lkml.org 
[lkml]   [2017]   [Aug]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v2 14/20] mm: Provide speculative fault infrastructure
From
Date
On 30/08/2017 07:03, Anshuman Khandual wrote:
> On 08/29/2017 07:15 PM, Peter Zijlstra wrote:
>> On Tue, Aug 29, 2017 at 03:18:25PM +0200, Laurent Dufour wrote:
>>> On 29/08/2017 14:04, Peter Zijlstra wrote:
>>>> On Tue, Aug 29, 2017 at 09:59:30AM +0200, Laurent Dufour wrote:
>>>>> On 27/08/2017 02:18, Kirill A. Shutemov wrote:
>>>>>>> +
>>>>>>> + if (unlikely(!vma->anon_vma))
>>>>>>> + goto unlock;
>>>>>>
>>>>>> It deserves a comment.
>>>>>
>>>>> You're right I'll add it in the next version.
>>>>> For the record, the root cause is that __anon_vma_prepare() requires the
>>>>> mmap_sem to be held because vm_next and vm_prev must be safe.
>>>>
>>>> But should that test not be:
>>>>
>>>> if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma))
>>>> goto unlock;
>>>>
>>>> Because !anon vmas will never have ->anon_vma set and you don't want to
>>>> exclude those.
>>>
>>> Yes in the case we later allow non anonymous vmas to be handled.
>>> Currently only anonymous vmas are supported so the check is good enough,
>>> isn't it ?
>>
>> That wasn't at all clear from reading the code. This makes it clear
>> ->anon_vma is only ever looked at for anonymous.
>>
>> And like Kirill says, we _really_ should start allowing some (if not
>> all) vm_ops. Large file based mappings aren't particularly rare.
>>
>> I'm not sure we want to introduce a white-list or just bite the bullet
>> and audit all ->fault() implementations. But either works and isn't
>> terribly difficult, auditing all is more work though.
>
> filemap_fault() is used as vma-vm_ops->fault() for most of the file
> systems. Changing it can enable speculative fault support for all of
> them. It will still exclude other driver based vma-vm_ops->fault()
> implementation. AFAICS, __lock_page_or_retry() function can drop
> mm->mmap_sem if the page could not be locked right away. As suggested
> by Peterz, making it understand FAULT_FLAG_SPECULATIVE should be good
> enough. The patch is lightly tested for file mappings on top of this
> series.

Hi Anshuman,

This sounds pretty good, except for the FAULT_FLAG_RETRY_NOWAIT's case I
mentioned in another mail.

The next step would be to find a way to discriminate between the vm_fault()
functions. Any idea ?

Thanks,
Laurent.

>
> diff --git a/mm/filemap.c b/mm/filemap.c
> index a497024..08f3042 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -1181,6 +1181,18 @@ int __lock_page_killable(struct page *__page)
> int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
> unsigned int flags)
> {
> + if (flags & FAULT_FLAG_SPECULATIVE) {
> + if (flags & FAULT_FLAG_KILLABLE) {
> + int ret;
> +
> + ret = __lock_page_killable(page);
> + if (ret)
> + return 0;
> + } else
> + __lock_page(page);
> + return 1;
> + }
> +
> if (flags & FAULT_FLAG_ALLOW_RETRY) {
> /*
> * CAUTION! In this case, mmap_sem is not released
> diff --git a/mm/memory.c b/mm/memory.c
> index 549d235..02347f3 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -3836,8 +3836,6 @@ static int handle_pte_fault(struct vm_fault *vmf)
> if (!vmf->pte) {
> if (vma_is_anonymous(vmf->vma))
> return do_anonymous_page(vmf);
> - else if (vmf->flags & FAULT_FLAG_SPECULATIVE)
> - return VM_FAULT_RETRY;
> else
> return do_fault(vmf);
> }
> @@ -4012,17 +4010,7 @@ int handle_speculative_fault(struct mm_struct *mm, unsigned long address,
> goto unlock;
> }
>
> - /*
> - * Can't call vm_ops service has we don't know what they would do
> - * with the VMA.
> - * This include huge page from hugetlbfs.
> - */
> - if (vma->vm_ops) {
> - trace_spf_vma_notsup(_RET_IP_, vma, address);
> - goto unlock;
> - }
> -
> - if (unlikely(!vma->anon_vma)) {
> + if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
> trace_spf_vma_notsup(_RET_IP_, vma, address);
> goto unlock;
> }
>

\
 
 \ /
  Last update: 2017-08-30 11:54    [W:0.879 / U:0.056 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site