On 6/20/24 06:48, Kundan Kumar wrote:
On 19/06/24 09:47AM, Hannes Reinecke wrote:
[ .. ]
Well. The issue here is that bvecs really only use the 'struct page'
entry as an address to the data; the page size itself is completely
immaterial. So from that perspective it doesn't really matter whether
we use 'struct page' or 'struct folio' to get to that address.
However, what matters is whether we _iterate_ over pages or folios.
The current workflow is to first allocate an array of pages,
call one of the _get_pages() variants, and the iterate over all
pages.
What we should be doing is to add _get_folios() variants, working
on folio batches and not pre-allocated arrays.
The XXX_get_pages() functions do page table walk and fill the pages
corresponding to a user space addr in pages array. The _get_folios()
variants shall return a folio_vec, rather than folio array, as every folio
entry will need a folio_offset and len per folio. If we convert to
_get_folios() variants, number of folios may be lesser than number of
pages. But we will need allocation of folio_vec array as a replacement
of pages array.
Well, actually I was thinking of using 'struct folio_batch' instead of
an array. There is an entire set of helpers here (pagevec.h) for
precisely this purpose.
But yes, we will end up with less folios than pages (which was kinda the
idea).
Am I missing something?
Down in the page table walk (gup_fast_pte_range), we fill the pages array
addr -> pte -> page. This shall be modified to fill a folio_vec array. The
page table walk also deals with huge pages, and looks like huge page
functions shall also be modified to fill folio_vec array. Also handling
the gup slow path will need a modification to fill the folio_vec array.
Yes. I did say it's a tall order. But it would have the advantage of
assembling large folios right at the start, so lower level (ie those
consuming the folio batch) would not need to worry of painstakingly
re-assemble folios from a list of pages.
Cheers,
Hannes
--
Dr. Hannes Reinecke Kernel Storage Architect
hare@xxxxxxx +49 911 74053 688
SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
