| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
Commit 3427445afd26 ("sched: Exclude cond_resched() from nested sleep
test") removed the task state check of __might_sleep() for
cond_resched_lock() because cond_resched_lock() is not a voluntary
scheduling point which blocks. It's a preemption point which requires the
lock holder to release the spin lock.
The same rationale applies to cond_resched_rwlock_read/write(), but those
were not touched.
Make it consistent and use the non-state checking __might_resched() there
as well.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20210923165357.991262778@linutronix.de
|
|
__might_sleep() vs. ___might_sleep() is hard to distinguish. Aside of that
the three underscore variant is exposed to provide a checkpoint for
rescheduling points which are distinct from blocking points.
They are semantically a preemption point which means that scheduling is
state preserving. A real blocking operation, e.g. mutex_lock(), wait*(),
which cannot preserve a task state which is not equal to RUNNING.
While technically blocking on a "sleeping" spinlock in RT enabled kernels
falls into the voluntary scheduling category because it has to wait until
the contended spin/rw lock becomes available, the RT lock substitution code
can semantically be mapped to a voluntary preemption because the RT lock
substitution code and the scheduler are providing mechanisms to preserve
the task state and to take regular non-lock related wakeups into account.
Rename ___might_sleep() to __might_resched() to make the distinction of
these functions clear.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20210923165357.928693482@linutronix.de
|
|
In the commit mentioned below, fscache was converted from slow-work to
workqueue. slow_work_enqueue() and slow_work_sleep_till_thread_needed()
did not use a per-CPU workqueue. They choose from two global waitqueues
depending on the SLOW_WORK_VERY_SLOW bit which was not set so it always
one waitqueue.
I can't find out how it is ensured that a waiter on certain CPU is woken
up be the other side. My guess is that the timeout in schedule_timeout()
ensures that it does not wait forever (or a random wake up).
fscache_object_sleep_till_congested() must be invoked from preemptible
context in order for schedule() to work. In this case this_cpu_ptr()
should complain with CONFIG_DEBUG_PREEMPT enabled except the thread is
bound to one CPU.
wake_up() wakes only one waiter and I'm not sure if it is guaranteed
that only one waiter exists.
Replace the per-CPU waitqueue with one global waitqueue.
Fixes: 8b8edefa2fffb ("fscache: convert object to use workqueue instead of slow-work")
Reported-by: Gregor Beck <gregor.beck@gmail.com>
Cc: stable-rt@vger.kernel.org
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20211029083839.xwwt7jgzru3kcpii@linutronix.de
|
|
The MNT_WRITE_HOLD flag is used to hold back any new writers while the
mount point is about to be made read-only. __mnt_want_write() then loops
with disabled preemption until this flag disappears. Callers of
mnt_hold_writers() (which sets the flag) hold the spinlock_t of
mount_lock (seqlock_t) which disables preemption on !PREEMPT_RT and
ensures the task is not scheduled away so that the spinning side spins
for a long time.
On PREEMPT_RT the spinlock_t does not disable preemption and so it is
possible that the task setting MNT_WRITE_HOLD is preempted by task with
higher priority which then spins infinitely waiting for MNT_WRITE_HOLD
to get removed.
Acquire mount_lock::lock which is held by setter of MNT_WRITE_HOLD. This
will PI-boost the owner and wait until the lock is dropped and which
means that MNT_WRITE_HOLD is cleared again.
Link: https://lkml.kernel.org/r/20211025152218.opvcqfku2lhqvp4o@linutronix.de
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
|
|
The softirq implementation on PREEMPT_RT does not provide do_softirq().
The other user of do_softirq() is replaced with a local_bh_disable()
+ enable() around the possible raise-softirq invocation. This can not be
done here because migration_cpu_stop() is invoked with disabled
preemption.
Wake the softirq thread on PREEMPT_RT if there are any pending softirqs.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20210927073814.x5h6osr4dgiu44sc@linutronix.de
|
|
__raise_softirq_irqoff() adds a bit to the pending sofirq mask and this
is it. The softirq won't be handled in a deterministic way but randomly
when an interrupt fires and handles softirq in its irq_exit() routine or
if something randomly checks and handles pending softirqs in the call
chain before the CPU goes idle.
