sched/topology: fix the issue groups don't span domain->span for NUMA diameter > 2

As long as NUMA diameter > 2, building sched_domain by sibling's child
domain will definitely create a sched_domain with sched_group which will
span out of the sched_domain:

               +------+         +------+        +-------+       +------+
               | node |  12     |node  | 20     | node  |  12   |node  |
               |  0   +---------+1     +--------+ 2     +-------+3     |
               +------+         +------+        +-------+       +------+

domain0        node0            node1            node2          node3

domain1        node0+1          node0+1          node2+3        node2+3
                                                 +
domain2        node0+1+2                         |
             group: node0+1                      |
               group:node2+3 <-------------------+

when node2 is added into the domain2 of node0, kernel is using the child
domain of node2's domain2, which is domain1(node2+3). Node 3 is outside
the span of the domain including node0+1+2.

This will make load_balance() run based on screwed avg_load and group_type
in the sched_group spanning out of the sched_domain, and it also makes
select_task_rq_fair() pick an idle CPU outside the sched_domain.

Real servers which suffer from this problem include Kunpeng920 and 8-node
Sun Fire X4600-M2, at least.

Here we move to use the *child* domain of the *child* domain of node2's
domain2 as the new added sched_group. At the same, we re-use the lower
level sgc directly.
               +------+         +------+        +-------+       +------+
               | node |  12     |node  | 20     | node  |  12   |node  |
               |  0   +---------+1     +--------+ 2     +-------+3     |
               +------+         +------+        +-------+       +------+

domain0        node0            node1          +- node2          node3
                                               |
domain1        node0+1          node0+1        | node2+3        node2+3
                                               |
domain2        node0+1+2                       |
             group: node0+1                    |
               group:node2 <-------------------+

While the lower level sgc is re-used, this patch only changes the remote
sched_groups for those sched_domains playing grandchild trick, therefore,
sgc->next_update is still safe since it's only touched by CPUs that have
the group span as local group. And sgc->imbalance is also safe because
sd_parent remains the same in load_balance and LB only tries other CPUs
from the local group.
Moreover, since local groups are not touched, they are still getting
roughly equal size in a TL. And should_we_balance() only matters with
local groups, so the pull probability of those groups are still roughly
equal.

Tested by the below topology:
qemu-system-aarch64  -M virt -nographic \
 -smp cpus=8 \
 -numa node,cpus=0-1,nodeid=0 \
 -numa node,cpus=2-3,nodeid=1 \
 -numa node,cpus=4-5,nodeid=2 \
 -numa node,cpus=6-7,nodeid=3 \
 -numa dist,src=0,dst=1,val=12 \
 -numa dist,src=0,dst=2,val=20 \
 -numa dist,src=0,dst=3,val=22 \
 -numa dist,src=1,dst=2,val=22 \
 -numa dist,src=2,dst=3,val=12 \
 -numa dist,src=1,dst=3,val=24 \
 -m 4G -cpu cortex-a57 -kernel arch/arm64/boot/Image

w/o patch, we get lots of "groups don't span domain->span":
[    0.802139] CPU0 attaching sched-domain(s):
[    0.802193]  domain-0: span=0-1 level=MC
[    0.802443]   groups: 0:{ span=0 cap=1013 }, 1:{ span=1 cap=979 }
[    0.802693]   domain-1: span=0-3 level=NUMA
[    0.802731]    groups: 0:{ span=0-1 cap=1992 }, 2:{ span=2-3 cap=1943 }
[    0.802811]    domain-2: span=0-5 level=NUMA
[    0.802829]     groups: 0:{ span=0-3 cap=3935 }, 4:{ span=4-7 cap=3937 }
[    0.802881] ERROR: groups don't span domain->span
[    0.803058]     domain-3: span=0-7 level=NUMA
[    0.803080]      groups: 0:{ span=0-5 mask=0-1 cap=5843 }, 6:{ span=4-7 mask=6-7 cap=4077 }
[    0.804055] CPU1 attaching sched-domain(s):
[    0.804072]  domain-0: span=0-1 level=MC
[    0.804096]   groups: 1:{ span=1 cap=979 }, 0:{ span=0 cap=1013 }
[    0.804152]   domain-1: span=0-3 level=NUMA
[    0.804170]    groups: 0:{ span=0-1 cap=1992 }, 2:{ span=2-3 cap=1943 }
[    0.804219]    domain-2: span=0-5 level=NUMA
[    0.804236]     groups: 0:{ span=0-3 cap=3935 }, 4:{ span=4-7 cap=3937 }
[    0.804302] ERROR: groups don't span domain->span
[    0.804520]     domain-3: span=0-7 level=NUMA
[    0.804546]      groups: 0:{ span=0-5 mask=0-1 cap=5843 }, 6:{ span=4-7 mask=6-7 cap=4077 }
[    0.804677] CPU2 attaching sched-domain(s):
[    0.804687]  domain-0: span=2-3 level=MC
[    0.804705]   groups: 2:{ span=2 cap=934 }, 3:{ span=3 cap=1009 }
[    0.804754]   domain-1: span=0-3 level=NUMA
[    0.804772]    groups: 2:{ span=2-3 cap=1943 }, 0:{ span=0-1 cap=1992 }
[    0.804820]    domain-2: span=0-5 level=NUMA
[    0.804836]     groups: 2:{ span=0-3 mask=2-3 cap=3991 }, 4:{ span=0-1,4-7 mask=4-5 cap=5985 }
[    0.804944] ERROR: groups don't span domain->span
[    0.805108]     domain-3: span=0-7 level=NUMA
[    0.805134]      groups: 2:{ span=0-5 mask=2-3 cap=5899 }, 6:{ span=0-1,4-7 mask=6-7 cap=6125 }
[    0.805223] CPU3 attaching sched-domain(s):
[    0.805232]  domain-0: span=2-3 level=MC
[    0.805249]   groups: 3:{ span=3 cap=1009 }, 2:{ span=2 cap=934 }
[    0.805319]   domain-1: span=0-3 level=NUMA
[    0.805336]    groups: 2:{ span=2-3 cap=1943 }, 0:{ span=0-1 cap=1992 }
[    0.805383]    domain-2: span=0-5 level=NUMA
[    0.805399]     groups: 2:{ span=0-3 mask=2-3 cap=3991 }, 4:{ span=0-1,4-7 mask=4-5 cap=5985 }
[    0.805458] ERROR: groups don't span domain->span
[    0.805605]     domain-3: span=0-7 level=NUMA
[    0.805626]      groups: 2:{ span=0-5 mask=2-3 cap=5899 }, 6:{ span=0-1,4-7 mask=6-7 cap=6125 }
[    0.805712] CPU4 attaching sched-domain(s):
[    0.805721]  domain-0: span=4-5 level=MC
[    0.805738]   groups: 4:{ span=4 cap=984 }, 5:{ span=5 cap=924 }
[    0.805787]   domain-1: span=4-7 level=NUMA
[    0.805803]    groups: 4:{ span=4-5 cap=1908 }, 6:{ span=6-7 cap=2029 }
[    0.805851]    domain-2: span=0-1,4-7 level=NUMA
[    0.805867]     groups: 4:{ span=4-7 cap=3937 }, 0:{ span=0-3 cap=3935 }
[    0.805915] ERROR: groups don't span domain->span
[    0.806108]     domain-3: span=0-7 level=NUMA
[    0.806130]      groups: 4:{ span=0-1,4-7 mask=4-5 cap=5985 }, 2:{ span=0-3 mask=2-3 cap=3991 }
[    0.806214] CPU5 attaching sched-domain(s):
[    0.806222]  domain-0: span=4-5 level=MC
[    0.806240]   groups: 5:{ span=5 cap=924 }, 4:{ span=4 cap=984 }
[    0.806841]   domain-1: span=4-7 level=NUMA
[    0.806866]    groups: 4:{ span=4-5 cap=1908 }, 6:{ span=6-7 cap=2029 }
[    0.806934]    domain-2: span=0-1,4-7 level=NUMA
[    0.806953]     groups: 4:{ span=4-7 cap=3937 }, 0:{ span=0-3 cap=3935 }
[    0.807004] ERROR: groups don't span domain->span
[    0.807312]     domain-3: span=0-7 level=NUMA
[    0.807386]      groups: 4:{ span=0-1,4-7 mask=4-5 cap=5985 }, 2:{ span=0-3 mask=2-3 cap=3991 }
[    0.807686] CPU6 attaching sched-domain(s):
[    0.807710]  domain-0: span=6-7 level=MC
[    0.807750]   groups: 6:{ span=6 cap=1017 }, 7:{ span=7 cap=1012 }
[    0.807840]   domain-1: span=4-7 level=NUMA
[    0.807870]    groups: 6:{ span=6-7 cap=2029 }, 4:{ span=4-5 cap=1908 }
[    0.807952]    domain-2: span=0-1,4-7 level=NUMA
[    0.807985]     groups: 6:{ span=4-7 mask=6-7 cap=4077 }, 0:{ span=0-5 mask=0-1 cap=5843 }
[    0.808045] ERROR: groups don't span domain->span
[    0.808257]     domain-3: span=0-7 level=NUMA
[    0.808571]      groups: 6:{ span=0-1,4-7 mask=6-7 cap=6125 }, 2:{ span=0-5 mask=2-3 cap=5899 }
[    0.808848] CPU7 attaching sched-domain(s):
[    0.808860]  domain-0: span=6-7 level=MC
[    0.808880]   groups: 7:{ span=7 cap=1012 }, 6:{ span=6 cap=1017 }
[    0.808953]   domain-1: span=4-7 level=NUMA
[    0.808974]    groups: 6:{ span=6-7 cap=2029 }, 4:{ span=4-5 cap=1908 }
[    0.809034]    domain-2: span=0-1,4-7 level=NUMA
[    0.809055]     groups: 6:{ span=4-7 mask=6-7 cap=4077 }, 0:{ span=0-5 mask=0-1 cap=5843 }
[    0.809128] ERROR: groups don't span domain->span
[    0.810361]     domain-3: span=0-7 level=NUMA
[    0.810400]      groups: 6:{ span=0-1,4-7 mask=6-7 cap=5961 }, 2:{ span=0-5 mask=2-3 cap=5903 }

