthor_gem_update_reclaim_state_locked(bo, NULL);
+ mutex_unlock(&bo->base.gpuva.lock);
}
}
static void
panthor_gem_dev_map_cleanup_locked(struct panthor_gem_object *bo)
{
- dma_resv_assert_held(bo->base.resv);
+ bo_assert_locked_or_gone(bo);
if (!bo->dmap.sgt)
return;
@@ -200,9 +317,6 @@ panthor_gem_dev_map_get_sgt_locked(struct
panthor_gem_object *bo)
if (bo->dmap.sgt)
return bo->dmap.sgt;
- /* Pages stay around after they've been allocated. At least that stands
- * until we add a shrinker.
- */
ret = panthor_gem_backing_get_pages_locked(bo);
if (ret)
return ERR_PTR(ret);
@@ -336,11 +450,19 @@ static void panthor_gem_free_object(struct drm_gem_object
*obj)
if (drm_gem_is_imported(obj)) {
drm_prime_gem_destroy(obj, bo->dmap.sgt);
} else {
- dma_resv_lock(obj->resv, NULL);
+ /* The last ref on the GEM object can be released
+ * by the shrinker, which can't block on the resv
+ * lock acquisition. In practice, even if we were
+ * taking the lock, it wouldn't block because we're
+ * the last piece of code having a visibility on
+ * this GEM, but lockdep can't see that, so we've
+ * just tought the _cleanup_locked() helpers about
+ * this "being freed" exception, and we call those
+ * without the lock held here.
+ */
panthor_gem_vmap_cleanup_locked(bo);
panthor_gem_dev_map_cleanup_locked(bo);
panthor_gem_backing_cleanup_locked(bo);
- dma_resv_unlock(obj->resv);
}
drm_gem_object_release(obj);
@@ -535,6 +657,41 @@ void panthor_gem_unpin(struct panthor_gem_object *bo)
dma_resv_unlock(bo->base.resv);
}
+int panthor_gem_swapin_locked(struct panthor_gem_object *bo)
+{
+ struct sg_table *sgt;
+
+ dma_resv_assert_held(bo->base.resv);
+
+ if (drm_WARN_ON_ONCE(bo->base.dev, drm_gem_is_imported(&bo->base)))
+ return -EINVAL;
+
+ sgt = panthor_gem_dev_map_get_sgt_locked(bo);
+ if (IS_ERR(sgt))
+ return PTR_ERR(sgt);
+
+ return 0;
+}
+
+static void panthor_gem_evict_locked(struct panthor_gem_object *bo)
+{
+ dma_resv_assert_held(bo->base.resv);
+ lockdep_assert_held(&bo->base.gpuva.lock);
+
+ if (drm_WARN_ON_ONCE(bo->base.dev, drm_gem_is_imported(&bo->base)))
+ return;
+
+ if (drm_WARN_ON_ONCE(bo->base.dev,
refcount_read(&bo->backing.pin_count)))
+ return;
+
+ if (drm_WARN_ON_ONCE(bo->base.dev, !bo->backing.pages))
+ return;
+
+ panthor_gem_dev_map_cleanup_locked(bo);
+ panthor_gem_backing_cleanup_locked(bo);
+ panthor_gem_update_reclaim_state_locked(bo, NULL);
+}
+
static struct sg_table *panthor_gem_get_sg_table(struct drm_gem_object *obj)
{
struct panthor_gem_object *bo = to_panthor_bo(obj);
@@ -607,8 +764,12 @@ static int panthor_gem_mmap(struct drm_gem_object *obj,
struct vm_area_struct *v
if (!refcount_inc_not_zero(&bo->cmap.mmap_count)) {
dma_resv_lock(obj->resv, NULL);
- if (!refcount_inc_not_zero(&bo->cmap.mmap_count))
+ if (!refcount_inc_not_zero(&bo->cmap.mmap_count)) {
refcount_set(&bo->cmap.mmap_count, 1);
+ mutex_lock(&bo->base.gpuva.lock);
+ panthor_gem_update_reclaim_state_locked(bo, NULL);
+ mutex_unlock(&bo->base.gpuva.lock);
+ }
dma_resv_unlock(obj->resv);
}
@@ -692,6 +853,10 @@ static vm_fault_t blocking_page_setup(struct vm_fault
*vmf, unsigned int order,
} else {
struct page *page = bo->backing.pages[page_offset];
+ mutex_lock(&bo->base.gpuva.lock);
+ panthor_gem_update_reclaim_state_locked(bo, NULL);
+ mutex_unlock(&bo->base.gpuva.lock);
+
if (mmap_lock_held)
ret = insert_page(vmf, order, page);
else
@@ -734,7 +899,12 @@ static vm_fault_t panthor_gem_any_fault(struct vm_fault
*vmf, unsigned int order
/* Wait with the mmap lock released, if we're allowed to. */
drm_gem_object_get(&bo->base);
- mmap_read_unlock(vmf->vma->vm_mm);
+
+ if (vmf->flags & FAULT_FLAG_VMA_LOCK)
+ vma_end_read(vmf->vma);
+ else
+ mmap_read_unlock(vmf->vma->vm_mm);
+
ret = blocking_page_setup(vmf, order, bo, page_offset, false);
drm_gem_object_put(&bo->base);
return ret;
@@ -770,7 +940,9 @@ static void panthor_gem_vm_close(struct vm_area_struct *vma)
dma_resv_lock(bo->base.resv, NULL);
if (refcount_dec_and_test(&bo->cmap.mmap_count)) {
- /* Nothing to do, pages are reclaimed lazily. */
+ mutex_lock(&bo->base.gpuva.lock);
+ panthor_gem_update_reclaim_state_locked(bo, NULL);
+ mutex_unlock(&bo->base.gpuva.lock);
}
dma_resv_unlock(bo->base.resv);
@@ -810,6 +982,7 @@ panthor_gem_alloc_object(u32 flags)
if (!bo)
return ERR_PTR(-ENOMEM);
+ bo->reclaim_state = PANTHOR_GEM_UNRECLAIMABLE;
bo->base.funcs = &panthor_gem_funcs;
bo->flags = flags;
mutex_init(&bo->label.lock);
@@ -968,6 +1141,7 @@ panthor_gem_sync(struct drm_gem_object *obj, u32 type,
struct sg_table *sgt;
struct scatterlist *sgl;
unsigned int count;
+ int ret;
/* Make sure the range is in bounds. */
if (offset + size < offset || offset + size > bo->base.size)
