Message ID | 20170809200755.11234-8-tycho@docker.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On Wed, Aug 09, 2017 at 02:07:52PM -0600, Tycho Andersen wrote: > From: Juerg Haefliger <juerg.haefliger@hpe.com> > > If the page is unmapped by XPFO, a data cache flush results in a fatal > page fault. So don't flush in that case. Do you have an example callchain where that happens? We might need to shuffle things around to cater for that case. > @@ -30,7 +31,9 @@ void sync_icache_aliases(void *kaddr, unsigned long len) > unsigned long addr = (unsigned long)kaddr; > > if (icache_is_aliasing()) { > - __clean_dcache_area_pou(kaddr, len); > + /* Don't flush if the page is unmapped by XPFO */ > + if (!xpfo_page_is_unmapped(virt_to_page(kaddr))) > + __clean_dcache_area_pou(kaddr, len); > __flush_icache_all(); > } else { > flush_icache_range(addr, addr + len); I don't think this patch is correct. If data cache maintenance is required in the absence of XPFO, I don't see why it wouldn't be required in the presence of XPFO. I'm not immediately sure why the non-aliasing case misses data cache maintenance. I couldn't spot where that happens otherwise. On a more general note, in future it would be good to Cc the arm64 maintainers and the linux-arm-kernel mailing list for patches affecting arm64. Thanks, Mark.
On Sat, Aug 12, 2017 at 12:57:37PM +0100, Mark Rutland wrote: > On Wed, Aug 09, 2017 at 02:07:52PM -0600, Tycho Andersen wrote: > > From: Juerg Haefliger <juerg.haefliger@hpe.com> > > > > If the page is unmapped by XPFO, a data cache flush results in a fatal > > page fault. So don't flush in that case. > > Do you have an example callchain where that happens? We might need to shuffle > things around to cater for that case. > > > @@ -30,7 +31,9 @@ void sync_icache_aliases(void *kaddr, unsigned long len) > > unsigned long addr = (unsigned long)kaddr; > > > > if (icache_is_aliasing()) { > > - __clean_dcache_area_pou(kaddr, len); > > + /* Don't flush if the page is unmapped by XPFO */ > > + if (!xpfo_page_is_unmapped(virt_to_page(kaddr))) > > + __clean_dcache_area_pou(kaddr, len); > > __flush_icache_all(); > > } else { > > flush_icache_range(addr, addr + len); > > I don't think this patch is correct. If data cache maintenance is required in > the absence of XPFO, I don't see why it wouldn't be required in the presence of > XPFO. > > I'm not immediately sure why the non-aliasing case misses data cache > maintenance. I couldn't spot where that happens otherwise. From another look, it seems that flush_icache_range() performs the D-cache maintenance internally, so it's just a slightly misleading name. However, __flush_icache_all() does not, and does need separate D-cache maintenance. Thanks, Mark.
Hi Mark, First, thanks for taking a look! On Sat, Aug 12, 2017 at 12:57:37PM +0100, Mark Rutland wrote: > On Wed, Aug 09, 2017 at 02:07:52PM -0600, Tycho Andersen wrote: > > From: Juerg Haefliger <juerg.haefliger@hpe.com> > > > > If the page is unmapped by XPFO, a data cache flush results in a fatal > > page fault. So don't flush in that case. > > Do you have an example callchain where that happens? We might need to shuffle > things around to cater for that case. Here's one from the other branch (i.e. xpfo_page_is_unmapped() is true): [ 15.487293] CPU: 2 PID: 1633 Comm: plymouth Not tainted 4.13.0-rc4-c2+ #242 [ 15.487295] Hardware name: Hardkernel ODROID-C2 (DT) [ 15.487297] Call trace: [ 15.487313] [<ffff0000080884f0>] dump_backtrace+0x0/0x248 [ 15.487317] [<ffff00000808878c>] show_stack+0x14/0x20 [ 15.487324] [<ffff000008b3e1b8>] dump_stack+0x98/0xb8 [ 15.487329] [<ffff000008098bb4>] sync_icache_aliases+0x84/0x98 [ 