Let's track the mapcount of large folios in a single value. The mapcount
of a large folio currently corresponds to the sum of the entire mapcount
and all page mapcounts.
This sum is what we actually want to know in folio_mapcount() and it is
also sufficient for implementing folio_mapped().
With PTE-mapped THP becoming more important and more widely used, we want
to avoid looping over all pages of a folio just to obtain the mapcount of
large folios. The comment "In the common case, avoid the loop when no
pages mapped by PTE" in folio_total_mapcount() does no longer hold for
mTHP that are always mapped by PTE.
Further, we are planning on using folio_mapcount() more frequently, and
might even want to remove page mapcounts for large folios in some kernel
configs. Therefore, allow for reading the mapcount of large folios
efficiently and atomically without looping over any pages.
Maintain the mapcount also for hugetlb pages for simplicity. Use the new
mapcount to implement folio_mapcount() and folio_mapped(). Make
page_mapped() simply call folio_mapped(). We can now get rid of
folio_large_is_mapped().
_nr_pages_mapped is now only used in rmap code and for debugging purposes.
Keep folio_nr_pages_mapped() around, but document that its use should be
limited to rmap internals and debugging purposes.
This change implies one additional atomic add/sub whenever
mapping/unmapping (parts of) a large folio.
As we now batch RMAP operations for PTE-mapped THP during fork(), during
unmap/zap, and when PTE-remapping a PMD-mapped THP, and we adjust the
large mapcount for a PTE batch only once, the added overhead in the common
case is small. Only when unmapping individual pages of a large folio
(e.g., during COW), the overhead might be bigger in comparison, but it's
essentially one additional atomic operation.
Note that before the new mapcount would overflow, already our refcount
would overflow: each mapping requires a folio reference. Extend the
focumentation of folio_mapcount().
Link: https://lkml.kernel.org/r/20240409192301.907377-5-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Yin Fengwei <fengwei.yin@intel.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: John Paul Adrian Glaubitz <glaubitz@physik.fu-berlin.de>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Richard Chang <richardycc@google.com>
Cc: Rich Felker <dalias@libc.org>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yoshinori Sato <ysato@users.sourceforge.jp>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm: mapcount for large folios + page_mapcount() cleanups".
This series tracks the mapcount of large folios in a single value, so it
can be read efficiently and atomically, just like the mapcount of small
folios.
folio_mapcount() is then used in a couple more places, most notably to
reduce false negatives in folio_likely_mapped_shared(), and many users of
page_mapcount() are cleaned up (that's maybe why you got CCed on the full
series, sorry sh+xtensa folks! :) ).
The remaining s390x user and one KSM user of page_mapcount() are getting
removed separately on the list right now. I have patches to handle the
other KSM one, the khugepaged one and the kpagecount one; as they are not
as "obvious", I will send them out separately in the future. Once that is
all in place, I'm planning on moving page_mapcount() into
fs/proc/task_mmu.c, the remaining user for the time being (and we can
discuss at LSF/MM details on that :) ).
I proposed the mapcount for large folios (previously called total
mapcount) originally in part of [1] and I later included it in [2] where
it is a requirement. In the meantime, I changed the patch a bit so I
dropped all RB's. During the discussion of [1], Peter Xu correctly raised
that this additional tracking might affect the performance when PMD->PTE
remapping THPs. In the meantime. I addressed that by batching RMAP
operations during fork(), unmap/zap and when PMD->PTE remapping THPs.
Running some of my micro-benchmarks [3] (fork,munmap,cow-byte,remap) on 1
GiB of memory backed by folios with the same order, I observe the
following on an Intel(R) Xeon(R) Silver 4210R CPU @ 2.40GHz tuned for
reproducible results as much as possible:
Standard deviation is mostly < 1%, except for order-9, where it's < 2% for
fork() and munmap().
(1) Small folios are not affected (< 1%) in all 4 microbenchmarks.
(2) Order-4 folios are not affected (< 1%) in all 4 microbenchmarks. A bit
weird comapred to the other orders ...
(3) PMD->PTE remapping of order-9 THPs is not affected (< 1%)
(4) COW-byte (COWing a single page by writing a single byte) is not
affected for any order (< 1 %). The page copy_fault overhead dominates
everything.
