mirror of
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git
synced 2025-12-22 12:17:45 +01:00
master
34 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Gabriele Monaco
|
7dec062cfc |
timers/migration: Exclude isolated cpus from hierarchy
The timer migration mechanism allows active CPUs to pull timers from idle ones to improve the overall idle time. This is however undesired when CPU intensive workloads run on isolated cores, as the algorithm would move the timers from housekeeping to isolated cores, negatively affecting the isolation. Exclude isolated cores from the timer migration algorithm, extend the concept of unavailable cores, currently used for offline ones, to isolated ones: * A core is unavailable if isolated or offline; * A core is available if non isolated and online; A core is considered unavailable as isolated if it belongs to: * the isolcpus (domain) list * an isolated cpuset Except if it is: * in the nohz_full list (already idle for the hierarchy) * the nohz timekeeper core (must be available to handle global timers) CPUs are added to the hierarchy during late boot, excluding isolated ones, the hierarchy is also adapted when the cpuset isolation changes. Due to how the timer migration algorithm works, any CPU part of the hierarchy can have their global timers pulled by remote CPUs and have to pull remote timers, only skipping pulling remote timers would break the logic. For this reason, prevent isolated CPUs from pulling remote global timers, but also the other way around: any global timer started on an isolated CPU will run there. This does not break the concept of isolation (global timers don't come from outside the CPU) and, if considered inappropriate, can usually be mitigated with other isolation techniques (e.g. IRQ pinning). This effect was noticed on a 128 cores machine running oslat on the isolated cores (1-31,33-63,65-95,97-127). The tool monopolises CPUs, and the CPU with lowest count in a timer migration hierarchy (here 1 and 65) appears as always active and continuously pulls global timers, from the housekeeping CPUs. This ends up moving driver work (e.g. delayed work) to isolated CPUs and causes latency spikes: before the change: # oslat -c 1-31,33-63,65-95,97-127 -D 62s ... Maximum: 1203 10 3 4 ... 5 (us) after the change: # oslat -c 1-31,33-63,65-95,97-127 -D 62s ... Maximum: 10 4 3 4 3 ... 5 (us) The same behaviour was observed on a machine with as few as 20 cores / 40 threads with isocpus set to: 1-9,11-39 with rtla-osnoise-top. Signed-off-by: Gabriele Monaco <gmonaco@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: John B. Wyatt IV <jwyatt@redhat.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Link: https://patch.msgid.link/20251120145653.296659-8-gmonaco@redhat.com |
||
|
Gabriele Monaco
|
4c2374ed86 |
timers/migration: Use scoped_guard on available flag set/clear
Cleanup tmigr_clear_cpu_available() and tmigr_set_cpu_available() to prepare for easier checks on the available flag. Signed-off-by: Gabriele Monaco <gmonaco@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Link: https://patch.msgid.link/20251120145653.296659-4-gmonaco@redhat.com |
||
|
Gabriele Monaco
|
a048ca5f00 |
timers/migration: Add mask for CPUs available in the hierarchy
Keep track of the CPUs available for timer migration in a cpumask. This prepares the ground to generalise the concept of unavailable CPUs. Signed-off-by: Gabriele Monaco <gmonaco@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Link: https://patch.msgid.link/20251120145653.296659-3-gmonaco@redhat.com |
||
|
Gabriele Monaco
|
8312cab5ff |
timers/migration: Rename 'online' bit to 'available'
The timer migration hierarchy excludes offline CPUs via the tmigr_is_not_available function, which is essentially checking the online bit for the CPU. Rename the online bit to available and all references in function names and tracepoint to generalise the concept of available CPUs. Signed-off-by: Gabriele Monaco <gmonaco@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Link: https://patch.msgid.link/20251120145653.296659-2-gmonaco@redhat.com |
||
Jianyun Gao
|
4518767be9 |
time: Fix a few typos in time[r] related code comments
Signed-off-by: Jianyun Gao <jianyungao89@gmail.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Link: https://patch.msgid.link/20250927093411.1509275-1-jianyungao89@gmail.com |
||
Frederic Weisbecker
|
ba14500e4b |
timers/migration: Remove dead code handling idle CPU checking for remote timers
Idle migrators don't walk the whole tree in order to find out if there are timers to migrate because they recorded the next deadline to be verified within a single check in tmigr_requires_handle_remote(). Remove the related dead code and data. Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Link: https://patch.msgid.link/20251024132536.39841-7-frederic@kernel.org |
||
|
Frederic Weisbecker
|
93643b90d6 |
timers/migration: Remove unused "cpu" parameter from tmigr_get_group()
Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Link: https://patch.msgid.link/20251024132536.39841-6-frederic@kernel.org |
||
|
Frederic Weisbecker
|
3c8eb36e2a |
timers/migration: Assert that hotplug preparing CPU is part of stable active hierarchy
The CPU doing the prepare work for a remote target must be online from the tree point of view and its hierarchy must be active, otherwise propagating its active state up to the new root branch would be either incorrect or racy. Assert those conditions with more sanity checks. Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Link: https://patch.msgid.link/20251024132536.39841-5-frederic@kernel.org |
||
|
Frederic Weisbecker
|
5eb579dfd4 |
timers/migration: Fix imbalanced NUMA trees
When a CPU from a new node boots, the old root may happen to be
connected to the new root even if their node mismatch, as depicted in
the following scenario:
1) CPU 0 boots and creates the first group for node 0.
[GRP0:0]
node 0
|
CPU 0
2) CPU 1 from node 1 boots and creates a new top that corresponds to
node 1, but it also connects the old root from node 0 to the new root
from node 1 by mistake.
[GRP1:0]
node 1
/ \
/ \
[GRP0:0] [GRP0:1]
node 0 node 1
| |
CPU 0 CPU 1
3) This eventually leads to an imbalanced tree where some node 0 CPUs
migrate node 1 timers (and vice versa) way before reaching the
crossnode groups, resulting in more frequent remote memory accesses
than expected.
