When a client driver tries to obtain the HW iova for a given buffer,
the call is directed to SMMU layer. In mem_mgr each buffer entry
has it's own lock, but in SMMU it is a singleton lock for the
entire context bank. If there are multiple clients trying to
obtain iova for buffers from the same bank, it leads to waiting
on the same singleton mutex. To avoid this, when allocating
or mapping a buffer, save the iova addr of the buffer for a given
device in the mem_mgr bufq entry. When a client requests, check
the bufq entry for iova, redirect the call to SMMU only if the
address is not found in the mem_mgr. The change attempts to
reduce the number of CPU cycles in trying to obtain a buffer
iova.
CRs-Fixed: 3466755
Change-Id: I19b983ee59704eccf421ab808cbe885725571f5b
Signed-off-by: Karthik Anantha Ram <quic_kartanan@quicinc.com>
32c5a4da96