(5) fork() is mostly not affected (< 1%), except order-2, where we have
a slowdown of ~4%. Already for order-3 folios, we're down to a slowdown
of < 1%.
(6) munmap() sees a slowdown by < 3% for some orders (order-5,
order-6, order-9), but less for others (< 1% for order-4 and order-8,
< 2% for order-2, order-3, order-7).
Especially the fork() and munmap() benchmark are sensitive to each added
instruction and other system noise, so I suspect some of the change and
observed weirdness (order-4) is due to code layout changes and other
factors, but not really due to the added atomics.
So in the common case where we can batch, the added atomics don't really
make a big difference, especially in light of the recent improvements for
large folios that we recently gained due to batching. Surprisingly, for
some cases where we cannot batch (e.g., COW), the added atomics don't seem
to matter, because other overhead dominates.
My fork and munmap micro-benchmarks don't cover cases where we cannot
batch-process bigger parts of large folios. As this is not the common
case, I'm not worrying about that right now.
Future work is batching RMAP operations during swapout and folio
migration.
[1] https://lore.kernel.org/all/20230809083256.699513-1-david@redhat.com/
[2] https://lore.kernel.org/all/20231124132626.235350-1-david@redhat.com/
[3] https://gitlab.com/davidhildenbrand/scratchspace/-/raw/main/pte-mapped-folio-benchmarks.c?ref_type=heads
This patch (of 18):
Commit 53277bcf126d ("mm: support page_mapcount() on page_has_type()
pages") made it impossible to detect mapcount underflows by treating any
negative raw mapcount value as a mapcount of 0.
We perform such underflow checks in zap_present_folio_ptes() and
zap_huge_pmd(), which would currently no longer trigger.
Let's check against PAGE_MAPCOUNT_RESERVE instead by using
page_type_has_type(), like page_has_type() would, so we can still catch
some underflows.
[david@redhat.com: make page_mapcount() slighly more efficient]
Link: https://lkml.kernel.org/r/1af4fd61-7926-47c8-be45-833c0dbec08b@redhat.com
Link: https://lkml.kernel.org/r/20240409192301.907377-1-david@redhat.com
Link: https://lkml.kernel.org/r/20240409192301.907377-2-david@redhat.com
Fixes: 53277bcf126d ("mm: support page_mapcount() on page_has_type() pages")
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: John Paul Adrian Glaubitz <glaubitz@physik.fu-berlin.de>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Richard Chang <richardycc@google.com>
Cc: Rich Felker <dalias@libc.org>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yin Fengwei <fengwei.yin@intel.com>
Cc: Yoshinori Sato <ysato@users.sourceforge.jp>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
follow_pte() is now our main function to lookup PTEs in VM_PFNMAP/VM_IO
VMAs. Let's perform some more sanity checks to make this exported
function harder to abuse.
Further, extend the doc a bit, it still focuses on the KVM use case with
MMU notifiers. Drop the KVM+follow_pfn() comment, follow_pfn() is no
more, and we have other users nowadays.
Also extend the doc regarding refcounted pages and the interaction with
MMU notifiers.
KVM is one example that uses MMU notifiers and can deal with refcounted
pages properly. VFIO is one example that doesn't use MMU notifiers, and
to prevent use-after-free, rejects refcounted pages: pfn_valid(pfn) &&
!PageReserved(pfn_to_page(pfn)). Protection changes are less of a concern
for users like VFIO: the behavior is similar to longterm-pinning a page,
and getting the PTE protection changed afterwards.
The primary concern with refcounted pages is use-after-free, which callers
should be aware of.
Link: https://lkml.kernel.org/r/20240410155527.474777-4-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Alex Williamson <alex.williamson@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Fei Li <fei1.li@intel.com>
Cc: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Sean Christopherson <seanjc@google.com>
Cc: Yonghua Huang <yonghua.huang@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm: follow_pte() improvements and acrn follow_pte() fixes".
Patch #1 fixes a bunch of issues I spotted in the acrn driver. It
compiles, that's all I know. I'll appreciate some review and testing from
acrn folks.