@@ -994,9 +1168,21 @@ panthor_gem_sync(struct drm_gem_object *obj, u32 type,
if (size == 0)
return 0;
- sgt = panthor_gem_get_dev_sgt(bo);
- if (IS_ERR(sgt))
- return PTR_ERR(sgt);
+ ret = dma_resv_lock_interruptible(bo->base.resv, NULL);
+ if (ret)
+ return ret;
+
+ /* If there's no pages, there's no point pulling those back, bail out
early. */
+ if (!bo->backing.pages) {
+ ret = 0;
+ goto out_unlock;
+ }
+
+ sgt = panthor_gem_dev_map_get_sgt_locked(bo);
+ if (IS_ERR(sgt)) {
+ ret = PTR_ERR(sgt);
+ goto out_unlock;
+ }
for_each_sgtable_dma_sg(sgt, sgl, count) {
if (size == 0)
@@ -1040,7 +1226,11 @@ panthor_gem_sync(struct drm_gem_object *obj, u32 type,
dma_sync_single_for_cpu(dma_dev, paddr, len,
DMA_FROM_DEVICE);
}
- return 0;
+ ret = 0;
+
+out_unlock:
+ dma_resv_unlock(bo->base.resv);
+ return ret;
}
/**
@@ -1050,11 +1240,13 @@ panthor_gem_sync(struct drm_gem_object *obj, u32 type,
*/
void panthor_kernel_bo_destroy(struct panthor_kernel_bo *bo)
{
+ struct panthor_device *ptdev;
struct panthor_vm *vm;
if (IS_ERR_OR_NULL(bo))
return;
+ ptdev = container_of(bo->obj->dev, struct panthor_device, base);
vm = bo->vm;
panthor_kernel_bo_vunmap(bo);
@@ -1062,6 +1254,8 @@ void panthor_kernel_bo_destroy(struct panthor_kernel_bo
*bo)
to_panthor_bo(bo->obj)->exclusive_vm_root_gem !=
panthor_vm_root_gem(vm));
panthor_vm_unmap_range(vm, bo->va_node.start, bo->va_node.size);
panthor_vm_free_va(vm, &bo->va_node);
+ if (vm == panthor_fw_vm(ptdev))
+ panthor_gem_unpin(to_panthor_bo(bo->obj));
drm_gem_object_put(bo->obj);
panthor_vm_put(vm);
kfree(bo);
@@ -1110,6 +1304,12 @@ panthor_kernel_bo_create(struct panthor_device *ptdev,
struct panthor_vm *vm,
kbo->obj = &bo->base;
+ if (vm == panthor_fw_vm(ptdev)) {
+ ret = panthor_gem_pin(bo);
+ if (ret)
+ goto err_put_obj;
+ }
+
panthor_gem_kernel_bo_set_label(kbo, name);
/* The system and GPU MMU page size might differ, which becomes a
@@ -1121,7 +1321,7 @@ panthor_kernel_bo_create(struct panthor_device *ptdev,
struct panthor_vm *vm,
size = ALIGN(size, panthor_vm_page_size(vm));
ret = panthor_vm_alloc_va(vm, gpu_va, size, &kbo->va_node);
if (ret)
- goto err_put_obj;
+ goto err_unpin;
ret = panthor_vm_map_bo_range(vm, bo, 0, size, kbo->va_node.start,
vm_map_flags);
if (ret)
@@ -1133,6 +1333,10 @@ panthor_kernel_bo_create(struct panthor_device *ptdev,
struct panthor_vm *vm,
err_free_va:
panthor_vm_free_va(vm, &kbo->va_node);
+err_unpin:
+ if (vm == panthor_fw_vm(ptdev))
+ panthor_gem_unpin(bo);
+
err_put_obj:
drm_gem_object_put(&bo->base);
@@ -1141,6 +1345,241 @@ panthor_kernel_bo_create(struct panthor_device *ptdev,
struct panthor_vm *vm,
return ERR_PTR(ret);
}
+static bool can_swap(void)
+{
+ return get_nr_swap_pages() > 0;
+}
+
+static bool can_block(struct shrink_control *sc)
+{
+ /* If direct reclaim is allowed, we can always block.
+ * If kswapd reclaim is allowed, we can block, but only if we're called
+ * by the kswapd thread.
+ */
+ return (sc->gfp_mask & __GFP_DIRECT_RECLAIM) ||
+ ((sc->gfp_mask & __GFP_KSWAPD_RECLAIM) && current_is_kswapd());
+}
+
+static unsigned long
+panthor_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control
*sc)
+{
+ struct panthor_device *ptdev = shrinker->private_data;
+ unsigned long count;
+
+ /* We currently don't have a flag to tell when the content of a
+ * BO can be discarded.
+ */
+ if (!can_swap())
+ return 0;
+
+ /* This is racy, but that's okay because the returned count is just a
+ * hint. That's also what MSM is doing (no atomic var, it's relying on
+ * the fact unsigned long access is usually atomic), so if it's good
+ * enough for them, it's good enough for us too.
+ */
+ count = ptdev->reclaim.unused.count;
+ count += ptdev->reclaim.mmapped.count;
+
+ if (can_block(sc))
+ count += ptdev->reclaim.gpu_mapped_count;
+
+ return count ? count : SHRINK_EMPTY;
+}
+
+static bool panthor_gem_try_evict_no_resv_wait(struct drm_gem_object *obj,
+ struct ww_acquire_ctx *ticket)
+{
+ /*
+ * Track last locked entry for unwinding locks in error and
+ * success paths
+ */
+ struct panthor_gem_object *bo = to_panthor_bo(obj);
+ struct drm_gpuvm_bo *vm_bo, *last_locked = NULL;
+ enum panthor_gem_reclaim_state old_state;
+ int ret = 0;
+
+ /* To avoid potential lock ordering issue between bo_gpuva and
+ * mapping->i_mmap_rwsem, unmap the pages from CPU side before
+ * acquring the bo_gpuva lock. As the bo_resv lock is held, CPU
+ * page fault handler won't be able to map in the pages whilst
+ * eviction is in progress.
+ */
+ drm_vma_node_unmap(&bo->base.vma_node,
bo->base.dev->anon_inode->i_mapping);
+
+ /* We take this lock when walking the list to prevent
+ * insertion/deletion.
