mm/vma.h: Optimise vma_munmap_struct

The vma_munmap_struct has a hole of 4 bytes and pushes the struct to
three cachelines.  Relocating the three booleans upwards allows for the
struct to only use two cachelines (as reported by pahole on amd64).

Before:
struct vma_munmap_struct {
        struct vma_iterator *      vmi;                  /*     0     8 */
        struct vm_area_struct *    vma;                  /*     8     8 */
        struct vm_area_struct *    prev;                 /*    16     8 */
        struct vm_area_struct *    next;                 /*    24     8 */
        struct list_head *         uf;                   /*    32     8 */
        long unsigned int          start;                /*    40     8 */
        long unsigned int          end;                  /*    48     8 */
        long unsigned int          unmap_start;          /*    56     8 */
        /* --- cacheline 1 boundary (64 bytes) --- */
        long unsigned int          unmap_end;            /*    64     8 */
        int                        vma_count;            /*    72     4 */

        /* XXX 4 bytes hole, try to pack */

        long unsigned int          nr_pages;             /*    80     8 */
        long unsigned int          locked_vm;            /*    88     8 */
        long unsigned int          nr_accounted;         /*    96     8 */
        long unsigned int          exec_vm;              /*   104     8 */
        long unsigned int          stack_vm;             /*   112     8 */
        long unsigned int          data_vm;              /*   120     8 */
        /* --- cacheline 2 boundary (128 bytes) --- */
        bool                       unlock;               /*   128     1 */
        bool                       clear_ptes;           /*   129     1 */
        bool                       closed_vm_ops;        /*   130     1 */

        /* size: 136, cachelines: 3, members: 19 */
        /* sum members: 127, holes: 1, sum holes: 4 */
        /* padding: 5 */
        /* last cacheline: 8 bytes */
};

After:
struct vma_munmap_struct {
        struct vma_iterator *      vmi;                  /*     0     8 */
        struct vm_area_struct *    vma;                  /*     8     8 */
        struct vm_area_struct *    prev;                 /*    16     8 */
        struct vm_area_struct *    next;                 /*    24     8 */
        struct list_head *         uf;                   /*    32     8 */
        long unsigned int          start;                /*    40     8 */
        long unsigned int          end;                  /*    48     8 */
        long unsigned int          unmap_start;          /*    56     8 */
        /* --- cacheline 1 boundary (64 bytes) --- */
        long unsigned int          unmap_end;            /*    64     8 */
        int                        vma_count;            /*    72     4 */
        bool                       unlock;               /*    76     1 */
        bool                       clear_ptes;           /*    77     1 */
        bool                       closed_vm_ops;        /*    78     1 */

        /* XXX 1 byte hole, try to pack */

        long unsigned int          nr_pages;             /*    80     8 */
        long unsigned int          locked_vm;            /*    88     8 */
        long unsigned int          nr_accounted;         /*    96     8 */
        long unsigned int          exec_vm;              /*   104     8 */
        long unsigned int          stack_vm;             /*   112     8 */
        long unsigned int          data_vm;              /*   120     8 */

        /* size: 128, cachelines: 2, members: 19 */
        /* sum members: 127, holes: 1, sum holes: 1 */
};

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>

mm/vma: Drop incorrect comment from vms_gather_munmap_vmas()

The comment has been outdated since 6b73cff239e52 ("mm: change munmap
splitting order and move_vma()").  The move_vma() was altered to fix the
fragile state of the accounting since then.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm: Move may_expand_vm() check in mmap_region()

The may_expand_vm() check requires the count of the pages within the
munmap range.  Since this is needed for accounting and obtained later,
the reodering of ma_expand_vm() to later in the call stack, after the
vma munmap struct (vms) is initialised and the gather stage is
potentially run, will allow for a single loop over the vmas.  The gather
sage does not commit any work and so everything can be undone in the
case of a failure.

The MAP_FIXED page count is available after the vms_gather_munmap_vmas()
call, so use it instead of looping over the vmas twice.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

ipc/shm, mm: Drop do_vma_munmap()

The do_vma_munmap() wrapper existed for callers that didn't have a vma
iterator and needed to check the vma mseal status prior to calling the
underlying munmap().  All callers now use a vma iterator and since the
mseal check has been moved to do_vmi_align_munmap() and the vmas are
aligned, this function can just be called instead.

do_vmi_align_munmap() can no longer be static as ipc/shm is using it and
it is exported via the mm.h header.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm/mmap: Use vms accounted pages in mmap_region()

Change from nr_pages variable to vms.nr_accounted for the charged pages
calculation.  This is necessary for a future patch.

This also avoids checking security_vm_enough_memory_mm() if the amount
of memory won't change.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Cc: Kees Cook <kees@kernel.org>
Cc: linux-security-module@vger.kernel.org
Reviewed-by: Kees Cook <kees@kernel.org>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Acked-by: Paul Moore <paul@paul-moore.com> (LSM)

mm/mmap: Use PHYS_PFN in mmap_region()

Instead of shifting the length by PAGE_SIZE, use PHYS_PFN.  Also use the
existing local variable everywhere instead of some of the time.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>

mm/vma: Fix null pointer dereference in vms_abort_munmap_vmas()

Don't pass a NULL vma to the vma_iter_store(), instead set up the maple
state for the store and do it manually.  vma_iter_clear() cannot be used
as it needs preallocations.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reported-by: Dan Carpenter <dan.carpenter@linaro.org>

mm: Change failure of MAP_FIXED to restoring the gap on failure

Prior to call_mmap(), the vmas that will be replaced need to clear the
way for what may happen in the call_mmap().  This clean up work includes
clearing the ptes and calling the close() vm_ops.  Some users do more
setup than can be restored by calling the vm_ops open() function.  It is
safer to store the gap in the vma tree in these cases.

That is to say that the failure scenario that existed before the
MAP_FIXED gap exposure is restored as it is safer than trying to undo a
partial mapping.

Since abort_munmap_vmas() is only reattaching vmas with this change, the
function is renamed to reattach_vmas().

There is also a secondary failure that may occur if there is not enough
memory to store the gap.  In this case, the vmas are reattached and
resources freed.  If the system cannot complete the call_mmap() and
fails to allocate with GFP_KERNEL, then the system will print a warning
about the failure.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm/mmap: Avoid zeroing vma tree in mmap_region()

Instead of zeroing the vma tree and then overwriting the area, let the
area be overwritten and then clean up the gathered vmas using
vms_complete_munmap_vmas().

To ensure locking is downgraded correctly, the mm is set regardless of
MAP_FIXED or not (NULL vma).

If a driver is mapping over an existing vma, then clear the ptes before
the call_mmap() invocation.  This is done using the vms_clean_up_area()
helper.  If there is a close vm_ops, that must also be called to ensure
any cleanup is done before mapping over the area.  This also means that
calling open has been added to the abort of an unmap operation, for now.

Since vm_ops->open() and vm_ops->close() are not always undo each other
(state cleanup may exist in ->close() that is lost forever), the code
cannot be left in this way, but that change has been isolated to another
commit to make this point very obvious for traceability.

Temporarily keep track of the number of pages that will be removed and
reduce the charged amount.

This also drops the validate_mm() call in the vma_expand() function.
It is necessary to drop the validate as it would fail since the mm
map_count would be incorrect during a vma expansion, prior to the
cleanup from vms_complete_munmap_vmas().

Clean up the error handing of the vms_gather_munmap_vmas() by calling
the verification within the function.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm: Clean up unmap_region() argument list

With the only caller to unmap_region() being the error path of
mmap_region(), the argument list can be significantly reduced.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm/vma: Track start and end for munmap in vma_munmap_struct

Set the start and end address for munmap when the prev and next are
gathered.  This is needed to avoid incorrect addresses being used during
the vms_complete_munmap_vmas() function if the prev/next vma are
expanded.

Add a new helper vms_complete_pte_clear(), which is needed later and
will avoid growing the argument list to unmap_region() beyond the 9 it
already has.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm/mmap: Reposition vma iterator in mmap_region()

Instead of moving (or leaving) the vma iterator pointing at the previous
vma, leave it pointing at the insert location.  Pointing the vma
iterator at the insert location allows for a cleaner walk of the vma
tree for MAP_FIXED and the no expansion cases.

The vma_prev() call in the case of merging the previous vma is
equivalent to vma_iter_prev_range(), since the vma iterator will be
pointing to the location just before the previous vma.