Add a local_bh_disable/enable() around the IRQ-off section which will
handle pending softirqs.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20210930103754.2128949-1-bigeasy@linutronix.de
|
|
On PREEMPT_RT most items are processed as LAZY via softirq context.
Avoid to spin-wait for them because irq_work_sync() could have higher
priority and not allow the irq-work to be completed.
Wait additionally for !IRQ_WORK_HARD_IRQ irq_work items on PREEMPT_RT.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20211006111852.1514359-5-bigeasy@linutronix.de
|
|
The irq_work callback is invoked in hard IRQ context. By default all
callbacks are scheduled for invocation right away (given supported by
the architecture) except for the ones marked IRQ_WORK_LAZY which are
delayed until the next timer-tick.
While looking over the callbacks, some of them may acquire locks
(spinlock_t, rwlock_t) which are transformed into sleeping locks on
PREEMPT_RT and must not be acquired in hard IRQ context.
Changing the locks into locks which could be acquired in this context
will lead to other problems such as increased latencies if everything
in the chain has IRQ-off locks. This will not solve all the issues as
one callback has been noticed which invoked kref_put() and its callback
invokes kfree() and this can not be invoked in hardirq context.
Some callbacks are required to be invoked in hardirq context even on
PREEMPT_RT to work properly. This includes for instance the NO_HZ
callback which needs to be able to observe the idle context.
The callbacks which require to be run in hardirq have already been
marked. Use this information to split the callbacks onto the two lists
on PREEMPT_RT:
- lazy_list
Work items which are not marked with IRQ_WORK_HARD_IRQ will be added
to this list. Callbacks on this list will be invoked from a per-CPU
thread.
The handler here may acquire sleeping locks such as spinlock_t and
invoke kfree().
- raised_list
Work items which are marked with IRQ_WORK_HARD_IRQ will be added to
this list. They will be invoked in hardirq context and must not
acquire any sleeping locks.
The wake up of the per-CPU thread occurs from irq_work handler/
hardirq context. The thread runs with lowest RT priority to ensure it
runs before any SCHED_OTHER tasks do.
[bigeasy: melt tglx's irq_work_tick_soft() which splits irq_work_tick() into a
hard and soft variant. Collected fixes over time from Steven
Rostedt and Mike Galbraith. Move to per-CPU threads instead of
softirq as suggested by PeterZ.]
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20211007092646.uhshe3ut2wkrcfzv@linutronix.de
|
|
irq_work() triggers instantly an interrupt if supported by the
architecture. Otherwise the work will be processed on the next timer
tick. In worst case irq_work_sync() could spin up to a jiffy.
irq_work_sync() is usually used in tear down context which is fully
preemptible. Based on review irq_work_sync() is invoked from preemptible
context and there is one waiter at a time. This qualifies it to use
rcuwait for synchronisation.
Let irq_work_sync() synchronize with rcuwait if the architecture
processes irqwork via the timer tick.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20211006111852.1514359-3-bigeasy@linutronix.de
|
|
The push-IPI logic for RT tasks expects to be invoked from hardirq
context. One reason is that a RT task on the remote CPU would block the
softirq processing on PREEMPT_RT and so avoid pulling / balancing the RT
tasks as intended.
Annotate root_domain::rto_push_work as IRQ_WORK_HARD_IRQ.
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Ben Segall <bsegall@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Daniel Bristot de Oliveira <bristot@redhat.com>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20211006111852.1514359-2-bigeasy@linutronix.de
|
|
Since the rework, the statistics code always adds up the byte and packet
value(s). On 32bit architectures a seqcount_t is used in
gnet_stats_basic_sync to ensure that the 64bit values are not modified
during the read since two 32bit loads are required. The usage of a
seqcount_t requires a lock to ensure that only one writer is active at a
time. This lock leads to disabled preemption during the update.
The lack of disabling preemption is now creating a warning as reported
by Naresh since the query done by gnet_stats_copy_basic() is in
preemptible context.
For ___gnet_stats_copy_basic() there is no need to disable preemption
since the update is performed on stack and can't be modified by another
writer. Instead of disabling preemption, to avoid the warning,
simply create a read function to just read the values and return as u64.
Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org>
Fixes: 67c9e6270f301 ("net: sched: Protect Qdisc::bstats with u64_stats")
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20211021095919.bi3szpt3c2kcoiso@linutronix.de
|
|
__QDISC_STATE_RUNNING is only set/cleared from contexts owning qdisc lock.
Thus we can use less expensive bit operations, as we were doing
before commit f9eb8aea2a1e ("net_sched: transform qdisc running bit into a seqcount")
Fixes: 29cbcd858283 ("net: sched: Remove Qdisc::running sequence counter")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ahmed S. Darwish <a.darwish@linutronix.de>
Acked-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Tested-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
|
|
For non TCQ_F_NOLOCK qdisc, qdisc_run_begin() tries to set
__QDISC_STATE_RUNNING and should return true if the bit was not set.
test_and_set_bit() returns old bit value, therefore we need to invert.
Fixes: 29cbcd858283 ("net: sched: Remove Qdisc::running sequence counter")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ahmed S. Darwish <a.darwish@linutronix.de>
Tested-by: Ido Schimmel <idosch@nvidia.com>
Acked-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Tested-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
|
|
Eric reported that the rate estimator reads statics from the softirq
which in turn triggers a warning introduced in the statistics rework.
The warning is too cautious. The updates happen in the softirq context
so reads from softirq are fine since the writes can not be preempted.
The updates/writes happen during qdisc_run() which ensures one writer
and the softirq context.
The remaining bad context for reading statistics remains in hard-IRQ
because it may preempt a writer.
Fixes: 29cbcd8582837 ("net: sched: Remove Qdisc::running sequence counter")
Reported-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The Qdisc::running sequence counter has two uses:
1. Reliably reading qdisc's tc statistics while the qdisc is running
(a seqcount read/retry loop at gnet_stats_add_basic()).
2. As a flag, indicating whether the qdisc in question is running
(without any retry loops).
For the first usage, the Qdisc::running sequence counter write section,
qdisc_run_begin() => qdisc_run_end(), covers a much wider area than what
is actually needed: the raw qdisc's bstats update. A u64_stats sync
point was thus introduced (in previous commits) inside the bstats
structure itself. A local u64_stats write section is then started and
stopped for the bstats updates.
Use that u64_stats sync point mechanism for the bstats read/retry loop
at gnet_stats_add_basic().
For the second qdisc->running usage, a __QDISC_STATE_RUNNING bit flag,
accessed with atomic bitops, is sufficient. Using a bit flag instead of
a sequence counter at qdisc_run_begin/end() and qdisc_is_running() leads
to the SMP barriers implicitly added through raw_read_seqcount() and
write_seqcount_begin/end() getting removed. All call sites have been
surveyed though, and no required ordering was identified.
Now that the qdisc->running sequence counter is no longer used, remove
it.
Note, using u64_stats implies no sequence counter protection for 64-bit
architectures. This can lead to the qdisc tc statistics "packets" vs.
"bytes" values getting out of sync on rare occasions. The individual
values will still be valid.
Signed-off-by: Ahmed S. Darwish <a.darwish@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The only factor differentiating per-CPU bstats data type (struct
gnet_stats_basic_cpu) from the packed non-per-CPU one (struct
gnet_stats_basic_packed) was a u64_stats sync point inside the former.
The two data types are now equivalent: earlier commits added a u64_stats
sync point to the latter.
Combine both data types into "struct gnet_stats_basic_sync". This
eliminates redundancy and simplifies the bstats read/write APIs.
Use u64_stats_t for bstats "packets" and "bytes" data types. On 64-bit
architectures, u64_stats sync points do not use sequence counter
protection.
Signed-off-by: Ahmed S. Darwish <a.darwish@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The Qdisc::running sequence counter, used to protect Qdisc::bstats reads
from parallel writes, is in the process of being removed. Qdisc::bstats
read/writes will synchronize using an internal u64_stats sync point
instead.
Modify all bstats writes to use _bstats_update(). This ensures that
the internal u64_stats sync point is always acquired and released as
appropriate.
Signed-off-by: Ahmed S. Darwish <a.darwish@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The not-per-CPU variant of qdisc tc (traffic control) statistics,
Qdisc::gnet_stats_basic_packed bstats, is protected with Qdisc::running
sequence counter.
