On 12/25/20 9:50 PM, Al Viro wrote:
> On Fri, Dec 25, 2020 at 07:41:17PM -0700, Jens Axboe wrote:
>> On 12/17/20 9:19 AM, Jens Axboe wrote:
>>> Most callers check for non-zero return, and assume it's -ECHILD (which
>>> it always will be). One caller uses the actual error return. Clean this
>>> up and make it fully consistent, by having unlazy_walk() return a bool
>>> instead. Rename it to try_to_unlazy() and return true on success, and
>>> failure on error. That's easier to read.
>>
>> Al, were you planning on queuing this one up for 5.11 still? I'm fine
>> with holding for 5.12 as well, would just like to know what your plans
>> are. Latter goes for the whole series too, fwiw.
>
> Seeing that it has not sat in -next at all, what I'm going to do is
> to put it into 5.11-rc1-based branch. It's really been too late for
> something like that for this cycle and IME a topic branch started
> before the merges for previous cycle are over is too likely to require
> backmerges, if not outright rebases. So let's branch it at -rc1 and
> it'll go into #for-next from the very beginning.
That sounds fine, thanks Al. Will you queue up 1-3 in a stable branch?
Then I can just pull that into mine for the io_uring bits. I'll also
then post the stat part of this too, for separate review.
BTW, I wrote a man page addition for openat2(2) for RESOLVE_CACHED:
commit 3b381a6bc50da79eee6a3a1da330480d0cb46302
Author: Jens Axboe <axboe@xxxxxxxxx>
Date: Thu Dec 17 14:15:15 2020 -0700
man2/openat2.2: add RESOLVE_CACHED
RESOLVE_CACHED allows an application to attempt a cache-only open
of a file. If this isn't possible, the request will fail with
-1/EAGAIN and the caller should retry without RESOLVE_CACHED set.
This will generally happen from a different context, where a slower
open operation can be performed.
Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
diff --git a/man2/openat2.2 b/man2/openat2.2
index 3bda20620574..f5af3eee2cda 100644
--- a/man2/openat2.2
+++ b/man2/openat2.2
@@ -385,6 +385,17 @@ This may occur if, for example,
a system pathname that is used by an application is modified
(e.g., in a new distribution release)
so that a pathname component (now) contains a bind mount.
+.TP
+.B RESOLVE_CACHED
+Make the open operation fail unless all path components are already present
+in the kernels lookup cache.
+If any kind of revalidation or IO is needed to satisfy the lookup,
+.BR openat2 ()
+fails with the error
+.B EAGAIN.
+This is useful in providing a fast path open that can be performed without
+resorting to thread offload, or other mechanism that an application might
+use to offload slower operations.
.RE
.IP
If any bits other than those listed above are set in
@@ -421,6 +432,14 @@ The caller may choose to retry the
.BR openat2 ()
call.
.TP
+.B EAGAIN
+.BR RESOLVE_CACHED
+was set, and the open operation cannot be performed cached.
+The caller should retry without
+.B RESOLVE_CACHED
+set in
+.I how.resolve
+.TP
.B EINVAL
An unknown flag or invalid value was specified in
.IR how .
--
Jens Axboe
