The tasklist_lock can be a scalability bottleneck. For current tasks,
we don't need the tasklist_lock to protect tsk->sighand or tsk->signal.
If non-current callers become a bottleneck, we could use
lock_task_sighand().
Signed-off-by: Barret Rhoden <[email protected]>
---
kernel/sys.c | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/kernel/sys.c b/kernel/sys.c
index 8fdac0d90504..e56d1ae910af 100644
--- a/kernel/sys.c
+++ b/kernel/sys.c
@@ -1576,7 +1576,8 @@ int do_prlimit(struct task_struct *tsk, unsigned int resource,
}
/* protect tsk->signal and tsk->sighand from disappearing */
- read_lock(&tasklist_lock);
+ if (tsk != current)
+ read_lock(&tasklist_lock);
if (!tsk->sighand) {
retval = -ESRCH;
goto out;
@@ -1611,7 +1612,8 @@ int do_prlimit(struct task_struct *tsk, unsigned int resource,
IS_ENABLED(CONFIG_POSIX_TIMERS))
update_rlimit_cpu(tsk, new_rlim->rlim_cur);
out:
- read_unlock(&tasklist_lock);
+ if (tsk != current)
+ read_unlock(&tasklist_lock);
return retval;
}
--
2.34.1.173.g76aa8bc2d0-goog
Barret Rhoden <[email protected]> writes:
> The tasklist_lock can be a scalability bottleneck. For current tasks,
> we don't need the tasklist_lock to protect tsk->sighand or tsk->signal.
> If non-current callers become a bottleneck, we could use
> lock_task_sighand().
Do you have any numbers? As the entire point of this change is
performance it would be good to see how the performance changes.
Especially as a read_lock should not be too bad as it allows sharing,
nor do I expect reading or writing the rlimit values to be particularly
frequent. So some insight into what kinds of userspace patterns make
this a problem would be nice.
This change is a bit scary as it makes taking a lock conditional and
increases the probability of causing a locking mistake.
If you are going to make this change I would say that do_prlimit should
become static and taking the tasklist_lock should move into prlimit64.
Looking a little closer it looks like that update_rlimit_cpu should use
lock_task_sighand, and once lock_task_sighand is used there is actually
no need for the tasklist_lock at all. As holding the reference to tsk
guarantees that tsk->signal remains valid.
So I completely agree there are cleanups that can happen in this area.
Please make those and show numbers in how they improve things, instead
of making the code worse with a conditional lock.
Eric
> Signed-off-by: Barret Rhoden <[email protected]>
> ---
> kernel/sys.c | 6 ++++--
> 1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sys.c b/kernel/sys.c
> index 8fdac0d90504..e56d1ae910af 100644
> --- a/kernel/sys.c
> +++ b/kernel/sys.c
> @@ -1576,7 +1576,8 @@ int do_prlimit(struct task_struct *tsk, unsigned int resource,
> }
>
> /* protect tsk->signal and tsk->sighand from disappearing */
> - read_lock(&tasklist_lock);
> + if (tsk != current)
> + read_lock(&tasklist_lock);
> if (!tsk->sighand) {
> retval = -ESRCH;
> goto out;
> @@ -1611,7 +1612,8 @@ int do_prlimit(struct task_struct *tsk, unsigned int resource,
> IS_ENABLED(CONFIG_POSIX_TIMERS))
> update_rlimit_cpu(tsk, new_rlim->rlim_cur);
> out:
> - read_unlock(&tasklist_lock);
> + if (tsk != current)
> + read_unlock(&tasklist_lock);
> return retval;
> }
Hi -
On 12/13/21 5:34 PM, Eric W. Biederman wrote:
> Do you have any numbers? As the entire point of this change is
> performance it would be good to see how the performance changes.
>
> Especially as a read_lock should not be too bad as it allows sharing,
> nor do I expect reading or writing the rlimit values to be particularly
> frequent. So some insight into what kinds of userspace patterns make
> this a problem would be nice.
This was motivated by slowdowns we observed on a few machines running
tests in a cluster. AFAIK, there were a lot of small tests, many of
which mucked with process management syscalls while fork/joining other
tasks.