Patch #2+#3 improve follow_pte(), passing a VMA instead of the MM, adding
more sanity checks, and improving the documentation. Gave it a quick test
on x86-64 using VM_PAT that ends up using follow_pte().
This patch (of 3):
We currently miss handling various cases, resulting in a dangerous
follow_pte() (previously follow_pfn()) usage.
(1) We're not checking PTE write permissions.
Maybe we should simply always require pte_write() like we do for
pin_user_pages_fast(FOLL_WRITE)? Hard to tell, so let's check for
ACRN_MEM_ACCESS_WRITE for now.
(2) We're not rejecting refcounted pages.
As we are not using MMU notifiers, messing with refcounted pages is
dangerous and can result in use-after-free. Let's make sure to reject them.
(3) We are only looking at the first PTE of a bigger range.
We only lookup a single PTE, but memmap->len may span a larger area.
Let's loop over all involved PTEs and make sure the PFN range is
actually contiguous. Reject everything else: it couldn't have worked
either way, and rather made use access PFNs we shouldn't be accessing.
Link: https://lkml.kernel.org/r/20240410155527.474777-1-david@redhat.com
Link: https://lkml.kernel.org/r/20240410155527.474777-2-david@redhat.com
Fixes: 8a6e85f75a ("virt: acrn: obtain pa from VMA with PFNMAP flag")
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Alex Williamson <alex.williamson@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Fei Li <fei1.li@intel.com>
Cc: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Yonghua Huang <yonghua.huang@intel.com>
Cc: Sean Christopherson <seanjc@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The current implementation treats emulated memory devices, such as CXL1.1
type3 memory, as normal DRAM when they are emulated as normal memory
(E820_TYPE_RAM). However, these emulated devices have different
characteristics than traditional DRAM, making it important to distinguish
them. Thus, we modify the tiered memory initialization process to
introduce a delay specifically for CPUless NUMA nodes. This delay ensures
that the memory tier initialization for these nodes is deferred until HMAT
information is obtained during the boot process. Finally, demotion tables
are recalculated at the end.
* late_initcall(memory_tier_late_init);
Some device drivers may have initialized memory tiers between
`memory_tier_init()` and `memory_tier_late_init()`, potentially bringing
online memory nodes and configuring memory tiers. They should be
excluded in the late init.
* Handle cases where there is no HMAT when creating memory tiers
There is a scenario where a CPUless node does not provide HMAT
information. If no HMAT is specified, it falls back to using the
default DRAM tier.
* Introduce another new lock `default_dram_perf_lock` for adist
calculation In the current implementation, iterating through CPUlist
nodes requires holding the `memory_tier_lock`. However,
`mt_calc_adistance()` will end up trying to acquire the same lock,
leading to a potential deadlock. Therefore, we propose introducing a
standalone `default_dram_perf_lock` to protect `default_dram_perf_*`.
This approach not only avoids deadlock but also prevents holding a large
lock simultaneously.
* Upgrade `set_node_memory_tier` to support additional cases, including
default DRAM, late CPUless, and hot-plugged initializations. To cover
hot-plugged memory nodes, `mt_calc_adistance()` and
`mt_find_alloc_memory_type()` are moved into `set_node_memory_tier()` to
handle cases where memtype is not initialized and where HMAT information
is available.
* Introduce `default_memory_types` for those memory types that are not
initialized by device drivers. Because late initialized memory and
default DRAM memory need to be managed, a default memory type is created
for storing all memory types that are not initialized by device drivers
and as a fallback.
Link: https://lkml.kernel.org/r/20240405000707.2670063-3-horenchuang@bytedance.com
Signed-off-by: Ho-Ren (Jack) Chuang <horenchuang@bytedance.com>
Signed-off-by: Hao Xiang <hao.xiang@bytedance.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Cc: SeongJae Park <sj@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vishal Verma <vishal.l.verma@intel.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawie.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "Improved Memory Tier Creation for CPUless NUMA Nodes", v11.
