2003-07-03 08:14:09

by Kirill Korotaev

Subject: Again: Fix multithread coredump deadlock (patch Manfred Spraul)

Hi!

There was a patch some time ago included in linux-2.4.17-pre6 which fixed mmap
semaphore deadlock in do_coredump (double down_read() on mmap_sem).
This fix introduces down_write() on mmap_sem and uses get_user_pages()
function to avoid do_page_fault().
The question is why down_write() is used in elf_core_dump() (instead of
down_read())?

Kirill

2003-07-03 16:48:00

by Manfred Spraul

Subject: Re: Again: Fix multithread coredump deadlock (patch Manfred Spraul)

Kirill Korotaev wrote:

>Hi!
>
>There was a patch some time ago included in linux-2.4.17-pre6 which fixed mmap
>semaphore deadlock in do_coredump (double down_read() on mmap_sem).
>This fix introduces down_write() on mmap_sem and uses get_user_pages()
>function to avoid do_page_fault().
>The question is why down_write() is used in elf_core_dump() (instead of
>down_read())?
>
>
down_write is required to prevent expand_stack() from growing the stack
- expand_stack is called by the page fault handler under
down_read(&->mmap_sem) and changes vma->vm_start. A change of vm_start
between writing the program headers and the actual segment dump would
corrupt the coredump.

--
Manfred