Version 1 and 2 of L3C PMU also use different HID. Make use of
struct acpi_device_id::driver_data for version specific information
rather than judge the version register. This will help to
simplify the probe process and also a bit easier for extension.
Acked-by: Jonathan Cameron <jonathan.cameron@huawei.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Signed-off-by: Yushan Wang <wangyushan12@huawei.com>
Reviewed-by: Jonathan Cameron <jonathan.cameron@huawei.com>
Signed-off-by: Will Deacon <will@kernel.org>
Currently Uncore PMU framework assume one PMU device only have one
interrupt and will help register the interrupt handler. It cannot
support a PMU with multiple interrupt resources. An uncore PMU may
have multiple interrupts that can share the same handler. Export
hisi_uncore_pmu_isr() to allow drivers register the irq handler by
their own routine.
Acked-by: Jonathan Cameron <jonathan.cameron@huawei.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Signed-off-by: Yushan Wang <wangyushan12@huawei.com>
Reviewed-by: Jonathan Cameron <jonathan.cameron@huawei.com>
Signed-off-by: Will Deacon <will@kernel.org>
Event ID is only using the attr::config bit [7, 0] but we check the
event range using the whole 64bit field. It blocks the usage of the
rest field of attr::config. Relax the check by only using the
bit [7, 0].
Acked-by: Jonathan Cameron <jonathan.cameron@huawei.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Signed-off-by: Yushan Wang <wangyushan12@huawei.com>
Signed-off-by: Will Deacon <will@kernel.org>
This adds a new dynamic PMU to the Perf Events framework to program and
control the Uncore PMUs in Fujitsu chips.
This driver exports formatting and event information to sysfs so it can
be used by the perf user space tools with the syntaxes:
perf stat -e pci_iod0_pci0/ea-pci/ ls
perf stat -e pci_iod0_pci0/event=0x80/ ls
perf stat -e mac_iod0_mac0_ch0/ea-mac/ ls
perf stat -e mac_iod0_mac0_ch0/event=0x80/ ls
FUJITSU-MONAKA PMU Events Specification v1.1 URL:
https://github.com/fujitsu/FUJITSU-MONAKA
Reviewed-by: Yicong Yang <yangyicong@hisilicon.com>
Signed-off-by: Koichi Okuno <fj2767dz@fujitsu.com>
Signed-off-by: Will Deacon <will@kernel.org>
Since the referenced fixes commit, the kernel's .text section is only
mapped starting from _stext; the region [_text, _stext) is omitted. As a
result, other vmalloc/vmap allocations may use the virtual addresses
nominally in the range [_text, _stext). This address reuse confuses
multiple things:
1. crash_prepare_elf64_headers() sets up a segment in /proc/vmcore
mapping the entire range [_text, _end) to
[__pa_symbol(_text), __pa_symbol(_end)). Reading an address in
[_text, _stext) from /proc/vmcore therefore gives the incorrect
result.
2. Tools doing symbolization (either by reading /proc/kallsyms or based
on the vmlinux ELF file) will incorrectly identify vmalloc/vmap
allocations in [_text, _stext) as kernel symbols.
In practice, both of these issues affect the drgn debugger.
Specifically, there were cases where the vmap IRQ stacks for some CPUs
were allocated in [_text, _stext). As a result, drgn could not get the
stack trace for a crash in an IRQ handler because the core dump
contained invalid data for the IRQ stack address. The stack addresses
were also symbolized as being in the _text symbol.
Fix this by bringing back the mapping of [_text, _stext), but now make
it non-executable and read-only. This prevents other allocations from
using it while still achieving the original goal of not mapping
unpredictable data as executable. Other than the changed protection,
this is effectively a revert of the fixes commit.
Fixes: e2a073dde9 ("arm64: omit [_text, _stext) from permanent kernel mapping")
Cc: stable@vger.kernel.org
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Will Deacon <will@kernel.org>
Except 'xchk_setup_metapath_rtginode()' case, 'path' argument of
'xchk_setup_metapath_scan()' is a compile-time constant. So it may
be reasonable to use 'kstrdup_const()' / 'kree_const()' to manage
'path' field of 'struct xchk_metapath' in attempt to reuse .rodata
instance rather than making a copy. Compile tested only.
Signed-off-by: Dmitry Antipov <dmantipov@yandex.ru>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Carlos Maiolino <cem@kernel.org>
Only ranges inside the file system can be translated, and the file system
can be smaller than the containing device.
Fixes: f4ed930379 ("xfs: don't shut down the filesystem for media failures beyond end of log")
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Carlos Maiolino <cem@kernel.org>
Add a bt_nr_sectors to track the number of sector in each buftarg, and
replace the check that hard codes sb_dblock in xfs_buf_map_verify with
this new value so that it is correct for non-ddev buftargs. The
RT buftarg only has a superblock in the first block, so it is unlikely
to trigger this, or are we likely to ever have enough blocks in the
in-memory buftargs, but we might as well get the check right.