When a memory device, such as CXL1.1 type3 memory, is emulated as normal
memory (E820_TYPE_RAM), the memory device is indistinguishable from normal
DRAM in terms of memory tiering with the current implementation. The
current memory tiering assigns all detected normal memory nodes to the
same DRAM tier. This results in normal memory devices with different
attributions being unable to be assigned to the correct memory tier,
leading to the inability to migrate pages between different types of
memory.
https://lore.kernel.org/linux-mm/PH0PR08MB7955E9F08CCB64F23963B5C3A860A@PH0PR08MB7955.namprd08.prod.outlook.com/T/
This patchset automatically resolves the issues. It delays the
initialization of memory tiers for CPUless NUMA nodes until they obtain
HMAT information and after all devices are initialized at boot time,
eliminating the need for user intervention. If no HMAT is specified, it
falls back to using `default_dram_type`.
Example usecase:
We have CXL memory on the host, and we create VMs with a new system memory
device backed by host CXL memory. We inject CXL memory performance
attributes through QEMU, and the guest now sees memory nodes with
performance attributes in HMAT. With this change, we enable the guest
kernel to construct the correct memory tiering for the memory nodes.
This patch (of 2):
Since different memory devices require finding, allocating, and putting
memory types, these common steps are abstracted in this patch, enhancing
the scalability and conciseness of the code.
Link: https://lkml.kernel.org/r/20240405000707.2670063-1-horenchuang@bytedance.com
Link: https://lkml.kernel.org/r/20240405000707.2670063-2-horenchuang@bytedance.com
Signed-off-by: Ho-Ren (Jack) Chuang <horenchuang@bytedance.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawie.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Gregory Price <gourry.memverge@gmail.com>
Cc: Hao Xiang <hao.xiang@bytedance.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Ravi Jonnalagadda <ravis.opensrc@micron.com>
Cc: SeongJae Park <sj@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vishal Verma <vishal.l.verma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The hugetlb_cma code passes 0 in the order_per_bit argument to
cma_declare_contiguous_nid (the alignment, computed using the page order,
is correctly passed in).
This causes a bit in the cma allocation bitmap to always represent a 4k
page, making the bitmaps potentially very large, and slower.
It would create bitmaps that would be pretty big. E.g. for a 4k page
size on x86, hugetlb_cma=64G would mean a bitmap size of (64G / 4k) / 8
== 2M. With HUGETLB_PAGE_ORDER as order_per_bit, as intended, this
would be (64G / 2M) / 8 == 4k. So, that's quite a difference.
Also, this restricted the hugetlb_cma area to ((PAGE_SIZE <<
MAX_PAGE_ORDER) * 8) * PAGE_SIZE (e.g. 128G on x86) , since
bitmap_alloc uses normal page allocation, and is thus restricted by
MAX_PAGE_ORDER. Specifying anything about that would fail the CMA
initialization.
So, correctly pass in the order instead.
Link: https://lkml.kernel.org/r/20240404162515.527802-2-fvdl@google.com
Fixes: cf11e85fc0 ("mm: hugetlb: optionally allocate gigantic hugepages using cma")
Signed-off-by: Frank van der Linden <fvdl@google.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm/gup: consistently call it GUP-fast".
Some cleanups around function names, comments and the config option of
"GUP-fast" -- GUP without "lock" safety belts on.
With this cleanup it's easy to judge which functions are GUP-fast
specific. We now consistently call it "GUP-fast", avoiding mixing it with
"fast GUP", "lockless", or simply "gup" (which I always considered
confusing in the ode).
So the magic now happens in functions that contain "gup_fast", whereby
gup_fast() is the entry point into that magic. Comments consistently
reference either "GUP-fast" or "gup_fast()".
This patch (of 3):
Let's consistently call the "fast-only" part of GUP "GUP-fast" and rename
all relevant internal functions to start with "gup_fast", to make it
clearer that this is not ordinary GUP. The current mixture of "lockless",
"gup" and "gup_fast" is confusing.
Further, avoid the term "huge" when talking about a "leaf" -- for example,
we nowadays check pmd_leaf() because pmd_huge() is gone. For the
"hugepd"/"hugepte" stuff, it's part of the name ("is_hugepd"), so that
stays.
