Allocation of numa_distance uses memblock_phys_alloc_range() to limit
allocation to be below the last mapped page.
But NUMA initializaition runs after the direct map is populated and there
is also code in setup_arch() that adjusts memblock limit to reflect how
much memory is already mapped in the direct map.
Simplify the allocation of numa_distance and use plain memblock_alloc().
arch, mm: pull out allocation of NODE_DATA to generic code
Architectures that support NUMA duplicate the code that allocates
NODE_DATA on the node-local memory with slight variations in reporting of
the addresses where the memory was allocated.
Use x86 version as the basis for the generic alloc_node_data() function
and call this function in architecture specific numa initialization.
Round up node data size to SMP_CACHE_BYTES rather than to PAGE_SIZE like
x86 used to do since the bootmem era when allocation granularity was
PAGE_SIZE anyway.
There are no users of HAVE_ARCH_NODEDATA_EXTENSION left, so
arch_alloc_nodedata() and arch_refresh_nodedata() are not needed anymore.
Replace the call to arch_alloc_nodedata() in free_area_init() with a new
helper alloc_offline_node_data(), remove arch_refresh_nodedata() and
cleanup include/linux/memory_hotplug.h from the associated ifdefery.
arch, mm: move definition of node_data to generic code
Every architecture that supports NUMA defines node_data in the same way:
struct pglist_data *node_data[MAX_NUMNODES];
No reason to keep multiple copies of this definition and its forward
declarations, especially when such forward declaration is the only thing
in include/asm/mmzone.h for many architectures.
Add definition and declaration of node_data to generic code and drop
architecture-specific versions.
MIPS: loongson64: drop HAVE_ARCH_NODEDATA_EXTENSION
Commit f8f9f21c7848 ("MIPS: Fix build error for loongson64 and sgi-ip27")
added HAVE_ARCH_NODEDATA_EXTENSION to loongson64 to silence a compilation
error that happened because loongson64 didn't define array of pg_data_t as
node_data like most other architectures did.
After rename of __node_data to node_data arch_alloc_nodedata() and
HAVE_ARCH_NODEDATA_EXTENSION can be dropped from loongson64.
Since it was the only user of HAVE_ARCH_NODEDATA_EXTENSION config option
also remove this option from arch/mips/Kconfig.
Make definition of node_data match other architectures. This will allow
pulling declaration of node_data to the generic mm code in the following
commit.
Commit f8f9f21c7848 ("MIPS: Fix build error for loongson64 and sgi-ip27")
added HAVE_ARCH_NODEDATA_EXTENSION to sgi-ip27 to silence a compilation
error that happened because sgi-ip27 didn't define array of pg_data_t as
node_data like most other architectures did.
After addition of node_data array that matches other architectures and
after ensuring that offline nodes do not appear on node_possible_map, it
is safe to drop arch_alloc_nodedata() and HAVE_ARCH_NODEDATA_EXTENSION
from sgi-ip27.
Following the discussion about handling of CXL fixed memory windows on
arm64 [1] I decided to bite the bullet and move numa_memblks from x86 to
the generic code so they will be available on arm64/riscv and maybe on
loongarch sometime later.
While it could be possible to use memblock to describe CXL memory windows,
it currently lacks notion of unpopulated memory ranges and numa_memblks
does implement this.
Another reason to make numa_memblks generic is that both arch_numa (arm64
and riscv) and loongarch use trimmed copy of x86 code although there is no
fundamental reason why the same code cannot be used on all these
platforms. Having numa_memblks in mm/ will make it's interaction with
ACPI and FDT more consistent and I believe will reduce maintenance burden.
And with generic numa_memblks it is (almost) straightforward to enable
NUMA emulation on arm64 and riscv.
The first 9 commits in this series are cleanups that are not strictly
related to numa_memblks.
Commits 10-16 slightly reorder code in x86 to allow extracting numa_memblks
and NUMA emulation to the generic code.
Commits 17-19 actually move the code from arch/x86/ to mm/ and commits 20-22
does some aftermath cleanups.
Commit 23 updates of_numa_init() to return error of no NUMA nodes were
found in the device tree.
Commit 24 switches arch_numa to numa_memblks.
Commit 25 enables usage of phys_to_target_node() and
memory_add_physaddr_to_nid() with numa_memblks.
Commit 26 moves the description for numa=fake from x86 to admin-guide.
Kairui Song [Wed, 31 Jul 2024 06:49:21 +0000 (23:49 -0700)]
mm: swap: add a adaptive full cluster cache reclaim
Link all full cluster with one full list, and reclaim from it when the
allocation have ran out of all usable clusters.
There are many reason a folio can end up being in the swap cache while
having no swap count reference. So the best way to search for such slots
is still by iterating the swap clusters.
With the list as an LRU, iterating from the oldest cluster and keep them
rotating is a very doable and clean way to free up potentially not inuse
clusters.
When any allocation failure, try reclaim and rotate only one cluster.
This is adaptive for high order allocations they can tolerate fallback.
So this avoids latency, and give the full cluster list an fair chance to
get reclaimed. It release the usage stress for the fallback order 0
allocation or following up high order allocation.
If the swap device is getting very full, reclaim more aggresively to
ensure no OOM will happen. This ensures order 0 heavy workload won't go
OOM as order 0 won't fail if any cluster still have any space.