Fixes: 10616b806d ("xfs: fix _xfs_buf_find oops on blocks beyond the filesystem end")
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Carlos Maiolino <cem@kernel.org>
Our goal is to move towards enabling vmalloc-huge by default on arm64 so
as to reduce TLB pressure. Therefore, we need a way to analyze the portion
of block mappings in vmalloc space we can get on a production system; this
can be done through ptdump, but currently we disable vmalloc-huge if
CONFIG_PTDUMP_DEBUGFS is on. The reason is that lazy freeing of kernel
pagetables via vmap_try_huge_pxd() may race with ptdump, so ptdump
may dereference a bogus address.
To solve this, we need to synchronize ptdump_walk() and ptdump_check_wx()
with pud_free_pmd_page() and pmd_free_pte_page().
Since this race is very unlikely to happen in practice, we do not want to
penalize the vmalloc pagetable tearing path by taking the init_mm
mmap_lock. Therefore, we use static keys. ptdump_walk() and
ptdump_check_wx() are the pagetable walkers; they will enable the static
key - upon observing that, the vmalloc pagetable tearing path will get
patched in with an mmap_read_lock/unlock sequence. A combination of the
patched-in mmap_read_lock/unlock, the acquire semantics of
static_branch_inc(), and the barriers in __flush_tlb_kernel_pgtable()
ensures that ptdump will never get a hold on the address of a freed PMD
or PTE table.
We can verify the correctness of the algorithm via the following litmus
test (thanks to James Houghton and Will Deacon):
AArch64 ptdump
Variant=Ifetch
{
uint64_t pud=0xa110c;
uint64_t pmd;
0:X0=label:"P1:L0"; 0:X1=instr:"NOP"; 0:X2=lock; 0:X3=pud; 0:X4=pmd;
1:X1=0xdead; 1:X2=lock; 1:X3=pud; 1:X4=pmd;
}
P0 | P1 ;
(* static_key_enable *) | (* pud_free_pmd_page *) ;
STR W1, [X0] | LDR X9, [X3] ;
DC CVAU,X0 | STR XZR, [X3] ;
DSB ISH | DSB ISH ;
IC IVAU,X0 | ISB ;
DSB ISH | ;
ISB | (* static key *) ;
| L0: ;
(* mmap_lock *) | B out1 ;
Lwlock: | ;
MOV W7, #1 | (* mmap_lock *) ;
SWPA W7, W8, [X2] | Lrlock: ;
| MOV W7, #1 ;
| SWPA W7, W8, [X2] ;
(* walk pgtable *) | ;
LDR X9, [X3] | (* mmap_unlock *) ;
CBZ X9, out0 | STLR WZR, [X2] ;
EOR X10, X9, X9 | ;
LDR X11, [X4, X10] | out1: ;
| EOR X10, X9, X9 ;
out0: | STR X1, [X4, X10] ;
exists (0:X8=0 /\ 1:X8=0 /\ (* Lock acquisitions succeed *)
0:X9=0xa110c /\ (* P0 sees the valid PUD ...*)
0:X11=0xdead) (* ... but the freed PMD *)
For an approximate written proof of why this algorithm works, please read
the code comment in [1], which is now removed for the sake of simplicity.
mm-selftests pass. No issues were observed while parallelly running
test_vmalloc.sh (which stresses the vmalloc subsystem),
and cat /sys/kernel/debug/{kernel_page_tables, check_wx_pages} in a loop.
Link: https://lore.kernel.org/all/20250723161827.15802-1-dev.jain@arm.com/ [1]
Reviewed-by: Ryan Roberts <ryan.roberts@arm.com>
Signed-off-by: Dev Jain <dev.jain@arm.com>
Signed-off-by: Will Deacon <will@kernel.org>
Neoverse-V3AE advertises support for BBML2 and is known to not raise
conflict aborts. So add it to the BBML2_NOABORT allow list.
However, just like Neoverse-V3, Neoverse-V3AE r0p0 and r0p1 suffer from
erratum #3053180, for which the workaround is to always observe
break-before-make requirements for affected revisions. Therefore only
add to the allow list from r0p2 onwards.
For more details see Software Developer Errata Notice (SDEN) document:
Neoverse V3AE (MP172) SDEN v9.0, erratum 3053180
https://developer.arm.com/documentation/SDEN-2615521/9-0/
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Signed-off-by: Will Deacon <will@kernel.org>
Handle the case when the aux buffer is going to be full and
data needs to be written to the data file. perf_aux_output_begin()
function checks if there is enough space depending on the values of
aux_wakeup and aux_watermark which is part of "struct perf_buffer".