What remains is the "external" interface:
* get_user_pages_fast_only()
* get_user_pages_fast()
* pin_user_pages_fast()
The high-level internal functions for GUP-fast (+slow fallback) are now:
* internal_get_user_pages_fast() -> gup_fast_fallback()
* lockless_pages_from_mm() -> gup_fast()
The basic GUP-fast walker functions:
* gup_pgd_range() -> gup_fast_pgd_range()
* gup_p4d_range() -> gup_fast_p4d_range()
* gup_pud_range() -> gup_fast_pud_range()
* gup_pmd_range() -> gup_fast_pmd_range()
* gup_pte_range() -> gup_fast_pte_range()
* gup_huge_pgd() -> gup_fast_pgd_leaf()
* gup_huge_pud() -> gup_fast_pud_leaf()
* gup_huge_pmd() -> gup_fast_pmd_leaf()
The weird hugepd stuff:
* gup_huge_pd() -> gup_fast_hugepd()
* gup_hugepte() -> gup_fast_hugepte()
The weird devmap stuff:
* __gup_device_huge_pud() -> gup_fast_devmap_pud_leaf()
* __gup_device_huge_pmd -> gup_fast_devmap_pmd_leaf()
* __gup_device_huge() -> gup_fast_devmap_leaf()
* undo_dev_pagemap() -> gup_fast_undo_dev_pagemap()
Helper functions:
* unpin_user_pages_lockless() -> gup_fast_unpin_user_pages()
* gup_fast_folio_allowed() is already properly named
* gup_fast_permitted() is already properly named
With "gup_fast()", we now even have a function that is referred to in
comment in mm/mmu_gather.c.
Link: https://lkml.kernel.org/r/20240402125516.223131-1-david@redhat.com
Link: https://lkml.kernel.org/r/20240402125516.223131-2-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Reviewed-by: Mike Rapoport (IBM) <rppt@kernel.org>
Reviewed-by: John Hubbard <jhubbard@nvidia.com>
Cc: Peter Xu <peterx@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Rework madvise_cold_or_pageout_pte_range() to avoid splitting any large
folio that is fully and contiguously mapped in the pageout/cold vm range.
This change means that large folios will be maintained all the way to swap
storage. This both improves performance during swap-out, by eliding the
cost of splitting the folio, and sets us up nicely for maintaining the
large folio when it is swapped back in (to be covered in a separate
series).
Folios that are not fully mapped in the target range are still split, but
note that behavior is changed so that if the split fails for any reason
(folio locked, shared, etc) we now leave it as is and move to the next pte
in the range and continue work on the proceeding folios. Previously any
failure of this sort would cause the entire operation to give up and no
folios mapped at higher addresses were paged out or made cold. Given
large folios are becoming more common, this old behavior would have likely
lead to wasted opportunities.
While we are at it, change the code that clears young from the ptes to use
ptep_test_and_clear_young(), via the new mkold_ptes() batch helper
function. This is more efficent than get_and_clear/modify/set, especially
for contpte mappings on arm64, where the old approach would require
unfolding/refolding and the new approach can be done in place.
Link: https://lkml.kernel.org/r/20240408183946.2991168-8-ryan.roberts@arm.com
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Reviewed-by: Barry Song <v-songbaohua@oppo.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Barry Song <21cnbao@gmail.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Gao Xiang <xiang@kernel.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Multi-size THP enables performance improvements by allocating large,
pte-mapped folios for anonymous memory. However I've observed that on an
arm64 system running a parallel workload (e.g. kernel compilation) across
many cores, under high memory pressure, the speed regresses. This is due
to bottlenecking on the increased number of TLBIs added due to all the
extra folio splitting when the large folios are swapped out.
Therefore, solve this regression by adding support for swapping out mTHP
without needing to split the folio, just like is already done for
PMD-sized THP. This change only applies when CONFIG_THP_SWAP is enabled,
and when the swap backing store is a non-rotating block device. These are
the same constraints as for the existing PMD-sized THP swap-out support.
Note that no attempt is made to swap-in (m)THP here - this is still done
page-by-page, like for PMD-sized THP. But swapping-out mTHP is a
prerequisite for swapping-in mTHP.
