Hi Shengyu, On Fri, 17 Feb 2023 22:14:13 +0800, Shengyu Qu wrote: > Seems we could use GFP_DMA32 flag to limit memory required by driver into > upper size range(actually using ZONE_DMA32 configured by device tree). Just > some driver modification needed. I don't think the GFP_DMA32 flag works with DmaBuf import. The buffer may be allocated by some other driver that is able to address more than 4G and imported into the RGA driver. In this case, limiting the allocations is not enough, but we would still need error handling in the map function for buffers that cannot be addressed by the RGA. I guess we need both, a limit for the allocation and error checking for the map. Michael > Maybe Nicolas could help testing? I would > > like to fix this, but I don't have much free time these days. > > Best regards, > > Shengyu > > > Hi, > > > > On Sun, 22 Jan 2023 00:50:37 +0800, Shengyu Qu wrote: > > > Since we have the over-4GB problem now, should we mark this problem as a > > > TODO or something? > > I am not really sure where to put such a TODO to make it visible for people > > that are running into the issue and to make sure that it is removed once it is > > fixed. > > > > Maybe it would be better to add error handling to the rga_buf_map function to > > fail if the address of the buffer that should be mapped has the upper 32 bit > > set and print a warning. Furthermore, the driver would be able to skip the > > buffer and prevent potential memory corruption caused by the erroneous > > mapping. > > > > Unfortunately, I don't have hardware that allows me to test this. > > > > Michael