This sequence counter is used for reliably protecting bstats reads from
parallel writes. Meanwhile, the seqcount's write section covers a much
wider area than bstats update: qdisc_run_begin() => qdisc_run_end().
That read/write section asymmetry can lead to needless retries of the
read section. To prepare for removing the Qdisc::running sequence
counter altogether, introduce a u64_stats sync point inside bstats
instead.
Modify _bstats_update() to start/end the bstats u64_stats write
section.
For bisectability, and finer commits granularity, the bstats read
section is still protected with a Qdisc::running read/retry loop and
qdisc_run_begin/end() still starts/ends that seqcount write section.
Once all call sites are modified to use _bstats_update(), the
Qdisc::running seqcount will be removed and bstats read/retry loop will
be modified to utilize the internal u64_stats sync point.
Note, using u64_stats implies no sequence counter protection for 64-bit
architectures. This can lead to the statistics "packets" vs. "bytes"
values getting out of sync on rare occasions. The individual values will
still be valid.
[bigeasy: Minor commit message edits, init all gnet_stats_basic_packed.]
Signed-off-by: Ahmed S. Darwish <a.darwish@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Allow to directly set a u64_stats_t value which is used to provide an init
function which sets it directly to zero intead of memset() the value.
Add u64_stats_set() to the u64_stats API.
[bigeasy: commit message. ]
Signed-off-by: Ahmed S. Darwish <a.darwish@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The gnet_stats_queue::qlen member is only used in the SMP-case.
qdisc_qstats_qlen_backlog() needs to add qdisc_qlen() to qstats.qlen to
have the same value as that provided by qdisc_qlen_sum().
gnet_stats_copy_queue() needs to overwritte the resulting qstats.qlen
field whith the caller submitted qlen value. It might be differ from the
submitted value.
Let both functions use gnet_stats_add_queue() and remove unused
__gnet_stats_copy_queue().
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
gnet_stats_add_basic() and gnet_stats_add_queue() add up the statistics
so they can be used directly for both the per-CPU and global case.
gnet_stats_add_queue() copies either Qdisc's per-CPU
gnet_stats_queue::qlen or the global member. The global
gnet_stats_queue::qlen isn't touched in the per-CPU case so there is no
need to consider it in the global-case.
In the per-CPU case, the sum of global gnet_stats_queue::qlen and
the per-CPU gnet_stats_queue::qlen was assigned to sch->q.qlen and
sch->qstats.qlen. Now both fields are copied individually.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This function will replace __gnet_stats_copy_queue(). It reads all
arguments and adds them into the passed gnet_stats_queue argument.
In contrast to __gnet_stats_copy_queue() it also copies the qlen member.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
__gnet_stats_copy_basic() always assigns the value to the bstats
argument overwriting the previous value. The later added per-CPU version
always accumulated the values in the returning gnet_stats_basic_packed
argument.
Based on review there are five users of that function as of today:
- est_fetch_counters(), ___gnet_stats_copy_basic()
memsets() bstats to zero, single invocation.
- mq_dump(), mqprio_dump(), mqprio_dump_class_stats()
memsets() bstats to zero, multiple invocation but does not use the
function due to !qdisc_is_percpu_stats().
Add the values in __gnet_stats_copy_basic() instead overwriting. Rename
the function to gnet_stats_add_basic() to make it more obvious.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
leaf classes of ETS qdiscs are served in strict priority or deficit round
robin (DRR), depending on the value of 'nstrict'. Since this value can be
changed while traffic is running, we need to be sure that the active list
of DRR classes can be updated at any time, so:
1) call INIT_LIST_HEAD(&alist) on all leaf classes in .init(), before the
first packet hits any of them.
2) ensure that 'alist' is not overwritten with zeros when a leaf class is
no more strict priority nor DRR (i.e. array elements beyond 'nbands').
Link: https://lore.kernel.org/netdev/YS%2FoZ+f0Nr8eQkzH@dcaratti.users.ipa.redhat.com
Suggested-by: Cong Wang <cong.wang@bytedance.com>
Signed-off-by: Davide Caratti <dcaratti@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
|
|
The kcov code mixes local_irq_save() and spin_lock() in
kcov_remote_{start|end}(). This creates a warning on PREEMPT_RT because
local_irq_save() disables interrupts and spin_lock_t is turned into a
sleeping lock which can not be acquired in a section with disabled
interrupts.
