[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <7b69abe5-3782-965c-ec82-5baef84e2d06@126.com>
Date: Mon, 17 Jun 2024 17:50:27 +0800
From: yangge1116 <yangge1116@....com>
To: David Hildenbrand <david@...hat.com>, Matthew Wilcox <willy@...radead.org>
Cc: Baolin Wang <baolin.wang@...ux.alibaba.com>, akpm@...ux-foundation.org,
linux-mm@...ck.org, linux-kernel@...r.kernel.org, liuzixing@...on.cn
Subject: Re: [PATCH] mm/gup: don't check page lru flag before draining it
在 2024/6/12 下午3:32, David Hildenbrand 写道:
> On 11.06.24 13:20, yangge1116 wrote:
>>
>>
>> 在 2024/6/9 上午12:03, David Hildenbrand 写道:
>>> On 08.06.24 17:15, Matthew Wilcox wrote:
>>>> On Sat, Jun 08, 2024 at 12:38:49PM +0800, yangge1116 wrote:
>>>>> Can we add a PG_lru_batch flag to determine whether a page is in lru
>>>>> batch?
>>>>> If we can, seems this problem will be easier.
>>>>
>>>> Page flags are in short supply. You'd need a really good
>>>> justification.
>>>>
>>>
>>> A flag would not be able to handle the "part of multiple LRU batches"
>>> that should currently possible (when to clear the flag?). Well, if we
>>> have to keep supporting that. If we only to be part in a single LRU
>>> batch, a new flag could work and we could still allow isolating a folio
>>> from LRU while in some LRU batch.
>>
>> Yes, before adding a folio to LRU batch, check whether the folio has
>> been added. Add the folio to LRU batch only if the folio has not been
>> added.
>>
>>>
>>> If we could handle it using the existing flags, that would of course be
>>> better (wondering if we could store more information in the existing
>>> flags by using a different encoding for the different states).
>>
>> If a folio contains more than one page, the folio will not be added to
>> LRU batch. Can we use folio_test_large(folio) to filter?
>>
>> if (!folio_test_large(folio) && drain_allow) {
>> lru_add_drain_all();
>> drain_allow = false;
>> }
>
> I think we should do better than this, and not do arbitrary
> lru_add_drain_all() calls.
>
Thanks, I've got another idea.
If we add GUP_PIN_COUNTING_BIAS to folio's ref count before adding to
LRU batch, we can use folio_maybe_dma_pinned(folio) to check whether the
folio is in LRU batch. I wonder if it's feasible?
static void folio_batch_add_and_move(struct folio_batch *fbatch,
struct folio *folio, move_fn_t move_fn)
{
if (!folio_test_large(folio)) {
folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
if (folio_batch_add(fbatch, folio) && !lru_cache_disabled())
return;
}
folio_batch_move_lru(fbatch, move_fn);
}
if (!folio_test_large(folio) && folio_maybe_dma_pinned(folio) &&
drain_allow) {
lru_add_drain_all();
drain_allow = false;
}
Powered by blists - more mailing lists