Revision 209887

[Alan Cox]

Mark Tinguely wrote:
> Alan Cox wrote:
>> Mark Tinguely wrote:
>>> pmap_qenter() is pretty expensive on the ARM processor with VIPT 
>>> caches, because remapping a KVA can potentially cause the shared 
>>> mappings to have their cache turned back on and then immediately 
>>> turned back off and force a cache flush. Revision 209887 makes 
>>> amd64/i386 pmap_qenter() make smart decisions on the remapping which 
>>> would help the ARM.
>>>
>>> Looking at the buffer code 2 or 3 weeks ago, and  noticed that most 
>>> of the pmap_qenter() calls were done wisely. At that time, I was 
>>> wondering if we could/should fix the few places that involved a page 
>>> by page loop doing bogus_page substitutions and then calling 
>>> pmap_qenter() for a full buffer re-map instead of the changed page. 
>>> It appeared to me that we call pmap_qenter() L*B times where L is 
>>> the number of pages in the buffer and B is the number of bogus pages 
>>> in the buffer. We only need to change B pages. I can understand it 
>>> would not be worth the change if typically the number of pages in a 
>>> buffer and the number of bogus pages in the buffer is small.
>>
>> I think that vfs_unbusy_pages(), which is used exclusively by NFS, is 
>> the only remaining case.  I see no reason not to change it.  Go for it.
>>
>> Alan
> Gee thanks :)
>
> I will do a complete search of pmap_qenter(). Unless I am reading it 
> wrong, a quick example is a B_VMIO case in routine bufdone_finish() in 
> source file sys/kern/vfs_bio.c:
>

I thought that I already fixed bufdone_finish() to do at most one call 
to pmap_qenter() in HEAD..., but vfs_unbusy_pages() still looks like 
what you describe below.

>     for (i = 0; i < bp->b_npages; i++) {
>        if (bogus_page) {
>            find/substitution page
>             pmap_qenter() the entire buffer rather than just the 
> single page
>        }
>       other page operations
>    }
>

Alan