Re: [RFC PATCH 1/2] NFSD: Add aggressive write throttling control

[Chuck Lever <cel@kernel.org>]

On 1/7/26 2:55 AM, Christoph Hellwig wrote:
> On Fri, Dec 19, 2025 at 09:11:04AM -0500, Chuck Lever wrote:
>> From: Chuck Lever <chuck.lever@oracle.com>
>>
>> On NFS servers with fast network links but slow storage, clients can
>> generate WRITE requests faster than the server can flush payloads to
>> durable storage. This can push the server into memory exhaustion as
>> dirty pages accumulate across hundreds of concurrent NFSD threads.
>>
>> The existing dirty page throttling (balance_dirty_pages()) uses
>> per-task accounting with default ratelimits that allow each thread
>> to dirty ~32 pages before throttling occurs. With many NFSD threads,
>> this allows significant dirty page accumulation before any
>> throttling kicks in.
> 
> What makes NFSD so special here vs say a userspace process with a bunch
> of threads?  Also what is the actual problem we're trying to solve?

The problem, as I see it, is that the system is not providing enough
backpressure to slow down noisy clients, allowing them to overwhelm
the server's memory with UNSTABLE WRITE traffic.

This is the same issue, IMO, that Mike's direct I/O is attempting to
address. Our implementation of UNSTABLE WRITE is a denial-of-service
vector.


> I kinda hate having this stuff in NFSD when there's nothing specific
> about nfs serving here.
Don't worry too much about that, these patches are obviously not in any
kind of merge-able shape yet. We do need to understand the metabolism of
UNSTABLE WRITEs, in particular, to get a clear picture of what needs to
be controlled to make the server autonomously stable.


-- 
Chuck Lever