This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
[PATCH]Fix that GDB will get hang on Windows when using pipe to get stdout and stderr from stub
- From: "Terry Guo" <terry dot guo at arm dot com>
- To: <gdb-patches at sourceware dot org>
- Cc: <eliz at gnu dot org>, "Joey Ye" <Joey dot Ye at arm dot com>, "Matthew Gretton-Dann" <Matthew dot Gretton-Dann at arm dot com>, "'Pedro Alves'" <palves at redhat dot com>, <daniel dot jacobowitz at gmail dot com>
- Date: Tue, 26 Jun 2012 09:13:14 +0800
- Subject: [PATCH]Fix that GDB will get hang on Windows when using pipe to get stdout and stderr from stub
Hi,
I noticed a cross-built MINGW arm-none-eabi GDB will get hang on Windows
when use pipe to get stderr and stdout from stub. The command used to start
stub in GDB is "target extended-remote |
stub-that-write-stderr-before-stdout". For my case, after send
"$vFlashDone#ea" to stub, GDB get hang. The GDB source show that GDB will
keep waiting for ACK message from stdout of stub, after send the packet.
Unfortunately my stub will write some kind of log information into stderr
and this action takes place before stub write ACK message to its stdout. So
the only pipe is occupied by stderr which is waiting for GDB to consume,
while GDB keep waiting for message from the stdout which hasn't pipe to use.
We finally end up with a deadlock on pipe between GDB/stderr/stdout.
The following patch can avoid such deadlock by letting GDB also probe and
consume stderr when waiting for stdout. Please review and comment.
The Linux version GDB hasn't such issue. I think it's because we use
different way to handle PIPE as stated in functions pipe_open and
pipe_windows_open. For Linux we have two socketpair kind pipes, one for
stdout and one for stderr. While for windows, we only have one pipe which is
created by _pipe function.
BR,
Terry
2012-06-25 Terry Guo <terry.guo@arm.com>
* ser_base (ser_base_read_error_fd): New function.
(do_ser_base_readchar): Poll error file descriptor as well as
standard output.
(generic_readchar): Refactor error handling.
diff --git a/gdb/ser-base.c b/gdb/ser-base.c
index 368afa6..ee6db54 100644
--- a/gdb/ser-base.c
+++ b/gdb/ser-base.c
@@ -223,6 +223,63 @@ ser_base_wait_for (struct serial *scb, int timeout)
}
}
+/* Read any error output we might have. */
+
+void
+ser_base_read_error_fd (struct serial *scb, int close_fd)
+{
+ if (scb->error_fd != -1)
+ {
+ ssize_t s;
+ char buf[81];
+
+ for (;;)
+ {
+ char *current;
+ char *newline;
+ int to_read = 80;
+ int num_bytes = -1;
+
+ if (scb->ops->avail)
+ num_bytes = (scb->ops->avail)(scb, scb->error_fd);
+
+ if (num_bytes != -1)
+ to_read = (num_bytes < to_read) ? num_bytes : to_read;
+
+ if (to_read == 0)
+ break;
+
+ s = read (scb->error_fd, &buf, to_read);
+ if ((s == -1) || (s == 0 && !close_fd))
+ break;
+
+ if (s == 0 && close_fd)
+ {
+ /* End of file. */
+ close (scb->error_fd);
+ scb->error_fd = -1;
+ break;
+ }
+
+ /* In theory, embedded newlines are not a problem.
+ But for MI, we want each output line to have just
+ one newline for legibility. So output things
+ in newline chunks. */
+ buf[s] = '\0';
+ current = buf;
+ while ((newline = strstr (current, "\n")) != NULL)
+ {
+ *newline = '\0';
+ fputs_unfiltered (current, gdb_stderr);
+ fputs_unfiltered ("\n", gdb_stderr);
+ current = newline + 1;
+ }
+
+ fputs_unfiltered (current, gdb_stderr);
+ }
+ }
+}
+
/* Read a character with user-specified timeout. TIMEOUT is number of
seconds
to wait, or -1 to wait forever. Use timeout of 0 to effect a poll.
Returns
char if successful. Returns -2 if timeout expired, EOF if line dropped
@@ -273,6 +330,11 @@ do_ser_base_readchar (struct serial *scb, int timeout)
status = SERIAL_TIMEOUT;
break;
}
+
+ /* We also need to check and consume the stderr because it could
+ come before the stdout for some stubs. If we just sit and wait
+ for stdout, we would hit a deadlock for that case. */
+ ser_base_read_error_fd (scb, 0);
}
if (status < 0)
@@ -344,53 +406,7 @@ generic_readchar (struct serial *scb, int timeout,
}
}
/* Read any error output we might have. */
- if (scb->error_fd != -1)
- {
- ssize_t s;
- char buf[81];
-
- for (;;)
- {
- char *current;
- char *newline;
- int to_read = 80;
-
- int num_bytes = -1;
- if (scb->ops->avail)
- num_bytes = (scb->ops->avail)(scb, scb->error_fd);
- if (num_bytes != -1)
- to_read = (num_bytes < to_read) ? num_bytes : to_read;
-
- if (to_read == 0)
- break;
-
- s = read (scb->error_fd, &buf, to_read);
- if (s == -1)
- break;
- if (s == 0)
- {
- /* EOF */
- close (scb->error_fd);
- scb->error_fd = -1;
- break;
- }
-
- /* In theory, embedded newlines are not a problem.
- But for MI, we want each output line to have just
- one newline for legibility. So output things
- in newline chunks. */
- buf[s] = '\0';
- current = buf;
- while ((newline = strstr (current, "\n")) != NULL)
- {
- *newline = '\0';
- fputs_unfiltered (current, gdb_stderr);
- fputs_unfiltered ("\n", gdb_stderr);
- current = newline + 1;
- }
- fputs_unfiltered (current, gdb_stderr);
- }
- }
+ ser_base_read_error_fd (scb, 1);
reschedule (scb);
return ch;