The main change here is to improve the swap entry allocator so that it can
allocate any power-of-2 number of contiguous entries between [1, (1 <<
PMD_ORDER)]. This is done by allocating a cluster for each distinct order
and allocating sequentially from it until the cluster is full. This
ensures that we don't need to search the map and we get no fragmentation
due to alignment padding for different orders in the cluster. If there is
no current cluster for a given order, we attempt to allocate a free
cluster from the list. If there are no free clusters, we fail the
allocation and the caller can fall back to splitting the folio and
allocates individual entries (as per existing PMD-sized THP fallback).
The per-order current clusters are maintained per-cpu using the existing
infrastructure. This is done to avoid interleving pages from different
tasks, which would prevent IO being batched. This is already done for the
order-0 allocations so we follow the same pattern.
As is done for order-0 per-cpu clusters, the scanner now can steal order-0
entries from any per-cpu-per-order reserved cluster. This ensures that
when the swap file is getting full, space doesn't get tied up in the
per-cpu reserves.
This change only modifies swap to be able to accept any order mTHP. It
doesn't change the callers to elide doing the actual split. That will be
done in separate changes.
Link: https://lkml.kernel.org/r/20240408183946.2991168-6-ryan.roberts@arm.com
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Barry Song <21cnbao@gmail.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: Gao Xiang <xiang@kernel.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
struct percpu_cluster stores the index of cpu's current cluster and the
offset of the next entry that will be allocated for the cpu. These two
pieces of information are redundant because the cluster index is just
(offset / SWAPFILE_CLUSTER). The only reason for explicitly keeping the
cluster index is because the structure used for it also has a flag to
indicate "no cluster". However this data structure also contains a spin
lock, which is never used in this context, as a side effect the code
copies the spinlock_t structure, which is questionable coding practice in
my view.
So let's clean this up and store only the next offset, and use a sentinal
value (SWAP_NEXT_INVALID) to indicate "no cluster". SWAP_NEXT_INVALID is
chosen to be 0, because 0 will never be seen legitimately; The first page
in the swap file is the swap header, which is always marked bad to prevent
it from being allocated as an entry. This also prevents the cluster to
which it belongs being marked free, so it will never appear on the free
list.
This change saves 16 bytes per cpu. And given we are shortly going to
extend this mechanism to be per-cpu-AND-per-order, we will end up saving
16 * 9 = 144 bytes per cpu, which adds up if you have 256 cpus in the
system.
Link: https://lkml.kernel.org/r/20240408183946.2991168-4-ryan.roberts@arm.com
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Barry Song <21cnbao@gmail.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: Gao Xiang <xiang@kernel.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Now that we no longer have a convenient flag in the cluster to determine
if a folio is large, free_swap_and_cache() will take a reference and lock
a large folio much more often, which could lead to contention and (e.g.)
failure to split large folios, etc.
Let's solve that problem by batch freeing swap and cache with a new
function, free_swap_and_cache_nr(), to free a contiguous range of swap
entries together. This allows us to first drop a reference to each swap
slot before we try to release the cache folio. This means we only try to
release the folio once, only taking the reference and lock once - much
better than the previous 512 times for the 2M THP case.
Contiguous swap entries are gathered in zap_pte_range() and
madvise_free_pte_range() in a similar way to how present ptes are already
gathered in zap_pte_range().
While we are at it, let's simplify by converting the return type of both
functions to void. The return value was used only by zap_pte_range() to
print a bad pte, and was ignored by everyone else, so the extra reporting
wasn't exactly guaranteed. We will still get the warning with most of the
information from get_swap_device(). With the batch version, we wouldn't
know which pte was bad anyway so could print the wrong one.
[ryan.roberts@arm.com: fix a build warning on parisc]
Link: https://lkml.kernel.org/r/20240409111840.3173122-1-ryan.roberts@arm.com
Link: https://lkml.kernel.org/r/20240408183946.2991168-3-ryan.roberts@arm.com
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Barry Song <21cnbao@gmail.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Gao Xiang <xiang@kernel.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "Swap-out mTHP without splitting", v7.
This series adds support for swapping out multi-size THP (mTHP) without
needing to first split the large folio via
split_huge_page_to_list_to_order(). It closely follows the approach
already used to swap-out PMD-sized THP.
There are a couple of reasons for swapping out mTHP without splitting:
- Performance: It is expensive to split a large folio and under
extreme memory pressure some workloads regressed performance when
using 64K mTHP vs 4K small folios because of this extra cost in the
swap-out path. This series not only eliminates the regression but
makes it faster to swap out 64K mTHP vs 4K small folios.
