newlib/winsup/cygwin/DevNotes

533 lines
23 KiB
Plaintext

2013-01-31 cgf-000022
While researching the lftp behavior reported here:
http://cygwin.com/ml/cygwin/2013-01/msg00390.html
after a frenzy of rewriting sigflush handling to avoid blocking in the
signal thread (which is now and should ever have been illegal), it
dawned on me that we're not supposed to be flushing the tty input buffer
every time a signal is received. We're supposed to do this only when
the user hits a character (e.g., CTRL-C) which initiates a signal
action. So, I removed sigflush from sigpacket::process and moved it to
tc ()->kill_pgrp (). This function should only be called to send
signals related to the tty so this should have the desired effect.
2013-01-11 cgf-000021
Apparently I got the signal handling semantics of select() wrong again
even though I would have sworn that I tested this on Linux and Windows.
select() is apparently *always* interrupted by a signal and *never*
restarts. Hopefully, between the comment added to the code and this
note, I'll not make this mistake again.
2013-01-02 cgf-000020
(This entry should have been checked in with the changes but... I forgot)
This is a fairly big revamp of the way that windows signals are handled.
The intent is that all signal decisions should be made by the signal
thread; not by the exception handler.
This required the ability to pass information from the exception handler
to the signal thread so, a si_cyg field was added to siginfo_t. This
contains information needed to generate a "core dump". Hmm. Haven't
checked to see if this breaks Cygwin's hardly-ever-used real core dump
facility.
Anyway, I moved signal_exit back into exceptions.cc and removed it from
the sigpacket class. This function is now treated like a signal handler
function - Cygwin will cause it to be dispatched in the context of
whatever thread caught the signal. signal_exit also makes the
determination about when to write a stackdump.
The signal-handler thread will no longer ever attempt to exit. It will
just keep processing signals (it will not process real signals after
Cygwin stops shutting down, however). This should make it impossible
for the signal thread to ever block waiting for the process lock since
it now never grabs the process lock. The signal-handler thread will
notify gdb when it gets a signal now but, in theory, gdb should see the
context of the thread which received the signal, not the signal-handler
thread.
2012-12-28 cgf-000019
(I forgot to mention that cgf-000018 was reverted. Although I never saw
a hang from this, I couldn't convince myself that one wasn't possible.)
This fix attempts to correct a deadlock where, when a true Windows
signal arrives, Windows creates a thread which "does stuff" and attempts
to exit. In the process of exiting Cygwin grabs the process lock. If
the signal thread has seen the signal and wants to exit, it can't
because the newly-created thread now holds it. But, since the new
thread is relying on the signal thread to release its process lock,
it exits and the process lock is never released.
To fix this, I removed calls to _cygtls::signal_exit in favor of
flagging that we were exiting by setting signal_exit_code (almost forgot
to mark that NO_COPY: that would have been fun). The new function
setup_signal_exit() now handles setting things up so that ReadFile loop
in wait_sig will do the right thing when it terminates. This function
may just Sleep indefinitely if a signal is being sent from a thread
other than the signal thread. wait_sig() was changed so that it will
essentially drop into asychronous-read-mode when a signal which exits
has been detected. The ReadFile loop is exited when we know that the
process is supposed to be exiting and there is nothing else in the
signal queue.
Although I never actually saw this happen, exit_thread() was also
changed to release the process lock and just sleep indefintely if it is
detected that we are exiting.
2012-12-21 cgf-000018
Re: cgf-000017
It occurred to me that just getting the process lock during
DLL_THREAD_DETACH in dll_entry() might be adequate to fix this
problem. It's certainly much less intrusive.
There are potential deadlock problems with grabbing a lock in
this code, though, so this check-in will be experimental.
2012-12-21 cgf-000017
The changes in this set are to work around the issue noted here:
http://cygwin.com/ml/cygwin/2012-12/threads.html#00140
The problem is, apparently, that the return value of an ExitThread()
will take precedence over the return value of TerminateProcess/ExitProcess
if the thread is the last one exiting. That's rather amazing...
