mbox series

[RFC,v2,00/21] mm/zsmalloc: Split zsdesc from struct page

Message ID 20230713042037.980211-1-42.hyeyoo@gmail.com (mailing list archive)
Headers show
Series mm/zsmalloc: Split zsdesc from struct page | expand

Message

Hyeonggon Yoo July 13, 2023, 4:20 a.m. UTC
v1: https://lore.kernel.org/linux-mm/20230220132218.546369-1-42.hyeyoo@gmail.com

v1 -> v2:
 - rebased to the latest mm-unstable, resulting in some patches dropped
 - adjusted comments from Mike Rapoport, defining helpers when
   converting its users

The purpose of this series is to define own memory descriptor for zsmalloc,
instead of re-using various fields of struct page. This is a part of the
effort to reduce the size of struct page to unsigned long and enable
dynamic allocation of memory descriptors.

While [1] outlines this ultimate objective, the current use of struct page
is highly dependent on its definition, making it challenging to separately
allocate memory descriptors.

Therefore, this series introduces new descriptor for zsmalloc, called
zsdesc. It overlays struct page for now, but will eventually be allocated
independently in the future. And apart from dynamic allocation of descriptors,
this is a nice cleanup.

This work is also available at:
	https://gitlab.com/hyeyoo/linux/-/tree/separate_zsdesc_rfc-v2

[1] State Of The Page, August 2022
https://lore.kernel.org/lkml/YvV1KTyzZ+Jrtj9x@casper.infradead.org

Hyeonggon Yoo (21):
  mm/zsmalloc: create new struct zsdesc
  mm/zsmalloc: add utility functions for zsdesc
  mm/zsmalloc: replace first_page to first_zsdesc in struct zspage
  mm/zsmalloc: add alternatives of frequently used helper functions
  mm/zsmalloc: convert {try,}lock_zspage() to use zsdesc
  mm/zsmalloc: convert __zs_{map,unmap}_object() to use zsdesc
  mm/zsmalloc: convert obj_to_location() and its users to use zsdesc
  mm/zsmalloc: convert obj_malloc() to use zsdesc
  mm/zsmalloc: convert create_page_chain() and its user to use zsdesc
  mm/zsmalloc: convert obj_allocated() and related helpers to use zsdesc
  mm/zsmalloc: convert init_zspage() to use zsdesc
  mm/zsmalloc: convert obj_to_page() and zs_free() to use zsdesc
  mm/zsmalloc: convert reset_page() to reset_zsdesc()
  mm/zsmalloc: convert zs_page_{isolate,migrate,putback} to use zsdesc
  mm/zsmalloc: convert __free_zspage() to use zsdesc
  mm/zsmalloc: convert location_to_obj() to use zsdesc
  mm/zsmalloc: convert migrate_zspage() to use zsdesc
  mm/zsmalloc: convert get_zspage() to take zsdesc
  mm/zsmalloc: convert SetZsPageMovable() to use zsdesc
  mm/zsmalloc: remove now unused helper functions
  mm/zsmalloc: convert {get,set}_first_obj_offset() to use zsdesc

 mm/zsmalloc.c | 574 +++++++++++++++++++++++++++++++-------------------
 1 file changed, 360 insertions(+), 214 deletions(-)

Comments

Sergey Senozhatsky July 20, 2023, 7:18 a.m. UTC | #1
On (23/07/13 13:20), Hyeonggon Yoo wrote:
> The purpose of this series is to define own memory descriptor for zsmalloc,
> instead of re-using various fields of struct page. This is a part of the
> effort to reduce the size of struct page to unsigned long and enable
> dynamic allocation of memory descriptors.
> 
> While [1] outlines this ultimate objective, the current use of struct page
> is highly dependent on its definition, making it challenging to separately
> allocate memory descriptors.

I glanced through the series and it all looks pretty straight forward to
me. I'll have a closer look. And we definitely need Minchan to ACK it.

> Therefore, this series introduces new descriptor for zsmalloc, called
> zsdesc. It overlays struct page for now, but will eventually be allocated
> independently in the future.

