[2/3] elf: Set l_contiguous to 1 for the main map in more cases
Checks
| Context |
Check |
Description |
| dj/TryBot-apply_patch |
success
|
Patch applied to master at the time it was sent
|
Commit Message
l_contiguous was not initialized at all for the main map and
always 0. This commit adds code to check if the LOAD segments
are adjacent to each other, and sets l_contiguous accordingly.
This helps _dl_find_object because it is more efficient if the
main mapping is contiguous.
Note that not all (PIE or non-PIE) binaries are contiguous in this
way because BFD ld creates executables with LOAD holes:
ELF LOAD segments creating holes in the process image on GNU/Linux
https://sourceware.org/pipermail/binutils/2022-January/119082.html
https://sourceware.org/bugzilla/show_bug.cgi?id=28743
---
elf/rtld.c | 25 +++++++++++++++++++++++++
1 file changed, 25 insertions(+)
Comments
On Mon, Jan 3, 2022 at 9:12 AM Florian Weimer via Libc-alpha
<libc-alpha@sourceware.org> wrote:
>
> l_contiguous was not initialized at all for the main map and
> always 0. This commit adds code to check if the LOAD segments
> are adjacent to each other, and sets l_contiguous accordingly.
> This helps _dl_find_object because it is more efficient if the
> main mapping is contiguous.
>
> Note that not all (PIE or non-PIE) binaries are contiguous in this
> way because BFD ld creates executables with LOAD holes:
>
> ELF LOAD segments creating holes in the process image on GNU/Linux
> https://sourceware.org/pipermail/binutils/2022-January/119082.html
> https://sourceware.org/bugzilla/show_bug.cgi?id=28743
> ---
> elf/rtld.c | 25 +++++++++++++++++++++++++
> 1 file changed, 25 insertions(+)
>
> diff --git a/elf/rtld.c b/elf/rtld.c
> index ba6e31377d..53293f5b13 100644
> --- a/elf/rtld.c
> +++ b/elf/rtld.c
> @@ -1144,6 +1144,22 @@ rtld_setup_main_map (struct link_map *main_map)
> main_map->l_map_start = ~0;
> /* And it was opened directly. */
> ++main_map->l_direct_opencount;
> + main_map->l_contiguous = 1;
> +
> + /* A PT_LOAD segment at an unexpected address will clear the
> + l_contiguous flag. The ELF specification says that PT_LOAD
> + segments need to be sorted in in increasing order, but perhaps
> + not all executables follow this requirement. Having l_contiguous
> + equal to 1 is just an optimization, so the code below does not
> + try to sort the segments in case they are unordered.
> +
> + There is one corner case in which l_contiguous is not set to 1,
> + but where it could be set: If a PIE (ET_DYN) binary is loaded by
> + glibc itself (not the kernel), it is always contiguous due to the
> + way the glibc loader works. However, the kernel loader may still
> + create holes in this case, and the code here still uses 0
> + conservatively for the glibc-loaded case, too. */
> + ElfW(Addr) expected_load_address = 0;
>
> /* Scan the program header table for the dynamic section. */
> for (const ElfW(Phdr) *ph = phdr; ph < &phdr[phnum]; ++ph)
> @@ -1207,12 +1223,21 @@ rtld_setup_main_map (struct link_map *main_map)
> if (main_map->l_map_start > mapstart)
> main_map->l_map_start = mapstart;
>
> + if (main_map->l_contiguous && expected_load_address != 0
> + && expected_load_address != mapstart)
> + main_map->l_contiguous = 0;
> +
> /* Also where it ends. */
> allocend = main_map->l_addr + ph->p_vaddr + ph->p_memsz;
> if (main_map->l_map_end < allocend)
> main_map->l_map_end = allocend;
> if ((ph->p_flags & PF_X) && allocend > main_map->l_text_end)
> main_map->l_text_end = allocend;
> +
> + /* The next expected address is the page following this load
> + segment. */
> + expected_load_address = ((allocend + GLRO(dl_pagesize) - 1)
> + & ~(GLRO(dl_pagesize) - 1));
> }
> break;
>
> --
> 2.33.1
>
LGTM.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Thanks.
@@ -1144,6 +1144,22 @@ rtld_setup_main_map (struct link_map *main_map)
main_map->l_map_start = ~0;
/* And it was opened directly. */
++main_map->l_direct_opencount;
+ main_map->l_contiguous = 1;
+
+ /* A PT_LOAD segment at an unexpected address will clear the
+ l_contiguous flag. The ELF specification says that PT_LOAD
+ segments need to be sorted in in increasing order, but perhaps
+ not all executables follow this requirement. Having l_contiguous
+ equal to 1 is just an optimization, so the code below does not
+ try to sort the segments in case they are unordered.
+
+ There is one corner case in which l_contiguous is not set to 1,
+ but where it could be set: If a PIE (ET_DYN) binary is loaded by
+ glibc itself (not the kernel), it is always contiguous due to the
+ way the glibc loader works. However, the kernel loader may still
+ create holes in this case, and the code here still uses 0
+ conservatively for the glibc-loaded case, too. */
+ ElfW(Addr) expected_load_address = 0;
/* Scan the program header table for the dynamic section. */
for (const ElfW(Phdr) *ph = phdr; ph < &phdr[phnum]; ++ph)
@@ -1207,12 +1223,21 @@ rtld_setup_main_map (struct link_map *main_map)
if (main_map->l_map_start > mapstart)
main_map->l_map_start = mapstart;
+ if (main_map->l_contiguous && expected_load_address != 0
+ && expected_load_address != mapstart)
+ main_map->l_contiguous = 0;
+
/* Also where it ends. */
allocend = main_map->l_addr + ph->p_vaddr + ph->p_memsz;
if (main_map->l_map_end < allocend)
main_map->l_map_end = allocend;
if ((ph->p_flags & PF_X) && allocend > main_map->l_text_end)
main_map->l_text_end = allocend;
+
+ /* The next expected address is the page following this load
+ segment. */
+ expected_load_address = ((allocend + GLRO(dl_pagesize) - 1)
+ & ~(GLRO(dl_pagesize) - 1));
}
break;