This change needs to export abort_munmap_vmas() from mm/vma.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm/vma: Support vma == NULL in init_vma_munmap()

Adding support for a NULL vma means the init_vma_munmap() can be
initialized for a less error-prone process when calling
vms_complete_munmap_vmas() later on.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm/vma: Expand mmap_region() munmap call

Open code the do_vmi_align_munmap() call so that it can be broken up
later in the series.

This requires exposing a few more vma operations.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm/vma: Inline munmap operation in mmap_region()

mmap_region is already passed sanitized addr and len, so change the
call to do_vmi_munmap() to do_vmi_align_munmap() and inline the other
checks.

The inlining of the function and checks is an intermediate step in the
series so future patches are easier to follow.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm/vma: Extract validate_mm() from vma_complete()

vma_complete() will need to be called during an unsafe time to call
validate_mm().  Extract the call in all places now so that only one
location can be modified in the next change.

Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm/vma: fix bookkeeping checks

Switch the order of the checking.

Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>

mm/vma: Change munmap to use vma_munmap_struct() for accounting and surrounding
vmas

Clean up the code by changing the munmap operation to use a structure
for the accounting and munmap variables.

Since remove_mt() is only called in one location and the contents will
be reduced to almost nothing.  The remains of the function can be added
to vms_complete_munmap_vmas().

Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>

mm/vma: Introduce vma_munmap_struct for use in munmap operations

Use a structure to pass along all the necessary information and counters
involved in removing vmas from the mm_struct.

Update vmi_ function names to vms_ to indicate the first argument
type change.

Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm/vma: Extract the gathering of vmas from do_vmi_align_munmap()

Create vmi_gather_munmap_vmas() to handle the gathering of vmas into a
detached maple tree for removal later.  Part of the gathering is the
splitting of vmas that span the boundary.

Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>

mm/vma: Introduce vmi_complete_munmap_vmas()

Extract all necessary operations that need to be completed after the vma
maple tree is updated from a munmap() operation.  Extracting this makes
the later patch in the series easier to understand.

Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>

mm/vma: Introduce abort_munmap_vmas()

Extract clean up of failed munmap() operations from
do_vmi_align_munmap().  This simplifies later patches in the series.

It is worth noting that the mas_for_each() loop now has a different
upper limit.  This should not change the number of vmas visited for
reattaching to the main vma tree (mm_mt), as all vmas are reattached in
both scenarios.

Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>

mm/vma: Correctly position vma_iterator in __split_vma()

The vma iterator may be left pointing to the newly created vma.  This
happens when inserting the new vma at the end of the old vma
(!new_below).

The incorrect position in the vma iterator is not exposed currently
since the vma iterator is repositioned in the munmap path and is not
reused in any of the other paths.

This has limited impact in the current code, but is required for future
changes.

Fixes: b2b3b886738f ("mm: don't use __vma_adjust() in __split_vma()")
Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Lorenzo Stoakes <lstoakes@gmail.com>

selftests/mm: add more mseal traversal tests

Add more mseal traversal tests across VMAs, where we could possibly
screw up sealing checks. These test more across-vma traversal for
mprotect, munmap and madvise. Particularly, we test for the case where a
regular VMA is followed by a sealed VMA.

Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>

mm: Remove can_modify_mm()

With no more users in the tree, we can finally remove can_modify_mm().

Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>

mseal: Replace can_modify_mm_madv with a vma variant

Replace can_modify_mm_madv() with a single vma variant, and associated
checks in madvise.

While we're at it, also invert the order of checks in:
 if (unlikely(is_ro_anon(vma) && !can_modify_vma(vma))

Checking if we can modify the vma itself (through vm_flags) is
certainly cheaper than is_ro_anon() due to arch_vma_access_permitted()
looking at e.g pkeys registers (with extra branches) in some
architectures.

This patch allows for partial madvise success when finding a sealed VMA,
which historically has been allowed in Linux.

Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>

mm/mremap: Replace can_modify_mm with can_modify_vma

Delegate all can_modify checks to the proper places. Unmap checks are
done in do_unmap (et al). The source VMA check is done purposefully
before unmapping, to keep the original mseal semantics.

Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>

mm/mprotect: Replace can_modify_mm with can_modify_vma

Avoid taking an extra trip down the mmap tree by checking the vmas
directly. mprotect (per POSIX) tolerates partial failure.

Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>

mm/munmap: Replace can_modify_mm with can_modify_vma

We were doing an extra mmap tree traversal just to check if the entire
range is modifiable. This can be done when we iterate through the VMAs
instead.

Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>

mm: Move can_modify_vma to mm/vma.h

Move can_modify_vma to vma.h so it can be inlined properly (with
the intent to remove can_modify_mm callsites).

Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>

mm: optimization on page allocation when CMA enabled

According to current CMA utilization policy, an alloc_pages(GFP_USER)
could 'steal' UNMOVABLE & RECLAIMABLE page blocks via the help of CMA(pass
zone_watermark_ok by counting CMA in but use U&R in rmqueue), which could
lead to following alloc_pages(GFP_KERNEL) fail.  Solving this by
introducing second watermark checking for GFP_MOVABLE, which could have
the allocation use CMA when proper.

-- Free_pages(30MB)
|
|
-- WMARK_LOW(25MB)
|
-- Free_CMA(12MB)
|
|
--

Link: https://lkml.kernel.org/r/20231016071245.2865233-1-zhaoyang.huang@unisoc.com
Link: https://lkml.kernel.org/r/1683782550-25799-1-git-send-email-zhaoyang.huang@unisoc.com
Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: ke.wang <ke.wang@unisoc.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Zhaoyang Huang <huangzhaoyang@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm/memory-failure: fix VM_BUG_ON_PAGE(PagePoisoned(page)) when unpoison memory

When I did memory failure tests recently, below panic occurs:

page dumped because: VM_BUG_ON_PAGE(PagePoisoned(page))
kernel BUG at include/linux/page-flags.h:616!
Oops: invalid opcode: 0000 [#1] PREEMPT SMP NOPTI
CPU: 3 PID: 720 Comm: bash Not tainted 6.10.0-rc1-00195-g148743902568 #40
RIP: 0010:unpoison_memory+0x2f3/0x590
RSP: 0018:ffffa57fc8787d60 EFLAGS: 00000246
RAX: 0000000000000037 RBX: 0000000000000009 RCX: ffff9be25fcdc9c8
RDX: 0000000000000000 RSI: 0000000000000027 RDI: ffff9be25fcdc9c0
RBP: 0000000000300000 R08: ffffffffb4956f88 R09: 0000000000009ffb
R10: 0000000000000284 R11: ffffffffb4926fa0 R12: ffffe6b00c000000
R13: ffff9bdb453dfd00 R14: 0000000000000000 R15: fffffffffffffffe
FS:  00007f08f04e4740(0000) GS:ffff9be25fcc0000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000564787a30410 CR3: 000000010d4e2000 CR4: 00000000000006f0
Call Trace:
 <TASK>
 unpoison_memory+0x2f3/0x590
 simple_attr_write_xsigned.constprop.0.isra.0+0xb3/0x110
 debugfs_attr_write+0x42/0x60
 full_proxy_write+0x5b/0x80
 vfs_write+0xd5/0x540
 ksys_write+0x64/0xe0
 do_syscall_64+0xb9/0x1d0
 entry_SYSCALL_64_after_hwframe+0x77/0x7f
RIP: 0033:0x7f08f0314887
RSP: 002b:00007ffece710078 EFLAGS: 00000246 ORIG_RAX: 0000000000000001
RAX: ffffffffffffffda RBX: 0000000000000009 RCX: 00007f08f0314887
RDX: 0000000000000009 RSI: 0000564787a30410 RDI: 0000000000000001
RBP: 0000564787a30410 R08: 000000000000fefe R09: 000000007fffffff
R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000009
R13: 00007f08f041b780 R14: 00007f08f0417600 R15: 00007f08f0416a00
 </TASK>
Modules linked in: hwpoison_inject
---[ end trace 0000000000000000 ]---
RIP: 0010:unpoison_memory+0x2f3/0x590
RSP: 0018:ffffa57fc8787d60 EFLAGS: 00000246
RAX: 0000000000000037 RBX: 0000000000000009 RCX: ffff9be25fcdc9c8
RDX: 0000000000000000 RSI: 0000000000000027 RDI: ffff9be25fcdc9c0
RBP: 0000000000300000 R08: ffffffffb4956f88 R09: 0000000000009ffb
R10: 0000000000000284 R11: ffffffffb4926fa0 R12: ffffe6b00c000000
R13: ffff9bdb453dfd00 R14: 0000000000000000 R15: fffffffffffffffe
FS:  00007f08f04e4740(0000) GS:ffff9be25fcc0000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000564787a30410 CR3: 000000010d4e2000 CR4: 00000000000006f0
Kernel panic - not syncing: Fatal exception
Kernel Offset: 0x31c00000 from 0xffffffff81000000 (relocation range: 0xffffffff80000000-0xffffffffbfffffff)
---[ end Kernel panic - not syncing: Fatal exception ]---