So I don't expect zsmalloc memory usage increase. On one hand for each
physical page that zspage consists of we will allocate zsdesc (extra bytes),
but at the same time struct page gets slimmer. So we should be even, or
am I wrong?
Yosry Ahmed July 20, 2023, 7:54 a.m. UTC | #2
On Thu, Jul 20, 2023 at 12:18 AM Sergey Senozhatsky
<senozhatsky@chromium.org> wrote:
>
> On (23/07/13 13:20), Hyeonggon Yoo wrote:
> > The purpose of this series is to define own memory descriptor for zsmalloc,
> > instead of re-using various fields of struct page. This is a part of the
> > effort to reduce the size of struct page to unsigned long and enable
> > dynamic allocation of memory descriptors.
> >
> > While [1] outlines this ultimate objective, the current use of struct page
> > is highly dependent on its definition, making it challenging to separately
> > allocate memory descriptors.
>
> I glanced through the series and it all looks pretty straight forward to
> me. I'll have a closer look. And we definitely need Minchan to ACK it.
>
> > Therefore, this series introduces new descriptor for zsmalloc, called
> > zsdesc. It overlays struct page for now, but will eventually be allocated
> > independently in the future.
>
> So I don't expect zsmalloc memory usage increase. On one hand for each
> physical page that zspage consists of we will allocate zsdesc (extra bytes),
> but at the same time struct page gets slimmer. So we should be even, or
> am I wrong?

Well, it depends. Here is my understanding (which may be completely wrong):

The end goal would be to have an 8-byte memdesc for each order-0 page,
and then allocate a specialized struct per-folio according to the use
case. In this case, we would have a memdesc and a zsdesc for each
order-0 page. If sizeof(zsdesc) is 64 bytes (on 64-bit), then it's a
net loss. The savings only start kicking in with higher order folios.
As of now, zsmalloc only uses order-0 pages as far as I can tell, so
the usage would increase if I understand correctly.

It seems to me though the sizeof(zsdesc) is actually 56 bytes (on
64-bit), so sizeof(zsdesc) + sizeof(memdesc) would be equal to the
current size of struct page. If that's true, then there is no loss,
and there's potential gain if we start using higher order folios in
zsmalloc in the future.

(That is of course unless we want to maintain cache line alignment for
the zsdescs, then we might end up using 64 bytes anyway).
Hyeonggon Yoo July 20, 2023, 11:34 a.m. UTC | #3
On Thu, Jul 20, 2023 at 4:55 PM Yosry Ahmed <yosryahmed@google.com> wrote:
>
> On Thu, Jul 20, 2023 at 12:18 AM Sergey Senozhatsky
> <senozhatsky@chromium.org> wrote:
> >
> > On (23/07/13 13:20), Hyeonggon Yoo wrote:
> > > The purpose of this series is to define own memory descriptor for zsmalloc,
> > > instead of re-using various fields of struct page. This is a part of the
> > > effort to reduce the size of struct page to unsigned long and enable
> > > dynamic allocation of memory descriptors.
> > >
> > > While [1] outlines this ultimate objective, the current use of struct page
> > > is highly dependent on its definition, making it challenging to separately
> > > allocate memory descriptors.
> >
> > I glanced through the series and it all looks pretty straight forward to
> > me. I'll have a closer look. And we definitely need Minchan to ACK it.
> >
> > > Therefore, this series introduces new descriptor for zsmalloc, called
> > > zsdesc. It overlays struct page for now, but will eventually be allocated
> > > independently in the future.
> >
> > So I don't expect zsmalloc memory usage increase. On one hand for each
> > physical page that zspage consists of we will allocate zsdesc (extra bytes),
> > but at the same time struct page gets slimmer. So we should be even, or
> > am I wrong?
>
> Well, it depends. Here is my understanding (which may be completely wrong):
>
> The end goal would be to have an 8-byte memdesc for each order-0 page,
> and then allocate a specialized struct per-folio according to the use
> case. In this case, we would have a memdesc and a zsdesc for each
> order-0 page. If sizeof(zsdesc) is 64 bytes (on 64-bit), then it's a
> net loss. The savings only start kicking in with higher order folios.
> As of now, zsmalloc only uses order-0 pages as far as I can tell, so
> the usage would increase if I understand correctly.

I partially agree with you that the point of memdesc stuff is
allocating a use-case specific
descriptor per folio. but I thought the primary gain from memdesc was
from anon and file pages
(where high order pages are more usable), rather than zsmalloc.

