On Tue, Apr 14, 2020 at 04:28:10PM +0200, Greg Kroah-Hartman wrote:
On Tue, Apr 14, 2020 at 04:18:47PM +0200, �rjan Eide wrote:
Only sync the sg-list of an Ion dma-buf attachment when the attachment is actually mapped on the device.
dma-bufs may be synced at any time. It can be reached from user space via DMA_BUF_IOCTL_SYNC, so there are no guarantees from callers on when syncs may be attempted, and dma_buf_end_cpu_access() and dma_buf_begin_cpu_access() may not be paired.
Since the sg_list's dma_address isn't set up until the buffer is used on the device, and dma_map_sg() is called on it, the dma_address will be NULL if sync is attempted on the dma-buf before it's mapped on a device.
Before v5.0 (commit 55897af63091 ("dma-direct: merge swiotlb_dma_ops into the dma_direct code")) this was a problem as the dma-api (at least the swiotlb_dma_ops on arm64) would use the potentially invalid dma_address. How that failed depended on how the device handled physical address 0. If 0 was a valid address to physical ram, that page would get flushed a lot, while the actual pages in the buffer would not get synced correctly. While if 0 is an invalid physical address it may cause a fault and trigger a crash.
In v5.0 this was incidentally fixed by commit 55897af63091 ("dma-direct: merge swiotlb_dma_ops into the dma_direct code"), as this moved the dma-api to use the page pointer in the sg_list, and (for Ion buffers at least) this will always be valid if the sg_list exists at all.
But, this issue is re-introduced in v5.3 with commit 449fa54d6815 ("dma-direct: correct the physical addr in dma_direct_sync_sg_for_cpu/device") moves the dma-api back to the old behaviour and picks the dma_address that may be invalid.
dma-buf core doesn't ensure that the buffer is mapped on the device, and thus have a valid sg_list, before calling the exporter's begin_cpu_access.
Signed-off-by: �rjan Eide orjan.eide@arm.com
drivers/staging/android/ion/ion.c | 12 ++++++++++++ 1 file changed, 12 insertions(+)
Resubmit without disclaimer, sorry about that.
This seems to be part of a bigger issue where dma-buf exporters assume that their dma-buf begin_cpu_access and end_cpu_access callbacks have a certain guaranteed behavior, which isn't ensured by dma-buf core.
This patch fixes this in ion only, but it also needs to be fixed for other exporters, either handled like this in each exporter, or in dma-buf core before calling into the exporters.
diff --git a/drivers/staging/android/ion/ion.c b/drivers/staging/android/ion/ion.c index 38b51eace4f9..7b752ba0cb6d 100644 --- a/drivers/staging/android/ion/ion.c +++ b/drivers/staging/android/ion/ion.c
Now that we have the dma-buff stuff in the tree, do we even need the ion code in the kernel anymore? Can't we delete it now?
It looks like the new dma-heaps have the same issue as ion. The heap-helpers also do dma_sync_sg_for_device() unconditionally on end_cpu_access which may happen before dma_map_sg(), leading to use of the 0 dma_address in the sg list of a, yet unmapped, attachment.
It could be fixed in dma-heaps just like this patch does for ion. Is patch a valid way to fix this problem? Or, should this rather be handled in dma-buf core by tracking the mapped state of attachments there?