linux-block.git
2 months agomm/damon/core: introduce damos->ops_filters
SeongJae Park [Tue, 4 Mar 2025 21:19:05 +0000 (13:19 -0800)]
mm/damon/core: introduce damos->ops_filters

Patch series "mm/damon: make allow filters after reject filters useful and
intuitive".

DAMOS filters do allow or reject elements of memory for given DAMOS scheme
only if those match the filter criterias.  For elements that don't match
any DAMOS filter, 'allowing' is the default behavior.  This makes
allow-filters that don't have any reject-filter after them meaningless
sources of overhead.  The decision was made to keep the behavior
consistent with that before the introduction of allow-filters.  This,
however, makes usage of DAMOS filters confusing and inefficient.  It is
more intuitive and still consistent behavior to reject by default unless
there is no filter at all or the last filter is a reject filter.  Update
the filtering logic in the way and update documents to clarify the
behavior.

Note that this is changing the old behavior.  But the old behavior for the
problematic filter combination was definitely confusing, inefficient and
anyway useless.  Also, the behavior has relatively recently introduced.
It is difficult to anticipate any user that depends on the behavior.
Hence this is not a user-breaking behavior change but an obvious
improvement.

This patch (of 9):

DAMOS filters can be categorized into two groups depending on which layer
they are handled, namely core layer and ops layer.  The groups are
important because the filtering behavior depends on evaluation sequence of
filters, and core layer-handled filters are evaluated before operations
layer-handled ones.

The behavior is clearly documented, but the implementation is bit
inefficient and complicated.  All filters are maintained in a single list
(damos->filters) in mix.  Filters evaluation logics in core layer and
operations layer iterates all the filters on the list, while skipping
filters that should be not handled by the layer of the logic.  It is
inefficient.  Making future extensions having differentiations for filters
of different handling layers will also be complicated.

Add a new list that will be used for having all operations layer-handled
DAMOS filters to DAMOS scheme data structure.  Also add the support of its
initialization and basic traversal functions.

Link: https://lkml.kernel.org/r/20250304211913.53574-1-sj@kernel.org
Link: https://lkml.kernel.org/r/20250304211913.53574-2-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agowriteback: fix calculations in trace_balance_dirty_pages() for cgwb
Tang Yizhou [Tue, 4 Mar 2025 11:03:18 +0000 (19:03 +0800)]
writeback: fix calculations in trace_balance_dirty_pages() for cgwb

In the commit dcc25ae76eb7 ("writeback: move global_dirty_limit into
wb_domain") of the cgroup writeback backpressure propagation patchset,
Tejun made some adaptations to trace_balance_dirty_pages() for cgroup
writeback.  However, this adaptation was incomplete and Tejun missed
further adaptation in the subsequent patches.

In the cgroup writeback scenario, if sdtc in balance_dirty_pages() is
assigned to mdtc, then upon entering trace_balance_dirty_pages(),
__entry->limit should be assigned based on the dirty_limit of the
corresponding memcg's wb_domain, rather than global_wb_domain.

To address this issue and simplify the implementation, introduce a 'limit'
field in struct dirty_throttle_control to store the hard_limit value
computed in wb_position_ratio() by calling hard_dirty_limit().  This field
will then be used in trace_balance_dirty_pages() to assign the value to
__entry->limit.

Link: https://lkml.kernel.org/r/20250304110318.159567-4-yizhou.tang@shopee.com
Fixes: dcc25ae76eb7 ("writeback: move global_dirty_limit into wb_domain")
Signed-off-by: Tang Yizhou <yizhou.tang@shopee.com>
Acked-by: Tejun Heo <tj@kernel.org>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Jan Kara <jack@suse.cz>
Cc: "Masami Hiramatsu (Google)" <mhiramat@kernel.org>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agowriteback: rename variables in trace_balance_dirty_pages()
Tang Yizhou [Tue, 4 Mar 2025 11:03:17 +0000 (19:03 +0800)]
writeback: rename variables in trace_balance_dirty_pages()

Rename bdi_setpoint and bdi_dirty in the tracepoint to wb_setpoint and
wb_dirty, respectively. These changes were omitted by Tejun in the cgroup
writeback patchset.

Link: https://lkml.kernel.org/r/20250304110318.159567-3-yizhou.tang@shopee.com
Signed-off-by: Tang Yizhou <yizhou.tang@shopee.com>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Jan Kara <jack@suse.cz>
Cc: "Masami Hiramatsu (Google)" <mhiramat@kernel.org>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agowriteback: let trace_balance_dirty_pages() take struct dtc as parameter
Tang Yizhou [Tue, 4 Mar 2025 11:03:16 +0000 (19:03 +0800)]
writeback: let trace_balance_dirty_pages() take struct dtc as parameter

Patch series "Fix calculations in trace_balance_dirty_pages() for cgwb", v2.

In my experiment, I found that the output of trace_balance_dirty_pages()
in the cgroup writeback scenario was strange because
trace_balance_dirty_pages() always uses global_wb_domain.dirty_limit for
related calculations instead of the dirty_limit of the corresponding
memcg's wb_domain.

The basic idea of the fix is to store the hard dirty limit value computed
in wb_position_ratio() into struct dirty_throttle_control and use it for
calculations in trace_balance_dirty_pages().

This patch (of 3):

Currently, trace_balance_dirty_pages() already has 12 parameters.  In the
patch #3, I initially attempted to introduce an additional parameter.
However, in include/linux/trace_events.h, bpf_trace_run12() only supports
up to 12 parameters and bpf_trace_run13() does not exist.

To reduce the number of parameters in trace_balance_dirty_pages(), we can
make it accept a pointer to struct dirty_throttle_control as a parameter.
To achieve this, we need to move the definition of struct
dirty_throttle_control from mm/page-writeback.c to
include/linux/writeback.h.

Link: https://lkml.kernel.org/r/20250304110318.159567-1-yizhou.tang@shopee.com
Link: https://lkml.kernel.org/r/20250304110318.159567-2-yizhou.tang@shopee.com
Signed-off-by: Tang Yizhou <yizhou.tang@shopee.com>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Jan Kara <jack@suse.cz>
Cc: "Masami Hiramatsu (Google)" <mhiramat@kernel.org>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Tang Yizhou <yizhou.tang@shopee.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm: hugetlb: log time needed to allocate hugepages
Thomas Prescher [Thu, 27 Feb 2025 22:45:07 +0000 (23:45 +0100)]
mm: hugetlb: log time needed to allocate hugepages

Having this information allows users to easily tune the
hugepages_node_threads parameter.

Link: https://lkml.kernel.org/r/20250227-hugepage-parameter-v2-3-7db8c6dc0453@cyberus-technology.de
Signed-off-by: Thomas Prescher <thomas.prescher@cyberus-technology.de>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm: hugetlb: add hugetlb_alloc_threads cmdline option
Thomas Prescher [Thu, 27 Feb 2025 22:45:06 +0000 (23:45 +0100)]
mm: hugetlb: add hugetlb_alloc_threads cmdline option

Add a command line option that enables control of how many threads should
be used to allocate huge pages.

[akpm@linux-foundation.org: tidy up a comment]
Link: https://lkml.kernel.org/r/20250227-hugepage-parameter-v2-2-7db8c6dc0453@cyberus-technology.de
Signed-off-by: Thomas Prescher <thomas.prescher@cyberus-technology.de>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm: hugetlb: improve parallel huge page allocation time
Thomas Prescher [Thu, 27 Feb 2025 22:45:05 +0000 (23:45 +0100)]
mm: hugetlb: improve parallel huge page allocation time

Patch series "Add a command line option that enables control of how many
threads should be used to allocate huge pages", v2.

Allocating huge pages can take a very long time on servers with terabytes
of memory even when they are allocated at boot time where the allocation
happens in parallel.

Before this series, the kernel used a hard coded value of 2 threads per
NUMA node for these allocations.  This value might have been good enough
in the past but it is not sufficient to fully utilize newer systems.

This series changes the default so the kernel uses 25% of the available
hardware threads for these allocations.  In addition, we allow the user
that wish to micro-optimize the allocation time to override this value via
a new kernel parameter.

We tested this on 2 generations of Xeon CPUs and the results show a big
improvement of the overall allocation time.

+-----------------------+-------+-------+-------+-------+-------+
| threads               |   8   |   16  |   32  |   64  |   128 |
+-----------------------+-------+-------+-------+-------+-------+
| skylake      144 cpus |   44s |   22s |   16s |   19s |   20s |
| cascade lake 192 cpus |   39s |   20s |   11s |   10s |    9s |
+-----------------------+-------+-------+-------+-------+-------+

On skylake, we see an improvment of 2.75x when using 32 threads, on
cascade lake we can get even better at 4.3x when we use 128 threads.

This speedup is quite significant and users of large machines like these
should have the option to make the machines boot as fast as possible.

This patch (of 3):

Before this patch, the kernel currently used a hard coded value of 2
threads per NUMA node for these allocations.

This patch changes this policy and the kernel now uses 25% of the
available hardware threads for the allocations.

Link: https://lkml.kernel.org/r/20250227-hugepage-parameter-v2-0-7db8c6dc0453@cyberus-technology.de
Link: https://lkml.kernel.org/r/20250227-hugepage-parameter-v2-1-7db8c6dc0453@cyberus-technology.de
Signed-off-by: Thomas Prescher <thomas.prescher@cyberus-technology.de>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomemcg: bypass root memcg check for skmem charging
Shakeel Butt [Fri, 28 Feb 2025 02:23:54 +0000 (18:23 -0800)]
memcg: bypass root memcg check for skmem charging

The root memcg is never associated with a socket in mem_cgroup_sk_alloc,
so there is no need to check if the given memcg is root for the skmem
charging code path.

Link: https://lkml.kernel.org/r/20250228022354.2624249-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agopage_counter: reduce struct page_counter size
Shakeel Butt [Fri, 28 Feb 2025 07:58:08 +0000 (23:58 -0800)]
page_counter: reduce struct page_counter size

The struct page_counter has explicit padding for better cache alignment.
The commit c6f53ed8f213a ("mm, memcg: cg2 memory{.swap,}.peak write
handlers") added a field to the struct page_counter and accidently
increased its size.  Let's move the failcnt field which is v1-only field
to the same cacheline of usage to reduce the size of struct page_counter.

Link: https://lkml.kernel.org/r/20250228075808.207484-4-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agopage_counter: track failcnt only for legacy cgroups
Shakeel Butt [Fri, 28 Feb 2025 07:58:07 +0000 (23:58 -0800)]
page_counter: track failcnt only for legacy cgroups

Currently page_counter tracks failcnt for counters used by v1 and v2
controllers.  However failcnt is only exported for v1 deployment and thus
there is no need to maintain it in v2.  The oom report does expose failcnt
for memory and swap in v2 but v2 already maintains MEMCG_MAX and
MEMCG_SWAP_MAX event counters which can be used.

Link: https://lkml.kernel.org/r/20250228075808.207484-3-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomemcg: don't call propagate_protected_usage() for v1
Shakeel Butt [Fri, 28 Feb 2025 07:58:06 +0000 (23:58 -0800)]
memcg: don't call propagate_protected_usage() for v1

Patch series "page_counter cleanup and size reduction".

Commit c6f53ed8f213a ("mm, memcg: cg2 memory{.swap,}.peak write handlers")
accidently increased the size of struct page_counter.  This series
rearrange the fields to reduce its size and also has some cleanups.

This patch (of 3):

Memcg-v1 does not support memory protection (min/low) and thus there is no
need to track protected memory usage for it.

Link: https://lkml.kernel.org/r/20250228075808.207484-1-shakeel.butt@linux.dev
Link: https://lkml.kernel.org/r/20250228075808.207484-2-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/page_alloc: clarify should_claim_block() commentary
Brendan Jackman [Fri, 28 Feb 2025 09:52:18 +0000 (09:52 +0000)]
mm/page_alloc: clarify should_claim_block() commentary

There's lots of text here but it's a little hard to follow, this is an
attempt to break it up and align its structure more closely with the code.

Reword the top-level function comment to just explain what question the
function answers from the point of view of the caller.

Break up the internal logic into different sections that can have their
own commentary describing why that part of the rationale is present.

Note the page_group_by_mobility_disabled logic is not explained in the
commentary, that is outside the scope of this patch...

Link: https://lkml.kernel.org/r/20250228-clarify-steal-v4-2-cb2ef1a4e610@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/page_alloc: clarify terminology in migratetype fallback code
Brendan Jackman [Fri, 28 Feb 2025 09:52:17 +0000 (09:52 +0000)]
mm/page_alloc: clarify terminology in migratetype fallback code

Patch series "mm/page_alloc: Some clarifications for migratetype
fallback", v4.

A couple of patches to try and make the code easier to follow.

This patch (of 2):

This code is rather confusing because:

 1. "Steal" is sometimes used to refer to the general concept of
    allocating from a from a block of a fallback migratetype
    (steal_suitable_fallback()) but sometimes it refers specifically to
    converting a whole block's migratetype (can_steal_fallback()).

