glibc/elf/dl-sort-maps.c

305 lines
10 KiB
C
Raw Normal View History

/* Sort array of link maps according to dependencies.
Copyright (C) 2017-2022 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
The GNU C Library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with the GNU C Library; if not, see
Prefer https to http for gnu.org and fsf.org URLs Also, change sources.redhat.com to sourceware.org. This patch was automatically generated by running the following shell script, which uses GNU sed, and which avoids modifying files imported from upstream: sed -ri ' s,(http|ftp)(://(.*\.)?(gnu|fsf|sourceware)\.org($|[^.]|\.[^a-z])),https\2,g s,(http|ftp)(://(.*\.)?)sources\.redhat\.com($|[^.]|\.[^a-z]),https\2sourceware.org\4,g ' \ $(find $(git ls-files) -prune -type f \ ! -name '*.po' \ ! -name 'ChangeLog*' \ ! -path COPYING ! -path COPYING.LIB \ ! -path manual/fdl-1.3.texi ! -path manual/lgpl-2.1.texi \ ! -path manual/texinfo.tex ! -path scripts/config.guess \ ! -path scripts/config.sub ! -path scripts/install-sh \ ! -path scripts/mkinstalldirs ! -path scripts/move-if-change \ ! -path INSTALL ! -path locale/programs/charmap-kw.h \ ! -path po/libc.pot ! -path sysdeps/gnu/errlist.c \ ! '(' -name configure \ -execdir test -f configure.ac -o -f configure.in ';' ')' \ ! '(' -name preconfigure \ -execdir test -f preconfigure.ac ';' ')' \ -print) and then by running 'make dist-prepare' to regenerate files built from the altered files, and then executing the following to cleanup: chmod a+x sysdeps/unix/sysv/linux/riscv/configure # Omit irrelevant whitespace and comment-only changes, # perhaps from a slightly-different Autoconf version. git checkout -f \ sysdeps/csky/configure \ sysdeps/hppa/configure \ sysdeps/riscv/configure \ sysdeps/unix/sysv/linux/csky/configure # Omit changes that caused a pre-commit check to fail like this: # remote: *** error: sysdeps/powerpc/powerpc64/ppc-mcount.S: trailing lines git checkout -f \ sysdeps/powerpc/powerpc64/ppc-mcount.S \ sysdeps/unix/sysv/linux/s390/s390-64/syscall.S # Omit change that caused a pre-commit check to fail like this: # remote: *** error: sysdeps/sparc/sparc64/multiarch/memcpy-ultra3.S: last line does not end in newline git checkout -f sysdeps/sparc/sparc64/multiarch/memcpy-ultra3.S
2019-09-07 05:40:42 +00:00
<https://www.gnu.org/licenses/>. */
elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645) This second patch contains the actual implementation of a new sorting algorithm for shared objects in the dynamic loader, which solves the slow behavior that the current "old" algorithm falls into when the DSO set contains circular dependencies. The new algorithm implemented here is simply depth-first search (DFS) to obtain the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1 bitfield is added to struct link_map to more elegantly facilitate such a search. The DFS algorithm is applied to the input maps[nmap-1] backwards towards maps[0]. This has the effect of a more "shallow" recursion depth in general since the input is in BFS. Also, when combined with the natural order of processing l_initfini[] at each node, this creates a resulting output sorting closer to the intuitive "left-to-right" order in most cases. Another notable implementation adjustment related to this _dl_sort_maps change is the removing of two char arrays 'used' and 'done' in _dl_close_worker to represent two per-map attributes. This has been changed to simply use two new bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes do along the way. Tunable support for switching between different sorting algorithms at runtime is also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1 (old algorithm) and 2 (new DFS algorithm) has been added. At time of commit of this patch, the default setting is 1 (old algorithm). Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com> Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
2021-10-21 13:41:22 +00:00
#include <assert.h>
#include <ldsodefs.h>
elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645) This second patch contains the actual implementation of a new sorting algorithm for shared objects in the dynamic loader, which solves the slow behavior that the current "old" algorithm falls into when the DSO set contains circular dependencies. The new algorithm implemented here is simply depth-first search (DFS) to obtain the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1 bitfield is added to struct link_map to more elegantly facilitate such a search. The DFS algorithm is applied to the input maps[nmap-1] backwards towards maps[0]. This has the effect of a more "shallow" recursion depth in general since the input is in BFS. Also, when combined with the natural order of processing l_initfini[] at each node, this creates a resulting output sorting closer to the intuitive "left-to-right" order in most cases. Another notable implementation adjustment related to this _dl_sort_maps change is the removing of two char arrays 'used' and 'done' in _dl_close_worker to represent two per-map attributes. This has been changed to simply use two new bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes do along the way. Tunable support for switching between different sorting algorithms at runtime is also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1 (old algorithm) and 2 (new DFS algorithm) has been added. At time of commit of this patch, the default setting is 1 (old algorithm). Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com> Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
2021-10-21 13:41:22 +00:00
#include <elf/dl-tunables.h>
elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645) This second patch contains the actual implementation of a new sorting algorithm for shared objects in the dynamic loader, which solves the slow behavior that the current "old" algorithm falls into when the DSO set contains circular dependencies. The new algorithm implemented here is simply depth-first search (DFS) to obtain the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1 bitfield is added to struct link_map to more elegantly facilitate such a search. The DFS algorithm is applied to the input maps[nmap-1] backwards towards maps[0]. This has the effect of a more "shallow" recursion depth in general since the input is in BFS. Also, when combined with the natural order of processing l_initfini[] at each node, this creates a resulting output sorting closer to the intuitive "left-to-right" order in most cases. Another notable implementation adjustment related to this _dl_sort_maps change is the removing of two char arrays 'used' and 'done' in _dl_close_worker to represent two per-map attributes. This has been changed to simply use two new bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes do along the way. Tunable support for switching between different sorting algorithms at runtime is also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1 (old algorithm) and 2 (new DFS algorithm) has been added. At time of commit of this patch, the default setting is 1 (old algorithm). Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com> Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
2021-10-21 13:41:22 +00:00
/* Note: this is the older, "original" sorting algorithm, being used as
default up to 2.35.
elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645) This second patch contains the actual implementation of a new sorting algorithm for shared objects in the dynamic loader, which solves the slow behavior that the current "old" algorithm falls into when the DSO set contains circular dependencies. The new algorithm implemented here is simply depth-first search (DFS) to obtain the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1 bitfield is added to struct link_map to more elegantly facilitate such a search. The DFS algorithm is applied to the input maps[nmap-1] backwards towards maps[0]. This has the effect of a more "shallow" recursion depth in general since the input is in BFS. Also, when combined with the natural order of processing l_initfini[] at each node, this creates a resulting output sorting closer to the intuitive "left-to-right" order in most cases. Another notable implementation adjustment related to this _dl_sort_maps change is the removing of two char arrays 'used' and 'done' in _dl_close_worker to represent two per-map attributes. This has been changed to simply use two new bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes do along the way. Tunable support for switching between different sorting algorithms at runtime is also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1 (old algorithm) and 2 (new DFS algorithm) has been added. At time of commit of this patch, the default setting is 1 (old algorithm). Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com> Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
2021-10-21 13:41:22 +00:00
Sort array MAPS according to dependencies of the contained objects.