The kcov_remote_lock is used to synchronize the access to the hash-list
kcov_remote_map. The local_irq_save() block protects access to the
per-CPU data kcov_percpu_data.
There no compelling reason to change the lock type to raw_spin_lock_t to
make it work with local_irq_save(). Changing it would require to move
memory allocation (in kcov_remote_add()) and deallocation outside of the
locked section.
Adding an unlimited amount of entries to the hashlist will increase the
IRQ-off time during lookup. It could be argued that this is debug code
and the latency does not matter. There is however no need to do so and
it would allow to use this facility in an RT enabled build.
Using a local_lock_t instead of local_irq_save() has the befit of adding
a protection scope within the source which makes it obvious what is
protected. On a !PREEMPT_RT && !LOCKDEP build the local_lock_irqsave()
maps directly to local_irq_save() so there is overhead at runtime.
Replace the local_irq_save() section with a local_lock_t.
Reported-by: Clark Williams <williams@redhat.com>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20210830172627.267989-6-bigeasy@linutronix.de
|
|
kcov_remote_start() may need to allocate memory in the in_task() case
(otherwise per-CPU memory has been pre-allocated) and therefore requires
enabled interrupts.
The interrupts are enabled before checking if the allocation is required
so if no allocation is required then the interrupts are needlessly
enabled and disabled again.
Enable interrupts only if memory allocation is performed.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20210830172627.267989-5-bigeasy@linutronix.de
|
|
During boot kcov allocates per-CPU memory which is used later if remote/
softirq processing is enabled.
Allocate the per-CPU memory on the CPU local node to avoid cross node
memory access.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20210830172627.267989-4-bigeasy@linutronix.de
|
|
The example code uses the variable `ip' but never declares it.
Declare `ip' as a 64bit variable which is the same type as the array
from which it loads its value.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20210830172627.267989-3-bigeasy@linutronix.de
|
|
The first example code has includes at the top, the following two
example share that part. The last example (remote coverage collection)
requires the linux/types.h header file due its __aligned_u64 usage.
Add the linux/types.h to the top most example and a comment that the
header files from above are required as it is done in the second
example.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20210830172627.267989-2-bigeasy@linutronix.de
|
|
PREEMPT_RT preempts softirqs and the current implementation avoids
do_softirq_own_stack() and only uses __do_softirq().
Disable the unused softirqs stacks on PREEMPT_RT to safe some memory and
ensure that do_softirq_own_stack() is not used which is not expected.
[bigeasy: commit description.]
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20210924161245.2357247-1-bigeasy@linutronix.de
|
|
PREEMPT_RT
TRANSPARENT_HUGEPAGE:
There are potential non-deterministic delays to an RT thread if a critical
memory region is not THP-aligned and a non-RT buffer is located in the same
hugepage-aligned region. It's also possible for an unrelated thread to migrate
pages belonging to an RT task incurring unexpected page faults due to memory
defragmentation even if khugepaged is disabled.
Regular HUGEPAGEs are not affected by this can be used.
NUMA_BALANCING:
There is a non-deterministic delay to mark PTEs PROT_NONE to gather NUMA fault
samples, increased page faults of regions even if mlocked and non-deterministic
delays when migrating pages.
[Mel Gorman worded 99% of the commit description].
Link: https://lore.kernel.org/all/20200304091159.GN3818@techsingularity.net/
Link: https://lore.kernel.org/all/20211026165100.ahz5bkx44lrrw5pt@linutronix.de/
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Link: https://lore.kernel.org/r/20211028143327.hfbxjze7palrpfgp@linutronix.de
|
|
sg_miter_stop() checks for disabled preemption before unmapping a page
via kunmap_atomic(). The kernel doc mentions under context that
preemption must be disabled if SG_MITER_ATOMIC is set.
There is no active requirement for the caller to have preemption
disabled before invoking sg_mitter_stop(). The sg_mitter_*()
implementation itself has no such requirement.
In fact, preemption is disabled by kmap_atomic() as part of
sg_miter_next() and remains disabled as long as there is an active
SG_MITER_ATOMIC mapping. This is a consequence of kmap_atomic() and not
a requirement for sg_mitter_*() itself.
