On Sun, Jan 03, 2021 at 03:09:14PM -0500, Andrea Arcangeli wrote: > Hello Mike, > > On Sun, Jan 03, 2021 at 03:47:53PM +0200, Mike Rapoport wrote: > > Thanks for the logs, it seems that implicitly adding reserved regions to > > memblock.memory wasn't that bright idea :) > > Would it be possible to somehow clean up the hack then? > > The only difference between the clean solution and the hack is that > the hack intended to achieved the exact same, but without adding the > reserved regions to memblock.memory. I didn't consider adding reserved regions to memblock.memory as a clean solution, this was still a hack, but I didn't think that things are that fragile. I still think we cannot rely on memblock.reserved to detect memory/zone/node sizes and the boot failure reported here confirms this. > The comment on that problematic area says the reserved area cannot be > used for DMA because of some unexplained hw issue, and that doing so > prevents booting, but since the area got reserved, even with the clean > solution, it shouldn't have never been used for DMA? > > So I can only imagine that the physical memory region is way more > problematic than just for DMA. It sounds like that anything that > touches it, including the CPU, will hang the system, not just DMA. It > sounds somewhat similar to the other e820 direct mapping issue on x86? My understanding is that the boot failed because when I implicitly added the reserved region to memblock.memory the memory size seen by free_area_init() jumped from 2G to 4G because the reserved area was close to 4G. The very first allocation would get a chunk from slightly below of 4G and as there is no real memory there, the kernel would crash. > If you want to test the hack on the arm board to check if it boots you > can use the below commit: > > https://git.kernel.org/pub/scm/linux/kernel/git/andrea/aa.git/commit/?id=c3ea2633015104ce0df33dcddbc36f57de1392bc My take is your solution would boot with this memory configuration, but I still don't think that using memblock.reserved for zone/node sizing is correct. > Thanks, > Andrea > -- Sincerely yours, Mike.