If FOR_FINI is true, this is called for finishing an object. */
static void
_dl_sort_maps_original (struct link_map **maps, unsigned int nmaps,
bool force_first, bool for_fini)
{
elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645) This second patch contains the actual implementation of a new sorting algorithm for shared objects in the dynamic loader, which solves the slow behavior that the current "old" algorithm falls into when the DSO set contains circular dependencies. The new algorithm implemented here is simply depth-first search (DFS) to obtain the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1 bitfield is added to struct link_map to more elegantly facilitate such a search. The DFS algorithm is applied to the input maps[nmap-1] backwards towards maps[0]. This has the effect of a more "shallow" recursion depth in general since the input is in BFS. Also, when combined with the natural order of processing l_initfini[] at each node, this creates a resulting output sorting closer to the intuitive "left-to-right" order in most cases. Another notable implementation adjustment related to this _dl_sort_maps change is the removing of two char arrays 'used' and 'done' in _dl_close_worker to represent two per-map attributes. This has been changed to simply use two new bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes do along the way. Tunable support for switching between different sorting algorithms at runtime is also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1 (old algorithm) and 2 (new DFS algorithm) has been added. At time of commit of this patch, the default setting is 1 (old algorithm). Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com> Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
2021-10-21 13:41:22 +00:00
/* Allows caller to do the common optimization of skipping the first map,
usually the main binary. */
maps += force_first;
nmaps -= force_first;
elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645) This second patch contains the actual implementation of a new sorting algorithm for shared objects in the dynamic loader, which solves the slow behavior that the current "old" algorithm falls into when the DSO set contains circular dependencies. The new algorithm implemented here is simply depth-first search (DFS) to obtain the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1 bitfield is added to struct link_map to more elegantly facilitate such a search. The DFS algorithm is applied to the input maps[nmap-1] backwards towards maps[0]. This has the effect of a more "shallow" recursion depth in general since the input is in BFS. Also, when combined with the natural order of processing l_initfini[] at each node, this creates a resulting output sorting closer to the intuitive "left-to-right" order in most cases. Another notable implementation adjustment related to this _dl_sort_maps change is the removing of two char arrays 'used' and 'done' in _dl_close_worker to represent two per-map attributes. This has been changed to simply use two new bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes do along the way. Tunable support for switching between different sorting algorithms at runtime is also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1 (old algorithm) and 2 (new DFS algorithm) has been added. At time of commit of this patch, the default setting is 1 (old algorithm). Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com> Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
2021-10-21 13:41:22 +00:00
/* A list of one element need not be sorted. */
if (nmaps <= 1)
return;
unsigned int i = 0;
uint16_t seen[nmaps];
memset (seen, 0, nmaps * sizeof (seen[0]));
while (1)
{
/* Keep track of which object we looked at this round. */
++seen[i];
struct link_map *thisp = maps[i];
if (__glibc_unlikely (for_fini))
{
/* Do not handle ld.so in secondary namespaces and objects which
are not removed. */
if (thisp != thisp->l_real || thisp->l_idx == -1)
goto skip;
}
/* Find the last object in the list for which the current one is
a dependency and move the current object behind the object
with the dependency. */
unsigned int k = nmaps - 1;
while (k > i)
{
struct link_map **runp = maps[k]->l_initfini;
if (runp != NULL)
/* Look through the dependencies of the object. */
while (*runp != NULL)
if (__glibc_unlikely (*runp++ == thisp))
{
move:
/* Move the current object to the back past the last
object with it as the dependency. */
memmove (&maps[i], &maps[i + 1],
(k - i) * sizeof (maps[0]));
maps[k] = thisp;
if (seen[i + 1] > nmaps - i)
{
++i;
goto next_clear;
}
uint16_t this_seen = seen[i];
memmove (&seen[i], &seen[i + 1], (k - i) * sizeof (seen[0]));
seen[k] = this_seen;
goto next;
}
if (__glibc_unlikely (for_fini && maps[k]->l_reldeps != NULL))
{
unsigned int m = maps[k]->l_reldeps->act;
struct link_map **relmaps = &maps[k]->l_reldeps->list[0];
/* Look through the relocation dependencies of the object. */
while (m-- > 0)
if (__glibc_unlikely (relmaps[m] == thisp))
{
/* If a cycle exists with a link time dependency,
preserve the latter. */
struct link_map **runp = thisp->l_initfini;
if (runp != NULL)
while (*runp != NULL)
if (__glibc_unlikely (*runp++ == maps[k]))
goto ignore;
goto move;
}
ignore:;
}
--k;
}
skip:
if (++i == nmaps)
break;
next_clear:
memset (&seen[i], 0, (nmaps - i) * sizeof (seen[0]));
next:;
}
}
elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645) This second patch contains the actual implementation of a new sorting algorithm for shared objects in the dynamic loader, which solves the slow behavior that the current "old" algorithm falls into when the DSO set contains circular dependencies. The new algorithm implemented here is simply depth-first search (DFS) to obtain the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1 bitfield is added to struct link_map to more elegantly facilitate such a search. The DFS algorithm is applied to the input maps[nmap-1] backwards towards maps[0]. This has the effect of a more "shallow" recursion depth in general since the input is in BFS. Also, when combined with the natural order of processing l_initfini[] at each node, this creates a resulting output sorting closer to the intuitive "left-to-right" order in most cases. Another notable implementation adjustment related to this _dl_sort_maps change is the removing of two char arrays 'used' and 'done' in _dl_close_worker to represent two per-map attributes. This has been changed to simply use two new bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes do along the way. Tunable support for switching between different sorting algorithms at runtime is also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1 (old algorithm) and 2 (new DFS algorithm) has been added. At time of commit of this patch, the default setting is 1 (old algorithm). Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com> Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
2021-10-21 13:41:22 +00:00
#if !HAVE_TUNABLES
/* In this case, just default to the original algorithm. */
strong_alias (_dl_sort_maps_original, _dl_sort_maps);
#else
/* We use a recursive function due to its better clarity and ease of
implementation, as well as faster execution speed. We already use
alloca() for list allocation during the breadth-first search of
dependencies in _dl_map_object_deps(), and this should be on the
same order of worst-case stack usage.
