On Fri, Jun 2, 2023 at 5:23 AM Yunsheng Lin <[email protected]> wrote:
>
> On 2023/6/2 2:14, Alexander Duyck wrote:
> > On Wed, May 31, 2023 at 5:19 AM Yunsheng Lin <[email protected]> wrote:
>
> ...
>
> >>
> >>>
> >>> If we have to have both version I would much rather just have some
> >>> inline calls in the header wrapped in one #ifdef for
> >>> PAGE_POOL_DMA_USE_PP_FRAG_COUNT that basically are a wrapper for
> >>> page_pool pages treated as pp_frag.
> >>
> >> Do you have a good name in mind for that wrapper.
> >> In addition to the naming, which API should I use when I am a driver
> >> author wanting to add page pool support?
> >
> > When I usually have to deal with these sort of things I just rename
> > the original with a leading underscore or two and then just name the
> > inline the same as the original function.
>
> Ok, will follow the pattern if it is really necessary.
>
> >
> >>>
> >>>> size = ALIGN(size, dma_get_cache_alignment());
> >>>> - *offset = pool->frag_offset;
> >>>>
> >>>
> >>> If we are going to be allocating mono-frag pages they should be
> >>> allocated here based on the size check. That way we aren't discrupting
> >>> the performance for the smaller fragments and the code below could
> >>> function undisturbed.
> >>
> >> It is to allow possible optimization as below.
> >
> > What optimization? From what I can tell you are taking extra steps for
> > non-page pool pages.
>
> I will talk about the optimization later.
>
> According to my defination in this patchset:
> frag page: page alloced from page_pool_alloc_frag() with page->pp_frag_count
> being greater than one.
> non-frag page:page alloced return from both page_pool_alloc_frag() and
> page_pool_alloc_pages() with page->pp_frag_count being one.
>
> I assume the above 'non-page pool pages' refer to what I call as 'non-frag
> page' alloced return from both page_pool_alloc_frag(), right? And it is
> still about doing the (size << 1 > max_size)' checking at the begin instead
> of at the middle right now to avoid extra steps for 'non-frag page' case?
Yeah, the non-page I was referring to were you mono-frag pages.
> >
> >>>
> >>>> - if (page && *offset + size > max_size) {
> >>>> + if (page) {
> >>>> + *offset = pool->frag_offset;
> >>>> +
> >>>> + if (*offset + size <= max_size) {
> >>>> + pool->frag_users++;
> >>>> + pool->frag_offset = *offset + size;
> >>>> + alloc_stat_inc(pool, fast);
> >>>> + return page;
> >>
> >> Note that we still allow frag page here when '(size << 1 > max_size)'.
>
> This is the optimization I was taking about: suppose we start
> from a clean state with 64K page size, if page_pool_alloc_frag()
> is called with size being 2K and then 34K, we only need one page
> to satisfy caller's need as we do the '*offset + size > max_size'
> checking before the '(size << 1 > max_size)' checking.
The issue is the unaccounted for waste. We are supposed to know the
general size of the frags being used so we can compute truesize. If
for example you are using an order 3 page and you are splitting it
between a 2K and a 17K fragment the 2K fragments will have a massive
truesize underestimate that can lead to memory issues if those smaller
fragments end up holding onto the pages.
As such we should try to keep the small fragments away from anything
larger than half of the page.
> As you mentioned below, it is at the cost of evicting the previously
> fragmented page, I thought about keeping it when implementing, but I
> am not sure evicting it is really matter if the previously fragmented
> page does not pass the testing by '*offset + size > max_size'?
>
> Or maybe we should adjust the code a litte bit as below to allow the
> optimization I mentioned without the cost of evicting the previously
> fragmented page?
Really we should just not be mixing smaller fragments w/ larger ones
for the above truesize reasons.
> struct page *page_pool_alloc_frag(struct page_pool *pool,
> unsigned int *offset,
> unsigned int size, gfp_t gfp)
> {
> unsigned int max_size = PAGE_SIZE << pool->p.order;
> struct page *page = pool->frag_page;
>
> if (unlikely(size > max_size))
> return NULL;
>
> if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT)
> goto alloc_non_frag;
>
> size = ALIGN(size, dma_get_cache_alignment());
>
> if (page && pool->frag_offset <= max_size) {
> *offset = pool->frag_offset;
> pool->frag_users++;
> pool->frag_offset += size;
> alloc_stat_inc(pool, fast);
> return page;
> }
>
> if (unlikely((size << 1) > max_size))
> goto alloc_non_frag;
>
> if (page) {
> page = page_pool_drain_frag(pool, page);
> if (page) {
> alloc_stat_inc(pool, fast);
> goto frag_reset;
> }
> }
>
> page = page_pool_alloc_pages(pool, gfp);
> if (unlikely(!page))
> return NULL;
>
> pool->frag_page = page;
> frag_reset:
> pool->frag_users = 1;
> *offset = 0;
> pool->frag_offset = size;
> page_pool_fragment_page(page, BIAS_MAX);
> return page;
>
> alloc_non_frag:
> *offset = 0;
> return page_pool_alloc_pages(pool, gfp);
> }
>
> >
> > You are creating what I call a mono-frag. I am not a huge fan.
