9:12 a.m.
Hi all,
This pertains to 21.4.24 and earlier; I have not checked the 21.5 branch.
There is a very rusty API called connec_to_file_descriptor() in
process-unix.c. I am using it in an emodule.
It works fine, from what I can tell. The problem I am encountering
now, is that when the in fd is a pipe, and the pipe has
less-than-internal-buffer-length it blocks on a subsequent read()
system call.
I am currently not sure how to "fix it". The culprit seems to be the
if in Lstream_read(), inside the while loop. And the result is that
it'll block on the read in filedesc_reader().
It's very wrong to adjust connect_to_file_descriptor() to make the resulting fd
stream allow quits when read()ing. That just pushes the bug into
user-visible space.
Adding poll or select into the filedesc reader seems like the wrong
solution, at first glance.
Ideally, there should be a way to tell XEmacs to read() exactly once,
but that also means getting rid of the while loop in Lstream_read()
from what I can tell.
Which may not be a good idea overall.
Is there a way to fix this problem with minimal refactoring? Or am I
cursed to have to fix all the internal buffering system in XEmacs to
make connect_to_file_descriptor() usable with pipes?
--
Johann
I'm not from the internet, I just work there.
11:49 a.m.
On Sat, Dec 9, 2017 at 5:12 PM, Johann 'Myrkraverk' Oskarsson
<johann(a)myrkraverk.com> wrote:
There is a very rusty API called connec_to_file_descriptor() in
process-unix.c. I am using it in an emodule.
It works fine, from what I can tell. The problem I am encountering
now, is that when the in fd is a pipe, and the pipe has
less-than-internal-buffer-length it blocks on a subsequent read()
system call.
Ok, so I "worked around this problem" by setting the read end of the
pipe to nonblocking mode before calling connect_to_file_descriptor().
I don't really call that an acceptable solution. It'll be much better if
XEmacs takes care of that on its own. Possibly inside
unix_init_process_io_handles()
though I don't know if it'll have any adverse effect on normal process
handling, but I'll try to experiment with that some day.
--
Johann
I'm not from the internet, I just work there.
3:48 p.m.
On Sat, Dec 9, 2017 at 7:49 PM, Johann 'Myrkraverk' Oskarsson
<johann(a)myrkraverk.com> wrote:
Ok, so I "worked around this problem" by setting the read
end of the
pipe to nonblocking mode before calling connect_to_file_descriptor().
This seems to work fine for me. The pre-patched version is just one line,
so I paste it here in its entirety.
static void
unix_init_process_io_handles (Lisp_Process *p, void* in, void* out, int flags)
{
UNIX_DATA(p)->infd = (int)in;
/* In the event where the input file descriptor is a pipe, XEmacs as of
* 21.4.24 will hang on the read() system call when the pipe doesn't
* have enough to fill the incoming buffer.
*
* This will likely affect other kinds of blocking file descriptors also,
* so we unconditionally set the input fd to non-blocking mode.
*/
{
int f = fcntl( UNIX_DATA(p)->infd, F_SETFL, O_NONBLOCK );
if ( f == -1 ) {
perror( "Setting pipe to nonblocking mode" );
}
}
}
Enjoy,
--
Johann
I'm not from the internet, I just work there.
8:29 a.m.
New subject: [COMMIT] Set input FDs to non-blocking unconditionally, thank you Myrkraverk
Ar an deichiú lá de mí na Nollaig, scríobh Johann 'Myrkraverk' Oskarsson:
On Sat, Dec 9, 2017 at 7:49 PM, Johann 'Myrkraverk'
Oskarsson
<johann(a)myrkraverk.com> wrote:
> Ok, so I "worked around this problem" by setting the read end of the
> pipe to nonblocking mode before calling connect_to_file_descriptor().
>
This seems to work fine for me. The pre-patched version is just one line,
so I paste it here in its entirety.
I’ve gone ahead and applied that to 21.5. I have occasionally got hangs
especially when VM is using an external program for qp-decoding, and I suspect
your issue is at the root of them. I don’t see any downside to your change,
XEmacs code further up the call stack needs to be able to handle smaller reads
anyway. Thank you!
APPROVE COMMIT
NOTE: This patch has been committed
# HG changeset patch
# User Aidan Kehoe <kehoea(a)parhasard.net>
# Date 1513067231 0
# Tue Dec 12 08:27:11 2017 +0000
# Node ID affc195870d60c7f3f61d304cc01f205f63432d8
# Parent 0f7f0274157b50c71683368d0614b8912e2b70d7
Set input FDs to non-blocking unconditionally, thank you Myrkraverk
src/ChangeLog addition:
2017-12-12 Aidan Kehoe <kehoea(a)parhasard.net>
* process-unix.c (unix_init_process_io_handles):
Set input FDs to non-blocking unconditionally, see recent patch
posted to xemacs-beta from Johann "Myrkraverk" Oskarsson.
diff -r 0f7f0274157b -r affc195870d6 src/ChangeLog
--- a/src/ChangeLog Wed Dec 06 20:52:28 2017 +0000
+++ b/src/ChangeLog Tue Dec 12 08:27:11 2017 +0000
＠＠ -1,3 +1,9 ＠＠
+2017-12-12 Aidan Kehoe <kehoea(a)parhasard.net>
+
+ * process-unix.c (unix_init_process_io_handles):
+ Set input FDs to non-blocking unconditionally, see recent patch
+ posted to xemacs-beta from Johann "Myrkraverk" Oskarsson.
+
2017-12-02 Aidan Kehoe <kehoea(a)parhasard.net>
* bytecode.c (set_compiled_function_arglist):
diff -r 0f7f0274157b -r affc195870d6 src/process-unix.c
--- a/src/process-unix.c Wed Dec 06 20:52:28 2017 +0000
+++ b/src/process-unix.c Tue Dec 12 08:27:11 2017 +0000
＠＠ -831,6 +831,25 ＠＠
/* if sizeof(EMACS_INT) > sizeof(int) this truncates the value */
UNIX_DATA(p)->infd = (EMACS_INT) in;
UNIX_DATA(p)->errfd = (EMACS_INT) err;
+
+ /* In the event where the input file descriptor is a pipe, XEmacs as of
+ * 21.4.24 will hang on the read() system call when the pipe doesn't
+ * have enough to fill the incoming buffer.
+ *
+ * This will likely affect other kinds of blocking file descriptors also, so
+ * we unconditionally set the input fd to non-blocking mode. See:
+ *
http://mid.xemacs.org/CAA_SOw__WYNiUO=4hHhiwg0LXq7++9Xr=tnBMgN5Uc8iAa+mvg...
+ * and the related thread.
+ */
+ {
+ int f = fcntl (UNIX_DATA(p)->infd, F_SETFL, O_NONBLOCK);
+ if (f == -1)
+ {
+ warn_when_safe (Qio_error, lisp_strerror (errno),
+ "failed setting pipe (fd %ld) to nonblocking mode",
+ (EMACS_INT) in);
+ }
+ }
}
/* Move the file descriptor FD so that its number is not less than MIN. *
--
‘As I sat looking up at the Guinness ad, I could never figure out /
How your man stayed up on the surfboard after forty pints of stout’
(C. Moore)
2537
days inactive
2540
days old
3 comments
2 participants
participants (2)
-
Aidan Kehoe -
Johann 'Myrkraverk' Oskarsson