w/ patch, we don't get "groups don't span domain->span" any more:
[    1.486271] CPU0 attaching sched-domain(s):
[    1.486820]  domain-0: span=0-1 level=MC
[    1.500924]   groups: 0:{ span=0 cap=980 }, 1:{ span=1 cap=994 }
[    1.515717]   domain-1: span=0-3 level=NUMA
[    1.515903]    groups: 0:{ span=0-1 cap=1974 }, 2:{ span=2-3 cap=1989 }
[    1.516989]    domain-2: span=0-5 level=NUMA
[    1.517124]     groups: 0:{ span=0-3 cap=3963 }, 4:{ span=4-5 cap=1949 }
[    1.517369]     domain-3: span=0-7 level=NUMA
[    1.517423]      groups: 0:{ span=0-5 mask=0-1 cap=5912 }, 6:{ span=4-7 mask=6-7 cap=4054 }
[    1.520027] CPU1 attaching sched-domain(s):
[    1.520097]  domain-0: span=0-1 level=MC
[    1.520184]   groups: 1:{ span=1 cap=994 }, 0:{ span=0 cap=980 }
[    1.520429]   domain-1: span=0-3 level=NUMA
[    1.520487]    groups: 0:{ span=0-1 cap=1974 }, 2:{ span=2-3 cap=1989 }
[    1.520687]    domain-2: span=0-5 level=NUMA
[    1.520744]     groups: 0:{ span=0-3 cap=3963 }, 4:{ span=4-5 cap=1949 }
[    1.520948]     domain-3: span=0-7 level=NUMA
[    1.521038]      groups: 0:{ span=0-5 mask=0-1 cap=5912 }, 6:{ span=4-7 mask=6-7 cap=4054 }
[    1.522068] CPU2 attaching sched-domain(s):
[    1.522348]  domain-0: span=2-3 level=MC
[    1.522606]   groups: 2:{ span=2 cap=1003 }, 3:{ span=3 cap=986 }
[    1.522832]   domain-1: span=0-3 level=NUMA
[    1.522885]    groups: 2:{ span=2-3 cap=1989 }, 0:{ span=0-1 cap=1974 }
[    1.523043]    domain-2: span=0-5 level=NUMA
[    1.523092]     groups: 2:{ span=0-3 mask=2-3 cap=4037 }, 4:{ span=4-5 cap=1949 }
[    1.523302]     domain-3: span=0-7 level=NUMA
[    1.523352]      groups: 2:{ span=0-5 mask=2-3 cap=5986 }, 6:{ span=0-1,4-7 mask=6-7 cap=6102 }
[    1.523748] CPU3 attaching sched-domain(s):
[    1.523774]  domain-0: span=2-3 level=MC
[    1.523825]   groups: 3:{ span=3 cap=986 }, 2:{ span=2 cap=1003 }
[    1.524009]   domain-1: span=0-3 level=NUMA
[    1.524086]    groups: 2:{ span=2-3 cap=1989 }, 0:{ span=0-1 cap=1974 }
[    1.524281]    domain-2: span=0-5 level=NUMA
[    1.524331]     groups: 2:{ span=0-3 mask=2-3 cap=4037 }, 4:{ span=4-5 cap=1949 }
[    1.524534]     domain-3: span=0-7 level=NUMA
[    1.524586]      groups: 2:{ span=0-5 mask=2-3 cap=5986 }, 6:{ span=0-1,4-7 mask=6-7 cap=6102 }
[    1.524847] CPU4 attaching sched-domain(s):
[    1.524873]  domain-0: span=4-5 level=MC
[    1.524954]   groups: 4:{ span=4 cap=958 }, 5:{ span=5 cap=991 }
[    1.525105]   domain-1: span=4-7 level=NUMA
[    1.525153]    groups: 4:{ span=4-5 cap=1949 }, 6:{ span=6-7 cap=2006 }
[    1.525368]    domain-2: span=0-1,4-7 level=NUMA
[    1.525428]     groups: 4:{ span=4-7 cap=3955 }, 0:{ span=0-1 cap=1974 }
[    1.532726]     domain-3: span=0-7 level=NUMA
[    1.532811]      groups: 4:{ span=0-1,4-7 mask=4-5 cap=6003 }, 2:{ span=0-3 mask=2-3 cap=4037 }
[    1.534125] CPU5 attaching sched-domain(s):
[    1.534159]  domain-0: span=4-5 level=MC
[    1.534303]   groups: 5:{ span=5 cap=991 }, 4:{ span=4 cap=958 }
[    1.534490]   domain-1: span=4-7 level=NUMA
[    1.534572]    groups: 4:{ span=4-5 cap=1949 }, 6:{ span=6-7 cap=2006 }
[    1.534734]    domain-2: span=0-1,4-7 level=NUMA
[    1.534783]     groups: 4:{ span=4-7 cap=3955 }, 0:{ span=0-1 cap=1974 }
[    1.536057]     domain-3: span=0-7 level=NUMA
[    1.536430]      groups: 4:{ span=0-1,4-7 mask=4-5 cap=6003 }, 2:{ span=0-3 mask=2-3 cap=3896 }
[    1.536815] CPU6 attaching sched-domain(s):
[    1.536846]  domain-0: span=6-7 level=MC
[    1.536934]   groups: 6:{ span=6 cap=1005 }, 7:{ span=7 cap=1001 }
[    1.537144]   domain-1: span=4-7 level=NUMA
[    1.537262]    groups: 6:{ span=6-7 cap=2006 }, 4:{ span=4-5 cap=1949 }
[    1.537553]    domain-2: span=0-1,4-7 level=NUMA
[    1.537613]     groups: 6:{ span=4-7 mask=6-7 cap=4054 }, 0:{ span=0-1 cap=1805 }
[    1.537872]     domain-3: span=0-7 level=NUMA
[    1.537998]      groups: 6:{ span=0-1,4-7 mask=6-7 cap=6102 }, 2:{ span=0-5 mask=2-3 cap=5845 }
[    1.538448] CPU7 attaching sched-domain(s):
[    1.538505]  domain-0: span=6-7 level=MC
[    1.538586]   groups: 7:{ span=7 cap=1001 }, 6:{ span=6 cap=1005 }
[    1.538746]   domain-1: span=4-7 level=NUMA
[    1.538798]    groups: 6:{ span=6-7 cap=2006 }, 4:{ span=4-5 cap=1949 }
[    1.539048]    domain-2: span=0-1,4-7 level=NUMA
[    1.539111]     groups: 6:{ span=4-7 mask=6-7 cap=4054 }, 0:{ span=0-1 cap=1805 }
[    1.539571]     domain-3: span=0-7 level=NUMA
[    1.539610]      groups: 6:{ span=0-1,4-7 mask=6-7 cap=6102 }, 2:{ span=0-5 mask=2-3 cap=5845 }

Signed-off-by: Barry Song <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Tested-by: Meelis Roos <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

cpu/hotplug: Add cpuhp_invoke_callback_range()

Factorizing and unifying cpuhp callback range invocations, especially for
the hotunplug path, where two different ways of decrementing were used. The
first one, decrements before the callback is called:

 cpuhp_thread_fun()
     state = st->state;
     st->state--;
     cpuhp_invoke_callback(state);

The second one, after:

 take_down_cpu()|cpuhp_down_callbacks()
     cpuhp_invoke_callback(st->state);
     st->state--;

This is problematic for rolling back the steps in case of error, as
depending on the decrement, the rollback will start from N or N-1. It also
makes tracing inconsistent, between steps run in the cpuhp thread and
the others.

Additionally, avoid useless cpuhp_thread_fun() loops by skipping empty
steps.

Signed-off-by: Vincent Donnefort <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

cpu/hotplug: CPUHP_BRINGUP_CPU failure exception

The atomic states (between CPUHP_AP_IDLE_DEAD and CPUHP_AP_ONLINE) are
triggered by the CPUHP_BRINGUP_CPU step. If the latter fails, no atomic
state can be rolled back.

DEAD callbacks too can't fail and disallow recovery. As a consequence,
during hotunplug, the fail injection interface should prohibit all states
from CPUHP_BRINGUP_CPU to CPUHP_ONLINE.

Signed-off-by: Vincent Donnefort <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

cpu/hotplug: Allowing to reset fail injection

Currently, the only way of resetting the fail injection is to trigger a
hotplug, hotunplug or both. This is rather annoying for testing
and, as the default value for this file is -1, it seems pretty natural to
let a user write it.

Signed-off-by: Vincent Donnefort <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched/pelt: Fix task util_est update filtering

Being called for each dequeue, util_est reduces the number of its updates
by filtering out when the EWMA signal is different from the task util_avg
by less than 1%. It is a problem for a sudden util_avg ramp-up. Due to the
decay from a previous high util_avg, EWMA might now be close enough to
the new util_avg. No update would then happen while it would leave
ue.enqueued with an out-of-date value.

Taking into consideration the two util_est members, EWMA and enqueued for
the filtering, ensures, for both, an up-to-date value.

This is for now an issue only for the trace probe that might return the
stale value. Functional-wise, it isn't a problem, as the value is always
accessed through max(enqueued, ewma).

This problem has been observed using LISA's UtilConvergence:test_means on
the sd845c board.

No regression observed with Hackbench on sd845c and Perf-bench sched pipe
on hikey/hikey960.

Signed-off-by: Vincent Donnefort <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Dietmar Eggemann <[email protected]>
Reviewed-by: Vincent Guittot <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched/fair: Fix shift-out-of-bounds in load_balance()

Syzbot reported a handful of occurrences where an sd->nr_balance_failed can
grow to much higher values than one would expect.

A successful load_balance() resets it to 0; a failed one increments
it. Once it gets to sd->cache_nice_tries + 3, this *should* trigger an
active balance, which will either set it to sd->cache_nice_tries+1 or reset
it to 0. However, in case the to-be-active-balanced task is not allowed to
run on env->dst_cpu, then the increment is done without any further
modification.

This could then be repeated ad nauseam, and would explain the absurdly high
values reported by syzbot (86, 149). VincentG noted there is value in
letting sd->cache_nice_tries grow, so the shift itself should be
fixed. That means preventing:

  """
  If the value of the right operand is negative or is greater than or equal
  to the width of the promoted left operand, the behavior is undefined.
  """

Thus we need to cap the shift exponent to
  BITS_PER_TYPE(typeof(lefthand)) - 1.

I had a look around for other similar cases via coccinelle:

  @expr@
  position pos;
  expression E1;
  expression E2;
  @@
  (
  E1 >> E2@pos
  |
  E1 >> E2@pos
  )

  @cst depends on expr@
  position pos;
  expression expr.E1;
  constant cst;
  @@
  (
  E1 >> cst@pos
  |
  E1 << cst@pos
  )

  @script:python depends on !cst@
  pos << expr.pos;
  exp << expr.E2;
  @@
  # Dirty hack to ignore constexpr
  if exp.upper() != exp:
     coccilib.report.print_report(pos[0], "Possible UB shift here")

The only other match in kernel/sched is rq_clock_thermal() which employs
sched_thermal_decay_shift, and that exponent is already capped to 10, so
that one is fine.

Fixes: 5a7f55590467 ("sched/fair: Relax constraint on task's load during load balance")
Reported-by: [email protected]
Signed-off-by: Valentin Schneider <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Link: http://lore.kernel.org/r/[email protected]

sched/fair: use lsub_positive in cpu_util_next()

The sub_positive local version is saving an explicit load-store and is
enough for the cpu_util_next() usage.

Signed-off-by: Vincent Donnefort <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Quentin Perret <[email protected]>
Reviewed-by: Dietmar Eggemann <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched/fair: Fix task utilization accountability in compute_energy()

find_energy_efficient_cpu() (feec()) computes for each perf_domain (pd) an
energy delta as follows:

  feec(task)
    for_each_pd
      base_energy = compute_energy(task, -1, pd)
        -> for_each_cpu(pd)
           -> cpu_util_next(cpu, task, -1)

      energy_delta = compute_energy(task, dst_cpu, pd)
        -> for_each_cpu(pd)
           -> cpu_util_next(cpu, task, dst_cpu)
      energy_delta -= base_energy

Then it picks the best CPU as being the one that minimizes energy_delta.

cpu_util_next() estimates the CPU utilization that would happen if the
task was placed on dst_cpu as follows:

  max(cpu_util + task_util, cpu_util_est + _task_util_est)

The task contribution to the energy delta can then be either:

  (1) _task_util_est, on a mostly idle CPU, where cpu_util is close to 0
      and _task_util_est > cpu_util.
  (2) task_util, on a mostly busy CPU, where cpu_util > _task_util_est.

  (cpu_util_est doesn't appear here. It is 0 when a CPU is idle and
   otherwise must be small enough so that feec() takes the CPU as a
   potential target for the task placement)

This is problematic for feec(), as cpu_util_next() might give an unfair
advantage to a CPU which is mostly busy (2) compared to one which is
mostly idle (1). _task_util_est being always bigger than task_util in
feec() (as the task is waking up), the task contribution to the energy
might look smaller on certain CPUs (2) and this breaks the energy
comparison.