>
> Do you mean 'mono-frag' as the unifying of frag and non-frag
> page handling by assuming all pages in page pool having one
> frag user initially in patch 1?
> If yes, please let's continue the discussion in pacth 1 so that
> we don't have to restart the discussion again.
>
> Or is there some other obvious concern about 'mono-frag' I missed?
The main concern I have is having mono-frags and truly fragmented
pages mixing too much which is making a bit of a mess of things.
If we are going to use these fragments we really need to be able to
tell the truesize of these mono-frag pages. What I want to avoid is
seeing drivers abuse this to allocate huge swaths of memory but not
accounting for it in the skb_truesize.
On 2023/6/2 23:57, Alexander Duyck wrote:
> On Fri, Jun 2, 2023 at 5:23 AM Yunsheng Lin <[email protected]> wrote:
...
>>
>> According to my defination in this patchset:
>> frag page: page alloced from page_pool_alloc_frag() with page->pp_frag_count
>> being greater than one.
>> non-frag page:page alloced return from both page_pool_alloc_frag() and
>> page_pool_alloc_pages() with page->pp_frag_count being one.
>>
>> I assume the above 'non-page pool pages' refer to what I call as 'non-frag
>> page' alloced return from both page_pool_alloc_frag(), right? And it is
>> still about doing the (size << 1 > max_size)' checking at the begin instead
>> of at the middle right now to avoid extra steps for 'non-frag page' case?
>
> Yeah, the non-page I was referring to were you mono-frag pages.
I was using 'frag page' and 'non-frag page' per the defination above,
and you were using 'mono-frag' mostly and 'non-page' sometimes.
I am really confused by them as I felt like I got what they meant and
then I was lost when you used them in the next comment. I really hope
that you could describe what do you mean in more detailed by using
'mono-frag pages' and 'non-page', so that we can choose the right
naming to continue the discussion without further misunderstanding
and confusion.
>
>>>
>>>>>
>>>>>> - if (page && *offset + size > max_size) {
>>>>>> + if (page) {
>>>>>> + *offset = pool->frag_offset;
>>>>>> +
>>>>>> + if (*offset + size <= max_size) {
>>>>>> + pool->frag_users++;
>>>>>> + pool->frag_offset = *offset + size;
>>>>>> + alloc_stat_inc(pool, fast);
>>>>>> + return page;
>>>>
>>>> Note that we still allow frag page here when '(size << 1 > max_size)'.
>>
>> This is the optimization I was taking about: suppose we start
>> from a clean state with 64K page size, if page_pool_alloc_frag()
>> is called with size being 2K and then 34K, we only need one page
>> to satisfy caller's need as we do the '*offset + size > max_size'
>> checking before the '(size << 1 > max_size)' checking.
>
> The issue is the unaccounted for waste. We are supposed to know the
> general size of the frags being used so we can compute truesize. If
Note, for case of veth and virtio_net, the driver may only know the
current frag size when calling page_pool_alloc_frag(), it does not
konw what is the size of the frags will be used next time, how exactly
are we going to compute the truesize for cases with different frag
size? As far as I can tell, we may only do something like virtio_net
is doing with 'page_frag' for the last frag as below, for other frags,
the truesize may need to take accounting to the aligning requirement:
https://elixir.bootlin.com/linux/v6.3.5/source/drivers/net/virtio_net.c#L1638
> for example you are using an order 3 page and you are splitting it
> between a 2K and a 17K fragment the 2K fragments will have a massive
> truesize underestimate that can lead to memory issues if those smaller
> fragments end up holding onto the pages.
>
> As such we should try to keep the small fragments away from anything
> larger than half of the page.
IMHO, doing the above only alleviate the problem. How is above splitting
different from splitting it evently with 16 2K fragments, and when 15 frag
is released, we still have the last 2K fragment holding onto 32K memory,
doesn't that also cause massive truesize underestimate? Not to mention that
for system with 64K page size.
In RFC patch below, 'page_pool_frag' is used to report the truesize, but
I was thinking both 'page_frag' and 'page_frag_cache' both have a similiar
problem, so I dropped it in V1 and left that as a future improvement.
I can pick it up again if 'truesize' is really the concern, but we have to
align on how to compute the truesize here first.
https://patchwork.kernel.org/project/netdevbpf/patch/[email protected]/
On Fri, Jun 2, 2023 at 9:20 PM Yunsheng Lin <[email protected]> wrote:
>
> On 2023/6/2 23:57, Alexander Duyck wrote:
> > On Fri, Jun 2, 2023 at 5:23 AM Yunsheng Lin <[email protected]> wrote:
>
> ...
