> From: Jeff Trawick [mailto:[EMAIL PROTECTED]
> Sent: Thursday, February 27, 2003 2:41 PM
> Sander Striker wrote:
[...]
>> rv = apr_mmap_delete(mm);
>>
>> [failure, rv == 22 (EINVAL?)]
>>
>> My guess is that munmap doesn't like the 0 passed in for size. It
>> strikes me as odd though that mmap doesn't seem to care about that.
>>
>> Is there a known issue with mmapping 0 sized files?
>>
>> FYI, I'm on Linux 2.4.20, glibc 2.3.1.
>
> does strace show that it is kernel behavior in question rather than
> something questionable that APR may be doing?
Something questionable at both ends ;)
kernel returns NULL on an mmap of a 0 sized file (even though the mmap
manpage tells us that'll never happen):
mmap2(NULL, 0, PROT_READ, MAP_SHARED, 5, 0) = 0
apr calls this on apr_mmap_delete (because it stored the NULL in apr_mmap_t.mm):
munmap(0, 0) = -1 EINVAL (Invalid argument)
This patch seems like a working fix:
Index: mmap/unix/mmap.c
===================================================================
RCS file: /home/cvs/apr/mmap/unix/mmap.c,v
retrieving revision 1.47
diff -u -r1.47 mmap.c
--- mmap/unix/mmap.c 7 Jan 2003 00:52:55 -0000 1.47
+++ mmap/unix/mmap.c 27 Feb 2003 15:12:59 -0000
@@ -99,7 +99,10 @@
#ifdef BEOS
rv = delete_area(mm->area);
#else
- rv = munmap(mm->mm, mm->size);
+ if (mm->mm != NULL)
+ rv = munmap(mm->mm, mm->size);
+ else
+ rv = APR_SUCCESS;
#endif
mm->mm = (void *)-1;
Someone with some more mmap-fu, please review.
Thanks,
Sander
