On Tue, Feb 08, 2022 at 11:07:50AM +0100, Nicolas Saenz Julienne wrote:
> The page allocator's per-cpu page lists (pcplists) are currently
> protected using local_locks. While performance savvy, this doesn't allow
> for remote access to these structures. CPUs requiring system-wide
> changes to the per-cpu lists get around this by scheduling
> workers on each CPU. That said, some setups like NOHZ_FULL CPUs,
> aren't well suited to this since they can't handle interruptions
> of any sort.
>
> To mitigate this, replace the current draining mechanism with one that
> allows remotely draining the lists:
>
> - Each CPU now has two pcplists pointers: one that points to a pcplists
> instance that is in-use, 'pcp->lp', another that points to an idle
> and empty instance, 'pcp->drain'. CPUs access their local pcplists
> through 'pcp->lp' and the pointer is dereferenced atomically.
>
> - When a CPU decides it needs to empty some remote pcplists, it'll
> atomically exchange the remote CPU's 'pcp->lp' and 'pcp->drain'
> pointers. A remote CPU racing with this will either have:
>
> - An old 'pcp->lp' reference, it'll soon be emptied by the drain
> process, we just have to wait for it to finish using it.
>
> - The new 'pcp->lp' reference, that is, an empty pcplists instance.
> rcu_replace_pointer()'s release semantics ensures any prior
> changes will be visible by the remote CPU, for example: changes
> to 'pcp->high' and 'pcp->batch' when disabling the pcplists.
>
> - The CPU that started the drain can now wait for an RCU grace period
> to make sure the remote CPU is done using the old pcplists.
> synchronize_rcu() counts as a full memory barrier, so any changes the
> local CPU makes to the soon to be drained pcplists will be visible to
> the draining CPU once it returns.
>
> - Then the CPU can safely free the old pcplists. Nobody else holds a
> reference to it. Note that concurrent access to the remote pcplists
> drain is protected by the 'pcpu_drain_mutex'.
>
> >From an RCU perspective, we're only protecting access to the pcplists
> pointer, the drain operation is the writer and the local_lock critical
> sections are the readers. RCU guarantees atomicity both while
> dereferencing the pcplists pointer and replacing it. It also checks for
> RCU critical section/locking correctness, as all readers have to hold
> their per-cpu pagesets local_lock, which also counts as a critical
> section from RCU's perspective.
>
> >From a performance perspective, on production configurations, the patch
> adds an extra dereference to all hot paths (under such circumstances
> rcu_dereference() will simplify to READ_ONCE()). Extensive measurements
> have been performed on different architectures to ascertain the
> performance impact is minimal. Most platforms don't see any difference
> and the worst-case scenario shows a 1-3% degradation on a page
> allocation micro-benchmark. See cover letter for in-depth results.
>
> Accesses to the pcplists like the ones in mm/vmstat.c don't require RCU
> supervision since they can handle outdated data, but they do use
> rcu_access_pointer() to avoid compiler weirdness make sparse happy.
>
> Note that special care has been taken to verify there are no races with
> the memory hotplug code paths. Notably with calls to zone_pcp_reset().
> As Mel Gorman explains in a previous patch[1]: "The existing hotplug
> paths guarantees the pcplists cannot be used after zone_pcp_enable()
> [the one in offline_pages()]. That should be the case already because
> all the pages have been freed and there is no page to put on the PCP
> lists."
>
> All in all, this technique allows for remote draining on all setups with
> an acceptable performance impact. It benefits all sorts of use cases:
> low-latency, real-time, HPC, idle systems, KVM guests.
>
> [1] 8ca559132a2d ("mm/memory_hotplug: remove broken locking of zone PCP
> structures during hot remove")
>
> Signed-off-by: Nicolas Saenz Julienne <nsaenzju@xxxxxxxxxx>
> ---
>
> Changes since RFC:
> - Avoid unnecessary spin_lock_irqsave/restore() in free_pcppages_bulk()
> - Add more detail to commit and code comments.
> - Use synchronize_rcu() instead of synchronize_rcu_expedited(), the RCU
> documentation says to avoid it unless really justified. I don't think
> it's justified here, if we can schedule and join works, waiting for
> an RCU grace period is OK.
https://patchwork.ozlabs.org/project/netdev/patch/1306228052.3026.16.camel@edumazet-laptop/
Adding 100ms to direct reclaim path might be problematic. It will also
slowdown kcompactd (note it'll call drain_all_pages for each zone).
> - Avoid sparse warnings by using rcu_access_pointer() and
> rcu_dereference_protected().
>
> include/linux/mmzone.h | 22 +++++-
> mm/page_alloc.c | 155 ++++++++++++++++++++++++++---------------
> mm/vmstat.c | 6 +-
> 3 files changed, 120 insertions(+), 63 deletions(-)
>
> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> index b4cb85d9c6e8..b0b593fd8e48 100644
> --- a/include/linux/mmzone.h
> +++ b/include/linux/mmzone.h
> @@ -388,13 +388,31 @@ struct per_cpu_pages {
> short expire; /* When 0, remote pagesets are drained */
> #endif
>
> - struct pcplists *lp;
> + /*
> + * As a rule of thumb, any access to struct per_cpu_pages's 'lp' has
> + * happen with the pagesets local_lock held and using
> + * rcu_dereference_check(). If there is a need to modify both
> + * 'lp->count' and 'lp->lists' in the same critical section 'pcp->lp'
> + * can only be derefrenced once. See for example:
Typo.
