From: Pedro Alves <pedro@palves.net>
To: gdb-patches@sourceware.org
Subject: Re: [PATCH 01/25] Don't use pthread_mutex_t in gdb.base/step-over-clone.c
Date: Wed, 13 Jul 2022 22:35:09 +0100 [thread overview]
Message-ID: <55a0ce6e-1a6e-bdba-543d-9b3cd807e810@palves.net> (raw)
In-Reply-To: <20220620225419.382221-2-pedro@palves.net>
I'm dropping this patch from a v2 of this series, because I've since remembered
that current glibc is eliminating separte libpthread, merging it into the main library, and
gdb will load libthread_db.so even with non-threaded programs, so that printing errno
always works... This all means that with new enough glibc, I think we'd get back the
warnings...
Pedro Alves
On 2022-06-20 11:53 p.m., Pedro Alves wrote:
> I noticed this in gdb.log after running gdb.base/step-over-clone.exp:
>
> ...
> gdbserver: PID mismatch! Expected 1790818, got 1790817
> gdbserver: Cannot find thread after clone.
> gdbserver: PID mismatch! Expected 1790819, got 1790817
> gdbserver: Cannot find thread after clone.
> gdbserver: PID mismatch! Expected 1790820, got 1790817
> gdbserver: Cannot find thread after clone.
> ...
>
> Those "PID mismatch" come from gdbserver/thread_db.c. The problem is
> that the testcase program is testing raw clone, which bypasses
> libpthread entirely and leaves libthread_db confused. The testcase is
> linking with pthreads because it wants to use pthread_mutex_t for
> synchronization between the clones. Mixing pthreads and raw clone is
> just something we shouldn't do, however.
>
> My first thought was to fix this by using an atomic decrement
> (__atomic_fetch_sub) instead of a mutex, for synchronization.
> However, on some archs, that may require linking with -latomic, which
> can itself pull in libpthread.
>
> My next idea, is to make each thread write to its own "I'm ready"
> variable, such that we can't actually have read-modify-write races.
> This is what this patch does.
>
> Change-Id: Id418978ac86bfa6d51d0af1e1625a86cdd039a20
> ---
> gdb/testsuite/gdb.base/step-over-clone.c | 69 +++++++++++++-------
> gdb/testsuite/gdb.base/step-over-syscall.exp | 7 +-
> 2 files changed, 45 insertions(+), 31 deletions(-)
>
> diff --git a/gdb/testsuite/gdb.base/step-over-clone.c b/gdb/testsuite/gdb.base/step-over-clone.c
> index c0f67af188b..8a56b492e5e 100644
> --- a/gdb/testsuite/gdb.base/step-over-clone.c
> +++ b/gdb/testsuite/gdb.base/step-over-clone.c
> @@ -19,7 +19,7 @@
> #include <stdlib.h>
> #include <unistd.h>
> #include <sched.h>
> -#include <pthread.h>
> +#include <signal.h>
>
> static void
> marker ()
> @@ -27,30 +27,55 @@ marker ()
>
> #define STACK_SIZE 0x1000
>
> -/* These are used to signal that the threads have started correctly. The
> - GLOBAL_THREAD_COUNT is set to the number of threads in main, then
> - decremented (under a lock) in each new thread. */
> -pthread_mutex_t global_lock = PTHREAD_MUTEX_INITIALIZER;
> -int global_thread_count = 0;
> +#define NUM_THREADS 6
> +
> +/* This is used to signal that the threads have started correctly. We
> + can't use a single global updated by all thread guarded by a
> + pthread mutex, or anything pthread related for the matter, since we
> + are using raw clone. A single global updated with atomics
> + (__atomic_fetch* etc.) instead of a pthread mutex would sound
> + appealing, but we avoid that too because for some architectures,
> + we'd have to link with -latomic, which itself links with
> + -lpthread... So instead have one array with one element per
> + thread, and each thread only ever writes to its own array element.