The user chooses SG_MITER_ATOMIC because it uses the API in a context
where blocking is not possible or blocking is possible but he chooses a
lower weight mapping which is not available on all CPUs and so it might
need less overhead to setup at a price that now preemption will be
disabled.
The kmap_atomic() implementation on PREEMPT_RT does not disable
preemption. It simply disables CPU migration to ensure that the task
remains on the same CPU while the caller remains preemptible. This in
turn triggers the warning in sg_miter_stop() because preemption is
allowed.
The PREEMPT_RT and !PREEMPT_RT implementation of kmap_atomic() disable
pagefaults as a requirement. It is sufficient to check for this instead
of disabled preemption.
Check for disabled pagefault handler in the SG_MITER_ATOMIC case. Remove
the "preemption disabled" part from the kernel doc as the sg_milter*()
implementation does not care.
[bigeasy: commit description. ]
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20211015211409.cqopacv3pxdwn2ty@linutronix.de
|
|
drain_local_pages_wq() disables preemption to avoid CPU migration during
CPU hotplug and can't use cpus_read_lock().
Using migrate_disable() works here, too. The scheduler won't take the
CPU offline until the task left the migrate-disable section.
The problem with disabled preemption here is that drain_local_pages()
acquires locks which are turned into sleeping locks on PREEMPT_RT and
can't be acquired with disabled preemption.
Use migrate_disable() in drain_local_pages_wq().
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20211015210933.viw6rjvo64qtqxn4@linutronix.de
|
|
Memory allocators may disable interrupts or preemption as part of the
allocation and freeing process. For PREEMPT_RT it is important that
these sections remain deterministic and short and therefore don't depend
on the size of the memory to allocate/ free or the inner state of the
algorithm.
Until v3.12-RT the SLAB allocator was an option but involved several
changes to meet all the requirements. The SLUB design fits better with
PREEMPT_RT model and so the SLAB patches were dropped in the 3.12-RT
patchset. Comparing the two allocator, SLUB outperformed SLAB in both
throughput (time needed to allocate and free memory) and the maximal
latency of the system measured with cyclictest during hackbench.
SLOB was never evaluated since it was unlikely that it preforms better
than SLAB. During a quick test, the kernel crashed with SLOB enabled
during boot.
Disable SLAB and SLOB on PREEMPT_RT.
[bigeasy: commit description.]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Link: https://lore.kernel.org/r/20211015210336.gen3tib33ig5q2md@linutronix.de
|
|
crypto_disable_simd_for_test() disables preemption in order to receive a
stable per-CPU variable which it needs to modify in order to alter
crypto_simd_usable() results.
This can also be achived by migrate_disable() which forbidds CPU
migrations but allows the task to be preempted. The latter is important
for PREEMPT_RT since operation like skcipher_walk_first() may allocate
memory which must not happen with disabled preemption on PREEMPT_RT.
Use migrate_disable() in crypto_disable_simd_for_test() to achieve a
stable per-CPU pointer.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20210928115401.441339-1-bigeasy@linutronix.de
|
|
The variables names read_lock and write_lock can clash with functions used for
read/writer locks.
Rename read_lock to read_access and write_lock to write_access to avoid a name
collision.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lkml.kernel.org/r/20210806152551.qio7c3ho6pexezup@linutronix.de
|
|
napi_busy_loop() disables preemption and performs a NAPI poll. We can't acquire
sleeping locks with disabled preemption which would be required while
__napi_poll() invokes the callback of the driver.
A threaded interrupt performing the NAPI-poll can be preempted on PREEMPT_RT.
A RT thread on another CPU may observe NAPIF_STATE_SCHED bit set and busy-spin
until it is cleared or its spin time runs out. Given it is the task with the
highest priority it will never observe the NEED_RESCHED bit set.
In this case the time is better spent by simply sleeping.
The NET_RX_BUSY_POLL is disabled by default (the system wide sysctls for
poll/read are set to zero). Disabling NET_RX_BUSY_POLL on PREEMPT_RT to avoid
wrong locking context in case it is used.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20211001145841.2308454-1-bigeasy@linutronix.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
For efficiency reasons, zsmalloc is using a slim `handle'. The value is
the address of a memory allocation of 4 or 8 bytes depending on the size
of the long data type. The lowest bit in that allocated memory is used
as a bit spin lock.