+ */
+ /* We can only trylock in that path, because
+ * - allocation might happen while some of these locks are held
+ * - lock ordering is different in other paths
+ * vm_resv -> bo_resv -> bo_gpuva
+ * vs
+ * bo_resv -> bo_gpuva -> vm_resv
+ *
+ * If we fail to lock that's fine, we back off and will get
+ * back to it later.
+ */
+ if (!mutex_trylock(&bo->base.gpuva.lock))
+ return false;
+
+ drm_gem_for_each_gpuvm_bo(vm_bo, obj) {
+ struct dma_resv *resv = drm_gpuvm_resv(vm_bo->vm);
+
+ if (resv == obj->resv)
+ continue;
+
+ if (!dma_resv_trylock(resv)) {
+ ret = -EDEADLK;
+ goto out_unlock;
+ }
+
+ last_locked = vm_bo;
+ }
+
+ /* Update the state before trying to evict the buffer, if the state was
+ * updated to something that's harder to reclaim (higher value in the
+ * enum), skip it (will be processed when the relevant LRU is).
+ */
+ panthor_gem_update_reclaim_state_locked(bo, &old_state);
+ if (old_state < bo->reclaim_state) {
+ ret = -EAGAIN;
+ goto out_unlock;
+ }
+
+ /* Couldn't teardown the GPU mappings? Skip. */
+ ret = panthor_vm_evict_bo_mappings_locked(bo);
+ if (ret)
+ goto out_unlock;
+
+ /* If everything went fine, evict the object. */
+ panthor_gem_evict_locked(bo);
+
+out_unlock:
+ if (last_locked) {
+ drm_gem_for_each_gpuvm_bo(vm_bo, obj) {
+ struct dma_resv *resv = drm_gpuvm_resv(vm_bo->vm);
+
+ if (resv == obj->resv)
+ continue;
+
+ dma_resv_unlock(resv);
+
+ if (last_locked == vm_bo)
+ break;
+ }
+ }
+ mutex_unlock(&bo->base.gpuva.lock);
+
+ return ret == 0;
+}
+
+static bool panthor_gem_try_evict(struct drm_gem_object *obj,
+ struct ww_acquire_ctx *ticket)
+{
+ struct panthor_gem_object *bo = to_panthor_bo(obj);
+
+ /* Wait was too long, skip. */
+ if (dma_resv_wait_timeout(obj->resv, DMA_RESV_USAGE_BOOKKEEP, false,
10) <= 0)
+ return false;
+
+ return panthor_gem_try_evict_no_resv_wait(&bo->base, ticket);
+}
+
+static unsigned long
+panthor_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
+{
+ struct panthor_device *ptdev = shrinker->private_data;
+ unsigned long remaining = 0;
+ unsigned long freed = 0;
+
+ if (!can_swap())
+ goto out;
+
+ freed += drm_gem_lru_scan(&ptdev->reclaim.unused,
+ sc->nr_to_scan - freed, &remaining,
+ panthor_gem_try_evict_no_resv_wait, NULL);
+ if (freed >= sc->nr_to_scan)
+ goto out;
+
+ freed += drm_gem_lru_scan(&ptdev->reclaim.mmapped,
+ sc->nr_to_scan - freed, &remaining,
+ panthor_gem_try_evict_no_resv_wait, NULL);
+ if (freed >= sc->nr_to_scan)
+ goto out;
+
+ if (!can_block(sc))
+ goto out;
+
+ freed += panthor_mmu_reclaim_priv_bos(ptdev, sc->nr_to_scan - freed,
+ &remaining,
panthor_gem_try_evict);
+ if (freed >= sc->nr_to_scan)
+ goto out;
+
+ freed += drm_gem_lru_scan(&ptdev->reclaim.gpu_mapped_shared,
+ sc->nr_to_scan - freed, &remaining,
+ panthor_gem_try_evict, NULL);
+
+out:
+#ifdef CONFIG_DEBUG_FS
+ /* This is racy, but that's okay, because this is just debugfs
+ * reporting and doesn't need to be accurate.
+ */
+ ptdev->reclaim.nr_pages_reclaimed_on_last_scan = freed;
+#endif
+
+ /* If there are things to reclaim, try a couple times before giving up.
*/
+ if (!freed && remaining > 0 &&
+ atomic_inc_return(&ptdev->reclaim.retry_count) < 2)
+ return 0;
+
+ atomic_set(&ptdev->reclaim.retry_count, 0);
+
+ if (freed)
+ return freed;
+
+ /* There's nothing left to reclaim, or the resources are contended.