[GRP2:0]
NUMA_NO_NODE
/ \
[GRP1:0] [GRP1:1]
node 1 node 0
/ \ |
/ \ [...]
[GRP0:0] [GRP0:1]
node 0 node 1
| |
CPU 0... CPU 1...
A balanced tree should only contain groups having children that belong
to the same node:
[GRP2:0]
NUMA_NO_NODE
/ \
[GRP1:0] [GRP1:0]
node 0 node 1
/ \ / \
/ \ / \
[GRP0:0] [...] [...] [GRP0:1]
node 0 node 1
| |
CPU 0... CPU 1...
In order to fix this, the hierarchy must be unfolded up to the crossnode
level as soon as a node mismatch is detected. For example the stage 2
above should lead to this layout:
[GRP2:0]
NUMA_NO_NODE
/ \
[GRP1:0] [GRP1:1]
node 0 node 1
/ \
/ \
[GRP0:0] [GRP0:1]
node 0 node 1
| |
CPU 0 CPU 1
This means that not only GRP1:0 must be created but also GRP1:1 and
GRP2:0 in order to prepare a balanced tree for next CPUs to boot.
Fixes:
|
||
|
Frederic Weisbecker
|
fa9620355d |
timers/migration: Remove locking on group connection
Initializing the tmc's group, the group's number of children and the
group's parent can all be done without locking because:
1) Reading the group's parent and its group mask is done locklessly.
2) The connections prepared for a given CPU hierarchy are visible to the
target CPU once online, thanks to the CPU hotplug enforced memory
ordering.
3) In case of a newly created upper level, the new root and its
connections and initialization are made visible by the CPU which made
the connections. When that CPUs goes idle in the future, the new link
is published by tmigr_inactive_up() through the atomic RmW on
->migr_state.
4) If CPUs were still walking up the active hierarchy, they could observe
the new root earlier. In this case the ordering is enforced by an
early initialization of the group mask and by barriers that maintain
address dependency as explained in:
|
||
|
Frederic Weisbecker
|
6c181b5667 |
timers/migration: Convert "while" loops to use "for"
Both the "do while" and "while" loops in tmigr_setup_groups() eventually mimic the behaviour of "for" loops. Simplify accordingly. Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Link: https://patch.msgid.link/20251024132536.39841-2-frederic@kernel.org |
||
Petr Tesarik
|
ff56a3e2a8 |
timers/migration: Clean up the loop in tmigr_quick_check()
Make the logic easier to follow:
- Remove the final return statement, which is never reached, and move the
actual walk-terminating return statement out of the do-while loop.
- Remove the else-clause to reduce indentation. If a non-lonely group is
encountered during the walk, the loop is immediately terminated with a
return statement anyway; no need for an else.
Signed-off-by: Petr Tesarik <ptesarik@suse.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Link: https://lore.kernel.org/all/20250606124818.455560-1-ptesarik@suse.com
|
||
|
Frederic Weisbecker
|
868c9037df |
timers/migration: Fix off-by-one root mis-connection
Before attaching a new root to the old root, the children counter of the new root is checked to verify that only the upcoming CPU's top group have been connected to it. However since the recently added commit |
||
Linus Torvalds
|
f200c315da |
Merge tag 'timers-core-2025-01-21' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull timer and timekeeping updates from Thomas Gleixner: - Just boring cleanups, typo and comment fixes and trivial optimizations * tag 'timers-core-2025-01-21' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: timers/migration: Simplify top level detection on group setup timers: Optimize get_timer_[this_]cpu_base() timekeeping: Remove unused ktime_get_fast_timestamps() timer/migration: Fix kernel-doc warnings for union tmigr_state tick/broadcast: Add kernel-doc for function parameters hrtimers: Update the return type of enqueue_hrtimer() clocksource/wdtest: Print time values for short udelay(1) posix-timers: Fix typo in __lock_timer() vdso: Correct typo in PAGE_SHIFT comment |
||
|
Frederic Weisbecker
|
dcf6230555 |
timers/migration: Simplify top level detection on group setup
Having a single group on a given level is enough to know this is the top level, because a root has to have at least two children, unless that root is the only group and the children are actual CPUs. Simplify the test in tmigr_setup_groups() accordingly. Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Link: https://lore.kernel.org/all/20250114231507.21672-5-frederic@kernel.org |
||
|
Frederic Weisbecker
|
922efd298b |
timers/migration: Annotate accesses to ignore flag
The group's ignore flag is: _ read under the group's lock (idle entry, remote expiry) _ turned on/off under the group's lock (idle entry, remote expiry) _ turned on locklessly on idle exit When idle entry or remote expiry clear the "ignore" flag of a group, the operation must be synchronized against other concurrent idle entry or remote expiry to make sure the related group timer is never missed. To enforce this synchronization, both "ignore" clear and read are performed under the group lock. On the contrary, whether idle entry or remote expiry manage to observe the "ignore" flag turned on by a CPU exiting idle is a matter of optimization. If that flag set is missed or cleared concurrently, the worst outcome is a migrator wasting time remotely handling a "ghost" timer. This is why the ignore flag can be set locklessly. Unfortunately, the related lockless accesses are bare and miss appropriate annotations. KCSAN rightfully complains: BUG: KCSAN: data-race in __tmigr_cpu_activate / print_report write to 0xffff88842fc28004 of 1 bytes by task 0 on cpu 0: __tmigr_cpu_activate tmigr_cpu_activate timer_clear_idle tick_nohz_restart_sched_tick tick_nohz_idle_exit do_idle cpu_startup_entry kernel_init do_initcalls clear_bss reserve_bios_regions common_startup_64 read to 0xffff88842fc28004 of 1 bytes by task 0 on cpu 1: print_report kcsan_report_known_origin kcsan_setup_watchpoint tmigr_next_groupevt tmigr_update_events tmigr_inactive_up __walk_groups+0x50/0x77 walk_groups __tmigr_cpu_deactivate tmigr_cpu_deactivate __get_next_timer_interrupt timer_base_try_to_set_idle tick_nohz_stop_tick tick_nohz_idle_stop_tick cpuidle_idle_call do_idle Although the relevant accesses could be marked as data_race(), the "ignore" flag being read several times within the same tmigr_update_events() function is confusing and error prone. Prefer reading it once in that function and make use of similar/paired accesses elsewhere with appropriate comments when necessary. Reported-by: kernel test robot <oliver.sang@intel.com> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Link: https://lore.kernel.org/all/20250114231507.21672-4-frederic@kernel.org Closes: https://lore.kernel.org/oe-lkp/202501031612.62e0c498-lkp@intel.com |
||
|
Frederic Weisbecker
|
de3ced72a7 |
timers/migration: Enforce group initialization visibility to tree walkers
Commit 2522c84db513 ("timers/migration: Fix another race between hotplug
and idle entry/exit") fixed yet another race between idle exit and CPU
hotplug up leading to a wrong "0" value migrator assigned to the top
level. However there is yet another situation that remains unhandled:
[GRP0:0]
migrator = TMIGR_NONE
active = NONE
groupmask = 1
/ \ \
0 1 2..7
idle idle idle
0) The system is fully idle.