And I believe enabling a memory descriptor per folio would be
impossible (or inefficient)
if zsmalloc and other subsystems are using struct page in the current
way (or please tell me I'm wrong?)

So I expect the primary gain would be from high-order anon/file folios,
while this series is a prerequisite for them to work sanely.

> It seems to me though the sizeof(zsdesc) is actually 56 bytes (on
> 64-bit), so sizeof(zsdesc) + sizeof(memdesc) would be equal to the
> current size of struct page. If that's true, then there is no loss,

Yeah, zsdesc would be 56 bytes on 64 bit CPUs as memcg_data field is
not used in zsmalloc.
More fields in the current struct page might not be needed in the
future, although it's hard to say at the moment.
but it's not a loss.

> and there's potential gain if we start using higher order folios in
> zsmalloc in the future.

AFAICS zsmalloc should work even when the system memory is fragmented,
so we may implement fallback allocation (as currently discussed in
large anon folios thread).

It might work, but IMHO the purpose of this series is to enable memdesc
for large anon/file folios, rather than seeing a large gain in zsmalloc itself.
(But even in zsmalloc, it's not a loss)

> (That is of course unless we want to maintain cache line alignment for
> the zsdescs, then we might end up using 64 bytes anyway).

we already don't require cache line alignment for struct page. the current
alignment requirement is due to SLUB's cmpxchg128 operation, not cache
line alignment.

I might be wrong in some aspects, so please tell me if I am.
And thank you and Sergey for taking a look at this!
--
Hyeonggon
Yosry Ahmed July 20, 2023, 6:31 p.m. UTC | #4
On Thu, Jul 20, 2023 at 4:34 AM Hyeonggon Yoo <42.hyeyoo@gmail.com> wrote:
>
> On Thu, Jul 20, 2023 at 4:55 PM Yosry Ahmed <yosryahmed@google.com> wrote:
> >
> > On Thu, Jul 20, 2023 at 12:18 AM Sergey Senozhatsky
> > <senozhatsky@chromium.org> wrote:
> > >
> > > On (23/07/13 13:20), Hyeonggon Yoo wrote:
> > > > The purpose of this series is to define own memory descriptor for zsmalloc,
> > > > instead of re-using various fields of struct page. This is a part of the
> > > > effort to reduce the size of struct page to unsigned long and enable
> > > > dynamic allocation of memory descriptors.
> > > >
> > > > While [1] outlines this ultimate objective, the current use of struct page
> > > > is highly dependent on its definition, making it challenging to separately
> > > > allocate memory descriptors.
> > >
> > > I glanced through the series and it all looks pretty straight forward to
> > > me. I'll have a closer look. And we definitely need Minchan to ACK it.
> > >
> > > > Therefore, this series introduces new descriptor for zsmalloc, called
> > > > zsdesc. It overlays struct page for now, but will eventually be allocated
> > > > independently in the future.
> > >
> > > So I don't expect zsmalloc memory usage increase. On one hand for each
> > > physical page that zspage consists of we will allocate zsdesc (extra bytes),
> > > but at the same time struct page gets slimmer. So we should be even, or
> > > am I wrong?
> >
> > Well, it depends. Here is my understanding (which may be completely wrong):
> >
> > The end goal would be to have an 8-byte memdesc for each order-0 page,
> > and then allocate a specialized struct per-folio according to the use
> > case. In this case, we would have a memdesc and a zsdesc for each
> > order-0 page. If sizeof(zsdesc) is 64 bytes (on 64-bit), then it's a
> > net loss. The savings only start kicking in with higher order folios.
> > As of now, zsmalloc only uses order-0 pages as far as I can tell, so
> > the usage would increase if I understand correctly.
>
> I partially agree with you that the point of memdesc stuff is
> allocating a use-case specific
> descriptor per folio. but I thought the primary gain from memdesc was
> from anon and file pages
> (where high order pages are more usable), rather than zsmalloc.
>
> And I believe enabling a memory descriptor per folio would be
> impossible (or inefficient)
> if zsmalloc and other subsystems are using struct page in the current
> way (or please tell me I'm wrong?)
>
> So I expect the primary gain would be from high-order anon/file folios,
> while this series is a prerequisite for them to work sanely.

