A complete reworking of Valgrind's handling of system calls and signals,
with the aim of making it more robust, more correct and perhaps faster.
This patch removes the need to poll blocking syscalls, by adding a proxy
LWP for each application thread. This LWP is a kernel thread whose job
is to run all (potentially) blocking syscalls, and also to handle signals.
This allows the kernel to do more of the work of dealing with signals,
so on kernels which do this properly (2.6), Valgrind's behavious is a
lot more posix compliant. On base 2.4 kernels, we emulate some of the
missing 2.6 functionality.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@1918 a5019735-40e9-0310-863c-91ae7b9d1cf9
diff --git a/NOTES.syscalls b/NOTES.syscalls
new file mode 100644
index 0000000..86e3231
--- /dev/null
+++ b/NOTES.syscalls
@@ -0,0 +1,195 @@
+- works on stock 2.4 kernels, but the scheduler loop must poll
+- works on RH9 2.4.20-18.9 kernel, but doesn't seem quite as stable
+ as 2.5/2.6
+ for pending signals rather than relying on the kernel delivering them
+ to the right place.
+- most tested on 2.6.0-test1 and up
+
+- running job-control programs (ie, bash) under Valgrind won't work
+ properly without a kernel patch (as of 2.6.0-test2-mm2). This is because
+ threads in a thread group don't follow the thread group leader's changes
+ in process group ID, and they can't change it for themselves.
+
+- SA_NOCLDWAIT doesn't work properly if the program is actually blocked
+ in wait4() when SIGCHLD arrives; the wait4() will return details for
+ the exiting child. In other circumstances children should be quietly reaped.
+ [ This may be fixable when running under RH2.4 and 2.6, since we can
+ set NOCLDWAIT in the kernel's state without risk of losing our child
+ threads. ]
+
+- 2.4 has somewhat disfunctional thread/signal interactions, so many test
+ do not work as well under 2.4. In general, it should be no worse than
+ the old signal code. I don't intend spending a lot of time fixing this
+ because 2.6 is nearly ready for widespread use.
+
+TODO:
+
+- support application use of clone(). Interesting question is which
+ options do we support? Do we need to implement futex as well, or can
+ we just use the kernel's implementation?
+
+========================================
+Testing
+
+I've been testing with the Posix test suite:
+http://sourceforge.net/projects/posixtest/, version 1.2.0.
+
+----------------------------------------
+Expected failures:
+
+conformance/interfaces/sigwaitinfo/6-1.test
+ pthread_kill() calls the tkill syscall, which causes a code of
+ SI_TKILL rather than the SI_USER which this test expects.
+
+conformance/interfaces/sigrelse/3-*.test
+ glibc bug in sigrelse(), which fails without Valgrind too.
+
+conformance/interfaces/pthread_barrier_*/*
+ Valgrind's libpthreads doesn't implement pthread_barrier_*.
+ (There are some passes, but I don't know why.)
+
+conformance/interfaces/pthread_cond_timedwait/2-3
+ This test is just completely broken. It does expose a problem
+ in Valgrind's mutex implementation - it is too dependent on
+ the client code not doing stupid stuff. This test makes
+ Valgrind have an assertion failure.
+
+conformance/interfaces/pthread_condattr_getpshared/*
+ pthread_condattr_getpshared not implemented
+
+conformance/interfaces/pthread_condattr_setpshared/*
+ pthread_condattr_setpshared not implemented
+
+conformance/interfaces/pthread_key_create/speculative/5-1
+ Valgrind should cope with key overload
+
+conformance/interfaces/pthread_mutex_timedlock/*
+ not implemented
+
+conformance/interfaces/pthread_rwlock_rdlock/2-1:
+ relies on pthread_setschedparam
+
+conformance/interfaces/pthread_rwlock_timedrdlock/*
+ valgrind's libpthread.so: UNIMPLEMENTED FUNCTION: pthread_rwlock_timedrdlock
+
+conformance/interfaces/pthread_rwlockattr_getpshared/*
+ pthread_rwlockattr_getpshared not implemented
+
+conformance/interfaces/pthread_rwlockattr_setpshared/*
+ pthread_rwlockattr_setpshared not implemented
+
+conformance/interfaces/sched_rr_get_interval/*
+ syscall 161 (sched_rr_get_interval) not implemented
+
+conformance/interfaces/sigaction/21-1
+ Subtle problem: if app specifies SA_NOCLDWAIT on their SIGCHLD
+ signal handler, Valgrind will attempt to catch the SIGCHLD and
+ wait4() on all the children before returning to the app.
+ However, if the app was running a wait4() at the time the
+ SIGCHLD arrives, it will get the child's status. Quite what
+ the app is doing running wait4() when it explicitly asked for
+ it to be useless, I'm not sure...