[GRP0:0]
migrator = CPU 0
active = CPU 0
groupmask = 1
/ \ \
0 1 2..7
active idle idle
1) CPU 0 is activating. It has done the cmpxchg on the top's ->migr_state
but it hasn't yet returned to __walk_groups().
[GRP0:0]
migrator = CPU 0
active = CPU 0, CPU 1
groupmask = 1
/ \ \
0 1 2..7
active active idle
2) CPU 1 is activating. CPU 0 stays the migrator (still stuck in
__walk_groups(), delayed by #VMEXIT for example).
[GRP1:0]
migrator = TMIGR_NONE
active = NONE
groupmask = 1
/ \
[GRP0:0] [GRP0:1]
migrator = CPU 0 migrator = TMIGR_NONE
active = CPU 0, CPU1 active = NONE
groupmask = 1 groupmask = 2
/ \ \
0 1 2..7 8
active active idle !online
3) CPU 8 is preparing to boot. CPUHP_TMIGR_PREPARE is being ran by CPU 1
which has created the GRP0:1 and the new top GRP1:0 connected to GRP0:1
and GRP0:0. CPU 1 hasn't yet propagated its activation up to GRP1:0.
[GRP1:0]
migrator = GRP0:0
active = GRP0:0
groupmask = 1
/ \
[GRP0:0] [GRP0:1]
migrator = CPU 0 migrator = TMIGR_NONE
active = CPU 0, CPU1 active = NONE
groupmask = 1 groupmask = 2
/ \ \
0 1 2..7 8
active active idle !online
4) CPU 0 finally resumed after its #VMEXIT. It's in __walk_groups()
returning from tmigr_cpu_active(). The new top GRP1:0 is visible and
fetched and the pre-initialized groupmask of GRP0:0 is also visible.
As a result tmigr_active_up() is called to GRP1:0 with GRP0:0 as active
and migrator. CPU 0 is returning to __walk_groups() but suffers again
a #VMEXIT.
[GRP1:0]
migrator = GRP0:0
active = GRP0:0
groupmask = 1
/ \
[GRP0:0] [GRP0:1]
migrator = CPU 0 migrator = TMIGR_NONE
active = CPU 0, CPU1 active = NONE
groupmask = 1 groupmask = 2
/ \ \
0 1 2..7 8
active active idle !online
5) CPU 1 propagates its activation of GRP0:0 to GRP1:0. This has no
effect since CPU 0 did it already.
[GRP1:0]
migrator = GRP0:0
active = GRP0:0, GRP0:1
groupmask = 1
/ \
[GRP0:0] [GRP0:1]
migrator = CPU 0 migrator = CPU 8
active = CPU 0, CPU1 active = CPU 8
groupmask = 1 groupmask = 2
/ \ \ \
0 1 2..7 8
active active idle active
6) CPU 1 links CPU 8 to its group. CPU 8 boots and goes through
CPUHP_AP_TMIGR_ONLINE which propagates activation.
[GRP2:0]
migrator = TMIGR_NONE
active = NONE
groupmask = 1
/ \
[GRP1:0] [GRP1:1]
migrator = GRP0:0 migrator = TMIGR_NONE
active = GRP0:0, GRP0:1 active = NONE
groupmask = 1 groupmask = 2
/ \
[GRP0:0] [GRP0:1] [GRP0:2]
migrator = CPU 0 migrator = CPU 8 migrator = TMIGR_NONE
active = CPU 0, CPU1 active = CPU 8 active = NONE
groupmask = 1 groupmask = 2 groupmask = 0
/ \ \ \
0 1 2..7 8 64
active active idle active !online
7) CPU 64 is booting. CPUHP_TMIGR_PREPARE is being ran by CPU 1
which has created the GRP1:1, GRP0:2 and the new top GRP2:0 connected to
GRP1:1 and GRP1:0. CPU 1 hasn't yet propagated its activation up to
GRP2:0.