Inorder to maintain where to write to aux buffer, add two fields
to "struct vpa_pmu_buf". Field "threshold" to indicate total possible
DTL entries that can be contained in aux buffer and field "full" to
indicate anytime when buffer is full. In perf_aux_output_end, there
is check to see if wake up is needed based on aux head value.
In vpa_dtl_capture_aux(), check if there is enough space to contain the
DTL data. If not, save the data for available memory and set full to true.
Set head of private aux to zero when buffer is full so that next data
will be copied to beginning of the buffer. The address used for copying
to aux is "aux_copy_buf + buf->head". So once buffer is full, set head
to zero, so that next time it will be written from start of the buffer.
Signed-off-by: Athira Rajeev <atrajeev@linux.ibm.com>
Tested-by: Tejas Manhas <tejas05@linux.ibm.com>
Tested-by: Venkat Rao Bagalkote <venkat88@linux.ibm.com>
Signed-off-by: Madhavan Srinivasan <maddy@linux.ibm.com>
Link: https://patch.msgid.link/20250915102947.26681-7-atrajeev@linux.ibm.com
vpa dtl pmu has one hrtimer added per vpa-dtl pmu thread. When the
hrtimer expires, in the timer handler, code is added to save the DTL
data to perf event record via vpa_dtl_capture_aux() function.
The DTL (Dispatch Trace Log) contains information
about dispatch/preempt, enqueue time etc. We directly copy the DTL
buffer data as part of auxiliary buffer. Data will be written to
disk only when the allocated buffer is full.
By this approach, all the DTL data will be present as-is in the
perf.data. The data will be post-processed in perf tools side when doing
perf report/perf script and this will avoid time taken to create samples
in the kernel space.
To corelate each DTL entry with other events across CPU's, we need to
map timebase from "struct dtl_entry" which phyp provides with boot
timebase. This also needs timebase frequency. Define "struct boottb_freq"
to save these details.
Added changes to capture the Dispatch Trace Log details to AUX buffer
in vpa_dtl_dump_sample_data(). Boot timebase and frequency needs to be
saved only at once, added field to indicate this as part of
"vpa_pmu_buf" structure.
perf_aux_output_begin: This function is called before writing to AUX
area. This returns the pointer to aux area private structure, ie
"struct vpa_pmu_buf". The function obtains the output handle
(used in perf_aux_output_end). when capture completes in
vpa_dtl_capture_aux(), call perf_aux_output_end() to commit the recorded
data. perf_aux_output_end() is called to move the aux->head of
"struct perf_buffer" to indicate size of data in aux buffer.
aux_tail will be moved in perf tools side when writing the data from
aux buffer to perf.data file in disk.
It is responsiblity of PMU driver to make sure data is copied between
perf_aux_output_begin and perf_aux_output_end.
Signed-off-by: Athira Rajeev <atrajeev@linux.ibm.com>
Tested-by: Tejas Manhas <tejas05@linux.ibm.com>
Tested-by: Venkat Rao Bagalkote <venkat88@linux.ibm.com>
Signed-off-by: Madhavan Srinivasan <maddy@linux.ibm.com>
Link: https://patch.msgid.link/20250915102947.26681-6-atrajeev@linux.ibm.com
vpa dtl pmu has one hrtimer added per vpa-dtl pmu thread. When the
hrtimer expires, in the timer handler, code is added to save the DTL
data to perf event record. DTL (Dispatch Trace Log) contains information
about dispatch/preempt, enqueue time etc. We directly copy the DTL
buffer data as part of auxiliary buffer and it will be postprocessed
later. To enable the support for aux buffer, add the PMU callbacks for
setup_aux and free_aux.
In setup_aux, set up pmu-private data structures for an AUX
area. rb_alloc_aux uses "alloc_pages_node" and returns pointer to each
page address. Map these pages to contiguous space using vmap and use
that as base address. The aux private data structure ie,
"struct vpa_pmu_buf" mainly saves:
1. buf->base: aux buffer base address
2. buf->head: offset from base address where data will be written to.
3. buf->size: Size of allocated memory
free_aux will free pmu-private AUX data structures.
Signed-off-by: Athira Rajeev <atrajeev@linux.ibm.com>
Tested-by: Tejas Manhas <tejas05@linux.ibm.com>
Tested-by: Venkat Rao Bagalkote <venkat88@linux.ibm.com>
Signed-off-by: Madhavan Srinivasan <maddy@linux.ibm.com>
Link: https://patch.msgid.link/20250915102947.26681-5-atrajeev@linux.ibm.com
The pseries Shared Processor Logical Partition(SPLPAR) machines
can retrieve a log of dispatch and preempt events from the
hypervisor using data from Disptach Trace Log(DTL) buffer.
With this information, user can retrieve when and why each dispatch &
preempt has occurred. Added an interface to expose the Virtual Processor
Area(VPA) DTL counters via perf.
