| Message ID | 550025E0.7030008@suse.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Alexander Graf <agraf@suse.de> writes: > Unfortunately however, the 32bit glibc has a bogus mmap() implementation > that hard codes 4k page size. Which is correct (it's not a page size). Andreas.
On Wednesday 11 March 2015 06:24:16 Alexander Graf wrote: > So after recompiling all of the distribution with newer binutils we now > have an openSUSE Factory tree that has 64k aligned 32bit binaries. > > Unfortunately however, the 32bit glibc has a bogus mmap() implementation > that hard codes 4k page size. > > With the patch below applied to glibc, I can successfully run 32bit user > space on Seattle with 64k PAGE_SIZE though. So I guess we'll need to fix > up glibc next. > > Do you know of anyone who's fluent enough in 32bit ARM assembly to > convert the hard coded assumptions in there to instead use a variable > that takes the actual host page size into account? I believe this is a kernel bug, and the kernel API for 32-bit emulation should always take the pgoff argument in 4KB units instead of PAGE_SIZE units, see the implementation of sys_mmap2 in arch/powerpc/kernel/sys_ppc32.c. All user space programs that call mmap2 still need to make sure that their arguments are PAGE_SIZE aligned, but the libc need not care about this here. Arnd
> Am 11.03.2015 um 07:47 schrieb Arnd Bergmann <arnd@arndb.de>: > >> On Wednesday 11 March 2015 06:24:16 Alexander Graf wrote: >> So after recompiling all of the distribution with newer binutils we now >> have an openSUSE Factory tree that has 64k aligned 32bit binaries. >> >> Unfortunately however, the 32bit glibc has a bogus mmap() implementation >> that hard codes 4k page size. >> >> With the patch below applied to glibc, I can successfully run 32bit user >> space on Seattle with 64k PAGE_SIZE though. So I guess we'll need to fix >> up glibc next. >> >> Do you know of anyone who's fluent enough in 32bit ARM assembly to >> convert the hard coded assumptions in there to instead use a variable >> that takes the actual host page size into account? > > I believe this is a kernel bug, and the kernel API for 32-bit emulation > should always take the pgoff argument in 4KB units instead of PAGE_SIZE > units, see the implementation of sys_mmap2 in > arch/powerpc/kernel/sys_ppc32.c. > > All user space programs that call mmap2 still need to make sure that > their arguments are PAGE_SIZE aligned, but the libc need not care > about this here. Awesome, that makes my life a lot easier. Let me cook up a patch that fixes the compat mmap call later today :). Alex
Hi, On 03/11/2015 08:47 AM, Arnd Bergmann wrote: > On Wednesday 11 March 2015 06:24:16 Alexander Graf wrote: >> So after recompiling all of the distribution with newer binutils we now >> have an openSUSE Factory tree that has 64k aligned 32bit binaries. >> >> Unfortunately however, the 32bit glibc has a bogus mmap() implementation >> that hard codes 4k page size. >> >> With the patch below applied to glibc, I can successfully run 32bit user >> space on Seattle with 64k PAGE_SIZE though. So I guess we'll need to fix >> up glibc next. >> >> Do you know of anyone who's fluent enough in 32bit ARM assembly to >> convert the hard coded assumptions in there to instead use a variable >> that takes the actual host page size into account? > > I believe this is a kernel bug, and the kernel API for 32-bit emulation > should always take the pgoff argument in 4KB units instead of PAGE_SIZE > units, see the implementation of sys_mmap2 in > arch/powerpc/kernel/sys_ppc32.c. > > All user space programs that call mmap2 still need to make sure that > their arguments are PAGE_SIZE aligned, but the libc need not care > about this here. What is the correct behavior for /proc/pid/pagemap, /proc/kpagecount, and /proc/kpageflags for a AArch32 process running on an AArch64 kernel with non-4K translation granule? Actual page frame number or units-of-4K frame number? Thanks, Chris
On Monday 16 March 2015 10:16:37 Christopher Covington wrote: > > On 03/11/2015 08:47 AM, Arnd Bergmann wrote: > > On Wednesday 11 March 2015 06:24:16 Alexander Graf wrote: > >> So after recompiling all of the distribution with newer binutils we now > >> have an openSUSE Factory tree that has 64k aligned 32bit binaries. > >> > >> Unfortunately however, the 32bit glibc has a bogus mmap() implementation > >> that hard codes 4k page size. > >> > >> With the patch below applied to glibc, I can successfully run 32bit user > >> space on Seattle with 64k PAGE_SIZE though. So I guess we'll need to fix > >> up glibc next. > >> > >> Do you know of anyone who's fluent enough in 32bit ARM assembly to > >> convert the hard coded assumptions in there to instead use a variable > >> that takes the actual host page size into account? > > > > I believe this is a kernel bug, and the kernel API for 32-bit emulation > > should always take the pgoff argument in 4KB units instead of PAGE_SIZE > > units, see the implementation of sys_mmap2 in > > arch/powerpc/kernel/sys_ppc32.c. > > > > All user space programs that call mmap2 still need to make sure that > > their arguments are PAGE_SIZE aligned, but the libc need not care > > about this here. > > What is the correct behavior for /proc/pid/pagemap, /proc/kpagecount, and > /proc/kpageflags for a AArch32 process running on an AArch64 kernel with > non-4K translation granule? Actual page frame number or units-of-4K frame number? > Not sure, see what PowerPC does in that case. Arnd
Index: glibc-2.21/sysdeps/unix/sysv/linux/arm/mmap.S =================================================================== --- glibc-2.21.orig/sysdeps/unix/sysv/linux/arm/mmap.S +++ glibc-2.21/sysdeps/unix/sysv/linux/arm/mmap.S @@ -36,7 +36,7 @@ ENTRY (__mmap) /* convert offset to pages */ movs ip, r5, lsl #20 bne .Linval - mov r5, r5, lsr #12 + mov r5, r5, lsr #16 /* do the syscall */ DO_CALL (mmap2, 0) Index: glibc-2.21/sysdeps/unix/sysv/linux/arm/mmap64.S =================================================================== --- glibc-2.21.orig/sysdeps/unix/sysv/linux/arm/mmap64.S +++ glibc-2.21/sysdeps/unix/sysv/linux/arm/mmap64.S @@ -43,9 +43,9 @@ ENTRY (__mmap64) cfi_rel_offset (r4, 0) cfi_remember_state movs r4, ip, lsl $20 @ check that offset is page-aligned - mov ip, ip, lsr $12 + mov ip, ip, lsr $16 it eq - movseq r4, r5, lsr $12 @ check for overflow + movseq r4, r5, lsr $16 @ check for overflow bne .Linval ldr r4, [sp, $8] @ load fd orr r5, ip, r5, lsl $20 @ compose page offset