[GRP2:0]
migrator = 0 (!!!)
active = NONE
groupmask = 1
/ \
[GRP1:0] [GRP1:1]
migrator = GRP0:0 migrator = TMIGR_NONE
active = GRP0:0, GRP0:1 active = NONE
groupmask = 1 groupmask = 2
/ \
[GRP0:0] [GRP0:1] [GRP0:2]
migrator = CPU 0 migrator = CPU 8 migrator = TMIGR_NONE
active = CPU 0, CPU1 active = CPU 8 active = NONE
groupmask = 1 groupmask = 2 groupmask = 0
/ \ \ \
0 1 2..7 8 64
active active idle active !online
8) CPU 0 finally resumed after its #VMEXIT. It's in __walk_groups()
returning from tmigr_cpu_active(). The new top GRP2:0 is visible and
fetched but the pre-initialized groupmask of GRP1:0 is not because no
ordering made its initialization visible. As a result tmigr_active_up()
may be called to GRP2:0 with a "0" child's groumask. Leaving the timers
ignored for ever when the system is fully idle.
The race is highly theoretical and perhaps impossible in practice but
the groupmask of the child is not the only concern here as the whole
initialization of the child is not guaranteed to be visible to any
tree walker racing against hotplug (idle entry/exit, remote handling,
etc...). Although the current code layout seem to be resilient to such
hazards, this doesn't tell much about the future.
Fix this with enforcing address dependency between group initialization
and the write/read to the group's parent's pointer. Fortunately that
doesn't involve any barrier addition in the fast paths.
Fixes:
|
||
|
Frederic Weisbecker
|
b729cc1ec2 |
timers/migration: Fix another race between hotplug and idle entry/exit
Commit |
||
Anna-Maria Behnsen
|
f004bf9de0 |
timers/migration: Fix grammar in comment
Signed-off-by: Anna-Maria Behnsen <anna-maria@linutronix.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Link: https://lore.kernel.org/r/20240716-tmigr-fixes-v4-8-757baa7803fe@linutronix.de |
||
|
Anna-Maria Behnsen
|
2367e28e23 |
timers/migration: Spare write when nothing changed
The wakeup value is written unconditionally in tmigr_cpu_new_timer(). When there was no new next timer expiry that needs to be propagated, then the value that was read before is written. This is not required. Move the write to the place where wakeup value is changed changed. Signed-off-by: Anna-Maria Behnsen <anna-maria@linutronix.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Link: https://lore.kernel.org/r/20240716-tmigr-fixes-v4-7-757baa7803fe@linutronix.de |
||
|
Anna-Maria Behnsen
|
835a9a67f5 |
timers/migration: Rename childmask by groupmask to make naming more obvious
childmask in the group reflects the mask that is required to 'reference' this group in the parent. When reading childmask, this might be confusing, as this suggests, that this is the mask of the child of the group. Clarify this by renaming childmask in the tmigr_group and tmc_group by groupmask. Signed-off-by: Anna-Maria Behnsen <anna-maria@linutronix.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Link: https://lore.kernel.org/r/20240716-tmigr-fixes-v4-6-757baa7803fe@linutronix.de |
||
|
Anna-Maria Behnsen
|
d47be58984 |
timers/migration: Read childmask and parent pointer in a single place
Reading the childmask and parent pointer is required when propagating changes through the hierarchy. At the moment this reads are spread all over the place which makes it harder to follow. Move those reads to a single place to keep code clean. Signed-off-by: Anna-Maria Behnsen <anna-maria@linutronix.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Link: https://lore.kernel.org/r/20240716-tmigr-fixes-v4-5-757baa7803fe@linutronix.de |
||
|
Anna-Maria Behnsen
|
3ba111032b |
timers/migration: Use a single struct for hierarchy walk data
Two different structs are defined for propagating data from one to another level when walking the hierarchy. Several struct members exist in both structs which makes generalization harder. Merge those two structs into a single one and use it directly in walk_groups() and the corresponding function pointers instead of introducing pointer casting all over the place. Signed-off-by: Anna-Maria Behnsen <anna-maria@linutronix.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Link: https://lore.kernel.org/r/20240716-tmigr-fixes-v4-4-757baa7803fe@linutronix.de |
||
|
Anna-Maria Behnsen
|
9250674152 |
timers/migration: Improve tracing
Trace points of inactive and active propagation are located at the end of the related functions. The interesting information of those trace points is the updated group state. When trace points are not located directly at the place where group state changed, order of trace points in traces could be confusing. Move inactive and active propagation trace points directly after update of group state values. Signed-off-by: Anna-Maria Behnsen <anna-maria@linutronix.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Link: https://lore.kernel.org/r/20240716-tmigr-fixes-v4-3-757baa7803fe@linutronix.de |
||
|
Anna-Maria Behnsen
|
10a0e6f3d3 |
timers/migration: Move hierarchy setup into cpuhotplug prepare callback
When a CPU comes online the first time, it is possible that a new top level
group will be created. In general all propagation is done from the bottom
to top. This minimizes complexity and prevents possible races. But when a
new top level group is created, the formely top level group needs to be
connected to the new level. This is the only time, when the direction to
propagate changes is changed: the changes are propagated from top (new top
level group) to bottom (formerly top level group).
This introduces two races (see (A) and (B)) as reported by Frederic:
(A) This race happens, when marking the formely top level group as active,
but the last active CPU of the formerly top level group goes idle. Then
it's likely that formerly group is no longer active, but marked
nevertheless as active in new top level group:
[GRP0:0]
migrator = 0
active = 0
nextevt = KTIME_MAX
/ \
0 1 .. 7
active idle
0) Hierarchy has for now only 8 CPUs and CPU 0 is the only active CPU.
[GRP1:0]
migrator = TMIGR_NONE
active = NONE
nextevt = KTIME_MAX
\
[GRP0:0] [GRP0:1]
migrator = 0 migrator = TMIGR_NONE
active = 0 active = NONE
nextevt = KTIME_MAX nextevt = KTIME_MAX
/ \
0 1 .. 7 8
active idle !online
1) CPU 8 is booting and creates a new group in first level GRP0:1 and
therefore also a new top group GRP1:0. For now the setup code proceeded
only until the connected between GRP0:1 to the new top group. The
connection between CPU8 and GRP0:1 is not yet established and CPU 8 is
still !online.