The following events are available and exposed in sysfs:
vpa_dtl/dtl_cede/ - Trace voluntary (OS initiated) virtual processor waits
vpa_dtl/dtl_preempt/ - Trace time slice preempts
vpa_dtl/dtl_fault/ - Trace virtual partition memory page faults.
vpa_dtl/dtl_all/ - Trace all (dtl_cede/dtl_preempt/dtl_fault)
Added interface defines supported event list, config fields for the
event attributes and their corresponding bit values which are exported
via sysfs. User could use the standard perf tool to access perf events
exposed via vpa-dtl pmu.
The VPA DTL PMU counters do not interrupt on overflow or generate any
PMI interrupts. Therefore, the kernel needs to poll the counters, added
hrtimer code to do that. The timer interval can be provided by user via
sample_period field in nano seconds. There is one hrtimer added per
vpa-dtl pmu thread.
To ensure there are no other conflicting dtl users (example: debugfs dtl
or /proc/powerpc/vcpudispatch_stats), interface added code to use
"down_write_trylock" call to take the dtl_access_lock. The dtl_access_lock
is defined in dtl.h file. Also added global reference count variable called
"dtl_global_refc", to ensure dtl data can be captured per-cpu. Code also
added global lock called "dtl_global_lock" to avoid race condition.
Signed-off-by: Kajol Jain <kjain@linux.ibm.com>
Tested-by: Tejas Manhas <tejas05@linux.ibm.com>
Tested-by: Venkat Rao Bagalkote <venkat88@linux.ibm.com>
Signed-off-by: Madhavan Srinivasan <maddy@linux.ibm.com>
Link: https://patch.msgid.link/20250915102947.26681-3-atrajeev@linux.ibm.com
The "apm,xgene-ahci-v2" compatible has been in use for a long time, but
was undocumented. It doesn't require clocks or phys.
Remove the "apm,xgene-ahci-pcie" compatible which isn't used anywhere
while we're here.
Signed-off-by: Rob Herring (Arm) <robh@kernel.org>
Signed-off-by: Damien Le Moal <dlemoal@kernel.org>
Currently the error handling of run_delalloc_nocow() modifies
@cur_offset to handle different parts of the delalloc range.
However the error handling can always be split into 3 parts:
1) The range with ordered extents allocated (OE cleanup)
We have to cleanup the ordered extents and unlock the folios.
2) The range that have been cleaned up (skip)
For now it's only for the range of fallback_to_cow().
We should not touch it at all.
3) The range that is not yet touched (untouched)
We have to unlock the folios and clear any reserved space.
This 3 ranges split has the same principle as cow_file_range(), however
the NOCOW/COW handling makes the above 3 range split much more complex:
a) Failure immediately after a successful OE allocation
Thus no @cow_start nor @cow_end set.
start cur_offset end
| OE cleanup | untouched |
b) Failure after hitting a COW range but before calling
fallback_to_cow()
start cow_start cur_offset end
| OE Cleanup | untouched |
c) Failure to call fallback_to_cow()
start cow_start cow_end end
| OE Cleanup | skip | untouched |
Instead of modifying @cur_offset, do proper range calculation for
OE-cleanup and untouched ranges using above 3 cases with proper range
charts.
This avoid updating @cur_offset, as it will an extra output for debug
purposes later, and explain the behavior easier.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The ref_tracker infrastructure aids debugging but is not enabled by
default as it has a performance impact. Add mount option 'ref_tracker'
so it can be selectively enabled on a filesystem. Currently it track
references of 'delayed inodes'.
Signed-off-by: Leo Martins <loemra.dev@gmail.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We are seeing soft lockups in kill_all_delayed_nodes due to a
delayed_node with a lingering reference count of 1. Printing at this
point will reveal the guilty stack trace. If the delayed_node has no
references there should be no output.
Signed-off-by: Leo Martins <loemra.dev@gmail.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Add ref_tracker infrastructure for struct btrfs_delayed_node.
It is a response to the largest btrfs related crash in our fleet. We're
seeing soft lockups in btrfs_kill_all_delayed_nodes() that seem to be a
result of delayed_nodes not being released properly.
A ref_tracker object is allocated on reference count increases and freed
on reference count decreases. The ref_tracker object stores a stack
trace of where it is allocated. The ref_tracker_dir object is embedded
in btrfs_delayed_node and keeps track of all current and some old/freed
ref_tracker objects. When a leak is detected we can print the stack
traces for all ref_trackers that have not yet been freed.