Right, I agree with that, sorry if I wasn't clear. I meant that
generally speaking, we see gains from memdesc from higher order
folios, so for zsmalloc specifically we probably won't see seeing any
savings, and *might* see some extra usage (which I might be wrong
about, see below).

>
> > It seems to me though the sizeof(zsdesc) is actually 56 bytes (on
> > 64-bit), so sizeof(zsdesc) + sizeof(memdesc) would be equal to the
> > current size of struct page. If that's true, then there is no loss,
>
> Yeah, zsdesc would be 56 bytes on 64 bit CPUs as memcg_data field is
> not used in zsmalloc.
> More fields in the current struct page might not be needed in the
> future, although it's hard to say at the moment.
> but it's not a loss.

Is page->memcg_data something that we can drop? Aren't there code
paths that will check page->memcg_data even for kernel pages (e.g.
__folio_put() -> __folio_put_small() -> mem_cgroup_uncharge() ) ?

>
> > and there's potential gain if we start using higher order folios in
> > zsmalloc in the future.
>
> AFAICS zsmalloc should work even when the system memory is fragmented,
> so we may implement fallback allocation (as currently discussed in
> large anon folios thread).

Of course, any usage of higher order folios in zsmalloc must have a
fallback logic, although it might be simpler for zsmalloc than anon
folios. I agree that's off topic here.

>
> It might work, but IMHO the purpose of this series is to enable memdesc
> for large anon/file folios, rather than seeing a large gain in zsmalloc itself.
> (But even in zsmalloc, it's not a loss)
>
> > (That is of course unless we want to maintain cache line alignment for
> > the zsdescs, then we might end up using 64 bytes anyway).
>
> we already don't require cache line alignment for struct page. the current
> alignment requirement is due to SLUB's cmpxchg128 operation, not cache
> line alignment.

I thought we want  struct page to be cache line aligned (to avoid
having to fetch two cache lines for one struct page), but I can easily
be wrong.

>
> I might be wrong in some aspects, so please tell me if I am.
> And thank you and Sergey for taking a look at this!

Thanks to you for doing the work!

> --
> Hyeonggon
Hyeonggon Yoo July 20, 2023, 9:33 p.m. UTC | #5
On Fri, Jul 21, 2023 at 3:31 AM Yosry Ahmed <yosryahmed@google.com> wrote:
>
> On Thu, Jul 20, 2023 at 4:34 AM Hyeonggon Yoo <42.hyeyoo@gmail.com> wrote:
> >
> > On Thu, Jul 20, 2023 at 4:55 PM Yosry Ahmed <yosryahmed@google.com> wrote:
> > >
> > > On Thu, Jul 20, 2023 at 12:18 AM Sergey Senozhatsky
> > > <senozhatsky@chromium.org> wrote:
> > > >
> > > > On (23/07/13 13:20), Hyeonggon Yoo wrote:
> > > > > The purpose of this series is to define own memory descriptor for zsmalloc,
> > > > > instead of re-using various fields of struct page. This is a part of the
> > > > > effort to reduce the size of struct page to unsigned long and enable
> > > > > dynamic allocation of memory descriptors.
> > > > >
> > > > > While [1] outlines this ultimate objective, the current use of struct page
> > > > > is highly dependent on its definition, making it challenging to separately
> > > > > allocate memory descriptors.
> > > >
> > > > I glanced through the series and it all looks pretty straight forward to
> > > > me. I'll have a closer look. And we definitely need Minchan to ACK it.
> > > >
> > > > > Therefore, this series introduces new descriptor for zsmalloc, called
> > > > > zsdesc. It overlays struct page for now, but will eventually be allocated
> > > > > independently in the future.
> > > >
> > > > So I don't expect zsmalloc memory usage increase. On one hand for each
> > > > physical page that zspage consists of we will allocate zsdesc (extra bytes),
> > > > but at the same time struct page gets slimmer. So we should be even, or
> > > > am I wrong?
> > >
> > > Well, it depends. Here is my understanding (which may be completely wrong):
> > >
> > > The end goal would be to have an 8-byte memdesc for each order-0 page,
> > > and then allocate a specialized struct per-folio according to the use
> > > case. In this case, we would have a memdesc and a zsdesc for each
> > > order-0 page. If sizeof(zsdesc) is 64 bytes (on 64-bit), then it's a
> > > net loss. The savings only start kicking in with higher order folios.
> > > As of now, zsmalloc only uses order-0 pages as far as I can tell, so
> > > the usage would increase if I understand correctly.
> >
> > I partially agree with you that the point of memdesc stuff is
> > allocating a use-case specific
> > descriptor per folio. but I thought the primary gain from memdesc was
> > from anon and file pages
> > (where high order pages are more usable), rather than zsmalloc.
> >
> > And I believe enabling a memory descriptor per folio would be
> > impossible (or inefficient)
> > if zsmalloc and other subsystems are using struct page in the current
> > way (or please tell me I'm wrong?)
> >
> > So I expect the primary gain would be from high-order anon/file folios,
> > while this series is a prerequisite for them to work sanely.
>
> Right, I agree with that, sorry if I wasn't clear. I meant that
> generally speaking, we see gains from memdesc from higher order
> folios, so for zsmalloc specifically we probably won't see seeing any
> savings, and *might* see some extra usage (which I might be wrong
> about, see below).