> + We make the array have sig_atomic_t elements so that the elements
> + are portably naturally aligned and free from data races on all
> + archs, when different threads write to different elements. In
> + practice, "int" would work too, as accesses to int are pretty much
> + garanteed to be atomic on all Linux systems, but sig_atomic_t is
> + explicit. */
> +volatile sig_atomic_t thread_started[NUM_THREADS];
>
> static int
> -clone_fn (void *unused)
> +clone_fn (void *started)
> {
> /* Signal that this thread has started correctly. */
> - if (pthread_mutex_lock (&global_lock) != 0)
> - abort ();
> - global_thread_count--;
> - if (pthread_mutex_unlock (&global_lock) != 0)
> - abort ();
> + *(volatile sig_atomic_t *) started = 1;
>
> return 0;
> }
>
> +/* Return true if all threads have started. */
> +
> +static int
> +all_threads_started (void)
> +{
> + int i;
> +
> + /* Force full memory barrier so that caches are flushed and
> + THREAD_STARTED is refetched. */
> + __sync_synchronize ();
> + for (i = 0; i < NUM_THREADS; i++)
> + if (thread_started[i] == 0)
> + return 0;
> + return 1;
> +}
> +
> int
> main (void)
> {
> int i, pid;
> - unsigned char *stack[6];
> + unsigned char *stack[NUM_THREADS];
>
> /* Due to bug gdb/19675 the cloned thread _might_ try to reenter main
> (this depends on where the displaced instruction is placed for
> @@ -62,18 +87,16 @@ main (void)
> else
> abort ();
>
> - for (i = 0; i < (sizeof (stack) / sizeof (stack[0])); i++)
> + for (i = 0; i < NUM_THREADS; i++)
> stack[i] = malloc (STACK_SIZE);
>
> - global_thread_count = (sizeof (stack) / sizeof (stack[0]));
> -
> - for (i = 0; i < (sizeof (stack) / sizeof (stack[0])); i++)
> + for (i = 0; i < NUM_THREADS; i++)
> {
> pid = clone (clone_fn, stack[i] + STACK_SIZE, CLONE_FILES | CLONE_VM,
> - NULL);
> + (void *) &thread_started[i]);
> }
>
> - for (i = 0; i < (sizeof (stack) / sizeof (stack[0])); i++)
> + for (i = 0; i < NUM_THREADS; i++)
> free (stack[i]);
>
> /* Set an alarm so we don't end up stuck waiting for threads that might
> @@ -81,12 +104,8 @@ main (void)
> alarm (120);
>
> /* Now wait for all the threads to start up. */
> - while (global_thread_count != 0)
> - {
> - /* Force memory barrier so GLOBAL_THREAD_COUNT will be refetched. */
> - asm volatile ("" ::: "memory");
> - sleep (1);
> - }
> + while (!all_threads_started ())
> + sleep (1);
>
> /* Call marker, this is what GDB is waiting for. */
> marker ();
> diff --git a/gdb/testsuite/gdb.base/step-over-syscall.exp b/gdb/testsuite/gdb.base/step-over-syscall.exp
> index 788f6e3f5d0..e87d391cd5f 100644
> --- a/gdb/testsuite/gdb.base/step-over-syscall.exp
> +++ b/gdb/testsuite/gdb.base/step-over-syscall.exp
> @@ -241,12 +241,7 @@ proc step_over_syscall { syscall } {
>
> set testfile "step-over-$syscall"
>
> - set options [list debug]
> - if { $syscall == "clone" } {
> - lappend options "pthreads"
> - }
> -
> - if [build_executable ${testfile}.exp ${testfile} ${testfile}.c $options] {
> + if [build_executable ${testfile}.exp ${testfile} ${testfile}.c {debug}] {
> untested "failed to compile"
> return -1
> }
>
next prev parent reply other threads:[~2022-07-13 21:35 UTC|newest]
Thread overview: 47+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-06-20 22:53 [PATCH 00/25] Step over thread clone and thread exit Pedro Alves
