Patch series "ptrace: introduce PTRACE_SET_SYSCALL_INFO API", v7.
PTRACE_SET_SYSCALL_INFO is a generic ptrace API that complements
PTRACE_GET_SYSCALL_INFO by letting the ptracer modify details of system
calls the tracee is blocked in.
This API allows ptracers to obtain and modify system call details in a
straightforward and architecture-agnostic way, providing a consistent way
of manipulating the system call number and arguments across architectures.
As in case of PTRACE_GET_SYSCALL_INFO, PTRACE_SET_SYSCALL_INFO also does
not aim to address numerous architecture-specific system call ABI
peculiarities, like differences in the number of system call arguments for
such system calls as pread64 and preadv.
The current implementation supports changing only those bits of system
call information that are used by strace system call tampering, namely,
syscall number, syscall arguments, and syscall return value.
Support of changing additional details returned by
PTRACE_GET_SYSCALL_INFO, such as instruction pointer and stack pointer,
could be added later if needed, by using struct ptrace_syscall_info.flags
to specify the additional details that should be set. Currently, "flags"
and "reserved" fields of struct ptrace_syscall_info must be initialized
with zeroes; "arch", "instruction_pointer", and "stack_pointer" fields are
currently ignored.
PTRACE_SET_SYSCALL_INFO currently supports only PTRACE_SYSCALL_INFO_ENTRY,
PTRACE_SYSCALL_INFO_EXIT, and PTRACE_SYSCALL_INFO_SECCOMP operations.
Other operations could be added later if needed.
Ideally, PTRACE_SET_SYSCALL_INFO should have been introduced along with
PTRACE_GET_SYSCALL_INFO, but it didn't happen. The last straw that
convinced me to implement PTRACE_SET_SYSCALL_INFO was apparent failure to
provide an API of changing the first system call argument on riscv
architecture [1].
ptrace(2) man page:
long ptrace(enum __ptrace_request request, pid_t pid, void *addr, void *data);
...
PTRACE_SET_SYSCALL_INFO
Modify information about the system call that caused the stop.
The "data" argument is a pointer to struct ptrace_syscall_info
that specifies the system call information to be set.
The "addr" argument should be set to sizeof(struct ptrace_syscall_info)).
This patch (of 6):
hexagon is the only architecture that provides HAVE_ARCH_TRACEHOOK but
doesn't define syscall_set_return_value(). Since this function is going
to be needed on all HAVE_ARCH_TRACEHOOK architectures to implement
PTRACE_SET_SYSCALL_INFO API, add it on hexagon, too.
Link: https://lore.kernel.org/all/59505464-c84a-403d-972f-d4b2055eeaac@gmail.com/ [1]
Link: https://lkml.kernel.org/r/20250303111953.GB24170@strace.io
Signed-off-by: Dmitry V. Levin <ldv@strace.io>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Alexey Gladkov (Intel) <legion@kernel.org>
Cc: Andreas Larsson <andreas@gaisler.com>
Cc: anton ivanov <anton.ivanov@cambridgegreys.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Brian Cain <bcain@quicinc.com>
Cc: Charlie Jenkins <charlie@rivosinc.com>
Cc: Christian Borntraeger <borntraeger@linux.ibm.com>
Cc: Christian Zankel <chris@zankel.net>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Davide Berardi <berardi.dav@gmail.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Dinh Nguyen <dinguyen@kernel.org>
Cc: Eugene Syromyatnikov <evgsyr@gmail.com>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Guo Ren <guoren@kernel.org>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Helge Deller <deller@gmx.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Huacai Chen <chenhuacai@kernel.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Berg <johannes@sipsolutions.net>
Cc: John Paul Adrian Glaubitz <glaubitz@physik.fu-berlin.de>
Cc: Jonas Bonn <jonas@southpole.se>
Cc: Maciej W. Rozycki <macro@orcam.me.uk>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Michal Simek <monstr@monstr.eu>
Cc: Mike Frysinger <vapier@gentoo.org>
Cc: Naveen N Rao <naveen@kernel.org>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Renzo Davoi <renzo@cs.unibo.it>
Cc: Richard Weinberger <richard@nod.at>
Cc: Rich Felker <dalias@libc.org>
Cc: Russel King <linux@armlinux.org.uk>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Stafford Horne <shorne@gmail.com>
Cc: Stefan Kristiansson <stefan.kristiansson@saunalahti.fi>
Cc: Sven Schnelle <svens@linux.ibm.com>
Cc: Thomas Gleinxer <tglx@linutronix.de>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Vineet Gupta <vgupta@kernel.org>
Cc: WANG Xuerui <kernel@xen0n.name>
Cc: Will Deacon <will@kernel.org>
Cc: Yoshinori Sato <ysato@users.sourceforge.jp>
Cc: Eugene Syromiatnikov <esyr@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Replace cluster_swap_free_nr() with swap_entries_put_[map/cache]() to
remove repeat code and leverage batch-remove for entries with last flag.