This issue is, moreover, not sporadic. By starving idle CPUs, it keeps
their cpu_util < _task_util_est (1) while others will maintain cpu_util >
_task_util_est (2).

Fix this problem by always using max(task_util, _task_util_est) as a task
contribution to the energy (ENERGY_UTIL). The new estimated CPU
utilization for the energy would then be:

  max(cpu_util, cpu_util_est) + max(task_util, _task_util_est)

compute_energy() still needs to know which OPP would be selected if the
task would be migrated in the perf_domain (FREQUENCY_UTIL). Hence,
cpu_util_next() is still used to estimate the maximum util within the pd.

Signed-off-by: Vincent Donnefort <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Quentin Perret <[email protected]>
Reviewed-by: Dietmar Eggemann <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched/fair: Reduce the window for duplicated update

Start to update last_blocked_load_update_tick to reduce the possibility
of another cpu starting the update one more time

Signed-off-by: Vincent Guittot <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched/fair: Trigger the update of blocked load on newly idle cpu

Instead of waking up a random and already idle CPU, we can take advantage
of this_cpu being about to enter idle to run the ILB and update the
blocked load.

Signed-off-by: Vincent Guittot <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched/fair: Reorder newidle_balance pulled_task tests

Reorder the tests and skip useless ones when no load balance has been
performed and rq lock has not been released.

Signed-off-by: Vincent Guittot <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched/fair: Merge for each idle cpu loop of ILB

Remove the specific case for handling this_cpu outside for_each_cpu() loop
when running ILB. Instead we use for_each_cpu_wrap() and start with the
next cpu after this_cpu so we will continue to finish with this_cpu.

update_nohz_stats() is now used for this_cpu too and will prevents
unnecessary update. We don't need a special case for handling the update of
nohz.next_balance for this_cpu anymore because it is now handled by the
loop like others.

Signed-off-by: Vincent Guittot <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched/fair: Remove unused parameter of update_nohz_stats

idle load balance is the only user of update_nohz_stats and doesn't use
force parameter. Remove it

Signed-off-by: Vincent Guittot <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched/fair: Remove unused return of _nohz_idle_balance

The return of _nohz_idle_balance() is not used anymore so we can remove
it

Signed-off-by: Vincent Guittot <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched/fair: Remove update of blocked load from newidle_balance

newidle_balance runs with both preempt and irq disabled which prevent
local irq to run during this period. The duration for updating the
blocked load of CPUs varies according to the number of CPU cgroups
with non-decayed load and extends this critical period to an uncontrolled
level.

Remove the update from newidle_balance and trigger a normal ILB that
will take care of the update instead.

This reduces the IRQ latency from O(nr_cgroups * nr_nohz_cpus) to
O(nr_cgroups).

Signed-off-by: Vincent Guittot <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

kcov: Remove kcov include from sched.h and move it to its users.

The recent addition of in_serving_softirq() to kconv.h results in
compile failure on PREEMPT_RT because it requires
task_struct::softirq_disable_cnt. This is not available if kconv.h is
included from sched.h.

It is not needed to include kconv.h from sched.h. All but the net/ user
already include the kconv header file.

Move the include of the kconv.h header from sched.h it its users.
Additionally include sched.h from kconv.h to ensure that everything
task_struct related is available.

Signed-off-by: Sebastian Andrzej Siewior <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Acked-by: Johannes Berg <[email protected]>
Acked-by: Andrey Konovalov <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched: Simplify migration_cpu_stop()

Since, when ->stop_pending, only the stopper can uninstall
p->migration_pending. This could simplify a few ifs, because:

  (pending != NULL) => (pending == p->migration_pending)

Also, the fatty comment above affine_move_task() probably needs a bit
of gardening.

Signed-off-by: Valentin Schneider <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>

sched/membarrier: fix missing local execution of ipi_sync_rq_state()

The function sync_runqueues_membarrier_state() should copy the
membarrier state from the @mm received as parameter to each runqueue
currently running tasks using that mm.

However, the use of smp_call_function_many() skips the current runqueue,
which is unintended. Replace by a call to on_each_cpu_mask().

Fixes: 227a4aadc75b ("sched/membarrier: Fix p->mm->membarrier_state racy load")
Reported-by: Nadav Amit <[email protected]>
Signed-off-by: Mathieu Desnoyers <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Cc: [email protected] # 5.4.x+
Link: https://lore.kernel.org/r/[email protected]

sched: Simplify set_affinity_pending refcounts

Now that we have set_affinity_pending::stop_pending to indicate if a
stopper is in progress, and we have the guarantee that if that stopper
exists, it will (eventually) complete our @pending we can simplify the
refcount scheme by no longer counting the stopper thread.

Fixes: 6d337eab041d ("sched: Fix migrate_disable() vs set_cpus_allowed_ptr()")
Cc: [email protected]
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched: Fix affine_move_task() self-concurrency

Consider:

   sched_setaffinity(p, X); sched_setaffinity(p, Y);

Then the first will install p->migration_pending = &my_pending; and
issue stop_one_cpu_nowait(pending); and the second one will read
p->migration_pending and _also_ issue: stop_one_cpu_nowait(pending),
the _SAME_ @pending.

This causes stopper list corruption.

Add set_affinity_pending::stop_pending, to indicate if a stopper is in
progress.

Fixes: 6d337eab041d ("sched: Fix migrate_disable() vs set_cpus_allowed_ptr()")
Cc: [email protected]
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched: Optimize migration_cpu_stop()

When the purpose of migration_cpu_stop() is to migrate the task to
'any' valid CPU, don't migrate the task when it's already running on a
valid CPU.

Fixes: 6d337eab041d ("sched: Fix migrate_disable() vs set_cpus_allowed_ptr()")
Cc: [email protected]
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched: Collate affine_move_task() stoppers

The SCA_MIGRATE_ENABLE and task_running() cases are almost identical,
collapse them to avoid further duplication.

Fixes: 6d337eab041d ("sched: Fix migrate_disable() vs set_cpus_allowed_ptr()")
Cc: [email protected]
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched: Simplify migration_cpu_stop()

When affine_move_task() issues a migration_cpu_stop(), the purpose of
that function is to complete that @pending, not any random other
p->migration_pending that might have gotten installed since.

This realization much simplifies migration_cpu_stop() and allows
further necessary steps to fix all this as it provides the guarantee
that @pending's stopper will complete @pending (and not some random
other @pending).

Fixes: 6d337eab041d ("sched: Fix migrate_disable() vs set_cpus_allowed_ptr()")
Cc: [email protected]
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

sched: Fix migration_cpu_stop() requeueing

When affine_move_task(p) is called on a running task @p, which is not
otherwise already changing affinity, we'll first set
p->migration_pending and then do:

 stop_one_cpu(cpu_of_rq(rq), migration_cpu_stop, &arg);

This then gets us to migration_cpu_stop() running on the CPU that was
previously running our victim task @p.

If we find that our task is no longer on that runqueue (this can
happen because of a concurrent migration due to load-balance etc.),
then we'll end up at the:

} else if (dest_cpu < 1 || pending) {

branch. Which we'll take because we set pending earlier. Here we first
check if the task @p has already satisfied the affinity constraints,
if so we bail early [A]. Otherwise we'll reissue migration_cpu_stop()
onto the CPU that is now hosting our task @p:

stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
    &pending->arg, &pending->stop_work);

Except, we've never initialized pending->arg, which will be all 0s.

This then results in running migration_cpu_stop() on the next CPU with
arg->p == NULL, which gives the by now obvious result of fireworks.

The cure is to change affine_move_task() to always use pending->arg,
furthermore we can use the exact same pattern as the
SCA_MIGRATE_ENABLE case, since we'll block on the pending->done
completion anyway, no point in adding yet another completion in
stop_one_cpu().

This then gives a clear distinction between the two
migration_cpu_stop() use cases:

  - sched_exec() / migrate_task_to() : arg->pending == NULL
  - affine_move_task() : arg->pending != NULL;

And we can have it ignore p->migration_pending when !arg->pending. Any
stop work from sched_exec() / migrate_task_to() is in addition to stop
works from affine_move_task(), which will be sufficient to issue the
completion.

Fixes: 6d337eab041d ("sched: Fix migrate_disable() vs set_cpus_allowed_ptr()")
Cc: [email protected]
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Reviewed-by: Valentin Schneider <[email protected]>
Link: https://lkml.kernel.org/r/[email protected]

Linux 5.12-rc2

Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma

Pull rdma fixes from Jason Gunthorpe:
 "Nothing special here, though Bob's regression fixes for rxe would have
  made it before the rc cycle had there not been such strong winter
  weather!

   - Fix corner cases in the rxe reference counting cleanup that are
     causing regressions in blktests for SRP

   - Two kdoc fixes so W=1 is clean

   - Missing error return in error unwind for mlx5

   - Wrong lock type nesting in IB CM"

* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma:
  RDMA/rxe: Fix errant WARN_ONCE in rxe_completer()
  RDMA/rxe: Fix extra deref in rxe_rcv_mcast_pkt()
  RDMA/rxe: Fix missed IB reference counting in loopback
  RDMA/uverbs: Fix kernel-doc warning of _uverbs_alloc
  RDMA/mlx5: Set correct kernel-doc identifier
  IB/mlx5: Add missing error code
  RDMA/rxe: Fix missing kconfig dependency on CRYPTO
  RDMA/cm: Fix IRQ restore in ib_send_cm_sidr_rep

Merge tag 'gcc-plugins-v5.12-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/kees/linux

Pull gcc-plugins fixes from Kees Cook:
 "Tiny gcc-plugin fixes for v5.12-rc2. These issues are small but have
  been reported a couple times now by static analyzers, so best to get
  them fixed to reduce the noise. :)

   - Fix coding style issues (Jason Yan)"

* tag 'gcc-plugins-v5.12-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/kees/linux:
  gcc-plugins: latent_entropy: remove unneeded semicolon
  gcc-plugins: structleak: remove unneeded variable 'ret'

Merge tag 'pstore-v5.12-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/kees/linux

Pull pstore fixes from Kees Cook:

 - Rate-limit ECC warnings (Dmitry Osipenko)

 - Fix error path check for NULL (Tetsuo Handa)

* tag 'pstore-v5.12-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/kees/linux:
  pstore/ram: Rate-limit "uncorrectable error in header" message
  pstore: Fix warning in pstore_kill_sb()

Merge tag 'for-5.12/dm-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm

Pull device mapper fixes from Mike Snitzer:
 "Fix DM verity target's optional Forward Error Correction (FEC) for
  Reed-Solomon roots that are unaligned to block size"

* tag 'for-5.12/dm-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm:
  dm verity: fix FEC for RS roots unaligned to block size
  dm bufio: subtract the number of initial sectors in dm_bufio_get_device_size

Merge tag 'block-5.12-2021-03-05' of git://git.kernel.dk/linux-block

Pull block fixes from Jens Axboe:

 - NVMe fixes:
      - more device quirks (Julian Einwag, Zoltán Böszörményi, Pascal
        Terjan)
      - fix a hwmon error return (Daniel Wagner)
      - fix the keep alive timeout initialization (Martin George)
      - ensure the model_number can't be changed on a used subsystem
        (Max Gurtovoy)

 - rsxx missing -EFAULT on copy_to_user() failure (Dan)

 - rsxx remove unused linux.h include (Tian)

 - kill unused RQF_SORTED (Jean)

 - updated outdated BFQ comments (Joseph)

 - revert work-around commit for bd_size_lock, since we removed the
   offending user in this merge window (Damien)

* tag 'block-5.12-2021-03-05' of git://git.kernel.dk/linux-block:
  nvmet: model_number must be immutable once set
  nvme-fabrics: fix kato initialization
  nvme-hwmon: Return error code when registration fails
  nvme-pci: add quirks for Lexar 256GB SSD
  nvme-pci: mark Kingston SKC2000 as not supporting the deepest power state
  nvme-pci: mark Seagate Nytro XM1440 as QUIRK_NO_NS_DESC_LIST.
  rsxx: Return -EFAULT if copy_to_user() fails
  block/bfq: update comments and default value in docs for fifo_expire
  rsxx: remove unused including <linux/version.h>
  block: Drop leftover references to RQF_SORTED
  block: revert "block: fix bd_size_lock use"

Merge tag 'io_uring-5.12-2021-03-05' of git://git.kernel.dk/linux-block

Pull io_uring fixes from Jens Axboe:
 "A bit of a mix between fallout from the worker change, cleanups and
  reductions now possible from that change, and fixes in general. In
  detail:

   - Fully serialize manager and worker creation, fixing races due to
     that.

