[lkml]   [2010]   [Jun]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Subjecttrying to understand READ_META, READ_SYNC, WRITE_SYNC & co
Hi Jens,

I'm trying to understand the "sync" and "meta" flags we add to bios /
requests and the more I look into the more I get confused. Let's start
with the simple bits:


Only really used for READ_META. There's a WRITE_META that I
added a while ago, but the actual users of it are gone again

The READ_META is used by ext3 and ext4 for reading inodes,
and for all reads of directory blocks. XFS has a place that
sets READ_META, but it's unreachable now. GFS2 on the other
hand uses REQ_META directly, and ORs it into just about every
other kind of READ/WRITE request: WRITE_BARRIER,



READ_SYNC is used by GFS2 in a few spots, but always together
with REQ_META, and by jfs and xfs for the code reading the log
during recovery.
(S)WRITE_SYNC_PLUG is used for ->writepage instances with
the WB_SYNC_ALL sync_mode, and random pieces of journal /
metadata I/O in gfs2 and jbd(2) as well as fsync_buffers_list().
Note that the SWRITE_SYNC_PLUG uses actually degenerate to WRITE
in ll_rw_block.c due to what seems to be a bug in that code.
WRITE_SYNC is used by xfs for the log write if barriers are not
supported, by jfs for all log writes, by gfs2 for at least some
log writes, by btrfs for data writeout in some cases.
SWRITE_SYNC is entirely unused.
WRITE_ODIRECT_PLUG is used for O_DIRECT writes, and
WRITE_BARRIER is used for anything using barriers.


SWRITE variants. See the description above for details.



Now how do we use these flags in the block layer?


The only place where we ever use this flag is inside the
cfq scheduler. In cfq_choose_req we use it to give a meta
request priority over one that doesn't have it. But before
that we already do the same preference check with rw_is_sync,
which evaluates to true for requests with that are either
reads or have REQ_SYNC set. So for reads the REQ_META flag
here effectively is a no-op, and for writes it gives less
priority than REQ_SYNC.
In addition to that we use it to account for pending metadata
requests in cfq_rq_enqueued/cfq_remove_request which gets
checked in cfq_should_preempt to give priority to a meta
request if the other queue doesn't have any pending meta
requests. But again this priority comes after a similar
check for sync requests that checks if the other queue has
been marked to have sync requests pending.


This one is used in quite a few places, with many of them
obsfucated by macros like rw_is_sync, rq_is_sync and
cfq_bio_sync. In general all uses seem to imply giving
a write request the same priority as a read request and
treat it as synchronous. I could not spot a place where
it actually has any effect on reads.


Only used inside the cfq I/O scheduler in one place:
cfq_completed_request. Only used for READ or synchronous
request. I can't figure out what's actually going on in
details here.


Used to call __generic_unplug_device at the end of
__make_reques, and one probably wrong usage in drbd (but given
what a big piece of junk drbd is I refuse to step into that).

That's how far I got to understand this.

Now the big questions:

What eactly is the REQ_META flag for? Even after going
through all the things I mentioned above I don't quite
understand it. It's only used for reads, and doesn't seem
to give any major priority. Should it be used for all metadata
reads or just some? Currently it's not actually very widely

Should we allow REQ_META on writes, but ignore it except for
blktrace? If the answer above is we want to tag all metadata
reads as REQ_META that would allow easily spotting all metadata
I/O in blktrace.

Is there any point in keeping the READ_SYNC? If the users
really want to keep the direct unplugging semantics we can
just write READ | REQ_UNPLUG, but even that seems to be wrong
for the log recvoery callers.

Similarly is there a point in keeping WRITE_SYNC? Most callers
really want WRITE_SYNC_PLUG semantics, and the current naming
is extremtly confusing. I'd much rather see the callers
specify explicitly using REQ_UNPLUG if they want to unplug
the queue.

Why do O_DIRECT writes not want to set REQ_NOIDLE (and that
exactly does REQ_NOIDLE mean anyway). It's the only sync writes
that do not set it, so if this special case went away we
could get rid of the flag and key it off REQ_SYNC.

Do we really need to keep the SWRITE flags? I'd rather
make that an additional flag to ll_rw_block, or even better
split ll_rw_block into two helpers for the guaranteed locking
or not case.

 \ /
  Last update: 2010-06-21 11:51    [W:0.082 / U:31.360 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site