| Message ID | 20250213143558.10504-1-farosas@suse.de (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [RFC] elfload: Fix alignment when unmapping excess reservation | expand |
On 2/13/25 06:35, Fabiano Rosas wrote: > When complying with the alignment requested in the ELF and unmapping > the excess reservation, having align_end not aligned to the guest page > causes the unmap to be rejected by the alignment check at > target_munmap and later brk adjustments hit an EEXIST. > > Fix by aligning the start of region to be unmapped. > > Fixes: c81d1fafa6 ("linux-user: Honor elf alignment when placing images") > Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1913 > Signed-off-by: Fabiano Rosas <farosas@suse.de> > --- > In the bug there was mention of the vdso landing in the wrong spot, > but I don't see evidence of this in my testing. Looking at the > addresses in the bug report, there seems to have been a mistake > because I don't see an overlap there either. > --- > linux-user/elfload.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/linux-user/elfload.c b/linux-user/elfload.c > index a2c152e5ad..05ee5e74fd 100644 > --- a/linux-user/elfload.c > +++ b/linux-user/elfload.c > @@ -3351,7 +3351,7 @@ static void load_elf_image(const char *image_name, const ImageSource *src, > > if (align_size != reserve_size) { > abi_ulong align_addr = ROUND_UP(load_addr, align); > - abi_ulong align_end = align_addr + reserve_size; > + abi_ulong align_end = TARGET_PAGE_ALIGN(align_addr + reserve_size); > abi_ulong load_end = load_addr + align_size; Both align_end and load_end must be aligned. r~
On 2/15/25 11:06, Richard Henderson wrote: > On 2/13/25 06:35, Fabiano Rosas wrote: >> When complying with the alignment requested in the ELF and unmapping >> the excess reservation, having align_end not aligned to the guest page >> causes the unmap to be rejected by the alignment check at >> target_munmap and later brk adjustments hit an EEXIST. >> >> Fix by aligning the start of region to be unmapped. >> >> Fixes: c81d1fafa6 ("linux-user: Honor elf alignment when placing images") >> Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1913 >> Signed-off-by: Fabiano Rosas <farosas@suse.de> >> --- >> In the bug there was mention of the vdso landing in the wrong spot, >> but I don't see evidence of this in my testing. Looking at the >> addresses in the bug report, there seems to have been a mistake >> because I don't see an overlap there either. >> --- >> linux-user/elfload.c | 2 +- >> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> diff --git a/linux-user/elfload.c b/linux-user/elfload.c >> index a2c152e5ad..05ee5e74fd 100644 >> --- a/linux-user/elfload.c >> +++ b/linux-user/elfload.c >> @@ -3351,7 +3351,7 @@ static void load_elf_image(const char *image_name, const >> ImageSource *src, >> if (align_size != reserve_size) { >> abi_ulong align_addr = ROUND_UP(load_addr, align); >> - abi_ulong align_end = align_addr + reserve_size; >> + abi_ulong align_end = TARGET_PAGE_ALIGN(align_addr + reserve_size); >> abi_ulong load_end = load_addr + align_size; > > Both align_end and load_end must be aligned. Bah, hit ctrl-enter, not enter. I'll fix and queue. r~
13.02.2025 17:35, Fabiano Rosas wrote: > When complying with the alignment requested in the ELF and unmapping > the excess reservation, having align_end not aligned to the guest page > causes the unmap to be rejected by the alignment check at > target_munmap and later brk adjustments hit an EEXIST. > > Fix by aligning the start of region to be unmapped. > > Fixes: c81d1fafa6 ("linux-user: Honor elf alignment when placing images") > Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1913 > Signed-off-by: Fabiano Rosas <farosas@suse.de> Is this a qemu-stable material? That issue was quite hot.. Thanks, /mjt
Michael Tokarev <mjt@tls.msk.ru> writes: > 13.02.2025 17:35, Fabiano Rosas wrote: >> When complying with the alignment requested in the ELF and unmapping >> the excess reservation, having align_end not aligned to the guest page >> causes the unmap to be rejected by the alignment check at >> target_munmap and later brk adjustments hit an EEXIST. >> >> Fix by aligning the start of region to be unmapped. >> >> Fixes: c81d1fafa6 ("linux-user: Honor elf alignment when placing images") >> Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1913 >> Signed-off-by: Fabiano Rosas <farosas@suse.de> > > Is this a qemu-stable material? That issue was quite hot.. Yes, I think it's good for stable. > > Thanks, > > /mjt
diff --git a/linux-user/elfload.c b/linux-user/elfload.c index a2c152e5ad..05ee5e74fd 100644 --- a/linux-user/elfload.c +++ b/linux-user/elfload.c @@ -3351,7 +3351,7 @@ static void load_elf_image(const char *image_name, const ImageSource *src, if (align_size != reserve_size) { abi_ulong align_addr = ROUND_UP(load_addr, align); - abi_ulong align_end = align_addr + reserve_size; + abi_ulong align_end = TARGET_PAGE_ALIGN(align_addr + reserve_size); abi_ulong load_end = load_addr + align_size; if (align_addr != load_addr) {
When complying with the alignment requested in the ELF and unmapping the excess reservation, having align_end not aligned to the guest page causes the unmap to be rejected by the alignment check at target_munmap and later brk adjustments hit an EEXIST. Fix by aligning the start of region to be unmapped. Fixes: c81d1fafa6 ("linux-user: Honor elf alignment when placing images") Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1913 Signed-off-by: Fabiano Rosas <farosas@suse.de> --- In the bug there was mention of the vdso landing in the wrong spot, but I don't see evidence of this in my testing. Looking at the addresses in the bug report, there seems to have been a mistake because I don't see an overlap there either. --- linux-user/elfload.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)