   - Clean up some naming that had gone stale.

   - SQPOLL fixes.

   - Fix race condition around task_work rework that went into this
     merge window.

   - Implement unshare. Used for when the original task does unshare(2)
     or setuid/seteuid and friends, drops the original workers and forks
     new ones.

   - Drop the only remaining piece of state shuffling we had left, which
     was cred. Move it into issue instead, and we can drop all of that
     code too.

   - Kill f_op->flush() usage. That was such a nasty hack that we had
     out of necessity, we no longer need it.

   - Following from ->flush() removal, we can also drop various bits of
     ctx state related to SQPOLL and cancelations.

   - Fix an issue with IOPOLL retry, which originally was fallout from a
     filemap change (removing iov_iter_revert()), but uncovered an issue
     with iovec re-import too late.

   - Fix an issue with system suspend.

   - Use xchg() for fallback work, instead of cmpxchg().

   - Properly destroy io-wq on exec.

   - Add create_io_thread() core helper, and use that in io-wq and
     io_uring. This allows us to remove various silly completion events
     related to thread setup.

   - A few error handling fixes.

  This should be the grunt of fixes necessary for the new workers, next
  week should be quieter. We've got a pending series from Pavel on
  cancelations, and how tasks and rings are indexed. Outside of that,
  should just be minor fixes. Even with these fixes, we're still killing
  a net ~80 lines"

* tag 'io_uring-5.12-2021-03-05' of git://git.kernel.dk/linux-block: (41 commits)
  io_uring: don't restrict issue_flags for io_openat
  io_uring: make SQPOLL thread parking saner
  io-wq: kill hashed waitqueue before manager exits
  io_uring: clear IOCB_WAITQ for non -EIOCBQUEUED return
  io_uring: don't keep looping for more events if we can't flush overflow
  io_uring: move to using create_io_thread()
  kernel: provide create_io_thread() helper
  io_uring: reliably cancel linked timeouts
  io_uring: cancel-match based on flags
  io-wq: ensure all pending work is canceled on exit
  io_uring: ensure that threads freeze on suspend
  io_uring: remove extra in_idle wake up
  io_uring: inline __io_queue_async_work()
  io_uring: inline io_req_clean_work()
  io_uring: choose right tctx->io_wq for try cancel
  io_uring: fix -EAGAIN retry with IOPOLL
  io-wq: fix error path leak of buffered write hash map
  io_uring: remove sqo_task
  io_uring: kill sqo_dead and sqo submission halting
  io_uring: ignore double poll add on the same waitqueue head
  ...

Merge tag 'pm-5.12-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm

Pull power management fixes from Rafael Wysocki:
 "These fix the usage of device links in the runtime PM core code and
  update the DTPM (Dynamic Thermal Power Management) feature added
  recently.

  Specifics:

   - Make the runtime PM core code avoid attempting to suspend supplier
     devices before updating the PM-runtime status of a consumer to
     'suspended' (Rafael Wysocki).

   - Fix DTPM (Dynamic Thermal Power Management) root node
     initialization and label that feature as EXPERIMENTAL in Kconfig
     (Daniel Lezcano)"

* tag 'pm-5.12-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
  powercap/drivers/dtpm: Add the experimental label to the option description
  powercap/drivers/dtpm: Fix root node initialization
  PM: runtime: Update device status before letting suppliers suspend

Merge tag 'acpi-5.12-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm

Pull ACPI fix from Rafael Wysocki:
 "Make the empty stubs of some helper functions used when CONFIG_ACPI is
  not set actually match those functions (Andy Shevchenko)"

* tag 'acpi-5.12-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
  ACPI: bus: Constify is_acpi_node() and friends (part 2)

Merge tag 'iommu-fixes-v5.12-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/joro/iommu

Pull iommu fixes from Joerg Roedel:

 - Fix a sleeping-while-atomic issue in the AMD IOMMU code

 - Disable lazy IOTLB flush for untrusted devices in the Intel VT-d
   driver

 - Fix status code definitions for Intel VT-d

 - Fix IO Page Fault issue in Tegra IOMMU driver

* tag 'iommu-fixes-v5.12-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/joro/iommu:
  iommu/vt-d: Fix status code for Allocate/Free PASID command
  iommu: Don't use lazy flush for untrusted device
  iommu/tegra-smmu: Fix mc errors on tegra124-nyan
  iommu/amd: Fix sleeping in atomic in increase_address_space()

Merge tag 'for-5.12-rc1-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux

Pull btrfs fixes from David Sterba:
 "More regression fixes and stabilization.

  Regressions:

   - zoned mode
      - count zone sizes in wider int types
      - fix space accounting for read-only block groups

   - subpage: fix page tail zeroing

  Fixes:

   - fix spurious warning when remounting with free space tree

   - fix warning when creating a directory with smack enabled

   - ioctl checks for qgroup inheritance when creating a snapshot

   - qgroup
      - fix missing unlock on error path in zero range
      - fix amount of released reservation on error
      - fix flushing from unsafe context with open transaction,
        potentially deadlocking

   - minor build warning fixes"

* tag 'for-5.12-rc1-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux:
  btrfs: zoned: do not account freed region of read-only block group as zone_unusable
  btrfs: zoned: use sector_t for zone sectors
  btrfs: subpage: fix the false data csum mismatch error
  btrfs: fix warning when creating a directory with smack enabled
  btrfs: don't flush from btrfs_delayed_inode_reserve_metadata
  btrfs: export and rename qgroup_reserve_meta
  btrfs: free correct amount of space in btrfs_delayed_inode_reserve_metadata
  btrfs: fix spurious free_space_tree remount warning
  btrfs: validate qgroup inherit for SNAP_CREATE_V2 ioctl
  btrfs: unlock extents in btrfs_zero_range in case of quota reservation errors
  btrfs: ref-verify: use 'inline void' keyword ordering

Merge tag 'devicetree-fixes-for-5.12-1' of git://git.kernel.org/pub/scm/linux/kernel/git/robh/linux

Pull devicetree fixes from Rob Herring:

 - Another batch of graph and video-interfaces schema conversions

 - Drop DT header symlink for dropped C6X arch

 - Fix bcm2711-hdmi schema error

* tag 'devicetree-fixes-for-5.12-1' of git://git.kernel.org/pub/scm/linux/kernel/git/robh/linux:
  dt-bindings: media: Use graph and video-interfaces schemas, round 2
  dts: drop dangling c6x symlink
  dt-bindings: bcm2711-hdmi: Fix broken schema

Merge tag 'trace-v5.12-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace

Pull tracing fixes from Steven Rostedt:
 "Functional fixes:

   - Fix big endian conversion for arm64 in recordmcount processing

   - Fix timestamp corruption in ring buffer on discarding events

   - Fix memory leak in __create_synth_event()

   - Skip selftests if tracing is disabled as it will cause them to
     fail.

  Non-functional fixes:

   - Fix help text in Kconfig

   - Remove duplicate prototype for trace_empty()

   - Fix stale comment about the trace_event_call flags.

  Self test update:

   - Add more information to the validation output of when a corrupt
     timestamp is found in the ring buffer, and also trigger a warning
     to make sure that tests catch it"

* tag 'trace-v5.12-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace:
  tracing: Fix comment about the trace_event_call flags
  tracing: Skip selftests if tracing is disabled
  tracing: Fix memory leak in __create_synth_event()
  ring-buffer: Add a little more information and a WARN when time stamp going backwards is detected
  ring-buffer: Force before_stamp and write_stamp to be different on discard
  tracing: Fix help text of TRACEPOINT_BENCHMARK in Kconfig
  tracing: Remove duplicate declaration from trace.h
  ftrace: Have recordmcount use w8 to read relp->r_info in arm64_is_fake_mcount

RDMA/rxe: Fix errant WARN_ONCE in rxe_completer()

In rxe_comp.c in rxe_completer() the function free_pkt() did not clear skb
which triggered a warning at 'done:' and could possibly at 'exit:'. The
WARN_ONCE() calls are not actually needed.  The call to free_pkt() is
moved to the end to clearly show that all skbs are freed.

Fixes: 899aba891cab ("RDMA/rxe: Fix FIXME in rxe_udp_encap_recv()")
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Bob Pearson <[email protected]>
Signed-off-by: Jason Gunthorpe <[email protected]>

RDMA/rxe: Fix extra deref in rxe_rcv_mcast_pkt()

rxe_rcv_mcast_pkt() dropped a reference to ib_device when no error
occurred causing an underflow on the reference counter.  This code is
cleaned up to be clearer and easier to read.

Fixes: 899aba891cab ("RDMA/rxe: Fix FIXME in rxe_udp_encap_recv()")
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Bob Pearson <[email protected]>
Signed-off-by: Jason Gunthorpe <[email protected]>

RDMA/rxe: Fix missed IB reference counting in loopback

When the noted patch below extending the reference taken by
rxe_get_dev_from_net() in rxe_udp_encap_recv() until each skb is freed it
was not matched by a reference in the loopback path resulting in
underflows.

Fixes: 899aba891cab ("RDMA/rxe: Fix FIXME in rxe_udp_encap_recv()")
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Bob Pearson <[email protected]>
Signed-off-by: Jason Gunthorpe <[email protected]>

io_uring: don't restrict issue_flags for io_openat

45d189c606292 ("io_uring: replace force_nonblock with flags") did
something strange for io_openat() slicing all issue_flags but
IO_URING_F_NONBLOCK. Not a bug for now, but better to just forward the
flags.

Signed-off-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

Merge tag 'nvme-5.12-2021-03-05' of git://git.infradead.org/nvme into block-5.12

Pull NVMe fixes from Christoph:

"nvme fixes for 5.12:

 - more device quirks (Julian Einwag, Zoltán Böszörményi, Pascal Terjan)
 - fix a hwmon error return (Daniel Wagner)
 - fix the keep alive timeout initialization (Martin George)
 - ensure the model_number can't be changed on a used subsystem
   (Max Gurtovoy)"

* tag 'nvme-5.12-2021-03-05' of git://git.infradead.org/nvme:
  nvmet: model_number must be immutable once set
  nvme-fabrics: fix kato initialization
  nvme-hwmon: Return error code when registration fails
  nvme-pci: add quirks for Lexar 256GB SSD
  nvme-pci: mark Kingston SKC2000 as not supporting the deepest power state
  nvme-pci: mark Seagate Nytro XM1440 as QUIRK_NO_NS_DESC_LIST.

io_uring: make SQPOLL thread parking saner

We have this weird true/false return from parking, and then some of the
callers decide to look at that. It can lead to unbalanced parks and
sqd locking. Have the callers check the thread status once it's parked.
We know we have the lock at that point, so it's either valid or it's NULL.