We're hitting a BUG_ON in PF_ANY():

PAGEFLAG(HWPoison, hwpoison, PF_ANY)

#define PF_ANY(page, enforce) PF_POISONED_CHECK(page)

#define PF_POISONED_CHECK(page) ({ \
VM_BUG_ON_PGFLAGS(PagePoisoned(page), page); \
page; })

#define PAGE_POISON_PATTERN -1l
static inline int PagePoisoned(const struct page *page)
{
return READ_ONCE(page->flags) == PAGE_POISON_PATTERN;
}

The offlined pages will have page->flags set to PAGE_POISON_PATTERN
while pfn is still valid:

offline_pages
  remove_pfn_range_from_zone
    page_init_poison
      memset(page, PAGE_POISON_PATTERN, size);

The root cause is that unpoison_memory() tries to check the PG_HWPoison
flags of an uninitialized page. So VM_BUG_ON_PAGE(PagePoisoned(page)) is
triggered. This can be reproduced by below steps:
1.Offline memory block:
 echo offline > /sys/devices/system/memory/memory12/state
2.Get offlined memory pfn:
 page-types -b n -rlN
3.Write pfn to unpoison-pfn
 echo <pfn> > /sys/kernel/debug/hwpoison/unpoison-pfn

Link: https://lkml.kernel.org/r/20240712064249.3882707-1-linmiaohe@huawei.com
Fixes: f165b378bbdf ("mm: uninitialized struct page poisoning sanity checking")
Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm/hugetlb_cgroup: introduce peak and rsvd.peak to v2

Introduce peak and rsvd.peak to v2 to show the historical maximum usage of
resources, as in some scenarios it is necessary to configure the value of
max/rsvd.max based on the peak usage of resources.

Since HugeTLB doesn't support page reclaim, enforcing the limit at page
fault time implies that, the application will get SIGBUS signal if it
tries to fault in HugeTLB pages beyond its limit.  Therefore the
application needs to know exactly how many HugeTLB pages it uses before
hand, and the sysadmin needs to make sure that there are enough available
on the machine for all the users to avoid processes getting SIGBUS.

When running some open-source software, it may not be possible to know the
exact amount of hugetlb it consumes, so cannot correctly configure the max
value.  If there is a peak metric, we can run the open-source software
first and then configure the max based on the peak value.  In cgroup v1,
the hugetlb controller provides the max_usage_in_bytes and
rsvd.max_usage_in_bytes interface to display the historical maximum usage,
so introduce peak and rsvd.peak to v2 to address this issue.

Link: https://lkml.kernel.org/r/20240702125728.2743143-1-xiujianfeng@huawei.com
Signed-off-by: Xiu Jianfeng <xiujianfeng@huawei.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: add lazyfree folio to lru tail

Replace lruvec_add_folio with lruvec_add_folio_tail in the lru_lazyfree_fn:

1. The lazy-free folio is added to the LRU_INACTIVE_FILE list.  If
   it's moved to the LRU tail, it allows for faster release lazy-free
   folio and reduces the impact on file refault.

2. When mglru is enabled, the lazy-free folio is reclaimabled and
   should be added using lru_gen_add_folio(lruvec, folio, true) instead of
   lru_gen_add_folio(lruvec, folio, false) for adding to gen.

With the change in place, workingset_refault_file is reduced by 33% in the
continuous startup testing of the applications in the Android system.

Link: https://lkml.kernel.org/r/f29f64e29c08427b95e3df30a5770056@honor.com
Signed-off-by: gao xu <gaoxu2@hihonor.com>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: David Hildenbrand <david@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: memory_hotplug: unify Huge/LRU/non-LRU movable folio isolation

Use the isolate_folio_to_list() to unify hugetlb/LRU/non-LRU folio
isolation, which cleanup code a bit and save a few calls to
compound_head().

Link: https://lkml.kernel.org/r/20240816090435.888946-6-wangkefeng.wang@huawei.com
Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: migrate: add isolate_folio_to_list()

Add isolate_folio_to_list() helper to try to isolate HugeTLB, no-LRU
movable and LRU folios to a list, which will be reused by
do_migrate_range() from memory hotplug soon, also drop the
mf_isolate_folio() since we could directly use new helper in the
soft_offline_in_use_page().

Link: https://lkml.kernel.org/r/20240816090435.888946-5-wangkefeng.wang@huawei.com
Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: memory_hotplug: check hwpoisoned page firstly in do_migrate_range()

The commit b15c87263a69 ("hwpoison, memory_hotplug: allow hwpoisoned pages
to be offlined") don't handle the hugetlb pages, the endless loop still
occur if offline a hwpoison hugetlb, luckly, with the commit e591ef7d96d6
("mm,hwpoison,hugetlb,memory_hotplug: hotremove memory section with
hwpoisoned hugepage") section with hwpoisoned hugepage"), the
HPageMigratable of hugetlb page will be clear, and the hwpoison hugetlb
page will be skipped in scan_movable_pages(), so the endless loop issue is
fixed.

However if the HPageMigratable() check passed(without reference and lock),
the hugetlb page may be hwpoisoned, it won't cause issue since the
hwpoisoned page will be handled correctly in the next movable pages scan
loop, and it will be isolated in do_migrate_range() but fails to migrate.
In order to avoid the unnecessary isolation and unify all hwpoisoned page
handling, let's unconditionally check hwpoison firstly, and if it is a
hwpoisoned hugetlb page, try to unmap it as the catch all safety net like
normal page does.

Link: https://lkml.kernel.org/r/20240816090435.888946-4-wangkefeng.wang@huawei.com
Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: memory-failure: add unmap_posioned_folio()

Add unmap_posioned_folio() helper which will be reused by
do_migrate_range() from memory hotplug soon.

Link: https://lkml.kernel.org/r/20240816090435.888946-3-wangkefeng.wang@huawei.com
Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: memory_hotplug: remove head variable in do_migrate_range()

Patch series "mm: memory_hotplug: improve do_migrate_range()", v2.

Unify hwpoisoned page handling and isolation of HugeTLB/LRU/non-LRU
movable page, also convert to use folios in do_migrate_range().

This patch (of 2):

Directly use a folio for HugeTLB and THP when calculate the next pfn, then
remove unused head variable.

Link: https://lkml.kernel.org/r/20240816090435.888946-1-wangkefeng.wang@huawei.com
Link: https://lkml.kernel.org/r/20240816090435.888946-2-wangkefeng.wang@huawei.com
Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm/rmap: use folio->_mapcount for small folios

We have some cases left whereby we operate on small folios and still refer
to page->_mapcount.  Let's just use folio->_mapcount instead, which
currently still overlays page->_mapcount, so no change.

This change will make it easier to later spot any remaining users of
page->_mapcount that target tail pages.

Link: https://lkml.kernel.org/r/20240816103246.719209-1-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

selftests: test_zswap: add test for hierarchical zswap.writeback

Ensure that zswap.writeback check goes up the cgroup tree.

Link: https://lkml.kernel.org/r/20240816144344.18135-2-me@yhndnzj.com
Signed-off-by: Mike Yuan <me@yhndnzj.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm/memcontrol: respect zswap.writeback setting from parent cg too

Currently, the behavior of zswap.writeback wrt.  the cgroup hierarchy
seems a bit odd.  Unlike zswap.max, it doesn't honor the value from parent
cgroups.  This surfaced when people tried to globally disable zswap
writeback, i.e.  reserve physical swap space only for hibernation [1] -
disabling zswap.writeback only for the root cgroup results in subcgroups
with zswap.writeback=3D1 still performing writeback.