Kairui Song [Wed, 31 Jul 2024 06:49:20 +0000 (23:49 -0700)]
mm: swap: relaim the cached parts that got scanned
This commit implements reclaim during scan for cluster allocator.
Cluster scanning were unable to reuse SWAP_HAS_CACHE slots, which could
result in low allocation success rate or early OOM.
So to ensure maximum allocation success rate, integrate reclaiming with
scanning. If found a range of suitable swap slots but fragmented due to
HAS_CACHE, just try to reclaim the slots.
Kairui Song [Wed, 31 Jul 2024 06:49:19 +0000 (23:49 -0700)]
mm: swap: add a fragment cluster list
Now swap cluster allocator arranges the clusters in LRU style, so the
"cold" cluster stay at the head of nonfull lists are the ones that were
used for allocation long time ago and still partially occupied. So if
allocator can't find enough contiguous slots to satisfy an high order
allocation, it's unlikely there will be slot being free on them to satisfy
the allocation, at least in a short period.
As a result, nonfull cluster scanning will waste time repeatly scanning
the unusable head of the list.
Also, multiple CPUs could content on the same head cluster of nonfull
list. Unlike free clusters which are removed from the list when any CPU
starts using it, nonfull cluster stays on the head.
So introduce a new list frag list, all scanned nonfull clusters will be
moved to this list. Both for avoiding repeated scanning and contention.
Frag list is still used as fallback for allocations, so if one CPU failed
to allocate one order of slots, it can still steal other CPU's clusters.
And order 0 will favor the fragmented clusters to better protect nonfull
clusters
If any slots on a fragment list are being freed, move the fragment list
back to nonfull list indicating it worth another scan on the cluster.
Compared to scan upon freeing a slot, this keep the scanning lazy and save
some CPU if there are still other clusters to use.
It may seems unneccessay to keep the fragmented cluster on list at all if
they can't be used for specific order allocation. But this will start to
make sense once reclaim dring scanning is ready.
Kairui Song [Wed, 31 Jul 2024 06:49:18 +0000 (23:49 -0700)]
mm: swap: allow cache reclaim to skip slot cache
Currently we free the reclaimed slots through slot cache even if the slot
is required to be empty immediately. As a result the reclaim caller will
see the slot still occupied even after a successful reclaim, and need to
keep reclaiming until slot cache get flushed. This caused ineffective or
over reclaim when SWAP is under stress.
So introduce a new flag allowing the slot to be emptied bypassing the slot
cache.
Kairui Song [Wed, 31 Jul 2024 06:49:17 +0000 (23:49 -0700)]
mm: swap: skip slot cache on freeing for mTHP
Currently when we are freeing mTHP folios from swap cache, we free then
one by one and put each entry into swap slot cache. Slot cache is
designed to reduce the overhead by batching the freeing, but mTHP swap
entries are already continuous so they can be batch freed without it
already, it saves litle overhead, or even increase overhead for larger
mTHP.
What's more, mTHP entries could stay in swap cache for a while.
Contiguous swap entry is an rather rare resource so releasing them
directly can help improve mTHP allocation success rate when under
pressure.
Chris Li [Wed, 31 Jul 2024 06:49:15 +0000 (23:49 -0700)]
mm: swap: separate SSD allocation from scan_swap_map_slots()
Previously the SSD and HDD share the same swap_map scan loop in
scan_swap_map_slots(). This function is complex and hard to flow the
execution flow.
scan_swap_map_try_ssd_cluster() can already do most of the heavy lifting
to locate the candidate swap range in the cluster. However it needs to go
back to scan_swap_map_slots() to check conflict and then perform the
allocation.
When scan_swap_map_try_ssd_cluster() failed, it still depended on the
scan_swap_map_slots() to do brute force scanning of the swap_map. When
the swapfile is large and almost full, it will take some CPU time to go
through the swap_map array.
Get rid of the cluster allocation dependency on the swap_map scan loop in
scan_swap_map_slots(). Streamline the cluster allocation code path. No
more conflict checks.
For order 0 swap entry, when run out of free and nonfull list. It will
allocate from the higher order nonfull cluster list.
Users should see less CPU time spent on searching the free swap slot when
swapfile is almost full.
Chris Li [Wed, 31 Jul 2024 06:49:14 +0000 (23:49 -0700)]
mm: swap: mTHP allocate swap entries from nonfull list
Track the nonfull cluster as well as the empty cluster on lists. Each
order has one nonfull cluster list.
The cluster will remember which order it was used during new cluster
allocation.
When the cluster has free entry, add to the nonfull[order] list. When
the free cluster list is empty, also allocate from the nonempty list of
that order.
This improves the mTHP swap allocation success rate.
There are limitations if the distribution of numbers of different orders
of mTHP changes a lot. e.g. there are a lot of nonfull cluster assign to
order A while later time there are a lot of order B allocation while very
little allocation in order A. Currently the cluster used by order A will
not reused by order B unless the cluster is 100% empty.
Chris Li [Wed, 31 Jul 2024 06:49:13 +0000 (23:49 -0700)]
mm: swap: swap cluster switch to double link list
Patch series "mm: swap: mTHP swap allocator base on swap cluster order",
v5.
This is the short term solutions "swap cluster order" listed in my "Swap
Abstraction" discussion slice 8 in the recent LSF/MM conference.