Give up now. */
+ return SHRINK_STOP;
+}
+
+int panthor_gem_shrinker_init(struct panthor_device *ptdev)
+{
+ struct shrinker *shrinker;
+ int ret;
+
+ ret = drmm_mutex_init(&ptdev->base, &ptdev->reclaim.lock);
+ if (ret)
+ return ret;
+
+ INIT_LIST_HEAD(&ptdev->reclaim.vms);
+ drm_gem_lru_init(&ptdev->reclaim.unused, &ptdev->reclaim.lock);
+ drm_gem_lru_init(&ptdev->reclaim.mmapped, &ptdev->reclaim.lock);
+ drm_gem_lru_init(&ptdev->reclaim.gpu_mapped_shared,
&ptdev->reclaim.lock);
+ ptdev->reclaim.gpu_mapped_count = 0;
+
+ /* Teach lockdep about lock ordering wrt. shrinker: */
+ fs_reclaim_acquire(GFP_KERNEL);
+ might_lock(&ptdev->reclaim.lock);
+ fs_reclaim_release(GFP_KERNEL);
+
+ shrinker = shrinker_alloc(0, "drm-panthor-gem");
+ if (!shrinker)
+ return -ENOMEM;
+
+ shrinker->count_objects = panthor_gem_shrinker_count;
+ shrinker->scan_objects = panthor_gem_shrinker_scan;
+ shrinker->private_data = ptdev;
+ ptdev->reclaim.shrinker = shrinker;
+
+ shrinker_register(shrinker);
+ return 0;
+}
+
+void panthor_gem_shrinker_unplug(struct panthor_device *ptdev)
+{
+ if (ptdev->reclaim.shrinker)
+ shrinker_free(ptdev->reclaim.shrinker);
+}
+
#ifdef CONFIG_DEBUG_FS
struct gem_size_totals {
size_t size;
@@ -1184,6 +1623,7 @@ static void panthor_gem_debugfs_bo_print(struct
panthor_gem_object *bo,
struct seq_file *m,
struct gem_size_totals *totals)
{
+ enum panthor_gem_reclaim_state reclaim_state = bo->reclaim_state;
unsigned int refcount = kref_read(&bo->base.refcount);
char creator_info[32] = {};
size_t resident_size;
@@ -1219,6 +1659,8 @@ static void panthor_gem_debugfs_bo_print(struct
panthor_gem_object *bo,
totals->size += bo->base.size;
totals->resident += resident_size;
+ if (reclaim_state != PANTHOR_GEM_UNRECLAIMABLE)
+ totals->reclaimable += resident_size;
}
static void panthor_gem_debugfs_print_bos(struct panthor_device *ptdev,
@@ -1259,10 +1701,42 @@ static struct drm_info_list panthor_gem_debugfs_list[]
= {
{ "gems", panthor_gem_show_bos, 0, NULL },
};
+static int shrink_get(void *data, u64 *val)
+{
+ struct panthor_device *ptdev =
+ container_of(data, struct panthor_device, base);
+
+ *val = ptdev->reclaim.nr_pages_reclaimed_on_last_scan;
+ return 0;
+}
+
+static int shrink_set(void *data, u64 val)
+{
+ struct panthor_device *ptdev =
+ container_of(data, struct panthor_device, base);
+ struct shrink_control sc = {
+ .gfp_mask = GFP_KERNEL,
+ .nr_to_scan = val,
+ };
+
+ fs_reclaim_acquire(GFP_KERNEL);
+ if (ptdev->reclaim.shrinker)
+ panthor_gem_shrinker_scan(ptdev->reclaim.shrinker, &sc);
+ fs_reclaim_release(GFP_KERNEL);
+
+ return 0;
+}
+
+DEFINE_DEBUGFS_ATTRIBUTE(panthor_gem_debugfs_shrink_fops,
+ shrink_get, shrink_set,
+ "0x%08llx\n");
+
void panthor_gem_debugfs_init(struct drm_minor *minor)
{
drm_debugfs_create_files(panthor_gem_debugfs_list,
ARRAY_SIZE(panthor_gem_debugfs_list),
minor->debugfs_root, minor);
+ debugfs_create_file("shrink", 0600, minor->debugfs_root,
+ minor->dev, &panthor_gem_debugfs_shrink_fops);
}
#endif
diff --git a/drivers/gpu/drm/panthor/panthor_gem.h
b/drivers/gpu/drm/panthor/panthor_gem.h
index c0a18dca732c..ae0491d0b121 100644
--- a/drivers/gpu/drm/panthor/panthor_gem.h
+++ b/drivers/gpu/drm/panthor/panthor_gem.h
@@ -1,6 +1,7 @@
/* SPDX-License-Identifier: GPL-2.0 or MIT */
/* Copyright 2019 Linaro, Ltd, Rob Herring <[email protected]> */
/* Copyright 2023 Collabora ltd. */
+/* Copyright 2025 ARM Limited. All rights reserved. */
#ifndef __PANTHOR_GEM_H__
#define __PANTHOR_GEM_H__
@@ -93,6 +94,65 @@ struct panthor_gem_dev_map {
struct sg_table *sgt;
};
+/**
+ * enum panthor_gem_reclaim_state - Reclaim state of a GEM object
+ *
+ * This is defined in descending reclaimability order and some part
+ * of the code depends on that.
+ */
+enum panthor_gem_reclaim_state {
+ /**
+ * @PANTHOR_GEM_UNUSED: GEM is currently unused
+ *
+ * This can happen when the GEM was previously vmap-ed, mmap-ed,
+ * and/or GPU mapped and got unmapped. Because pages are lazily
+ * returned to the shmem layer, we want to keep a list of such
+ * BOs, because they should be fairly easy to reclaim (no need
+ * to wait for GPU to be done, and no need to tear down user
+ * mappings either).
+ */
+ PANTHOR_GEM_UNUSED,
+
+ /**
+ * @PANTHOR_GEM_MMAPPED: GEM is currently mmap-ed
+ *
+ * When a GEM has pages allocated and the mmap_count is > 0, the
+ * GEM is placed in the mmapped list. This comes right after
+ * unused because we can relatively easily tear down user mappings.
+ */
+ PANTHOR_GEM_MMAPPED,
+
+ /**
+ * @PANTHOR_GEM_GPU_MAPPED_SINGLE_VM: GEM is GPU mapped to only one VM
+ *
+ * When a GEM is mapped to a single VM, reclaim requests have more
+ * chances to succeed, because we only need to synchronize against
+ * a single GPU context. This is more annoying than reclaiming
+ * mmap-ed pages still, because we have to wait for in-flight jobs
+ * to land, and we might not be able to acquire all necessary locks
+ * at reclaim time either.
+ */
+ PANTHOR_GEM_GPU_MAPPED_SINGLE_VM,
+
+ /**
+ * @PANTHOR_GEM_GPU_MAPPED_MULTI_VM: GEM is GPU mapped to multiple VMs
+ *
+ * Like PANTHOR_GEM_GPU_MAPPED_SINGLE_VM, but the synchronization across
+ * VMs makes such BOs harder to reclaim.
+ */
+ PANTHOR_GEM_GPU_MAPPED_MULTI_VM,
+
+ /**
+ * @PANTHOR_GEM_UNRECLAIMABLE: GEM can't be reclaimed
+ *
+ * Happens when the GEM memory is pinned. It's also the state all GEM
+ * objects start in, because no memory is allocated until explicitly
+ * requested by a CPU or GPU map, meaning there's nothing to reclaim
+ * until such an allocation happens.
+ */
+ PANTHOR_GEM_UNRECLAIMABLE,
+};
+
/**
* struct panthor_gem_object - Driver specific GEM object.
*/
@@ -109,6 +169,9 @@ struct panthor_gem_object {
/** @dmap: Device mapping state */
struct panthor_gem_dev_map dmap;
+ /** @reclaim_state: Cached reclaim state */
+ enum panthor_gem_reclaim_state reclaim_state;
+
/**
* @exclusive_vm_root_gem: Root GEM of the exclusive VM this GEM object
* is attached to.