Yeah, even if I said, "oh, we don't necessarily need to use extra
memory for zsdesc"
below, a slight increase wouldn't hurt too much in that perspective,
because there
will be savings from other users of memdesc.

> > > It seems to me though the sizeof(zsdesc) is actually 56 bytes (on
> > > 64-bit), so sizeof(zsdesc) + sizeof(memdesc) would be equal to the
> > > current size of struct page. If that's true, then there is no loss,
> >
> > Yeah, zsdesc would be 56 bytes on 64 bit CPUs as memcg_data field is
> > not used in zsmalloc.
> > More fields in the current struct page might not be needed in the
> > future, although it's hard to say at the moment.
> > but it's not a loss.
>
> Is page->memcg_data something that we can drop? Aren't there code
> paths that will check page->memcg_data even for kernel pages (e.g.
> __folio_put() -> __folio_put_small() -> mem_cgroup_uncharge() ) ?

zsmalloc pages are not accounted for via __GFP_ACCOUNT,
and IIUC the current implementation of zswap memcg charging does not
use memcg_data
either - so I think it can be dropped.

I think we don't want to increase memdesc to 16 bytes by adding memcg_data.
It should be in use-case specific descriptors if it can be charged to memcg?

> > > and there's potential gain if we start using higher order folios in
> > > zsmalloc in the future.
> >
> > AFAICS zsmalloc should work even when the system memory is fragmented,
> > so we may implement fallback allocation (as currently discussed in
> > large anon folios thread).
>
> Of course, any usage of higher order folios in zsmalloc must have a
> fallback logic, although it might be simpler for zsmalloc than anon
> folios. I agree that's off topic here.
> > It might work, but IMHO the purpose of this series is to enable memdesc
> > for large anon/file folios, rather than seeing a large gain in zsmalloc itself.
> > (But even in zsmalloc, it's not a loss)
> >
> > > (That is of course unless we want to maintain cache line alignment for
> > > the zsdescs, then we might end up using 64 bytes anyway).
> >
> > we already don't require cache line alignment for struct page. the current
> > alignment requirement is due to SLUB's cmpxchg128 operation, not cache
> > line alignment.
>
> I thought we want  struct page to be cache line aligned (to avoid
> having to fetch two cache lines for one struct page), but I can easily
> be wrong.

Right. I admit that even if it's not required to be cache line
aligned, it is 64 bytes
in commonly used configurations. and changing it could affect some workloads.

But I think for zsdesc it would be better not to align by cache line
size, before
observing degradations due to alignment. By the time zsmalloc is intensively
used, it shouldn't be a huge issue.

> > I might be wrong in some aspects, so please tell me if I am.
> > And thank you and Sergey for taking a look at this!
>
> Thanks to you for doing the work!

