On Wednesday 04 May 2016 07:17 PM, Peter Zijlstra wrote:
> On Wed, May 04, 2016 at 04:07:40PM +0530, Vineet Gupta wrote:
>> Is preemption disabling a requirement of kmap_atomic() callers independent of
>> where page is or is it only needed when page is in highmem and can trigger page
>> faults or TLB Misses between kmap_atomic() and kunmap_atomic and wants protection
>> against reschedules etc.
> Traditionally kmap_atomic() disables preemption; and the reason is that
> the returned pointer must stay valid. This had a side effect in that it
> also disabled pagefaults.
But how could the ptr possibly get invalid. Say despite the disable calls, we
could actually take the page fault (or TLB Miss on ARC) - the pagefault_disable()
only makes do_page_fault() do reduced handling vs. calling handle_mm_fault() etc.
It is essentially restricting the fault handling to a kernel mode fixup only.
Now if we didn't do disable, on ARC the semantics of do_page_fault() are still the
same - since the address would be for fixmap which is handled under "kernel" only
category as well.
void do_page_fault(unsigned long address, struct pt_regs *regs)
{
if (address >= VMALLOC_START) {
ret = handle_kernel_vaddr_fault(address);
return;
...
if (faulthandler_disabled() || !mm)
goto no_context;
...
> We've since de-coupled the pagefault from the preemption thing, so you
> could disable pagefaults while leaving preemption enabled.
Right - I've seen that patch set from David H.
> ...
>
> If you want a fast-slow path splt, you can easily do something like:
>
> static inline void *kmap_atomic(struct page *page)
> {
> preempt_disable();
> pagefault_disable();
> if (!PageHighMem(page))
> return page_address(page);
>
> return __kmap_atomic(page);
> }
I actually want to return early for !PageHighMem and avoid the pointless 2
LD-ADD-ST to memory for map and 2 LD-SUB-ST for unmap for regular pages for such
cases.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href