 2. can_steal_fallback() sounds as though it's answering the question "am
    I functionally permitted to allocate from that other type" but in
    fact it is encoding a heuristic preference.

 3. The same piece of data has different names in different places:
    can_steal vs whole_block. This reinforces point 2 because it looks
    like the different names reflect a shift in intent from "am I
    allowed to steal" to "do I want to steal", but no such shift exists.

Fix 1. by avoiding the term "steal" in ambiguous contexts. Start using
the term "claim" to refer to the special case of stealing the entire
block.

Fix 2. by using "should" instead of "can", and also rename its
parameters and add some commentary to make it more explicit what they
mean.

Fix 3. by adopting the new "claim" terminology universally for this
set of variables.

Link: https://lkml.kernel.org/r/20250228-clarify-steal-v4-0-cb2ef1a4e610@google.com
Link: https://lkml.kernel.org/r/20250228-clarify-steal-v4-1-cb2ef1a4e610@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoRevert "x86/xen: allow nesting of same lazy mode"
Ryan Roberts [Mon, 3 Mar 2025 14:15:39 +0000 (14:15 +0000)]
Revert "x86/xen: allow nesting of same lazy mode"

Commit 49147beb0ccb ("x86/xen: allow nesting of same lazy mode") was added
as a solution for a core-mm code change where
arch_[enter|leave]_lazy_mmu_mode() started to be called in a nested
manner; see commit bcc6cc832573 ("mm: add default definition of
set_ptes()").

However, now that we have fixed the API to avoid nesting, we no longer
need this capability in the x86 implementation.

Additionally, from code review, I don't believe the fix was ever robust in
the case of preemption occurring while in the nested lazy mode.  The
implementation usually deals with preemption by calling
arch_leave_lazy_mmu_mode() from xen_start_context_switch() for the
outgoing task if we are in the lazy mmu mode.  Then in
xen_end_context_switch(), it restarts the lazy mode by calling
arch_enter_lazy_mmu_mode() for an incoming task that was in the lazy mode
when it was switched out.  But arch_leave_lazy_mmu_mode() will only unwind
a single level of nesting.  If we are in the double nest, then it's not
fully unwound and per-cpu variables are left in a bad state.

So the correct solution is to remove the possibility of nesting from the
higher level (which has now been done) and remove this x86-specific
solution.

Link: https://lkml.kernel.org/r/20250303141542.3371656-6-ryan.roberts@arm.com
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Acked-by: David Hildenbrand <david@redhat.com>
Acked-by: Juergen Gross <jgross@suse.com>
Cc: Andreas Larsson <andreas@gaisler.com>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Juegren Gross <jgross@suse.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Thomas Gleinxer <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agosparc/mm: avoid calling arch_enter/leave_lazy_mmu() in set_ptes
Ryan Roberts [Mon, 3 Mar 2025 14:15:38 +0000 (14:15 +0000)]
sparc/mm: avoid calling arch_enter/leave_lazy_mmu() in set_ptes

With commit 1a10a44dfc1d ("sparc64: implement the new page table range
API") set_ptes was added to the sparc architecture.  The implementation
included calling arch_enter/leave_lazy_mmu() calls.

The patch removes the usage of arch_enter/leave_lazy_mmu() since this
implies nesting of lazy mmu regions which is not supported.  Without this
fix, lazy mmu mode is effectively disabled because we exit the mode after
the first set_ptes:

remap_pte_range()
  -> arch_enter_lazy_mmu()
  -> set_ptes()
      -> arch_enter_lazy_mmu()
      -> arch_leave_lazy_mmu()
  -> arch_leave_lazy_mmu()

Powerpc suffered the same problem and fixed it in a corresponding way with
commit 47b8def9358c ("powerpc/mm: Avoid calling
arch_enter/leave_lazy_mmu() in set_ptes").

Link: https://lkml.kernel.org/r/20250303141542.3371656-5-ryan.roberts@arm.com
Fixes: 1a10a44dfc1d ("sparc64: implement the new page table range API")
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Acked-by: David Hildenbrand <david@redhat.com>
Acked-by: Andreas Larsson <andreas@gaisler.com>
Acked-by: Juergen Gross <jgross@suse.com>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Juegren Gross <jgross@suse.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Thomas Gleinxer <tglx@linutronix.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agosparc/mm: disable preemption in lazy mmu mode
Ryan Roberts [Mon, 3 Mar 2025 14:15:37 +0000 (14:15 +0000)]
sparc/mm: disable preemption in lazy mmu mode

Since commit 38e0edb15bd0 ("mm/apply_to_range: call pte function with lazy
updates") it's been possible for arch_[enter|leave]_lazy_mmu_mode() to be
called without holding a page table lock (for the kernel mappings case),
and therefore it is possible that preemption may occur while in the lazy
mmu mode.  The Sparc lazy mmu implementation is not robust to preemption
since it stores the lazy mode state in a per-cpu structure and does not
attempt to manage that state on task switch.

Powerpc had the same issue and fixed it by explicitly disabling preemption
in arch_enter_lazy_mmu_mode() and re-enabling in
arch_leave_lazy_mmu_mode().  See commit b9ef323ea168 ("powerpc/64s:
Disable preemption in hash lazy mmu mode").

Given Sparc's lazy mmu mode is based on powerpc's, let's fix it in the
same way here.

Link: https://lkml.kernel.org/r/20250303141542.3371656-4-ryan.roberts@arm.com
Fixes: 38e0edb15bd0 ("mm/apply_to_range: call pte function with lazy updates")
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Acked-by: David Hildenbrand <david@redhat.com>
Acked-by: Andreas Larsson <andreas@gaisler.com>
Acked-by: Juergen Gross <jgross@suse.com>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Juegren Gross <jgross@suse.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Thomas Gleinxer <tglx@linutronix.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agofs/proc/task_mmu: reduce scope of lazy mmu region
Ryan Roberts [Mon, 3 Mar 2025 14:15:36 +0000 (14:15 +0000)]
fs/proc/task_mmu: reduce scope of lazy mmu region

Update the way arch_[enter|leave]_lazy_mmu_mode() is called in
pagemap_scan_pmd_entry() to follow the normal pattern of holding the ptl
for user space mappings.  As a result the scope is reduced to only the pte
table, but that's where most of the performance win is.

While I believe there wasn't technically a bug here, the original scope
made it easier to accidentally nest or, worse, accidentally call something
like kmap() which would expect an immediate mode pte modification but it
would end up deferred.

Link: https://lkml.kernel.org/r/20250303141542.3371656-3-ryan.roberts@arm.com
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Acked-by: David Hildenbrand <david@redhat.com>
Acked-by: Juergen Gross <jgross@suse.com>
Cc: Andreas Larsson <andreas@gaisler.com>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Juegren Gross <jgross@suse.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Thomas Gleinxer <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm: fix lazy mmu docs and usage
Ryan Roberts [Mon, 3 Mar 2025 14:15:35 +0000 (14:15 +0000)]
mm: fix lazy mmu docs and usage

Patch series "Fix lazy mmu mode", v2.

I'm planning to implement lazy mmu mode for arm64 to optimize vmalloc.  As
part of that, I will extend lazy mmu mode to cover kernel mappings in
vmalloc table walkers.  While lazy mmu mode is already used for kernel
mappings in a few places, this will extend it's use significantly.

Having reviewed the existing lazy mmu implementations in powerpc, sparc
and x86, it looks like there are a bunch of bugs, some of which may be
more likely to trigger once I extend the use of lazy mmu.  So this series
attempts to clarify the requirements and fix all the bugs in advance of
that series.  See patch #1 commit log for all the details.

This patch (of 5):

The docs, implementations and use of arch_[enter|leave]_lazy_mmu_mode() is
a bit of a mess (to put it politely).  There are a number of issues
related to nesting of lazy mmu regions and confusion over whether the
task, when in a lazy mmu region, is preemptible or not.  Fix all the
issues relating to the core-mm.  Follow up commits will fix the
arch-specific implementations.  3 arches implement lazy mmu; powerpc,
sparc and x86.

When arch_[enter|leave]_lazy_mmu_mode() was first introduced by commit
6606c3e0da53 ("[PATCH] paravirt: lazy mmu mode hooks.patch"), it was
expected that lazy mmu regions would never nest and that the appropriate
page table lock(s) would be held while in the region, thus ensuring the
region is non-preemptible.  Additionally lazy mmu regions were only used
during manipulation of user mappings.

Commit 38e0edb15bd0 ("mm/apply_to_range: call pte function with lazy
updates") started invoking the lazy mmu mode in apply_to_pte_range(),
which is used for both user and kernel mappings.  For kernel mappings the
region is no longer protected by any lock so there is no longer any
guarantee about non-preemptibility.  Additionally, for RT configs, the
holding the PTL only implies no CPU migration, it doesn't prevent
preemption.

Commit bcc6cc832573 ("mm: add default definition of set_ptes()") added
arch_[enter|leave]_lazy_mmu_mode() to the default implementation of
set_ptes(), used by x86.  So after this commit, lazy mmu regions can be
nested.  Additionally commit 1a10a44dfc1d ("sparc64: implement the new
page table range API") and commit 9fee28baa601 ("powerpc: implement the
new page table range API") did the same for the sparc and powerpc
set_ptes() overrides.

powerpc couldn't deal with preemption so avoids it in commit b9ef323ea168
("powerpc/64s: Disable preemption in hash lazy mmu mode"), which
explicitly disables preemption for the whole region in its implementation.
x86 can support preemption (or at least it could until it tried to add
support nesting; more on this below).  Sparc looks to be totally broken in
the face of preemption, as far as I can tell.

powerpc can't deal with nesting, so avoids it in commit 47b8def9358c
("powerpc/mm: Avoid calling arch_enter/leave_lazy_mmu() in set_ptes"),
which removes the lazy mmu calls from its implementation of set_ptes().
x86 attempted to support nesting in commit 49147beb0ccb ("x86/xen: allow
nesting of same lazy mode") but as far as I can tell, this breaks its
support for preemption.

In short, it's all a mess; the semantics for
arch_[enter|leave]_lazy_mmu_mode() are not clearly defined and as a result
the implementations all have different expectations, sticking plasters and
bugs.

arm64 is aiming to start using these hooks, so let's clean everything up
before adding an arm64 implementation.  Update the documentation to state
that lazy mmu regions can never be nested, must not be called in interrupt
context and preemption may or may not be enabled for the duration of the
region.  And fix the generic implementation of set_ptes() to avoid
nesting.

arch-specific fixes to conform to the new spec will proceed this one.

These issues were spotted by code review and I have no evidence of issues
being reported in the wild.

Link: https://lkml.kernel.org/r/20250303141542.3371656-1-ryan.roberts@arm.com
Link: https://lkml.kernel.org/r/20250303141542.3371656-2-ryan.roberts@arm.com
Fixes: bcc6cc832573 ("mm: add default definition of set_ptes()")
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Acked-by: David Hildenbrand <david@redhat.com>
Acked-by: Juergen Gross <jgross@suse.com>
Cc: Andreas Larsson <andreas@gaisler.com>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Juegren Gross <jgross@suse.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Thomas Gleinxer <tglx@linutronix.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoDocs/admin-guide/mm/damon/usage: add intervals_goal directory on the hierarchy
SeongJae Park [Mon, 3 Mar 2025 22:17:26 +0000 (14:17 -0800)]
Docs/admin-guide/mm/damon/usage: add intervals_goal directory on the hierarchy

Document DAMON sysfs interface usage for DAMON sampling and aggregation
intervals auto-tuning.

Link: https://lkml.kernel.org/r/20250303221726.484227-9-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoDocs/ABI/damon: document intervals auto-tuning ABI
SeongJae Park [Mon, 3 Mar 2025 22:17:25 +0000 (14:17 -0800)]
Docs/ABI/damon: document intervals auto-tuning ABI

Document the DAMON user-space ABI for DAMON sampling and aggregation
intervals auto-tuning.

Link: https://lkml.kernel.org/r/20250303221726.484227-8-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoDocs/mm/damon/design: document for intervals auto-tuning
SeongJae Park [Mon, 3 Mar 2025 22:17:24 +0000 (14:17 -0800)]
Docs/mm/damon/design: document for intervals auto-tuning

Document the design of DAMON sampling and aggregation intervals
auto-tuning.

[sj@kernel.org: fix a typo on 'intervals auto-tuning' section]
Link: https://lkml.kernel.org/r/20250305182744.56125-1-sj@kernel.org
Link: https://lkml.kernel.org/r/20250303221726.484227-7-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/damon/sysfs: implement a command to update auto-tuned monitoring intervals
SeongJae Park [Mon, 3 Mar 2025 22:17:23 +0000 (14:17 -0800)]
mm/damon/sysfs: implement a command to update auto-tuned monitoring intervals

DAMON kernel API callers can show auto-tuned sampling and aggregation
intervals from the monmitoring attributes data structure.  That can be
useful for debugging or tuning of the feature.  DAMON user-space ABI users
has no way to see that, though.  Implement a new DAMON sysfs interface
command, namely 'update_tuned_intervals', for the purpose.  If the command
is written to the kdamond state file, the tuned sampling and aggregation
intervals will be updated to the corresponding sysfs interface files.

