On 12 Nov 2020, at 14:04, Trond Myklebust wrote:
> On Thu, 2020-11-12 at 13:39 -0500, Benjamin Coddington wrote:
>>
>>
>> On 12 Nov 2020, at 13:26, Trond Myklebust wrote:
>>
>>> On Thu, 2020-11-12 at 16:51 +0000, Trond Myklebust wrote:
>>>>
>>>> I was going to ask you if perhaps reverting Scott's commit
>>>> 07b5ce8ef2d8
>>>> ("NFS: Make nfs_readdir revalidate less often") might help here?
>>>> My thinking is that will trigger more cache invalidations when
>>>> the
>>>> directory is changing underneath us, and will now trigger
>>>> uncached
>>>> readdir in those situations.
>>>>
>>>>>
>>>
>>> IOW, the suggestion would be to apply something like the following
>>> on
>>> top of the existing readdir patchset:
>>
>> I'm all for this approach, but - I'm rarely seeing the mapping-
>>> nrpages == 0
>> since the cache is dropped by a process in nfs_readdir() that
>> immediately
>> starts filling the cache again.
>
> That's why I moved the check in readdir_search_pagecache. Unless that
> process has set desc->dir_cookie == 0, then that should prevent the
> refilling.
My pathological benchmarking does send another process in with
desc->dir_cookie == 0.
I'm doing fork(), while(getdents) every second with acdirmin=1 and a listing
that takes longer than 1 second to complete uncached, while touching the
directory every second.
Ben