@@ -190,6 +253,11 @@ struct sg_table *
panthor_gem_get_dev_sgt(struct panthor_gem_object *bo);
int panthor_gem_pin(struct panthor_gem_object *bo);
void panthor_gem_unpin(struct panthor_gem_object *bo);
+int panthor_gem_swapin_locked(struct panthor_gem_object *bo);
+void panthor_gem_update_reclaim_state_locked(struct panthor_gem_object *bo,
+ enum panthor_gem_reclaim_state
*old_state);
+int panthor_gem_shrinker_init(struct panthor_device *ptdev);
+void panthor_gem_shrinker_unplug(struct panthor_device *ptdev);
void panthor_gem_bo_set_label(struct drm_gem_object *obj, const char *label);
void panthor_gem_kernel_bo_set_label(struct panthor_kernel_bo *bo, const char
*label);
diff --git a/drivers/gpu/drm/panthor/panthor_mmu.c
b/drivers/gpu/drm/panthor/panthor_mmu.c
index b78f694ca92b..fa8b31df85c9 100644
--- a/drivers/gpu/drm/panthor/panthor_mmu.c
+++ b/drivers/gpu/drm/panthor/panthor_mmu.c
@@ -1,6 +1,7 @@
// SPDX-License-Identifier: GPL-2.0 or MIT
/* Copyright 2019 Linaro, Ltd, Rob Herring <[email protected]> */
/* Copyright 2023 Collabora ltd. */
+/* Copyright 2025 ARM Limited. All rights reserved. */
#include <drm/drm_debugfs.h>
#include <drm/drm_drv.h>
@@ -131,6 +132,9 @@ struct panthor_vma {
* Only map related flags are accepted.
*/
u32 flags;
+
+ /** @evicted: True if the VMA has been evicted. */
+ bool evicted;
};
/**
@@ -198,6 +202,9 @@ struct panthor_vm_op_ctx {
* to allocate in ::run_job().
*/
struct sg_table *sgt;
+
+ /** @map.bo: the BO being mapped. */
+ struct panthor_gem_object *bo;
} map;
};
@@ -380,6 +387,18 @@ struct panthor_vm {
/** @locked_region.size: Size of the locked region. */
u64 size;
} locked_region;
+
+ /** @reclaim: Fields related to BO reclaim. */
+ struct {
+ /** @reclaim.lru: LRU of BOs that are only mapped to this VM. */
+ struct drm_gem_lru lru;
+
+ /**
+ * @reclaim.lru_node: Node used to insert the VM in
+ * panthor_device::reclaim::vms.
+ */
+ struct list_head lru_node;
+ } reclaim;
};
/**
@@ -684,6 +703,16 @@ int panthor_vm_active(struct panthor_vm *vm)
if (refcount_inc_not_zero(&vm->as.active_cnt))
goto out_dev_exit;
+ /* As soon as active is called, we place the VM at the end of the VM
LRU.
+ * If something fails after that, the only downside is that this VM that
+ * never became active in the first place will be reclaimed last, but
+ * that's an acceptable trade-off.
+ */
+ mutex_lock(&ptdev->reclaim.lock);
+ if (vm->reclaim.lru.count)
+ list_move_tail(&vm->reclaim.lru_node, &ptdev->reclaim.vms);
+ mutex_unlock(&ptdev->reclaim.lock);
+
/* Make sure we don't race with lock/unlock_region() calls
* happening around VM bind operations.
*/
@@ -1079,7 +1108,15 @@ static void panthor_vm_bo_free(struct drm_gpuvm_bo
*vm_bo)
{
struct panthor_gem_object *bo = to_panthor_bo(vm_bo->obj);
- panthor_gem_unpin(bo);
+ /* We couldn't call this when we unlinked, because the resv lock can't
+ * be taken in the dma signalling path, so call it now.
+ */
+ dma_resv_lock(bo->base.resv, NULL);
+ mutex_lock(&bo->base.gpuva.lock);
+ panthor_gem_update_reclaim_state_locked(bo, NULL);
+ mutex_unlock(&bo->base.gpuva.lock);
+ dma_resv_unlock(bo->base.resv);
+
kfree(vm_bo);
}
@@ -1088,6 +1125,16 @@ static void panthor_vm_cleanup_op_ctx(struct
panthor_vm_op_ctx *op_ctx,
{
u32 remaining_pt_count = op_ctx->rsvd_page_tables.count -
op_ctx->rsvd_page_tables.ptr;
+ u32 op_type = op_ctx->flags & DRM_PANTHOR_VM_BIND_OP_TYPE_MASK;
+
+ /* If this is a map operation and no BO is attached, we're being called
+ * from vm_bo_validate() and we can't acquire the VM lock because it's
+ * already held. In that case, we just skip the deferred vm_bo cleanup,
+ * which is fine, because the vm_bo validation is not calling
+ * drm_gpuvm_bo_put_deferred().
