mmap bug on Windows 9x
On Windows 9x/ME different calls to mmap sometimes produce the same address (without that memory being unmapped in the meantime, at least not by application code). Here's a trace of the application calls to mmap (on WME): try mmap($0, $400038, ..., MAP_ANON, ...); success, address=$8395 try mmap($8395, $2cfd0, ..., MAP_FIXED|MAP_FILE, imagefile, 0); failed: Not enough core try mmap($83d52000, $400038, ..., MAP_ANON, ...); success, address=$8395 try mmap($83d52000, $4000, ..., MAP_ANON, ...); success, address=$83d51000 try mmap($83d56000, $4000, ..., MAP_ANON, ...); success, address=$83d55000 try mmap($83d5a000, $3c00, ..., MAP_ANON, ...); success, address=$83d59000 try mmap($83d5e000, $3a00, ..., MAP_ANON, ...); success, address=$83466000 try mmap($8346b000, $4, ..., MAP_ANON, ...); success, address=$83d6 try mmap($83da1000, $4, ..., MAP_ANON, ...); success, address=$83d6 Note that the last two calls produce the same address. This happens with cygwin1.dll versions 1.5.5-1 and 1.5.10-3 (and probably also with 1.3.22-1). On Windows 2000 (with cygwin1.dll 1.3.22) the same application produces the following trace: try mmap($0, $400038, ..., MAP_ANON, ...); success, address=$65 try mmap($65, $2cfd0, ..., MAP_FIXED|MAP_FILE, imagefile, 0); failed: Value too large for defined data type try mmap($a52000, $400038, ..., MAP_ANON, ...); success, address=$a6 try mmap($e62000, $4000, ..., MAP_ANON, ...); success, address=$a51000 try mmap($a56000, $4000, ..., MAP_ANON, ...); success, address=$a55000 try mmap($a5a000, $3c00, ..., MAP_ANON, ...); success, address=$a59000 try mmap($a5e000, $3a00, ..., MAP_ANON, ...); success, address=$e61000 try mmap($e66000, $4, ..., MAP_ANON, ...); success, address=$e7 try mmap($eb1000, $4, ..., MAP_ANON, ...); success, address=$eb and everything works fine. In this posting I just want to check if the bug is already known. If not, I will condense the application to something that exhibits the bug. Or, if you feel adventurous, you can deal with the full application: http://www.complang.tuwien.ac.at/forth/gforth/gforth-0.6.2.tar.gz (also available from GNU mirrors); you get the trace output above if you build gforth with gcc-3.3 and call it with --debug. Thanks for your work on Cygwin. - anton -- Unsubscribe info: http://cygwin.com/ml/#unsubscribe-simple Problem reports: http://cygwin.com/problems.html Documentation: http://cygwin.com/docs.html FAQ: http://cygwin.com/faq/
Re: mmap bug on Windows 9x
Anton Ertl wrote:
>
> On Windows 9x/ME different calls to mmap sometimes produce the same
> address (without that memory being unmapped in the meantime, at least
> not by application code).
You find a condensed test program below.
On Windows ME with cygwin1.dll 1.5.10 it outputs:
try mmap($0, $4, ..., MAP_ANON, ...); success, address=$833ca000
try mmap($8340b000, $4, ..., MAP_ANON, ...); success, address=$833ca000
Note that the result addresses are the same.
On Windows 2000 with cygwin1.dll 1.3.22 it outputs:
try mmap($0, $4, ..., MAP_ANON, ...); success, address=$64
try mmap($681000, $4, ..., MAP_ANON, ...); success, address=$68
i.e., non-overlapping memory ranges, as it should.
I believe that this is a problem that both cygwin1.dlls have on
Windows ME/9x (and not on NT/2K/XP), not something arising from the
difference between cygwin versions (we see a problem that probably
arises from this bug also with the Gforth-for-Windows package that
includes cygwin1.dll 1.3.22).
