From: Sam Day <[email protected]> On low memory boards like the DragonBoard410c (1GiB), if memory fragmentation is severe enough, large submits will fail when allocating the msm_gem_submit bookkeeping struct.
Mesa was consistently triggering this issue when running its full piglit suite on the DB410c - the vbo-subdata-many drawrangeelements test submits ~32k BOs, which results in submit_create() attempting to allocate ~1.05MB, which requires an order-9 (2MiB) allocation. When fragmentation is severe enough, this fails. Switch allocation to kvzalloc() so that larger submits can fall back to virtually contiguous memory. The msm_gem_submit struct is only used on the CPU side for bookkeeping that submit. Since submits are implicitly limited to KMALLOC_MAX_SIZE with the current kzalloc() path, we make this limit explicit with the switch to kvzalloc(). Link: https://gitlab.freedesktop.org/mesa/mesa/-/work_items/16414 Assisted-by: LLM Signed-off-by: Sam Day <[email protected]> --- The kvzalloc switch was tested and validated on my DB410c, the explicit bounds check was added later and compile-tested only. Performance has not been benchmarked. To: Rob Clark <[email protected]> To: Dmitry Baryshkov <[email protected]> To: Abhinav Kumar <[email protected]> To: Jessica Zhang <[email protected]> To: Sean Paul <[email protected]> To: Marijn Suijten <[email protected]> To: David Airlie <[email protected]> To: Simona Vetter <[email protected]> Cc: [email protected] Cc: [email protected] Cc: [email protected] Cc: [email protected] Signed-off-by: Sam Day <[email protected]> --- drivers/gpu/drm/msm/msm_gem_submit.c | 15 +++++++++++---- 1 file changed, 11 insertions(+), 4 deletions(-) diff --git a/drivers/gpu/drm/msm/msm_gem_submit.c b/drivers/gpu/drm/msm/msm_gem_submit.c index 5862db05297a7..1631bdeb64b20 100644 --- a/drivers/gpu/drm/msm/msm_gem_submit.c +++ b/drivers/gpu/drm/msm/msm_gem_submit.c @@ -42,14 +42,21 @@ static struct msm_gem_submit *submit_create(struct drm_device *dev, sz = size_add(struct_size(submit, bos, nr_bos), array_size(sizeof(submit->cmd[0]), nr_cmds)); - submit = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN); + /* We previously allocated submit with kmalloc(), but that's + * problematic on low-memory boards with high fragmentation. Now + * vmalloc can be used so we retain the previous implicit sz limit. + */ + if (sz > KMALLOC_MAX_SIZE) + return ERR_PTR(-ENOMEM); + + submit = kvzalloc(sz, GFP_KERNEL | __GFP_NOWARN); if (!submit) return ERR_PTR(-ENOMEM); submit->hw_fence = msm_fence_alloc(); if (IS_ERR(submit->hw_fence)) { ret = PTR_ERR(submit->hw_fence); - kfree(submit); + kvfree(submit); return ERR_PTR(ret); } @@ -57,7 +64,7 @@ static struct msm_gem_submit *submit_create(struct drm_device *dev, drm_client_id); if (ret) { kfree(submit->hw_fence); - kfree(submit); + kvfree(submit); return ERR_PTR(ret); } @@ -121,7 +128,7 @@ void __msm_gem_submit_destroy(struct kref *kref) for (i = 0; i < submit->nr_cmds; i++) kfree(submit->cmd[i].relocs); - kfree(submit); + kvfree(submit); } static int submit_lookup_objects(struct msm_gem_submit *submit, --- base-commit: 72d3fcf802c45d00b300f25b848a93c3a2bd7c7e change-id: 20260928-msmdrm-submit-vmalloc-02569e622ae6 Best regards, -- Sam Day <[email protected]>