2022-06-20 22:53 ` [PATCH 01/25] Don't use pthread_mutex_t in gdb.base/step-over-clone.c Pedro Alves
2022-07-13 21:35 ` Pedro Alves [this message]
2022-06-20 22:53 ` [PATCH 02/25] displaced step: pass down target_waitstatus instead of gdb_signal Pedro Alves
2022-06-20 22:53 ` [PATCH 03/25] linux-nat: introduce pending_status_str Pedro Alves
2022-06-20 22:53 ` [PATCH 04/25] Step over clone syscall w/ breakpoint, TARGET_WAITKIND_THREAD_CLONED Pedro Alves
2022-06-20 22:53 ` [PATCH 05/25] Support clone events in the remote protocol Pedro Alves
2022-06-20 22:54 ` [PATCH 06/25] Thread options & clone events (core + remote) Pedro Alves
2022-06-20 22:54 ` [PATCH 07/25] Thread options & clone events (native Linux) Pedro Alves
2022-06-20 22:54 ` [PATCH 08/25] Thread options & clone events (Linux GDBserver) Pedro Alves
2022-06-20 22:54 ` [PATCH 09/25] gdbserver: Hide and don't detach pending clone children Pedro Alves
2022-06-20 22:54 ` [PATCH 10/25] Remove gdb/19675 kfails (displaced stepping + clone) Pedro Alves
2022-06-20 22:54 ` [PATCH 11/25] Add test for stepping over clone syscall Pedro Alves
2022-06-20 22:54 ` [PATCH 12/25] all-stop/synchronous RSP support thread-exit events Pedro Alves
2022-06-20 22:54 ` [PATCH 13/25] Introduce GDB_TO_EXIT thread option, fix step-over-thread-exit Pedro Alves
2022-06-20 22:54 ` [PATCH 14/25] Implement GDB_TO_EXIT support for Linux GDBserver Pedro Alves
2022-06-20 22:54 ` [PATCH 15/25] Implement GDB_TO_EXIT support for native Linux Pedro Alves
2022-06-20 22:54 ` [PATCH 16/25] gdb: clear step over information on thread exit (PR gdb/27338) Pedro Alves
2022-06-20 22:54 ` [PATCH 17/25] stop_all_threads: (re-)enable async before waiting for stops Pedro Alves
2022-06-20 22:54 ` [PATCH 18/25] gdbserver: Queue no-resumed event after thread exit Pedro Alves
2022-06-20 22:54 ` [PATCH 19/25] Don't resume new threads if scheduler-locking is in effect Pedro Alves
2022-06-21 11:07 ` Eli Zaretskii
2022-07-11 14:20 ` Pedro Alves
2022-07-11 15:44 ` Eli Zaretskii
2022-07-11 16:09 ` Pedro Alves
2022-07-11 16:30 ` Eli Zaretskii
2022-07-11 16:38 ` Pedro Alves
2022-07-11 17:00 ` Eli Zaretskii
2022-07-11 17:48 ` Pedro Alves
2022-07-11 17:50 ` Eli Zaretskii
2022-07-11 18:18 ` Pedro Alves
2022-07-11 18:29 ` Eli Zaretskii
2022-07-11 19:39 ` Pedro Alves
2022-07-12 16:08 ` Eli Zaretskii
2022-07-12 17:14 ` Pedro Alves
2022-06-20 22:54 ` [PATCH 20/25] Tighten gdb.threads/no-unwaited-for-left.exp regexps Pedro Alves
2022-07-13 21:32 ` Pedro Alves
2022-06-20 22:54 ` [PATCH 21/25] Report thread exit event for leader if reporting thread exit events Pedro Alves
2022-06-20 22:54 ` [PATCH 22/25] Ignore failure to read PC when resuming Pedro Alves
2022-06-20 22:54 ` [PATCH 23/25] gdb/testsuite/lib/my-syscalls.S: Refactor new SYSCALL macro Pedro Alves
2022-06-20 22:54 ` [PATCH 24/25] Testcases for stepping over thread exit syscall (PR gdb/27338) Pedro Alves
2022-06-20 22:54 ` [PATCH 25/25] Document remote clone events, and QThreadOptions packet Pedro Alves
2022-06-21 12:07 ` Eli Zaretskii
2022-07-11 15:19 ` Pedro Alves
2022-07-11 16:09 ` Eli Zaretskii
2022-07-11 16:54 ` Pedro Alves
2022-07-11 17:02 ` Eli Zaretskii
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=55a0ce6e-1a6e-bdba-543d-9b3cd807e810@palves.net \
--to=pedro@palves.net \
--cc=gdb-patches@sourceware.org \
/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).