>
> >>
> >> According to my defination in this patchset:
> >> frag page: page alloced from page_pool_alloc_frag() with page->pp_frag_count
> >> being greater than one.
> >> non-frag page:page alloced return from both page_pool_alloc_frag() and
> >> page_pool_alloc_pages() with page->pp_frag_count being one.
> >>
> >> I assume the above 'non-page pool pages' refer to what I call as 'non-frag
> >> page' alloced return from both page_pool_alloc_frag(), right? And it is
> >> still about doing the (size << 1 > max_size)' checking at the begin instead
> >> of at the middle right now to avoid extra steps for 'non-frag page' case?
> >
> > Yeah, the non-page I was referring to were you mono-frag pages.
>
> I was using 'frag page' and 'non-frag page' per the defination above,
> and you were using 'mono-frag' mostly and 'non-page' sometimes.
> I am really confused by them as I felt like I got what they meant and
> then I was lost when you used them in the next comment. I really hope
> that you could describe what do you mean in more detailed by using
> 'mono-frag pages' and 'non-page', so that we can choose the right
> naming to continue the discussion without further misunderstanding
> and confusion.
I will try to be consistent about this going forward:
non-fragmented - legacy page pool w/o page frags
mono-frag - after this page page pool w/o frags
fragmented - before/after this patch w/ frags
> >
> >>>
> >>>>>
> >>>>>> - if (page && *offset + size > max_size) {
> >>>>>> + if (page) {
> >>>>>> + *offset = pool->frag_offset;
> >>>>>> +
> >>>>>> + if (*offset + size <= max_size) {
> >>>>>> + pool->frag_users++;
> >>>>>> + pool->frag_offset = *offset + size;
> >>>>>> + alloc_stat_inc(pool, fast);
> >>>>>> + return page;
> >>>>
> >>>> Note that we still allow frag page here when '(size << 1 > max_size)'.
> >>
> >> This is the optimization I was taking about: suppose we start
> >> from a clean state with 64K page size, if page_pool_alloc_frag()
> >> is called with size being 2K and then 34K, we only need one page
> >> to satisfy caller's need as we do the '*offset + size > max_size'
> >> checking before the '(size << 1 > max_size)' checking.
> >
> > The issue is the unaccounted for waste. We are supposed to know the
> > general size of the frags being used so we can compute truesize. If
>
> Note, for case of veth and virtio_net, the driver may only know the
> current frag size when calling page_pool_alloc_frag(), it does not
> konw what is the size of the frags will be used next time, how exactly
> are we going to compute the truesize for cases with different frag
> size? As far as I can tell, we may only do something like virtio_net
> is doing with 'page_frag' for the last frag as below, for other frags,
> the truesize may need to take accounting to the aligning requirement:
> https://elixir.bootlin.com/linux/v6.3.5/source/drivers/net/virtio_net.c#L1638
Yeah, that is more-or-less what I am getting at. This is why drivers
will tend to want to allocate a mono-frag themselves and then take
care of adding additional fragmentation as needed. If you are
abstracting that away from the driver then it makes it much harder to
track that truesize.
> > for example you are using an order 3 page and you are splitting it
> > between a 2K and a 17K fragment the 2K fragments will have a massive
> > truesize underestimate that can lead to memory issues if those smaller
> > fragments end up holding onto the pages.
> >
> > As such we should try to keep the small fragments away from anything
> > larger than half of the page.
>
> IMHO, doing the above only alleviate the problem. How is above splitting
> different from splitting it evently with 16 2K fragments, and when 15 frag
> is released, we still have the last 2K fragment holding onto 32K memory,
> doesn't that also cause massive truesize underestimate? Not to mention that
> for system with 64K page size.
Yes, that is a known issue. That is why I am not wanting us to further
exacerbate the issue.
> In RFC patch below, 'page_pool_frag' is used to report the truesize, but
> I was thinking both 'page_frag' and 'page_frag_cache' both have a similiar
> problem, so I dropped it in V1 and left that as a future improvement.
>
> I can pick it up again if 'truesize' is really the concern, but we have to
> align on how to compute the truesize here first.
>
> https://patchwork.kernel.org/project/netdevbpf/patch/[email protected]/
I am assuming this is the same one you mentioned in the other patch.
As I said the problem is the remainder is being ignored. The logic
should be pushed to the drivers to handle the truesize and is one of
the reasons why the expectation is that either the driver will use
something like a fixed constant size if it is using the raw page pool
fragments, or if it is going to do random sized chunks then it will
track the size of the chunks of the page is it using and assign the
remainders to the last fragment used in a given page.