Here is a common example of taking a reference to a delayed_node and
freeing it with ref_tracker.
struct btrfs_ref_tracker tracker;
struct btrfs_delayed_node *node;
node = btrfs_get_delayed_node(inode, &tracker);
// use delayed_node...
btrfs_release_delayed_node(node, &tracker);
There are two special cases where the delayed_node reference is "long
lived", meaning that the thread that takes the reference and the thread
that releases the reference are different. The 'inode_cache_tracker'
tracks the delayed_node stored in btrfs_inode. The 'node_list_tracker'
tracks the delayed_node stored in the btrfs_delayed_root
node_list/prepare_list. These trackers are embedded in the
btrfs_delayed_node.
btrfs_ref_tracker and btrfs_ref_tracker_dir are wrappers that either
compile to the corresponding ref_tracker structs or empty structs
depending on CONFIG_BTRFS_DEBUG. There are also btrfs wrappers for
the ref_tracker API.
Signed-off-by: Leo Martins <loemra.dev@gmail.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We're almost done cleaning misused int/bool parameters. Convert a bunch
of them, found by manual grepping. Note that btrfs_sync_fs() needs an
int as it's mandated by the struct super_operations prototype.
Reviewed-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>
Remove CONFIG_BTRFS_FS_REF_VERIFY Kconfig and add it as part of
CONFIG_BTRFS_DEBUG. This should not be impactful to the performance
of debug. The struct btrfs_ref takes an additional u64, btrfs_fs_info
takes an additional spinlock_t and rb_root. All of the ref_verify logic
is still protected by a mount option.
Signed-off-by: Leo Martins <loemra.dev@gmail.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Currently we manually check the block size against 3 different values:
- 4K
- PAGE_SIZE
- MIN_BLOCKSIZE
Those 3 values can match or differ from each other. This makes it
pretty complex to output the supported block sizes.
Considering we're going to add block size > page size support soon, this
can make the support block size sysfs attribute much harder to
implement.
To make it easier, factor out a helper, btrfs_supported_blocksize() to
do a simple check for the block size.
Then utilize it in the two locations:
- btrfs_validate_super()
This is very straightforward
- supported_sectorsizes_show()
Iterate through all valid block sizes, and only output supported ones.
This is to make future full range block sizes support much easier.
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
[BEHAVIOR DIFFERENCE BETWEEN COMPRESSION ALGOS]
Currently LZO compression algorithm will check if we're making the
compressed data larger after compressing more than 2 blocks.
But zlib and zstd do the same checks after compressing more than 8192
bytes.
This is not a big deal, but since we're already supporting larger block
size (e.g. 64K block size if page size is also 64K), this check is not
suitable for all block sizes.
For example, if our page and block size are both 16KiB, and after the
first block compressed using zlib, the resulted compressed data is
slightly larger than 16KiB, we will immediately abort the compression.
This makes zstd and zlib compression algorithms to behave slightly
different from LZO, which only aborts after compressing two blocks.
[ENHANCEMENT]
To unify the behavior, only abort the compression after compressing at
least two blocks.
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
For the 3 supported compression algorithms, two of them (zstd and zlib)
are already grabbing the btrfs inode for error messages.
It's more common to pass btrfs_inode and grab the address space from it.
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The hint block group selection in the extent allocator is wrong in the
first place, as it can select the dedicated data relocation block group for
the normal data allocation.
Since we separated the normal data space_info and the data relocation
space_info, we can easily identify a block group is for data relocation or
not. Do not choose it for the normal data allocation.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com>
Signed-off-by: David Sterba <dsterba@suse.com>
If you run a workload with:
- a cgroup that does tons of parallel data reading, with a working set
much larger than its memory limit
- a second cgroup that writes relatively fewer files, with overwrites,
with no memory limit
(see full code listing at the bottom for a reproducer)
Then what quickly occurs is:
- we have a large number of threads trying to read the csum tree
- we have a decent number of threads deleting csums running delayed refs
- we have a large number of threads in direct reclaim and thus high
memory pressure
The result of this is that we writeback the csum tree repeatedly mid
transaction, to get back the extent_buffer folios for reclaim. As a
result, we repeatedly COW the csum tree for the delayed refs that are
deleting csums. This means repeatedly write locking the higher levels of
the tree.
As a result of this, we achieve an unpleasant priority inversion. We
have:
- a high degree of contention on the csum root node (and other upper
nodes) eb rwsem
- a memory starved cgroup doing tons of reclaim on CPU.
- many reader threads in the memory starved cgroup "holding" the sem
as readers, but not scheduling promptly. i.e., task __state == 0, but
not running on a cpu.
- btrfs_commit_transaction stuck trying to acquire the sem as a writer.
(running delayed_refs, deleting csums for unreferenced data extents)
This results in arbitrarily long transactions. This then results in
seriously degraded performance for any cgroup using the filesystem (the
victim cgroup in the script).
It isn't an academic problem, as we see this exact problem in production
at Meta with one cgroup over its memory limit ruining btrfs performance
for the whole system, stalling critical system services that depend on
btrfs syncs.
The underlying scheduling "problem" with global rwsems is sort of thorny
and apparently well known and was discussed at LPC 2024, for example.