For the fix, I replaced all calls to ExitThread with exit_thread(). The
exit_thread function, creates a handle to the current thread and sends
it to a packet via sig_send(__SIGTHREADEXIT). Then it acquires the
process lock and calls ExitThread.
wait_sig will then wait for the handle, indicating that the thread has
exited, and, when that has happened, removes the process lock on behalf
of the now-defunct thread. wait_sig will now also avoid actually
exiting since it could trigger the same problem.
Holding process_lock should prevent threads from exiting while a Cygwin
process is shutting down. They will just block forever in that case -
just like wait_sig.
2012-08-17 cgf-000016
While debugging another problem I finally noticed that
sigpacket::process was unconditionally calling tls->set_siginfo prior to
calling setup_handler even though setup_handler could fail. In the
event of two successive signals, that would cause the second signal's
info to overwrite the first even though the signal handler for the first
would eventually be called. Doh.
Fixing this required passing the sigpacket si field into setup_handler.
Making setup_handler part of the sigpacket class seemed to make a lot of
sense so that's what I did. Then I passed the si element into
interrupt_setup so that the infodata structure could be filled out prior
to arming the signal.
The other changes checked in here eliminate the ResetEvent for
signal_arrived since previous changes to cygwait should handle the
case of spurious signal_arrived detection. Since signal_arrived is
not a manual-reset event, we really should just let the appropriate
WFMO handle it. Otherwise, there is a race where a signal comes in
a "split second" after WFMO responds to some other event. Resetting
the signal_arrived would cause any subsequent WFMO to never be
triggered. My current theory is that this is what is causing:
http://cygwin.com/ml/cygwin/2012-08/msg00310.html
2012-08-15 cgf-000015
RIP cancelable_wait. Yay.
2012-08-09 cgf-000014
So, apparently I got it somewhat right before wrt signal handling.
Checking on linux, it appears that signals will be sent to a thread
which can accept the signal. So resurrecting and extending the
"find_tls" function is in order. This function will return the tls
of any thread which 1) is waiting for a signal with sigwait*() or
2) has the signal unmasked.
In redoing this it became obvious that I had the class designation wrong
for the threadlist handling so I moved the manipulation of the global
threadlist into the cygheap where it logically belongs.
2012-07-21 cgf-000013
These changes reflect a revamp of the "wait for signal" functionality
which has existed in Cygwin through several signal massages.
We now create a signal event only when a thread is waiting for a signal
and arm it only for that thread. The "set_signal_arrived" function is
used to establish the event and set it in a location referencable by
the caller.
I still do not handle all of the race conditions. What happens when
a signal comes in just after a WF?O succeeds for some other purpose? I
suspect that it will arm the next WF?O call and the subsequent call to
call_signal_handler could cause a function to get an EINTR when possibly
it shouldn't have.
I haven't yet checked all of the test cases for the URL listed in the
previous entry.
Baby steps.
2012-06-12 cgf-000012
These changes are the preliminary for redoing the way threads wait for
signals. The problems are shown by the test case mentioned here:
http://cygwin.com/ml/cygwin/2012-05/msg00434.html
I've known that the signal handling in threads wasn't quite right for
some time. I lost all of my thread signal tests in the great "rm -r"
debacle of a few years ago and have been less than enthusiastic about
redoing everything (I had PCTS tests and everything). But it really is
time to redo this signal handling to make it more like it is supposed to
be.
This change should not introduce any new behavior. Things should
continue to behave as before. The major differences are a change in the
arguments to cancelable_wait and cygwait now uses cancelable_wait and,
so, the returns from cygwait now mirror cancelable_wait.
The next change will consolidate cygwait and cancelable_wait into one
cygwait function.
2012-06-02 cgf-000011
The refcnt handling was tricky to get right but I had convinced myself
that the refcnt's were always incremented/decremented under a lock.
Corinna's 2012-05-23 change to refcnt exposed a potential problem with
dup handling where the fdtab could be updated while not locked.
That should be fixed by this change but, on closer examination, it seems
like there are many places where it is possible for the refcnt to be
updated while the fdtab is not locked since the default for
cygheap_fdget is to not lock the fdtab (and that should be the default -
you can't have read holding a lock).
Since refcnt was only ever called with 1 or -1, I broke it up into two
functions but kept the Interlocked* operation. Incrementing a variable
should not be as racy as adding an arbitrary number to it but we have
InterlockedIncrement/InterlockedDecrement for a reason so I kept the
Interlocked operation here.