When commit 845982eb264bc "mm: swap: allow storage of all mTHP orders" is
introduced, it only allocates the mTHP swap entries from the new empty
cluster list. It has a fragmentation issue reported by Barry.
The reason is that all the empty clusters have been exhausted while there
are plenty of free swap entries in the cluster that are not 100% free.
Remember the swap allocation order in the cluster. Keep track of the per
order non full cluster list for later allocation.
This series gives the swap SSD allocation a new separate code path from
the HDD allocation. The new allocator use cluster list only and do not
global scan swap_map[] without lock any more.
This streamline the swap allocation for SSD. The code matches the
execution flow much better.
User impact: For users that allocate and free mix order mTHP swapping, It
greatly improves the success rate of the mTHP swap allocation after the
initial phase.
It also performs faster when the swapfile is close to full, because the
allocator can get the non full cluster from a list rather than scanning a
lot of swap_map entries.
Previously, the swap cluster used a cluster index as a pointer to
construct a custom single link list type "swap_cluster_list". The next
cluster pointer is shared with the cluster->count. It prevents puting the
non free cluster into a list.
Change the cluster to use the standard double link list instead. This
allows tracing the nonfull cluster in the follow up patch. That way, it
is faster to get to the nonfull cluster of that order.
Remove the cluster getter/setter for accessing the cluster struct member.
The list operation is protected by the swap_info_struct->lock.
Change cluster code to use "struct swap_cluster_info *" to reference the
cluster rather than by using index. That is more consistent with the list
manipulation. It avoids the repeat adding index to the cluser_info. The
code is easier to understand.
Remove the cluster next pointer is NULL flag, the double link list can
handle the empty list pretty well.
The "swap_cluster_info" struct is two pointer bigger, because 512 swap
entries share one swap_cluster_info struct, it has very little impact on
the average memory usage per swap entry. For 1TB swapfile, the swap
cluster data structure increases from 8MB to 24MB.
Other than the list conversion, there is no real function change in this
patch.
Shakeel Butt [Wed, 14 Aug 2024 22:00:21 +0000 (15:00 -0700)]
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.
Shakeel Butt [Wed, 14 Aug 2024 22:00:20 +0000 (15:00 -0700)]
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.
Shakeel Butt [Wed, 14 Aug 2024 22:00:19 +0000 (15:00 -0700)]
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.
Shakeel Butt [Wed, 14 Aug 2024 22:00:18 +0000 (15:00 -0700)]
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.
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.
Shakeel Butt [Thu, 15 Aug 2024 05:04:50 +0000 (22:04 -0700)]
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.
Shakeel Butt [Thu, 15 Aug 2024 05:04:49 +0000 (22:04 -0700)]
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().
Shakeel Butt [Thu, 15 Aug 2024 05:04:47 +0000 (22:04 -0700)]
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.
Ryan Roberts [Wed, 14 Aug 2024 02:02:47 +0000 (14:02 +1200)]
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:
Sidhartha Kumar [Wed, 14 Aug 2024 16:19:43 +0000 (12:19 -0400)]
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.
Sidhartha Kumar [Wed, 14 Aug 2024 16:19:42 +0000 (12:19 -0400)]
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().
Sidhartha Kumar [Wed, 14 Aug 2024 16:19:41 +0000 (12:19 -0400)]
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.
Sidhartha Kumar [Wed, 14 Aug 2024 16:19:40 +0000 (12:19 -0400)]
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.
Sidhartha Kumar [Wed, 14 Aug 2024 16:19:38 +0000 (12:19 -0400)]
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.
Sidhartha Kumar [Wed, 14 Aug 2024 16:19:37 +0000 (12:19 -0400)]
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().
Sidhartha Kumar [Wed, 14 Aug 2024 16:19:36 +0000 (12:19 -0400)]
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.
Sidhartha Kumar [Wed, 14 Aug 2024 16:19:33 +0000 (12:19 -0400)]
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.
Sidhartha Kumar [Wed, 14 Aug 2024 16:19:32 +0000 (12:19 -0400)]
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().
Sidhartha Kumar [Wed, 14 Aug 2024 16:19:31 +0000 (12:19 -0400)]
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.
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.
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.
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.
Muchun Song [Wed, 14 Aug 2024 09:34:15 +0000 (17:34 +0800)]
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.
Shakeel Butt [Tue, 13 Aug 2024 21:53:58 +0000 (14:53 -0700)]
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.
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.
Linus Torvalds [Tue, 20 Aug 2024 22:14:47 +0000 (15:14 -0700)]
mm: remove legacy install_special_mapping() code
All relevant architectures had already been converted to the new interface
(which just has an underscore in front of the name - not very imaginative
naming), this just force-converts the stragglers.
The modern interface is almost identical to the old one, except instead of
the page pointer it takes a "struct vm_special_mapping" that describes the
mapping (and contains the page pointer as one member), and it returns the
resulting 'vma' instead of just the error code.
Getting rid of the old interface also gets rid of some special casing,
which had caused problems with the mremap extensions to "struct
vm_special_mapping".
Michael Ellerman [Mon, 12 Aug 2024 08:26:05 +0000 (18:26 +1000)]
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.
Michael Ellerman [Mon, 12 Aug 2024 08:26:04 +0000 (18:26 +1000)]
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.