The inconsistency became more noticeable after I introduced the
MemoryZSwapWriteback=3D systemd unit setting [2] for controlling the knob.
The patch assumed that the kernel would enforce the value of parent
cgroups.  It could probably be workarounded from systemd's side, by going
up the slice unit tree and inheriting the value.  Yet I think it's more
sensible to make it behave consistently with zswap.max and friends.

[1] https://wiki.archlinux.org/title/Power_management/Suspend_and_hibernate=
#Disable_zswap_writeback_to_use_the_swap_space_only_for_hibernation
[2] https://github.com/systemd/systemd/pull/31734

Link: https://lkml.kernel.org/r/20240816144344.18135-1-me@yhndnzj.com
Signed-off-by: Mike Yuan <me@yhndnzj.com>
Reviewed-by: Nhat Pham <nphamcs@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

memcg: initiate deprecation of pressure_level

The pressure_level in memcg v1 provides memory pressure notifications to
the user space.  At the moment it provides notifications for three levels
of memory pressure i.e.  low, medium and critical, which are defined based
on internal memory reclaim implementation details.  More specifically the
ratio of scanned and reclaimed pages during a memory reclaim.  However
this is not robust as there are workloads with mostly unreclaimable user
memory or kernel memory.

For v2, the users can use PSI for memory pressure status of the system or
the cgroup.  Let's start the deprecation process for pressure_level and
add warnings to gather the info on how the current users are using this
interface and how they can be used to PSI.

Link: https://lkml.kernel.org/r/20240814220021.3208384-5-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Reviewed-by: T.J. Mercier <tjmercier@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

memcg: initiate deprecation of oom_control

The oom_control provides functionality to disable memcg oom-killer,
notifications on oom-kill and reading the stats regarding oom-kills.  This
interface was mainly introduced to provide functionality for userspace
oom-killers.  However it is not robust enough and only supports OOM
handling in the page fault path.

For v2, the users can use the combination of memory.events notifications,
memory.high and PSI to provide userspace OOM-killing functionality.
Actually LMKD in Android and OOMd in systemd and Meta infrastructure
already use PSI in combination with other stats to implement userspace
OOM-killing.

Let's start the deprecation process for v1 and gather the info on how the
current users are using this interface and work on providing a more robust
functionality in v2.

Link: https://lkml.kernel.org/r/20240814220021.3208384-4-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Reviewed-by: T.J. Mercier <tjmercier@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

memcg: initiate deprecation of v1 soft limit

Memcg v1 provides soft limit functionality for the best effort memory
sharing between multiple workloads on a system.  It is usually triggered
through kswapd and at the moment does not reclaim kernel memory.

Memcg v2 provides more straightforward best effort (memory.low) and hard
protection (memory.min) functionalities.  Let's initiate the deprecation
of soft limit from v1 and gather if v2 needs something more to move the
existing v1 users to v2 regarding soft limit.

Link: https://lkml.kernel.org/r/20240814220021.3208384-3-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Reviewed-by: T.J. Mercier <tjmercier@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

memcg: initiate deprecation of v1 tcp accounting

Patch series "memcg: initiate deprecation of v1 features", v2.

Start the deprecation process of the memcg v1 features which we discussed
during LSFMMBPF 2024 [1].  For now add the warnings to collect the
information on how the current users are using these features.  Next we
will work on providing better alternatives in v2 (if needed) and fully
deprecate these features.

Link: https://lwn.net/Articles/974575
This patch (of 4):

Memcg v1 provides opt-in TCP memory accounting feature.  However it is
mostly unused due to its performance impact on the network traffic.  In
v2, the TCP memory is accounted in the regular memory usage and is
transparent to the users but they can observe the TCP memory usage through
memcg stats.

Let's initiate the deprecation process of memcg v1's tcp accounting
functionality and add warnings to gather if there are any users and if
there are, collect how they are using it and plan to provide them better
alternative in v2.

Link: https://lkml.kernel.org/r/20240814220021.3208384-1-shakeel.butt@linux.dev
Link: https://lkml.kernel.org/r/20240814220021.3208384-2-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Reviewed-by: T.J. Mercier <tjmercier@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

memcg: make PGPGIN and PGPGOUT v1 only

Currently PGPGIN and PGPGOUT are used and exposed in the memcg v1 only
code.  So, let's put them under CONFIG_MEMCG_V1.

Link: https://lkml.kernel.org/r/20240815050453.1298138-8-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 <roman.gushchin@linux.dev>
Cc: T.J. Mercier <tjmercier@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

memcg: allocate v1 event percpu only on v1 deployment

Currently memcg->events_percpu gets allocated on v2 deployments.  Let's
move the allocation to v1 only codebase.  This is not needed in v2.

Link: https://lkml.kernel.org/r/20240815050453.1298138-7-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 <roman.gushchin@linux.dev>
Cc: T.J. Mercier <tjmercier@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

memcg: make v1 only functions static

The functions memcg1_charge_statistics() and memcg1_check_events() are
never used outside of v1 source file.  So, make them static.

Link: https://lkml.kernel.org/r/20240815050453.1298138-6-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 <roman.gushchin@linux.dev>
Cc: T.J. Mercier <tjmercier@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

memcg: move v1 events and statistics code to v1 file

Currently the common code path for charge commit, swapout and batched
uncharge are executing v1 only code which is completely useless for the v2
deployments where CONFIG_MEMCG_V1 is disabled.  In addition, it is mucking
with IRQs which might be slow on some architectures.  Let's move all of
this code to v1 only code and remove them from v2 only deployments.

Link: https://lkml.kernel.org/r/20240815050453.1298138-5-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 <roman.gushchin@linux.dev>
Cc: T.J. Mercier <tjmercier@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

memcg: move mem_cgroup_charge_statistics to v1 code

There are no callers of mem_cgroup_charge_statistics() in the v2 code
base, so move it to the v1 only code and rename it to
memcg1_charge_statistics().

Link: https://lkml.kernel.org/r/20240815050453.1298138-4-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 <roman.gushchin@linux.dev>
Cc: T.J. Mercier <tjmercier@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

memcg: move mem_cgroup_event_ratelimit to v1 code

There are no callers of mem_cgroup_event_ratelimit() in the v2 code.  Move
it to v1 only code and rename it to memcg1_event_ratelimit().

Link: https://lkml.kernel.org/r/20240815050453.1298138-3-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 <roman.gushchin@linux.dev>
Cc: T.J. Mercier <tjmercier@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

memcg: move v1 only percpu stats in separate struct

Patch series "memcg: further decouple v1 code from v2".

Some of the v1 code is still in v2 code base due to v1 fields in the
struct memcg_vmstats_percpu.  This field decouples those fileds from v2
struct and move all the related code into v1 only code base.

This patch (of 7):

At the moment struct memcg_vmstats_percpu contains two v1 only fields
which consumes memory even when CONFIG_MEMCG_V1 is not enabled.  In
addition there are v1 only functions accessing them and are in the main
memcontrol source file and can not be moved to v1 only source file due to
these fields.  Let's move these fields into their own struct.  Later
patches will move the functions accessing them to v1 source file and only
allocate these fields when CONFIG_MEMCG_V1 is enabled.

Link: https://lkml.kernel.org/r/20240815050453.1298138-1-shakeel.butt@linux.dev
Link: https://lkml.kernel.org/r/20240815050453.1298138-2-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: T.J. Mercier <tjmercier@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: use get_oder() and check size is is_power_of_2

Using get_order() is more robust according to Baolin.  It is also better
to filter illegal size such as 3KB, 16KB according to David.

Link: https://lkml.kernel.org/r/20240814224635.43272-1-21cnbao@gmail.com
Signed-off-by: Barry Song <v-songbaohua@oppo.com>
Suggested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Suggested-by: David Hildenbrand <david@redhat.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: override mTHP "enabled" defaults at kernel cmdline

Add thp_anon= cmdline parameter to allow specifying the default enablement
of each supported anon THP size.  The parameter accepts the following
format and can be provided multiple times to configure each size:

thp_anon=<size>,<size>[KMG]:<value>;<size>-<size>[KMG]:<value>

An example:

thp_anon=16K-64K:always;128K,512K:inherit;256K:madvise;1M-2M:never

See Documentation/admin-guide/mm/transhuge.rst for more details.

Configuring the defaults at boot time is useful to allow early user
space to take advantage of mTHP before its been configured through
sysfs.

