- Wednesday, March 23, 2016

release-team@conference.openafs.org

Wednesday, March 23, 2016< ^ >

Room Configuration

Room Occupants

GMT+0
[14:03:08] mvita joins the room
[14:04:47] meffie joins the room
[14:07:05] mvita leaves the room
[14:07:25] mvita joins the room
[14:10:13] Jeffrey Altman leaves the room
[14:17:51] Jeffrey Altman joins the room
[14:58:16] wiesand joins the room
[15:01:18] <wiesand> Hello
[15:01:33] <mvita> hi Stephan
[15:03:13] <meffie> ug.
[15:03:42] <wiesand> ug?
[15:03:53] <meffie> bit by the flu bug.
[15:04:54] <wiesand> bad
[15:06:55] <wiesand> I’m not sure I understand Ben’s addition to what has become 12228
[15:07:08] <mvita> looking
[15:07:43] <wiesand> The part in osi_vnodeops.c
[15:08:34] <wiesand> If there is a default_file_splice_read, we fall back to the old code which is considered unsafe
[15:09:04] <wiesand> If there is, we leave the fields unassigned, which is also considered unsafe?
[15:14:16] <meffie> he discusses it on 12217
[15:14:23] <mvita> sorry, my brain is a bit fuzzy, running on very little sleep
[15:14:55] <wiesand> I’ll bring it up in gerrit
[15:15:38] <wiesand> I’m not even sure what to discuss today.
[15:15:50] <wiesand> We’re basically where we were 3 weeks ago.
[15:16:25] <wiesand> + merge conflicts due to the changes smuggled in behind gerrit’s back
[15:17:22] <jhutz@jis.mit.edu/owl> I'm not sure what you mean by "unsafe".
If there is a default_file_splice_read, we _don't_ fill the ops fields,
presumably because the presence of default splice_read (and splice_write)
means Linux will use them if the fields are left NULL.
If there's no default_file_splice_read, we still have to fill them with
something, which turns out to be the generic ones.
[15:20:04] <wiesand> we still fill .splice_write with iter_file_splice_write if that exists?
[15:20:40] <jhutz@jis.mit.edu/owl> Hm. Was that part of the original problem?
[15:21:20] <wiesand> I guess. But I may be wrong.
[15:23:18] <jhutz@jis.mit.edu/owl> It strikes me that this patchset actually contains two changes:
(1) splice() may be risky due to -ERESTARTSYS, so don't use it;
(2) generic_file_splice_read/write go away in newer kernels, so use the
default implementation instead (but not in older kernels where it doesn't
exist).
[15:24:47] <jhutz@jis.mit.edu/owl> Err, wrt (1), specifically don't use it in the implementation of storeproc
on a cache file.
[15:26:59] <jhutz@jis.mit.edu/owl> These are fundamentally separate -- one is about the interface we as a
filesystem provider present to Linux's VFS layer as a filesytem, while the
other is about what we do as a filesystem _consumer_ managing the cache.
Having them in the same commit makes it harder to tell they are separate.
(or, if they're actually not, the commit message should explain why)
[15:31:52] <wiesand> Thanks for the comments.
[15:32:40] <wiesand> The sad fact is that we no longer have a contributor who is able to sort these things out efficiently.
[15:37:04] <mvita> stepping away for a bit, brb
[15:39:00] <wiesand> Ok. As I said, not much to discuss anyway.
[16:00:58] <wiesand> I have to run. I’ll have to rebase a lot, and then should probably beg for reviews on -devel. If that doesn’t help, it may be time to give up.
[16:01:36] wiesand leaves the room
[16:59:12] meffie leaves the room
[18:29:27] Jeffrey Altman leaves the room
[18:29:27] Jeffrey Altman joins the room
[20:52:25] mvita leaves the room