After removing cluster_swap_free_nr, only functions with "_nr" suffix
could free entries spanning cross clusters. Add corresponding description
in comment of swap_entries_put_map_nr() as is first function with "_nr"
suffix and have a non-suffix variant function swap_entries_put_map().
Link: https://lkml.kernel.org/r/20250325162528.68385-9-shikemeng@huaweicloud.com
Signed-off-by: Kemeng Shi <shikemeng@huaweicloud.com>
Reviewed-by: Tim Chen <tim.c.chen@linux.intel.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Kairui Song <kasong@tencent.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
1. Factor out general swap_entries_put_map() helper to drop entries
belonging to one cluster. If entries are last map, free entries in
batch, otherwise put entries with cluster lock acquired and released
only once.
2. Iterate and call swap_entries_put_map() for each cluster in
swap_entries_put_nr() to leverage batch-remove for last map belonging
to one cluster and reduce lock acquire/release in fallback case.
3. As swap_entries_put_nr() won't handle SWAP_HSA_CACHE drop, rename
it to swap_entries_put_map_nr().
4. As we won't drop each entry invidually with swap_entry_put() now,
do reclaim in free_swap_and_cache_nr() because
swap_entries_put_map_nr() is general routine to drop reference and the
relcaim work should only be done in free_swap_and_cache_nr(). Remove
stale comment accordingly.
Link: https://lkml.kernel.org/r/20250325162528.68385-7-shikemeng@huaweicloud.com
Signed-off-by: Kemeng Shi <shikemeng@huaweicloud.com>
Reviewed-by: Tim Chen <tim.c.chen@linux.intel.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Kairui Song <kasong@tencent.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
In swap_entry_put_locked(), we will set slot to SWAP_HAS_CACHE before
using swap_entries_free() to do actual swap entry freeing. This introduce
an unnecessary intermediate state. By using swap_entries_free() in
swap_entry_put_locked(), we can eliminate the need to set slot to
SWAP_HAS_CACHE. This change would make the behavior of
swap_entry_put_locked() more consistent with other put() operations which
will do actual free work after put last reference.
Link: https://lkml.kernel.org/r/20250325162528.68385-4-shikemeng@huaweicloud.com
Signed-off-by: Kemeng Shi <shikemeng@huaweicloud.com>
Reviewed-by: Tim Chen <tim.c.chen@linux.intel.com>
Reviewed-by: Kairui Song <kasong@tencent.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "Minor cleanups and improvements to swap freeing code", v4.
This series contains some cleanups and improvements which are made
during learning swapfile. Here is a summary of the changes:
1. Function naming improvments.
- Use "put" instead of "free" to name functions which only do actual
free when count drops to zero.
- Use "entry" to name function only frees one swap slot. Use
"entries" to name function could may free multi swap slots within one
cluster. Use "_nr" suffix to name function which could free multi
swap slots spanning cross multi clusters.
