Do we really need clean dcache/icache at the current stage?
That will introduce more code work, so far, iput() will put
those unreferenced inodes into superblock lru list. To free
the inodes inside a specific directory, it seems we do not
have a handy API to use, and need
modify iput() to recognize our situation, and collect those
inodes into our list rather than superblock lru list. Maybe
we stay at current stage now, since it is simple and could
gain the major benefits, leave the dcache/icache cleaning
to do in the future?
On 2013/12/31 5:33, Dave Hansen wrote:
> On 12/30/2013 11:40 AM, Andreas Dilger wrote:
>> On Dec 30, 2013, at 12:18, Dave Hansen <[email protected]> wrote:
>>> Why is this necessary to do in the kernel? Why not leave it to
>>> userspace to walk the filesystem(s)?
>>
>> I would suspect that trying to do it in userspace would be quite bad. It would require traversing the whole directory tree to issue cache flushed for each subdirectory, but it doesn't know when to stop traversal. That would mean the "cache flush" would turn into "cache pollute" and cause a lot of disk IO for subdirectories not in cache to begin with.
>
> That makes sense for dentries at least and is a pretty good reason.
> Probably good enough to to include some text in the patch description.
> ;) Perhaps: "We need this interface because we have no way of
> determining what is in the dcache from userspace, and we do not want
> userspace to pollute the dcache going and looking for page cache to evict."
>
> One other thing that bothers me: POSIX_FADV_DONTNEED on a directory
> seems like it should do something with the _directory_. It should undo
> the kernel's caching that happens as a result of readdir().
>
> Should this also be trying to drop the dentry/inode entries like "echo 2
>> .../drop_caches" does?
On 01/02/2014 04:44 AM, Li Wang wrote:
> Do we really need clean dcache/icache at the current stage?
> That will introduce more code work, so far, iput() will put
> those unreferenced inodes into superblock lru list. To free
> the inodes inside a specific directory, it seems we do not
> have a handy API to use, and need
> modify iput() to recognize our situation, and collect those
> inodes into our list rather than superblock lru list. Maybe
> we stay at current stage now, since it is simple and could
> gain the major benefits, leave the dcache/icache cleaning
> to do in the future?
<sigh> top posting....
I read your response as "that's the right thing to do, but it's too much
work". Fair enough. But if we're going to take the lazy hack approach
here, maybe we should do this through some other interface than a
syscall where we're stuck with the behavior.