+ */
+ bool skip_deferred_cleanup = op_type == DRM_PANTHOR_VM_BIND_OP_TYPE_MAP
&&
+ !op_ctx->map.bo;
if (remaining_pt_count) {
kmem_cache_free_bulk(pt_cache, remaining_pt_count,
@@ -1100,10 +1147,16 @@ static void panthor_vm_cleanup_op_ctx(struct
panthor_vm_op_ctx *op_ctx,
if (op_ctx->map.vm_bo)
drm_gpuvm_bo_put_deferred(op_ctx->map.vm_bo);
+ if (op_ctx->map.bo) {
+ panthor_gem_unpin(op_ctx->map.bo);
+ drm_gem_object_put(&op_ctx->map.bo->base);
+ }
+
for (u32 i = 0; i < ARRAY_SIZE(op_ctx->preallocated_vmas); i++)
kfree(op_ctx->preallocated_vmas[i]);
- drm_gpuvm_bo_deferred_cleanup(&vm->base);
+ if (!skip_deferred_cleanup)
+ drm_gpuvm_bo_deferred_cleanup(&vm->base);
}
static void
@@ -1259,18 +1312,17 @@ static int panthor_vm_prepare_map_op_ctx(struct
panthor_vm_op_ctx *op_ctx,
if (ret)
goto err_cleanup;
+ drm_gem_object_get(&bo->base);
+ op_ctx->map.bo = bo;
+
sgt = panthor_gem_get_dev_sgt(bo);
if (IS_ERR(sgt)) {
- panthor_gem_unpin(bo);
ret = PTR_ERR(sgt);
goto err_cleanup;
}
- op_ctx->map.sgt = sgt;
-
preallocated_vm_bo = drm_gpuvm_bo_create(&vm->base, &bo->base);
if (!preallocated_vm_bo) {
- panthor_gem_unpin(bo);
ret = -ENOMEM;
goto err_cleanup;
}
@@ -1289,6 +1341,13 @@ static int panthor_vm_prepare_map_op_ctx(struct
panthor_vm_op_ctx *op_ctx,
dma_resv_unlock(panthor_vm_resv(vm));
}
+ /* And finally update the BO state. */
+ dma_resv_lock(bo->base.resv, NULL);
+ mutex_lock(&bo->base.gpuva.lock);
+ panthor_gem_update_reclaim_state_locked(bo, NULL);
+ mutex_unlock(&bo->base.gpuva.lock);
+ dma_resv_unlock(bo->base.resv);
+
return 0;
err_cleanup:
@@ -1876,6 +1935,10 @@ static void panthor_vm_free(struct drm_gpuvm *gpuvm)
struct panthor_vm *vm = container_of(gpuvm, struct panthor_vm, base);
struct panthor_device *ptdev = vm->ptdev;
+ mutex_lock(&ptdev->reclaim.lock);
+ list_del_init(&vm->reclaim.lru_node);
+ mutex_unlock(&ptdev->reclaim.lock);
+
mutex_lock(&vm->heaps.lock);
if (drm_WARN_ON(&ptdev->base, vm->heaps.pool))
panthor_heap_pool_destroy(vm->heaps.pool);
@@ -2089,7 +2152,7 @@ static int panthor_gpuva_sm_step_map(struct drm_gpuva_op
*op, void *priv)
panthor_vma_init(vma, op_ctx->flags & PANTHOR_VM_MAP_FLAGS);
ret = panthor_vm_map_pages(vm, op->map.va.addr,
flags_to_prot(vma->flags),
- op_ctx->map.sgt, op->map.gem.offset,
+ op_ctx->map.bo->dmap.sgt, op->map.gem.offset,
op->map.va.range);
if (ret) {
panthor_vm_op_ctx_return_vma(op_ctx, vma);
@@ -2173,21 +2236,27 @@ static int panthor_gpuva_sm_step_remap(struct
drm_gpuva_op *op,
* atomicity. panthor_vm_lock_region() bails out early if the new region
* is already part of the locked region, so no need to do this check
here.
*/
- panthor_vm_lock_region(vm, unmap_start, unmap_range);
- panthor_vm_unmap_pages(vm, unmap_start, unmap_range);
+ if (!unmap_vma->evicted) {
+ panthor_vm_lock_region(vm, unmap_start, unmap_range);
+ panthor_vm_unmap_pages(vm, unmap_start, unmap_range);
+ }
if (op->remap.prev) {
struct panthor_gem_object *bo =
to_panthor_bo(op->remap.prev->gem.obj);
u64 offset = op->remap.prev->gem.offset + unmap_start -
op->remap.prev->va.addr;
u64 size = op->remap.prev->va.addr + op->remap.prev->va.range -
unmap_start;
- ret = panthor_vm_map_pages(vm, unmap_start,
flags_to_prot(unmap_vma->flags),
- bo->dmap.sgt, offset, size);
- if (ret)
- return ret;
+ if (!unmap_vma->evicted) {
+ ret = panthor_vm_map_pages(vm, unmap_start,
+
flags_to_prot(unmap_vma->flags),
+ bo->dmap.sgt, offset, size);
+ if (ret)
+ return ret;
+ }
prev_vma = panthor_vm_op_ctx_get_vma(op_ctx);
panthor_vma_init(prev_vma, unmap_vma->flags);
+ prev_vma->evicted = unmap_vma->evicted;
}
if (op->remap.next) {
@@ -2195,13 +2264,17 @@ static int panthor_gpuva_sm_step_remap(struct
drm_gpuva_op *op,
u64 addr = op->remap.next->va.addr;
u64 size = unmap_start + unmap_range - op->remap.next->va.addr;
- ret = panthor_vm_map_pages(vm, addr,
flags_to_prot(unmap_vma->flags),
- bo->dmap.sgt,
op->remap.next->gem.offset, size);
- if (ret)
- return ret;
+ if (!unmap_vma->evicted) {
+ ret = panthor_vm_map_pages(vm, addr,
flags_to_prot(unmap_vma->flags),
+ bo->dmap.sgt,
op->remap.next->gem.offset,
+ size);
+ if (ret)
+ return ret;
+ }
next_vma = panthor_vm_op_ctx_get_vma(op_ctx);
panthor_vma_init(next_vma, unmap_vma->flags);
+ next_vma->evicted = unmap_vma->evicted;
}
drm_gpuva_remap(prev_vma ? &prev_vma->base : NULL,
@@ -2231,19 +2304,230 @@ static int panthor_gpuva_sm_step_unmap(struct
drm_gpuva_op *op,
struct panthor_vma *unmap_vma = container_of(op->unmap.va, struct
panthor_vma, base);
struct panthor_vm *vm = priv;
- panthor_vm_unmap_pages(vm, unmap_vma->base.va.addr,
- unmap_vma->base.va.range);
+ if (!unmap_vma->evicted) {
+ panthor_vm_unmap_pages(vm, unmap_vma->base.va.addr,
+ unmap_vma->base.va.range);
+ }
+
drm_gpuva_unmap(&op->unmap);
panthor_vma_unlink(unmap_vma);
return 0;
}
+void panthor_vm_update_bo_reclaim_lru_locked(struct panthor_gem_object *bo)
+{
+ struct panthor_device *ptdev = container_of(bo->base.dev, struct
panthor_device, base);
+ struct panthor_vm *vm = NULL;
+ struct drm_gpuvm_bo *vm_bo;
+
+ dma_resv_assert_held(bo->base.resv);
+ lockdep_assert_held(&bo->base.gpuva.lock);
+
+ drm_gem_for_each_gpuvm_bo(vm_bo, &bo->base) {
+ if (vm_bo->evicted)
+ continue;
+
+ /* We're only supposed to have one non-evicted vm_bo in the
list if we get
+ * there.