2. Eliminate the need to set swap slot to intermediate SWAP_HAS_CACHE
value before do actual free by using swap_entry_range_free()
3. Add helpers swap_entries_put_map() and swap_entries_put_cache() as
a general-purpose routine to free swap entries within a single cluster
which will try batch-remove first and fallback to put eatch entry
indvidually with cluster lock acquired/released only once. By using
these helpers, we could remove repeated code, levarage batch-remove in
more cases and aoivd to acquire/release cluster lock for each single
swap entry.
This patch (of 8):
In __swap_entry_free[_locked] and __swap_entries_free, we decrease count
first and only free swap entry if count drops to zero. This behavior is
more akin to a put() operation rather than a free() operation. Therefore,
rename these functions with "put" instead of "free". Additionally, add
"_nr" suffix to swap_entries_put to indicate the input range may span swap
clusters.
Link: https://lkml.kernel.org/r/20250325162528.68385-1-shikemeng@huaweicloud.com
Link: https://lkml.kernel.org/r/20250325162528.68385-2-shikemeng@huaweicloud.com
Signed-off-by: Kemeng Shi <shikemeng@huaweicloud.com>
Reviewed-by: Tim Chen <tim.c.chen@linux.intel.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Kairui Song <kasong@tencent.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
When handing slab objects, we use obj_cgroup_[un]charge() for (un)charging
and mod_objcg_state() to account NR_SLAB_[UN]RECLAIMABLE_B. All these
operations use the percpu stock for performance. However with the calls
being separate, the stock_lock is taken twice in each case.
By refactoring the code, we can turn mod_objcg_state() into
__account_obj_stock() which is called on a stock that's already locked and
validated. On the charging side we can call this function from
consume_obj_stock() when it succeeds, and refill_obj_stock() in the
fallback. We just expand parameters of these functions as necessary. The
uncharge side from __memcg_slab_free_hook() is just the call to
refill_obj_stock().
Other callers of obj_cgroup_[un]charge() (i.e. not slab) simply pass the
extra parameters as NULL/zeroes to skip the __account_obj_stock()
operation.
In __memcg_slab_post_alloc_hook() we now charge each object separately,
but that's not a problem as we did call mod_objcg_state() for each object
separately, and most allocations are non-bulk anyway. This could be
improved by batching all operations until slab_pgdat(slab) changes.
Some preliminary benchmarking with a kfree(kmalloc()) loop of 10M
iterations with/without __GFP_ACCOUNT:
Before the patch:
kmalloc/kfree !memcg: 581390144 cycles
kmalloc/kfree memcg: 783689984 cycles
After the patch:
kmalloc/kfree memcg: 658723808 cycles
More than half of the overhead of __GFP_ACCOUNT relative to
non-accounted case seems eliminated.
Link: https://lkml.kernel.org/r/20250404013913.1663035-9-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
obj_cgroup_release is called when all the references to the objcg have
been released i.e. no more memory objects are pointing to it. Most
probably objcg->memcg will be pointing to some ancestor memcg. In
obj_cgroup_release(), the kernel calls obj_cgroup_uncharge_pages() which
refills the local stock.
There is no need to refill the local stock with some ancestor memcg and
flush the local stock. Let's decouple obj_cgroup_release() from the local
stock by uncharging instead of refilling. One additional benefit of this
change is that it removes the requirement to only call obj_cgroup_put()
outside of local_lock.
Link: https://lkml.kernel.org/r/20250404013913.1663035-6-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Currently drain_obj_stock() can potentially call __refill_stock which
accesses local cpu stock and thus requires memcg stock's local_lock.
However if we look at the code paths leading to drain_obj_stock(), there
is never a good reason to refill the memcg stock at all from it.
At the moment, drain_obj_stock can be called from reclaim, hotplug cpu
teardown, mod_objcg_state() and refill_obj_stock(). For reclaim and
hotplug there is no need to refill. For the other two paths, most
probably the newly switched objcg would be used in near future and thus no
need to refill stock with the older objcg.
In addition, __refill_stock() from drain_obj_stock() happens on rare
cases, so performance is not really an issue. Let's just uncharge
directly instead of refill which will also decouple drain_obj_stock from
local cpu stock and local_lock requirements.