Note: the '*rpo' parameter is supposed to point to one past the
last element of the array where we save the sort results, and is
decremented before storing the current map at each level. */
static void
dfs_traversal (struct link_map ***rpo, struct link_map *map,
bool *do_reldeps)
{
elf: Fix DFS sorting algorithm for LD_TRACE_LOADED_OBJECTS with missing libraries (BZ #28868) On _dl_map_object the underlying file is not opened in trace mode (in other cases where the underlying file can't be opened, _dl_map_object quits with an error). If there any missing libraries being processed, they will not be considered on final nlist size passed on _dl_sort_maps later in the function. And it is then used by _dl_sort_maps_dfs on the stack allocated working maps: 222 /* Array to hold RPO sorting results, before we copy back to maps[]. */ 223 struct link_map *rpo[nmaps]; 224 225 /* The 'head' position during each DFS iteration. Note that we start at 226 one past the last element due to first-decrement-then-store (see the 227 bottom of above dfs_traversal() routine). */ 228 struct link_map **rpo_head = &rpo[nmaps]; However while transversing the 'l_initfini' on dfs_traversal it will still consider the l_faked maps and thus update rpo more times than the allocated working 'rpo', overflowing the stack object. As suggested in bugzilla, one option would be to avoid sorting the maps for trace mode. However I think ignoring l_faked object does make sense (there is one less constraint to call the sorting function), it allows a slight less stack usage for trace, and it is slight simpler solution. The tests does trigger the stack overflow, however I tried to make it more generic to check different scenarios or missing objects. Checked on x86_64-linux-gnu. Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
2022-02-08 18:22:49 +00:00
/* _dl_map_object_deps ignores l_faked objects when calculating the
number of maps before calling _dl_sort_maps, ignore them as well. */
if (map->l_visited || map->l_faked)
elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645) This second patch contains the actual implementation of a new sorting algorithm for shared objects in the dynamic loader, which solves the slow behavior that the current "old" algorithm falls into when the DSO set contains circular dependencies. The new algorithm implemented here is simply depth-first search (DFS) to obtain the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1 bitfield is added to struct link_map to more elegantly facilitate such a search. The DFS algorithm is applied to the input maps[nmap-1] backwards towards maps[0]. This has the effect of a more "shallow" recursion depth in general since the input is in BFS. Also, when combined with the natural order of processing l_initfini[] at each node, this creates a resulting output sorting closer to the intuitive "left-to-right" order in most cases. Another notable implementation adjustment related to this _dl_sort_maps change is the removing of two char arrays 'used' and 'done' in _dl_close_worker to represent two per-map attributes. This has been changed to simply use two new bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes do along the way. Tunable support for switching between different sorting algorithms at runtime is also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1 (old algorithm) and 2 (new DFS algorithm) has been added. At time of commit of this patch, the default setting is 1 (old algorithm). Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com> Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
2021-10-21 13:41:22 +00:00
return;
map->l_visited = 1;
if (map->l_initfini)
{
for (int i = 0; map->l_initfini[i] != NULL; i++)
{
struct link_map *dep = map->l_initfini[i];
if (dep->l_visited == 0
&& dep->l_main_map == 0)
dfs_traversal (rpo, dep, do_reldeps);
}
}
if (__glibc_unlikely (do_reldeps != NULL && map->l_reldeps != NULL))
{
/* Indicate that we encountered relocation dependencies during
traversal. */
*do_reldeps = true;
for (int m = map->l_reldeps->act - 1; m >= 0; m--)
{
struct link_map *dep = map->l_reldeps->list[m];
if (dep->l_visited == 0
&& dep->l_main_map == 0)
dfs_traversal (rpo, dep, do_reldeps);
}
}
*rpo -= 1;
**rpo = map;
}
/* Topologically sort array MAPS according to dependencies of the contained
objects. */
static void
_dl_sort_maps_dfs (struct link_map **maps, unsigned int nmaps,