No problem! :)
Yosry Ahmed July 20, 2023, 9:38 p.m. UTC | #6
<snip>
>
> > > > It seems to me though the sizeof(zsdesc) is actually 56 bytes (on
> > > > 64-bit), so sizeof(zsdesc) + sizeof(memdesc) would be equal to the
> > > > current size of struct page. If that's true, then there is no loss,
> > >
> > > Yeah, zsdesc would be 56 bytes on 64 bit CPUs as memcg_data field is
> > > not used in zsmalloc.
> > > More fields in the current struct page might not be needed in the
> > > future, although it's hard to say at the moment.
> > > but it's not a loss.
> >
> > Is page->memcg_data something that we can drop? Aren't there code
> > paths that will check page->memcg_data even for kernel pages (e.g.
> > __folio_put() -> __folio_put_small() -> mem_cgroup_uncharge() ) ?
>
> zsmalloc pages are not accounted for via __GFP_ACCOUNT,

Yeah, but the code in the free path above will check page->memcg_data
nonetheless to check if it is charged. I think to drop memcg_data we
need to enlighten the code that some pages do not even have memcg_data
at all, no?

> and IIUC the current implementation of zswap memcg charging does not
> use memcg_data
> either - so I think it can be dropped.

My question is more about the generic mm code expecting to see
page->memcg_data in every page, even if it is not actually used
(zero).

>
> I think we don't want to increase memdesc to 16 bytes by adding memcg_data.
> It should be in use-case specific descriptors if it can be charged to memcg?
>
<snip>
Hyeonggon Yoo July 20, 2023, 9:52 p.m. UTC | #7
On Fri, Jul 21, 2023 at 6:39 AM Yosry Ahmed <yosryahmed@google.com> wrote:
>
> <snip>
> >
> > > > > It seems to me though the sizeof(zsdesc) is actually 56 bytes (on
> > > > > 64-bit), so sizeof(zsdesc) + sizeof(memdesc) would be equal to the
> > > > > current size of struct page. If that's true, then there is no loss,
> > > >
> > > > Yeah, zsdesc would be 56 bytes on 64 bit CPUs as memcg_data field is
> > > > not used in zsmalloc.
> > > > More fields in the current struct page might not be needed in the
> > > > future, although it's hard to say at the moment.
> > > > but it's not a loss.
> > >
> > > Is page->memcg_data something that we can drop? Aren't there code
> > > paths that will check page->memcg_data even for kernel pages (e.g.
> > > __folio_put() -> __folio_put_small() -> mem_cgroup_uncharge() ) ?
> >
> > zsmalloc pages are not accounted for via __GFP_ACCOUNT,
>
> Yeah, but the code in the free path above will check page->memcg_data
> nonetheless to check if it is charged.

Right.

> I think to drop memcg_data we need to enlighten the code that some pages
> do not even have memcg_data at all

I agree with you. It should be one of the milestones for all of this to work.
It won't be complicated for the code to be aware of it, because there will be
a freeing (and uncharging if need) routine per type of descriptors.
Yosry Ahmed July 20, 2023, 9:57 p.m. UTC | #8
On Thu, Jul 20, 2023 at 2:52 PM Hyeonggon Yoo <42.hyeyoo@gmail.com> wrote:
>
> On Fri, Jul 21, 2023 at 6:39 AM Yosry Ahmed <yosryahmed@google.com> wrote:
> >
> > <snip>
> > >
> > > > > > It seems to me though the sizeof(zsdesc) is actually 56 bytes (on
> > > > > > 64-bit), so sizeof(zsdesc) + sizeof(memdesc) would be equal to the
> > > > > > current size of struct page. If that's true, then there is no loss,
> > > > >
> > > > > Yeah, zsdesc would be 56 bytes on 64 bit CPUs as memcg_data field is
> > > > > not used in zsmalloc.
> > > > > More fields in the current struct page might not be needed in the
> > > > > future, although it's hard to say at the moment.
> > > > > but it's not a loss.
> > > >
> > > > Is page->memcg_data something that we can drop? Aren't there code
> > > > paths that will check page->memcg_data even for kernel pages (e.g.
> > > > __folio_put() -> __folio_put_small() -> mem_cgroup_uncharge() ) ?
> > >
> > > zsmalloc pages are not accounted for via __GFP_ACCOUNT,
> >
> > Yeah, but the code in the free path above will check page->memcg_data
> > nonetheless to check if it is charged.
>
> Right.
>
> > I think to drop memcg_data we need to enlighten the code that some pages
> > do not even have memcg_data at all
>
> I agree with you. It should be one of the milestones for all of this to work.
> It won't be complicated for the code to be aware of it, because there will be
> a freeing (and uncharging if need) routine per type of descriptors.

Right.

For this patch series, do we need to maintain memcg_data in zsdec to
avoid any subtle problems?