+
+conformance/interfaces/sigaction/17-{3,6,8,12}
+ (2.4) These fail under 2.4 because they deal with SIGSEGV, SIGBUS
+ and SIGILL. These signals can only be delivered if there's a
+ thread immediately ready to handle them, but cannot be left
+ pending indefinitely. These tests hang forever because the
+ signal is discarded rather than delivered.
+
+conformance/interfaces/sigqueue/{1,4,8}-1
+ (2.4) Signals that we route manually do not have queued data
+ associated with them - they are routed with tkill. Also
+ pending signals are only kept in a mask, not in a queue, so
+ there can only be one at a time.
+
+----------------------------------------
+
+Still to investigate:
+
+conformance/interfaces/pthread_detach/4-1
+
++conformance/interfaces/pthread_rwlock_rdlock/4-1: execution: FAILED: Output:
++main: attempt write lock
++main: acquired write lock
++sig_thread: attemp read lock
++main: fire SIGUSR1 to sig_thread
++SIGUSR1 was not caught by sig_thread
+
+
++conformance/interfaces/pthread_rwlock_unlock/4-1: execution: FAILED: Output:
++Test FAILED: Incorrect error code, expected 0 or EINVAL, got 1
+
++conformance/interfaces/pthread_rwlock_wrlock/2-1: execution: FAILED: Output:
++main: attempt write lock
++sig_thread: attempt write lock
++main: fire SIGUSR1 to sig_thread
++The signal handler did not get called.
+
++conformance/interfaces/pthread_rwlock_wrlock/3-1: execution: FAILED: Output:
++
++sched status:
++
++Thread 1: status = WaitCV, associated_mx = 0x40115910, associated_cv = 0x401158E0
++==11243== at 0x40102962: pthread_cond_wait (vg_libpthread.c:1093)
++==11243== by 0x40104976: __pthread_rwlock_wrlock (vg_libpthread.c:2619)
++==11243== by 0x8048588: main (3-1.c:53)
++==11243== by 0x4013DA46: __libc_start_main (in /lib/libc-2.3.2.so)
++
++==11243== Warning: pthread scheduler exited due to deadlock
++
++valgrind: vg_main.c:1619 (vgPlain_main): Assertion `vgPlain_threads[vgPlain_last_run_tid].status == VgTs_Runnable' failed.
++
++sched status:
++
++Thread 1: status = WaitCV, associated_mx = 0x40115910, associated_cv = 0x401158E0
++==11243== at 0x40102962: pthread_cond_wait (vg_libpthread.c:1093)
++==11243== by 0x40104976: __pthread_rwlock_wrlock (vg_libpthread.c:2619)
++==11243== by 0x8048588: main (3-1.c:53)
++==11243== by 0x4013DA46: __libc_start_main (in /lib/libc-2.3.2.so)
+
+
++conformance/interfaces/sem_close/1-1.test:
+/home/jeremy/bk/valgrind/syscalls/coregrind/.in_place/libpthread.so.0:
+version `GLIBC_2.1.1' not found (required by
+conformance/interfaces/sem_close/1-1.test)
+
++conformance/interfaces/sem_timedwait/6-1: execution: FAILED: Output:
++TEST FAILED
++conformance/interfaces/sem_timedwait/6-2: execution: FAILED: Output:
++TEST FAILED
+
++conformance/interfaces/sem_timedwait/9-1: execution: FAILED: Output:
++In handler
++TEST FAILED: errno != EINTR
+
+
+conformance/interfaces/sigaction/10-1:
+ Used to work. Mysterious. Works everywhere except in the test harness...
+
+
++conformance/interfaces/sigpause/1-2: execution: FAILED: Output:
++
++valgrind: vg_mylibc.c:1324 (vgPlain_read_millisecond_timer): Assertion `rdtsc_now > rdtsc_cal_end_raw' failed.
++
++sched status:
++
++Thread 1: status = Sleeping, associated_mx = 0x0, associated_cv = 0x0
++==19929== at 0x401D6765: __GI___libc_nanosleep (in /lib/libc-2.3.2.so)
++==19929== by 0x80485C1: main (1-2.c:65)
++==19929== by 0x4013DA46: __libc_start_main (in /lib/libc-2.3.2.so)
++==19929== by 0x8048494: ??? (start.S:81)
++
++Thread 2: status = WaitSys, associated_mx = 0x0, associated_cv = 0x0
++==19929== at 0x40150796: __libc_sigsuspend (in /lib/libc-2.3.2.so)
++==19929== by 0x401509B3: __GI___sigpause (in /lib/libc-2.3.2.so)
++==19929== by 0x804857C: a_thread_func (1-2.c:48)
++==19929== by 0x40102099: thread_wrapper (vg_libpthread.c:667)
+
+
+
+----------------------------------------
+
+Fixes:
+conformance/interfaces/pthread_detach/4-2
+ This fails under NPTL, but passes under Valgrind