Link: https://lkml.kernel.org/r/20250303221726.484227-6-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/damon/sysfs: commit intervals tuning goal
SeongJae Park [Mon, 3 Mar 2025 22:17:22 +0000 (14:17 -0800)]
mm/damon/sysfs: commit intervals tuning goal

Connect DAMON sysfs interface for sampling and aggregation intervals
auto-tuning with DAMON core API, so that users can really use the feature
using the sysfs files.

Link: https://lkml.kernel.org/r/20250303221726.484227-5-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/damon/sysfs: implement intervals tuning goal directory
SeongJae Park [Mon, 3 Mar 2025 22:17:21 +0000 (14:17 -0800)]
mm/damon/sysfs: implement intervals tuning goal directory

Implement DAMON sysfs interface directory and its files for setting DAMON
sampling and aggregation intervals auto-tuning goal.

Link: https://lkml.kernel.org/r/20250303221726.484227-4-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/damon/core: implement intervals auto-tuning
SeongJae Park [Mon, 3 Mar 2025 22:17:20 +0000 (14:17 -0800)]
mm/damon/core: implement intervals auto-tuning

Implement the DAMON sampling and aggregation intervals auto-tuning
mechanism as briefly described on 'struct damon_intervals_goal'.  The core
part for deciding the direction and amount of the changes is implemented
reusing the feedback loop function which is being used for DAMOS quotas
auto-tuning.  Unlike the DAMOS quotas auto-tuning use case, limit the
maximum decreasing amount after the adjustment to 50% of the current
value, though.  This is because the intervals have no good merits at rapid
reductions since it could unnecessarily increase the monitoring overhead.

Link: https://lkml.kernel.org/r/20250303221726.484227-3-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/damon: add data structure for monitoring intervals auto-tuning
SeongJae Park [Mon, 3 Mar 2025 22:17:19 +0000 (14:17 -0800)]
mm/damon: add data structure for monitoring intervals auto-tuning

Patch series "mm/damon: auto-tune aggregation interval".

DAMON requires time-consuming and repetitive aggregation interval tuning.
Introduce a feature for automating it using a feedback loop that aims an
amount of observed access events, like auto-exposing cameras.

Background: Access Frequency Monitoring and Aggregation Interval
================================================================

DAMON checks if each memory element (damon_region) is accessed or not for
every user-specified time interval called 'sampling interval'.  It
aggregates the check intervals on per-element counter called
'nr_accesses'.  DAMON users can read the counters to get the access
temperature of a given element.  The counters are reset for every another
user-specified time interval called 'aggregation interval'.

This can be illustrated as DAMON continuously capturing a snapshot of
access events that happen and captured within the last aggregation
interval.  This implies the aggregation interval plays a key role for the
quality of the snapshots, like the camera exposure time.  If it is too
short, the amount of access events that happened and captured for each
snapshot is small, so each snapshot will show no many interesting things
but just a cold and dark world with hopefuly one pale blue dot or two.  If
it is too long, too many events are aggregated in a single shot, so each
snapshot will look like world of flames, or Muspellheim.  It will be
difficult to find practical insights in both cases.

Problem: Time Consuming and Repetitive Tuning
=============================================

The appropriate length of the aggregation interval depends on how
frequently the system and workloads are making access events that DAMON
can observe.  Hence, users have to tune the interval with excessive amount
of tests with the target system and workloads.  If the system and
workloads are changed, the tuning should be done again.  If the
characteristic of the workloads is dynamic, it becomes more challenging.
It is therefore time-consuming and repetitive.

The tuning challenge mainly stems from the wrong question.  It is not
asking users what quality of monitoring results they want, but how DAMON
should operate for their hidden goal.  To make the right answer, users
need to fully understand DAMON's mechanisms and the characteristics of
their workloads.  Users shouldn't be asked to understand the underlying
mechanism.  Understanding the characteristics of the workloads shouldn't
be the role of users but DAMON.

Aim-oriented Feedback-driven Auto-Tuning
=========================================

Fortunately, the appropriate length of the aggregation interval can be
inferred using a feedback loop.  If the current snapshots are showing no
much intresting information, in other words, if it shows only rare access
events, increasing the aggregation interval helps, and vice versa.  We
tested this theory on a few real-world workloads, and documented one of
the experience with an official DAMON monitoring intervals tuning
guideline.  Since it is a simple theory that requires repeatable tries, it
can be a good job for machines.

Based on the guideline's theory, we design an automation of aggregation
interval tuning, in a way similar to that of camera auto-exposure feature.
It defines the amount of interesting information as the ratio of
DAMON-observed access events that DAMON actually observed to theoretical
maximum amount of it within each snapshot.  Events are accounted in byte
and sampling attempts granularity.  For example, let's say there is a
region of 'X' bytes size.  DAMON tried access check smapling for the
region 'Y' times in total for a given aggregation.  Among the 'Y'
attempts, 'Z' times it shown positive results.  Then, the theoritical
maximum number of access events for the region is 'X * Y'.  And the number
of access events that DAMON has observed for the region is 'X * Z'.  The
abount of the interesting information is '(X * Z / X * Y)'.  Note that
each snapshot would have multiple regions.

Users can set an arbitrary value of the ratio as their target.  Once the
target is set, the automation periodically measures the current value of
the ratio and increase or decrease the aggregation interval if the ratio
value is lower or higher than the target.  The amount of the change is
proportion to the distance between the current adn the target values.

To avoid auto-tuning goes too long way, let users set the minimum and the
maximum aggregation interval times.  Changing only aggregation interval
while sampling interval is kept makes the maximum level of access
frequency in each snapshot, or discernment of regions inconsistent.  Also,
unnecessarily short sampling interval causes meaningless monitoring
overhed.  The automation therefore adjusts the sampling interval together
with aggregation interval, while keeping the ratio between the two
intervals.  Users can set the ratio, or the discernment.

Discussion
==========

The modified question (aimed amount of access events, or lights, in each
snapshot) is easy to answer by both the users and the kernel.  If users
are interested in finding more cold regions, the value should be lower,
and vice versa.  If users have no idea, kernel can suggest a fair default
value based on some theories and experiments.  For example, based on the
Pareto principle (80/20 rule), we could expect 20% target ratio will
capture 80% of real access events.  Since 80% might be too high, applying
the rule once again, 4% (20% * 20%) may capture about 56% (80% * 80%) of
real access events.

Sampling to aggregation intervals ratio and min/max aggregation intervals
are also arguably easy to answer.  What users want is discernment of
regions for efficient system operation, for examples, X amount of colder
regions or Y amount of warmer regions, not exactly how many times each
cache line is accessed in nanoseconds degree.  The appropriate min/max
aggregation interval can relatively naively set, and may better to set for
aimed monitoring overhead.  Since sampling interval is directly deciding
the overhead, setting it based on the sampling interval can be easy.  With
my experiences, I'd argue the intervals ratio 0.05, and 5 milliseconds to
20 seconds sampling interval range (100 milliseconds to 400 seconds
aggregation interval) can be a good default suggestion.

Evaluation
==========

On a machine running a real world server workload, I ran DAMON to monitor
its physical address space for about 23 hours, with this feature turned
on.  We set it to tune sampling interval in a range from 5 milliseconds to
10 seconds, aiming 4 % DAMON-observed access ratio per three aggregation
intervals.  The exact command I used is as below.

    damo start --monitoring_intervals_goal 4% 3 5ms 10s --damos_action stat

During the test run, DAMON continuously updated sampling and aggregation
intervals as designed, within the given range.  For all the time, DAMON
was able to find the intervals that meets the target access events ratio
in the given intervals range (sampling interval between 5 milliseconds and
10 seconds).

For most of the time, tuned sampling interval was converged in 300-400
milliseconds.  It made only small amount of changes within the range.  The
average of the tuned sampling interval during the test was about 380
milliseconds.

The workload periodically gets less load and decreases its CPU usage.
Presumably this also caused it making less memory access events.
Reactively to such event,s DAMON also increased the intervals as expected.
It was still able to find the optimum interval that satisfying the target
access ratio within the given intervals range.  Usually it was converged
to about 5 seconds.  Once the workload gets normal amount of load again,
DAMON reactively reduced the intervals to the normal range.

I collected and visualized DAMON's monitoring results on the server a few
times.  Every time the visualized access pattern looked not biased to only
cold or hot pages but diverse and balanced.  Let me show some of the
snapshots that I collected at the nearly end of the test (after about 23
hours have passed since starting DAMON on the server).

The recency histogram looks as below.  Please note that this visualization
shows only a very coarse grained information.  For more details about the
visualization format, please refer to DAMON user-space tool
documentation[1].

    # ./damo report access --style recency-sz-hist --tried_regions_of 0 0 0 --access_rate 0 0
    <last accessed time (us)> <total size>
    [-19 h 7 m 45.514 s, -17 h 12 m 58.963 s)  6.198 GiB  |****                |
    [-17 h 12 m 58.963 s, -15 h 18 m 12.412 s) 0 B        |                    |
    [-15 h 18 m 12.412 s, -13 h 23 m 25.860 s) 0 B        |                    |
    [-13 h 23 m 25.860 s, -11 h 28 m 39.309 s) 0 B        |                    |
    [-11 h 28 m 39.309 s, -9 h 33 m 52.757 s)  0 B        |                    |
    [-9 h 33 m 52.757 s, -7 h 39 m 6.206 s)    0 B        |                    |
    [-7 h 39 m 6.206 s, -5 h 44 m 19.654 s)    0 B        |                    |
    [-5 h 44 m 19.654 s, -3 h 49 m 33.103 s)   0 B        |                    |
    [-3 h 49 m 33.103 s, -1 h 54 m 46.551 s)   0 B        |                    |
    [-1 h 54 m 46.551 s, -0 ns)                16.967 GiB |*********           |
    [-0 ns, --6886551440000 ns)                38.835 GiB |********************|
    memory bw estimate: 9.425 GiB per second
    total size: 62.000 GiB

It shows about 38 GiB of memory was accessed at least once within last
aggregation interval (given ~300 milliseconds tuned sampling interval,
this is about six seconds).  This is about 61 % of the total memory.  In
other words, DAMON found warmest 61 % memory of the system.  The number is
particularly interesting given our Pareto principle based theory for the
tuning goal value.  We set it as 20 % of 20 % (4 %), thinking it would
capture 80 % of 80 % (64 %) real access events.  And it foudn 61 % hot
memory, or working set.  Nevertheless, to make the theory clearer, much
more discussion and tests would be needed.  At the moment, nonetheless, we
can say making the target value higher helps finding more hot memory
regions.

The histogram also shows an amount of cold memory.  About 17 GiB memory of
the system has not accessed at least for last aggregation interval (about
six seconds), and at most for about last two hours.  The real longest
unaccessed time of the 17 GiB memory was about 19 minutes, though.  This
is a limitation of this visualization format.

It further found very cold 6 GiB memory.  It has not accessed at least for
last 17 hours and at most 19 hours.

What about hot memory distribution?  To see this, I capture and visualize
the snapshot in access temperature histogram.  Again, please refer to the
DAMON user-space tool documentation[1] for the format and what access
temperature mean.  Both the visualization and metric shows only very
coarse grained and limited information.  The resulting histogram look like
below.

    # ./damo report access --style temperature-sz-hist --tried_regions_of 0 0 0
    <temperature> <total size>
    [-6,840,763,776,000, -5,501,580,939,800) 6.198 GiB  |***                 |
    [-5,501,580,939,800, -4,162,398,103,600) 0 B        |                    |
    [-4,162,398,103,600, -2,823,215,267,400) 0 B        |                    |
    [-2,823,215,267,400, -1,484,032,431,200) 0 B        |                    |
    [-1,484,032,431,200, -144,849,595,000)   0 B        |                    |
    [-144,849,595,000, 1,194,333,241,200)    55.802 GiB |********************|
    [1,194,333,241,200, 2,533,516,077,400)   4.000 KiB  |*                   |
    [2,533,516,077,400, 3,872,698,913,600)   4.000 KiB  |*                   |
    [3,872,698,913,600, 5,211,881,749,800)   8.000 KiB  |*                   |
    [5,211,881,749,800, 6,551,064,586,000)   12.000 KiB |*                   |
    [6,551,064,586,000, 7,890,247,422,200)   4.000 KiB  |*                   |
    memory bw estimate: 5.178 GiB per second
    total size: 62.000 GiB

We can see most of the memory is in similar access temperature range, and
definitely some pages are extremely hot.

To see the picture in more detail, let's capture and visualize the
snapshot per DAMON-region, sorted by their access temperature.  The total
number of the regions was about 300.  Due to the limited space, I'm
showing only a few parts of the output here.