In the meantime, I'll be mulling over whether the refcnt operations are
actually safe as they are. Maybe just ensuring that they are atomically
updated is enough since they control the destruction of an fh. If I got
the ordering right with incrementing and decrementing then that should
be adequate.
2012-06-02 cgf-000010
<1.7.16>
- Fix emacs problem which exposed an issue with Cygwin's select() function.
If a signal arrives while select is blocking and the program longjmps
out of the signal handler then threads and memory may be left hanging.
Fixes: http://cygwin.com/ml/cygwin/2012-05/threads.html#00275
</1.7.16>
This was try #4 or #5 to get select() signal handling working right.
It's still not there but it should now at least not leak memory or
threads.
I mucked with the interface between cygwin_select and select_stuff::wait
so that the "new" loop in select_stuff::wait() was essentially moved
into the caller. cygwin_select now uses various enum states to decide
what to do. It builds the select linked list at the beginning of the
loop, allowing wait() to tear everything down and restart. This is
necessary before calling a signal handler because the signal handler may
longjmp away.
I initially had this all coded up to use a special signal_cleanup
callback which could be called when a longjmp is called in a signal
handler. And cygwin_select() set up and tore down this callback. Once
I got everything compiling it, of course, dawned on me that just because
you call a longjmp in a signal handler it doesn't mean that you are
jumping *out* of the signal handler. So, if the signal handler invokes
the callback and returns it will be very bad for select(). Hence, this
slower, but hopefully more correct implementation.
(I still wonder if some sort of signal cleanup callback might still
be useful in the future)
TODO: I need to do an audit of other places where this problem could be
occurring.
As alluded to above, select's signal handling is still not right. It
still acts as if it could call a signal handler from something other
than the main thread but, AFAICT, from my STC, this doesn't seem to be
the case. It might be worthwhile to extend cygwait to just magically
figure this out and not even bother using w4[0] for scenarios like this.
2012-05-16 cgf-000009
<1.7.16>
- Fix broken console mouse handling. Reported here:
http://cygwin.com/ml/cygwin/2012-05/msg00360.html
</1.7.16>
I did a cvs annotate on smallprint.cc and see that the code to translate
%characters > 127 to 0x notation was in the 1.1 revision. Then I
checked the smallprint.c predecessor. It was in the 1.1 version of that
program too, which means that this odd change has probably been around
since <= 2000.
Since __small_sprintf is supposed to emulate sprintf, I got rid of the
special case handling. This may affect fhandler_socket::bind. If so, we
should work around this problem there rather than keeping this strange
hack in __small_printf.
2012-05-14 cgf-000008
<1.7.16>
- Fix hang when zero bytes are written to a pty using
Windows WriteFile or equivalent. Fixes:
http://cygwin.com/ml/cygwin/2012-05/msg00323.html
</1.7.16>
cgf-000002, as usual, fixed one thing while breaking another. See
Larry's predicament in: http://goo.gl/oGEr2 .
The problem is that zero byte writes to the pty pipe caused the dread
end-of-the-world-as-we-know-it problem reported on the mailing list
where ReadFile reads zero bytes even though there is still more to read
on the pipe. This is because that change caused a 'record' to be read
and a record can be zero bytes.
I was never really keen about using a throwaway buffer just to get a
count of the number of characters available to be read in the pty pipe.
On closer reading of the documentation for PeekNamedPipe it seemed like
the sixth argument to PeekNamedPipe should return what I needed without
using a buffer. And, amazingly, it did, except that the problem still
remained - a zero byte message still screwed things up.
So, we now detect the case where there is zero bytes available as a
message but there are bytes available in the pipe. In that scenario,
return the bytes available in the pipe rather than the message length of
zero. This could conceivably cause problems with pty pipe handling in
this scenario but since the only way this scenario could possibly happen
is when someone is writing zero bytes using WriteFile to a pty pipe, I'm
ok with that.
2012-05-14 cgf-000007
<1.7.16>
- Fix invocation of strace from a cygwin process. Fixes:
http://cygwin.com/ml/cygwin/2012-05/msg00292.html
</1.7.16>
The change in cgf-000004 introduced a problem for processes which load
cygwin1.dll dynamically. strace.exe is the most prominent example of
this.