The usage of the bit spin lock is problematic because with the bit spin
lock held zsmalloc acquires a rwlock_t and spinlock_t which are both
sleeping locks on PREEMPT_RT and therefore must not be acquired with
disabled preemption.
There is a patch which extends the handle on PREEMPT_RT so that a full
spinlock_t fits (even with lockdep enabled) and then eliminates the bit
spin lock. I'm not sure how sensible zsmalloc on PREEMPT_RT is given
that it is used to store compressed user memory.
Disable ZSMALLOC on PREEMPT_RT. If there is need for it, we can try to
get it to work.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20210923170121.1860133-1-bigeasy@linutronix.de
|
|
In case the command line option "efi=noruntime" is default at built-time, the user
could overwrite its state by `efi=runtime' and allow it again.
This is useful on PREEMPT_RT where "efi=noruntime" is default and the
user might need to alter the boot order for instance.
Acked-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20210924134919.1913476-3-bigeasy@linutronix.de
|
|
Based on measurements the EFI functions get_variable /
get_next_variable take up to 2us which looks okay.
The functions get_time, set_time take around 10ms. These 10ms are too
much. Even one ms would be too much.
Ard mentioned that SetVariable might even trigger larger latencies if
the firmware will erase flash blocks on NOR.
The time-functions are used by efi-rtc and can be triggered during
run-time (either via explicit read/write or ntp sync).
The variable write could be used by pstore.
These functions can be disabled without much of a loss. The poweroff /
reboot hooks may be provided by PSCI.
Disable EFI's runtime wrappers on PREEMPT_RT.
This was observed on "EFI v2.60 by SoftIron Overdrive 1000".
Acked-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20210924134919.1913476-2-bigeasy@linutronix.de
|
|
lock_is_held_type(, 1) detects acquired read locks. It only recognized
locks acquired with lock_acquire_shared(). Read locks acquired with
lock_acquire_shared_recursive() are not recognized because a `2' is
stored as the read value.
Rework the check to additionally recognise lock's read value one and two
as a read held lock.
Fixes: e918188611f07 ("locking: More accurate annotations for read_lock()")
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Boqun Feng <boqun.feng@gmail.com>
Acked-by: Waiman Long <longman@redhat.com>
Link: https://lkml.kernel.org/r/20210903084001.lblecrvz4esl4mrr@linutronix.de
|
|
The support for misrouted IRQs is used on old / legacy systems and is
not feasible on PREEMPT_RT.
Polling for interrupts reduces the overall system performance.
Additionally the interrupt latency depends on the polling frequency and
delays are not desired for real time workloads.
Disable IRQ polling on PREEMPT_RT and let the user know that it is not
enabled. The compiler will optimize the real fixup/poll code out.
[ bigeasy: Update changelog and switch to IS_ENABLED() ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/r/20210917223841.c6j6jcaffojrnot3@linutronix.de
|
|
With enabled threaded interrupts the nouveau driver reported the
following:
| Chain exists of:
| &mm->mmap_lock#2 --> &device->mutex --> &cpuset_rwsem
|
| Possible unsafe locking scenario:
|
| CPU0 CPU1
| ---- ----
| lock(&cpuset_rwsem);
| lock(&device->mutex);
| lock(&cpuset_rwsem);
| lock(&mm->mmap_lock#2);
The device->mutex is nvkm_device::mutex.
Unblocking the lockchain at `cpuset_rwsem' is probably the easiest
thing to do. Move the priority assignment to the start of the newly
created thread.
Fixes: 710da3c8ea7df ("sched/core: Prevent race condition between cpuset and __sched_setscheduler()")
Reported-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
[bigeasy: Patch description]
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/a23a826af7c108ea5651e73b8fbae5e653f16e86.camel@gmx.de
|
|
With enabled threaded interrupts the nouveau driver reported the
following:
| Chain exists of:
| &mm->mmap_lock#2 --> &device->mutex --> &cpuset_rwsem
|
| Possible unsafe locking scenario:
|
| CPU0 CPU1
| ---- ----
| lock(&cpuset_rwsem);
| lock(&device->mutex);
| lock(&cpuset_rwsem);
| lock(&mm->mmap_lock#2);
The device->mutex is nvkm_device::mutex.
