From: Mark Wielaard <mark@klomp.org>
To: elfutils-devel@sourceware.org
Cc: hsm2@rice.edu
Subject: Re: [PATCH 12/16] libdw: Make libdw_find_split_unit thread-safe
Date: Wed, 11 Oct 2023 19:17:16 +0200 [thread overview]
Message-ID: <20231011171716.GO728@gnu.wildebeest.org> (raw)
In-Reply-To: <20231010134300.53830-12-mark@klomp.org>
Hi Heather,
On Tue, Oct 10, 2023 at 03:42:56PM +0200, Mark Wielaard wrote:
> From: Heather McIntyre <hsm2@rice.edu>
>
> * (try_split_file): Use eu_tsearch.
> Add lock for cu->split.
>
> Signed-off-by: Heather S. McIntyre <hsm2@rice.edu>
> Signed-off-by: Mark Wielaard <mark@klomp.org>
> ---
> libdw/libdw_find_split_unit.c | 43 +++++++++++++++++++++++++++++------
> 1 file changed, 36 insertions(+), 7 deletions(-)
>
> diff --git a/libdw/libdw_find_split_unit.c b/libdw/libdw_find_split_unit.c
> index 533f8072..a288e9bd 100644
> --- a/libdw/libdw_find_split_unit.c
> +++ b/libdw/libdw_find_split_unit.c
> @@ -34,13 +34,19 @@
> #include "libelfP.h"
>
> #include <limits.h>
> -#include <search.h>
> +#include <eu-search.h>
> #include <stdlib.h>
> #include <string.h>
> #include <sys/types.h>
> #include <sys/stat.h>
> #include <fcntl.h>
>
> +/* __libdw_link_skel_split() modifies "skel->split = split"
> + and "split->split = skel".
> + "cu->split" is read at multiple locations and conditionally updated.
> + Mutual exclusion is enforced to prevent a race. */
> +rwlock_define(static, cu_split_lock);
__libdw_link_skel_split is defined in libdw/libdwP.h and is called in
try_split_file. (It is also called in src/readelf.c but that is a
terrible hack, so ignore that for now.)
A Dwarf_CU is created (see libdw/libdw_findcu.c) with split set to
(Dwarf_CU *) -1 to indicate the split dwarf is not yet set. If
cu->split is NULL it means the split dwarf was searched for, but not
found.
> static void
> try_split_file (Dwarf_CU *cu, const char *dwo_path)
> {
> @@ -57,7 +63,7 @@ try_split_file (Dwarf_CU *cu, const char *dwo_path)
> if (split->unit_type == DW_UT_split_compile
> && cu->unit_id8 == split->unit_id8)
> {
> - if (tsearch (split->dbg, &cu->dbg->split_tree,
> + if (eu_tsearch (split->dbg, &cu->dbg->split_tree,
> __libdw_finddbg_cb) == NULL)
> {
> /* Something went wrong. Don't link. */
> @@ -65,9 +71,13 @@ try_split_file (Dwarf_CU *cu, const char *dwo_path)
> break;
> }
>
> + rwlock_wrlock(cu_split_lock);
> +
> /* Link skeleton and split compile units. */
> __libdw_link_skel_split (cu, split);
>
> + rwlock_unlock(cu_split_lock);
> +
Is this locking wide enough? It seems like multiple threads can race
to this code and set different Dwarfs (which means they'll leak). See
below.
> /* We have everything we need from this ELF
> file. And we are going to close the fd to
> not run out of file descriptors. */
> @@ -75,8 +85,13 @@ try_split_file (Dwarf_CU *cu, const char *dwo_path)
> break;
> }
> }
> +
> + rwlock_rdlock(cu_split_lock);
> +
> if (cu->split == (Dwarf_CU *) -1)
> dwarf_end (split_dwarf);
> +
> + rwlock_unlock(cu_split_lock);
This means __libdw_link_skel_split wasn't called above (so the
split_dwarf created by dwarf_begin didn't contain the expected CU).
> }
> /* Always close, because we don't want to run out of file
> descriptors. See also the elf_fcntl ELF_C_FDDONE call
> @@ -89,9 +104,13 @@ Dwarf_CU *
> internal_function
> __libdw_find_split_unit (Dwarf_CU *cu)
> {
> + rwlock_rdlock(cu_split_lock);
> + Dwarf_CU *cu_split_local = cu->split;
> + rwlock_unlock(cu_split_lock);
> +
> /* Only try once. */
> - if (cu->split != (Dwarf_CU *) -1)
> - return cu->split;
> + if (cu_split_local != (Dwarf_CU *) -1)
> + return cu_split_local;
>
> /* We need a skeleton unit with a comp_dir and [GNU_]dwo_name attributes.
