On Tue, Mar 16, 2021 at 10:08:10AM +0100, David Hildenbrand wrote: > On 16.03.21 09:58, Liang, Liang (Leo) wrote: > > [AMD Public Use] > > > > Hi David, > > > > root@scbu-Chachani:~# cat /proc/mtrr > > reg00: base=0x000000000 ( 0MB), size= 2048MB, count=1: write-back > > reg01: base=0x0ffe00000 ( 4094MB), size= 2MB, count=1: write-protect > > reg02: base=0x100000000 ( 4096MB), size= 16MB, count=1: write-protect > > ^ there it is > > https://wiki.osdev.org/MTRR > > "Reads allocate cache lines on a cache miss. All writes update main memory. > > Cache lines are not allocated on a write miss. Write hits invalidate the > cache line and update main memory. " > > AFAIU, writes completely bypass caches and store directly to main mamory. If > there are cache lines from a previous read, they are invalidated. So I think > especially slow will be read(addr), write(addr), read(addr), ... which is > what we have in the kstream benchmark. > > > The question is: > > who sets this up without owning the memory? > Is the memory actually special/slow or is that setting wrong? I really doubt that 16M at 0x100000000 in a system with 8G RAM would *physically* differ from the neighbouring memory. > Buggy firmware/BIOS? > Buggy device driver? [ 0.000027] MTRR default type: uncachable [ 0.000028] MTRR fixed ranges enabled: [ 0.000030] 00000-9FFFF write-back [ 0.000031] A0000-BFFFF uncachable [ 0.000032] C0000-FFFFF write-through [ 0.000033] MTRR variable ranges enabled: [ 0.000034] 0 base 000000000000 mask FFFF80000000 write-back [ 0.000036] 1 base 0000FFE00000 mask FFFFFFE00000 write-protect [ 0.000037] 2 base 000100000000 mask FFFFFF000000 write-protect As we have the range at 0x100000000 write-protected reported that early in boot I'd say it's BIOS. The question is how to reliably detect that this is a bogus setting... [ 0.000038] 3 base 0000FFDE0000 mask FFFFFFFE0000 write-protect [ 0.000039] 4 base 0000FF000000 mask FFFFFFF80000 write-protect [ 0.000040] 5 disabled [ 0.000041] 6 disabled [ 0.000042] 7 disabled [ 0.000042] TOM2: 0000000280000000 aka 10240M -- Sincerely yours, Mike. _______________________________________________ amd-gfx mailing list amd-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/amd-gfx
