Odd IA32 mmap calls - Is this fixed
Hi,
The mummap is failing to unmap the pages allocated by mmap when a ia32 application ( 4k Pagesize) is run on ia64 system( 16k pagesize). This happens when the offset in mmap does not fall on 16k boundary. The problem seems to be similar to the problem " Odd IA32 mmap calls" (Calls to mmap/munmap where the length is a fraction of the page size cause a memory leak, eventually running out of VM for the IA32 process) listed in the ToDo list of ia64 homepage.I am doing the test on sles-8 running 2.4.19 kernel. Is this problem fixed in the 2.6 test kernel ? If it is not fixed are there anyone working on it to fix the problem?
Thanks & Regards,
Shiju
________________________________________________________________________
Want to chat instantly with your online friends? Get the FREE Yahoo! Messenger http://mail.messenger.yahoo.co.uk
There is no "safe" solution that I know of. The attached patch will fix this particular problem, but might create new ones. Try it at your own risk.
-Arun
Quoted 6 lines · Shiju A Mathew
Hi,
The mummap is failing to unmap the pages allocated by mmap when a ia32 application ( 4k Pagesize) is run on ia64 system( 16k pagesize). This happens when the offset in mmap does not fall on 16k boundary. The problem seems to be similar to the problem " Odd IA32
Quoted 8 lines
mmap calls" (Calls to mmap/munmap where the length is a fraction of the page size cause a memory leak, eventually running out of VM for the IA32 process) listed in the ToDo list of ia64 homepage.I am doing the test on sles-8 running 2.4.19 kernel. Is this problem fixed in the 2.6 test kernel ? If it is not fixed are there anyone working on it to fix the problem?
arch/ia64/ia32/sys_ia32.c +2 -2
@@ -549,8 +559,8 @@if (start > end)return -EINVAL ; - start =PAGE_ALIGN (start); - end =PAGE_START (end); + start =PAGE_START (start); + end =PAGE_ALIGN (end);if (start >= end)return 0 ;
Arun-
I'm pretty sure your patch with cause things to break. You will wind up freeing memory that the IA32 process is still using, causing bad things to happen.
Actually, there is a solution but it's kind of icky. The idea is you have to maintain knowledge of which 4K area of the IA32 prcesses's IA64 page is truly in use. When all 4K areas of the IA64 page are freed you can free the IA64 page.
This not as bad as it could be because you only have to keep track of the first & last IA64 pages that are used for each IA32 mmap request so the overhead isn't that outrageous. I submitted a patch to do this ages ago but I guess it got lost somehow.
Quoted 35 lines · Arun Sharma
There is no "safe" solution that I know of. The attached patch will fix this particular problem, but might create new ones. Try it at your own risk.
-Arun
Quoted 14 lines · Shiju A Mathew
Hi,
The mummap is failing to unmap the pages allocated by mmap when a ia32 application ( 4k Pagesize) is run on ia64 system( 16k pagesize). This happens when the offset in mmap does not fall on 16k boundary. The problem seems to be similar to the problem " Odd IA32 mmap calls" (Calls to mmap/munmap where the length is a fraction of the page size cause a memory leak, eventually running out of VM for the IA32 process) listed in the ToDo list of ia64 homepage.I am doing the test on sles-8 running 2.4.19 kernel. Is this problem fixed in the 2.6 test kernel ? If it is not fixed are there anyone working on it to fix the problem?arch/ia64/ia32/sys_ia32.c +2
-2@@ -549,8 +559,8 @@if (start > end)return -EINVAL ; - start =PAGE_ALIGN (start); - end =PAGE_START (end); + start =PAGE_START (start); + end =PAGE_ALIGN (end);if (start >= end)return 0 ;
Signature · 4 lines
Don Dugger "Censeo Toto nos in Kansa esse decisse." - D. Gale [email protected] Ph: 303/447-2201
Quoted 6 lines · On 11/12/2003 11:45 AM, [email protected] wrote:
Arun-
I'm pretty sure your patch with cause things to break. You will wind up freeing memory that the IA32 process is still using, causing
badthings to happen.
Yes, of course. May be the disclaimer wasn't prominent enough :)
Quoted 10 lines
Actually, there is a solution but it's kind of icky. The idea is you have to maintain knowledge of which 4K area of the IA32 prcesses's IA64 page is truly in use. When all 4K areas of the IA64 page are freed you can free the IA64 page.
This not as bad as it could be because you only have to keep track of the first & last IA64 pages that are used for each IA32 mmap request so the overhead isn't that outrageous. I submitted a patch to do this ages ago but I guess it got lost somehow.
Is this the patch you're talking about ?
http://www.gelato.unsw.edu.au/linux-ia64/0203/3099.html
-Arun
Arun-
Yep, that's the one. Looks like I dropped ball on that one and never got around to making the suggested changes. Looks like you just inherited the problem :-)
Quoted 8 lines · Arun Sharma
...
Is this the patch you're talking about ?
http://www.gelato.unsw.edu.au/linux-ia64/0203/3099.html
-Arun
Signature · 4 lines
Don Dugger "Censeo Toto nos in Kansa esse decisse." - D. Gale [email protected] Ph: 303/447-2201