Fix race with parking on thread exit. We need to be careful here with
ordering of the sdq->lock and the IO_SQ_THREAD_SHOULD_PARK bit.

Rename sqd->completion to sqd->parked to reflect that this is the only
thing this completion event doesn.

Signed-off-by: Jens Axboe <[email protected]>

io-wq: kill hashed waitqueue before manager exits

If we race with shutting down the io-wq context and someone queueing
a hashed entry, then we can exit the manager with it armed. If it then
triggers after the manager has exited, we can have a use-after-free where
io_wqe_hash_wake() attempts to wake a now gone manager process.

Move the killing of the hashed write queue into the manager itself, so
that we know we've killed it before the task exits.

Fixes: e941894eae31 ("io-wq: make buffered file write hashed work map per-ctx")
Signed-off-by: Jens Axboe <[email protected]>

io_uring: clear IOCB_WAITQ for non -EIOCBQUEUED return

The callback can only be armed, if we get -EIOCBQUEUED returned. It's
important that we clear the WAITQ bit for other cases, otherwise we can
queue for async retry and filemap will assume that we're armed and
return -EAGAIN instead of just blocking for the IO.

Cc: [email protected] # 5.9+
Signed-off-by: Jens Axboe <[email protected]>

io_uring: don't keep looping for more events if we can't flush overflow

It doesn't make sense to wait for more events to come in, if we can't
even flush the overflow we already have to the ring. Return -EBUSY for
that condition, just like we do for attempts to submit with overflow
pending.

Cc: [email protected] # 5.11
Signed-off-by: Jens Axboe <[email protected]>

io_uring: move to using create_io_thread()

This allows us to do task creation and setup without needing to use
completions to try and synchronize with the starting thread. Get rid of
the old io_wq_fork_thread() wrapper, and the 'wq' and 'worker' startup
completion events - we can now do setup before the task is running.

Signed-off-by: Jens Axboe <[email protected]>

Merge branch 'powercap'

* powercap:
  powercap/drivers/dtpm: Add the experimental label to the option description
  powercap/drivers/dtpm: Fix root node initialization

nvmet: model_number must be immutable once set

In case we have already established connection to nvmf target, it
shouldn't be allowed to change the model_number. E.g. if someone will
identify ctrl and get model_number of "my_model" later on will change
the model_numbel via configfs to "my_new_model" this will break the NVMe
specification for "Get Log Page – Persistent Event Log" that refers to
Model Number as: "This field contains the same value as reported in the
Model Number field of the Identify Controller data structure, bytes
63:24."

Although it doesn't mentioned explicitly that this field can't be
changed, we can assume it.

So allow setting this field only once: using configfs or in the first
identify ctrl operation.

Signed-off-by: Max Gurtovoy <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>

nvme-fabrics: fix kato initialization

Currently kato is initialized to NVME_DEFAULT_KATO for both
discovery & i/o controllers. This is a problem specifically
for non-persistent discovery controllers since it always ends
up with a non-zero kato value. Fix this by initializing kato
to zero instead, and ensuring various controllers are assigned
appropriate kato values as follows:

non-persistent controllers  - kato set to zero
persistent controllers      - kato set to NVMF_DEV_DISC_TMO
                              (or any positive int via nvme-cli)
i/o controllers             - kato set to NVME_DEFAULT_KATO
                              (or any positive int via nvme-cli)

Signed-off-by: Martin George <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>

nvme-hwmon: Return error code when registration fails

The hwmon pointer wont be NULL if the registration fails. Though the
exit code path will assign it to ctrl->hwmon_device. Later
nvme_hwmon_exit() will try to free the invalid pointer. Avoid this by
returning the error code from hwmon_device_register_with_info().

Fixes: ed7770f66286 ("nvme/hwmon: rework to avoid devm allocation")
Signed-off-by: Daniel Wagner <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>

nvme-pci: add quirks for Lexar 256GB SSD

Add the NVME_QUIRK_NO_NS_DESC_LIST and NVME_QUIRK_IGNORE_DEV_SUBNQN
quirks for this buggy device.

Reported and tested in https://bugs.mageia.org/show_bug.cgi?id=28417

Signed-off-by: Pascal Terjan <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>

nvme-pci: mark Kingston SKC2000 as not supporting the deepest power state

My 2TB SKC2000 showed the exact same symptoms that were provided
in 538e4a8c57 ("nvme-pci: avoid the deepest sleep state on
Kingston A2000 SSDs"), i.e. a complete NVME lockup that needed
cold boot to get it back.

According to some sources, the A2000 is simply a rebadged
SKC2000 with a slightly optimized firmware.

Adding the SKC2000 PCI ID to the quirk list with the same workaround
as the A2000 made my laptop survive a 5 hours long Yocto bootstrap
buildfest which reliably triggered the SSD lockup previously.

Signed-off-by: Zoltán Böszörményi <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>

nvme-pci: mark Seagate Nytro XM1440 as QUIRK_NO_NS_DESC_LIST.

The kernel fails to fully detect these SSDs, only the character devices
are present:

[   10.785605] nvme nvme0: pci function 0000:04:00.0
[   10.876787] nvme nvme1: pci function 0000:81:00.0
[   13.198614] nvme nvme0: missing or invalid SUBNQN field.
[   13.198658] nvme nvme1: missing or invalid SUBNQN field.
[   13.206896] nvme nvme0: Shutdown timeout set to 20 seconds
[   13.215035] nvme nvme1: Shutdown timeout set to 20 seconds
[   13.225407] nvme nvme0: 16/0/0 default/read/poll queues
[   13.233602] nvme nvme1: 16/0/0 default/read/poll queues
[   13.239627] nvme nvme0: Identify Descriptors failed (8194)
[   13.246315] nvme nvme1: Identify Descriptors failed (8194)

Adding the NVME_QUIRK_NO_NS_DESC_LIST fixes this problem.

BugLink: https://bugzilla.kernel.org/show_bug.cgi?id=205679
Signed-off-by: Julian Einwag <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
Reviewed-by: Keith Busch <[email protected]>

Merge tag 'drm-fixes-2021-03-05' of git://anongit.freedesktop.org/drm/drm

Pull drm fixes from Dave Airlie:
 "More may show up but this is what I have at this stage: just a single
  nouveau regression fix, and a bunch of amdgpu fixes.

  amdgpu:
   - S0ix fix
   - Handle new NV12 SKU
   - Misc power fixes
   - Display uninitialized value fix
   - PCIE debugfs register access fix

  nouveau:
   - regression fix for gk104"

* tag 'drm-fixes-2021-03-05' of git://anongit.freedesktop.org/drm/drm:
  drm/amdgpu: fix parameter error of RREG32_PCIE() in amdgpu_regs_pcie
  drm/amd/display: fix the return of the uninitialized value in ret
  drm/amdgpu: enable BACO runpm by default on sienna cichlid and navy flounder
  drm/amd/pm: correct Arcturus mmTHM_BACO_CNTL register address
  drm/amdgpu/swsmu/vangogh: Only use RLCPowerNotify msg for disable
  drm/amdgpu/pm: make unsupported power profile messages debug
  drm/amdgpu:disable VCN for Navi12 SKU
  drm/amdgpu: Only check for S0ix if AMD_PMC is configured
  drm/nouveau/fifo/gk104-gp1xx: fix creation of sw class

Merge tag 'mkp-scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/mkp/scsi

Pull iSCSI fixes from Martin Petersen:
 "Three fixes for missed iSCSI verification checks (and make the sysfs
  files use "sysfs_emit()" - that's what it is there for)"

* tag 'mkp-scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/mkp/scsi:
  scsi: iscsi: Verify lengths on passthrough PDUs
  scsi: iscsi: Ensure sysfs attributes are limited to PAGE_SIZE
  scsi: iscsi: Restrict sessions and handles to admin capabilities

Merge tag 'amd-drm-fixes-5.12-2021-03-03' of https://gitlab.freedesktop.org/agd5f/linux into drm-fixes

amd-drm-fixes-5.12-2021-03-03:

amdgpu:
- S0ix fix
- Handle new NV12 SKU
- Misc power fixes
- Display uninitialized value fix
- PCIE debugfs register access fix

Signed-off-by: Dave Airlie <[email protected]>
From: Alex Deucher <[email protected]>
Link: https://patchwork.freedesktop.org/patch/msgid/[email protected]

Merge branch '00.00-inst' of git://github.com/skeggsb/linux into drm-fixes

A single regression fix here that I noticed while testing a bunch of
boards for something else, not sure where this got lost!  Prevents 3D
driver from initialising on some GPUs.

Signed-off-by: Dave Airlie <[email protected]>
From: Ben Skeggs <[email protected]>
Link: https://patchwork.freedesktop.org/patch/msgid/CACAvsv5gmq14BrDmkMncfd=tHVSSaU89BdBEWfs6Jy-aRz03GQ@mail.gmail.com

scsi: iscsi: Verify lengths on passthrough PDUs

Open-iSCSI sends passthrough PDUs over netlink, but the kernel should be
verifying that the provided PDU header and data lengths fall within the
netlink message to prevent accessing beyond that in memory.

Cc: [email protected]
Reported-by: Adam Nichols <[email protected]>
Reviewed-by: Lee Duncan <[email protected]>
Reviewed-by: Mike Christie <[email protected]>
Signed-off-by: Chris Leech <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>

scsi: iscsi: Ensure sysfs attributes are limited to PAGE_SIZE

As the iSCSI parameters are exported back through sysfs, it should be
enforcing that they never are more than PAGE_SIZE (which should be more
than enough) before accepting updates through netlink.

Change all iSCSI sysfs attributes to use sysfs_emit().

Cc: [email protected]
Reported-by: Adam Nichols <[email protected]>
Reviewed-by: Lee Duncan <[email protected]>
Reviewed-by: Greg Kroah-Hartman <[email protected]>
Reviewed-by: Mike Christie <[email protected]>
Signed-off-by: Chris Leech <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>

scsi: iscsi: Restrict sessions and handles to admin capabilities

Protect the iSCSI transport handle, available in sysfs, by requiring
CAP_SYS_ADMIN to read it. Also protect the netlink socket by restricting
reception of messages to ones sent with CAP_SYS_ADMIN. This disables
normal users from being able to end arbitrary iSCSI sessions.

Cc: [email protected]
Reported-by: Adam Nichols <[email protected]>
Reviewed-by: Chris Leech <[email protected]>
Reviewed-by: Mike Christie <[email protected]>
Signed-off-by: Lee Duncan <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>

kernel: provide create_io_thread() helper

Provide a generic helper for setting up an io_uring worker. Returns a
task_struct so that the caller can do whatever setup is needed, then call
wake_up_new_task() to kick it into gear.

Add a kernel_clone_args member, io_thread, which tells copy_process() to
mark the task with PF_IO_WORKER.

Signed-off-by: Jens Axboe <[email protected]>

io_uring: reliably cancel linked timeouts

Linked timeouts are fired asynchronously (i.e. soft-irq), and use
generic cancellation paths to do its stuff, including poking into io-wq.
The problem is that it's racy to access tctx->io_wq, as
io_uring_task_cancel() and others may be happening at this exact moment.
Mark linked timeouts with REQ_F_INLIFGHT for now, making sure there are
no timeouts before io-wq destraction.

Signed-off-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

io_uring: cancel-match based on flags

Instead of going into request internals, like checking req->file->f_op,
do match them based on REQ_F_INFLIGHT, it's set only when we want it to
be reliably cancelled.

