On Mon 09-12-19 14:24:22, Logan Gunthorpe wrote: > > > On 2019-12-09 1:41 p.m., Michal Hocko wrote: > > On Mon 09-12-19 13:24:19, Logan Gunthorpe wrote: > >> > >> > >> On 2019-12-09 12:23 p.m., David Hildenbrand wrote: > >>> On 09.12.19 20:13, Logan Gunthorpe wrote: > >>>> devm_memremap_pages() is currently used by the PCI P2PDMA code to create > >>>> struct page mappings for IO memory. At present, these mappings are created > >>>> with PAGE_KERNEL which implies setting the PAT bits to be WB. However, on > >>>> x86, an mtrr register will typically override this and force the cache > >>>> type to be UC-. In the case firmware doesn't set this register it is > >>>> effectively WB and will typically result in a machine check exception > >>>> when it's accessed. > >>>> > >>>> Other arches are not currently likely to function correctly seeing they > >>>> don't have any MTRR registers to fall back on. > >>>> > >>>> To solve this, add an argument to arch_add_memory() to explicitly > >>>> set the pgprot value to a specific value. > >>>> > >>>> Of the arches that support MEMORY_HOTPLUG: x86_64, s390 and arm64 is a > >>>> simple change to pass the pgprot_t down to their respective functions > >>>> which set up the page tables. For x86_32, set the page tables explicitly > >>>> using _set_memory_prot() (seeing they are already mapped). For sh, reject > >>>> anything but PAGE_KERNEL settings -- this should be fine, for now, seeing > >>>> sh doesn't support ZONE_DEVICE anyway. > >>>> > >>>> Cc: Dan Williams <dan.j.williams@xxxxxxxxx> > >>>> Cc: David Hildenbrand <david@xxxxxxxxxx> > >>>> Cc: Michal Hocko <mhocko@xxxxxxxx> > >>>> Signed-off-by: Logan Gunthorpe <logang@xxxxxxxxxxxx> > >>>> --- > >>>> arch/arm64/mm/mmu.c | 4 ++-- > >>>> arch/ia64/mm/init.c | 5 ++++- > >>>> arch/powerpc/mm/mem.c | 4 ++-- > >>>> arch/s390/mm/init.c | 4 ++-- > >>>> arch/sh/mm/init.c | 5 ++++- > >>>> arch/x86/mm/init_32.c | 7 ++++++- > >>>> arch/x86/mm/init_64.c | 4 ++-- > >>>> include/linux/memory_hotplug.h | 2 +- > >>>> mm/memory_hotplug.c | 2 +- > >>>> mm/memremap.c | 2 +- > >>>> 10 files changed, 25 insertions(+), 14 deletions(-) > >>>> > >>>> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c > >>>> index 60c929f3683b..48b65272df15 100644 > >>>> --- a/arch/arm64/mm/mmu.c > >>>> +++ b/arch/arm64/mm/mmu.c > >>>> @@ -1050,7 +1050,7 @@ int p4d_free_pud_page(p4d_t *p4d, unsigned long addr) > >>>> } > >>>> > >>>> #ifdef CONFIG_MEMORY_HOTPLUG > >>>> -int arch_add_memory(int nid, u64 start, u64 size, > >>>> +int arch_add_memory(int nid, u64 start, u64 size, pgprot_t prot, > >>>> struct mhp_restrictions *restrictions) > >>> > >>> Can we fiddle that into "struct mhp_restrictions" instead? > >> > >> Yes, if that's what people want, it's pretty trivial to do. I chose not > >> to do it that way because it doesn't get passed down to add_pages() and > >> it's not really a "restriction". If I don't hear any objections, I will > >> do that for v2. > > > > I do agree that restriction is not the best fit. But I consider prot > > argument to complicate the API to all users even though it is not really > > clear whether we are going to have many users really benefiting from it. > > Look at the vmalloc API and try to find how many users of __vmalloc do > > not use PAGE_KERNEL. > > > > So I can see two options. One of them is to add arch_add_memory_prot > > that would allow to have give and extra prot argument or simply call > > an arch independent API to change the protection after arch_add_memory. > > The later sounds like much less code. The memory shouldn't be in use by > > anybody at that stage yet AFAIU. Maybe there even is an API like that. > > Yes, well, we tried something like this by calling set_memory_wc() > inside memremap_pages(); but on large bars (tens of GB) it was too slow > (taking several seconds to complete) and on some hosts actually hit CPU > watchdog errors. Which looks like something to fix independently. > So at the very least we'd have to add some cpu_relax() calls to that > path. And it's also the case that set_memory_wc() is x86 only right now. > So we'd have to create a new general interface to walk and fixup page > tables for all arches. > > But, in my opinion, setting up all those page tables twice is too large > of an overhead and it's better to just add them correctly the first > time. The changes I propose to do this aren't really a lot of code and > probably less than creating a new interface for all arches. OK, fair enough. Then I would suggest going with arch_add_memory_prot then unless there is a wider disagreement witht that. -- Michal Hocko SUSE Labs