Unblocking the lockchain at `cpuset_rwsem' is probably the easiest
thing to do. Move the priority reset to the start of the newly
created thread.
Fixes: 710da3c8ea7df ("sched/core: Prevent race condition between cpuset and __sched_setscheduler()")
Reported-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/a23a826af7c108ea5651e73b8fbae5e653f16e86.camel@gmx.de
|
|
rcutorture is generating some nesting scenarios that are not compatible on PREEMPT_RT.
For example:
preempt_disable();
rcu_read_lock_bh();
preempt_enable();
rcu_read_unlock_bh();
The problem here is that on PREEMPT_RT the bottom halves have to be
disabled and enabled in preemptible context.
Reorder locking: start with BH locking and continue with then with
disabling preemption or interrupts. In the unlocking do it reverse by
first enabling interrupts and preemption and BH at the very end.
Ensure that on PREEMPT_RT BH locking remains unchanged if in
non-preemptible context.
Link: https://lkml.kernel.org/r/20190911165729.11178-6-swood@redhat.com
Link: https://lkml.kernel.org/r/20210819182035.GF4126399@paulmck-ThinkPad-P17-Gen-1
Signed-off-by: Scott Wood <swood@redhat.com>
[bigeasy: Drop ATOM_BH, make it only about changing BH in atomic
context. Allow enabling RCU in IRQ-off section. Reword commit message.]
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20210820074236.2zli4nje7bof62rh@linutronix.de
|
|
With PREEMPT_RT enabled all hrtimers callbacks will be invoked in
softirq mode unless they are explicitly marked as HRTIMER_MODE_HARD.
During boot kthread_bind() is used for the creation of per-CPU threads
and then hangs in wait_task_inactive() if the ksoftirqd is not
yet up and running.
The hang disappeared since commit
26c7295be0c5e ("kthread: Do not preempt current task if it is going to call schedule()")
but enabling function trace on boot reliably leads to the freeze on boot
behaviour again.
The timer in wait_task_inactive() can not be directly used by an user
interface to abuse it and create a mass wake of several tasks at the
same time which would to long sections with disabled interrupts.
Therefore it is safe to make the timer HRTIMER_MODE_REL_HARD.
Switch the timer to HRTIMER_MODE_REL_HARD.
Cc: stable-rt@vger.kernel.org
Link: https://lkml.kernel.org/r/20210826170408.vm7rlj7odslshwch@linutronix.de
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
|
|
There is msleep in pr_flush(). If call WARN() in the early boot
stage such as in early_initcall, pr_flush() will run into msleep
when process scheduler is not ready yet. And then the system will
sleep forever.
Before the system_state is SYSTEM_RUNNING, make sure DO NOT sleep
in pr_flush().
Fixes: c0b395bd0fe3("printk: add pr_flush()")
Signed-off-by: Chao Qin <chao.qin@intel.com>
Signed-off-by: Lili Li <lili.li@intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: John Ogness <john.ogness@linutronix.de>
Signed-off-by: John Ogness <john.ogness@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/lkml/20210719022649.3444072-1-chao.qin@intel.com
|
|
Provide a function to allow waiting for console printers to catch
up to the latest logged message.
Use pr_flush() to give console printers a chance to finish in
critical situations if no atomic console is available. For now
pr_flush() is only used in the most common error paths:
panic(), print_oops_end_marker(), report_bug(), kmsg_dump().
Signed-off-by: John Ogness <john.ogness@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
|
|
If earlyprintk is used, a boot console will print directly to the
console immediately. The boot console will unregister itself as soon
as a non-boot console registers. However, the non-boot console does
not begin printing until its kthread has started. Since this happens
much later, there is a long pause in the console output. If the
ringbuffer is small, messages could even be dropped during the
pause.
Add a new CON_HANDOVER console flag to be used internally by printk
in order to track which non-boot console took over from a boot
console. If handover consoles have implemented write_atomic(), they
are allowed to print directly to the console until their kthread can
take over.
Signed-off-by: John Ogness <john.ogness@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
|
|
Since printing occurs either atomically or from the printing
kthread, there is no need for any deferring or tracking possible
recursion paths. Remove all printk defer functions and context
tracking.
Signed-off-by: John Ogness <john.ogness@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
|