Based on a cycles profile, it looked like ~87% of the time was spent in
the kernel, ~42% of which was just trying to get *some* spinlock
(queued_spin_lock_slowpath, not necessarily the tasklist_lock).
The big offenders (with rough percentages in cycles of the overall trace):
- do_wait 11%
- setpriority 8% (potential future patch)
- kill 8%
- do_exit 5%
- clone 3%
- prlimit64 2% (this patch)
- getrlimit 1% (this patch)
Even though do_prlimit was using a read_lock, it was still contending on
the internal queued_spin_lock.
The prlimit was only 3% of the total. This patch was more of a "oh,
this doesn't *need* the tasklist_lock for p == current" - can we fix
that? I actually don't even know often those prlimit64 calls had p ==
current.
setpriority was a bigger one too - is the tasklist lock only needed for
the PGRP ops? (I thought so based on where the tasklist_lock is write
locked and the comment on task_pgrp()). If so, I could do that in
another patch.
> This change is a bit scary as it makes taking a lock conditional and
> increases the probability of causing a locking mistake.
I definitely see how making the code more brittle might not be worth the
small win. If this is more "damage" than "cleanup", then I can drop it.
> If you are going to make this change I would say that do_prlimit should
> become static and taking the tasklist_lock should move into prlimit64.
>
>
> Looking a little closer it looks like that update_rlimit_cpu should use
> lock_task_sighand, and once lock_task_sighand is used there is actually
> no need for the tasklist_lock at all. As holding the reference to tsk
> guarantees that tsk->signal remains valid.
Maybe do both? unconditionally grab lock_task_sighand (instead of
tasklist_lock) in prlimit64.
> So I completely agree there are cleanups that can happen in this area.
> Please make those and show numbers in how they improve things, instead
> of making the code worse with a conditional lock.
Unfortunately, I can't easily get a "before and after" on this change.
The motivating issue popped up sporadically, but getting it to happen in
a setup under *my* control is organizationally a pain. So I understand
if you wouldn't want the patch for that reason. Ideally, the changes
would make the code easier to follow and clearer about why we're locking.
If you're OK with two patches that 1) grab lock_task_sighand in
prlimit64 and 2) moving the read_lock in {set,get}priority into the PGRP
cases (assuming I was correct on that), I can send them out.
If it's too much of a risk/ugliness for not clear enough gain (in code
quality or performance), I'm fine with dropping it.
Thanks for looking,
Barret
Barret Rhoden <[email protected]> writes:
> Hi -
>
> On 12/13/21 5:34 PM, Eric W. Biederman wrote:
>> Do you have any numbers? As the entire point of this change is
>> performance it would be good to see how the performance changes.
>>
>> Especially as a read_lock should not be too bad as it allows sharing,
>> nor do I expect reading or writing the rlimit values to be particularly
>> frequent. So some insight into what kinds of userspace patterns make
>> this a problem would be nice.
>
> This was motivated by slowdowns we observed on a few machines running
> tests in a cluster. AFAIK, there were a lot of small tests, many of
> which mucked with process management syscalls while fork/joining other
> tasks.
>
> Based on a cycles profile, it looked like ~87% of the time was spent
> in the kernel, ~42% of which was just trying to get *some* spinlock
> (queued_spin_lock_slowpath, not necessarily the tasklist_lock).
>
> The big offenders (with rough percentages in cycles of the overall trace):
>
> - do_wait 11%
> - setpriority 8% (potential future patch)
> - kill 8%
> - do_exit 5%
> - clone 3%
> - prlimit64 2% (this patch)
> - getrlimit 1% (this patch)
>
> Even though do_prlimit was using a read_lock, it was still contending
> on the internal queued_spin_lock.
>
> The prlimit was only 3% of the total. This patch was more of a "oh,
> this doesn't *need* the tasklist_lock for p == current" - can we fix
> that? I actually don't even know often those prlimit64 calls had p ==
> current.