+ */
+ drm_WARN_ON(&ptdev->base, vm);
+ vm = container_of(vm_bo->vm, struct panthor_vm, base);
+
+ mutex_lock(&ptdev->reclaim.lock);
+ drm_gem_lru_move_tail_locked(&vm->reclaim.lru, &bo->base);
+ if (list_empty(&vm->reclaim.lru_node))
+ list_move(&vm->reclaim.lru_node, &ptdev->reclaim.vms);
+ mutex_unlock(&ptdev->reclaim.lock);
+ }
+}
+
+int panthor_vm_evict_bo_mappings_locked(struct panthor_gem_object *bo)
+{
+ struct drm_gpuvm_bo *vm_bo;
+ int ret = 0;
+
+ drm_gem_for_each_gpuvm_bo(vm_bo, &bo->base) {
+ struct panthor_vm *vm = container_of(vm_bo->vm, struct
panthor_vm, base);
+ struct drm_gpuva *va;
+
+ /* Skip already evicted GPU mappings. */
+ if (vm_bo->evicted)
+ continue;
+
+ if (!mutex_trylock(&vm->op_lock))
+ return -EDEADLK;
+
+ drm_gpuvm_bo_evict(vm_bo, true);
+ drm_gpuvm_bo_for_each_va(va, vm_bo) {
+ struct panthor_vma *vma = container_of(va, struct
panthor_vma, base);
+
+ if (vma->evicted)
+ continue;
+
+ /* If something fail in the middle of a VM_BO eviction,
the VM_BO
+ * is considered fully evicted, but some of its VMAs
might still be
+ * active. That's okay because the pages won't be
released if this
+ * function returns an error.
+ *
+ * On the next job targeting this VM, the partially
evicted VM_BO
+ * will be validated, causing all its evicted VMAs to
be repopulated
+ * before the job runs. So no GPU fault expected.
+ */
+ ret = panthor_vm_lock_region(vm, va->va.addr,
va->va.range);
+ if (ret)
+ break;
+
+ panthor_vm_unmap_pages(vm, va->va.addr, va->va.range);
+ panthor_vm_unlock_region(vm);
+ vma->evicted = true;
+ }
+
+ mutex_unlock(&vm->op_lock);
+
+ if (ret)
+ break;
+ }
+
+ return ret;
+}
+
+static struct panthor_vma *select_evicted_vma(struct drm_gpuvm_bo *vm_bo,
+ struct panthor_vm_op_ctx *op_ctx)
+{
+ struct panthor_vm *vm = container_of(vm_bo->vm, struct panthor_vm,
base);
+ struct panthor_vma *first_evicted_vma = NULL;
+ struct drm_gpuva *va;
+
+ /* Take op_lock to protect against va insertion/removal. */
+ mutex_lock(&vm->op_lock);
+ drm_gpuvm_bo_for_each_va(va, vm_bo) {
+ struct panthor_vma *vma = container_of(va, struct panthor_vma,
base);
+
+ if (vma->evicted) {
+ first_evicted_vma = vma;
+ panthor_vm_init_op_ctx(op_ctx, va->va.range,
va->va.addr, vma->flags);
+ op_ctx->map.bo_offset = va->gem.offset;
+ break;
+ }
+ }
+ mutex_unlock(&vm->op_lock);
+
+ return first_evicted_vma;
+}
+
+static int remap_evicted_vma(struct drm_gpuvm_bo *vm_bo,
+ struct panthor_vma *evicted_vma,
+ struct panthor_vm_op_ctx *op_ctx)
+{
+ struct panthor_vm *vm = container_of(vm_bo->vm, struct panthor_vm,
base);
+ struct panthor_gem_object *bo = to_panthor_bo(vm_bo->obj);
+ struct drm_gpuva *va;
+ bool found = false;
+ int ret;
+
+ ret = panthor_vm_op_ctx_prealloc_pts(op_ctx);
+ if (ret)
+ goto out_cleanup;
+
+ /* Take op_lock to protect against va insertion/removal. Note that the
+ * evicted_vma selection was done with the same lock held, but we had
+ * to release it so we can allocate PTs, because this very same lock
+ * is taken in a DMA-signalling path.
+ */
+ mutex_lock(&vm->op_lock);
+ drm_gpuvm_bo_for_each_va(va, vm_bo) {
+ struct panthor_vma *vma = container_of(va, struct panthor_vma,
base);
+
+ if (vma != evicted_vma)
+ continue;
+
+ /* Because we had to release the lock between the evicted_vma
selection
+ * and its repopulation, we can't rely solely on pointer
equality (the
+ * VMA might have been freed and a new one allocated at the
same address).
+ * If the evicted bit is still set, we're sure it's our VMA,
because
+ * population/eviction is serialized with the BO resv lock.
