Re: [PULL 00/15] tcg patch queue

[Richard Henderson <richard.henderson@linaro.org>]

On 3/30/23 03:37, Joel Stanley wrote:
> On Tue, 28 Mar 2023 at 22:59, Richard Henderson
> <richard.henderson@linaro.org> wrote:
>>
>> The following changes since commit d37158bb2425e7ebffb167d611be01f1e9e6c86f:
>>
>>    Update version for v8.0.0-rc2 release (2023-03-28 20:43:21 +0100)
>>
>> are available in the Git repository at:
>>
>>    https://gitlab.com/rth7680/qemu.git tags/pull-tcg-20230328
>>
>> for you to fetch changes up to 87e303de70f93bf700f58412fb9b2c3ec918c4b5:
>>
>>    softmmu: Restore use of CPU watchpoint for all accelerators (2023-03-28 15:24:06 -0700)
>>
>> ----------------------------------------------------------------
>> Use a local version of GTree [#285]
>> Fix page_set_flags vs the last page of the address space [#1528]
>> Re-enable gdbstub breakpoints under KVM
>>
>> ----------------------------------------------------------------
>> Emilio Cota (2):
>>        util: import GTree as QTree
>>        tcg: use QTree instead of GTree
>>
>> Philippe Mathieu-Daudé (3):
>>        softmmu: Restrict cpu_check_watchpoint / address_matches to TCG accel
>>        softmmu/watchpoint: Add missing 'qemu/error-report.h' include
>>        softmmu: Restore use of CPU watchpoint for all accelerators
>>
>> Richard Henderson (10):
>>        linux-user: Diagnose misaligned -R size
>>        accel/tcg: Pass last not end to page_set_flags
>>        accel/tcg: Pass last not end to page_reset_target_data
>>        accel/tcg: Pass last not end to PAGE_FOR_EACH_TB
>>        accel/tcg: Pass last not end to page_collection_lock
>>        accel/tcg: Pass last not end to tb_invalidate_phys_page_range__locked
>>        accel/tcg: Pass last not end to tb_invalidate_phys_range
>>        linux-user: Pass last not end to probe_guest_base
>>        include/exec: Change reserved_va semantics to last byte
>>        linux-user/arm: Take more care allocating commpage
> 
> Thanks for getting these fixes merged.
> 
> This last one (4f5c67f8df7f26e559509c68c45e652709edd23f) causes a
> regression for me. On ppc64le, qemu-arm now segfaults. If I revert
> this one I can run executables without the assert.
> 
> The segfault looks like this:
> 
> #0  0x00000001001e44fc in stl_he_p (v=5, ptr=0x240450ffc) at
> /home/joel/qemu/include/qemu/bswap.h:260
> #1  stl_le_p (v=5, ptr=0x240450ffc) at /home/joel/qemu/include/qemu/bswap.h:302
> #2  init_guest_commpage () at ../linux-user/elfload.c:460
> #3  probe_guest_base (image_name=image_name@entry=0x1003c72e0
> <real_exec_path> "/home/joel/hello",
>      guest_loaddr=guest_loaddr@entry=65536,
> guest_hiaddr=guest_hiaddr@entry=17411743) at
> ../linux-user/elfload.c:2818
> #4  0x00000001001e50d4 in load_elf_image (image_name=0x1003c72e0
> <real_exec_path> "/home/joel/hello",
>      image_fd=<optimised out>, info=info@entry=0x7fffffffe7e8,
> pinterp_name=pinterp_name@entry=0x7fffffffe558,
>      bprm_buf=bprm_buf@entry=0x7fffffffe8d0 "\177ELF\001\001\001") at
> ../linux-user/elfload.c:3108
> #5  0x00000001001e5434 in load_elf_binary (bprm=0x7fffffffe8d0,
> info=0x7fffffffe7e8) at ../linux-user/elfload.c:3548
> #6  0x00000001001e85bc in loader_exec (fdexec=<optimised out>,
> filename=<optimised out>, argv=<optimised out>,
>      envp=<optimised out>, regs=0x7fffffffe888, infop=0x7fffffffe7e8,
> bprm=0x7fffffffe8d0) at ../linux-user/linuxload.c:155
> #7  0x0000000100046c7c in main (argc=<optimised out>,
> argv=0x7ffffffff1c8, envp=<optimised out>) at ../linux-user/main.c:892

Gah!  I've exposed the same sort of overflow conditions within target_mmap and friends.  I 
think the only short-term solution for 8.0 is to revert the last patch.


r~