    # ./damo report access --style hot --tried_regions_of 0 0 0
    heatmap: 00000000888888889999999888888888888888888888888888888888888888888888888888888888
    # min/max temperatures: -6,827,258,184,000, 17,589,052,500, column size: 793.600 MiB
     |999999999999999999999999999999999999999| 4.000 KiB   access 100 % 18 h 9 m 43.918 s
     |999999999999999999999999999999999999999| 8.000 KiB   access 100 % 17 h 56 m 5.351 s
     |999999999999999999999999999999999999999| 4.000 KiB   access 100 % 15 h 24 m 19.634 s
     |999999999999999999999999999999999999999| 4.000 KiB   access 100 % 14 h 10 m 55.606 s
     |999999999999999999999999999999999999999| 4.000 KiB   access 100 % 11 h 34 m 18.993 s
    [...]
               |99999999999999999999999999999| 8.000 KiB   access 100 % 1 m 27.945 s
               |11111111111111111111111111111| 80.000 KiB  access 15 %  1 m 21.180 s
               |00000000000000000000000000000| 24.000 KiB  access 5 %   1 m 21.180 s
               |00000000000000000000000000000| 5.919 GiB   access 10 %  1 m 14.415 s
               |99999999999999999999999999999| 12.000 KiB  access 100 % 1 m 7.650 s
    [...]
                                           |0| 4.000 KiB   access 5 %   0 ns
                                           |0| 12.000 KiB  access 5 %   0 ns
                                           |0| 188.000 KiB access 0 %   0 ns
                                           |0| 24.000 KiB  access 0 %   0 ns
                                           |0| 48.000 KiB  access 0 %   0 ns
    [...]
             |0000000000000000000000000000000| 8.000 KiB   access 0 %   6 m 45.901 s
            |00000000000000000000000000000000| 36.000 KiB  access 0 %   7 m 26.491 s
            |00000000000000000000000000000000| 4.000 KiB   access 0 %   12 m 37.682 s
           |000000000000000000000000000000000| 8.000 KiB   access 0 %   18 m 9.168 s
           |000000000000000000000000000000000| 16.000 KiB  access 0 %   19 m 3.288 s
    |0000000000000000000000000000000000000000| 6.198 GiB   access 0 %   18 h 57 m 52.582 s
    memory bw estimate: 8.798 GiB per second
    total size: 62.000 GiB

We can see DAMON found small and extremely hot regions that accessed for
all access check sampling (once per about 300 milliseconds) for more than
10 hours.  The access temperature rapidly decreases.  DAMON was also able
to find small and big regions that not accessed for up to about 19
minutes.  It even found an outlier cold region of 6 GiB that not accessed
for about 19 hours.  It is unclear what the outlier region is, as of this
writing.

For the testing, DAMON was consuming about 0.1% of single CPU time.  This
is again expected results, since DAMON was using about 370 milliseconds
sampling interval in most case.

    # ps -p $kdamond_pid -o %cpu
    %CPU
     0.1

I also ran similar tests against kernel build workload and an in-memory
cache workload benchmark[2].  Detialed results including tuned intervals
and captured access pattern were of course different sicne those depend on
the workloads.  But the auto-tuning feature was always working as expected
like the above results for the real world workload.

To wrap up, with intervals auto-tuning feature, DAMON was able to capture
access pattern snapshots of a quality on a real world server workload.
The auto-tuning feature was able to adaptively react to the dynamic access
patterns of the workload and reliably provide consistent monitoring
results without manual human interventions.  Also, the auto-tuning made
DAMON consumes only necessary amount of resource for the required quality.

References
==========

[1] https://github.com/damonitor/damo/blob/next/USAGE.md#access-report-styles
[2] https://github.com/facebookresearch/DCPerf/blob/main/packages/tao_bench/README.md

This patch (of 8):

Add data structures for DAMON sampling and aggregation intervals automatic
tuning that aims specific amount of DAMON-observed access events per
snapshot.  In more detail, define the data structure for the tuning goal,
link it to the monitoring attributes data structure so that DAMON kernel
API callers can make the request, and update parameters setup DAMON
function to respect the new parameter.

Link: https://lkml.kernel.org/r/20250303221726.484227-1-sj@kernel.org
Link: https://lkml.kernel.org/r/20250303221726.484227-2-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/list_lru: make the case where mlru is NULL as unlikely
Zeng Jingxiang [Thu, 27 Feb 2025 08:22:23 +0000 (16:22 +0800)]
mm/list_lru: make the case where mlru is NULL as unlikely

In the following memcg_list_lru_alloc() function, mlru here is almost
always NULL, so in most cases this should save a function call, mark mlru
as unlikely to optimize the code, and reusing the mlru for the next
attempt when the tree insertion fails.

        do {
                xas_lock_irqsave(&xas, flags);
                if (!xas_load(&xas) && !css_is_dying(&pos->css)) {
                        xas_store(&xas, mlru);
                        if (!xas_error(&xas))
                                mlru = NULL;
                }
                xas_unlock_irqrestore(&xas, flags);
        } while (xas_nomem(&xas, GFP_KERNEL));
>       if (mlru)
                kfree(mlru);

Link: https://lkml.kernel.org/r/20250227082223.1173847-1-jingxiangzeng.cas@gmail.com
Signed-off-by: Zeng Jingxiang <linuszeng@tencent.com>
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202412290924.UTP7GH2Z-lkp@intel.com/
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Muchun Song <muchun.song@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Chengming Zhou <chengming.zhou@linux.dev>
Cc: Jingxiang Zeng <linuszeng@tencent.com>
Cc: Kairui Song <kasong@tencent.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm: rename GENERIC_PTDUMP and PTDUMP_CORE
Anshuman Khandual [Wed, 26 Feb 2025 12:24:04 +0000 (17:54 +0530)]
mm: rename GENERIC_PTDUMP and PTDUMP_CORE

Platforms subscribe into generic ptdump implementation via GENERIC_PTDUMP.
But generic ptdump gets enabled via PTDUMP_CORE.  These configs
combination is confusing as they sound very similar and does not
differentiate between platform's feature subscription and feature
enablement for ptdump.  Rename the configs as ARCH_HAS_PTDUMP and PTDUMP
making it more clear and improve readability.

Link: https://lkml.kernel.org/r/20250226122404.1927473-6-anshuman.khandual@arm.com
Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
Reviewed-by: Christophe Leroy <christophe.leroy@csgroup.eu> (powerpc)
Acked-by: Catalin Marinas <catalin.marinas@arm.com> [arm64]
Cc: Will Deacon <will@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Steven Price <steven.price@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm: make DEBUG_WX depdendent on GENERIC_PTDUMP
Anshuman Khandual [Wed, 26 Feb 2025 12:24:03 +0000 (17:54 +0530)]
mm: make DEBUG_WX depdendent on GENERIC_PTDUMP

DEBUG_WX selects PTDUMP_CORE without even ensuring that the given platform
implements GENERIC_PTDUMP.  This problem has been latent until now, as all
the platforms subscribing ARCH_HAS_DEBUG_WX also subscribe GENERIC_PTDUMP.

Link: https://lkml.kernel.org/r/20250226122404.1927473-5-anshuman.khandual@arm.com
Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Reviewed-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agodocs: arm64: drop PTDUMP config options from ptdump.rst
Anshuman Khandual [Wed, 26 Feb 2025 12:24:02 +0000 (17:54 +0530)]
docs: arm64: drop PTDUMP config options from ptdump.rst

Both GENERIC_PTDUMP and PTDUMP_CORE are not user selectable config
options.  Just drop these from documentation.

Link: https://lkml.kernel.org/r/20250226122404.1927473-4-anshuman.khandual@arm.com
Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
Suggested-by: Steven Price <steven.price@arm.com>
Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoarch/powerpc: drop GENERIC_PTDUMP from mpc885_ads_defconfig
Anshuman Khandual [Wed, 26 Feb 2025 12:24:01 +0000 (17:54 +0530)]
arch/powerpc: drop GENERIC_PTDUMP from mpc885_ads_defconfig

GENERIC_PTDUMP gets selected on powerpc explicitly and hence can be
dropped off from mpc885_ads_defconfig.  Replace with CONFIG_PTDUMP_DEBUGFS
instead.

Link: https://lkml.kernel.org/r/20250226122404.1927473-3-anshuman.khandual@arm.com
Fixes: e084728393a5 ("powerpc/ptdump: Convert powerpc to GENERIC_PTDUMP")
Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
Suggested-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Reviewed-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Steven Price <steven.price@arm.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoconfigs: drop GENERIC_PTDUMP from debug.config
Anshuman Khandual [Wed, 26 Feb 2025 12:24:00 +0000 (17:54 +0530)]
configs: drop GENERIC_PTDUMP from debug.config

Patch series "mm: Rework generic PTDUMP configs", v3.

The series reworks generic PTDUMP configs before eventually renaming them
after some basic cleanups first.

This patch (of 5):

The platforms that support GENERIC_PTDUMP select the config explicitly.
But enabling this feature on platforms that don't really support - does
nothing or might cause a build failure.  Hence just drop GENERIC_PTDUMP
from generic debug.config

Link: https://lkml.kernel.org/r/20250226122404.1927473-1-anshuman.khandual@arm.com
Link: https://lkml.kernel.org/r/20250226122404.1927473-2-anshuman.khandual@arm.com
Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/mmu_notifier: use MMU_NOTIFY_CLEAR in remove_device_exclusive_entry()
David Hildenbrand [Wed, 26 Feb 2025 13:22:57 +0000 (14:22 +0100)]
mm/mmu_notifier: use MMU_NOTIFY_CLEAR in remove_device_exclusive_entry()

Let's limit the use of MMU_NOTIFY_EXCLUSIVE to the case where we convert a
present PTE to device-exclusive.  For the other case, we can simply use
MMU_NOTIFY_CLEAR, because it really is clearing the device-exclusive entry
first, to then install the present entry.

Update the documentation of MMU_NOTIFY_EXCLUSIVE, to document the single
use case more thoroughly.

If ever required, we could add a separate MMU_NOTIFY_CLEAR_EXCLUSIVE; for
now using MMU_NOTIFY_CLEAR seems to be sufficient.

Link: https://lkml.kernel.org/r/20250226132257.2826043-6-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/memory: document restore_exclusive_pte()
David Hildenbrand [Wed, 26 Feb 2025 13:22:56 +0000 (14:22 +0100)]
mm/memory: document restore_exclusive_pte()

Let's document how this function is to be used, and why the folio lock is
involved.

Link: https://lkml.kernel.org/r/20250226132257.2826043-5-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/memory: pass folio and pte to restore_exclusive_pte()
David Hildenbrand [Wed, 26 Feb 2025 13:22:55 +0000 (14:22 +0100)]
mm/memory: pass folio and pte to restore_exclusive_pte()

Let's pass the folio and the pte to restore_exclusive_pte(), so we can
avoid repeated page_folio() and ptep_get().  To do that, pass the pte to
try_restore_exclusive_pte() and use a folio in there already.

While at it, just avoid the "swp_entry_t entry" variable in
try_restore_exclusive_pte() and add a folio-locked check to
restore_exclusive_pte().

Link: https://lkml.kernel.org/r/20250226132257.2826043-4-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Alistair Popple <apopple@nvidia.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/memory: remove PageAnonExclusive sanity-check in restore_exclusive_pte()
David Hildenbrand [Wed, 26 Feb 2025 13:22:54 +0000 (14:22 +0100)]
mm/memory: remove PageAnonExclusive sanity-check in restore_exclusive_pte()

In commit b832a354d787 ("mm/memory: page_add_anon_rmap() ->
folio_add_anon_rmap_pte()") we accidentally changed the sanity check to
essentially ignore anonymous folio by mis-placing the "!" ...  but we
really always only get anonymous folios in restore_exclusive_pte().

However, in the meantime we removed the separate "writable
device-exclusive entries" and always detect if the PTE can be writable
using can_change_pte_writable() -- which also consults PageAnonExclusive.

So let's just get rid of this sanity check completely.

Link: https://lkml.kernel.org/r/20250226132257.2826043-3-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agolib/test_hmm: make dmirror_atomic_map() consume a single page
David Hildenbrand [Wed, 26 Feb 2025 13:22:53 +0000 (14:22 +0100)]
lib/test_hmm: make dmirror_atomic_map() consume a single page

Patch series "mm: cleanups for device-exclusive entries (hmm)", v2.

Some smaller device-exclusive cleanups I have lying around.

This patch (of 5):

The caller now always passes a single page; let's simplify, and return "0"
on success.

Link: https://lkml.kernel.org/r/20250226132257.2826043-1-david@redhat.com
Link: https://lkml.kernel.org/r/20250226132257.2826043-2-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Alistair Popple <apopple@nvidia.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm: assert the folio is locked in folio_start_writeback()
Matthew Wilcox (Oracle) [Wed, 26 Feb 2025 15:36:12 +0000 (15:36 +0000)]
mm: assert the folio is locked in folio_start_writeback()

The folio must be locked when we start writeback in order to prevent
writeback from being started twice on the same folio.  I don't expect this
to catch any problems, but it should be good documentation.