bool force_first __attribute__ ((unused)), bool for_fini)
elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645) This second patch contains the actual implementation of a new sorting algorithm for shared objects in the dynamic loader, which solves the slow behavior that the current "old" algorithm falls into when the DSO set contains circular dependencies. The new algorithm implemented here is simply depth-first search (DFS) to obtain the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1 bitfield is added to struct link_map to more elegantly facilitate such a search. The DFS algorithm is applied to the input maps[nmap-1] backwards towards maps[0]. This has the effect of a more "shallow" recursion depth in general since the input is in BFS. Also, when combined with the natural order of processing l_initfini[] at each node, this creates a resulting output sorting closer to the intuitive "left-to-right" order in most cases. Another notable implementation adjustment related to this _dl_sort_maps change is the removing of two char arrays 'used' and 'done' in _dl_close_worker to represent two per-map attributes. This has been changed to simply use two new bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes do along the way. Tunable support for switching between different sorting algorithms at runtime is also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1 (old algorithm) and 2 (new DFS algorithm) has been added. At time of commit of this patch, the default setting is 1 (old algorithm). Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com> Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
2021-10-21 13:41:22 +00:00
{
for (int i = nmaps - 1; i >= 0; i--)
maps[i]->l_visited = 0;
/* We apply DFS traversal for each of maps[i] until the whole total order
is found and we're at the start of the Reverse-Postorder (RPO) sequence,
which is a topological sort.
We go from maps[nmaps - 1] backwards towards maps[0] at this level.
Due to the breadth-first search (BFS) ordering we receive, going
backwards usually gives a more shallow depth-first recursion depth,
adding more stack usage safety. Also, combined with the natural
processing order of l_initfini[] at each node during DFS, this maintains
an ordering closer to the original link ordering in the sorting results
under most simpler cases.
Another reason we order the top level backwards, it that maps[0] is
usually exactly the main object of which we're in the midst of
_dl_map_object_deps() processing, and maps[0]->l_initfini[] is still
blank. If we start the traversal from maps[0], since having no
dependencies yet filled in, maps[0] will always be immediately
incorrectly placed at the last place in the order (first in reverse).
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. */
/* Array to hold RPO sorting results, before we copy back to maps[]. */
struct link_map *rpo[nmaps];
/* The 'head' position during each DFS iteration. Note that we start at
one past the last element due to first-decrement-then-store (see the
bottom of above dfs_traversal() routine). */
struct link_map **rpo_head = &rpo[nmaps];
bool do_reldeps = false;
bool *do_reldeps_ref = (for_fini ? &do_reldeps : NULL);
for (int i = nmaps - 1; i >= 0; i--)
{
dfs_traversal (&rpo_head, maps[i], do_reldeps_ref);
/* We can break early if all objects are already placed. */
if (rpo_head == rpo)
goto end;
}
assert (rpo_head == rpo);
end:
/* Here we may do a second pass of sorting, using only l_initfini[]
static dependency links. This is avoided if !FOR_FINI or if we didn't
find any reldeps in the first DFS traversal.
The reason we do this is: while it is unspecified how circular
dependencies should be handled, the presumed reasonable behavior is to
have destructors to respect static dependency links as much as possible,
overriding reldeps if needed. And the first sorting pass, which takes
l_initfini/l_reldeps links equally, may not preserve this priority.
Hence we do a 2nd sorting pass, taking only DT_NEEDED links into account
(see how the do_reldeps argument to dfs_traversal() is NULL below). */
if (do_reldeps)
{
for (int i = nmaps - 1; i >= 0; i--)
rpo[i]->l_visited = 0;
struct link_map **maps_head = &maps[nmaps];
for (int i = nmaps - 1; i >= 0; i--)
{
dfs_traversal (&maps_head, rpo[i], NULL);
/* We can break early if all objects are already placed.
