This is the mail archive of the gdb-patches@sourceware.org mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Windows semihosting (remote-fileio) fix for console reads


First of all, a disclaimer.  I will absolutely understand if someone wants
to draw the line on broken system support on the wrong side of this patch.
It's incredibly bizarre.  I'm going to let it sit for comments for a bit; if
you object, please let me know.  I don't really know how I could move the
ugly bit to a Windows-specific file, but maybe we can come up with
something.

Here's the problem: a customer tried to use the remote file I/O protocol to
read from standard input.  Trivial little program, just a read() from stdin
followed by a write() to stdout.  Read was returning -1.  A well-placed
printf discovered that this was actually ENOMEM.

By trial and error, we worked out that on the system we were using a read of
26609 bytes from the console, or more, would always return ENOMEM.  Given
26608 bytes or less, it would succeed normally.  Please don't ask me what
that number means.  No, it does not vary with the size of the malloced
buffer.

Read is always allowed to return a short value, and it was already
hardcoding a buffer size of 32767.  So this changes that buffer to 8K,
with a comment about the arbitrary number.

I don't believe there's anywhere else we do big reads from stdin.
Normally we want one character at a time.

-- 
Daniel Jacobowitz
CodeSourcery

2006-06-10  Daniel Jacobowitz  <dan@codesourcery.com>

	gdb/
	* remote-fileio.c (remote_fileio_func_read): Limit console
	reads to 8K.

Index: remote-fileio.c
===================================================================
RCS file: /cvs/src/src/gdb/remote-fileio.c,v
retrieving revision 1.18
diff -u -p -r1.18 remote-fileio.c
--- remote-fileio.c	8 Jun 2006 19:08:22 -0000	1.18
+++ remote-fileio.c	10 Jun 2006 18:11:51 -0000
@@ -729,7 +729,7 @@ remote_fileio_func_read (char *buf)
 	  static char *remaining_buf = NULL;
 	  static int remaining_length = 0;
 
-	  buffer = (gdb_byte *) xmalloc (32768);
+	  buffer = (gdb_byte *) xmalloc (8192);
 	  if (remaining_buf)
 	    {
 	      remote_fio_no_longjmp = 1;
@@ -751,7 +751,16 @@ remote_fileio_func_read (char *buf)
 	    }
 	  else
 	    {
-	      ret = ui_file_read (gdb_stdtargin, (char *) buffer, 32767);
+	      /* NOTE drow/2006-06-10: Windows (mingw32) has a truly
+		 bizarre bug.  If a handle is backed by a real console
+		 device, overly large reads from the console will fail
+		 and set errno == ENOMEM.  On a Windows Server 2003
+		 system where I tested, reading 26608 bytes from the
+		 console was OK, but anything about 26609 bytes would
+		 fail.  So, we limit this read to something smaller
+		 than that - by a safe margin, in case the limit
+		 depends on system resources or version.  */
+	      ret = ui_file_read (gdb_stdtargin, (char *) buffer, 8191);
 	      remote_fio_no_longjmp = 1;
 	      if (ret > 0 && (size_t)ret > length)
 		{


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]