Link: https://lkml.kernel.org/r/20250226153614.3774896-1-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agosamples/damon: a typo in the kconfig - sameple
Seongjun Kim [Wed, 26 Feb 2025 18:42:04 +0000 (10:42 -0800)]
samples/damon: a typo in the kconfig - sameple

There is a typo in the Kconfig file of the damon sample module.  Correct
it: s/sameple/sample/

Link: https://lkml.kernel.org/r/20250226184204.29370-1-sj@kernel.org
Signed-off-by: Seongjun Kim <bus710@gmail.com>
Signed-off-by: SeongJae Park <sj@kernel.org>
Reviewed-by: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/page_alloc: warn on nr_reserved_highatomic underflow
Brendan Jackman [Tue, 25 Feb 2025 18:45:09 +0000 (18:45 +0000)]
mm/page_alloc: warn on nr_reserved_highatomic underflow

As documented in the comment this underflow should not happen.  The
locking has indeed changed here since the comment was written, see the
migratetype hygiene patches[0].  However, those changes made the locking
_safer_, so the underflow _really_ shouldn't happen now.  So upgrade the
comment to a warning.

[0] https://lore.kernel.org/all/20240320180429.678181-7-hannes@cmpxchg.org/T/#m3da87e6cc3348a4640aa298137bc9f8f61b76c84

Link: https://lkml.kernel.org/r/20250225-warn-underflow-v1-1-3dc542941d3a@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agovmalloc: drop Christoph from Reviewers
Christoph Hellwig [Mon, 24 Feb 2025 16:30:33 +0000 (08:30 -0800)]
vmalloc: drop Christoph from Reviewers

I haven't been doing as much review as I should.  As part of reducing my
inbox flow drop me from the official Reviewers.  I might still chime in on
patches occasionally.

Link: https://lkml.kernel.org/r/20250224163033.350072-1-hch@lst.de
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm, swap: simplify folio swap allocation
Kairui Song [Thu, 13 Mar 2025 16:59:35 +0000 (00:59 +0800)]
mm, swap: simplify folio swap allocation

With slot cache gone, clean up the allocation helpers even more.
folio_alloc_swap will be the only entry for allocation and adding the
folio to swap cache (except suspend), making it opposite of
folio_free_swap.

Link: https://lkml.kernel.org/r/20250313165935.63303-8-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Baoquan He <bhe@redhat.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm, swap: remove swap slot cache
Kairui Song [Thu, 13 Mar 2025 16:59:34 +0000 (00:59 +0800)]
mm, swap: remove swap slot cache

Slot cache is no longer needed now, removing it and all related code.

- vm-scalability with: `usemem --init-time -O -y -x -R -31 1G`,
12G memory cgroup using simulated pmem as SWAP (32G pmem, 32 CPUs),
16 test runs for each case, measuring the total throughput:

                      Before (KB/s) (stdev)  After (KB/s) (stdev)
Random (4K):          424907.60 (24410.78)   414745.92  (34554.78)
Random (64K):         163308.82 (11635.72)   167314.50  (18434.99)
Sequential (4K, !-R): 6150056.79 (103205.90) 6321469.06 (115878.16)

The performance changes are below noise level.

- Build linux kernel with make -j96, using 4K folio with 1.5G memory
cgroup limit and 64K folio with 2G memory cgroup limit, on top of tmpfs,
12 test runs, measuring the system time:

                  Before (s) (stdev)  After (s) (stdev)
make -j96 (4K):   6445.69 (61.95)     6408.80 (69.46)
make -j96 (64K):  6841.71 (409.04)    6437.99 (435.55)

Similar to above, 64k mTHP case showed a slight improvement.

Link: https://lkml.kernel.org/r/20250313165935.63303-7-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm, swap: use percpu cluster as allocation fast path
Kairui Song [Thu, 13 Mar 2025 16:59:33 +0000 (00:59 +0800)]
mm, swap: use percpu cluster as allocation fast path

Current allocation workflow first traverses the plist with a global lock
held, after choosing a device, it uses the percpu cluster on that swap
device.  This commit moves the percpu cluster variable out of being tied
to individual swap devices, making it a global percpu variable, and will
be used directly for allocation as a fast path.

The global percpu cluster variable will never point to a HDD device, and
allocations on a HDD device are still globally serialized.

This improves the allocator performance and prepares for removal of the
slot cache in later commits.  There shouldn't be much observable behavior
change, except one thing: this changes how swap device allocation rotation
works.

Currently, each allocation will rotate the plist, and because of the
existence of slot cache (one order 0 allocation usually returns 64
entries), swap devices of the same priority are rotated for every 64 order
0 entries consumed.  High order allocations are different, they will
bypass the slot cache, and so swap device is rotated for every 16K, 32K,
or up to 2M allocation.

The rotation rule was never clearly defined or documented, it was changed
several times without mentioning.

After this commit, and once slot cache is gone in later commits, swap
device rotation will happen for every consumed cluster.  Ideally non-HDD
devices will be rotated if 2M space has been consumed for each order.
Fragmented clusters will rotate the device faster, which seems OK.  HDD
devices is rotated for every allocation regardless of the allocation
order, which should be OK too and trivial.

This commit also slightly changes allocation behaviour for slot cache.
The new added cluster allocation fast path may allocate entries from
different device to the slot cache, this is not observable from user
space, only impact performance very slightly, and slot cache will be just
gone in next commit, so this can be ignored.

Link: https://lkml.kernel.org/r/20250313165935.63303-6-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Baoquan He <bhe@redhat.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm, swap: don't update the counter up-front
Kairui Song [Thu, 13 Mar 2025 16:59:32 +0000 (00:59 +0800)]
mm, swap: don't update the counter up-front

The counter update before allocation design was useful to avoid
unnecessary scan when device is full, so it will abort early if the
counter indicates the device is full.  But that is an uncommon case, and
now scanning of a full device is very fast, so the up-front update is not
helpful any more.

Remove it and simplify the slot allocation logic.

Link: https://lkml.kernel.org/r/20250313165935.63303-5-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm, swap: avoid redundant swap device pinning
Kairui Song [Thu, 13 Mar 2025 16:59:31 +0000 (00:59 +0800)]
mm, swap: avoid redundant swap device pinning

Currently __read_swap_cache_async() has get/put_swap_device() calls to
increase/decrease a swap device reference to prevent swapoff.  While some
of its callers have already held the swap device reference, e.g in
do_swap_page() and shmem_swapin_folio() where __read_swap_cache_async()
will finally called.  Now there are only two callers not holding a swap
device reference, so make them hold a reference instead.  And drop the
get/put_swap_device calls in __read_swap_cache_async.  This should reduce
the overhead for swap in during page fault slightly.

Link: https://lkml.kernel.org/r/20250313165935.63303-4-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm, swap: drop the flag TTRS_DIRECT
Kairui Song [Thu, 13 Mar 2025 16:59:30 +0000 (00:59 +0800)]
mm, swap: drop the flag TTRS_DIRECT

This flag exists temporarily to allow the allocator to bypass the slot
cache during freeing, so reclaiming one slot will free the slot
immediately.

But now we have already removed slot cache usage on freeing, so this flag
has no effect now.

Link: https://lkml.kernel.org/r/20250313165935.63303-3-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm, swap: avoid reclaiming irrelevant swap cache
Kairui Song [Thu, 13 Mar 2025 16:59:29 +0000 (00:59 +0800)]
mm, swap: avoid reclaiming irrelevant swap cache

Patch series "mm, swap: remove swap slot cache", v3.

Slot cache was initially introduced by commit 67afa38e012e ("mm/swap: add
cache for swap slots allocation") to reduce the lock contention of
si->lock.

Previous series "mm, swap: rework of swap allocator locks" [1] removed
swap slot cache for freeing path as freeing path no longer touches
si->lock in most cased.  Allocation path also have slight to none
contention on si->lock since that series, but slot cache still helps to
reduce other overheads, like counters and the plist.

This series removes the slot cache from allocation path too, by using the
cluster as allocation fast path and also reduce other overheads.

Now slot cache is completely gone, the code is much simplified without
obvious feature or performance change, also clean up related workaround.
Also this should avoid other potential issues, e.g.  the long pinning of
swap slots: swap slot cache pins swap slots with HAS_CACHE, causing
reclaim or allocation fail to use these slots on scanning.

The only behavior change is the swap device allocation rotation mechanism,
as explained in the patch "mm, swap: use percpu cluster as allocation fast
path".

Test results are looking good after deleting the swap slot cache:

- vm-scalability with: `usemem --init-time -O -y -x -R -31 1G`,
12G memory cgroup using simulated pmem as SWAP (32G pmem, 32 CPUs),
16 test runs for each case, measuring the total throughput:

                      Before (KB/s) (stdev)  After (KB/s) (stdev)
Random (4K):          424907.60 (24410.78)   414745.92  (34554.78)
Random (64K):         163308.82 (11635.72)   167314.50  (18434.99)
Sequential (4K, !-R): 6150056.79 (103205.90) 6321469.06 (115878.16)

- Build linux kernel with make -j96, using 4K folio with 1.5G memory
cgroup limit and 64K folio with 2G memory cgroup limit, on top of tmpfs,
12 test runs, measuring the system time:

                  Before (s) (stdev)  After (s) (stdev)
make -j96 (4K):   6445.69 (61.95)     6408.80 (69.46)
make -j96 (64K):  6841.71 (409.04)    6437.99 (435.55)

The performance is unchanged, slightly better in some cases.

[1] https://lore.kernel.org/linux-mm/20250113175732.48099-1-ryncsn@gmail.com/

This patch (of 7):

Swap allocator will do swap cache reclaim to recycle HAS_CACHE slots for
allocation.  It initiates the reclaim from the offset to be reclaimed and
looks up the corresponding folio.  The lookup process is lockless, so it's
possible the folio will be removed from the swap cache and given a
different swap entry before the reclaim locks the folio.  If it happens,
the reclaim will end up reclaiming an irrelevant folio, and return wrong
return value.

This shouldn't cause any problem with correctness or stability, but it is
indeed confusing and unexpected, and will increase fragmentation, decrease
performance.

Fix this by checking whether the folio is still pointing to the offset the
allocator want to reclaim before reclaiming it.

Link: https://lkml.kernel.org/r/20250313165935.63303-1-ryncsn@gmail.com
Link: https://lkml.kernel.org/r/20250313165935.63303-2-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kairui Song <kasong@tencent.com>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm: make page_mapped_in_vma() hugetlb walk aware
Jane Chu [Mon, 24 Feb 2025 21:14:45 +0000 (14:14 -0700)]
mm: make page_mapped_in_vma() hugetlb walk aware

When a process consumes a UE in a page, the memory failure handler
attempts to collect information for a potential SIGBUS.  If the page is an
anonymous page, page_mapped_in_vma(page, vma) is invoked in order to

  1. retrieve the vaddr from the process' address space,

  2. verify that the vaddr is indeed mapped to the poisoned page,
     where 'page' is the precise small page with UE.

It's been observed that when injecting poison to a non-head subpage of an
anonymous hugetlb page, no SIGBUS shows up, while injecting to the head
page produces a SIGBUS.  The cause is that, though hugetlb_walk() returns
a valid pmd entry (on x86), but check_pte() detects mismatch between the
head page per the pmd and the input subpage.  Thus the vaddr is considered
not mapped to the subpage and the process is not collected for SIGBUS
purpose.  This is the calling stack:

      collect_procs_anon
        page_mapped_in_vma
          page_vma_mapped_walk
            hugetlb_walk
              huge_pte_lock
                check_pte

check_pte() header says that it
"check if [pvmw->pfn, @pvmw->pfn + @pvmw->nr_pages) is mapped at the @pvmw->pte"
but practically works only if pvmw->pfn is the head page pfn at pvmw->pte.
Hindsight acknowledging that some pvmw->pte could point to a hugepage of
some sort such that it makes sense to make check_pte() work for hugepage.

Link: https://lkml.kernel.org/r/20250224211445.2663312-1-jane.chu@oracle.com
Signed-off-by: Jane Chu <jane.chu@oracle.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: linmiaohe <linmiaohe@huawei.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Peter Xu <peterx@redhat.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm: page_alloc: group fallback functions together
Johannes Weiner [Tue, 25 Feb 2025 00:08:26 +0000 (19:08 -0500)]
mm: page_alloc: group fallback functions together

The way the fallback rules are spread out makes them hard to follow.  Move
the functions next to each other at least.

Link: https://lkml.kernel.org/r/20250225001023.1494422-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm: page_alloc: remove remnants of unlocked migratetype updates
Johannes Weiner [Tue, 25 Feb 2025 00:08:25 +0000 (19:08 -0500)]
mm: page_alloc: remove remnants of unlocked migratetype updates

The freelist hygiene patches made migratetype accesses fully protected
under the zone->lock.  Remove remnants of handling the race conditions
that existed before from the MIGRATE_HIGHATOMIC code.