15.487332] [<ffff000008098c74>] __sync_icache_dcache+0x64/0x88 [ 15.487337] [<ffff0000081d4814>] alloc_set_pte+0x4ec/0x6b8 [ 15.487342] [<ffff00000819d920>] filemap_map_pages+0x350/0x360 [ 15.487344] [<ffff0000081d4ccc>] do_fault+0x28c/0x568 [ 15.487347] [<ffff0000081d67a0>] __handle_mm_fault+0x410/0xd08 [ 15.487350] [<ffff0000081d7164>] handle_mm_fault+0xcc/0x1a8 [ 15.487352] [<ffff000008098580>] do_page_fault+0x270/0x380 [ 15.487355] [<ffff00000808128c>] do_mem_abort+0x3c/0x98 [ 15.487358] Exception stack(0xffff800061dabe20 to 0xffff800061dabf50) [ 15.487362] be20: 0000000000000000 0000800062e19000 ffffffffffffffff 0000ffff8f64ddc8 [ 15.487365] be40: ffff800061dabe80 ffff000008238810 ffff800061d80330 0000000000000018 [ 15.487368] be60: ffffffffffffffff 0000ffff8f5ba958 ffff800061d803d0 ffff800067132e18 [ 15.487370] be80: 0000000000000000 ffff800061d80d08 0000000000000000 0000000000000019 [ 15.487373] bea0: 000000002bd3d0f0 0000000000000000 0000000000000019 ffff800067132e00 [ 15.487376] bec0: 0000000000000000 0000ffff8f657220 0000000000000000 0000000000000000 [ 15.487379] bee0: 8080808000000000 0000000000000000 0000000080808080 fefefeff6f6b6467 [ 15.487381] bf00: 7f7f7f7f7f7f7f7f 000000002bd3fb40 0101010101010101 0000000000000020 [ 15.487384] bf20: 00000000004072b0 00000000004072e0 0000000000000000 0000ffff8f6b2000 [ 15.487386] bf40: 0000ffff8f66b190 0000ffff8f576380 [ 15.487389] [<ffff000008082b74>] el0_da+0x20/0x24 > > @@ -30,7 +31,9 @@ void sync_icache_aliases(void *kaddr, unsigned long len) > > unsigned long addr = (unsigned long)kaddr; > > > > if (icache_is_aliasing()) { > > - __clean_dcache_area_pou(kaddr, len); > > + /* Don't flush if the page is unmapped by XPFO */ > > + if (!xpfo_page_is_unmapped(virt_to_page(kaddr))) > > + __clean_dcache_area_pou(kaddr, len); > > __flush_icache_all(); > > } else { > > flush_icache_range(addr, addr + len); > > I don't think this patch is correct. If data cache maintenance is required in > the absence of XPFO, I don't see why it wouldn't be required in the presence of > XPFO. Ok. I suppose we could do re-map like we do for dma; or is there some re-arrangement of things you can see that would help? > I'm not immediately sure why the non-aliasing case misses data cache > maintenance. I couldn't spot where that happens otherwise. > > On a more general note, in future it would be good to Cc the arm64 maintainers > and the linux-arm-kernel mailing list for patches affecting arm64. Yes, I thought about doing that for the series, but since it has x86 patches too, I didn't want to spam everyone :). I'll just add x86/arm lists to CC in the patches in the future. If there's some better way, let me know. Tycho
On Mon, Aug 14, 2017 at 02:27:27PM -0600, Tycho Andersen wrote: > Hi Mark, > > First, thanks for taking a look! > > On Sat, Aug 12, 2017 at 12:57:37PM +0100, Mark Rutland wrote: > > On Wed, Aug 09, 2017 at 02:07:52PM -0600, Tycho Andersen wrote: > > > From: Juerg Haefliger <juerg.haefliger@hpe.com> > > > > > > If the page is unmapped by XPFO, a data cache flush results in a fatal > > > page fault. So don't flush in that case. > > > > Do you have an example callchain where that happens? We might need to shuffle > > things around to cater for that case. > > Here's one from the other branch (i.e. xpfo_page_is_unmapped() is true): > > [ 15.487293] CPU: 2 PID: 1633 Comm: plymouth Not tainted 4.13.0-rc4-c2+ #242 > [ 15.487295] Hardware name: Hardkernel ODROID-C2 (DT) > [ 15.487297] Call trace: > [ 15.487313] [<ffff0000080884f0>] dump_backtrace+0x0/0x248 > [ 15.487317] [<ffff00000808878c>] show_stack+0x14/0x20 > [ 15.487324] [<ffff000008b3e1b8>] dump_stack+0x98/0xb8 > [ 15.487329] [<ffff000008098bb4>] sync_icache_aliases+0x84/0x98 > [ 15.487332] [<ffff000008098c74>] __sync_icache_dcache+0x64/0x88 > [ 15.487337] [<ffff0000081d4814>] alloc_set_pte+0x4ec/0x6b8 > [ 15.487342] [<ffff00000819d920>] filemap_map_pages+0x350/0x360 > [ 15.487344] [<ffff0000081d4ccc>] do_fault+0x28c/0x568 > [ 15.487347] [<ffff0000081d67a0>] __handle_mm_fault+0x410/0xd08 > [ 15.487350] [<ffff0000081d7164>] handle_mm_fault+0xcc/0x1a8 > [ 15.487352] [<ffff000008098580>] do_page_fault+0x270/0x380 > [ 15.487355] [<ffff00000808128c>] do_mem_abort+0x3c/0x98 > [ 15.487358] Exception stack(0xffff800061dabe20 to 0xffff800061dabf50) > [ 15.487362] be20: 0000000000000000 0000800062e19000 ffffffffffffffff 0000ffff8f64ddc8 > [ 15.487365] be40: ffff800061dabe80 ffff000008238810 ffff800061d80330 0000000000000018 > [ 15.487368] be60: ffffffffffffffff 0000ffff8f5ba958 ffff800061d803d0 ffff800067132e18 > [ 15.487370] be80: 0000000000000000 ffff800061d80d08 0000000000000000 0000000000000019 > [ 15.487373] bea0: 000000002bd3d0f0 0000000000000000 0000000000000019 ffff800067132e00 > [ 15.487376] bec0: 0000000000000000 0000ffff8f657220 0000000000000000 0000000000000000 > [ 15.487379] bee0: 8080808000000000 0000000000000000 0000000080808080 fefefeff6f6b6467 > [ 15.487381] bf00: 7f7f7f7f7f7f7f7f 000000002bd3fb40 0101010101010101 0000000000000020 > [ 15.487384] bf20: 00000000004072b0 00000000004072e0 0000000000000000 0000ffff8f6b2000 > [ 15.487386] bf40: 0000ffff8f66b190 0000ffff8f576380 > [ 15.487389] [<ffff000008082b74>] el0_da+0x20/0x24 > > > > @@ -30,7 +31,9 @@ void sync_icache_aliases(void *kaddr, unsigned long len) > > > unsigned long addr = (unsigned long)kaddr; > > > > > > if (icache_is_aliasing()) { > > > - __clean_dcache_area_pou(kaddr, len); > > > + /* Don't flush if the page is unmapped by XPFO */ > > > + if (!xpfo_page_is_unmapped(virt_to_page(kaddr))) > > > + __clean_dcache_area_pou(kaddr, len); > > > __flush_icache_all(); > > > } else { > > > flush_icache_range(addr, addr + len); > > > > I don't think this patch is correct. If data cache maintenance is required in > > the absence of XPFO, I don't see why it wouldn't be required in the presence of > > XPFO. > > Ok. I suppose we could do re-map like we do for dma; or is there some > re-arrangement of things you can see that would help? Creating a temporary mapping (which I guess you do for DMA) should work. We might be able to perform the maintenance before we unmap the linear map alias, but I guess this might not be possible if the that logic is too far removed from the PTE manipulation. > > On a more general note, in future it would be good to Cc the arm64 maintainers > > and the linux-arm-kernel mailing list for patches affecting arm64. > > Yes, I thought about doing that for the series, but since it has x86 patches > too, I didn't want to spam everyone :). I'll just add x86/arm lists to CC in > the patches in the future. If there's some better way, let me know. That sounds fine; thanks. Mark.
diff --git a/arch/arm64/mm/flush.c b/arch/arm64/mm/flush.c index 21a8d828cbf4..e17a063b2df2 100644 --- a/arch/arm64/mm/flush.c +++ b/arch/arm64/mm/flush.c @@ -20,6 +20,7 @@ #include <linux/export.h> #include <linux/mm.h> #include <linux/pagemap.h> +#include <linux/xpfo.h> #include <asm/cacheflush.h> #include <asm/cache.h> @@ -30,7 +31,9 @@ void sync_icache_aliases(void *kaddr, unsigned long len) unsigned long addr = (unsigned long)kaddr; if (icache_is_aliasing()) { - __clean_dcache_area_pou(kaddr, len); + /* Don't flush if the page is unmapped by XPFO */ + if (!xpfo_page_is_unmapped(virt_to_page(kaddr))) + __clean_dcache_area_pou(kaddr, len); __flush_icache_all(); } else { flush_icache_range(addr, addr + len);