As a result, our main lever in the short term is just trying to reduce
contention on our various rwsems with an eye to reducing the frequency
of write locking, to avoid disabling the read lock fast acquisition path.
Luckily, it seems likely that many reads are for old extents written
many transactions ago, and that for those we *can* in fact search the
commit root. The commit_root_sem only gets taken write once, near the
end of transaction commit, no matter how much memory pressure there is,
so we have much less contention between readers and writers.
This change detects when we are trying to read an old extent (according
to extent map generation) and then wires that through bio_ctrl to the
btrfs_bio, which unfortunately isn't allocated yet when we have this
information. When we go to lookup the csums in lookup_bio_sums we can
check this condition on the btrfs_bio and do the commit root lookup
accordingly.
Note that a single bio_ctrl might collect a few extent_maps into a single
bio, so it is important to track a maximum generation across all the
extent_maps used for each bio to make an accurate decision on whether it
is valid to look in the commit root. If any extent_map is updated in the
current generation, we can't use the commit root.
To test and reproduce this issue, I used the following script and
accompanying C program (to avoid bottlenecks in constantly forking
thousands of dd processes):
====== big-read.c ======
#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <unistd.h>
#include <errno.h>
#define BUF_SZ (128 * (1 << 10UL))
int read_once(int fd, size_t sz) {
char buf[BUF_SZ];
size_t rd = 0;
int ret = 0;
while (rd < sz) {
ret = read(fd, buf, BUF_SZ);
if (ret < 0) {
if (errno == EINTR)
continue;
fprintf(stderr, "read failed: %d\n", errno);
return -errno;
} else if (ret == 0) {
break;
} else {
rd += ret;
}
}
return rd;
}
int read_loop(char *fname) {
int fd;
struct stat st;
size_t sz = 0;
int ret;
while (1) {
fd = open(fname, O_RDONLY);
if (fd == -1) {
perror("open");
return 1;
}
if (!sz) {
if (!fstat(fd, &st)) {
sz = st.st_size;
} else {
perror("stat");
return 1;
}
}
ret = read_once(fd, sz);
close(fd);
}
}
int main(int argc, char *argv[]) {
int fd;
struct stat st;
off_t sz;
char *buf;
int ret;
if (argc != 2) {
fprintf(stderr, "Usage: %s <filename>\n", argv[0]);
return 1;
}
return read_loop(argv[1]);
}
====== repro.sh ======
#!/usr/bin/env bash
SCRIPT=$(readlink -f "$0")
DIR=$(dirname "$SCRIPT")
dev=$1
mnt=$2
shift
shift
CG_ROOT=/sys/fs/cgroup
BAD_CG=$CG_ROOT/bad-nbr
GOOD_CG=$CG_ROOT/good-nbr
NR_BIGGOS=1
NR_LITTLE=10
NR_VICTIMS=32
NR_VILLAINS=512
START_SEC=$(date +%s)
_elapsed() {
echo "elapsed: $(($(date +%s) - $START_SEC))"
}
_stats() {
local sysfs=/sys/fs/btrfs/$(findmnt -no UUID $dev)
echo "================"
date
_elapsed
cat $sysfs/commit_stats
cat $BAD_CG/memory.pressure
}
_setup_cgs() {
echo "+memory +cpuset" > $CG_ROOT/cgroup.subtree_control
mkdir -p $GOOD_CG
mkdir -p $BAD_CG
echo max > $BAD_CG/memory.max
# memory.high much less than the working set will cause heavy reclaim
echo $((1 << 30)) > $BAD_CG/memory.high
# victims get a subset of villain CPUs
echo 0 > $GOOD_CG/cpuset.cpus
echo 0,1,2,3 > $BAD_CG/cpuset.cpus
}
_kill_cg() {
local cg=$1
local attempts=0
echo "kill cgroup $cg"
[ -f $cg/cgroup.procs ] || return
while true; do
attempts=$((attempts + 1))
echo 1 > $cg/cgroup.kill
sleep 1
procs=$(wc -l $cg/cgroup.procs | cut -d' ' -f1)
[ $procs -eq 0 ] && break
done
rmdir $cg
echo "killed cgroup $cg in $attempts attempts"
}
_biggo_vol() {
echo $mnt/biggo_vol.$1
}
_biggo_file() {
echo $(_biggo_vol $1)/biggo
}
_subvoled_biggos() {
total_sz=$((10 << 30))
per_sz=$((total_sz / $NR_VILLAINS))
dd_count=$((per_sz >> 20))
echo "create $NR_VILLAINS subvols with a file of size $per_sz bytes for a total of $total_sz bytes."
for i in $(seq $NR_VILLAINS)
do
btrfs subvol create $(_biggo_vol $i) &>/dev/null
dd if=/dev/zero of=$(_biggo_file $i) bs=1M count=$dd_count &>/dev/null
done
echo "done creating subvols."