Link: https://lkml.kernel.org/r/20250225001023.1494422-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm: page_alloc: don't steal single pages from biggest buddy
Johannes Weiner [Tue, 25 Feb 2025 00:08:24 +0000 (19:08 -0500)]
mm: page_alloc: don't steal single pages from biggest buddy

The fallback code searches for the biggest buddy first in an attempt to
steal the whole block and encourage type grouping down the line.

The approach used to be this:

- Non-movable requests will split the largest buddy and steal the
  remainder. This splits up contiguity, but it allows subsequent
  requests of this type to fall back into adjacent space.

- Movable requests go and look for the smallest buddy instead. The
  thinking is that movable requests can be compacted, so grouping is
  less important than retaining contiguity.

c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block
conversion") enforces freelist type hygiene, which restricts stealing to
either claiming the whole block or just taking the requested chunk; no
additional pages or buddy remainders can be stolen any more.

The patch mishandled when to switch to finding the smallest buddy in that
new reality.  As a result, it may steal the exact request size, but from
the biggest buddy.  This causes fracturing for no good reason.

Fix this by committing to the new behavior: either steal the whole block,
or fall back to the smallest buddy.

Remove single-page stealing from steal_suitable_fallback().  Rename it to
try_to_steal_block() to make the intentions clear.  If this fails, always
fall back to the smallest buddy.

The following is from 4 runs of mmtest's thpchallenge.  "Pollute" is
single page fallback, "steal" is conversion of a partially used block.
The numbers for free block conversions (omitted) are comparable.

     vanilla       patched

@pollute[unmovable from reclaimable]:   27   106
@pollute[unmovable from movable]:   82    46
@pollute[reclaimable from unmovable]:  256    83
@pollute[reclaimable from movable]:   46     8
@pollute[movable from unmovable]: 4841   868
@pollute[movable from reclaimable]: 5278 12568

@steal[unmovable from reclaimable]:   11    12
@steal[unmovable from movable]:  113    49
@steal[reclaimable from unmovable]:   19    34
@steal[reclaimable from movable]:   47    21
@steal[movable from unmovable]:  250   183
@steal[movable from reclaimable]:   81    93

The allocator appears to do a better job at keeping stealing and polluting
to the first fallback preference.  As a result, the numbers for "from
movable" - the least preferred fallback option, and most detrimental to
compactability - are down across the board.

Link: https://lkml.kernel.org/r/20250225001023.1494422-2-hannes@cmpxchg.org
Fixes: c0cd6f557b90 ("mm: page_alloc: fix freelist movement during block conversion")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agotools/selftests: add guard region test for /proc/$pid/pagemap
Lorenzo Stoakes [Fri, 21 Feb 2025 12:05:23 +0000 (12:05 +0000)]
tools/selftests: add guard region test for /proc/$pid/pagemap

Add a test to the guard region self tests to assert that the
/proc/$pid/pagemap information now made availabile to the user correctly
identifies and reports guard regions.

As a part of this change, update vm_util.h to add the new bit (note there
is no header file in the kernel where this is exposed, the user is
expected to provide their own mask) and utilise the helper functions there
for pagemap functionality.

[lorenzo.stoakes@oracle.com: fixup define name]
Link: https://lkml.kernel.org/r/32e83941-e6f5-42ee-9292-a44c16463cf1@lucifer.local
Link: https://lkml.kernel.org/r/164feb0a43ae72650e6b20c3910213f469566311.1740139449.git.lorenzo.stoakes@oracle.com
Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: "Paul E . McKenney" <paulmck@kernel.org>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agofs/proc/task_mmu: add guard region bit to pagemap
Lorenzo Stoakes [Fri, 21 Feb 2025 12:05:22 +0000 (12:05 +0000)]
fs/proc/task_mmu: add guard region bit to pagemap

Patch series "fs/proc/task_mmu: add guard region bit to pagemap".

Currently there is no means of determining whether a given page in a
mapping range is designated a guard region (as installed via madvise()
using the MADV_GUARD_INSTALL flag).

This is generally not an issue, but in some instances users may wish to
determine whether this is the case.

This series adds this ability via /proc/$pid/pagemap, updates the
documentation and adds a self test to assert that this functions
correctly.

This patch (of 2):

Currently there is no means by which users can determine whether a given
page in memory is in fact a guard region, that is having had the
MADV_GUARD_INSTALL madvise() flag applied to it.

This is intentional, as to provide this information in VMA metadata would
contradict the intent of the feature (providing a means to change fault
behaviour at a page table level rather than a VMA level), and would
require VMA metadata operations to scan page tables, which is
unacceptable.

In many cases, users have no need to reflect and determine what regions
have been designated guard regions, as it is the user who has established
them in the first place.

But in some instances, such as monitoring software, or software that
relies upon being able to ascertain the nature of mappings within a remote
process for instance, it becomes useful to be able to determine which
pages have the guard region marker applied.

This patch makes use of an unused pagemap bit (58) to provide this
information.

This patch updates the documentation at the same time as making the change
such that the implementation of the feature and the documentation of it
are tied together.

Link: https://lkml.kernel.org/r/cover.1740139449.git.lorenzo.stoakes@oracle.com
Link: https://lkml.kernel.org/r/521d99c08b975fb06a1e7201e971cc24d68196d1.1740139449.git.lorenzo.stoakes@oracle.com
Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: "Paul E . McKenney" <paulmck@kernel.org>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm: swap: remove stale comment of swap_reclaim_full_clusters()
Kemeng Shi [Sat, 22 Feb 2025 16:08:50 +0000 (00:08 +0800)]
mm: swap: remove stale comment of swap_reclaim_full_clusters()

swap_reclaim_full_clusters() has no return value now, just remove the
stale comment which says swap_reclaim_full_clusters() wil return a bool
value.

Link: https://lkml.kernel.org/r/20250222160850.505274-7-shikemeng@huaweicloud.com
Signed-off-by: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: Kairui Song <ryncsn@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm, swap: correct comment in swap_usage_sub()
Kemeng Shi [Sat, 22 Feb 2025 16:08:49 +0000 (00:08 +0800)]
mm, swap: correct comment in swap_usage_sub()

We will add si back to plist in swap_usage_sub(), just correct the wrong
comment which says we will remove si from plist in swap_usage_sub().

Link: https://lkml.kernel.org/r/20250222160850.505274-6-shikemeng@huaweicloud.com
Signed-off-by: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: Kairui Song <ryncsn@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm, swap: remove setting SWAP_MAP_BAD for discard cluster
Kemeng Shi [Sat, 22 Feb 2025 16:08:48 +0000 (00:08 +0800)]
mm, swap: remove setting SWAP_MAP_BAD for discard cluster

Before alloc from a cluster, we will aqcuire cluster's lock and make sure
it is usable by cluster_is_usable(), so there is no need to set
SWAP_MAP_BAD for cluster to be discarded.

Link: https://lkml.kernel.org/r/20250222160850.505274-5-shikemeng@huaweicloud.com
Signed-off-by: Kemeng Shi <shikemeng@huaweicloud.com>
Reviewed-by: Kairui Song <kasong@tencent.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoselftests/mm/mlock: print error on failure
Brendan Jackman [Tue, 11 Mar 2025 13:18:23 +0000 (13:18 +0000)]
selftests/mm/mlock: print error on failure

It's not really possible to start diagnosing this without knowing the
actual error.

Also update the mlock2 helper to behave like libc would by setting errno
and returning -1.

Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-12-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoselftests/mm: skip mlock tests if nobody user can't read it
Brendan Jackman [Tue, 11 Mar 2025 13:18:22 +0000 (13:18 +0000)]
selftests/mm: skip mlock tests if nobody user can't read it

If running from a directory that can't be read by unprivileged users,
executing on-fault-test via the nobody user will fail.

The kselftest build does give the file the correct permissions, but after
being installed it might be in a directory without global execute
permissions.

Since the script can't safely fix that, just skip if it happens.  Note
that the stderr of the `ls` command is unfiltered meaning the user sees a
"permission denied" error that can help inform them why the test was
skipped.

Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-11-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoselftests/mm: ensure uffd-wp-mremap gets pages of each size
Brendan Jackman [Tue, 11 Mar 2025 13:18:21 +0000 (13:18 +0000)]
selftests/mm: ensure uffd-wp-mremap gets pages of each size

This test allocates a page of every available size and doesn't have any
SKIP logic if the allocation fails.  So, ensure it's available and skip
the test if we can't do so.

Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-10-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoselftests/mm: drop unnecessary sudo usage
Brendan Jackman [Tue, 11 Mar 2025 13:18:20 +0000 (13:18 +0000)]
selftests/mm: drop unnecessary sudo usage

This script must be run as root anyway (see all the writing to privileged
files in /proc etc).

Remove the unnecessary use of sudo to avoid breaking on single-user
systems that don't have sudo.  This also avoids confusing readers.

Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-9-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoselftests/mm: skip gup_longterm tests on weird filesystems
Brendan Jackman [Tue, 11 Mar 2025 13:18:19 +0000 (13:18 +0000)]
selftests/mm: skip gup_longterm tests on weird filesystems

Some filesystems don't support ftruncate()ing unlinked files.  They return
ENOENT.  In that case, skip the test.

Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-8-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoselftests/mm: skip map_populate on weird filesystems
Brendan Jackman [Tue, 11 Mar 2025 13:18:18 +0000 (13:18 +0000)]
selftests/mm: skip map_populate on weird filesystems

It seems that 9pfs does not allow truncating unlinked files, Mark Brown
has noted that NFS may also behave this way.

It doesn't seem quite right to call this a "bug" but it's probably a
special enough case that it makes sense for the test to just SKIP if it
happens.

Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-7-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoselftests/mm: don't fail uffd-stress if too many CPUs
Brendan Jackman [Tue, 11 Mar 2025 13:18:17 +0000 (13:18 +0000)]
selftests/mm: don't fail uffd-stress if too many CPUs

This calculation divides a fixed parameter by an environment-dependent
parameter i.e.  the number of CPUs.

The simple way to avoid machine-specific failures here is to just put a
cap on the max value of the latter.

Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-6-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Suggested-by: Mateusz Guzik <mjguzik@gmail.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoselftests/mm: print some details when uffd-stress gets bad params
Brendan Jackman [Tue, 11 Mar 2025 13:18:16 +0000 (13:18 +0000)]
selftests/mm: print some details when uffd-stress gets bad params

So this can be debugged more easily.

Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-5-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoselftests/mm/uffd: rename nr_cpus -> nr_parallel
Brendan Jackman [Tue, 11 Mar 2025 13:18:15 +0000 (13:18 +0000)]
selftests/mm/uffd: rename nr_cpus -> nr_parallel

A later commit will bound this variable so it no longer necessarily
matches the number of CPUs.  Rename it appropriately.

Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-4-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoselftests/mm: skip uffd-wp-mremap if userfaultfd not available
Brendan Jackman [Tue, 11 Mar 2025 13:18:14 +0000 (13:18 +0000)]
selftests/mm: skip uffd-wp-mremap if userfaultfd not available

It's obvious that this should fail in that case, but still, save the
reader the effort of figuring out that they've run into this by just
SKIPping

Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-3-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoselftests/mm: skip uffd-stress if userfaultfd not available
Brendan Jackman [Tue, 11 Mar 2025 13:18:13 +0000 (13:18 +0000)]
selftests/mm: skip uffd-stress if userfaultfd not available

It's pretty obvious that the test wouldn't work if you don't have the
feature enabled.  But, it's still useful to SKIP instead of failing so the
reader can immediately tell that this is the reason why.

Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-2-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoselftests/mm: report errno when things fail in gup_longterm
Brendan Jackman [Tue, 11 Mar 2025 13:18:12 +0000 (13:18 +0000)]
selftests/mm: report errno when things fail in gup_longterm

Patch series "selftests/mm: Some cleanups from trying to run them", v4.

I never had much luck running mm selftests so I spent a few hours digging
into why.

Looks like most of the reason is missing SKIP checks, so this series is
just adding a bunch of those that I found.  I did not do anything like all
of them, just the ones I spotted in gup_longterm, gup_test, mmap,
userfaultfd and memfd_secret.

It's a bit unfortunate to have to skip those tests when ftruncate() fails,
but I don't have time to dig deep enough into it to actually make them
pass.  I have observed the issue on 9pfs and heard rumours that NFS has a
similar problem.

I'm now able to run these test groups successfully:

- mmap
- gup_test
- compaction
- migration
- page_frag
- userfaultfd
- mlock

I've never gone past "Waiting for hugetlb memory to get depleted", in the
hugetlb tests.  I don't know if they are stuck or if they would eventually
work if I was patient enough (testing on a 1G machine).  I have not
investigated further.

I had some issues with mlock tests failing due to -ENOSRCH from mlock2(),
I can no longer reproduce that though, things work OK now.