Link: https://lkml.kernel.org/r/20250404013913.1663035-3-shakeel.butt@linux.dev
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm/compaction: allow more aggressive proactive compaction",
v4.
Our goal is to keep memory usage of a VM low on the host. For that
reason, we use free page reporting which by default reports free pages of
order 9 and larger to the host to be freed. The feature works well only
if the memory in the guest is not fragmented below pages of order 9.
Proactive compaction can be reused to achieve defragmentation after some
parameter tweaking.
When the fragmentation score (lower is better) gets larger than the high
watermark, proactive compaction kicks in. Compaction stops when the score
goes below the low watermark (or no progress is made and backoff kicks
in). Let's define the difference between high and low watermarks as
leeway. Before these changes, the minimum possible value for low
watermark was 5 and the leeway was hardcoded to 10 (so minimum possible
value for high watermark was 15).
To test this, I created a VM with 19GB of memory and free page reporting
enabled. The VM was ~idle. I meassured the memory usage from inside the
guest (/proc/meminfo) and from the host (provided by the hypervisor).
Before:
https://drive.google.com/file/d/1Xw23lRry_PgEH3f6QRnSGvoHh2u9UHyI/view?usp=sharing
After:
https://drive.google.com/file/d/1wMhpIzepx6t44F70yCPA50n1S5V2AT-a/view?usp=sharing
This patch (of 2):
Previously a min cap of 5 has been set in the commit introducing proactive
compaction. This was to make sure users don't hurt themselves by setting
the proactiveness to 100 and making their system unresponsive. But the
compaction mechanism has a backoff mechanism that will sleep for 30s if no
progress is made, so I don't see a significant risk here. My system (19GB
of memory) has been perfectly fine with both watermarks hardcoded to 0.
Link: https://lkml.kernel.org/r/20250404111103.1994507-1-mclapinski@google.com
Link: https://lkml.kernel.org/r/20250404111103.1994507-2-mclapinski@google.com
Signed-off-by: Michal Clapinski <mclapinski@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The writeback interface supports a page_index=N parameter which performs
writeback of the given page. Since we rarely need to writeback just one
single page, the typical use case involves a number of writeback calls,
each performing writeback of one page:
echo page_index=100 > zram0/writeback
...
echo page_index=200 > zram0/writeback
echo page_index=500 > zram0/writeback
...
echo page_index=700 > zram0/writeback
One obvious downside of this is that it increases the number of syscalls.
Less obvious, but a significantly more important downside, is that when
given only one page to post-process zram cannot perform an optimal target
selection. This becomes a critical limitation when writeback_limit is
enabled, because under writeback_limit we want to guarantee the highest
memory savings hence we first need to writeback pages that release the
highest amount of zsmalloc pool memory.
This patch adds page_indexes=LOW-HIGH parameter to the writeback
interface:
echo page_indexes=100-200 page_indexes=500-700 > zram0/writeback
This gives zram a chance to apply an optimal target selection strategy on
each iteration of the writeback loop.
We also now permit multiple page_index parameters per call (previously
zram would recognize only one page_index) and a mix or single pages and
page ranges:
echo page_index=42 page_index=99 page_indexes=100-200 \
page_indexes=500-700 > zram0/writeback
Apart from that the patch also unifies parameters passing and resembles
other "modern" zram device attributes (e.g. recompression), while the old
interface used a mixed scheme: values-less parameters for mode and a
key=value format for page_index. We still support the "old" value-less
format for compatibility reasons.
[senozhatsky@chromium.org: simplify parse_page_index() range checks, per Brian]
nk: https://lkml.kernel.org/r/20250404015327.2427684-1-senozhatsky@chromium.org
[sozhatsky@chromium.org: fix uninitialized variable in zram_writeback_slots(), per Dan]
nk: https://lkml.kernel.org/r/20250409112611.1154282-1-senozhatsky@chromium.org
Link: https://lkml.kernel.org/r/20250327015818.4148660-1-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Reviewed-by: Brian Geffon <bgeffon@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Richard Chang <richardycc@google.com>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Refactors the si_meminfo_node() function by reducing redundant code and
improving readability.