Link: https://lkml.kernel.org/r/20240814020247.67297-1-21cnbao@gmail.com
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Co-developed-by: Barry Song <v-songbaohua@oppo.com>
Signed-off-by: Barry Song <v-songbaohua@oppo.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Lance Yang <ioworker0@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm/hugetlb: use __GFP_COMP for gigantic folios

Use __GFP_COMP for gigantic folios to greatly reduce not only the amount
of code but also the allocation and free time.

LOC (approximately): +60, -240

Allocate and free 500 1GB hugeTLB memory without HVO by:
  time echo 500 >/sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages
  time echo 0 >/sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages

       Before  After
Alloc  ~13s    ~10s
Free   ~15s    <1s

The above magnitude generally holds for multiple x86 and arm64 CPU models.

Link: https://lkml.kernel.org/r/20240814035451.773331-4-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Reported-by: Frank van der Linden <fvdl@google.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm/cma: add cma_{alloc,free}_folio()

With alloc_contig_range() and free_contig_range() supporting large folios,
CMA can allocate and free large folios too, by cma_alloc_folio() and
cma_free_folio().

Link: https://lkml.kernel.org/r/20240814035451.773331-3-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Cc: Frank van der Linden <fvdl@google.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm/contig_alloc: support __GFP_COMP

Patch series "mm/hugetlb: alloc/free gigantic folios", v2.

Use __GFP_COMP for gigantic folios can greatly reduce not only the amount
of code but also the allocation and free time.

Approximate LOC to mm/hugetlb.c: +60, -240

Allocate and free 500 1GB hugeTLB memory without HVO by:
  time echo 500 >/sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages
  time echo 0 >/sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages

       Before  After
Alloc  ~13s    ~10s
Free   ~15s    <1s

The above magnitude generally holds for multiple x86 and arm64 CPU
models.

Perf profile before:
  Alloc
    - 99.99% alloc_pool_huge_folio
       - __alloc_fresh_hugetlb_folio
          - 83.23% alloc_contig_pages_noprof
             - 47.46% alloc_contig_range_noprof
                - 20.96% isolate_freepages_range
                     16.10% split_page
                - 14.10% start_isolate_page_range
                - 12.02% undo_isolate_page_range

  Free
    - update_and_free_pages_bulk
       - 87.71% free_contig_range
          - 76.02% free_unref_page
             - 41.30% free_unref_page_commit
                - 32.58% free_pcppages_bulk
                   - 24.75% __free_one_page
               13.96% _raw_spin_trylock
         12.27% __update_and_free_hugetlb_folio

Perf profile after:
  Alloc
    - 99.99% alloc_pool_huge_folio
         alloc_gigantic_folio
       - alloc_contig_pages_noprof
          - 59.15% alloc_contig_range_noprof
             - 20.72% start_isolate_page_range
               20.64% prep_new_page
             - 17.13% undo_isolate_page_range

  Free
    - update_and_free_pages_bulk
       - __folio_put
       - __free_pages_ok
            7.46% free_tail_page_prepare
          - 1.97% free_one_page
               1.86% __free_one_page

This patch (of 3):

Support __GFP_COMP in alloc_contig_range().  When the flag is set, upon
success the function returns a large folio prepared by prep_new_page(),
rather than a range of order-0 pages prepared by split_free_pages() (which
is renamed from split_map_pages()).

alloc_contig_range() can be used to allocate folios larger than
MAX_PAGE_ORDER, e.g., gigantic hugeTLB folios.  So on the free path,
free_one_page() needs to handle that by split_large_buddy().

Link: https://lkml.kernel.org/r/20240814035451.773331-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20240814035451.773331-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Frank van der Linden <fvdl@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mmmemcg-provide-per-cgroup-counters-for-numa-balancing-operations-v4

added documentation of the memcg counters in cgroup-v2.rst

Link: https://lkml.kernel.org/r/20240814235122.252309-1-kaiyang2@cs.cmu.edu
Signed-off-by: Kaiyang Zhao <kaiyang2@cs.cmu.edu>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm,memcg: provide per-cgroup counters for NUMA balancing operations

The ability to observe the demotion and promotion decisions made by the
kernel on a per-cgroup basis is important for monitoring and tuning
containerized workloads on machines equipped with tiered memory.

Different containers in the system may experience drastically different
memory tiering actions that cannot be distinguished from the global
counters alone.

For example, a container running a workload that has a much hotter memory
accesses will likely see more promotions and fewer demotions, potentially
depriving a colocated container of top tier memory to such an extent that
its performance degrades unacceptably.

For another example, some containers may exhibit longer periods between
data reuse, causing much more numa_hint_faults than numa_pages_migrated.
In this case, tuning hot_threshold_ms may be appropriate, but the signal
can easily be lost if only global counters are available.

In the long term, we hope to introduce per-cgroup control of promotion and
demotion actions to implement memory placement policies in tiering.

This patch set adds seven counters to memory.stat in a cgroup:
numa_pages_migrated, numa_pte_updates, numa_hint_faults, pgdemote_kswapd,
pgdemote_khugepaged, pgdemote_direct and pgpromote_success.  pgdemote_*
and pgpromote_success are also available in memory.numa_stat.

count_memcg_events_mm() is added to count multiple event occurrences at
once, and get_mem_cgroup_from_folio() is added because we need to get a
reference to the memcg of a folio before it's migrated to track
numa_pages_migrated.  The accounting of PGDEMOTE_* is moved to
shrink_inactive_list() before being changed to per-cgroup.

Link: https://lkml.kernel.org/r/20240814174227.30639-1-kaiyang2@cs.cmu.edu
Signed-off-by: Kaiyang Zhao <kaiyang2@cs.cmu.edu>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Wei Xu <weixugc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: make write helper functions void

The return value of various write helper functions are not checked. We
can safely change the return type of these functions to be void.

Link: https://lkml.kernel.org/r/20240814161944.55347-18-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: remove unneeded mas_wr_walk() in mas_store_prealloc()

Users of mas_store_prealloc() enter this function with nodes already
preallocated. This means the store type must be already set. We can then
remove the call to mas_wr_store_type() and initialize the write state to
continue the partial walk that was done when determining the store type.

Link: https://lkml.kernel.org/r/20240814161944.55347-17-sidhartha.kumar@oracle.com
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: remove repeated sanity checks from write helper functions

These sanity checks are now redundant as they are already checked in
mas_wr_store_type(). We can remove them from mas_wr_append() and
mas_wr_node_store().

Link: https://lkml.kernel.org/r/20240814161944.55347-16-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: remove node allocations from various write helper functions

These write helper functions are all called from store paths which
preallocate enough nodes that will be needed for the write. There is no
more need to allocate within the functions themselves.

Link: https://lkml.kernel.org/r/20240814161944.55347-15-sidhartha.kumar@oracle.com
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: have mas_store() allocate nodes if needed

Not all users of mas_store() enter with nodes already preallocated.
Check for the MA_STATE_PREALLOC flag to decide whether to preallocate nodes
within mas_store() rather than relying on future write helper functions
to perform the allocations. This allows the write helper functions to be
simplified as they do not have to do checks to make sure there are
enough allocated nodes to perform the write.

Link: https://lkml.kernel.org/r/20240814161944.55347-14-sidhartha.kumar@oracle.com
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: remove mas_wr_modify()

There are no more users of the function, safely remove it.

Link: https://lkml.kernel.org/r/20240814161944.55347-13-sidhartha.kumar@oracle.com
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: simplify mas_commit_b_node()

The only callers of mas_commit_b_node() are those with store type of
wr_rebalance and wr_split_store. Use mas->store_type to dispatch to the
correct helper function. This allows the removal of mas_reuse_node() as
it is no longer used.

Link: https://lkml.kernel.org/r/20240814161944.55347-12-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: convert mas_insert() to preallocate nodes

By setting the store type in mas_insert(), we no longer need to use
mas_wr_modify() to determine the correct store function to use. Instead,
set the store type and call mas_wr_store_entry(). Also, pass in the
requested gfp flags to mas_insert() so they can be passed to the call to
mas_wr_preallocate().

Link: https://lkml.kernel.org/r/20240814161944.55347-11-sidhartha.kumar@oracle.com
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: use store type in mas_wr_store_entry()

When storing an entry, we can read the store type that was set from a
previous partial walk of the tree. Now that the type of store is known,
select the correct write helper function to use to complete the store.