Michael Ellerman [Mon, 12 Aug 2024 08:26:03 +0000 (18:26 +1000)]
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.
Michael Ellerman [Mon, 12 Aug 2024 08:26:02 +0000 (18:26 +1000)]
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.
Tianchen Ding [Mon, 12 Aug 2024 09:55:17 +0000 (17:55 +0800)]
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
Wei Yang [Mon, 12 Aug 2024 15:09:25 +0000 (15:09 +0000)]
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.
Wei Yang [Mon, 12 Aug 2024 15:09:24 +0000 (15:09 +0000)]
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.
Sidhartha Kumar [Mon, 12 Aug 2024 19:05:43 +0000 (15:05 -0400)]
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.
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.
Yu Zhao [Mon, 12 Aug 2024 22:48:23 +0000 (16:48 -0600)]
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.
yangge [Tue, 13 Aug 2024 09:52:23 +0000 (17:52 +0800)]
mm/swap: take folio refcount after testing the LRU flag
Whoever passes a folio to __folio_batch_add_and_move() must hold a
reference, otherwise something else would already be messed up. If the
folio is referenced, it will not be freed elsewhere, so we can safely
clear the folio's lru flag. As discussed with David in [1], we should
take the reference after testing the LRU flag, not before.
Takaya Saeki [Tue, 13 Aug 2024 10:03:12 +0000 (10:03 +0000)]
filemap: add trace events for get_pages, map_pages, and fault
To allow precise tracking of page caches accessed, add new tracepoints
that trigger when a process actually accesses them.
The ureadahead program used by ChromeOS traces the disk access of programs
as they start up at boot up. It uses mincore(2) or the
'mm_filemap_add_to_page_cache' trace event to accomplish this. It stores
this information in a "pack" file and on subsequent boots, it will read
the pack file and call readahead(2) on the information so that disk
storage can be loaded into RAM before the applications actually need it.
A problem we see is that due to the kernel's readahead algorithm that can
aggressively pull in more data than needed (to try and accomplish the same
goal) and this data is also recorded. The end result is that the pack
file contains a lot of pages on disk that are never actually used.
Calling readahead(2) on these unused pages can slow down the system boot
up times.
To solve this, add 3 new trace events, get_pages, map_pages, and fault.
These will be used to trace the pages are not only pulled in from disk,
but are actually used by the application. Only those pages will be stored
in the pack file, and this helps out the performance of boot up.
With the combination of these 3 new trace events and
mm_filemap_add_to_page_cache, we observed a reduction in the pack file by
7.3% - 20% on ChromeOS varying by device.
Peter Xu [Mon, 12 Aug 2024 18:12:25 +0000 (14:12 -0400)]
mm/mprotect: fix dax pud handlings
This is only relevant to the two archs that support PUD dax, aka, x86_64
and ppc64. PUD THPs do not yet exist elsewhere, and hugetlb PUDs do not
count in this case.
DAX have had PUD mappings for years, but change protection path never
worked. When the path is triggered in any form (a simple test program
would be: call mprotect() on a 1G dev_dax mapping), the kernel will report
"bad pud". This patch should fix that.
The new change_huge_pud() tries to keep everything simple. For example,
it doesn't optimize write bit as that will need even more PUD helpers.
It's not too bad anyway to have one more write fault in the worst case
once for 1G range; may be a bigger thing for each PAGE_SIZE, though.
Neither does it support userfault-wp bits, as there isn't such PUD
mappings that is supported; file mappings always need a split there.
The same to TLB shootdown: the pmd path (which was for x86 only) has the
trick of using _ad() version of pmdp_invalidate*() which can avoid one
redundant TLB, but let's also leave that for later. Again, the larger the
mapping, the smaller of such effect.
There's some difference on handling "retry" for change_huge_pud() (where
it can return 0): it isn't like change_huge_pmd(), as the pmd version is
safe with all conditions handled in change_pte_range() later, thanks to
Hugh's new pte_offset_map_lock(). In short, change_pte_range() is simply
smarter. For that, change_pud_range() will need proper retry if it races
with something else when a huge PUD changed from under us.
The last thing to mention is currently the PUD path ignores the huge pte
numa counter (NUMA_HUGE_PTE_UPDATES), not only because DAX is not
applicable to NUMA, but also that it's ambiguous on its own to decide how
to account pud in this case. In one earlier version of this patchset I
proposed to remove the counter as it doesn't even look right to do the
accounting as of now [1], but then a further discussion suggests we can
leave that for later, as that doesn't block this series if we choose to
ignore that counter. That's what this patch does, by ignoring it.
When at it, touch up the comment in pgtable_split_needed() to make it
generic to either pmd or pud file THPs.
Peter Xu [Mon, 12 Aug 2024 18:12:24 +0000 (14:12 -0400)]
mm/x86: add missing pud helpers
Some new helpers will be needed for pud entry updates soon. Introduce
these helpers by referencing the pmd ones. Namely:
- pudp_invalidate(): this helper invalidates a huge pud before a
split happens, so that the invalidated pud entry will make sure no
race will happen (either with software, like a concurrent zap, or
hardware, like a/d bit lost).
- pud_modify(): this helper applies a new pgprot to an existing huge
pud mapping.
For more information on why we need these two helpers, please refer to the
corresponding pmd helpers in the mprotect() code path.