+ */
+ if (vma->evicted)
+ found = true;
+
+ break;
+ }
+
+ if (found) {
+ vm->op_ctx = op_ctx;
+ ret = panthor_vm_lock_region(vm, evicted_vma->base.va.addr,
+ evicted_vma->base.va.range);
+ if (!ret) {
+ ret = panthor_vm_map_pages(vm,
evicted_vma->base.va.addr,
+
flags_to_prot(evicted_vma->flags),
+ bo->dmap.sgt,
+ evicted_vma->base.gem.offset,
+ evicted_vma->base.va.range);
+ if (!ret)
+ evicted_vma->evicted = false;
+
+ panthor_vm_unlock_region(vm);
+ }
+
+ vm->op_ctx = NULL;
+ }
+
+ mutex_unlock(&vm->op_lock);
+
+out_cleanup:
+ panthor_vm_cleanup_op_ctx(op_ctx, vm);
+ return ret;
+}
+
+static int panthor_vm_restore_vmas(struct drm_gpuvm_bo *vm_bo)
+{
+ struct panthor_vm *vm = container_of(vm_bo->vm, struct panthor_vm,
base);
+ struct panthor_gem_object *bo = to_panthor_bo(vm_bo->obj);
+ struct panthor_vm_op_ctx op_ctx;
+
+ if (drm_WARN_ON_ONCE(&vm->ptdev->base, !bo->dmap.sgt))
+ return -EINVAL;
+
+ for (struct panthor_vma *vma = select_evicted_vma(vm_bo, &op_ctx);
+ vma; vma = select_evicted_vma(vm_bo, &op_ctx)) {
+ int ret;
+
+ ret = remap_evicted_vma(vm_bo, vma, &op_ctx);
+ if (ret)
+ return ret;
+ }
+
+ return 0;
+}
+
+static int panthor_vm_bo_validate(struct drm_gpuvm_bo *vm_bo,
+ struct drm_exec *exec)
+{
+ struct panthor_gem_object *bo = to_panthor_bo(vm_bo->obj);
+ int ret;
+
+ ret = panthor_gem_swapin_locked(bo);
+ if (ret)
+ return ret;
+
+ ret = panthor_vm_restore_vmas(vm_bo);
+ if (ret)
+ return ret;
+
+ drm_gpuvm_bo_evict(vm_bo, false);
+ mutex_lock(&bo->base.gpuva.lock);
+ panthor_gem_update_reclaim_state_locked(bo, NULL);
+ mutex_unlock(&bo->base.gpuva.lock);
+ return 0;
+}
+
static const struct drm_gpuvm_ops panthor_gpuvm_ops = {
.vm_free = panthor_vm_free,
.vm_bo_free = panthor_vm_bo_free,
.sm_step_map = panthor_gpuva_sm_step_map,
.sm_step_remap = panthor_gpuva_sm_step_remap,
.sm_step_unmap = panthor_gpuva_sm_step_unmap,
+ .vm_bo_validate = panthor_vm_bo_validate,
};
/**
@@ -2458,6 +2742,8 @@ panthor_vm_create(struct panthor_device *ptdev, bool
for_mcu,
vm->kernel_auto_va.start = auto_kernel_va_start;
vm->kernel_auto_va.end = vm->kernel_auto_va.start + auto_kernel_va_size
- 1;
+ drm_gem_lru_init(&vm->reclaim.lru, &ptdev->reclaim.lock);
+ INIT_LIST_HEAD(&vm->reclaim.lru_node);
INIT_LIST_HEAD(&vm->node);
INIT_LIST_HEAD(&vm->as.lru_node);
vm->as.id = -1;
@@ -2805,7 +3091,78 @@ int panthor_vm_prepare_mapped_bos_resvs(struct drm_exec
*exec, struct panthor_vm
if (ret)
return ret;
- return drm_gpuvm_prepare_objects(&vm->base, exec, slot_count);
+ ret = drm_gpuvm_prepare_objects(&vm->base, exec, slot_count);
+ if (ret)
+ return ret;
+
+ return drm_gpuvm_validate(&vm->base, exec);
+}
+
+unsigned long
+panthor_mmu_reclaim_priv_bos(struct panthor_device *ptdev,
+ unsigned int nr_to_scan, unsigned long *remaining,
+ bool (*shrink)(struct drm_gem_object *,
+ struct ww_acquire_ctx *))
+{
+ unsigned long freed = 0;
+ LIST_HEAD(remaining_vms);
+ LIST_HEAD(vms);
+
+ mutex_lock(&ptdev->reclaim.lock);
+ list_splice_init(&ptdev->reclaim.vms, &vms);
+
+ while (freed < nr_to_scan) {
+ struct panthor_vm *vm;
+
+ vm = list_first_entry_or_null(&vms, typeof(*vm),
+ reclaim.lru_node);
+ if (!vm)
+ break;
+
+ if (!kref_get_unless_zero(&vm->base.kref)) {
+ list_del_init(&vm->reclaim.lru_node);
+ continue;
+ }
+
+ mutex_unlock(&ptdev->reclaim.lock);
+
+ freed += drm_gem_lru_scan(&vm->reclaim.lru, nr_to_scan - freed,
+ remaining, shrink, NULL);
+
+ mutex_lock(&ptdev->reclaim.lock);
+
+ /* If the VM is still in the temporary list, remove it so we
+ * can proceed with the next VM.
+ */
+ if (vm == list_first_entry_or_null(&vms, typeof(*vm),
reclaim.lru_node)) {
+ list_del_init(&vm->reclaim.lru_node);
+
+ /* Keep the VM around if there are still things to
+ * reclaim, so we can preserve the LRU order when
+ * re-inserting in ptdev->reclaim.vms at the end.
+ */
+ if (vm->reclaim.lru.count > 0)
+ list_add_tail(&vm->reclaim.lru_node,
&remaining_vms);
+ }
+
+ mutex_unlock(&ptdev->reclaim.lock);
+
+ panthor_vm_put(vm);
+
+ mutex_lock(&ptdev->reclaim.lock);
+ }
+
+ /* Re-insert VMs with remaining data to reclaim at the beginning of
+ * the LRU. Note that any activeness change on the VM that happened
+ * while we were reclaiming would have moved the VM out of our
+ * temporary [remaining_]vms list, meaning anything we re-insert here
+ * preserves the LRU order.
+ */
+ list_splice_tail(&vms, &remaining_vms);
+ list_splice(&remaining_vms, &ptdev->reclaim.vms);
+ mutex_unlock(&ptdev->reclaim.lock);
+
+ return freed;
}
/**
diff --git a/drivers/gpu/drm/panthor/panthor_mmu.h
b/drivers/gpu/drm/panthor/panthor_mmu.h
index 0e268fdfdb2f..3522fbbce369 100644
--- a/drivers/gpu/drm/panthor/panthor_mmu.h
+++ b/drivers/gpu/drm/panthor/panthor_mmu.h
@@ -1,6 +1,7 @@
/* SPDX-License-Identifier: GPL-2.0 or MIT */
/* Copyright 2019 Linaro, Ltd, Rob Herring <[email protected]> */
/* Copyright 2023 Collabora ltd. */
+/* Copyright 2025 ARM Limited. All rights reserved. */
#ifndef __PANTHOR_MMU_H__
#define __PANTHOR_MMU_H__
@@ -46,6 +47,13 @@ struct panthor_vm *panthor_vm_create(struct panthor_device
*ptdev, bool for_mcu,
u64 kernel_auto_va_start,
u64 kernel_auto_va_size);
+void panthor_vm_update_bo_reclaim_lru_locked(struct panthor_gem_object *bo);
+int panthor_vm_evict_bo_mappings_locked(struct panthor_gem_object *bo);
+unsigned long
+panthor_mmu_reclaim_priv_bos(struct panthor_device *ptdev,
+ unsigned int nr_to_scan, unsigned long *remaining,
+ bool (*shrink)(struct drm_gem_object *,
+ struct ww_acquire_ctx *));
int panthor_vm_prepare_mapped_bos_resvs(struct drm_exec *exec,
struct panthor_vm *vm,
u32 slot_count);
--
2.53.0