Also noinline mas_wr_spanning_store() to limit stack frame usage in
mas_wr_store_entry() as it allocates a maple_big_node on the stack.

Link: https://lkml.kernel.org/r/20240814161944.55347-10-sidhartha.kumar@oracle.com
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: print store type in mas_dump()

Knowing the store type of the maple state could be helpful for debugging.
Have mas_dump() print mas->store_type.

Link: https://lkml.kernel.org/r/20240814161944.55347-9-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: use mas_store_gfp() in mtree_store_range()

Refactor mtree_store_range() to use mas_store_gfp() which will abstract
the store, memory allocation, and error handling.

Link: https://lkml.kernel.org/r/20240814161944.55347-8-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: preallocate nodes in mas_erase()

Use mas_wr_preallocate() in mas_erase() to preallocate enough nodes to
complete the erase.  Add error handling by skipping the store if the
preallocation lead to some error besides no memory.

Link: https://lkml.kernel.org/r/20240814161944.55347-7-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: remove mas_destroy() from mas_nomem()

Separate call to mas_destroy() from mas_nomem() so we can check for no
memory errors without destroying the current maple state in
mas_store_gfp().  We then add calls to mas_destroy() to callers of
mas_nomem().

Link: https://lkml.kernel.org/r/20240814161944.55347-6-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: introduce mas_wr_store_type()

Introduce mas_wr_store_type() which will set the correct store type based
on a walk of the tree.  In mas_wr_node_store() the <= min_slots condition
is changed to < as if new_end is = to mt_min_slots then there is not
enough room.

mas_prealloc_calc() is also introduced to abstract the calculation used to
determine the number of nodes needed for a store operation.

In this change a call to mas_reset() is removed in the error case of
mas_prealloc().  This is only needed in the MA_STATE_REBALANCE case of
mas_destroy().  We can move the call to mas_reset() directly to
mas_destroy().

Also, add a test case to validate the order that we check the store type
in is correct.  This test models a vma expanding and then shrinking which
is part of the boot process.

Link: https://lkml.kernel.org/r/20240814161944.55347-5-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: move up mas_wr_store_setup() and mas_wr_prealloc_setup()

Subsequent patches require these definitions to be higher, no functional
changes intended.

Link: https://lkml.kernel.org/r/20240814161944.55347-4-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: introduce mas_wr_prealloc_setup()

Introduce a helper function, mas_wr_prealoc_setup(), that will set up a
maple write state in order to start a walk of a maple tree.

Link: https://lkml.kernel.org/r/20240814161944.55347-3-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: introduce store_type enum

Patch series "Introduce a store type enum for the Maple tree", v4.

================================ OVERVIEW ================================

This series implements two work items[3]: "aligning mas_store_gfp() with
mas_preallocate()" and "enum for store type".

mas_store_gfp() is modified to preallocate nodes.  This simplies many of
the write helper functions by allowing them to use mas_store_gfp() rather
than open coding node allocation and error handling.

The enum defines the following store types:

enum store_type {
    wr_invalid,
    wr_new_root,
    wr_store_root,
    wr_exact_fit,
    wr_spanning_store,
    wr_split_store,
    wr_rebalance,
    wr_append,
    wr_node_store,
    wr_slot_store,
};

In the current maple tree code, a walk down the tree is done in
mas_preallocate() to determine the number of nodes needed for this write.
After node allocation, mas_wr_store_entry() will perform another walk to
determine which write helper function to use to complete the write.

Rather than performing the second walk, we can store the type of write in
the maple write state during node allocation and read this field to
complete the write.

Patches 1-16 implement this store type feature.
Patch 17 is a cleanup patch to change functions that have unused return
types to be void.

================================ RESULTS =================================

Phoronix t-test-1 (Seconds < Lower Is Better)
    v6.10-rc6
        Threads: 1
            33.15

        Threads: 2
            10.81

    v6.10-rc6 + this series
            Threads: 1
            32.69

        Threads: 2
            10.45

Stress-ng mmap
                    6.10_base  store_type_v4
Duration User        2744.65     2769.40
Duration System     10862.69    10817.59
Duration Elapsed     1477.58     1478.35

================================ TESTING =================================

Testing was done with the maple tree test suite.  A new test case is also
added to validate the order in which we test for and assign the store
type.

[1]: https://lore.kernel.org/linux-mm/80926b22-a8d2-9992-eb5e-27e2c99cf460@google.com/T/#m81044feb66765265f8ca7f21e4b4b3725b18780a
[2]: https://lore.kernel.org/linux-mm/80926b22-a8d2-9992-eb5e-27e2c99cf460@google.com/T/#mb36c6526486638e82518c0f37a428fb279c84d8a
[3]: https://lists.infradead.org/pipermail/maple-tree/2023-December/003098.html

This patch (of 17):

Add a store_type enum that is stored in ma_state.  This will be used to
keep track of partial walks of the tree so that subsequent walks can pick
up where a previous walk left off.

Link: https://lkml.kernel.org/r/20240814161944.55347-1-sidhartha.kumar@oracle.com
Link: https://lkml.kernel.org/r/20240814161944.55347-2-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: kmem: add lockdep assertion to obj_cgroup_memcg

obj_cgroup_memcg() is supposed to safe to prevent the returned memory
cgroup from being freed only when the caller is holding the rcu read lock
or objcg_lock or cgroup_mutex.  It is very easy to ignore thoes conditions
when users call some upper APIs which call obj_cgroup_memcg() internally
like mem_cgroup_from_slab_obj() (See the link below).  So it is better to
add lockdep assertion to obj_cgroup_memcg() to find those issues ASAP.

Because there is no user of obj_cgroup_memcg() holding objcg_lock to make
the returned memory cgroup safe, do not add objcg_lock assertion (We
should export objcg_lock if we really want to do).  Additionally, this is
some internal implementation detail of memcg and should not be accessible
outside memcg code.

Some users like __mem_cgroup_uncharge() do not care the lifetime of the
returned memory cgroup, which just want to know if the folio is charged to
a memory cgroup, therefore, they do not need to hold the needed locks.  In
which case, introduce a new helper folio_memcg_charged() to do this.
Compare it to folio_memcg(), it could eliminate a memory access of
objcg->memcg for kmem, actually, a really small gain.

Link: https://lore.kernel.org/all/20240718083607.42068-1-songmuchun@bytedance.com/
Link: https://lkml.kernel.org/r/20240814093415.17634-1-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

kasan-simplify-and-clarify-makefile-v2

comments fixes per Miguel

Link: https://lkml.kernel.org/r/20240814161052.10374-1-andrey.konovalov@linux.dev
Signed-off-by: Andrey Konovalov <andreyknvl@gmail.com>
Reviewed-by: Miguel Ojeda <ojeda@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

kasan: simplify and clarify Makefile

When KASAN support was being added to the Linux kernel, GCC did not yet
support all of the KASAN-related compiler options.  Thus, the KASAN
Makefile had to probe the compiler for supported options.

Nowadays, the Linux kernel GCC version requirement is 5.1+, and thus we
don't need the probing of the -fasan-shadow-offset parameter: it exists in
all 5.1+ GCCs.

Simplify the KASAN Makefile to drop CFLAGS_KASAN_MINIMAL.

Also add a few more comments and unify the indentation.

Link: https://lkml.kernel.org/r/20240813224027.84503-1-andrey.konovalov@linux.dev
Signed-off-by: Andrey Konovalov <andreyknvl@gmail.com>
Reviewed-by: Miguel Ojeda <ojeda@kernel.org>
Cc: Alexander Potapenko <glider@google.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Marco Elver <elver@google.com>
Cc: Matthew Maurer <mmaurer@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

memcg: use ratelimited stats flush in the reclaim

The Meta prod is seeing large amount of stalls in memcg stats flush from
the memcg reclaim code path.  At the moment, this specific callsite is
doing a synchronous memcg stats flush.  The rstat flush is an expensive
and time consuming operation, so concurrent relaimers will busywait on the
lock potentially for a long time.  Actually this issue is not unique to
Meta and has been observed by Cloudflare [1] as well.  For the Cloudflare
case, the stalls were due to contention between kswapd threads running on
their 8 numa node machines which does not make sense as rstat flush is
global and flush from one kswapd thread should be sufficient for all.
Simply replace the synchronous flush with the ratelimited one.