[GRP1:0]
migrator = TMIGR_NONE
active = NONE
nextevt = KTIME_MAX
/ \
[GRP0:0] [GRP0:1]
migrator = 0 migrator = TMIGR_NONE
active = 0 active = NONE
nextevt = KTIME_MAX nextevt = KTIME_MAX
/ \
0 1 .. 7 8
active idle !online
2) Setup code now connects GRP0:0 to GRP1:0 and observes while in
tmigr_connect_child_parent() that GRP0:0 is not TMIGR_NONE. So it
prepares to call tmigr_active_up() on it. It hasn't done it yet.
[GRP1:0]
migrator = TMIGR_NONE
active = NONE
nextevt = KTIME_MAX
/ \
[GRP0:0] [GRP0:1]
migrator = TMIGR_NONE migrator = TMIGR_NONE
active = NONE active = NONE
nextevt = KTIME_MAX nextevt = KTIME_MAX
/ \
0 1 .. 7 8
idle idle !online
3) CPU 0 goes idle. Since GRP0:0->parent has been updated by CPU 8 with
GRP0:0->lock held, CPU 0 observes GRP1:0 after calling
tmigr_update_events() and it propagates the change to the top (no change
there and no wakeup programmed since there is no timer).
[GRP1:0]
migrator = GRP0:0
active = GRP0:0
nextevt = KTIME_MAX
/ \
[GRP0:0] [GRP0:1]
migrator = TMIGR_NONE migrator = TMIGR_NONE
active = NONE active = NONE
nextevt = KTIME_MAX nextevt = KTIME_MAX
/ \
0 1 .. 7 8
idle idle !online
4) Now the setup code finally calls tmigr_active_up() to and sets GRP0:0
active in GRP1:0
[GRP1:0]
migrator = GRP0:0
active = GRP0:0, GRP0:1
nextevt = KTIME_MAX
/ \
[GRP0:0] [GRP0:1]
migrator = TMIGR_NONE migrator = 8
active = NONE active = 8
nextevt = KTIME_MAX nextevt = KTIME_MAX
/ \ |
0 1 .. 7 8
idle idle active
5) Now CPU 8 is connected with GRP0:1 and CPU 8 calls tmigr_active_up() out
of tmigr_cpu_online().
[GRP1:0]
migrator = GRP0:0
active = GRP0:0
nextevt = T8
/ \
[GRP0:0] [GRP0:1]
migrator = TMIGR_NONE migrator = TMIGR_NONE
active = NONE active = NONE
nextevt = KTIME_MAX nextevt = T8
/ \ |
0 1 .. 7 8
idle idle idle
5) CPU 8 goes idle with a timer T8 and relies on GRP0:0 as the migrator.
But it's not really active, so T8 gets ignored.
--> The update which is done in third step is not noticed by setup code. So
a wrong migrator is set to top level group and a timer could get
ignored.
(B) Reading group->parent and group->childmask when an hierarchy update is
ongoing and reaches the formerly top level group is racy as those values
could be inconsistent. (The notation of migrator and active now slightly
changes in contrast to the above example, as now the childmasks are used.)
[GRP1:0]
migrator = TMIGR_NONE
active = 0x00
nextevt = KTIME_MAX
\
[GRP0:0] [GRP0:1]
migrator = TMIGR_NONE migrator = TMIGR_NONE
active = 0x00 active = 0x00
nextevt = KTIME_MAX nextevt = KTIME_MAX
childmask= 0 childmask= 1
parent = NULL parent = GRP1:0
/ \
0 1 .. 7 8
idle idle !online
childmask=1
1) Hierarchy has 8 CPUs. CPU 8 is at the moment in the process of onlining
but did not yet connect GRP0:0 to GRP1:0.
[GRP1:0]
migrator = TMIGR_NONE
active = 0x00
nextevt = KTIME_MAX
/ \
[GRP0:0] [GRP0:1]
migrator = TMIGR_NONE migrator = TMIGR_NONE
active = 0x00 active = 0x00
nextevt = KTIME_MAX nextevt = KTIME_MAX
childmask= 0 childmask= 1
parent = GRP1:0 parent = GRP1:0
/ \
0 1 .. 7 8
idle idle !online
childmask=1
2) Setup code (running on CPU 8) now connects GRP0:0 to GRP1:0, updates
parent pointer of GRP0:0 and ...
[GRP1:0]
migrator = TMIGR_NONE
active = 0x00
nextevt = KTIME_MAX
/ \
[GRP0:0] [GRP0:1]
migrator = 0x01 migrator = TMIGR_NONE
active = 0x01 active = 0x00
nextevt = KTIME_MAX nextevt = KTIME_MAX
childmask= 0 childmask= 1
parent = GRP1:0 parent = GRP1:0
/ \
0 1 .. 7 8
active idle !online
childmask=1
tmigr_walk.childmask = 0
3) ... CPU 0 comes active in the same time. As migrator in GRP0:0 was
TMIGR_NONE, childmask of GRP0:0 is stored in update propagation data
structure tmigr_walk (as update of childmask is not yet
visible/updated). And now ...
[GRP1:0]
migrator = TMIGR_NONE
active = 0x00
nextevt = KTIME_MAX
/ \
[GRP0:0] [GRP0:1]
migrator = 0x01 migrator = TMIGR_NONE
active = 0x01 active = 0x00
nextevt = KTIME_MAX nextevt = KTIME_MAX
childmask= 2 childmask= 1
parent = GRP1:0 parent = GRP1:0
/ \
0 1 .. 7 8
active idle !online
childmask=1
tmigr_walk.childmask = 0
4) ... childmask of GRP0:0 is updated by CPU 8 (still part of setup
code).
