| Message ID | 20190618095347.3850490-1-arnd@arndb.de |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [v2] page flags: prioritize kasan bits over last-cpuid | expand |
On 6/18/19 12:53 PM, Arnd Bergmann wrote: > ARM64 randdconfig builds regularly run into a build error, especially > when NUMA_BALANCING and SPARSEMEM are enabled but not SPARSEMEM_VMEMMAP: > > #error "KASAN: not enough bits in page flags for tag" > > The last-cpuid bits are already contitional on the available space, > so the result of the calculation is a bit random on whether they > were already left out or not. > > Adding the kasan tag bits before last-cpuid makes it much more likely > to end up with a successful build here, and should be reliable for > randconfig at least, as long as that does not randomize NR_CPUS > or NODES_SHIFT but uses the defaults. > > In order for the modified check to not trigger in the x86 vdso32 code > where all constants are wrong (building with -m32), enclose all the > definitions with an #ifdef. > Why not keep "#error "KASAN: not enough bits in page flags for tag"" under "#ifdef CONFIG_KASAN_SW_TAGS" ?
On Tue, Jun 18, 2019 at 4:30 PM Andrey Ryabinin <aryabinin@virtuozzo.com> wrote: > On 6/18/19 12:53 PM, Arnd Bergmann wrote: > > ARM64 randdconfig builds regularly run into a build error, especially > > when NUMA_BALANCING and SPARSEMEM are enabled but not SPARSEMEM_VMEMMAP: > > > > #error "KASAN: not enough bits in page flags for tag" > > > > The last-cpuid bits are already contitional on the available space, > > so the result of the calculation is a bit random on whether they > > were already left out or not. > > > > Adding the kasan tag bits before last-cpuid makes it much more likely > > to end up with a successful build here, and should be reliable for > > randconfig at least, as long as that does not randomize NR_CPUS > > or NODES_SHIFT but uses the defaults. > > > > In order for the modified check to not trigger in the x86 vdso32 code > > where all constants are wrong (building with -m32), enclose all the > > definitions with an #ifdef. > > > > Why not keep "#error "KASAN: not enough bits in page flags for tag"" under "#ifdef CONFIG_KASAN_SW_TAGS" ? I think I had meant the #error to leave out the mention of KASAN, as there might be other reasons for using up all the bits, but then I did not change it in the end. Should I remove the "KASAN" word or add the #ifdef when resending? Arnd
On 6/18/19 6:30 PM, Arnd Bergmann wrote: > On Tue, Jun 18, 2019 at 4:30 PM Andrey Ryabinin <aryabinin@virtuozzo.com> wrote: >> On 6/18/19 12:53 PM, Arnd Bergmann wrote: >>> ARM64 randdconfig builds regularly run into a build error, especially >>> when NUMA_BALANCING and SPARSEMEM are enabled but not SPARSEMEM_VMEMMAP: >>> >>> #error "KASAN: not enough bits in page flags for tag" >>> >>> The last-cpuid bits are already contitional on the available space, >>> so the result of the calculation is a bit random on whether they >>> were already left out or not. >>> >>> Adding the kasan tag bits before last-cpuid makes it much more likely >>> to end up with a successful build here, and should be reliable for >>> randconfig at least, as long as that does not randomize NR_CPUS >>> or NODES_SHIFT but uses the defaults. >>> >>> In order for the modified check to not trigger in the x86 vdso32 code >>> where all constants are wrong (building with -m32), enclose all the >>> definitions with an #ifdef. >>> >> >> Why not keep "#error "KASAN: not enough bits in page flags for tag"" under "#ifdef CONFIG_KASAN_SW_TAGS" ? > > I think I had meant the #error to leave out the mention of KASAN, as there > might be other reasons for using up all the bits, but then I did not change > it in the end. > > Should I remove the "KASAN" word or add the #ifdef when resending? It seems like changing the error message is a better choice. Don't forget to remove "for tag" as well.
diff --git a/include/linux/page-flags-layout.h b/include/linux/page-flags-layout.h index 1dda31825ec4..7d794a629822 100644 --- a/include/linux/page-flags-layout.h +++ b/include/linux/page-flags-layout.h @@ -32,6 +32,7 @@ #endif /* CONFIG_SPARSEMEM */ +#ifndef BUILD_VDSO32_64 /* * page->flags layout: * @@ -76,21 +77,23 @@ #define LAST_CPUPID_SHIFT 0 #endif -#if SECTIONS_WIDTH+ZONES_WIDTH+NODES_SHIFT+LAST_CPUPID_SHIFT <= BITS_PER_LONG - NR_PAGEFLAGS +#ifdef CONFIG_KASAN_SW_TAGS +#define KASAN_TAG_WIDTH 8 +#else +#define KASAN_TAG_WIDTH 0 +#endif + +#if SECTIONS_WIDTH+ZONES_WIDTH+NODES_SHIFT+LAST_CPUPID_SHIFT+KASAN_TAG_WIDTH \ + <= BITS_PER_LONG - NR_PAGEFLAGS #define LAST_CPUPID_WIDTH LAST_CPUPID_SHIFT #else #define LAST_CPUPID_WIDTH 0 #endif -#ifdef CONFIG_KASAN_SW_TAGS -#define KASAN_TAG_WIDTH 8 #if SECTIONS_WIDTH+NODES_WIDTH+ZONES_WIDTH+LAST_CPUPID_WIDTH+KASAN_TAG_WIDTH \ > BITS_PER_LONG - NR_PAGEFLAGS #error "KASAN: not enough bits in page flags for tag" #endif -#else -#define KASAN_TAG_WIDTH 0 -#endif /* * We are going to use the flags for the page to node mapping if its in @@ -104,4 +107,5 @@ #define LAST_CPUPID_NOT_IN_PAGE_FLAGS #endif +#endif #endif /* _LINUX_PAGE_FLAGS_LAYOUT */
ARM64 randdconfig builds regularly run into a build error, especially when NUMA_BALANCING and SPARSEMEM are enabled but not SPARSEMEM_VMEMMAP: #error "KASAN: not enough bits in page flags for tag" The last-cpuid bits are already contitional on the available space, so the result of the calculation is a bit random on whether they were already left out or not. Adding the kasan tag bits before last-cpuid makes it much more likely to end up with a successful build here, and should be reliable for randconfig at least, as long as that does not randomize NR_CPUS or NODES_SHIFT but uses the defaults. In order for the modified check to not trigger in the x86 vdso32 code where all constants are wrong (building with -m32), enclose all the definitions with an #ifdef. Fixes: 2813b9c02962 ("kasan, mm, arm64: tag non slab memory allocated via pagealloc") Reviewed-by: Andrey Konovalov <andreyknvl@google.com> Signed-off-by: Arnd Bergmann <arnd@arndb.de> --- Submitted v1 in March and never followed up on the build regression, which is fixed in this version. --- include/linux/page-flags-layout.h | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-) -- 2.20.0