The below memcpy is not needed in the do_reldeps case here,
since we wrote back to maps[] during DFS traversal. */
if (maps_head == maps)
return;
}
assert (maps_head == maps);
return;
}
memcpy (maps, rpo, sizeof (struct link_map *) * nmaps);
}
void
_dl_sort_maps_init (void)
{
int32_t algorithm = TUNABLE_GET (glibc, rtld, dynamic_sort, int32_t, NULL);
GLRO(dl_dso_sort_algo) = algorithm == 1 ? dso_sort_algorithm_original
: dso_sort_algorithm_dfs;
}
void
_dl_sort_maps (struct link_map **maps, unsigned int nmaps,
bool force_first, bool for_fini)
elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645) This second patch contains the actual implementation of a new sorting algorithm for shared objects in the dynamic loader, which solves the slow behavior that the current "old" algorithm falls into when the DSO set contains circular dependencies. The new algorithm implemented here is simply depth-first search (DFS) to obtain the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1 bitfield is added to struct link_map to more elegantly facilitate such a search. The DFS algorithm is applied to the input maps[nmap-1] backwards towards maps[0]. This has the effect of a more "shallow" recursion depth in general since the input is in BFS. Also, when combined with the natural order of processing l_initfini[] at each node, this creates a resulting output sorting closer to the intuitive "left-to-right" order in most cases. Another notable implementation adjustment related to this _dl_sort_maps change is the removing of two char arrays 'used' and 'done' in _dl_close_worker to represent two per-map attributes. This has been changed to simply use two new bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes do along the way. Tunable support for switching between different sorting algorithms at runtime is also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1 (old algorithm) and 2 (new DFS algorithm) has been added. At time of commit of this patch, the default setting is 1 (old algorithm). Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com> Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
2021-10-21 13:41:22 +00:00
{
/* It can be tempting to use a static function pointer to store and call
the current selected sorting algorithm routine, but experimentation
shows that current processors still do not handle indirect branches
that efficiently, plus a static function pointer will involve
PTR_MANGLE/DEMANGLE, further impairing performance of small, common
input cases. A simple if-case with direct function calls appears to
be the fastest. */
if (__glibc_likely (GLRO(dl_dso_sort_algo) == dso_sort_algorithm_original))
_dl_sort_maps_original (maps, nmaps, force_first, for_fini);
elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645) This second patch contains the actual implementation of a new sorting algorithm for shared objects in the dynamic loader, which solves the slow behavior that the current "old" algorithm falls into when the DSO set contains circular dependencies. The new algorithm implemented here is simply depth-first search (DFS) to obtain the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1 bitfield is added to struct link_map to more elegantly facilitate such a search. The DFS algorithm is applied to the input maps[nmap-1] backwards towards maps[0]. This has the effect of a more "shallow" recursion depth in general since the input is in BFS. Also, when combined with the natural order of processing l_initfini[] at each node, this creates a resulting output sorting closer to the intuitive "left-to-right" order in most cases. Another notable implementation adjustment related to this _dl_sort_maps change is the removing of two char arrays 'used' and 'done' in _dl_close_worker to represent two per-map attributes. This has been changed to simply use two new bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes do along the way. Tunable support for switching between different sorting algorithms at runtime is also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1 (old algorithm) and 2 (new DFS algorithm) has been added. At time of commit of this patch, the default setting is 1 (old algorithm). Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com> Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
2021-10-21 13:41:22 +00:00
else
_dl_sort_maps_dfs (maps, nmaps, force_first, for_fini);
elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645) This second patch contains the actual implementation of a new sorting algorithm for shared objects in the dynamic loader, which solves the slow behavior that the current "old" algorithm falls into when the DSO set contains circular dependencies. The new algorithm implemented here is simply depth-first search (DFS) to obtain the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1 bitfield is added to struct link_map to more elegantly facilitate such a search. The DFS algorithm is applied to the input maps[nmap-1] backwards towards maps[0]. This has the effect of a more "shallow" recursion depth in general since the input is in BFS. Also, when combined with the natural order of processing l_initfini[] at each node, this creates a resulting output sorting closer to the intuitive "left-to-right" order in most cases. Another notable implementation adjustment related to this _dl_sort_maps change is the removing of two char arrays 'used' and 'done' in _dl_close_worker to represent two per-map attributes. This has been changed to simply use two new bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes do along the way. Tunable support for switching between different sorting algorithms at runtime is also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1 (old algorithm) and 2 (new DFS algorithm) has been added. At time of commit of this patch, the default setting is 1 (old algorithm). Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com> Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
2021-10-21 13:41:22 +00:00
}
#endif /* HAVE_TUNABLES. */