When at it, simplify the pud_modify()/pmd_modify() comments on shadow
stack pgtable entries to reference pte_modify() to avoid duplicating the
whole paragraph three times.
Peter Xu [Mon, 12 Aug 2024 18:12:23 +0000 (14:12 -0400)]
mm/x86: implement arch_check_zapped_pud()
Introduce arch_check_zapped_pud() to sanity check shadow stack on PUD
zaps. It has the same logic as the PMD helper.
One thing to mention is, it might be a good idea to use page_table_check
in the future for trapping wrong setups of shadow stack pgtable entries
[1]. That is left for the future as a separate effort.
Peter Xu [Mon, 12 Aug 2024 18:12:22 +0000 (14:12 -0400)]
mm/x86: make pud_leaf() only care about PSE bit
When working on mprotect() on 1G dax entries, I hit an zap bad pud error
when zapping a huge pud that is with PROT_NONE permission.
Here the problem is x86's pud_leaf() requires both PRESENT and PSE bits
set to report a pud entry as a leaf, but that doesn't look right, as it's
not following the pXd_leaf() definition that we stick with so far, where
PROT_NONE entries should be reported as leaves.
To fix it, change x86's pud_leaf() implementation to only check against
PSE bit to report a leaf, irrelevant of whether PRESENT bit is set.
Peter Xu [Mon, 12 Aug 2024 18:12:21 +0000 (14:12 -0400)]
mm/powerpc: add missing pud helpers
Some new helpers will be needed for pud entry updates soon. Introduce
these helpers by referencing the pmd ones. Namely:
- pudp_invalidate(): this helper invalidates a huge pud before a split
happens, so that the invalidated pud entry will make sure no race will
happen (either with software, like a concurrent zap, or hardware, like
a/d bit lost).
- pud_modify(): this helper applies a new pgprot to an existing huge pud
mapping.
For more information on why we need these two helpers, please refer to the
corresponding pmd helpers in the mprotect() code path.
Peter Xu [Mon, 12 Aug 2024 18:12:20 +0000 (14:12 -0400)]
mm/mprotect: push mmu notifier to PUDs
mprotect() does mmu notifiers in PMD levels. It's there since 2014 of
commit a5338093bfb4 ("mm: move mmu notifier call from change_protection to
change_pmd_range").
At that time, the issue was that NUMA balancing can be applied on a huge
range of VM memory, even if nothing was populated. The notification can
be avoided in this case if no valid pmd detected, which includes either
THP or a PTE pgtable page.
Now to pave way for PUD handling, this isn't enough. We need to generate
mmu notifications even on PUD entries properly. mprotect() is currently
broken on PUD (e.g., one can easily trigger kernel error with dax 1G
mappings already), this is the start to fix it.
To fix that, this patch proposes to push such notifications to the PUD
layers.
There is risk on regressing the problem Rik wanted to resolve before, but I
think it shouldn't really happen, and I still chose this solution because
of a few reasons:
1) Consider a large VM that should definitely contain more than GBs of
memory, it's highly likely that PUDs are also none. In this case there
will have no regression.
2) KVM has evolved a lot over the years to get rid of rmap walks, which
might be the major cause of the previous soft-lockup. At least TDP MMU
already got rid of rmap as long as not nested (which should be the major
use case, IIUC), then the TDP MMU pgtable walker will simply see empty VM
pgtable (e.g. EPT on x86), the invalidation of a full empty region in
most cases could be pretty fast now, comparing to 2014.
3) KVM has explicit code paths now to even give way for mmu notifiers
just like this one, e.g. in commit d02c357e5bfa ("KVM: x86/mmu: Retry
fault before acquiring mmu_lock if mapping is changing"). It'll also
avoid contentions that may also contribute to a soft-lockup.
4) Stick with PMD layer simply don't work when PUD is there... We need
one way or another to fix PUD mappings on mprotect().
Pushing it to PUD should be the safest approach as of now, e.g. there's yet
no sign of huge P4D coming on any known archs.
Peter Xu [Mon, 12 Aug 2024 18:12:19 +0000 (14:12 -0400)]
mm/dax: dump start address in fault handler
Patch series "mm/mprotect: Fix dax puds", v5.
Dax supports pud pages for a while, but mprotect on puds was missing since
the start. This series tries to fix that by providing pud handling in
mprotect(). The goal is to add more types of pud mappings like hugetlb or
pfnmaps. This series paves way for it by fixing known pud entries.
Considering nobody reported this until when I looked at those other types
of pud mappings, I am thinking maybe it doesn't need to be a fix for
stable and this may not need to be backported. I would guess whoever
cares about mprotect() won't care 1G dax puds yet, vice versa. I hope
fixing that in new kernels would be fine, but I'm open to suggestions.
There're a few small things changed to teach mprotect work on PUDs. E.g.
it will need to start with dropping NUMA_HUGE_PTE_UPDATES which may stop
making sense when there can be more than one type of huge pte. OTOH,
we'll also need to push the mmu notifiers from pmd to pud layers, which
might need some attention but so far I think it's safe. For such details,
please refer to each patch's commit message.
The mprotect() pud process should be straightforward, as I kept it as
simple as possible. There's no NUMA handled as dax simply doesn't support
that. There's also no userfault involvements as file memory (even if work
with userfault-wp async mode) will need to split a pud, so pud entry
doesn't need to yet know userfault's existance (but hugetlb entries will;
that's also for later).