Signed-off-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

Merge tag 'tpmdd-next-v5.12-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/jarkko/linux-tpmdd

Pull tpm fixes Jarkko Sakkinen:
 "Three fixes for rc2"

* tag 'tpmdd-next-v5.12-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/jarkko/linux-tpmdd:
  tpm: Remove unintentional dump_stack() call
  tpm, tpm_tis: Decorate tpm_tis_gen_interrupt() with request_locality()
  tpm, tpm_tis: Decorate tpm_get_timeouts() with request_locality()

dm verity: fix FEC for RS roots unaligned to block size

Optional Forward Error Correction (FEC) code in dm-verity uses
Reed-Solomon code and should support roots from 2 to 24.

The error correction parity bytes (of roots lengths per RS block) are
stored on a separate device in sequence without any padding.

Currently, to access FEC device, the dm-verity-fec code uses dm-bufio
client with block size set to verity data block (usually 4096 or 512
bytes).

Because this block size is not divisible by some (most!) of the roots
supported lengths, data repair cannot work for partially stored parity
bytes.

This fix changes FEC device dm-bufio block size to "roots << SECTOR_SHIFT"
where we can be sure that the full parity data is always available.
(There cannot be partial FEC blocks because parity must cover whole
sectors.)

Because the optional FEC starting offset could be unaligned to this
new block size, we have to use dm_bufio_set_sector_offset() to
configure it.

The problem is easily reproduced using veritysetup, e.g. for roots=13:

  # create verity device with RS FEC
  dd if=/dev/urandom of=data.img bs=4096 count=8 status=none
  veritysetup format data.img hash.img --fec-device=fec.img --fec-roots=13 | awk '/^Root hash/{ print $3 }' >roothash

  # create an erasure that should be always repairable with this roots setting
  dd if=/dev/zero of=data.img conv=notrunc bs=1 count=8 seek=4088 status=none

  # try to read it through dm-verity
  veritysetup open data.img test hash.img --fec-device=fec.img --fec-roots=13 $(cat roothash)
  dd if=/dev/mapper/test of=/dev/null bs=4096 status=noxfer
  # wait for possible recursive recovery in kernel
  udevadm settle
  veritysetup close test

With this fix, errors are properly repaired.
  device-mapper: verity-fec: 7:1: FEC 0: corrected 8 errors
  ...

Without it, FEC code usually ends on unrecoverable failure in RS decoder:
  device-mapper: verity-fec: 7:1: FEC 0: failed to correct: -74
  ...

This problem is present in all kernels since the FEC code's
introduction (kernel 4.5).

It is thought that this problem is not visible in Android ecosystem
because it always uses a default RS roots=2.

Depends-on: a14e5ec66a7a ("dm bufio: subtract the number of initial sectors in dm_bufio_get_device_size")
Signed-off-by: Milan Broz <[email protected]>
Tested-by: Jérôme Carretero <[email protected]>
Reviewed-by: Sami Tolvanen <[email protected]>
Cc: [email protected] # 4.5+
Signed-off-by: Mike Snitzer <[email protected]>

dm bufio: subtract the number of initial sectors in dm_bufio_get_device_size

dm_bufio_get_device_size returns the device size in blocks. Before
returning the value, we must subtract the nubmer of starting
sectors. The number of starting sectors may not be divisible by block
size.

Note that currently, no target is using dm_bufio_set_sector_offset and
dm_bufio_get_device_size simultaneously, so this change has no effect.
However, an upcoming dm-verity-fec fix needs this change.

Signed-off-by: Mikulas Patocka <[email protected]>
Reviewed-by: Milan Broz <[email protected]>
Cc: [email protected]
Signed-off-by: Mike Snitzer <[email protected]>

Merge tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm

Pull KVM fixes from Paolo Bonzini:

 - Doc fixes

 - selftests fixes

 - Add runstate information to the new Xen support

 - Allow compiling out the Xen interface

 - 32-bit PAE without EPT bugfix

 - NULL pointer dereference bugfix

* tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm:
  KVM: SVM: Clear the CR4 register on reset
  KVM: x86/xen: Add support for vCPU runstate information
  KVM: x86/xen: Fix return code when clearing vcpu_info and vcpu_time_info
  selftests: kvm: Mmap the entire vcpu mmap area
  KVM: Documentation: Fix index for KVM_CAP_PPC_DAWR1
  KVM: x86: allow compiling out the Xen hypercall interface
  KVM: xen: flush deferred static key before checking it
  KVM: x86/mmu: Set SPTE_AD_WRPROT_ONLY_MASK if and only if PML is enabled
  KVM: x86: hyper-v: Fix Hyper-V context null-ptr-deref
  KVM: x86: remove misplaced comment on active_mmu_pages
  KVM: Documentation: rectify rst markup in kvm_run->flags
  Documentation: kvm: fix messy conversion from .txt to .rst

Merge tag 'for-linus-5.12b-rc2-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/xen/tip

Pull xen fixes from Juergen Gross:
 "Two security issues (XSA-367 and XSA-369)"

* tag 'for-linus-5.12b-rc2-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/xen/tip:
  xen: fix p2m size in dom0 for disabled memory hotplug case
  xen-netback: respect gnttab_map_refs()'s return value
  Xen/gnttab: handle p2m update errors on a per-slot basis

Merge tag 'sound-5.12-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai/sound

Pull sound fixes from Takashi Iwai:
 "Slightly bulky changes are seen at this time, mostly for dealing with
  the messed up Kconfig for ASoC Intel SOF stuff. The driver and its
  code was split to each module per platform now, which is far more
  straightforward. This should cover the randconfig problems, and more
  importantly, improve the actual device handling as well.

  Other than that, nothing particular stands out: the HDMI PCM
  assignment fix for Intel Tigerlake, MIPS n64 error handling fix, and
  the usual suspects, HD-audio / USB-audio quirks"

* tag 'sound-5.12-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai/sound: (21 commits)
  ALSA: hda/realtek: Apply dual codec quirks for MSI Godlike X570 board
  ALSA: hda/realtek: Add quirk for Intel NUC 10
  ALSA: hda/hdmi: let new platforms assign the pcm slot dynamically
  ALSA: hda/realtek: Add quirk for Clevo NH55RZQ
  ALSA: hda: intel-sdw-acpi: add missing include files
  ALSA: hda: move Intel SoundWire ACPI scan to dedicated module
  ASoC: SOF: Intel: SoundWire: simplify Kconfig
  ASoC: SOF: pci: move DSP_CONFIG use to platform-specific drivers
  ASoC: SOF: pci: split PCI into different drivers
  ASoC: SOF: ACPI: avoid reverse module dependency
  ASoC: soc-acpi: allow for partial match in parent name
  ALSA: hda: intel-nhlt: verify config type
  ALSA: hda: fix kernel-doc warnings
  ALSA: usb-audio: Fix Pioneer DJM devices URB_CONTROL request direction to set samplerate
  ALSA: usb-audio: use Corsair Virtuoso mapping for Corsair Virtuoso SE
  ALSA: hda/realtek: Enable headset mic of Acer SWIFT with ALC256
  ALSA: ctxfi: cthw20k2: fix mask on conf to allow 4 bits
  ALSA: usb-audio: Allow modifying parameters with succeeding hw_params calls
  ALSA: usb-audio: Drop bogus dB range in too low level
  ALSA: usb-audio: Don't abort even if the clock rate differs
  ...

btrfs: zoned: do not account freed region of read-only block group as zone_unusable

We migrate zone unusable bytes to read-only bytes when a block group is
set to read-only, and account all the free region as bytes_readonly.
Thus, we should not increase block_group->zone_unusable when the block
group is read-only.

Fixes: 169e0da91a21 ("btrfs: zoned: track unusable bytes for zones")
Reviewed-by: Johannes Thumshirn <[email protected]>
Signed-off-by: Naohiro Aota <[email protected]>
Signed-off-by: David Sterba <[email protected]>

btrfs: zoned: use sector_t for zone sectors

We need to use sector_t for zone_sectors, or it would set the zone size
to zero when the size >= 4GB (= 2^24 sectors) by shifting the
zone_sectors value by SECTOR_SHIFT. We're assuming zones sizes up to
8GiB.

Fixes: 5b316468983d ("btrfs: get zone information of zoned block devices")
Reviewed-by: Johannes Thumshirn <[email protected]>
Signed-off-by: Naohiro Aota <[email protected]>
Reviewed-by: David Sterba <[email protected]>
Signed-off-by: David Sterba <[email protected]>

tracing: Fix comment about the trace_event_call flags

In the declaration of the struct trace_event_call, the flags has the bits
defined in the comment above it. But these bits are also defined by the
TRACE_EVENT_FL_* enums just above the declaration of the struct. As the
comment about the flags in the struct has become stale and incorrect, just
replace it with a reference to the TRACE_EVENT_FL_* enum above.

Signed-off-by: Steven Rostedt (VMware) <[email protected]>

tracing: Skip selftests if tracing is disabled

If tracing is disabled for some reason (traceoff_on_warning, command line,
etc), the ftrace selftests are guaranteed to fail, as their results are
defined by trace data in the ring buffers. If the ring buffers are turned
off, the tests will fail, due to lack of data.

Because tracing being disabled is for a specific reason (warning, user
decided to, etc), it does not make sense to enable tracing to run the self
tests, as the test output may corrupt the reason for the tracing to be
disabled.

Instead, simply skip the self tests and report that they are being skipped
due to tracing being disabled.

Signed-off-by: Steven Rostedt (VMware) <[email protected]>

tracing: Fix memory leak in __create_synth_event()

kmemleak report:
unreferenced object 0xc5a6f708 (size 8):
  comm "ftracetest", pid 1209, jiffies 4294911500 (age 6.816s)
  hex dump (first 8 bytes):
    00 c1 3d 60 14 83 1f 8a                          ..=`....
  backtrace:
    [<f0aa4ac4>] __kmalloc_track_caller+0x2a6/0x460
    [<7d3d60a6>] kstrndup+0x37/0x70
    [<45a0e739>] argv_split+0x1c/0x120
    [<c17982f8>] __create_synth_event+0x192/0xb00
    [<0708b8a3>] create_synth_event+0xbb/0x150
    [<3d1941e1>] create_dyn_event+0x5c/0xb0
    [<5cf8b9e3>] trace_parse_run_command+0xa7/0x140
    [<04deb2ef>] dyn_event_write+0x10/0x20
    [<8779ac95>] vfs_write+0xa9/0x3c0
    [<ed93722a>] ksys_write+0x89/0xc0
    [<b9ca0507>] __ia32_sys_write+0x15/0x20
    [<7ce02d85>] __do_fast_syscall_32+0x45/0x80
    [<cb0ecb35>] do_fast_syscall_32+0x29/0x60
    [<2467454a>] do_SYSENTER_32+0x15/0x20
    [<9beaa61d>] entry_SYSENTER_32+0xa9/0xfc
unreferenced object 0xc5a6f078 (size 8):
  comm "ftracetest", pid 1209, jiffies 4294911500 (age 6.816s)
  hex dump (first 8 bytes):
    08 f7 a6 c5 00 00 00 00                          ........
  backtrace:
    [<bbac096a>] __kmalloc+0x2b6/0x470
    [<aa2624b4>] argv_split+0x82/0x120
    [<c17982f8>] __create_synth_event+0x192/0xb00
    [<0708b8a3>] create_synth_event+0xbb/0x150
    [<3d1941e1>] create_dyn_event+0x5c/0xb0
    [<5cf8b9e3>] trace_parse_run_command+0xa7/0x140
    [<04deb2ef>] dyn_event_write+0x10/0x20
    [<8779ac95>] vfs_write+0xa9/0x3c0
    [<ed93722a>] ksys_write+0x89/0xc0
    [<b9ca0507>] __ia32_sys_write+0x15/0x20
    [<7ce02d85>] __do_fast_syscall_32+0x45/0x80
    [<cb0ecb35>] do_fast_syscall_32+0x29/0x60
    [<2467454a>] do_SYSENTER_32+0x15/0x20
    [<9beaa61d>] entry_SYSENTER_32+0xa9/0xfc

In __create_synth_event(), while iterating field/type arguments, the
argv_split() will return array of atleast 2 elements even when zero
arguments(argc=0) are passed. for e.g. when there is double delimiter
or string ends with delimiter

To fix call argv_free() even when argc=0.

