On Thu, Jun 04, 2009 at 12:55:32AM -0700, Matt Helsley wrote: > On Thu, Jun 04, 2009 at 03:04:22AM +0400, Alexey Dobriyan wrote: > > On Sun, May 31, 2009 at 03:19:53PM -0700, Andrew Morton wrote: > > > On Mon, 1 Jun 2009 01:54:27 +0400 Alexey Dobriyan <adobriyan@xxxxxxxxx> wrote: > > > > > > > And BTW, there is something unnatural when executable path is attached > > > > to mm_struct(!) not task_struct, > > > > > > mm_struct is the central object for a heavyweight process. All threads > > > within that process share the same executable path (don't they?) so > > > attaching the executable path to the mm seems OK to me. > > > > OK, let's try this: > > > > > > [PATCH 1/9] exec_path 1/9: introduce ->exec_path and switch /proc/*/exe > > > > ->exec_path marks executable which is associated with running task. > > Binfmt loader decides which executable is such and can, in theory, > > assign anything. Unlike current status quo when first VM_EXECUTABLE mapping is > > sort of marks running executable. > > > > If executable unmaps its all VM_EXECUTABLE mappings, /proc/*/exe ceases > > to exists, ick! And userpsace can't even use MAP_EXECUTABLE. > > Suprising but intentional and unavoidable. More below.. > > > > > Tasks which aren't created by running clone(2) and execve(2) > > (read: kernel threads) get empty ->exec_path and > > > > ->exec_path is copied on clone(2) and put at do_exit() time. > > Doesn't this pin the vfs mount of the executable for the lifetime of > the task? > > That was one of Al Viro's objections to early revisions of the exe_file > patches. It's the reason the exe_file patches kept track of the number of > VM_EXECUTABLE mappings with num_exe_file_vmas. > > I've cc'd Al so he can confirm/deny my recollection of this. Basically > some programs need to be able to umount the filesystem that back their > executables. Being able to unmap these regions was effectively a > userspace API for unpinning these mounts. I needed to preserve that API, > hence the VMA ugliness of exe_file that you object to with the exe_file > patches. > > I think patches 2-7 look great and could be adapted to use exe_file instead > of ->exec_path. Well, except 5 and 6. Alternately, I think you could use the same VMA code with ->exec_path to avoid pinning the vfs mount. However, then it would probably be best to move it into the mm just like exe_file... > > Cheers, > -Matt Helsley -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html