This patch (of 7):
Currently the dax fault handler dumps the vma range when dynamic debugging
enabled. That's mostly not useful. Dump the (aligned) address instead
with the order info.
Yuanchu Xie [Tue, 13 Aug 2024 16:37:59 +0000 (09:37 -0700)]
mm: multi-gen LRU: ignore non-leaf pmd_young for force_scan=true
When non-leaf pmd accessed bits are available, MGLRU page table walks can
clear the non-leaf pmd accessed bit and ignore the accessed bit on the pte
if it's on a different node, skipping a generation update as well. If
another scan occurs on the same node as said skipped pte.
The non-leaf pmd accessed bit might remain cleared and the pte accessed
bits won't be checked. While this is sufficient for reclaim-driven aging,
where the goal is to select a reasonably cold page, the access can be
missed when aging proactively for workingset estimation of a node/memcg.
In more detail, get_pfn_folio returns NULL if the folio's nid != node
under scanning, so the page table walk skips processing of said pte. Now
the pmd_young flag on this pmd is cleared, and if none of the pte's are
accessed before another scan occurs on the folio's node, the pmd_young
check fails and the pte accessed bit is skipped.
Since force_scan disables various other optimizations, we check force_scan
to ignore the non-leaf pmd accessed bit.
Miao Wang [Tue, 13 Aug 2024 17:12:13 +0000 (01:12 +0800)]
mm: vmalloc: add optimization hint on page existence check
In commit 21e516b913c1 ("mm: vmalloc: dump page owner info if page is
already mapped"), a BUG_ON macro was changed into an if statement, where
the compiler optimization hint introduced in the BUG_ON macro was removed
along with this change. This patch adds back the hint.
Page isolation machinery doesn't know anything about unaccepted memory and
considers it non-free. It leads to alloc_contig_pages() failure.
Treat unaccepted memory as free and accept memory on pageblock isolation.
Once memory is accepted it becomes PageBuddy() and page isolation knows
how to deal with them.
Currently, the kernel only accepts memory in get_page_from_freelist(), but
there is another path that directly takes pages from free lists -
__alloc_page_bulk(). This function can consume all accepted memory and
will resort to __alloc_pages_noprof() if necessary.
Conditionally accepted in __alloc_pages_bulk().
The same issue may arise due to deferred page initialization. Kick the
deferred initialization machinery before abandoning the zone, as the
kernel does in get_page_from_freelist().
Patch series "mm: Fix several issues with unaccepted memory", v2.
The patchset addresses several issues related to unaccepted memory.
Pacth 1/7 preparatory cleanup.
Patch 2/7 ensures that __alloc_pages_bulk() will not exhaust all
accepted memory without accepting more.
Patches 3/7-5/7 are preparations for patch 6/7, which fixes
alloc_config_page() on machines with unaccepted memory. This allows, for
example, the allocation of gigantic pages at runtime.
Patch 7/7 enables the kernel to accept memory up to the promo watermark.
This patch (of 7):
Add dummy _deferred_grow_zone() for !DEFERRED_STRUCT_PAGE_INIT and remove
#ifdefs in two places.
Zi Yan [Fri, 9 Aug 2024 14:59:06 +0000 (10:59 -0400)]
mm/migrate: move common code to numa_migrate_check (was numa_migrate_prep)
do_numa_page() and do_huge_pmd_numa_page() share a lot of common code. To
reduce redundancy, move common code to numa_migrate_prep() and rename the
function to numa_migrate_check() to reflect its functionality.
Now do_huge_pmd_numa_page() also checks shared folios to set TNF_SHARED
flag.
Shakeel Butt [Fri, 9 Aug 2024 17:26:18 +0000 (10:26 -0700)]
memcg: replace memcg ID idr with xarray
At the moment memcg IDs are managed through IDR which requires external
synchronization mechanisms and makes the allocation code a bit awkward.
Let's switch to xarray and make the code simpler.
The unuse_pte_range() caller only wants the folio while do_swap_page()
wants both the page and the folio. Since do_swap_page() already has logic
for handling both the folio and the page, move the folio-to-page logic
there. This also lets us allocate larger folios in the SWP_SYNCHRONOUS_IO
path in future.
fs: remove calls to set and clear the folio error flag
Nobody checks the folio error flag any more, so we can stop setting and
clearing it. Also remove the documentation suggesting to not bother
setting the error bit.
Jianhui Zhou [Wed, 7 Aug 2024 07:44:48 +0000 (15:44 +0800)]
percpu: remove pcpu_alloc_size()
pcpu_alloc_size() was added in 7ac5c53e0073 "mm/percpu.c: introduce
pcpu_alloc_size()", which is used to get the allocated memory size in bpf.
However, pcpu_alloc_size() is no longer used in "bpf: Use c->unit_size to
select target cache during free" because its actuall allocated memory size
may change at runtime due to its slab merging mechanism. Therefore,
pcpu_alloc_size() can be removed.
mm/rmap: minimize folio->_nr_pages_mapped updates when batching PTE (un)mapping
It is not immediately obvious, but we can move the folio->_nr_pages_mapped
update out of the loop and reduce the number of atomic ops without
affecting the stats.