Moved the calculation of managed_pages inside the existing loop that
processes pgdat->node_zones, eliminating the need for a separate loop.
Simplified the logic by removing unnecessary preprocessor conditionals.
Ensured that both totalram, totalhigh, and other memory statistics are
consistently set without duplication.
This change results in cleaner and more efficient code without altering
functionality.
Link: https://lkml.kernel.org/r/20250325073803.852594-1-ye.liu@linux.dev
Signed-off-by: Ye Liu <liuye@kylinos.cn>
Acked-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "memory,x86,acpi: hotplug memory alignment advisement", v8.
When physical address regions are not aligned to memory block size, the
misaligned portion is lost (stranded capacity).
Block size (min/max/selected) is architecture defined. Most architectures
tend to use the minimum block size or some simplistic heurist. On x86,
memory block size increases up to 2GB, and is otherwise fitted to the
alignment of non-hotplug (i.e. not special purpose memory).
CXL exposes its memory for management through the ACPI CEDT (CXL Early
Detection Table) in a field called the CXL Fixed Memory Window. Per the
CXL specification, this memory must be aligned to at least 256MB.
When a CFMW aligns on a size less than the block size, this causes a loss
of up to 2GB per CFMW on x86. It is not uncommon for CFMW to be allocated
per-device - though this behavior is BIOS defined.
This patch set provides 3 things:
1) implement advise/query functions in driverse/base/memory.c to
report/query architecture agnostic hotplug block alignment advice.
2) update x86 memblock size logic to consider the hotplug advice
3) add code in acpi/numa/srat.c to report CFMW alignment advice
The advisement interfaces are design to be called during arch_init code
prior to allocator and smp_init. start_kernel will call these through
setup_arch() (via acpi and mm/init_64.c on x86), which occurs prior to
mm_core_init and smp_init - so no need for atomics.
There's an attempt to signal callers to advise() that query has already
occurred, but this is predicated on the notion that query actually occurs
(which presently only happens on the x86 arch). This is to assist
debugging future users. Otherwise, the advise() call has been marked
__init to help static discovery of bad call times.
Once query is called the first time, it will always return the same value.
Interfaces return -EBUSY and 0 respectively on systems without hotplug.
This patch (of 3):
Hotplug memory sources may have opinions on what the memblock size should
be - usually for alignment purposes. For example, CXL memory extents can
be 256MB with a matching alignment. If this size/alignment is smaller
than the block size, it can result in stranded capacity.
Implement memory_block_advise_max_size for use prior to allocator init,
for software to advise the system on the max block size.
Implement memory_block_probe_max_size for use by arch init code to
calculate the best block size. Use of advice is architecture defined.
The probe value can never change after first probe. Calls to advise after
probe will return -EBUSY to aid debugging.
On systems without hotplug, always return -ENODEV and 0 respectively.
Link: https://lkml.kernel.org/r/20250127153405.3379117-1-gourry@gourry.net
Link: https://lkml.kernel.org/r/20250127153405.3379117-2-gourry@gourry.net
Signed-off-by: Gregory Price <gourry@gourry.net>
Suggested-by: Ira Weiny <ira.weiny@intel.com>
Acked-by: David Hildenbrand <david@redhat.com>
Acked-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
Acked-by: Dan Williams <dan.j.williams@intel.com>
Tested-by: Fan Ni <fan.ni@samsung.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Acked-by: Oscar Salvador <osalvador@suse.de>
Cc: Alison Schofield <alison.schofield@intel.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Bruno Faccini <bfaccini@nvidia.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Haibo Xu <haibo1.xu@intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Len Brown <lenb@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Cc: Robert Richter <rrichter@amd.com>
Cc: Thomas Gleinxer <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
7775face20 ("memcg: killed threads should not invoke memcg OOM killer")
has added a bypass of the oom killer path for dying threads because a very
specific workload (described in the changelog) could hit "no killable
tasks" path. This itself is not fatal condition but it could be annoying
if this was a common case.
