I've been debugging why an n64 emulator ( Project64 ) gets "Failed to allocate memory" errors under wine 9.9.10 on linux kernel 2.6.
The program allocates a large chunk of size 0x20100000, which succeeds, and then another large 0x10100000 chunk, which fails.
WINEDEBUG=+relay,+virtual eventually brought me to look at the linux /proc/<pid>/maps to see if the address space was really fully allocated. Here's what I noticed:
Address start - end: 00400000-0554000 - Project64.exe address 00554000-20000000 ---p large mem allocation *** <gap> *** 55555000-573e0000 - loaded shared libraries 573e0000-774e0000 ---p mem alloc of 0x20100000 7bf00000-7c003000 - wine-pthread 7c003000-7c1ee000 - [heap] 7ffe0000-fffe0000 ---p large mem allocation ffff7000-fffff000 - [stack]
Wrote a test mmap binary to fill a process address space, it created a similar gap between the executable load address and the loaded shared libraries, eventually failing mmap calls when it hit the stack address space. When I updated the test app to explicitly mmap in the gap area via specifying the aligned start address, the mmap call succeeded.
I can't be the first person to notice this. Are there any TODO solutions for this ?
One linux-centric solution would be to look for gaps like this in /proc/<pid>/maps in the libs/wine/mmap.c and take advantage of this. Or maybe another would be to keep a table of address spaces returned in wine_anon_mmap as a guide, then validate with try_mmap_fixed as to where a next-allocation should occur ( would this solution work outside linux ) ?
Thanks, geoff Brimhall
__________________________________________________ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com
Geoff Brimhall wrote:
Wrote a test mmap binary to fill a process address space, it created a similar gap between the executable load address and the loaded shared libraries, eventually failing mmap calls when it hit the stack address space. When I updated the test app to explicitly mmap in the gap area via specifying the aligned start address, the mmap call succeeded.
I can't be the first person to notice this. Are there any TODO solutions for this ?
Wine's holey memory map has been discussed before:
http://www.winehq.com/pipermail/wine-devel/2006-February/044979.html
The "truely correct" solution is probably to implement Windows compatible memory allocation in the Linux kernel, but there's other hacks that could be done in user space. Just a matter of finding something that Alexandre is willing to live with.
A preload-ish type thing that overrode libc's mmap/munmap calls seems pausible to me.
Mike
On Wed, Mar 22, 2006 at 12:11:27AM -0800, Geoff Brimhall wrote:
I've been debugging why an n64 emulator ( Project64 ) gets "Failed to allocate memory" errors under wine 9.9.10 on linux kernel 2.6.
The program allocates a large chunk of size 0x20100000, which succeeds, and then another large 0x10100000 chunk, which fails.
Do you have ulimits set?
I can't be the first person to notice this. Are there any TODO solutions for this ?
No.
Ciao, Marcus
-
Geoff Brimhall -
Marcus Meissner -
Mike McCormack