diff options
author | Florian Weimer <fweimer@redhat.com> | 2022-09-20 11:00:42 +0200 |
---|---|---|
committer | Florian Weimer <fweimer@redhat.com> | 2022-09-20 11:00:42 +0200 |
commit | 1df71d32fe5f5905ffd5d100e5e9ca8ad6210891 (patch) | |
tree | 6a747a087b99a4a5e0fbdcbe8959d378f0996170 /elf/dl-sort-maps.c | |
parent | 766b73768b290b303f5b56268c6c0d588d5a9267 (diff) | |
download | glibc-1df71d32fe5f5905ffd5d100e5e9ca8ad6210891.zip glibc-1df71d32fe5f5905ffd5d100e5e9ca8ad6210891.tar.gz glibc-1df71d32fe5f5905ffd5d100e5e9ca8ad6210891.tar.bz2 |
elf: Implement force_first handling in _dl_sort_maps_dfs (bug 28937)
The implementation in _dl_close_worker requires that the first
element of l_initfini is always this very map (“We are always the
zeroth entry, and since we don't include ourselves in the
dependency analysis start at 1.”). Rather than fixing that
assumption, this commit adds an implementation of the force_first
argument to the new dependency sorting algorithm. This also means
that the directly dlopen'ed shared object is always initialized last,
which is the least surprising behavior in the presence of cycles.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Diffstat (limited to 'elf/dl-sort-maps.c')
-rw-r--r-- | elf/dl-sort-maps.c | 32 |
1 files changed, 23 insertions, 9 deletions
diff --git a/elf/dl-sort-maps.c b/elf/dl-sort-maps.c index 5b550b1..3e2a6a5 100644 --- a/elf/dl-sort-maps.c +++ b/elf/dl-sort-maps.c @@ -182,8 +182,9 @@ dfs_traversal (struct link_map ***rpo, struct link_map *map, static void _dl_sort_maps_dfs (struct link_map **maps, unsigned int nmaps, - bool force_first __attribute__ ((unused)), bool for_fini) + bool force_first, bool for_fini) { + struct link_map *first_map = maps[0]; for (int i = nmaps - 1; i >= 0; i--) maps[i]->l_visited = 0; @@ -208,14 +209,6 @@ _dl_sort_maps_dfs (struct link_map **maps, unsigned int nmaps, Adjusting the order so that maps[0] is last traversed naturally avoids this problem. - Further, the old "optimization" of skipping the main object at maps[0] - from the call-site (i.e. _dl_sort_maps(maps+1,nmaps-1)) is in general - no longer valid, since traversing along object dependency-links - may "find" the main object even when it is not included in the initial - order (e.g. a dlopen()'ed shared object can have circular dependencies - linked back to itself). In such a case, traversing N-1 objects will - create a N-object result, and raise problems. - To summarize, just passing in the full list, and iterating from back to front makes things much more straightforward. */ @@ -274,6 +267,27 @@ _dl_sort_maps_dfs (struct link_map **maps, unsigned int nmaps, } memcpy (maps, rpo, sizeof (struct link_map *) * nmaps); + + /* Skipping the first object at maps[0] is not valid in general, + since traversing along object dependency-links may "find" that + first object even when it is not included in the initial order + (e.g., a dlopen'ed shared object can have circular dependencies + linked back to itself). In such a case, traversing N-1 objects + will create a N-object result, and raise problems. Instead, + force the object back into first place after sorting. This naive + approach may introduce further dependency ordering violations + compared to rotating the cycle until the first map is again in + the first position, but as there is a cycle, at least one + violation is already present. */ + if (force_first && maps[0] != first_map) + { + int i; + for (i = 0; maps[i] != first_map; ++i) + ; + assert (i < nmaps); + memmove (&maps[1], maps, i * sizeof (maps[0])); + maps[0] = first_map; + } } void |