Messages in this thread | | | Date | Mon, 26 Jan 2009 21:09:57 +0100 | From | Ingo Molnar <> | Subject | Re: [git pull] cpus4096 tree, part 3 |
| |
* Andrew Morton <akpm@linux-foundation.org> wrote:
> but why did the current code pass testing at all??
i queued it up a week ago and beyond a same-day breakage i reported to Nick (and which he fixed) this commit was problem-free and passed all testing here.
Does it cause problems for you? If yes then please describe the kind of problems.
Note: i see that -mm modifies a few other details of the x86 pagefault handling path (there a pagefault-retry patch in there) - so there might be contextual interactions there. But this particular cleanup/improvement from Nick is working fine on a wide range of systems here.
Btw., regarding pagefault retry. The bits that are in -mm currently i find a bit ugly:
> +++ a/arch/x86/mm/fault.c > @@ -799,7 +799,7 @@ void __kprobes do_page_fault(struct pt_r > struct vm_area_struct *vma; > int write; > int fault; > - unsigned int retry_flag = FAULT_FLAG_RETRY; > + int retry_flag = 1; > > tsk = current; > mm = tsk->mm; > @@ -951,6 +951,7 @@ good_area: > } > > write |= retry_flag; > + > /* > * If for any reason at all we couldn't handle the fault, > * make sure we exit gracefully rather than endlessly redo > @@ -969,8 +970,8 @@ good_area: > * be removed or changed after the retry. > */ > if (fault & VM_FAULT_RETRY) { > - if (write & FAULT_FLAG_RETRY) { > - retry_flag &= ~FAULT_FLAG_RETRY; > + if (retry_flag) { > + retry_flag = 0; > goto retry; > } > BUG();
as this complicates every architecture with a 'can the fault be retried' logic and open-coded retry loop.
But that logic is rather repetitive and once an architecture filters out all its special in-kernel sources of faults and the hw quirks it has, the handling of pte faults is rather generic and largely offloaded into handle_pte_fault() already.
So when this patch was submitted a few weeks ago i suggested that retry should be done purely in mm/memory.c instead, and the low level code should at most be refactored to suit this method, but not complicated any further.
Any deep reasons for why such a more generic approach is not desirable?
Ingo
| |