On the other hand the bypass has some issues on its own. Without
triggering oom killer we won't be able to trigger async oom reclaim
(oom_reaper) which can operate on killed tasks as well as long as they
still have their mm available. This could be the case during futex
cleanup when the memory as pointed out by Johannes in [1]. The said case
is still not fully understood but let's drop this bypass that was mostly
driven by an artificial workload and allow dying tasks to go into oom
path. This will make the code easier to reason about and also help corner
cases where oom_reaper could help to release memory.
Link: https://lore.kernel.org/all/20241212183012.GB1026@cmpxchg.org/T/#u [1]
Link: https://lkml.kernel.org/r/20250402090117.130245-1-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Currently, zsmalloc, zswap's and zram's backend memory allocator, does not
enforce any policy for the allocation of memory for the compressed data,
instead just adopting the memory policy of the task entering reclaim, or
the default policy (prefer local node) if no such policy is specified.
This can lead to several pathological behaviors in multi-node NUMA
systems:
1. Systems with CXL-based memory tiering can encounter the following
inversion with zswap/zram: the coldest pages demoted to the CXL tier
can return to the high tier when they are reclaimed to compressed swap,
creating memory pressure on the high tier.
2. Consider a direct reclaimer scanning nodes in order of allocation
preference. If it ventures into remote nodes, the memory it compresses
there should stay there. Trying to shift those contents over to the
reclaiming thread's preferred node further *increases* its local
pressure, and provoking more spills. The remote node is also the most
likely to refault this data again. This undesirable behavior was
pointed out by Johannes Weiner in [1].
3. For zswap writeback, the zswap entries are organized in
node-specific LRUs, based on the node placement of the original pages,
allowing for targeted zswap writeback for specific nodes.
However, the compressed data of a zswap entry can be placed on a
different node from the LRU it is placed on. This means that reclaim
targeted at one node might not free up memory used for zswap entries in
that node, but instead reclaiming memory in a different node.
All of these issues will be resolved if the compressed data go to the same
node as the original page. This patch encourages this behavior by having
zswap and zram pass the node of the original page to zsmalloc, and have
zsmalloc prefer the specified node if we need to allocate new (zs)pages
for the compressed data.
Note that we are not strictly binding the allocation to the preferred
node. We still allow the allocation to fall back to other nodes when the
preferred node is full, or if we have zspages with slots available on a
different node. This is OK, and still a strict improvement over the
status quo:
1. On a system with demotion enabled, we will generally prefer
demotions over compressed swapping, and only swap when pages have
already gone to the lowest tier. This patch should achieve the desired
effect for the most part.
2. If the preferred node is out of memory, letting the compressed data
going to other nodes can be better than the alternative (OOMs, keeping
cold memory unreclaimed, disk swapping, etc.).
3. If the allocation go to a separate node because we have a zspage
with slots available, at least we're not creating extra immediate
memory pressure (since the space is already allocated).
3. While there can be mixings, we generally reclaim pages in same-node
batches, which encourage zspage grouping that is more likely to go to
the right node.
4. A strict binding would require partitioning zsmalloc by node, which
is more complicated, and more prone to regression, since it reduces the
storage density of zsmalloc. We need to evaluate the tradeoff and
benchmark carefully before adopting such an involved solution.
[1]: https://lore.kernel.org/linux-mm/20250331165306.GC2110528@cmpxchg.org/
[senozhatsky@chromium.org: coding-style fixes]
Link: https://lkml.kernel.org/r/mnvexa7kseswglcqbhlot4zg3b3la2ypv2rimdl5mh5glbmhvz@wi6bgqn47hge
Link: https://lkml.kernel.org/r/20250402204416.3435994-1-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Gregory Price <gourry@gourry.net>
Acked-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Chengming Zhou <chengming.zhou@linux.dev>
Acked-by: Sergey Senozhatsky <senozhatsky@chromium.org> [zram, zsmalloc]
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Yosry Ahmed <yosry.ahmed@linux.dev> [zswap/zsmalloc]
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Joanthan Cameron <Jonathan.Cameron@huawei.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