[GRP1:0]
migrator = 0x00
active = 0x00
nextevt = KTIME_MAX
/ \
[GRP0:0] [GRP0:1]
migrator = 0x01 migrator = TMIGR_NONE
active = 0x01 active = 0x00
nextevt = KTIME_MAX nextevt = KTIME_MAX
childmask= 2 childmask= 1
parent = GRP1:0 parent = GRP1:0
/ \
0 1 .. 7 8
active idle !online
childmask=1
tmigr_walk.childmask = 0
5) CPU 0 sees the connection to GRP1:0 and now propagates active state to
GRP1:0 but with childmask = 0 as stored in propagation data structure.
--> Now GRP1:0 always has a migrator as 0x00 != TMIGR_NONE and for all CPUs
it looks like GRP1:0 is always active.
To prevent those races, the setup of the hierarchy is moved into the
cpuhotplug prepare callback. The prepare callback is not executed by the
CPU which will come online, it is executed by the CPU which prepares
onlining of the other CPU. This CPU is active while it is connecting the
formerly top level to the new one. This prevents from (A) to happen and it
also prevents from any further walk above the formerly top level until that
active CPU becomes inactive, releasing the new ->parent and ->childmask
updates to be visible by any subsequent walk up above the formerly top
level hierarchy. This prevents from (B) to happen. The direction for the
updates is now forced to look like "from bottom to top".
However if the active CPU prevents from tmigr_cpu_(in)active() to walk up
with the update not-or-half visible, nothing prevents walking up to the new
top with a 0 childmask in tmigr_handle_remote_up() or
tmigr_requires_handle_remote_up() if the active CPU doing the prepare is
not the migrator. But then it looks fine because:
* tmigr_check_migrator() should just return false
* The migrator is active and should eventually observe the new childmask
at some point in a future tick.
Split setup functionality of online callback into the cpuhotplug prepare
callback and setup hotplug state. Change init call into early_initcall() to
make sure an already active CPU prepares everything for newly upcoming
CPUs. Reorder the code, that all prepare related functions are close to
each other and online and offline callbacks are also close together.
Fixes:
|
||
|
Anna-Maria Behnsen
|
facd40aa5c |
timers/migration: Do not rely always on group->parent
When reading group->parent without holding the group lock it is racy
against CPUs coming online the first time and thereby creating another
level of the hierarchy. This is not a problem when this value is read once
to decide whether to abort a propagation or not. The worst outcome is an
unnecessary/early CPU wake up. But it is racy when reading it several times
during a single 'action' (like activation, deactivation, checking for
remote timer expiry,...) and relying on the consitency of this value
without holding the lock. This happens at the moment e.g. in
tmigr_inactive_up() which is also calling tmigr_udpate_events(). Code relys
on group->parent not to change during this 'action'.
Update parent struct member description to explain the above only
once. Remove parent pointer checks when they are not mandatory (like update
of data->childmask). Remove a warning, which would be nice but the trigger
of this warning is not reliable and add expand the data structure member
description instead. Expand a comment, why it is safe to rely on parent
pointer here (inside hierarchy update).
Fixes:
|
||
Levi Yun
|
d7ad05c86e |
timers/migration: Prevent out of bounds access on failure
When tmigr_setup_groups() fails the level 0 group allocation, then the
cleanup derefences index -1 of the local stack array.
Prevent this by checking the loop condition first.
Fixes:
|
||
|
Anna-Maria Behnsen
|
7a96a84bfb |
timers/migration: Return early on deactivation
Commit |
||
|
Frederic Weisbecker
|
61f7fdf8fd |
timers/migration: Fix ignored event due to missing CPU update
When a group event is updated with its expiry unchanged but a different
CPU, that target change may go unnoticed and the event may be propagated
up with a stale CPU value. The following depicts a scenario that has
been actually observed:
[GRP2:0]
migrator = GRP1:1
active = GRP1:1
nextevt = TGRP1:0 (T0)
/ \
[GRP1:0] [GRP1:1]
migrator = NONE [...]
active = NONE
nextevt = TGRP0:0 (T0)
/ \
[GRP0:0] [...]
migrator = NONE
active = NONE
nextevt = T0
/ \
0 (T0) 1 (T1)
idle idle
0) The hierarchy has 3 levels. The left part (GRP1:0) is all idle,
including CPU 0 and CPU 1 which have a timer each: T0 and T1. They have
the same expiry value.
[GRP2:0]
migrator = GRP1:1
active = GRP1:1
nextevt = KTIME_MAX
/ \
[GRP1:0] [GRP1:1]
migrator = NONE [...]
active = NONE
nextevt = TGRP0:0 (T0)
/ \
[GRP0:0] [...]
migrator = NONE
active = NONE
nextevt = T0
/ \
0 (T0) 1 (T1)
idle idle
1) The migrator in GRP1:1 handles remotely T0. The event is dequeued
from the top and T0 executed.
[GRP2:0]
migrator = GRP1:1
active = GRP1:1
nextevt = KTIME_MAX
/ \
[GRP1:0] [GRP1:1]
migrator = NONE [...]
active = NONE
nextevt = TGRP0:0 (T0)
/ \
[GRP0:0] [...]
migrator = NONE
active = NONE
nextevt = T1
/ \
0 1 (T1)
idle idle
2) The migrator in GRP1:1 fetches the next timer for CPU 0 and finds
none. But it updates the events from its groups, starting with GRP0:0
which now has T1 as its next event. So far so good.