> The split unit will be the first in the dwo file and should have the
> @@ -116,7 +135,11 @@ __libdw_find_split_unit (Dwarf_CU *cu)
> free (dwo_path);
> }
>
> - if (cu->split == (Dwarf_CU *) -1)
> + rwlock_rdlock(cu_split_lock);
> + cu_split_local = cu->split;
> + rwlock_unlock(cu_split_lock);
> +
> + if (cu_split_local == (Dwarf_CU *) -1)
> {
> /* Try compdir plus dwo_name. */
> Dwarf_Attribute compdir;
It looks like multiple threads can end up here after having seen
cu->split/cu_split_local == (Dwarf_CU *) -1) Which means multiple
threads will enter try_split_file and might all call
__libdw_link_skel_split as mentioned above.
> @@ -138,9 +161,15 @@ __libdw_find_split_unit (Dwarf_CU *cu)
> }
> }
>
> + rwlock_wrlock(cu_split_lock);
> +
> /* If we found nothing, make sure we don't try again. */
> if (cu->split == (Dwarf_CU *) -1)
> - cu->split = NULL;
> + {
> + cu->split = NULL;
> + cu_split_local = cu->split;
> + }
>
> - return cu->split;
> + rwlock_unlock(cu_split_lock);
> + return cu_split_local;
> }
I think it would be better to keep the lock during the whole
__libdw_find_split_unit call.
Cheers,
Mark
next prev parent reply other threads:[~2023-10-11 17:17 UTC|newest]
Thread overview: 51+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-08-08 17:07 [PATCH] Fix thread-safety for elfutils Heather McIntyre
2023-08-21 22:08 ` John Mellor-Crummey
2023-08-25 14:10 ` Mark Wielaard
2023-10-10 13:40 ` Mark Wielaard
2023-10-10 13:42 ` [PATCH 01/16] lib: Add new once_define and once macros to eu-config.h Mark Wielaard
2023-10-10 13:42 ` [PATCH 02/16] libelf: Make elf_version thread-safe Mark Wielaard
2023-10-10 14:00 ` Mark Wielaard
2023-10-17 19:05 ` Heather McIntyre
2023-10-19 21:00 ` Mark Wielaard
2023-10-10 13:42 ` [PATCH 03/16] libelf: Fix deadlock in __libelf_readall Mark Wielaard
2023-10-10 15:06 ` Mark Wielaard
2023-10-17 19:11 ` Heather McIntyre
2023-11-09 13:26 ` Mark Wielaard
2023-10-10 13:42 ` [PATCH 04/16] libelf: Fix deadlock in elf_cntl Mark Wielaard
2023-10-10 15:23 ` Mark Wielaard
2023-10-17 19:14 ` Heather McIntyre
2023-10-10 13:42 ` [PATCH 05/16] libelf: Fix elf_end deadlock Mark Wielaard
2023-10-10 15:28 ` Mark Wielaard
2023-10-10 13:42 ` [PATCH 06/16] libelf: Make elf32_getchdr and elf64_getchdr thread-safe Mark Wielaard
2023-10-10 16:28 ` Mark Wielaard
2023-10-10 13:42 ` [PATCH 07/16] lib: Add eu_tsearch and eu_tfind Mark Wielaard
2023-10-10 16:51 ` Mark Wielaard
2023-10-17 20:52 ` Heather McIntyre
2023-10-10 13:42 ` [PATCH 08/16] libcpu: Change calls for tsearch/tfind to eu_tsearch/eu_tfind Mark Wielaard
2023-10-10 21:10 ` Mark Wielaard
2023-10-10 13:42 ` [PATCH 09/16] src: Use eu-search in nm and findtextrel Mark Wielaard
2023-10-10 21:25 ` Mark Wielaard
2023-10-17 19:20 ` Heather McIntyre
2023-10-10 13:42 ` [PATCH 10/16] libdw: make dwarf_getalt thread-safe Mark Wielaard
2023-10-10 22:02 ` Mark Wielaard
2023-10-17 19:25 ` Heather McIntyre
2023-10-10 13:42 ` [PATCH 11/16] libdw: Add locking around __libdw_dieabbrev for dwarf_hasattr Mark Wielaard
2023-10-11 15:10 ` Mark Wielaard
2023-10-17 19:57 ` Heather McIntyre
2023-10-19 22:06 ` Mark Wielaard
2023-10-10 13:42 ` [PATCH 12/16] libdw: Make libdw_find_split_unit thread-safe Mark Wielaard
2023-10-11 17:17 ` Mark Wielaard [this message]
2023-10-17 20:01 ` Heather McIntyre
2023-10-10 13:42 ` [PATCH 13/16] libdw: Make libdw_findcu thread-safe Mark Wielaard
2023-10-12 22:02 ` Mark Wielaard
2023-10-17 20:10 ` Heather McIntyre
2023-10-10 13:42 ` [PATCH 14/16] libdw,libdwfl: Use eu-search for thread-safety Mark Wielaard
2023-10-12 22:05 ` Mark Wielaard
2023-10-10 13:42 ` [PATCH 15/16] tests: Add eu-search tests Mark Wielaard
2023-10-13 14:38 ` Mark Wielaard
2023-10-10 13:43 ` [PATCH 16/16] configure: No longer mark --enable-thread-safety as EXPERIMENTAL Mark Wielaard
2023-10-12 22:09 ` Mark Wielaard
2023-10-10 13:54 ` [PATCH 01/16] lib: Add new once_define and once macros to eu-config.h Mark Wielaard
2023-10-14 15:39 ` [PATCH] Fix thread-safety for elfutils Mark Wielaard
2023-10-14 18:29 ` Heather McIntyre
2023-10-17 15:04 ` Mark Wielaard
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20231011171716.GO728@gnu.wildebeest.org \
--to=mark@klomp.org \
--cc=elfutils-devel@sourceware.org \
--cc=hsm2@rice.edu \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).