Skip to contentSkip to conversation

control placement of vDSO page

DISCUSSION John Reiser 2 participants 3 messages Dec 12, 2005

lore
Share & tools
lore
#0 lore raw

The vDSO page is getting in the way of applications on i686 (i386). There is a performance problem, a problem with a dangling pointer in the kernel, and a loss of control in the kernel<->user interface. A possible solution might be a new system call to move the vDSO page.

Performance: the kernel chooses the vDSO page after mapping the main program and the PT_INTERP [see fs/binfmt_elf.c]. If pre-linking chose that same page for some subsequent shared library, then the pre-linking is discarded, which costs time. In Fedora Core 4 i686 with random vDSO placement, the probability is about 7% that a main program with just one shared library must relocate libc.so.6. Cascading conflicts and costs are likely; a program with dozens of shared libs is almost certain to require expensive runtime relocation processing of shared libraries.

Dangling pointer: setup_frame() in arch/i386/kernel/signal.c uses

        restorer = current->mm->context.vdso + (long)&__kernel_sigreturn;

whenever !(.sa_flags & SA_RESTORER). Unfortunately, context.vdso is never changed, so if the user re-maps the vDSO page then the kernel has a dangling pointer which makes user return from signal unreliable.

Loss of control: in a 32-bit address there are at most one million 4KB pages. Maintaining contiguous ranges for large arrarys is a problem. Having no control over placement of the vDSO page makes things worse. Programs that audit or measure the running of other programs in the same address space, desire to place the vDSO page after mapping the programs that are being examined.

Possible solution: a new system call move_vdso(old_addr, new_addr, flags). Check that current vDSO was at old_addr, else error. Change vDSO page under control of flags like in mmap(): new_addr is hint (place to start search) or required address if MAP_FIXED. Return value is new vDSO address.

For some history and examples see
https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=162797

Comments?

Signature · 1 line
John Reiser, [email protected]
#1 lore raw
Quoted 5 lines · John Reiser

Possible solution: a new system call move_vdso(old_addr, new_addr, flags). Check that current vDSO was at old_addr, else error. Change vDSO page under control of flags like in mmap(): new_addr is hint (place to start search) or required address if MAP_FIXED. Return value is new vDSO address.

What about special casing the vDSO page in mremap() ?

Regards

Ingo Oeser

#2 lore raw
Quoted 10 lines · Ingo Oeser
Quoted 5 lines · John Reiser

Possible solution: a new system call move_vdso(old_addr, new_addr, flags). Check that current vDSO was at old_addr, else error. Change vDSO page under control of flags like in mmap(): new_addr is hint (place to start search) or required address if MAP_FIXED. Return value is new vDSO address.

What about special casing the vDSO page in mremap() ?

Hmmm... Where can the new address [hint] be given in the call to mremap(old_address, old_size, new_size, flags) ?
Also: probably the size is constant, perhaps even unknown to the user. It would be nice if the old page need not exist (could have been mmap()ed over) at the time of the move, although the proposed move_vdso() does ask for what the old_addr used to be, as a check on validity.

Signature · 1 line
John Reiser, [email protected]