[GRP2:0]
migrator = GRP1:1
active = GRP1:1
nextevt = KTIME_MAX
/ \
[GRP1:0] [GRP1:1]
migrator = NONE [...]
active = NONE
nextevt = TGRP0:0 (T0)
/ \
[GRP0:0] [...]
migrator = NONE
active = NONE
nextevt = T1
/ \
0 1 (T1)
idle idle
3) The migrator in GRP1:1 proceeds upward and updates the events in
GRP1:0. The child event TGRP0:0 is found queued with the same expiry
as before. And therefore it is left unchanged. However the target CPU
is not the same but that fact is ignored so TGRP0:0 still points to
CPU 0 when it should point to CPU 1.
[GRP2:0]
migrator = GRP1:1
active = GRP1:1
nextevt = TGRP1:0 (T0)
/ \
[GRP1:0] [GRP1:1]
migrator = NONE [...]
active = NONE
nextevt = TGRP0:0 (T0)
/ \
[GRP0:0] [...]
migrator = NONE
active = NONE
nextevt = T1
/ \
0 1 (T1)
idle idle
4) The propagation has reached the top level and TGRP1:0, having TGRP0:0
as its first event, also wrongly points to CPU 0. TGRP1:0 is added to
the top level group.
[GRP2:0]
migrator = GRP1:1
active = GRP1:1
nextevt = KTIME_MAX
/ \
[GRP1:0] [GRP1:1]
migrator = NONE [...]
active = NONE
nextevt = TGRP0:0 (T0)
/ \
[GRP0:0] [...]
migrator = NONE
active = NONE
nextevt = T1
/ \
0 1 (T1)
idle idle
5) The migrator in GRP1:1 dequeues the next event in top level pointing
to CPU 0. But since it actually doesn't see any real event in CPU 0, it
early returns.
6) T1 is left unhandled until either CPU 0 or CPU 1 wake up.
Some other bad scenario may involve trees with just two levels.
Fix this with unconditionally updating the CPU of the child event before
considering to early return while updating a queued event with an
unchanged expiry value.
Fixes:
|
||
|
Frederic Weisbecker
|
f55acb1e44 |
timers/migration: Fix endless timer requeue after idle interrupts
When a CPU is an idle migrator, but another CPU wakes up before it,
becomes an active migrator and handles the queue, the initial idle
migrator may end up endlessly reprogramming its clockevent, chasing ghost
timers forever such as in the following scenario:
[GRP0:0]
migrator = 0
active = 0
nextevt = T1
/ \
0 1
active idle (T1)
0) CPU 1 is idle and has a timer queued (T1), CPU 0 is active and is
the active migrator.
[GRP0:0]
migrator = NONE
active = NONE
nextevt = T1
/ \
0 1
idle idle (T1)
wakeup = T1
1) CPU 0 is now idle and is therefore the idle migrator. It has
programmed its next timer interrupt to handle T1.
[GRP0:0]
migrator = 1
active = 1
nextevt = KTIME_MAX
/ \
0 1
idle active
wakeup = T1
2) CPU 1 has woken up, it is now active and it has just handled its own
timer T1.
3) CPU 0 gets a timer interrupt to handle T1 but tmigr_handle_remote()
realize it is not the migrator anymore. So it early returns without
observing that T1 has been expired already and therefore without
updating its ->wakeup value.
4) CPU 0 goes into tmigr_cpu_new_timer() which also early returns
because it doesn't queue a timer of its own. So ->wakeup is left
unchanged and the next timer is programmed to fire now.
5) goto 3) forever
This results in timer interrupt storms in idle and also in nohz_full (as
observed in rcutorture's TREE07 scenario).
Fix this with forcing a re-evaluation of tmc->wakeup while trying
remote timer handling when the CPU isn't the migrator anymmore. The
check is inherently racy but in the worst case the CPU just races setting
the KTIME_MAX value that a remote expiry also tries to set.
Fixes:
|
||
|
Frederic Weisbecker
|
4b6f4c5a67 |
timer/migration: Remove buggy early return on deactivation
When a CPU enters into idle and deactivates itself from the timer
migration hierarchy without any global timer of its own to propagate,
the group event of that CPU is set to "ignore" and tmigr_update_events()
accordingly performs an early return without considering timers queued
by other CPUs.
If the hierarchy has a single level, and the CPU is the last one to
enter idle, it will ignore others' global timers, as in the following
layout:
[GRP0:0]
migrator = 0
active = 0
nextevt = T0i
/ \
0 1
active (T0i) idle (T1)
0) CPU 0 is active thus its event is ignored (the letter 'i') and so are
upper levels' events. CPU 1 is idle and has the timer T1 enqueued.
[GRP0:0]
migrator = NONE
active = NONE
nextevt = T0i
/ \
0 1
idle (T0i) idle (T1)
1) CPU 0 goes idle without global event queued. Therefore KTIME_MAX is
pushed as its next expiry and its own event kept as "ignore". As a result
tmigr_update_events() ignores T1 and CPU 0 goes to idle with T1
unhandled.
This isn't proper to single level hierarchy though. A similar issue,
although slightly different, may arise on multi-level:
[GRP1:0]
migrator = GRP0:0
active = GRP0:0
nextevt = T0:0i, T0:1
/ \
[GRP0:0] [GRP0:1]
migrator = 0 migrator = NONE
active = 0 active = NONE
nextevt = T0i nextevt = T2
/ \ / \
0 (T0i) 1 (T1) 2 (T2) 3
active idle idle idle
0) CPU 0 is active thus its event is ignored (the letter 'i') and so are
upper levels' events. CPU 1 is idle and has the timer T1 enqueued.
CPU 2 also has a timer. The expiry order is T0 (ignored) < T1 < T2
[GRP1:0]
migrator = GRP0:0
active = GRP0:0
nextevt = T0:0i, T0:1
/ \
[GRP0:0] [GRP0:1]
migrator = NONE migrator = NONE
active = NONE active = NONE
nextevt = T0i nextevt = T2
/ \ / \
0 (T0i) 1 (T1) 2 (T2) 3
idle idle idle idle
1) CPU 0 goes idle without global event queued. Therefore KTIME_MAX is
pushed as its next expiry and its own event kept as "ignore". As a result
tmigr_update_events() ignores T1. The change only propagated up to 1st
level so far.