Of the remaining tests there may be others that work fine, but there's no
convenient way to survey the whole output of run_vmtests.sh so I'm just
going test by test here.

In my spare moments I am slowly chipping away at a setup to run these
tests continuously in a reasonably hermetic QEMU environment via
virtme-ng:

https://github.com/bjackman/linux/blob/5fad4b9c592290f38e0f8bc73c9abb9c99d8787c/README.md

Hopefully that will eventually offer a way to provide a "canned"
environment where the tests are known to work, which can be fairly easily
reproduced by any developer.

This patch (of 12):

Just reporting failure doesn't tell you what went wrong.  This can fail in
different ways so report errno to help the reader get started debugging.

Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-0-dec210a658f5@google.com
Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-1-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: add might_sleep to zcomp API
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:28 +0000 (11:03 +0900)]
zram: add might_sleep to zcomp API

Explicitly state that zcomp compress/decompress must be called from
non-atomic context.

Link: https://lkml.kernel.org/r/20250303022425.285971-20-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: do not leak page on writeback_store error path
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:27 +0000 (11:03 +0900)]
zram: do not leak page on writeback_store error path

Ensure the page used for local object data is freed on error out path.

Link: https://lkml.kernel.org/r/20250303022425.285971-19-senozhatsky@chromium.org
Fixes: 330edc2bc059 (zram: rework writeback target selection strategy)
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: do not leak page on recompress_store error path
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:26 +0000 (11:03 +0900)]
zram: do not leak page on recompress_store error path

Ensure the page used for local object data is freed on error out path.

Link: https://lkml.kernel.org/r/20250303022425.285971-18-senozhatsky@chromium.org
Fixes: 3f909a60cec1 ("zram: rework recompress target selection strategy")
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: permit reclaim in zstd custom allocator
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:25 +0000 (11:03 +0900)]
zram: permit reclaim in zstd custom allocator

When configured with pre-trained compression/decompression dictionary
support, zstd requires custom memory allocator, which it calls internally
from compression()/decompression() routines.  That means allocation from
atomic context (either under entry spin-lock, or per-CPU local-lock or
both).  Now, with non-atomic zram read()/write(), those limitations are
relaxed and we can allow direct and indirect reclaim.

Link: https://lkml.kernel.org/r/20250303022425.285971-17-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: switch to new zsmalloc object mapping API
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:24 +0000 (11:03 +0900)]
zram: switch to new zsmalloc object mapping API

Use new read/write zsmalloc object API.  For cases when RO mapped object
spans two physical pages (requires temp buffer) compression streams now
carry around one extra physical page.

Link: https://lkml.kernel.org/r/20250303022425.285971-16-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozsmalloc: introduce new object mapping API
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:23 +0000 (11:03 +0900)]
zsmalloc: introduce new object mapping API

Current object mapping API is a little cumbersome.  First, it's
inconsistent, sometimes it returns with page-faults disabled and sometimes
with page-faults enabled.  Second, and most importantly, it enforces
atomicity restrictions on its users.  zs_map_object() has to return a
liner object address which is not always possible because some objects
span multiple physical (non-contiguous) pages.  For such objects zsmalloc
uses a per-CPU buffer to which object's data is copied before a pointer to
that per-CPU buffer is returned back to the caller.  This leads to
another, final, issue - extra memcpy().  Since the caller gets a pointer
to per-CPU buffer it can memcpy() data only to that buffer, and during
zs_unmap_object() zsmalloc will memcpy() from that per-CPU buffer to
physical pages that object in question spans across.

New API splits functions by access mode:
- zs_obj_read_begin(handle, local_copy)
  Returns a pointer to handle memory.  For objects that span two
  physical pages a local_copy buffer is used to store object's
  data before the address is returned to the caller.  Otherwise
  the object's page is kmap_local mapped directly.

- zs_obj_read_end(handle, buf)
  Unmaps the page if it was kmap_local mapped by zs_obj_read_begin().

- zs_obj_write(handle, buf, len)
  Copies len-bytes from compression buffer to handle memory
  (takes care of objects that span two pages).  This does not
  need any additional (e.g. per-CPU) buffers and writes the data
  directly to zsmalloc pool pages.

In terms of performance, on a synthetic and completely reproducible
test that allocates fixed number of objects of fixed sizes and
iterates over those objects, first mapping in RO then in RW mode:

OLD API
=======

3 first results out of 10

  369,205,778      instructions        #    0.80  insn per cycle
   40,467,926      branches            #  113.732 M/sec

  369,002,122      instructions        #    0.62  insn per cycle
   40,426,145      branches            #  189.361 M/sec

  369,036,706      instructions        #    0.63  insn per cycle
   40,430,860      branches            #  204.105 M/sec

[..]

NEW API
=======

3 first results out of 10

  265,799,293      instructions        #    0.51  insn per cycle
   29,834,567      branches            #  170.281 M/sec

  265,765,970      instructions        #    0.55  insn per cycle
   29,829,019      branches            #  161.602 M/sec

  265,764,702      instructions        #    0.51  insn per cycle
   29,828,015      branches            #  189.677 M/sec

[..]

T-test on all 10 runs
=====================

Difference at 95.0% confidence
   -1.03219e+08 +/- 55308.7
   -27.9705% +/- 0.0149878%
   (Student's t, pooled s = 58864.4)

The old API will stay around until the remaining users switch to the new
one.  After that we'll also remove zsmalloc per-CPU buffer and CPU hotplug
handling.

The split of map(RO) and map(WO) into read_{begin/end}/write is suggested
by Yosry Ahmed.

Link: https://lkml.kernel.org/r/20250303022425.285971-15-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Suggested-by: Yosry Ahmed <yosry.ahmed@linux.dev>
Reviewed-by: Yosry Ahmed <yosry.ahmed@linux.dev>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozsmalloc: sleepable zspage reader-lock
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:22 +0000 (11:03 +0900)]
zsmalloc: sleepable zspage reader-lock

In order to implement preemptible object mapping we need a zspage lock
that satisfies several preconditions:
- it should be reader-write type of a lock
- it should be possible to hold it from any context, but also being
  preemptible if the context allows it
- we never sleep while acquiring but can sleep while holding in read
  mode

An rwsemaphore doesn't suffice, due to atomicity requirements, rwlock
doesn't satisfy due to reader-preemptability requirement.  It's also worth
to mention, that per-zspage rwsem is a little too memory heavy (we can
easily have double digits megabytes used only on rwsemaphores).

Switch over from rwlock_t to a atomic_t-based implementation of a
reader-writer semaphore that satisfies all of the preconditions.

The spin-lock based zspage_lock is suggested by Hillf Danton.

Link: https://lkml.kernel.org/r/20250303022425.285971-14-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Suggested-by: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozsmalloc: rename pool lock
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:21 +0000 (11:03 +0900)]
zsmalloc: rename pool lock

The old name comes from the times when the pool did not have compaction
(defragmentation).  Rename it to ->lock because these days it synchronizes
not only migration.

Link: https://lkml.kernel.org/r/20250303022425.285971-13-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Reviewed-by: Yosry Ahmed <yosry.ahmed@linux.dev>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: move post-processing target allocation
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:20 +0000 (11:03 +0900)]
zram: move post-processing target allocation

Allocate post-processing target in place_pp_slot().  This simplifies
scan_slots_for_writeback() and scan_slots_for_recompress() loops because
we don't need to track pps pointer state anymore.  Previously we have to
explicitly NULL the point if it has been added to a post-processing bucket
or re-use previously allocated pointer otherwise and make sure we don't
leak the memory in the end.

We are also fine doing GFP_NOIO allocation, as post-processing can be
called under memory pressure so we better pick as many slots as we can as
soon as we can and start post-processing them, possibly saving the memory.
Allocation failure there is not fatal, we will post-process whatever we
put into the buckets on previous iterations.

Link: https://lkml.kernel.org/r/20250303022425.285971-12-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: rework recompression loop
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:19 +0000 (11:03 +0900)]
zram: rework recompression loop

This reworks recompression loop handling:

- set a rule that stream-put NULLs the stream pointer If the loop
  returns with a non-NULL stream then it's a successful recompression,
  otherwise the stream should always be NULL.

- do not count the number of recompressions Mark object as
  incompressible as soon as the algorithm with the highest priority failed
  to compress that object.

- count compression errors as resource usage Even if compression has
  failed, we still need to bump num_recomp_pages counter.

Link: https://lkml.kernel.org/r/20250303022425.285971-11-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: filter out recomp targets based on priority
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:18 +0000 (11:03 +0900)]
zram: filter out recomp targets based on priority

Do no select for post processing slots that are already compressed with
same or higher priority compression algorithm.

This should save some memory, as previously we would still put those
entries into corresponding post-processing buckets and filter them out
later in recompress_slot().

Link: https://lkml.kernel.org/r/20250303022425.285971-10-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: limit max recompress prio to num_active_comps
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:17 +0000 (11:03 +0900)]
zram: limit max recompress prio to num_active_comps

Use the actual number of algorithms zram was configure with instead of
theoretical limit of ZRAM_MAX_COMPS.

Also make sure that min prio is not above max prio.

Link: https://lkml.kernel.org/r/20250303022425.285971-9-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: remove writestall zram_stats member
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:16 +0000 (11:03 +0900)]
zram: remove writestall zram_stats member

There is no zsmalloc handle allocation slow path now and writestall is not
possible any longer.  Remove it from zram_stats.

Link: https://lkml.kernel.org/r/20250303022425.285971-8-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: add GFP_NOWARN to incompressible zsmalloc handle allocation
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:15 +0000 (11:03 +0900)]
zram: add GFP_NOWARN to incompressible zsmalloc handle allocation

We normally use __GFP_NOWARN for zsmalloc handle allocations, add it to
write_incompressible_page() allocation too.

Link: https://lkml.kernel.org/r/20250303022425.285971-7-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: remove second stage of handle allocation
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:14 +0000 (11:03 +0900)]
zram: remove second stage of handle allocation

Previously zram write() was atomic which required us to pass
__GFP_KSWAPD_RECLAIM to zsmalloc handle allocation on a fast path and
attempt a slow path allocation (with recompression) if the fast path
failed.

Since we are not in atomic context anymore we can permit direct reclaim
during handle allocation, and hence can have a single allocation path.
There is no slow path anymore so we don't unlock per-CPU stream (and don't
lose compressed data) which means that there is no need to do
recompression now (which should reduce CPU and battery usage).

Link: https://lkml.kernel.org/r/20250303022425.285971-6-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: remove max_comp_streams device attr
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:13 +0000 (11:03 +0900)]
zram: remove max_comp_streams device attr

max_comp_streams device attribute has been defunct since May 2016 when
zram switched to per-CPU compression streams, remove it.

Link: https://lkml.kernel.org/r/20250303022425.285971-5-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: remove unused crypto include
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:12 +0000 (11:03 +0900)]
zram: remove unused crypto include

We stopped using crypto API (for the time being), so remove its include
and replace CRYPTO_MAX_ALG_NAME with a local define.

Link: https://lkml.kernel.org/r/20250303022425.285971-4-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: permit preemption with active compression stream
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:11 +0000 (11:03 +0900)]
zram: permit preemption with active compression stream

Currently, per-CPU stream access is done from a non-preemptible (atomic)
section, which imposes the same atomicity requirements on compression
backends as entry spin-lock, and makes it impossible to use algorithms
that can schedule/wait/sleep during compression and decompression.

Switch to preemptible per-CPU model, similar to the one used in zswap.
Instead of a per-CPU local lock, each stream carries a mutex which is
locked throughout entire time zram uses it for compression or
decompression, so that cpu-dead event waits for zram to stop using a
particular per-CPU stream and release it.

Link: https://lkml.kernel.org/r/20250303022425.285971-3-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Suggested-by: Yosry Ahmed <yosry.ahmed@linux.dev>
Reviewed-by: Yosry Ahmed <yosry.ahmed@linux.dev>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agozram: sleepable entry locking
Sergey Senozhatsky [Mon, 3 Mar 2025 02:03:10 +0000 (11:03 +0900)]
zram: sleepable entry locking

Patch series "zsmalloc/zram: there be preemption", v10.

Currently zram runs compression and decompression in non-preemptible
sections, e.g.

    zcomp_stream_get()     // grabs CPU local lock
    zcomp_compress()

or

    zram_slot_lock()       // grabs entry spin-lock
    zcomp_stream_get()     // grabs CPU local lock
    zs_map_object()        // grabs rwlock and CPU local lock
    zcomp_decompress()

Potentially a little troublesome for a number of reasons.

