Re: [RFC v1][PATCH]page_fault retry with NOPAGE_RETRY

[Ingo Molnar]

* Ying Han <yinghan@xxxxxxxxxx> wrote:

> page fault retry with NOPAGE_RETRY

Interesting patch.

> Allow major faults to drop the mmap_sem read lock while waitting for
> synchronous disk read. This allows another thread which wishes to grab
> down_read(mmap_sem) to proceed while the current is waitting the disk IO.

Do you mean down_write()? down_read() can already be nested 
arbitrarily.

> The patch flags current->flags to PF_FAULT_MAYRETRY as identify that 
> the caller can tolerate the retry in the filemap_fault call patch.
> 
> Benchmark is done by mmap in huge file and spaw 64 thread each 
> faulting in pages in reverse order, the the result shows 8% 
> porformance hit with the patch.

I suspect we also want to see the cases where this change helps?

Also, constructs like this are pretty ugly:

> +#ifdef CONFIG_X86_64
> +asmlinkage
> +#endif
> +void do_page_fault(struct pt_regs *regs, unsigned long error_code)
> +{
> +     current->flags |= PF_FAULT_MAYRETRY;
> +     __do_page_fault(regs, error_code);
> +     current->flags &= ~PF_FAULT_MAYRETRY;
> +}

This seems to be unnecessary runtime overhead to pass in a flag to 
handle_mm_fault(). Why not extend the 'write' flag of 
handle_mm_fault() to also signal "arch is able to retry"?

Also, _if_ we decide that from-scratch pagefault retries are good, i 
see no reason why this should not be extended to all architectures:

The retry should happen purely in the MM layer - all information is 
available already, and much of do_page_fault() could generally be 
moved into mm/memory.c, with one or two arch-provided standard 
callbacks to express certain page fault quirks. (such as vm86 mode on 
x86)

(Such a design would allow more nice cleanups - handle_mm_fault() 
could inline inside the pagefault handler, etc.)

Also, a few small details. Please use this proper multi-line comment 
style:

> +			/*
> +			 * Page is already locked by someone else.
> +			 *
> +			 * We don't want to be holding down_read(mmap_sem)
> +			 * inside lock_page(). We use wait_on_page_lock here
> +			 * to just wait until the page is unlocked, but we
> +			 * don't really need
> +			 * to lock it.
> +			 */

Not this one:

> +	/* page may be available, but we have to restart the process
> +	 * because mmap_sem was dropped during the ->fault */

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