Skip to contentSkip to conversation

COW optimization on exec

DISCUSSION sidc7 3 participants 5 messages Mar 15, 2009

lore
Share & tools
lore

sidc7

author
#0 lore raw

The Linux kernel uses the COW optimization for fork, so the processes share the same pages, till on of the processes writes to the page. I was wondering, if I do a fork and do an exec immediately following the fork, will the COW optimization still be applied as it is most likely that the new process is going to write to the shared pages? So doing a COW will not give much benefit here, if it is done at all. Can anyone clarify if COW will be applied in such a case, for e.g. a command shell.

Signature · 2 lines
View this message in context: http://www.nabble.com/COW-optimization-on-exec-tp22519639p22519639.html
Sent from the linux-kernel mailing list archive at Nabble.com.
#1 lore raw
Quoted 8 lines · On Sat, Mar 14, 2009 at 07:57:54PM -0700, sidc7 wrote:

The Linux kernel uses the COW optimization for fork, so the processes share the same pages, till on of the processes writes to the page. I was wondering, if I do a fork and do an exec immediately following the fork, will the COW optimization still be applied as it is most likely that the new process is going to write to the shared pages? So doing a COW will not give much benefit here, if it is done at all. Can anyone clarify if COW will be applied in such a case, for e.g. a command shell.

COWing the pages is not much extrawork, it's handled with this code:

        /*
         * If it's a COW mapping, write protect it both
         * in the parent and the child
         */
        if (is_cow_mapping(vm_flags)) {
                ptep_set_wrprotect(src_mm, addr, src_pte);
                pte = pte_wrprotect(pte);
        }

you can find it in mm/memory.c::copy_one_pte(). The fault handler will then take care of it (it will notice that the pte is write-protected while the mapping itself allows writes) and then replace the page with a copy in the faulting process.

The real overhead is copying the page tables in the first place. But the kernel can not know whether exec() is soon to be called, so fork() always must provide the copy-whole-address-space semantics.

If the forking process knows in advance the child will exec immediately, it can use vfork() which doesn't copy the address space.

Hannes

author
#2 lore raw

Thanks for clearing this

Quoted 46 lines · Johannes Weiner-5 wrote:
Quoted 12 lines · On Sat, Mar 14, 2009 at 07:57:54PM -0700, sidc7 wrote:

The Linux kernel uses the COW optimization for fork, so the processes share
the same pages, till on of the processes writes to the page. I was wondering, if I do a fork and do an exec immediately following the fork, will the COW optimization still be applied as it is most likely that the new
process is going to write to the shared pages? So doing a COW will not give
much benefit here, if it is done at all. Can anyone clarify if COW will be
applied in such a case, for e.g. a command shell.

COWing the pages is not much extrawork, it's handled with this code:

        /*
         * If it's a COW mapping, write protect it both
         * in the parent and the child
         */
        if (is_cow_mapping(vm_flags)) {
                ptep_set_wrprotect(src_mm, addr, src_pte);
                pte = pte_wrprotect(pte);
        }

you can find it in mm/memory.c::copy_one_pte(). The fault handler will then take care of it (it will notice that the pte is write-protected while the mapping itself allows writes) and then replace the page with a copy in the faulting process.

The real overhead is copying the page tables in the first place. But the kernel can not know whether exec() is soon to be called, so fork() always must provide the copy-whole-address-space semantics.

If the forking process knows in advance the child will exec immediately, it can use vfork() which doesn't copy the address space.

Hannes
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Signature · 2 lines
View this message in context: http://www.nabble.com/COW-optimization-on-exec-tp22519639p22519869.html
Sent from the linux-kernel mailing list archive at Nabble.com.
#3 lore raw
In reply to sidc7
Quoted 5 lines · On Sunday 15 March 2009 13:57:54 sidc7 wrote:

The Linux kernel uses the COW optimization for fork, so the processes share the same pages, till on of the processes writes to the page. I was wondering, if I do a fork and do an exec immediately following the fork, will the COW optimization still be applied as it is most likely that the new process is going to write to the shared pages? So doing a COW will not

exec throws out all those pages, so go away without ever having been copied. However it still costs memory CPU time TLB flushes etc in order to set up the COW optimisation in the first place, so if you are doing fork/exec, then I think vfork is the better option.

author
#4 lore raw
Quoted 5 lines

exec throws out all those pages, so go away without ever having been copied. However it still costs memory CPU time TLB flushes etc in order to set up the COW optimisation in the first place, so if you are doing fork/exec, then I think vfork is the better option.

Thanks once again, if I do something like creating an array spanning over multiple pages, and then do a fork and in the child or the parent, write to the array elements, this should result in an actual cow, where the OS will actually do a copy of the pages as well in addition to the overhead required for setting up COW in the first place?

Thanks

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Signature · 2 lines
View this message in context: http://www.nabble.com/COW-optimization-on-exec-tp22519639p22520422.html
Sent from the linux-kernel mailing list archive at Nabble.com.