>
> setpriority was a bigger one too - is the tasklist lock only needed
> for the PGRP ops? (I thought so based on where the tasklist_lock is
> write locked and the comment on task_pgrp()). If so, I could do that
> in another patch.
That is my understanding. For setpriority to change everything
atomically it must hold the tasklist lock when dealing with more than
one process at a time.
>> This change is a bit scary as it makes taking a lock conditional and
>> increases the probability of causing a locking mistake.
>
> I definitely see how making the code more brittle might not be worth
> the small win. If this is more "damage" than "cleanup", then I can
> drop it.
>
>> If you are going to make this change I would say that do_prlimit should
>> become static and taking the tasklist_lock should move into prlimit64.
>>
>>
>> Looking a little closer it looks like that update_rlimit_cpu should use
>> lock_task_sighand, and once lock_task_sighand is used there is actually
>> no need for the tasklist_lock at all. As holding the reference to tsk
>> guarantees that tsk->signal remains valid.
>
> Maybe do both? unconditionally grab lock_task_sighand (instead of
> tasklist_lock) in prlimit64.
In update_rlimit_cpu use lock_task_sighand instead of unconditionally
grabbing sighand->siglock (because without tasklist_lock sighand might
be NULL).
Then do_prlimit64 can drop the tasklist lock and the test for
sighand == NULL.
This will address every prlimit64 case instead of just when updating
current. With prlimit64 in your profile I expect some of those are
non-current.
>> So I completely agree there are cleanups that can happen in this area.
>> Please make those and show numbers in how they improve things, instead
>> of making the code worse with a conditional lock.
>
> Unfortunately, I can't easily get a "before and after" on this
> change. The motivating issue popped up sporadically, but getting it to
> happen in a setup under *my* control is organizationally a pain. So I
> understand if you wouldn't want the patch for that reason. Ideally,
> the changes would make the code easier to follow and clearer about why
> we're locking.
Even a microbenchmark that stresses the lock and can show the
performance impact of the change you are making is useful.
Simply reorganizing and removing an unnecessary lock lowers the bar
because the code is then both simpler and pretty much by definition
faster. Although sometimes another lock is then hit and the contention
moves.
> If you're OK with two patches that 1) grab lock_task_sighand in
> prlimit64 and 2) moving the read_lock in {set,get}priority into the
> PGRP cases (assuming I was correct on that), I can send them out.
I think getpriority can only use the rcu_read_lock. I don't think it
has any atomicity guarantees.
For setpriority the single process case should be safe just use rcu.
I haven't read through all of what set_one_prio is doing to confirm it
is safe, but in principle it should be safe.
> If it's too much of a risk/ugliness for not clear enough gain (in code
> quality or performance), I'm fine with dropping it.
Removing the tasklist_lock where we can is definitely a clear gain.
Simply shoving tasklist_lock aside and making the code more complicated
is much less clear.
Plus anything you can benchmark (even microbenchmark) and show the
benefit of is welcome. Especially when you have indications that it
makes a difference in a larger context.
Eric
On Wed, Dec 15, 2021 at 01:42:32PM -0600, Eric W. Biederman wrote:
...
>
> > If it's too much of a risk/ugliness for not clear enough gain (in code
> > quality or performance), I'm fine with dropping it.
>
> Removing the tasklist_lock where we can is definitely a clear gain.
>
> Simply shoving tasklist_lock aside and making the code more complicated
> is much less clear.
>
> Plus anything you can benchmark (even microbenchmark) and show the
> benefit of is welcome. Especially when you have indications that it
> makes a difference in a larger context.
Thanks for looking into this, Eric! I must confess I've a vague memory
about this code. Still while you're talking about cleanup I wonder if
we should make do_prlimit() being a static function, not global as it
now.
On 12/15/21 14:42, Eric W. Biederman wrote:
> In update_rlimit_cpu use lock_task_sighand instead of unconditionally
> grabbing sighand->siglock (because without tasklist_lock sighand might
> be NULL).
this ended up being a minor complication, since update_rlimit_cpu()
could fail if the task was exiting, but i think i sorted it out.
i'll send out revised patchset shortly with this change, including
making do_prlimit() static.
thanks,
barret