[GRP1:0]
migrator = NONE
active = NONE
nextevt = T0:1
/ \
[GRP0:0] [GRP0:1]
migrator = NONE migrator = NONE
active = NONE active = NONE
nextevt = T0i nextevt = T2
/ \ / \
0 (T0i) 1 (T1) 2 (T2) 3
idle idle idle idle
2) The change now propagates up to the top. tmigr_update_events() finds
that the child event is ignored and thus removes it. The top level next
event is now T2 which is returned to CPU 0 as its next effective expiry
to take account for as the global idle migrator. However T1 has been
ignored along the way, leaving it unhandled.
Fix those issues with removing the buggy related early return. Ignored
child events must not prevent from evaluating the other events within
the same group.
Reported-by: Boqun Feng <boqun.feng@gmail.com>
Reported-by: Florian Fainelli <f.fainelli@gmail.com>
Reported-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Florian Fainelli <florian.fainelli@broadcom.com>
Link: https://lore.kernel.org/r/ZfOhB9ZByTZcBy4u@lothringen
|
||
|
Frederic Weisbecker
|
8ca1836769 |
timer/migration: Fix quick check reporting late expiry
When a CPU is the last active in the hierarchy and it tries to enter
into idle, the quick check looking up the next event towards cpuidle
heuristics may report a too late expiry, such as in the following
scenario:
[GRP1:0]
migrator = NONE
active = NONE
nextevt = T0:0, T0:1
/ \
[GRP0:0] [GRP0:1]
migrator = NONE migrator = NONE
active = NONE active = NONE
nextevt = T0, T1 nextevt = T2
/ \ / \
0 1 2 3
idle idle idle idle
0) The whole system is idle, and CPU 0 was the last migrator. CPU 0 has
a timer (T0), CPU 1 has a timer (T1) and CPU 2 has a timer (T2). The
expire order is T0 < T1 < T2.
[GRP1:0]
migrator = GRP0:0
active = GRP0:0
nextevt = T0:0(i), T0:1
/ \
[GRP0:0] [GRP0:1]
migrator = CPU0 migrator = NONE
active = CPU0 active = NONE
nextevt = T0(i), T1 nextevt = T2
/ \ / \
0 1 2 3
active idle idle idle
1) CPU 0 becomes active. The (i) means a now ignored timer.
[GRP1:0]
migrator = GRP0:0
active = GRP0:0
nextevt = T0:1
/ \
[GRP0:0] [GRP0:1]
migrator = CPU0 migrator = NONE
active = CPU0 active = NONE
nextevt = T1 nextevt = T2
/ \ / \
0 1 2 3
active idle idle idle
2) CPU 0 handles remote. No timer actually expired but ignored timers
have been cleaned out and their sibling's timers haven't been
propagated. As a result the top level's next event is T2 and not T1.
3) CPU 0 tries to enter idle without any global timer enqueued and calls
tmigr_quick_check(). The expiry of T2 is returned instead of the
expiry of T1.
When the quick check returns an expiry that is too late, the cpuidle
governor may pick up a C-state that is too deep. This may be result into
undesired CPU wake up latency if the next timer is actually close enough.
Fix this with assuming that expiries aren't sorted top-down while
performing the quick check. Pick up instead the earliest encountered one
while walking up the hierarchy.
|
||
|
Anna-Maria Behnsen
|
36e40df35d |
timer_migration: Add tracepoints
The timer pull logic needs proper debugging aids. Add tracepoints so the hierarchical idle machinery can be diagnosed. Signed-off-by: Anna-Maria Behnsen <anna-maria@linutronix.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Link: https://lore.kernel.org/r/20240222103403.31923-1-anna-maria@linutronix.de |
||
|
Anna-Maria Behnsen
|
7ee9887703 |
timers: Implement the hierarchical pull model
Placing timers at enqueue time on a target CPU based on dubious heuristics
does not make any sense:
1) Most timer wheel timers are canceled or rearmed before they expire.
2) The heuristics to predict which CPU will be busy when the timer expires
are wrong by definition.
So placing the timers at enqueue wastes precious cycles.
The proper solution to this problem is to always queue the timers on the
local CPU and allow the non pinned timers to be pulled onto a busy CPU at
expiry time.
Therefore split the timer storage into local pinned and global timers:
Local pinned timers are always expired on the CPU on which they have been
queued. Global timers can be expired on any CPU.
As long as a CPU is busy it expires both local and global timers. When a
CPU goes idle it arms for the first expiring local timer. If the first
expiring pinned (local) timer is before the first expiring movable timer,
then no action is required because the CPU will wake up before the first
movable timer expires. If the first expiring movable timer is before the
first expiring pinned (local) timer, then this timer is queued into an idle
timerqueue and eventually expired by another active CPU.
To avoid global locking the timerqueues are implemented as a hierarchy. The
lowest level of the hierarchy holds the CPUs. The CPUs are associated to
groups of 8, which are separated per node. If more than one CPU group
exist, then a second level in the hierarchy collects the groups. Depending
on the size of the system more than 2 levels are required. Each group has a
"migrator" which checks the timerqueue during the tick for remote expirable
timers.
If the last CPU in a group goes idle it reports the first expiring event in
the group up to the next group(s) in the hierarchy. If the last CPU goes
idle it arms its timer for the first system wide expiring timer to ensure
that no timer event is missed.
Signed-off-by: Anna-Maria Behnsen <anna-maria@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Link: https://lore.kernel.org/r/20240222103710.32582-1-anna-maria@linutronix.de
|