- anton
Here's the test program (compile and run with "gcc -O main.c; ./a):
-
#include
#include
#include
#include
#include
#include
#include
#ifndef STANDALONE
#include
#endif
#include
#include
#include
#include
#ifndef STANDALONE
#include
#endif
int debug=1;
#define HAVE_MMAP
typedef char *Address;
typedef int Cell;
int pagesize=4096;
static Address next_address=0;
void after_alloc(Address r, Cell size)
{
if (r != (Address)-1) {
if (debug)
fprintf(stderr, "success, address=$%lx\n", (long) r);
if (pagesize != 1)
next_address = (Address)(Cell)r)+size-1)&-pagesize)+2*pagesize); /* leave
one page unmapped */
} else {
if (debug)
fprintf(stderr, "failed: %s\n", strerror(errno));
}
}
#ifndef MAP_FAILED
#define MAP_FAILED ((Address) -1)
#endif
#ifndef MAP_FILE
# define MAP_FILE 0
#endif
#ifndef MAP_PRIVATE
# define MAP_PRIVATE 0
#endif
#if !defined(MAP_ANON) && defined(MAP_ANONYMOUS)
# define MAP_ANON MAP_ANONYMOUS
#endif
#if defined(HAVE_MMAP)
static Address alloc_mmap(Cell size)
{
Address r;
#if defined(MAP_ANON)
if (debug)
fprintf(stderr,"try mmap($%lx, $%lx, ..., MAP_ANON, ...); ", (long)next_address,
(long)size);
r = mmap(next_address, size, PROT_EXEC|PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE,
-1, 0);
#else /* !defined(MAP_ANON) */
/* Ultrix (at least) does not define MAP_FILE and MAP_PRIVATE (both are
apparently defaults) */
static int dev_zero=-1;
if (dev_zero == -1)
dev_zero = open("/dev/zero", O_RDONLY);
if (dev_zero == -1) {
r = MAP_FAILED;
if (debug)
fprintf(stderr, "open(\"/dev/zero\"...) failed (%s), no mmap; ",
strerror(errno));
} else {
if (debug)
fprintf(stderr,"try mmap($%lx, $%lx, ..., MAP_FILE, dev_zero, ...); ",
(long)next_address, (long)size);
r=mmap(next_address, size, PROT_EXEC|PROT_READ|PROT_WRITE, MAP_FILE|MAP_PRIVATE,
dev_zero, 0);
}
#endif /* !defined(MAP_ANON) */
after_alloc(r, size);
return r;
}
#endif
int main(int argc, char **argv, char **env)
{
debug=1;
alloc_mmap(0x4);
alloc_mmap(0x4);
exit(0);
}
--
Unsubscribe info: http://cygwin.com/ml/#unsubscribe-simple
Problem reports: http://cygwin.com/problems.html
Documentation: http://cygwin.com/docs.html
FAQ: http://cygwin.com/faq/
Re: mmap bug on Windows 9x
On Jul 16 09:12, Anton Ertl wrote: > Anton Ertl wrote: > > > > On Windows 9x/ME different calls to mmap sometimes produce the same > > address (without that memory being unmapped in the meantime, at least > > not by application code). > > You find a condensed test program below. > > On Windows ME with cygwin1.dll 1.5.10 it outputs: > > try mmap($0, $4, ..., MAP_ANON, ...); success, address=$833ca000 > try mmap($8340b000, $4, ..., MAP_ANON, ...); success, address=$833ca000 > > Note that the result addresses are the same. Thanks for the testcase! Wow, I'm impressed. Three serious bug reports in two days which all had a simple testcase(tm) attached! *Sob*, I'm moved to tears. Anyway, I found the cause of that problem. For some reason (moon phase or so) the mmap code didn't marked pages as used when running under 9x/Me. This could only be observed under a specific condition of mmapping anonymous private pages. I've applied a fix to the repository. Should be in the next developers snapshot from http://cygwin.com/snapshots/ Thanks for the report, Corinna -- Corinna Vinschen Please, send mails regarding Cygwin to Cygwin Co-Project Leader mailto:[EMAIL PROTECTED] Red Hat, Inc. -- Unsubscribe info: http://cygwin.com/ml/#unsubscribe-simple Problem reports: http://cygwin.com/problems.html Documentation: http://cygwin.com/docs.html FAQ: http://cygwin.com/faq/
