Re: [RFC PATCH 79/86] treewide: net: remove cond_resched()

[Eric Dumazet]

On Wed, Nov 8, 2023 at 12:09 AM Ankur Arora <ankur.a.arora@xxxxxxxxxx> wrote:
>
> There are broadly three sets of uses of cond_resched():
>
> 1.  Calls to cond_resched() out of the goodness of our heart,
>     otherwise known as avoiding lockup splats.
>
> 2.  Open coded variants of cond_resched_lock() which call
>     cond_resched().
>
> 3.  Retry or error handling loops, where cond_resched() is used as a
>     quick alternative to spinning in a tight-loop.
>
> When running under a full preemption model, the cond_resched() reduces
> to a NOP (not even a barrier) so removing it obviously cannot matter.
>
> But considering only voluntary preemption models (for say code that
> has been mostly tested under those), for set-1 and set-2 the
> scheduler can now preempt kernel tasks running beyond their time
> quanta anywhere they are preemptible() [1]. Which removes any need
> for these explicitly placed scheduling points.

What about RCU callbacks ? cond_resched() was helping a bit.

>
> The cond_resched() calls in set-3 are a little more difficult.
> To start with, given it's NOP character under full preemption, it
> never actually saved us from a tight loop.
> With voluntary preemption, it's not a NOP, but it might as well be --
> for most workloads the scheduler does not have an interminable supply
> of runnable tasks on the runqueue.
>
> So, cond_resched() is useful to not get softlockup splats, but not
> terribly good for error handling. Ideally, these should be replaced
> with some kind of timed or event wait.
> For now we use cond_resched_stall(), which tries to schedule if
> possible, and executes a cpu_relax() if not.
>
> Most of the uses here are in set-1 (some right after we give up a
> lock or enable bottom-halves, causing an explicit preemption check.)
>
> We can remove all of them.

A patch series of 86 is not reasonable.

596 files changed, 881 insertions(+), 2813 deletions(-)

If really cond_resched() becomes a nop (Nice !) ,
make this at the definition of cond_resched(),
and add there nice debugging.

Whoever needs to call a "real" cond_resched(), could call a
cond_resched_for_real()
(Please change the name, this is only to make a point)

Then let the removal happen whenever each maintainer decides, 6 months
later, without polluting lkml.

Imagine we have to revert this series in 1 month, how painful this
would be had we removed
~1400 cond_resched() calls all over the place, with many conflicts.

Thanks