}
_setup() {
[ -f .done ] && rm .done
findmnt -n $dev && exit 1
if [ -f .re-mkfs ]; then
mkfs.btrfs -f -m single -d single $dev >/dev/null || exit 2
else
echo "touch .re-mkfs to populate the test fs"
fi
mount -o noatime $dev $mnt || exit 3
[ -f .re-mkfs ] && _subvoled_biggos
_setup_cgs
}
_my_cleanup() {
echo "CLEANUP!"
_kill_cg $BAD_CG
_kill_cg $GOOD_CG
sleep 1
umount $mnt
}
_bad_exit() {
_err "Unexpected Exit! $?"
_stats
exit $?
}
trap _my_cleanup EXIT
trap _bad_exit INT TERM
_setup
# Use a lot of page cache reading the big file
_villain() {
local i=$1
echo $BASHPID > $BAD_CG/cgroup.procs
$DIR/big-read $(_biggo_file $i)
}
# Hit del_csum a lot by overwriting lots of small new files
_victim() {
echo $BASHPID > $GOOD_CG/cgroup.procs
i=0;
while (true)
do
local tmp=$mnt/tmp.$i
dd if=/dev/zero of=$tmp bs=4k count=2 >/dev/null 2>&1
i=$((i+1))
[ $i -eq $NR_LITTLE ] && i=0
done
}
_one_sync() {
echo "sync..."
before=$(date +%s)
sync
after=$(date +%s)
echo "sync done in $((after - before))s"
_stats
}
# sync in a loop
_sync() {
echo "start sync loop"
syncs=0
echo $BASHPID > $GOOD_CG/cgroup.procs
while true
do
[ -f .done ] && break
_one_sync
syncs=$((syncs + 1))
[ -f .done ] && break
sleep 10
done
if [ $syncs -eq 0 ]; then
echo "do at least one sync!"
_one_sync
fi
echo "sync loop done."
}
_sleep() {
local time=${1-60}
local now=$(date +%s)
local end=$((now + time))
while [ $now -lt $end ];
do
echo "SLEEP: $((end - now))s left. Sleep 10."
sleep 10
now=$(date +%s)
done
}
echo "start $NR_VILLAINS villains"
for i in $(seq $NR_VILLAINS)
do
_villain $i &
disown # get rid of annoying log on kill (done via cgroup anyway)
done
echo "start $NR_VICTIMS victims"
for i in $(seq $NR_VICTIMS)
do
_victim &
disown
done
_sync &
SYNC_PID=$!
_sleep $1
_elapsed
touch .done
wait $SYNC_PID
echo "OK"
exit 0
Without this patch, that reproducer:
- Ran for 6+ minutes instead of 60s
- Hung hundreds of threads in D state on the csum reader lock
- Got a commit stuck for 3 minutes
sync done in 388s
================
Wed Jul 9 09:52:31 PM UTC 2025
elapsed: 420
commits 2
cur_commit_ms 0
last_commit_ms 159446
max_commit_ms 159446
total_commit_ms 160058
some avg10=99.03 avg60=98.97 avg300=75.43 total=418033386
full avg10=82.79 avg60=80.52 avg300=59.45 total=324995274
419 hits state R, D comms big-read
btrfs_tree_read_lock_nested
btrfs_read_lock_root_node
btrfs_search_slot
btrfs_lookup_csum
btrfs_lookup_bio_sums
btrfs_submit_bbio
1 hits state D comms btrfs-transacti
btrfs_tree_lock_nested
btrfs_lock_root_node
btrfs_search_slot
btrfs_del_csums
__btrfs_run_delayed_refs
btrfs_run_delayed_refs
With the patch, the reproducer exits naturally, in 65s, completing a
pretty decent 4 commits, despite heavy memory pressure. Occasionally you
can still trigger a rather long commit (couple seconds) but never one
that is minutes long.
sync done in 3s
================
elapsed: 65
commits 4
cur_commit_ms 0
last_commit_ms 485
max_commit_ms 689
total_commit_ms 2453
some avg10=98.28 avg60=64.54 avg300=19.39 total=64849893
full avg10=74.43 avg60=48.50 avg300=14.53 total=48665168
some random rwalker samples showed the most common stack in reclaim,
rather than the csum tree:
145 hits state R comms bash, sleep, dd, shuf
shrink_folio_list
shrink_lruvec
shrink_node
do_try_to_free_pages
try_to_free_mem_cgroup_pages
reclaim_high
Link: https://lpc.events/event/18/contributions/1883/
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Boris Burkov <boris@bur.io>
Signed-off-by: David Sterba <dsterba@suse.com>
Now that btrfs_zone_finish_endio_workfn() is directly calling
do_zone_finish() the only caller of btrfs_zone_finish_endio() is
btrfs_finish_one_ordered().
btrfs_finish_one_ordered() already has error handling in-place so
btrfs_zone_finish_endio() can return an error if the block group lookup
fails.
