Re: [PATCH] linux-user: Fix loading of BSS segments

[Giuseppe Musacchio <thatlemon@gmail.com>]

On 19/12/20 11:55, Laurent Vivier wrote:
> Le 17/12/2020 à 11:17, Giuseppe Musacchio a écrit :
>> Some ELF binaries encode the .bss section as an extension of the data
>> ones by setting the segment p_memsz > p_filesz. Some other binaries take
>> a different route and encode it as a stand-alone PT_LOAD segment with
>> p_filesz = 0 and p_memsz > 0.
>>
>> Both the encodings are actually correct per ELF specification but the
>> ELF loader had some troubles in handling the former: with the old logic
>> it was very likely to get Qemu to crash in zero_bss when trying to
>> access unmapped memory.
>>
>> zero_bss isn't meant to allocate whole zero-filled segments but to
>> "complete" a previously mapped segment with the needed zero bits.
>>
>> The fix is pretty simple, if the segment is completely zero-filled we
>> simply allocate one or more pages (according to p_memsz) and avoid
>> calling zero_bss altogether.
> 
> So, the current code manages the bss segment when the memory page has already
> been allocated for the data segment by zeroing it:
> 
> +----------------------------------+
>  PAGE                              |
>  ----------+------------+          |
>  DATA      |   BSS      |          |
> 
> So your patch fixes the case when there is no data segment and thus no page
> to complete:
> 
> +----------------------------------+
>  PAGE                              |
>  ----------+                       |
>  BSS       |                       |
> 
> 
> But could we have a case where the BSS starts in a page allocated for the
> data segment but needs more pages?
> 
> +----------------------------------+----------------------------------+
>  PAGE                              | PAGE                             |
>  ------------------------+----------------------------+               |
>  DATA                    | BSS                        |               |
> 
> In this case we should also allocate the page, and the previous case is only a
> special case (data segment = 0) of this one.
> 

If I correctly understand your example here this case should be already
handled by zero_bss: the first chunk of .bss is mapped as part of the
vaddr_len mapping and is zeroed by the memset, while the other chunk/page
is mapped in the `host_map_start < host_end` clause.

> so something like (approxymately):
> 
> if (eppnt->p_filesz != 0) {
>    target_mmap()
>    if (vaddr_ef < vaddr_mem) {
>        zero_bss(vaddr_ef, MIN(vaddr_mem, vaddr_ps + vaddr_len))
>    }
> }
> if (vaddr_ps + vaddr_len < vaddr_mem) {
>   target_mmap(vaddr_ps + vaddr_len, vaddr_ps + vaddr_len - vaddr_mem - 1,
>               elf_prot, MAP_PRIVATE | MAP_FIXED | MAP_ANONYMOUS, -1, 0);
> }
> 
> I think your fix is correct, but I'm wondering if we can have something more
> generic, if we can cover an other possible case.
> 
> If you think we don't need to introduce more complexity for a case that can't
> happen I will queue your patch as is.
> 
> Thanks,
> Laurent
> 
>> Signed-off-by: Giuseppe Musacchio <thatlemon@gmail.com>
>> ---
>>  linux-user/elfload.c | 30 ++++++++++++++++++++----------
>>  1 file changed, 20 insertions(+), 10 deletions(-)
>>
>> diff --git a/linux-user/elfload.c b/linux-user/elfload.c
>> index 0b02a92602..a16c240e0f 100644
>> --- a/linux-user/elfload.c
>> +++ b/linux-user/elfload.c
>> @@ -2776,14 +2776,16 @@ static void load_elf_image(const char *image_name, int image_fd,
>>              vaddr = load_bias + eppnt->p_vaddr;
>>              vaddr_po = TARGET_ELF_PAGEOFFSET(vaddr);
>>              vaddr_ps = TARGET_ELF_PAGESTART(vaddr);
>> -            vaddr_len = TARGET_ELF_PAGELENGTH(eppnt->p_filesz + vaddr_po);
>> +
>> +            vaddr_ef = vaddr + eppnt->p_filesz;
>> +            vaddr_em = vaddr + eppnt->p_memsz;
>>  
>>              /*
>> -             * Some segments may be completely empty without any backing file
>> -             * segment, in that case just let zero_bss allocate an empty buffer
>> -             * for it.
>> +             * Some segments may be completely empty, with a non-zero p_memsz
>> +             * but no backing file segment.
>>               */
>>              if (eppnt->p_filesz != 0) {
>> +                vaddr_len = TARGET_ELF_PAGELENGTH(eppnt->p_filesz + vaddr_po);
>>                  error = target_mmap(vaddr_ps, vaddr_len, elf_prot,
>>                                      MAP_PRIVATE | MAP_FIXED,
>>                                      image_fd, eppnt->p_offset - vaddr_po);
>> @@ -2791,14 +2793,22 @@ static void load_elf_image(const char *image_name, int image_fd,
>>                  if (error == -1) {
>>                      goto exit_mmap;
>>                  }
>> -            }
>>  
>> -            vaddr_ef = vaddr + eppnt->p_filesz;
>> -            vaddr_em = vaddr + eppnt->p_memsz;
>> +                /*
>> +                 * If the load segment requests extra zeros (e.g. bss), map it.
>> +                 */
>> +                if (eppnt->p_filesz < eppnt->p_memsz) {
>> +                    zero_bss(vaddr_ef, vaddr_em, elf_prot);
>> +                }
>> +            } else if (eppnt->p_memsz != 0) {
>> +                vaddr_len = TARGET_ELF_PAGELENGTH(eppnt->p_memsz + vaddr_po);
>> +                error = target_mmap(vaddr_ps, vaddr_len, elf_prot,
>> +                                    MAP_PRIVATE | MAP_FIXED | MAP_ANONYMOUS,
>> +                                    -1, 0);
>>  
>> -            /* If the load segment requests extra zeros (e.g. bss), map it.  */
>> -            if (vaddr_ef < vaddr_em) {
>> -                zero_bss(vaddr_ef, vaddr_em, elf_prot);
>> +                if (error == -1) {
>> +                    goto exit_mmap;
>> +                }
>>              }
>>  
>>              /* Find the full program boundaries.  */
>>
>