Link: https://lkml.kernel.org/r/20210304094521.GA1826@cosmos
Signed-off-by: Vamshi K Sthambamkadi <[email protected]>
Signed-off-by: Steven Rostedt (VMware) <[email protected]>

ring-buffer: Add a little more information and a WARN when time stamp going backwards is detected

When the CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS is enabled, and the time
stamps are detected as not being valid, it reports information about the
write stamp, but does not show the before_stamp which is still useful
information. Also, it should give a warning once, such that tests detect
this happening.

Signed-off-by: Steven Rostedt (VMware) <[email protected]>

ring-buffer: Force before_stamp and write_stamp to be different on discard

Part of the logic of the new time stamp code depends on the before_stamp and
the write_stamp to be different if the write_stamp does not match the last
event on the buffer, as it will be used to calculate the delta of the next
event written on the buffer.

The discard logic depends on this, as the next event to come in needs to
inject a full timestamp as it can not rely on the last event timestamp in
the buffer because it is unknown due to events after it being discarded. But
by changing the write_stamp back to the time before it, it forces the next
event to use a full time stamp, instead of relying on it.

The issue came when a full time stamp was used for the event, and
rb_time_delta() returns zero in that case. The update to the write_stamp
(which subtracts delta) made it not change. Then when the event is removed
from the buffer, because the before_stamp and write_stamp still match, the
next event written would calculate its delta from the write_stamp, but that
would be wrong as the write_stamp is of the time of the event that was
discarded.

In the case that the delta change being made to write_stamp is zero, set the
before_stamp to zero as well, and this will force the next event to inject a
full timestamp and not use the current write_stamp.

Cc: [email protected]
Fixes: a389d86f7fd09 ("ring-buffer: Have nested events still record running time stamp")
Signed-off-by: Steven Rostedt (VMware) <[email protected]>

tracing: Fix help text of TRACEPOINT_BENCHMARK in Kconfig

It's "cond_resched()" not "cond_sched()".

Link: https://lkml.kernel.org/r/1863065.aFVDpXsuPd@devpool47
Signed-off-by: Rolf Eike Beer <[email protected]>
Signed-off-by: Steven Rostedt (VMware) <[email protected]>

tracing: Remove duplicate declaration from trace.h

A declaration of function "int trace_empty(struct trace_iterator *iter)"
shows up twice in the header file kernel/trace/trace.h

Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Yordan Karadzhov (VMware) <[email protected]>
Signed-off-by: Steven Rostedt (VMware) <[email protected]>

io-wq: ensure all pending work is canceled on exit

If we race on shutting down the io-wq, then we should ensure that any
work that was queued after workers shutdown is canceled. Harden the
add work check a bit too, checking for IO_WQ_BIT_EXIT and cancel if
it's set.

Add a WARN_ON() for having any work before we kill the io-wq context.

Reported-by: [email protected]
Signed-off-by: Jens Axboe <[email protected]>

io_uring: ensure that threads freeze on suspend

Alex reports that his system fails to suspend using 5.12-rc1, with the
following dump:

[  240.650300] PM: suspend entry (deep)
[  240.650748] Filesystems sync: 0.000 seconds
[  240.725605] Freezing user space processes ...
[  260.739483] Freezing of tasks failed after 20.013 seconds (3 tasks refusing to freeze, wq_busy=0):
[  260.739497] task:iou-mgr-446     state:S stack:    0 pid:  516 ppid:   439 flags:0x00004224
[  260.739504] Call Trace:
[  260.739507]  ? sysvec_apic_timer_interrupt+0xb/0x81
[  260.739515]  ? pick_next_task_fair+0x197/0x1cde
[  260.739519]  ? sysvec_reschedule_ipi+0x2f/0x6a
[  260.739522]  ? asm_sysvec_reschedule_ipi+0x12/0x20
[  260.739525]  ? __schedule+0x57/0x6d6
[  260.739529]  ? del_timer_sync+0xb9/0x115
[  260.739533]  ? schedule+0x63/0xd5
[  260.739536]  ? schedule_timeout+0x219/0x356
[  260.739540]  ? __next_timer_interrupt+0xf1/0xf1
[  260.739544]  ? io_wq_manager+0x73/0xb1
[  260.739549]  ? io_wq_create+0x262/0x262
[  260.739553]  ? ret_from_fork+0x22/0x30
[  260.739557] task:iou-mgr-517     state:S stack:    0 pid:  522 ppid:   439 flags:0x00004224
[  260.739561] Call Trace:
[  260.739563]  ? sysvec_apic_timer_interrupt+0xb/0x81
[  260.739566]  ? pick_next_task_fair+0x16f/0x1cde
[  260.739569]  ? sysvec_apic_timer_interrupt+0xb/0x81
[  260.739571]  ? asm_sysvec_apic_timer_interrupt+0x12/0x20
[  260.739574]  ? __schedule+0x5b7/0x6d6
[  260.739578]  ? del_timer_sync+0x70/0x115
[  260.739581]  ? schedule_timeout+0x211/0x356
[  260.739585]  ? __next_timer_interrupt+0xf1/0xf1
[  260.739588]  ? io_wq_check_workers+0x15/0x11f
[  260.739592]  ? io_wq_manager+0x69/0xb1
[  260.739596]  ? io_wq_create+0x262/0x262
[  260.739600]  ? ret_from_fork+0x22/0x30
[  260.739603] task:iou-wrk-517     state:S stack:    0 pid:  523 ppid:   439 flags:0x00004224
[  260.739607] Call Trace:
[  260.739609]  ? __schedule+0x5b7/0x6d6
[  260.739614]  ? schedule+0x63/0xd5
[  260.739617]  ? schedule_timeout+0x219/0x356
[  260.739621]  ? __next_timer_interrupt+0xf1/0xf1
[  260.739624]  ? task_thread.isra.0+0x148/0x3af
[  260.739628]  ? task_thread_unbound+0xa/0xa
[  260.739632]  ? task_thread_bound+0x7/0x7
[  260.739636]  ? ret_from_fork+0x22/0x30
[  260.739647] OOM killer enabled.
[  260.739648] Restarting tasks ... done.
[  260.740077] PM: suspend exit

Play nice and ensure that any thread we create will call try_to_freeze()
at an opportune time so that memory suspend can proceed. For the io-wq
worker threads, mark them as PF_NOFREEZE. They could potentially be
blocked for a long time.

Reported-by: Alex Xu (Hello71) <[email protected]>
Tested-by: Alex Xu (Hello71) <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

io_uring: remove extra in_idle wake up

io_dismantle_req() is always followed by io_put_task(), which already do
proper in_idle wake ups, so we can skip waking the owner task in
io_dismantle_req(). The rules are simpler now, do io_put_task() shortly
after ending a request, and it will be fine.

Signed-off-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

io_uring: inline __io_queue_async_work()

__io_queue_async_work() is only called from io_queue_async_work(),
inline it.

Signed-off-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

io_uring: inline io_req_clean_work()

Inline io_req_clean_work(), less code and easier to analyse
tctx dependencies and refs usage.

Signed-off-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

io_uring: choose right tctx->io_wq for try cancel

When we cancel SQPOLL, @task in io_uring_try_cancel_requests() will
differ from current. Use the right tctx from passed in @task, and don't
forget that it can be NULL when the io_uring ctx exits.

Signed-off-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

io_uring: fix -EAGAIN retry with IOPOLL

We no longer revert the iovec on -EIOCBQUEUED, see commit ab2125df921d,
and this started causing issues for IOPOLL on devies that run out of
request slots. Turns out what outside of needing a revert for those, we
also had a bug where we didn't properly setup retry inside the submission
path. That could cause re-import of the iovec, if any, and that could lead
to spurious results if the application had those allocated on the stack.

Catch -EAGAIN retry and make the iovec stable for IOPOLL, just like we do
for !IOPOLL retries.

Cc: <[email protected]> # 5.9+
Reported-by: Abaci Robot <[email protected]>
Reported-by: Xiaoguang Wang <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

io-wq: fix error path leak of buffered write hash map

The 'err' path should include the hash put, we already grabbed a reference
once we get that far.

Fixes: e941894eae31 ("io-wq: make buffered file write hashed work map per-ctx")
Signed-off-by: Jens Axboe <[email protected]>

io_uring: remove sqo_task

Now, sqo_task is used only for a warning that is not interesting anymore
since sqo_dead is gone, remove all of that including ctx->sqo_task.

Signed-off-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

io_uring: kill sqo_dead and sqo submission halting

As SQPOLL task doesn't poke into ->sqo_task anymore, there is no need to
kill the sqo when the master task exits. Before it was necessary to
avoid races accessing sqo_task->files with removing them.

Signed-off-by: Pavel Begunkov <[email protected]>
[axboe: don't forget to enable SQPOLL before exit, if started disabled]
Signed-off-by: Jens Axboe <[email protected]>

io_uring: ignore double poll add on the same waitqueue head

syzbot reports a deadlock, attempting to lock the same spinlock twice:

============================================
WARNING: possible recursive locking detected
5.11.0-syzkaller #0 Not tainted
--------------------------------------------
swapper/1/0 is trying to acquire lock:
ffff88801b2b1130 (&runtime->sleep){..-.}-{2:2}, at: spin_lock include/linux/spinlock.h:354 [inline]
ffff88801b2b1130 (&runtime->sleep){..-.}-{2:2}, at: io_poll_double_wake+0x25f/0x6a0 fs/io_uring.c:4960

but task is already holding lock:
ffff88801b2b3130 (&runtime->sleep){..-.}-{2:2}, at: __wake_up_common_lock+0xb4/0x130 kernel/sched/wait.c:137

other info that might help us debug this:
 Possible unsafe locking scenario:

       CPU0
       ----
  lock(&runtime->sleep);
  lock(&runtime->sleep);