One may raise a concern on potentially using 2 sec stale (at worst) stats
for heuristics like desirable inactive:active ratio and preferring
inactive file pages over anon pages but these specific heuristics do not
require very precise stats and also are ignored under severe memory
pressure.

More specifically for this code path, the stats are needed for two
specific heuristics:

1. Deactivate LRUs
2. Cache trim mode

The deactivate LRUs heuristic is to maintain a desirable inactive:active
ratio of the LRUs.  The specific stats needed are WORKINGSET_ACTIVATE* and
the hierarchical LRU size.  The WORKINGSET_ACTIVATE* is needed to check if
there is a refault since last snapshot and the LRU size are needed for the
desirable ratio between inactive and active LRUs.  See the table below on
how the desirable ratio is calculated.

/* total     target    max
 * memory    ratio     inactive
 * -------------------------------------
 *   10MB       1         5MB
 *  100MB       1        50MB
 *    1GB       3       250MB
 *   10GB      10       0.9GB
 *  100GB      31         3GB
 *    1TB     101        10GB
 *   10TB     320        32GB
 */

The desirable ratio only changes at the boundary of 1 GiB, 10 GiB, 100
GiB, 1 TiB and 10 TiB.  There is no need for the precise and accurate LRU
size information to calculate this ratio.  In addition, if deactivation is
skipped for some LRU, the kernel will force deactive on the severe memory
pressure situation.

For the cache trim mode, inactive file LRU size is read and the kernel
scales it down based on the reclaim iteration (file >> sc->priority) and
only checks if it is zero or not.  Again precise information is not
needed.

This patch has been running on Meta fleet for several months and we have
not observed any issues.  Please note that MGLRU is not impacted by this
issue at all as it avoids rstat flushing completely.

Link: https://lore.kernel.org/all/6ee2518b-81dd-4082-bdf5-322883895ffc@kernel.org
Link: https://lkml.kernel.org/r/20240813215358.2259750-1-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Jesper Dangaard Brouer <hawk@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: shmem: support large folio swap out

Shmem will support large folio allocation [1] [2] to get a better
performance, however, the memory reclaim still splits the precious large
folios when trying to swap out shmem, which may lead to the memory
fragmentation issue and can not take advantage of the large folio for
shmeme.

Moreover, the swap code already supports for swapping out large folio
without split, hence this patch set supports the large folio swap out for
shmem.

Note the i915_gem_shmem driver still need to be split when swapping, thus
add a new flag 'split_large_folio' for writeback_control to indicate
spliting the large folio.

[1] https://lore.kernel.org/all/cover.1717495894.git.baolin.wang@linux.alibaba.com/
[2] https://lore.kernel.org/all/20240515055719.32577-1-da.gomez@samsung.com/
Link: https://lkml.kernel.org/r/d80c21abd20e1b0f5ca66b330f074060fb2f082d.1723434324.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Chris Li <chrisl@kernel.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Pankaj Raghav <p.raghav@samsung.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: shmem: split large entry if the swapin folio is not large

Now the swap device can only swap-in order 0 folio, even though a large
folio is swapped out.  This requires us to split the large entry
previously saved in the shmem pagecache to support the swap in of small
folios.

Link: https://lkml.kernel.org/r/4a0f12f27c54a62eb4d9ca1265fed3a62531a63e.1723434324.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Chris Li <chrisl@kernel.org>
Cc: Daniel Gomez <da.gomez@samsung.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Pankaj Raghav <p.raghav@samsung.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: shmem: drop folio reference count using 'nr_pages' in shmem_delete_from_page_cache()

To support large folio swapin/swapout for shmem in the following patches,
drop the folio's reference count by the number of pages contained in the
folio when a shmem folio is deleted from shmem pagecache after adding into
swap cache.

Link: https://lkml.kernel.org/r/b371eadb27f42fc51261c51008fbb9a334985b4c.1723434324.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Chris Li <chrisl@kernel.org>
Cc: Daniel Gomez <da.gomez@samsung.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Pankaj Raghav <p.raghav@samsung.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: shmem: support large folio allocation for shmem_replace_folio()

To support large folio swapin for shmem in the following patches, add
large folio allocation for the new replacement folio in
shmem_replace_folio().  Moreover large folios occupy N consecutive entries
in the swap cache instead of using multi-index entries like the page
cache, therefore we should replace each consecutive entries in the swap
cache instead of using the shmem_replace_entry().

As well as updating statistics and folio reference count using the number
of pages in the folio.

Link: https://lkml.kernel.org/r/a41138ecc857ef13e7c5ffa0174321e9e2c9970a.1723434324.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Chris Li <chrisl@kernel.org>
Cc: Daniel Gomez <da.gomez@samsung.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Pankaj Raghav <p.raghav@samsung.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: shmem: use swap_free_nr() to free shmem swap entries

As a preparation for supporting shmem large folio swapout, use
swap_free_nr() to free some continuous swap entries of the shmem large
folio when the large folio was swapped in from the swap cache.  In
addition, the index should also be round down to the number of pages when
adding the swapin folio into the pagecache.

Link: https://lkml.kernel.org/r/342207fa679fc88a447dac2e101ad79e6050fe79.1723434324.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Chris Li <chrisl@kernel.org>
Cc: Daniel Gomez <da.gomez@samsung.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Pankaj Raghav <p.raghav@samsung.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: filemap: use xa_get_order() to get the swap entry order

In the following patches, shmem will support the swap out of large folios,
which means the shmem mappings may contain large order swap entries, so
using xa_get_order() to get the folio order of the shmem swap entry to
update the '*start' correctly.

Link: https://lkml.kernel.org/r/6876d55145c1cc80e79df7884aa3a62e397b101d.1723434324.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Chris Li <chrisl@kernel.org>
Cc: Daniel Gomez <da.gomez@samsung.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Pankaj Raghav <p.raghav@samsung.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: shmem: return number of pages beeing freed in shmem_free_swap

Both shmem_free_swap callers expect the number of pages being freed.  In
the large folios context, this needs to support larger values other than 0
(used as 1 page being freed) and -ENOENT (used as 0 pages being freed).
In preparation for large folios adoption, make shmem_free_swap routine
return the number of pages being freed.  So, returning 0 in this context,
means 0 pages being freed.

While we are at it, changing to use free_swap_and_cache_nr() to free large
order swap entry by Baolin Wang.

Link: https://lkml.kernel.org/r/9623e863c83d749d5ab407f6fdf0a8e5a3bdf052.1723434324.git.baolin.wang@linux.alibaba.com
Signed-off-by: Daniel Gomez <da.gomez@samsung.com>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Cc: Barry Song <baohua@kernel.org>
Cc: Chris Li <chrisl@kernel.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Pankaj Raghav <p.raghav@samsung.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: shmem: extend shmem_partial_swap_usage() to support large folio swap

To support shmem large folio swapout in the following patches, using
xa_get_order() to get the order of the swap entry to calculate the swap
usage of shmem.

Link: https://lkml.kernel.org/r/60b130b9fc3e422bb91293a172c2113c85e9233a.1723434324.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Chris Li <chrisl@kernel.org>
Cc: Daniel Gomez <da.gomez@samsung.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Pankaj Raghav <p.raghav@samsung.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: swap: extend swap_shmem_alloc() to support batch SWAP_MAP_SHMEM flag setting

Patch series "support large folio swap-out and swap-in for shmem", v5.

Shmem will support large folio allocation [1] [2] to get a better
performance, however, the memory reclaim still splits the precious large
folios when trying to swap-out shmem, which may lead to the memory
fragmentation issue and can not take advantage of the large folio for
shmeme.

Moreover, the swap code already supports for swapping out large folio
without split, and large folio swap-in[3] series is queued into
mm-unstable branch.  Hence this patch set also supports the large folio
swap-out and swap-in for shmem.

This patch (of 9):

To support shmem large folio swap operations, add a new parameter to
swap_shmem_alloc() that allows batch SWAP_MAP_SHMEM flag setting for shmem
swap entries.

While we are at it, using folio_nr_pages() to get the number of pages of
the folio as a preparation.