- Memory fragmentation avoidance: If we can avoid splitting a large
folio memory is less likely to become fragmented, making it easier to
re-allocate a large folio in future.
- Performance: Enables a separate series [7] to swap-in whole mTHPs,
which means we won't lose the TLB-efficiency benefits of mTHP once the
memory has been through a swap cycle.
I've done what I thought was the smallest change possible, and as a
result, this approach is only employed when the swap is backed by a
non-rotating block device (just as PMD-sized THP is supported today).
Discussion against the RFC concluded that this is sufficient.
Performance Testing
===================
I've run some swap performance tests on Ampere Altra VM (arm64) with 8
CPUs. The VM is set up with a 35G block ram device as the swap device and
the test is run from inside a memcg limited to 40G memory. I've then run
`usemem` from vm-scalability with 70 processes, each allocating and
writing 1G of memory. I've repeated everything 6 times and taken the mean
performance improvement relative to 4K page baseline:
| alloc size | baseline | + this series |
| | mm-unstable (~v6.9-rc1) | |
|:-----------|------------------------:|------------------------:|
| 4K Page | 0.0% | 1.3% |
| 64K THP | -13.6% | 46.3% |
| 2M THP | 91.4% | 89.6% |
So with this change, the 64K swap performance goes from a 14% regression to a
46% improvement. While 2M shows a small regression I'm confident that this is
just noise.
[1] https://lore.kernel.org/linux-mm/20231010142111.3997780-1-ryan.roberts@arm.com/
[2] https://lore.kernel.org/linux-mm/20231017161302.2518826-1-ryan.roberts@arm.com/
[3] https://lore.kernel.org/linux-mm/20231025144546.577640-1-ryan.roberts@arm.com/
[4] https://lore.kernel.org/linux-mm/20240311150058.1122862-1-ryan.roberts@arm.com/
[5] https://lore.kernel.org/linux-mm/20240327144537.4165578-1-ryan.roberts@arm.com/
[6] https://lore.kernel.org/linux-mm/20240403114032.1162100-1-ryan.roberts@arm.com/
[7] https://lore.kernel.org/linux-mm/20240304081348.197341-1-21cnbao@gmail.com/
[8] https://lore.kernel.org/linux-mm/CAGsJ_4yMOow27WDvN2q=E4HAtDd2PJ=OQ5Pj9DG+6FLWwNuXUw@mail.gmail.com/
[9] https://lore.kernel.org/linux-mm/579d5127-c763-4001-9625-4563a9316ac3@redhat.com/
This patch (of 7):
As preparation for supporting small-sized THP in the swap-out path,
without first needing to split to order-0, Remove the CLUSTER_FLAG_HUGE,
which, when present, always implies PMD-sized THP, which is the same as
the cluster size.
The only use of the flag was to determine whether a swap entry refers to a
single page or a PMD-sized THP in swap_page_trans_huge_swapped(). Instead
of relying on the flag, we now pass in order, which originates from the
folio's order. This allows the logic to work for folios of any order.
The one snag is that one of the swap_page_trans_huge_swapped() call sites
does not have the folio. But it was only being called there to shortcut a
call __try_to_reclaim_swap() in some cases. __try_to_reclaim_swap() gets
the folio and (via some other functions) calls
swap_page_trans_huge_swapped(). So I've removed the problematic call site
and believe the new logic should be functionally equivalent.
That said, removing the fast path means that we will take a reference and
trylock a large folio much more often, which we would like to avoid. The
next patch will solve this.
Removing CLUSTER_FLAG_HUGE also means we can remove split_swap_cluster()
which used to be called during folio splitting, since
split_swap_cluster()'s only job was to remove the flag.
Link: https://lkml.kernel.org/r/20240408183946.2991168-1-ryan.roberts@arm.com
Link: https://lkml.kernel.org/r/20240408183946.2991168-2-ryan.roberts@arm.com
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Chris Li <chrisl@kernel.org>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Barry Song <21cnbao@gmail.com>
Cc: Gao Xiang <xiang@kernel.org>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