For instance, this makes it impossible to use async compression algorithms
or/and H/W compression algorithms, which can wait for OP completion or
resource availability.  This also restricts what compression algorithms
can do internally, for example, zstd can allocate internal state memory
for C/D dictionaries:

do_fsync()
 do_writepages()
  zram_bio_write()
   zram_write_page()                          // become non-preemptible
    zcomp_compress()
     zstd_compress()
      ZSTD_compress_usingCDict()
       ZSTD_compressBegin_usingCDict_internal()
        ZSTD_resetCCtx_usingCDict()
         ZSTD_resetCCtx_internal()
          zstd_custom_alloc()                 // memory allocation

Not to mention that the system can be configured to maximize compression
ratio at a cost of CPU/HW time (e.g.  lz4hc or deflate with very high
compression level) so zram can stay in non-preemptible section (even under
spin-lock or/and rwlock) for an extended period of time.  Aside from
compression algorithms, this also restricts what zram can do.  One
particular example is zram_write_page() zsmalloc handle allocation, which
has an optimistic allocation (disallowing direct reclaim) and a
pessimistic fallback path, which then forces zram to compress the page one
more time.

This series changes zram to not directly impose atomicity restrictions on
compression algorithms (and on itself), which makes zram write() fully
preemptible; zram read(), sadly, is not always preemptible yet.  There are
still indirect atomicity restrictions imposed by zsmalloc().  One notable
example is object mapping API, which returns with: a) local CPU lock held
b) zspage rwlock held

First, zsmalloc's zspage lock is converted from rwlock to a special type
of RW-lookalike look with some extra guarantees/features.  Second, a new
handle mapping is introduced which doesn't use per-CPU buffers (and hence
no local CPU lock), does fewer memcpy() calls, but requires users to
provide a pointer to temp buffer for object copy-in (when needed).  Third,
zram is converted to the new zsmalloc mapping API and thus zram read()
becomes preemptible.

This patch (of 19):

Concurrent modifications of meta table entries is now handled by per-entry
spin-lock.  This has a number of shortcomings.

First, this imposes atomic requirements on compression backends.  zram can
call both zcomp_compress() and zcomp_decompress() under entry spin-lock,
which implies that we can use only compression algorithms that don't
schedule/sleep/wait during compression and decompression.  This, for
instance, makes it impossible to use some of the ASYNC compression
algorithms (H/W compression, etc.) implementations.

Second, this can potentially trigger watchdogs.  For example, entry
re-compression with secondary algorithms is performed under entry
spin-lock.  Given that we chain secondary compression algorithms and that
some of them can be configured for best compression ratio (and worst
compression speed) zram can stay under spin-lock for quite some time.

Having a per-entry mutex (or, for instance, a rw-semaphore) significantly
increases sizeof() of each entry and hence the meta table.  Therefore
entry locking returns back to bit locking, as before, however, this time
also preempt-rt friendly, because if waits-on-bit instead of
spinning-on-bit.  Lock owners are also now permitted to schedule, which is
a first step on the path of making zram non-atomic.

Link: https://lkml.kernel.org/r/20250303022425.285971-1-senozhatsky@chromium.org
Link: https://lkml.kernel.org/r/20250303022425.285971-2-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/folio_queue: delete __folio_order and use folio_order directly
Liu Ye [Wed, 12 Feb 2025 02:58:42 +0000 (10:58 +0800)]
mm/folio_queue: delete __folio_order and use folio_order directly

__folio_order is the same as folio_order, remove __folio_order and then
just include mm.h and use folio_order directly.

Link: https://lkml.kernel.org/r/20250212025843.80283-2-liuye@kylinos.cn
Signed-off-by: Liu Ye <liuye@kylinos.cn>
Reviewed-by: Shivank Garg <shivankg@amd.com>
Reviewed-by: Dev Jain <dev.jain@arm.com>
Acked-by: David Howells <dhowells@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/mincore: improve performance by adding an unlikely hint
Colin Ian King [Wed, 19 Feb 2025 08:36:07 +0000 (08:36 +0000)]
mm/mincore: improve performance by adding an unlikely hint

Adding an unlikely() hint on the masked start comparison error return path
improves run-time performance of the mincore system call.

Benchmarking on an i9-12900 shows an improvement of 7ns on mincore calls
on a 256KB mmap'd region where 50% of the pages we resident.  Improvement
was from ~970 ns down to 963 ns, so a small ~0.7% improvement.

Results based on running 20 tests with turbo disabled (to reduce clock
freq turbo changes), with 10 second run per test and comparing the number
of mincores calls per second.  The % standard deviation of the 20 tests
was ~0.10%, so results are reliable.

Link: https://lkml.kernel.org/r/20250219083607.5183-1-colin.i.king@gmail.com
Signed-off-by: Colin Ian King <colin.i.king@gmail.com>
Cc: Matthew Wilcow <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoDocs/mm/damon/design: document unmapped DAMOS filter type
SeongJae Park [Wed, 19 Feb 2025 22:01:46 +0000 (14:01 -0800)]
Docs/mm/damon/design: document unmapped DAMOS filter type

Document availability and meaning of unmapped DAMOS filter type on design
document.  Since introduction of the type requires no additional user ABI,
usage and ABI document need no update.

Link: https://lkml.kernel.org/r/20250219220146.133650-3-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/damon: implement a new DAMOS filter type for unmapped pages
SeongJae Park [Wed, 19 Feb 2025 22:01:45 +0000 (14:01 -0800)]
mm/damon: implement a new DAMOS filter type for unmapped pages

Patch series "mm/damon: introduce DAMOS filter type for unmapped pages".

User decides whether their memory will be mapped or unmapped.  It implies
that the two types of memory can have different characteristics and
management requirements.  Provide the DAMON-observaibility DAMOS-operation
capability for the different types by introducing a new DAMOS filter type
for unmapped pages.

This patch (of 2):

Implement yet another DAMOS filter type for unmapped pages on DAMON kernel
API, and add support of it from the physical address space DAMON
operations set (paddr).  Since it is for only unmapped pages, support from
the virtual address spaces DAMON operations set (vaddr) is not required.

Link: https://lkml.kernel.org/r/20250219220146.133650-1-sj@kernel.org
Link: https://lkml.kernel.org/r/20250219220146.133650-2-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agoarm/pgtable: remove duplicate included header file
Thorsten Blum [Wed, 19 Feb 2025 11:24:02 +0000 (12:24 +0100)]
arm/pgtable: remove duplicate included header file

The header file asm-generic/pgtable-nopud.h is included whether CONFIG_MMU
is defined or not.

Include it only once before the #ifndef/#else/#endif preprocessor
directives and remove the following make includecheck warning:

  asm-generic/pgtable-nopud.h is included more than once

Link: https://lkml.kernel.org/r/20250219112403.3959-2-thorsten.blum@linux.dev
Signed-off-by: Thorsten Blum <thorsten.blum@linux.dev>
Reviewed-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/hugetlb: move hugetlb CMA code in to its own file
Frank van der Linden [Fri, 28 Feb 2025 18:29:28 +0000 (18:29 +0000)]
mm/hugetlb: move hugetlb CMA code in to its own file

hugetlb.c contained a number of CONFIG_CMA ifdefs, and the code inside
them was large enough to merit being in its own file, so move it, cleaning
up things a bit.

Hide some direct variable access behind functions to accommodate the move.

No functional change intended.

Link: https://lkml.kernel.org/r/20250228182928.2645936-28-fvdl@google.com
Signed-off-by: Frank van der Linden <fvdl@google.com>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Joao Martins <joao.m.martins@oracle.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Cc: Usama Arif <usamaarif642@gmail.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/hugetlb: enable bootmem allocation from CMA areas
Frank van der Linden [Fri, 28 Feb 2025 18:29:27 +0000 (18:29 +0000)]
mm/hugetlb: enable bootmem allocation from CMA areas

If hugetlb_cma_only is enabled, we know that hugetlb pages can only be
allocated from CMA.  Now that there is an interface to do early
reservations from a CMA area (returning memblock memory), it can be used
to allocate hugetlb pages from CMA.

This also allows for doing pre-HVO on these pages (if enabled).

Make sure to initialize the page structures and associated data correctly.
Create a flag to signal that a hugetlb page has been allocated from CMA
to make things a little easier.

Some configurations of powerpc have a special hugetlb bootmem allocator,
so introduce a boolean arch_specific_huge_bootmem_alloc that returns true
if such an allocator is present.  In that case, CMA bootmem allocations
can't be used, so check that function before trying.

Link: https://lkml.kernel.org/r/20250228182928.2645936-27-fvdl@google.com
Signed-off-by: Frank van der Linden <fvdl@google.com>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Joao Martins <joao.m.martins@oracle.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Cc: Usama Arif <usamaarif642@gmail.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/hugetlb: add hugetlb_cma_only cmdline option
Frank van der Linden [Fri, 28 Feb 2025 18:29:26 +0000 (18:29 +0000)]
mm/hugetlb: add hugetlb_cma_only cmdline option

Add an option to force hugetlb gigantic pages to be allocated using CMA
only (if hugetlb_cma is enabled).  This avoids a fallback to allocation
from the rest of system memory if the CMA allocation fails.  This makes
the size of hugetlb_cma a hard upper boundary for gigantic hugetlb page
allocations.

This is useful because, with a large CMA area, the kernel's unmovable
allocations will have less room to work with and it is undesirable for new
hugetlb gigantic page allocations to be done from that remaining area.  It
will eat in to the space available for unmovable allocations, leading to
unwanted system behavior (OOMs because the kernel fails to do unmovable
allocations).

So, with this enabled, an administrator can force a hard upper bound for
runtime gigantic page allocations, and have more predictable system
behavior.

Link: https://lkml.kernel.org/r/20250228182928.2645936-26-fvdl@google.com
Signed-off-by: Frank van der Linden <fvdl@google.com>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Joao Martins <joao.m.martins@oracle.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Cc: Usama Arif <usamaarif642@gmail.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/cma: introduce interface for early reservations
Frank van der Linden [Fri, 28 Feb 2025 18:29:25 +0000 (18:29 +0000)]
mm/cma: introduce interface for early reservations

It can be desirable to reserve memory in a CMA area before it is
activated, early in boot.  Such reservations would effectively be memblock
allocations, but they can be returned to the CMA area later.  This
functionality can be used to allow hugetlb bootmem allocations from a
hugetlb CMA area.

A new interface, cma_reserve_early is introduced.  This allows for
pageblock-aligned reservations.  These reservations are skipped during the
initial handoff of pages in a CMA area to the buddy allocator.  The caller
is responsible for making sure that the page structures are set up, and
that the migrate type is set correctly, as with other memblock allocations
that stick around.  If the CMA area fails to activate (because it
intersects with multiple zones), the reserved memory is not given to the
buddy allocator, the caller needs to take care of that.

Link: https://lkml.kernel.org/r/20250228182928.2645936-25-fvdl@google.com
Signed-off-by: Frank van der Linden <fvdl@google.com>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Joao Martins <joao.m.martins@oracle.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Cc: Usama Arif <usamaarif642@gmail.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/cma: introduce a cma validate function
Frank van der Linden [Fri, 28 Feb 2025 18:29:24 +0000 (18:29 +0000)]
mm/cma: introduce a cma validate function

Define a function to check if a CMA area is valid, which means: do its
ranges not cross any zone boundaries.  Store the result in the newly
created flags for each CMA area, so that multiple calls are dealt with.

This allows for checking the validity of a CMA area early, which is needed
later in order to be able to allocate hugetlb bootmem pages from it with
pre-HVO.

Link: https://lkml.kernel.org/r/20250228182928.2645936-24-fvdl@google.com
Signed-off-by: Frank van der Linden <fvdl@google.com>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Joao Martins <joao.m.martins@oracle.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Cc: Usama Arif <usamaarif642@gmail.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agomm/cma: simplify zone intersection check
Frank van der Linden [Fri, 28 Feb 2025 18:29:23 +0000 (18:29 +0000)]
mm/cma: simplify zone intersection check

cma_activate_area walks all pages in the area, checking their zone
individually to see if the area resides in more than one zone.

Make this a little more efficient by using the recently introduced
pfn_range_intersects_zones() function.  Store the NUMA node id (if any) in
the cma structure to facilitate this.

Link: https://lkml.kernel.org/r/20250228182928.2645936-23-fvdl@google.com
Signed-off-by: Frank van der Linden <fvdl@google.com>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Joao Martins <joao.m.martins@oracle.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Cc: Usama Arif <usamaarif642@gmail.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2 months agox86/mm: set ARCH_WANT_HUGETLB_VMEMMAP_PREINIT
Frank van der Linden [Fri, 28 Feb 2025 18:29:22 +0000 (18:29 +0000)]
x86/mm: set ARCH_WANT_HUGETLB_VMEMMAP_PREINIT

Now that hugetlb bootmem pages are allocated earlier, and available for
section preinit (HVO-style), set ARCH_WANT_HUGETLB_VMEMMAP_PREINIT for
x86_64, so that is can be done.

This enables pre-HVO on x86_64.

Link: https://lkml.kernel.org/r/20250228182928.2645936-22-fvdl@google.com
Signed-off-by: Frank van der Linden <fvdl@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Joao Martins <joao.m.martins@oracle.com>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Cc: Usama Arif <usamaarif642@gmail.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>