Also as btrfs_zone_finish_endio() already checks for zoned filesystems and
returns early, there's no need to do this in the caller.
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Signed-off-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When btrfs_zone_finish_endio_workfn() is calling btrfs_zone_finish_endio()
it already has a pointer to the block group. Furthermore
btrfs_zone_finish_endio() does additional checks if the block group can be
finished or not.
But in the context of btrfs_zone_finish_endio_workfn() only the actual
call to do_zone_finish() is of interest, as the skipping condition when
there is still room to allocate from the block group cannot be checked.
Directly call do_zone_finish() on the block group.
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Signed-off-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: David Sterba <dsterba@suse.com>
There's one only one caller of unaccount_log_buffer() and both this
function and the caller are short, so move its code into the caller.
Reviewed-by: Boris Burkov <boris@bur.io>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Instead of extracting again the disk_bytenr and disk_num_bytes values from
the file extent item to pass to btrfs_qgroup_trace_extent(), use the key
local variable 'ins' which already has those values, reducing the size of
the source code.
Reviewed-by: Boris Burkov <boris@bur.io>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Instead of having an if statement to check for regular and prealloc
extents first, process them in a block, and then following with an else
statement to check for an inline extent, check for an inline extent first,
process it and jump to the 'update_inode' label, allowing us to avoid
having the code for processing regular and prealloc extents inside a
block, reducing the high indentation level by one and making the code
easier to read and avoid line splittings too.
Reviewed-by: Boris Burkov <boris@bur.io>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
At replay_one_extent(), we can jump to the code that updates the file
extent range and updates the inode when processing a file extent item that
represents a hole and we don't have the NO_HOLES feature enabled. This
helps reduce the high indentation level by one in replay_one_extent() and
avoid splitting some lines to make the code easier to read.
Reviewed-by: Boris Burkov <boris@bur.io>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
In the replay_one_buffer() log tree walk callback we return errors to the
log tree walk caller and then the caller aborts the transaction, if we
have one, or turns the fs into error state if we don't have one. While
this reduces code it makes it harder to figure out where exactly an error
came from. So add the transaction aborts after every failure inside the
replay_one_buffer() callback and the functions it calls, making it as
fine grained as possible, so that it helps figuring out why failures
happen.
Reviewed-by: Boris Burkov <boris@bur.io>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
If read_alloc_one_name() we explicitly return -ENOMEM and currently that
is fine since it's the only error read_alloc_one_name() can return for
now. However this is fragile and not future proof, so return instead what
read_alloc_one_name() returned.
Reviewed-by: Boris Burkov <boris@bur.io>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Instead of keep dereferencing the walk_control structure to extract the
transaction handle whenever is needed, do it once by storing it in a local
variable and then use the variable everywhere. This reduces code verbosity
and eliminates the need for some split lines.
Reviewed-by: Boris Burkov <boris@bur.io>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
In the process_one_buffer() log tree walk callback we return errors to the
log tree walk caller and then the caller aborts the transaction, if we
have one, or turns the fs into error state if we don't have one. While
this reduces code it makes it harder to figure out where exactly an error
came from. So add the transaction aborts after every failure inside the
process_one_buffer() callback, so that it helps figuring out why failures
happen.
Reviewed-by: Boris Burkov <boris@bur.io>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We do several things while walking a log tree (for replaying and for
freeing a log tree) like reading extent buffers and cleaning them up,
but we don't immediately abort the transaction, or turn the fs into an
error state, when one of these things fails. Instead we the transaction
abort or turn the fs into error state in the caller of the entry point
function that walks a log tree - walk_log_tree() - which means we don't
get to know exactly where an error came from.
Improve on this by doing a transaction abort / turn fs into error state
after each such failure so that when it happens we have a better
understanding where the failure comes from. This deliberately leaves
the transaction abort / turn fs into error state in the callers of
walk_log_tree() as to ensure we don't get into an inconsistent state in
case we forget to do it deeper in call chain. It also deliberately does
not do it after errors from the calls to the callback defined in
struct walk_control::process_func(), as we will do it later on another
patch.
Reviewed-by: Boris Burkov <boris@bur.io>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Yicong Yang
Koichi Okuno
Niranjan H Y
Omar Sandoval
Dmitry Antipov
Anshuman Khandual
Christoph Hellwig
Dev Jain
Ryan Roberts
Mark Rutland
Athira Rajeev
Kajol Jain
Aboorva Devarajan
Rob Herring (Arm)
Qu Wenruo
Leo Martins
David Sterba
Xichao Zhao
Naohiro Aota
Boris Burkov
Jiapeng Chong
Johannes Thumshirn
Filipe Manana