 *** DEADLOCK ***

 May be due to missing lock nesting notation

2 locks held by swapper/1/0:
 #0: ffff888147474908 (&group->lock){..-.}-{2:2}, at: _snd_pcm_stream_lock_irqsave+0x9f/0xd0 sound/core/pcm_native.c:170
 #1: ffff88801b2b3130 (&runtime->sleep){..-.}-{2:2}, at: __wake_up_common_lock+0xb4/0x130 kernel/sched/wait.c:137

stack backtrace:
CPU: 1 PID: 0 Comm: swapper/1 Not tainted 5.11.0-syzkaller #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
Call Trace:
 <IRQ>
 __dump_stack lib/dump_stack.c:79 [inline]
 dump_stack+0xfa/0x151 lib/dump_stack.c:120
 print_deadlock_bug kernel/locking/lockdep.c:2829 [inline]
 check_deadlock kernel/locking/lockdep.c:2872 [inline]
 validate_chain kernel/locking/lockdep.c:3661 [inline]
 __lock_acquire.cold+0x14c/0x3b4 kernel/locking/lockdep.c:4900
 lock_acquire kernel/locking/lockdep.c:5510 [inline]
 lock_acquire+0x1ab/0x730 kernel/locking/lockdep.c:5475
 __raw_spin_lock include/linux/spinlock_api_smp.h:142 [inline]
 _raw_spin_lock+0x2a/0x40 kernel/locking/spinlock.c:151
 spin_lock include/linux/spinlock.h:354 [inline]
 io_poll_double_wake+0x25f/0x6a0 fs/io_uring.c:4960
 __wake_up_common+0x147/0x650 kernel/sched/wait.c:108
 __wake_up_common_lock+0xd0/0x130 kernel/sched/wait.c:138
 snd_pcm_update_state+0x46a/0x540 sound/core/pcm_lib.c:203
 snd_pcm_update_hw_ptr0+0xa75/0x1a50 sound/core/pcm_lib.c:464
 snd_pcm_period_elapsed+0x160/0x250 sound/core/pcm_lib.c:1805
 dummy_hrtimer_callback+0x94/0x1b0 sound/drivers/dummy.c:378
 __run_hrtimer kernel/time/hrtimer.c:1519 [inline]
 __hrtimer_run_queues+0x609/0xe40 kernel/time/hrtimer.c:1583
 hrtimer_run_softirq+0x17b/0x360 kernel/time/hrtimer.c:1600
 __do_softirq+0x29b/0x9f6 kernel/softirq.c:345
 invoke_softirq kernel/softirq.c:221 [inline]
 __irq_exit_rcu kernel/softirq.c:422 [inline]
 irq_exit_rcu+0x134/0x200 kernel/softirq.c:434
 sysvec_apic_timer_interrupt+0x93/0xc0 arch/x86/kernel/apic/apic.c:1100
 </IRQ>
 asm_sysvec_apic_timer_interrupt+0x12/0x20 arch/x86/include/asm/idtentry.h:632
RIP: 0010:native_save_fl arch/x86/include/asm/irqflags.h:29 [inline]
RIP: 0010:arch_local_save_flags arch/x86/include/asm/irqflags.h:70 [inline]
RIP: 0010:arch_irqs_disabled arch/x86/include/asm/irqflags.h:137 [inline]
RIP: 0010:acpi_safe_halt drivers/acpi/processor_idle.c:111 [inline]
RIP: 0010:acpi_idle_do_entry+0x1c9/0x250 drivers/acpi/processor_idle.c:516
Code: dd 38 6e f8 84 db 75 ac e8 54 32 6e f8 e8 0f 1c 74 f8 e9 0c 00 00 00 e8 45 32 6e f8 0f 00 2d 4e 4a c5 00 e8 39 32 6e f8 fb f4 <9c> 5b 81 e3 00 02 00 00 fa 31 ff 48 89 de e8 14 3a 6e f8 48 85 db
RSP: 0018:ffffc90000d47d18 EFLAGS: 00000293
RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000000
RDX: ffff8880115c3780 RSI: ffffffff89052537 RDI: 0000000000000000
RBP: ffff888141127064 R08: 0000000000000001 R09: 0000000000000001
R10: ffffffff81794168 R11: 0000000000000000 R12: 0000000000000001
R13: ffff888141127000 R14: ffff888141127064 R15: ffff888143331804
 acpi_idle_enter+0x361/0x500 drivers/acpi/processor_idle.c:647
 cpuidle_enter_state+0x1b1/0xc80 drivers/cpuidle/cpuidle.c:237
 cpuidle_enter+0x4a/0xa0 drivers/cpuidle/cpuidle.c:351
 call_cpuidle kernel/sched/idle.c:158 [inline]
 cpuidle_idle_call kernel/sched/idle.c:239 [inline]
 do_idle+0x3e1/0x590 kernel/sched/idle.c:300
 cpu_startup_entry+0x14/0x20 kernel/sched/idle.c:397
 start_secondary+0x274/0x350 arch/x86/kernel/smpboot.c:272
 secondary_startup_64_no_verify+0xb0/0xbb

which is due to the driver doing poll_wait() twice on the same
wait_queue_head. That is perfectly valid, but from checking the rest
of the kernel tree, it's the only driver that does this.

We can handle this just fine, we just need to ignore the second addition
as we'll get woken just fine on the first one.

Cc: [email protected] # 5.8+
Fixes: 18bceab101ad ("io_uring: allow POLL_ADD with double poll_wait() users")
Reported-by: [email protected]
Signed-off-by: Jens Axboe <[email protected]>

io_uring: ensure that SQPOLL thread is started for exit

If we create it in a disabled state because IORING_SETUP_R_DISABLED is
set on ring creation, we need to ensure that we've kicked the thread if
we're exiting before it's been explicitly disabled. Otherwise we can run
into a deadlock where exit is waiting go park the SQPOLL thread, but the
SQPOLL thread itself is waiting to get a signal to start.

That results in the below trace of both tasks hung, waiting on each other:

INFO: task syz-executor458:8401 blocked for more than 143 seconds.
      Not tainted 5.11.0-next-20210226-syzkaller #0
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
task:syz-executor458 state:D stack:27536 pid: 8401 ppid:  8400 flags:0x00004004
Call Trace:
 context_switch kernel/sched/core.c:4324 [inline]
 __schedule+0x90c/0x21a0 kernel/sched/core.c:5075
 schedule+0xcf/0x270 kernel/sched/core.c:5154
 schedule_timeout+0x1db/0x250 kernel/time/timer.c:1868
 do_wait_for_common kernel/sched/completion.c:85 [inline]
 __wait_for_common kernel/sched/completion.c:106 [inline]
 wait_for_common kernel/sched/completion.c:117 [inline]
 wait_for_completion+0x168/0x270 kernel/sched/completion.c:138
 io_sq_thread_park fs/io_uring.c:7115 [inline]
 io_sq_thread_park+0xd5/0x130 fs/io_uring.c:7103
 io_uring_cancel_task_requests+0x24c/0xd90 fs/io_uring.c:8745
 __io_uring_files_cancel+0x110/0x230 fs/io_uring.c:8840
 io_uring_files_cancel include/linux/io_uring.h:47 [inline]
 do_exit+0x299/0x2a60 kernel/exit.c:780
 do_group_exit+0x125/0x310 kernel/exit.c:922
 __do_sys_exit_group kernel/exit.c:933 [inline]
 __se_sys_exit_group kernel/exit.c:931 [inline]
 __x64_sys_exit_group+0x3a/0x50 kernel/exit.c:931
 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46
 entry_SYSCALL_64_after_hwframe+0x44/0xae
RIP: 0033:0x43e899
RSP: 002b:00007ffe89376d48 EFLAGS: 00000246 ORIG_RAX: 00000000000000e7
RAX: ffffffffffffffda RBX: 00000000004af2f0 RCX: 000000000043e899
RDX: 000000000000003c RSI: 00000000000000e7 RDI: 0000000000000000
RBP: 0000000000000000 R08: ffffffffffffffc0 R09: 0000000010000000
R10: 0000000000008011 R11: 0000000000000246 R12: 00000000004af2f0
R13: 0000000000000001 R14: 0000000000000000 R15: 0000000000000001
INFO: task iou-sqp-8401:8402 can't die for more than 143 seconds.
task:iou-sqp-8401    state:D stack:30272 pid: 8402 ppid:  8400 flags:0x00004004
Call Trace:
 context_switch kernel/sched/core.c:4324 [inline]
 __schedule+0x90c/0x21a0 kernel/sched/core.c:5075
 schedule+0xcf/0x270 kernel/sched/core.c:5154
 schedule_timeout+0x1db/0x250 kernel/time/timer.c:1868
 do_wait_for_common kernel/sched/completion.c:85 [inline]
 __wait_for_common kernel/sched/completion.c:106 [inline]
 wait_for_common kernel/sched/completion.c:117 [inline]
 wait_for_completion+0x168/0x270 kernel/sched/completion.c:138
 io_sq_thread+0x27d/0x1ae0 fs/io_uring.c:6717
 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:294
INFO: task iou-sqp-8401:8402 blocked for more than 143 seconds.

Reported-by: [email protected]
Signed-off-by: Jens Axboe <[email protected]>

io_uring: replace cmpxchg in fallback with xchg

io_run_ctx_fallback() can use xchg() instead of cmpxchg(). It's simpler
and faster.

Signed-off-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

io_uring: fix __tctx_task_work() ctx race

There is an unlikely but possible race using a freed context. That's
because req->task_work.func() can free a request, but we won't
necessarily find a completion in submit_state.comp and so all ctx refs
may be put by the time we do mutex_lock(&ctx->uring_ctx);

There are several reasons why it can miss going through
submit_state.comp: 1) req->task_work.func() didn't complete it itself,
but punted to iowq (e.g. reissue) and it got freed later, or a similar
situation with it overflowing and getting flushed by someone else, or
being submitted to IRQ completion, 2) As we don't hold the uring_lock,
someone else can do io_submit_flush_completions() and put our ref.
3) Bugs and code obscurities, e.g. failing to propagate issue_flags
properly.

One example is as follows

  CPU1                                  |  CPU2
=======================================================================
@req->task_work.func()                  |
  -> @req overflwed,                    |
     so submit_state.comp,nr==0         |
                                        | flush overflows, and free @req
                                        | ctx refs == 0, free it
ctx is dead, but we do                  |
lock + flush + unlock           |

So take a ctx reference for each new ctx we see in __tctx_task_work(),
and do release it until we do all our flushing.

Fixes: 65453d1efbd2 ("io_uring: enable req cache for task_work items")
Reported-by: [email protected]
Signed-off-by: Pavel Begunkov <[email protected]>
[axboe: fold in my one-liner and fix ref mismatch]
Signed-off-by: Jens Axboe <[email protected]>

io_uring: kill io_uring_flush()

This was always a weird work-around or file referencing, and we don't
need it anymore. Get rid of it.

Signed-off-by: Jens Axboe <[email protected]>

io_uring: kill unnecessary io_run_ctx_fallback() in io_ring_exit_work()

We already run the fallback task_work in io_uring_try_cancel_requests(),
no need to duplicate at ring exit explicitly.

Signed-off-by: Jens Axboe <[email protected]>

io_uring: move cred assignment into io_issue_sqe()

If we move it in there, then we no longer have to care about it in io-wq.
This means we can drop the cred handling in io-wq, and we can drop the
REQ_F_WORK_INITIALIZED flag and async init functions as that was the last
user of it since we moved to the new workers. Then we can also drop
io_wq_work->creds, and just hold the personality u16 in there instead.

Suggested-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

io_uring: kill unnecessary REQ_F_WORK_INITIALIZED checks

We're no longer checking anything that requires the work item to be
initialized, as we're not carrying any file related state there.

Signed-off-by: Jens Axboe <[email protected]>

io_uring: remove unused argument 'tsk' from io_req_caches_free()

We prune the full cache regardless, get rid of the dead argument.

Signed-off-by: Jens Axboe <[email protected]>

io_uring: destroy io-wq on exec

Destroy current's io-wq backend and tctx on __io_uring_task_cancel(),
aka exec(). Looks it's not strictly necessary, because it will be done
at some point when the task dies and changes of creds/files/etc. are
handled, but better to do that earlier to free io-wq and not potentially
lock previous mm and other resources for the time being.

It's safe to do because we wait for all requests of the current task to
complete, so no request will use tctx afterwards. Note, that
io_uring_files_cancel() may leave some requests for later reaping, so it
leaves tctx intact, that's ok as the task is dying anyway.

Signed-off-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

io_uring: warn on not destroyed io-wq

Make sure that we killed an io-wq by the time a task is dead.

Signed-off-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>