Skip to content

Commit

Permalink
accel/habanalabs: postpone mem_mgr IDR destruction to hpriv_release()
Browse files Browse the repository at this point in the history
[ Upstream commit 2e8e9a8 ]

The memory manager IDR is currently destroyed when user releases the
file descriptor.
However, at this point the user context might be still held, and memory
buffers might be still in use.
Later on, calls to release those buffers will fail due to not finding
their handles in the IDR, leading to a memory leak.
To avoid this leak, split the IDR destruction from the memory manager
fini, and postpone it to hpriv_release() when there is no user context
and no buffers are used.

Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
  • Loading branch information
ttayar authored and gregkh committed May 24, 2023
1 parent 5524635 commit 840de32
Show file tree
Hide file tree
Showing 4 changed files with 23 additions and 1 deletion.
9 changes: 9 additions & 0 deletions drivers/accel/habanalabs/common/device.c
Expand Up @@ -423,6 +423,9 @@ static void hpriv_release(struct kref *ref)
mutex_destroy(&hpriv->ctx_lock);
mutex_destroy(&hpriv->restore_phase_mutex);

/* There should be no memory buffers at this point and handles IDR can be destroyed */
hl_mem_mgr_idr_destroy(&hpriv->mem_mgr);

/* Device should be reset if reset-upon-device-release is enabled, or if there is a pending
* reset that waits for device release.
*/
Expand Down Expand Up @@ -514,6 +517,10 @@ static int hl_device_release(struct inode *inode, struct file *filp)
}

hl_ctx_mgr_fini(hdev, &hpriv->ctx_mgr);

/* Memory buffers might be still in use at this point and thus the handles IDR destruction
* is postponed to hpriv_release().
*/
hl_mem_mgr_fini(&hpriv->mem_mgr);

hdev->compute_ctx_in_release = 1;
Expand Down Expand Up @@ -887,6 +894,7 @@ static int device_early_init(struct hl_device *hdev)

free_cb_mgr:
hl_mem_mgr_fini(&hdev->kernel_mem_mgr);
hl_mem_mgr_idr_destroy(&hdev->kernel_mem_mgr);
free_chip_info:
kfree(hdev->hl_chip_info);
free_prefetch_wq:
Expand Down Expand Up @@ -930,6 +938,7 @@ static void device_early_fini(struct hl_device *hdev)
mutex_destroy(&hdev->clk_throttling.lock);

hl_mem_mgr_fini(&hdev->kernel_mem_mgr);
hl_mem_mgr_idr_destroy(&hdev->kernel_mem_mgr);

kfree(hdev->hl_chip_info);

Expand Down
1 change: 1 addition & 0 deletions drivers/accel/habanalabs/common/habanalabs.h
Expand Up @@ -3861,6 +3861,7 @@ const char *hl_sync_engine_to_string(enum hl_sync_engine_type engine_type);

void hl_mem_mgr_init(struct device *dev, struct hl_mem_mgr *mmg);
void hl_mem_mgr_fini(struct hl_mem_mgr *mmg);
void hl_mem_mgr_idr_destroy(struct hl_mem_mgr *mmg);
int hl_mem_mgr_mmap(struct hl_mem_mgr *mmg, struct vm_area_struct *vma,
void *args);
struct hl_mmap_mem_buf *hl_mmap_mem_buf_get(struct hl_mem_mgr *mmg,
Expand Down
1 change: 1 addition & 0 deletions drivers/accel/habanalabs/common/habanalabs_drv.c
Expand Up @@ -237,6 +237,7 @@ int hl_device_open(struct inode *inode, struct file *filp)
out_err:
mutex_unlock(&hdev->fpriv_list_lock);
hl_mem_mgr_fini(&hpriv->mem_mgr);
hl_mem_mgr_idr_destroy(&hpriv->mem_mgr);
hl_ctx_mgr_fini(hpriv->hdev, &hpriv->ctx_mgr);
filp->private_data = NULL;
mutex_destroy(&hpriv->ctx_lock);
Expand Down
13 changes: 12 additions & 1 deletion drivers/accel/habanalabs/common/memory_mgr.c
Expand Up @@ -341,8 +341,19 @@ void hl_mem_mgr_fini(struct hl_mem_mgr *mmg)
"%s: Buff handle %u for CTX is still alive\n",
topic, id);
}
}

/* TODO: can it happen that some buffer is still in use at this point? */
/**
* hl_mem_mgr_idr_destroy() - destroy memory manager IDR.
* @mmg: parent unified memory manager
*
* Destroy the memory manager IDR.
* Shall be called when IDR is empty and no memory buffers are in use.
*/
void hl_mem_mgr_idr_destroy(struct hl_mem_mgr *mmg)
{
if (!idr_is_empty(&mmg->handles))
dev_crit(mmg->dev, "memory manager IDR is destroyed while it is not empty!\n");

idr_destroy(&mmg->handles);
}

0 comments on commit 840de32

Please sign in to comment.