Link: https://lkml.kernel.org/r/cover.1723434324.git.baolin.wang@linux.alibaba.com
Link: https://lkml.kernel.org/r/99f64115d04b285e009580eb177352c57119ffd0.1723434324.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: Barry Song <baohua@kernel.org>
Cc: Chris Li <chrisl@kernel.org>
Cc: Daniel Gomez <da.gomez@samsung.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Pankaj Raghav <p.raghav@samsung.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

powerpc/vdso: refactor error handling

Linus noticed that the error handling in __arch_setup_additional_pages()
fails to clear the mm VDSO pointer if _install_special_mapping() fails.
In practice there should be no actual bug, because if there's an error the
VDSO pointer is cleared later in arch_setup_additional_pages().

However it's no longer necessary to set the pointer before installing the
mapping.  Commit c1bab64360e6 ("powerpc/vdso: Move to
_install_special_mapping() and remove arch_vma_name()") reworked the code
so that the VMA name comes from the vm_special_mapping.name, rather than
relying on arch_vma_name().

So rework the code to only set the VDSO pointer once the mappings have
been installed correctly, and remove the stale comment.

Link: https://lkml.kernel.org/r/20240812082605.743814-4-mpe@ellerman.id.au
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Jeff Xu <jeffxu@google.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Pedro Falcato <pedro.falcato@gmail.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: remove arch_unmap()

Now that powerpc no longer uses arch_unmap() to handle VDSO unmapping,
there are no meaningful implementions left.  Drop support for it entirely,
and update comments which refer to it.

Link: https://lkml.kernel.org/r/20240812082605.743814-3-mpe@ellerman.id.au
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Acked-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Jeff Xu <jeffxu@google.com>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Pedro Falcato <pedro.falcato@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

powerpc/mm: handle VDSO unmapping via close() rather than arch_unmap()

Add a close() callback to the VDSO special mapping to handle unmapping of
the VDSO.  That will make it possible to remove the arch_unmap() hook
entirely in a subsequent patch.

Link: https://lkml.kernel.org/r/20240812082605.743814-2-mpe@ellerman.id.au
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Jeff Xu <jeffxu@google.com>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Pedro Falcato <pedro.falcato@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm: add optional close() to struct vm_special_mapping

Add an optional close() callback to struct vm_special_mapping.  It will be
used, by powerpc at least, to handle unmapping of the VDSO.

Although support for unmapping the VDSO was initially added for CRIU[1],
it is not desirable to guard that support behind
CONFIG_CHECKPOINT_RESTORE.

There are other known users of unmapping the VDSO which are not related to
CRIU, eg.  Valgrind [2] and void-ship [3].

The powerpc arch_unmap() hook has been in place for ~9 years, with no
ifdef, so there may be other unknown users that have come to rely on
unmapping the VDSO.  Even if the code was behind an ifdef, major distros
enable CHECKPOINT_RESTORE so users may not realise unmapping the VDSO
depends on that configuration option.

It's also undesirable to have such core mm behaviour behind a relatively
obscure CONFIG option.

Longer term the unmap behaviour should be standardised across
architectures, however that is complicated by the fact the VDSO pointer is
stored differently across architectures.  There was a previous attempt to
unify that handling [4], which could be revived.

See [5] for further discussion.

[1]: commit 83d3f0e90c6c ("powerpc/mm: tracking vDSO remap")
[2]: https://sourceware.org/git/?p=valgrind.git;a=commit;h=3a004915a2cbdcdebafc1612427576bf3321eef5
[3]: https://github.com/insanitybit/void-ship
[4]: https://lore.kernel.org/lkml/20210611180242.711399-17-dima@arista.com/
[5]: https://lore.kernel.org/linuxppc-dev/shiq5v3jrmyi6ncwke7wgl76ojysgbhrchsk32q4lbx2hadqqc@kzyy2igem256

Link: https://lkml.kernel.org/r/20240812082605.743814-1-mpe@ellerman.id.au
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Jeff Xu <jeffxu@google.com>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Pedro Falcato <pedro.falcato@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

kfence: save freeing stack trace at calling time instead of freeing time

For kmem_cache with SLAB_TYPESAFE_BY_RCU, the freeing trace stack at
calling kmem_cache_free() is more useful. While the following stack is
meaningless and provides no help:
  freed by task 46 on cpu 0 at 656.840729s:
   rcu_do_batch+0x1ab/0x540
   nocb_cb_wait+0x8f/0x260
   rcu_nocb_cb_kthread+0x25/0x80
   kthread+0xd2/0x100
   ret_from_fork+0x34/0x50
   ret_from_fork_asm+0x1a/0x30

Link: https://lkml.kernel.org/r/20240812095517.2357-1-dtcccc@linux.alibaba.com
Signed-off-by: Tianchen Ding <dtcccc@linux.alibaba.com>
Reviewed-by: Marco Elver <elver@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: fix comment typo with corresponding maple_status

In comment of function mas_start(), we list the return value of different
cases.  According to the comment context, tell the maple_status here is
more consistent with others.

Let's correct it with ma_active in the case it's a tree.

Link: https://lkml.kernel.org/r/20240812150925.31551-2-richard.weiyang@gmail.com
Signed-off-by: Wei Yang <richard.weiyang@gmail.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: fix comment typo of ma_root

In comment of mas_start(), we lists the return value for different cases.
In case of a single entry, we set mas->status to ma_root, while the
comment uses mas_root, which is not a maple_status.

Fix the typo according to the code.

Link: https://lkml.kernel.org/r/20240812150925.31551-1-richard.weiyang@gmail.com
Signed-off-by: Wei Yang <richard.weiyang@gmail.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: add test to replicate low memory race conditions

Add new callback fields to the userspace implementation of struct
kmem_cache.  This allows for executing callback functions in order to
further test low memory scenarios where node allocation is retried.

This callback can help test race conditions by calling a function when a
low memory event is tested.

Link: https://lkml.kernel.org/r/20240812190543.71967-2-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

maple_tree: reset mas->index and mas->last on write retries

The following scenario can result in a race condition:

Consider a node with the following indices and values

a<------->b<----------->c<--------->d
    0xA        NULL          0xB

CPU 1   CPU 2
      ---------         ---------
mas_set_range(a,b)
mas_erase()
-> range is expanded (a,c) because of null expansion

mas_nomem()
mas_unlock()
mas_store_range(b,c,0xC)

The node now looks like:

a<------->b<----------->c<--------->d
    0xA        0xC          0xB

mas_lock()
mas_erase() <------ range of erase is still (a,c)

The node is now NULL from (a,c) but the write from CPU 2 should have been
retained and range (b,c) should still have 0xC as its value.  We can fix
this by re-intializing to the original index and last.  This does not need
a cc: Stable as there are no users of the maple tree which use internal
locking and this condition is only possible with internal locking.

Link: https://lkml.kernel.org/r/20240812190543.71967-1-sidhartha.kumar@oracle.com
Signed-off-by: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm/hugetlb_vmemmap: batch HVO work when demoting

Batch the HVO work, including de-HVO of the source and HVO of the
destination hugeTLB folios, to speed up demotion.

After commit bd225530a4c7 ("mm/hugetlb_vmemmap: fix race with speculative
PFN walkers"), each request of HVO or de-HVO, batched or not, invokes
synchronize_rcu() once.  For example, when not batched, demoting one 1GB
hugeTLB folio to 512 2MB hugeTLB folios invokes synchronize_rcu() 513
times (1 de-HVO plus 512 HVO requests), whereas when batched, only twice
(1 de-HVO plus 1 HVO request).  And the performance difference between the
two cases is significant, e.g.,

  echo 2048kB >/sys/kernel/mm/hugepages/hugepages-1048576kB/demote_size
  time echo 100 >/sys/kernel/mm/hugepages/hugepages-1048576kB/demote

Before this patch:
  real     8m58.158s
  user     0m0.009s
  sys      0m5.900s

After this patch:
  real     0m0.900s
  user     0m0.000s
  sys      0m0.851s

Note that this patch changes the behavior of the `demote` interface when
de-HVO fails.  Before, the interface aborts immediately upon failure; now,
it tries to finish an entire batch, meaning it can make extra progress if
the rest of the batch contains folios that do not need to de-HVO.

Link: https://lkml.kernel.org/r/20240812224823.3914837-1-yuzhao@google.com
Fixes: bd225530a4c7 ("mm/hugetlb_vmemmap: fix race with speculative PFN walkers")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Reviewed-by: Muchun Song <muchun.song@linux.dev>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>