The important point to realize is that only removing the last PMD mapping
will result in _nr_pages_mapped going below ENTIRELY_MAPPED, not the
individual atomic_inc_return_relaxed() calls. Concurrent races with
removal of PMD mappings should be handled as expected, just like when we
would have such races right now on a single mapcount update.
In a simple munmap() microbenchmark [1] on 1 GiB of memory backed by the
same PTE-mapped folio size (only mapped by a single process such that they
will get completely unmapped), this change results in a speedup (positive
is good) per folio size on a x86-64 Intel machine of roughly (a bit of
noise expected):
Marco Elver [Mon, 5 Aug 2024 12:39:39 +0000 (14:39 +0200)]
kfence: introduce burst mode
Introduce burst mode, which can be configured with kfence.burst=$count,
where the burst count denotes the additional successive slab allocations
to be allocated through KFENCE for each sample interval.
The idea is that this can give developers an additional knob to make
KFENCE more aggressive when debugging specific issues of systems where
either rebooting or recompiling the kernel with KASAN is not possible.
Experiment: To assess the effectiveness of the new option, we randomly
picked a recent out-of-bounds [1] and use-after-free bug [2], each with a
reproducer provided by syzbot, that initially detected these bugs with
KASAN. We then tried to reproduce the bugs with KFENCE below.
With the Default and even the Aggressive configs the results are
unsurprising, given KFENCE has not been designed for deterministic bug
detection of small test cases.
However, when enabling burst mode with relatively large burst count,
KFENCE can start to detect heap memory-safety bugs even in simpler test
cases with high probability (in the above cases with ~80% probability).
Jann Horn [Mon, 5 Aug 2024 12:52:03 +0000 (14:52 +0200)]
mm: fix (harmless) type confusion in lock_vma_under_rcu()
There is a (harmless) type confusion in lock_vma_under_rcu(): After
vma_start_read(), we have taken the VMA lock but don't know yet whether
the VMA has already been detached and scheduled for RCU freeing. At this
point, ->vm_start and ->vm_end are accessed.
vm_area_struct contains a union such that ->vm_rcu uses the same memory as
->vm_start and ->vm_end; so accessing ->vm_start and ->vm_end of a
detached VMA is illegal and leads to type confusion between union members.
Fix it by reordering the vma->detached check above the address checks, and
document the rules for RCU readers accessing VMAs.
This will probably change the number of observed VMA_LOCK_MISS events
(since previously, trying to access a detached VMA whose ->vm_rcu has been
scheduled would bail out when checking the fault address against the
rcu_head members reinterpreted as VMA bounds).
Nhat Pham [Mon, 5 Aug 2024 23:22:43 +0000 (16:22 -0700)]
zswap: track swapins from disk more accurately
Currently, there are a couple of issues with our disk swapin tracking for
dynamic zswap shrinker heuristics:
1. We only increment the swapin counter on pivot pages. This means we
are not taking into account pages that also need to be swapped in,
but are already taken care of as part of the readahead window.
2. We are also incrementing when the pages are read from the zswap pool,
which is inaccurate.
This patch rectifies these issues by incrementing the counter whenever we
need to perform a non-zswap read. Note that we are slightly overcounting,
as a page might be read into memory by the readahead algorithm even though
it will not be neeeded by users - however, this is an acceptable
inaccuracy, as the readahead logic itself will adapt to these kind of
scenarios.
To test this change, I built the kernel under a cgroup with its memory.max
set to 2 GB:
With this change, the kernel CPU time reduces by a further 1.7%, and the
real time is reduced by another 3.3%, compared to just the second chance
algorithm by itself. The swapins count also reduces by another 13.85%.
Combinng the two changes, we reduce the real time by 10.32%, kernel CPU
time by 3%, and number of swapins by 64.12%.
To gauge the new scheme's ability to offload cold data, I ran another
benchmark, in which the kernel was built under a cgroup with memory.max
set to 3 GB, but with 0.5 GB worth of cold data allocated before each
build (in a shmem file).
Notice that we actually observe a 21% increase in the number of written
back pages - so the new scheme is just as good, if not better at
offloading pages from the zswap pool when they are cold. Build time
reduces by around 0.7% as a result.
Nhat Pham [Mon, 5 Aug 2024 23:22:42 +0000 (16:22 -0700)]
zswap: implement a second chance algorithm for dynamic zswap shrinker
Patch series "improving dynamic zswap shrinker protection scheme", v3.
When experimenting with the memory-pressure based (i.e "dynamic") zswap
shrinker in production, we observed a sharp increase in the number of
swapins, which led to performance regression. We were able to trace this
regression to the following problems with the shrinker's warm pages
protection scheme:
1. The protection decays way too rapidly, and the decaying is coupled with
zswap stores, leading to anomalous patterns, in which a small batch of
zswap stores effectively erase all the protection in place for the
warmer pages in the zswap LRU.
This observation has also been corroborated upstream by Takero Funaki
(in [1]).
2. We inaccurately track the number of swapped in pages, missing the
non-pivot pages that are part of the readahead window, while counting
the pages that are found in the zswap pool.
To alleviate these two issues, this patch series improve the dynamic zswap
shrinker in the following manner:
1. Replace the protection size tracking scheme with a second chance
algorithm. This new scheme removes the need for haphazard stats
decaying, and automatically adjusts the pace of pages aging with memory
pressure, and writeback rate with pool activities: slowing down when
the pool is dominated with zswpouts, and speeding up when the pool is
dominated with stale entries.