Since the parent handle is now closed for "non-Cygwin" processes, when
strace.exe tried to dynamically load cygwin1.dll, the handle was invalid
and child_info_spawn::handle_spawn couldn't use retrieve information
from the parent. This eventually led to a strace_printf error due to an
attempt to dereference an unavailable cygheap. Probably have to fix
this someday. You shouldn't use the cygheap while attempting to print
an error about the inavailability of said cygheap.
This was fixed by saving the parent pid in child_info_spawn and calling
OpenProcess for the parent pid and using that handle iff a process is
dynamically loaded.
2012-05-12 cgf-000006
<1.7.16>
- Fix hang when calling pthread_testcancel in a canceled thread.
Fixes some of: http://cygwin.com/ml/cygwin/2012-05/msg00186.html
</1.7.16>
This should fix the first part of the reported problem in the above
message. The cancel seemed to actually be working but, the fprintf
eventually ended up calling pthread_testcancel. Since we'd gotten here
via a cancel, it tried to recursively call the cancel handler causing a
recursive loop.
2012-05-12 cgf-000005
<1.7.16>
- Fix pipe creation problem which manifested as a problem creating a
fifo. Fixes: http://cygwin.com/ml/cygwin/2012-05/msg00253.html
</1.7.16>
My change on 2012-04-28 introduced a problem with fifos. The passed
in name was overwritten. This was because I wasn't properly keeping
track of the length of the generated pipe name when there was a
name passed in to fhandler_pipe::create.
There was also another problem in fhandler_pipe::create. Since fifos
use PIPE_ACCESS_DUPLEX and PIPE_ACCESS_DUPLEX is an or'ing of
PIPE_ACCESS_INBOUND and PIPE_ACCESS_OUTBOUND, using PIPE_ACCESS_OUTBOUND
as a "never-used" option for PIPE_ADD_PID in fhandler.h was wrong. So,
fifo creation attempted to add the pid of a pipe to the name which is
wrong for fifos.
2012-05-08 cgf-000004
The change for cgf-000003 introduced a new problem:
http://cygwin.com/ml/cygwin/2012-05/msg00154.html
http://cygwin.com/ml/cygwin/2012-05/msg00157.html
Since a handle associated with the parent is no longer being duplicated
into a non-cygwin "execed child", Windows is free to reuse the pid of
the parent when the parent exits. However, since we *did* duplicate a
handle pointing to the pid's shared memory area into the "execed child",
the shared memory for the pid was still active.
Since the shared memory was still available, if a new process reuses the
previous pid, Cygwin would detect that the shared memory was not created
and had a "PID_REAPED" flag. That was considered an error, and, so, it
would set procinfo to NULL and pinfo::thisproc would die since this
situation is not supposed to occur.
I fixed this in two ways:
1) If a shared memory region has a PID_REAPED flag then zero it and
reuse it. This should be safe since you are not really supposed to be
querying the shared memory region for anything after PID_REAPED has been
set.
2) Forego duping a copy of myself_pinfo if we're starting a non-cygwin
child for exec.
It seems like 2) is a common theme and an audit of all of the handles
that are being passed to non-cygwin children is in order for 1.7.16.
The other minor modification that was made in this change was to add the
pid of the failing process to fork error output. This helps slightly
when looking at strace output, even though in this case it was easy to
find what was failing by looking for '^---' when running the "stv"
strace dumper. That found the offending exception quickly.
2012-05-07 cgf-000003
<1.7.15>
Don't make Cygwin wait for all children of a non-cygwin child program.
Fixes: http://cygwin.com/ml/cygwin/2012-05/msg00063.html,
http://cygwin.com/ml/cygwin/2012-05/msg00075.html
</1.7.15>
This problem is due to a recent change which added some robustness and
speed to Cygwin's exec/spawn handling by not trying to force inheritance
every time a process is started. See ChangeLog entries starting on
2012-03-20, and multiple on 2012-03-21.
Making the handle inheritable meant that, as usual, there were problems
with non-Cygwin processes. When Cygwin "execs" a non-Cygwin process N,
all of its N + 1, N + 2, ... children will also inherit the handle.
That means that Cygwin will wait until all subprocesses have exited
before it returns.