2. Fix the tracking of the number of swapins to take into account
non-pivot pages in the readahead window.
With these two changes in place, in a kernel-building benchmark without
any cold data added, the number of swapins is reduced by 64.12%. This
translate to a 10.32% reduction in build time. We also observe a 3%
reduction in kernel CPU time.
In another benchmark, with cold data added (to gauge the new algorithm's
ability to offload cold data), the new second chance scheme outperforms
the old protection scheme by around 0.7%, and actually written back around
21% more pages to backing swap device. So the new scheme is just as good,
if not even better than the old scheme on this front as well.
Current zswap shrinker's heuristics to prevent overshrinking is brittle
and inaccurate, specifically in the way we decay the protection size (i.e
making pages in the zswap LRU eligible for reclaim).
We currently decay protection aggressively in zswap_lru_add() calls. This
leads to the following unfortunate effect: when a new batch of pages enter
zswap, the protection size rapidly decays to below 25% of the zswap LRU
size, which is way too low.
We have observed this effect in production, when experimenting with the
zswap shrinker: the rate of shrinking shoots up massively right after a
new batch of zswap stores. This is somewhat the opposite of what we want
originally - when new pages enter zswap, we want to protect both these new
pages AND the pages that are already protected in the zswap LRU.
Replace existing heuristics with a second chance algorithm
1. When a new zswap entry is stored in the zswap pool, its referenced
bit is set.
2. When the zswap shrinker encounters a zswap entry with the referenced
bit set, give it a second chance - only flips the referenced bit and
rotate it in the LRU.
3. If the shrinker encounters the entry again, this time with its
referenced bit unset, then it can reclaim the entry.
In this manner, the aging of the pages in the zswap LRUs are decoupled
from zswap stores, and picks up the pace with increasing memory pressure
(which is what we want).
The second chance scheme allows us to modulate the writeback rate based on
recent pool activities. Entries that recently entered the pool will be
protected, so if the pool is dominated by such entries the writeback rate
will reduce proportionally, protecting the workload's workingset.On the
other hand, stale entries will be written back quickly, which increases
the effective writeback rate.
The referenced bit is added at the hole after the `length` field of struct
zswap_entry, so there is no extra space overhead for this algorithm.
We will still maintain the count of swapins, which is consumed and
subtracted from the lru size in zswap_shrinker_count(), to further
penalize past overshrinking that led to disk swapins. The idea is that
had we considered this many more pages in the LRU active/protected, they
would not have been written back and we would not have had to swapped them
in.
To test this new heuristics, I built the kernel under a cgroup with
memory.max set to 2G, on a host with 36 cores:
David Gow [Sat, 3 Aug 2024 07:46:41 +0000 (15:46 +0800)]
mm: only enforce minimum stack gap size if it's sensible
The generic mmap_base code tries to leave a gap between the top of the
stack and the mmap base address, but enforces a minimum gap size (MIN_GAP)
of 128MB, which is too large on some setups. In particular, on arm tasks
without ADDR_LIMIT_32BIT, the STACK_TOP value is less than 128MB, so it's
impossible to fit such a gap in.
Only enforce this minimum if MIN_GAP < MAX_GAP, as we'd prefer to honour
MAX_GAP, which is defined proportionally, so scales better and always
leaves us with both _some_ stack space and some room for mmap.
This fixes the usercopy KUnit test suite on 32-bit arm, as it doesn't set
any personality flags so gets the default (in this case 26-bit) task size.
This test can be run with: ./tools/testing/kunit/kunit.py run --arch arm
usercopy --make_options LLVM=1
s390/mm/fault: convert do_secure_storage_access() from follow_page() to folio_walk
Let's get rid of another follow_page() user and perform the conversion
under PTL: Note that this is also what follow_page_pte() ends up doing.
Unfortunately we cannot currently optimize out the additional reference,
because arch_make_folio_accessible() must be called with a raised refcount
to protect against concurrent conversion to secure. We can just move the
arch_make_folio_accessible() under the PTL, like follow_page_pte() would.
We'll effectively drop the "writable" check implied by FOLL_WRITE:
follow_page_pte() would also not check that when calling
arch_make_folio_accessible(), so there is no good reason for doing that
here.
We'll lose the secretmem check from follow_page() as well, about which we
shouldn't really care.
s390/uv: convert gmap_destroy_page() from follow_page() to folio_walk
Let's get rid of another follow_page() user and perform the UV calls under
PTL -- which likely should be fine.
No need for an additional reference while holding the PTL:
uv_destroy_folio() and uv_convert_from_secure_folio() raise the refcount,
so any concurrent make_folio_secure() would see an unexpted reference and
cannot set PG_arch_1 concurrently.
Do we really need a writable PTE? Likely yes, because the "destroy" part
is, in comparison to the export, a destructive operation. So we'll keep
the writability check for now.
We'll lose the secretmem check from follow_page(). Likely we don't care
about that here.
mm/huge_memory: convert split_huge_pages_pid() from follow_page() to folio_walk
Let's remove yet another follow_page() user. Note that we have to do the
split without holding the PTL, after folio_walk_end(). We don't care
about losing the secretmem check in follow_page().
projects / linux.git / log