I was willing to make this a restriction of starting non-Cygwin
processes but the problem with allowing that is that it can cause the
creation of a "limbo" pid when N exits and N + 1 and friends are still
around. In this scenario, Cygwin dutifully notices that process N has
died and sets the exit code to indicate that but N's parent will wait on
rd_proc_pipe and will only return when every N + ... windows process
has exited.
The removal of cygheap::pid_handle was not related to the initial
problem that I set out to fix. The change came from the realization
that we were duping the current process handle into the child twice and
only needed to do it once. The current process handle is used by exec
to keep the Windows pid "alive" so that it will not be reused. So, now
we just close parent in child_info_spawn::handle_spawn iff we're not
execing.
In debugging this it bothered me that 'ps' identified a nonactive pid as
active. Part of the reason for this was the 'parent' handle in
child_info was opened in non-Cygwin processes, keeping the pid alive.
That has been kluged around (more changes after 1.7.15) but that didn't
fix the problem. On further investigation, this seems to be caused by
the fact that the shared memory region pid handles were still being
passed to non-cygwin children, keeping the pid alive in a limbo-like
fashion. This was easily fixed by having pinfo::init() consider a
memory region with PID_REAPED as not available. A more robust fix
should be considered for 1.7.15+ where these handles are not passed
to non-cygwin processes.
This fixed the problem where a pid showed up in the list after a user
does something like: "bash$ cmd /c start notepad" but, for some reason,
it does not fix the problem where "bash$ setsid cmd /c start notepad".
That bears investigation after 1.7.15 is released but it is not a
regression and so is not a blocker for the release.
2012-05-03 cgf-000002
<1.7.15>
Fix problem where too much input was attempted to be read from a
pty slave. Fixes: http://cygwin.com/ml/cygwin/2012-05/msg00049.html
</1.7.15>
My change on 2012/04/05 reintroduced the problem first described by:
http://cygwin.com/ml/cygwin/2011-10/threads.html#00445
The problem then was, IIRC, due to the fact that bytes sent to the pty
pipe were not written as records. Changing pipe to PIPE_TYPE_MESSAGE in
pipe.cc fixed the problem since writing lines to one side of the pipe
caused exactly that the number of characters to be read on the other
even if there were more characters in the pipe.
To debug this, I first replaced fhandler_tty.cc with the 1.258,
2012/04/05 version. The test case started working when I did that.
So, then, I replaced individual functions, one at a time, in
fhandler_tty.cc with their previous versions. I'd expected this to be a
problem with fhandler_pty_master::process_slave_output since that had
seen the most changes but was surprised to see that the culprit was
fhandler_pty_slave::read().
The reason was that I really needed the bytes_available() function to
return the number of bytes which would be read in the next operation
rather than the number of bytes available in the pipe. That's because
there may be a number of lines available to be read but the number of
bytes which will be read by ReadFile should reflect the mode of the pty
and, if there is a line to read, only the number of bytes in the line
should be seen as available for the next read.
Having bytes_available() return the number of bytes which would be read
seemed to fix the problem but it could subtly change the behavior of
other callers of this function. However, I actually think this is
probably a good thing since they probably should have been seeing the
line behavior.
2012-05-02 cgf-000001
<1.7.15>
Fix problem setting parent pid to 1 when process with children execs
itself. Fixes: http://cygwin.com/ml/cygwin/2012-05/msg00009.html
</1.7.15>
Investigating this problem with strace showed that ssh-agent was
checking the parent pid and getting a 1 when it shouldn't have. Other
stuff looked ok so I chose to consider this a smoking gun.
Going back to the version that the OP said did not have the problem, I
worked forward until I found where the problem first occurred -
somewhere around 2012-03-19. And, indeed, the getppid call returned the
correct value in the working version. That means that this stopped
working when I redid the way the process pipe was inherited around
this time period.
It isn't clear why (and I suspect I may have to debug this further at
some point) this hasn't always been a problem but I made the obvious fix.
We shouldn't have been setting ppid = 1 when we're about to pass off to
an execed process.
As I was writing this, I realized that it was necessary to add some
additional checks. Just checking for "have_execed" isn't enough. If
we've execed a non-cygwin process then it won't know how to deal with
any inherited children. So, always set ppid = 1 if we've execed a
non-cygwin process.