From: Vivek Goyal <vgoyal@redhat.com>
To: Amir Goldstein <amir73il@gmail.com>
Cc: overlayfs <linux-unionfs@vger.kernel.org>,
Miklos Szeredi <miklos@szeredi.hu>
Subject: Re: [PATCH 11/11] ovl: Put barriers to order oi->__upperdentry and OVL_METACOPY update
Date: Thu, 19 Oct 2017 11:59:32 -0400 [thread overview]
Message-ID: <20171019155932.GA24029@redhat.com> (raw)
In-Reply-To: <CAOQ4uxjfkTfwEi7iwca7sHfsQvWK_foinBYjP1OjVTjDeEKq+A@mail.gmail.com>
On Thu, Oct 19, 2017 at 06:39:57PM +0300, Amir Goldstein wrote:
> On Thu, Oct 19, 2017 at 6:22 PM, Vivek Goyal <vgoyal@redhat.com> wrote:
> > On Thu, Oct 19, 2017 at 06:08:32PM +0300, Amir Goldstein wrote:
> >> On Thu, Oct 19, 2017 at 5:58 PM, Vivek Goyal <vgoyal@redhat.com> wrote:
> >> > On Thu, Oct 19, 2017 at 04:21:46PM +0300, Amir Goldstein wrote:
> >> ...
> >> >>
> >> >> Process 2 will get lower dentry on open for read at 8AM
> >> >> Process 1 will copy up file at 9AM (on CPU1)
> >> >> Process 2 will open same file for read at 9AM (on CPU2)
> >> >> Does it matter if process 2 gets lower or upper dentry? No.
> >> >> It only matter that IF process 2 gets an upper dentry, that
> >> >> this dentry is consistent, so it only matters that IF __upperdentry
> >> >> is visible to CPU2 AND OVL_UPPER_DATA flag is visible to
> >> >> CPU2 then dentry and its inode are consistent.
> >> >
> >> > That's a good point. So if OVL_UPPER_DATA update is not visible on CPU2
> >> > yet, then CPU1 will use lower dentry. And this is equivalent to as if file
> >> > copy up has not taken place yet.
> >> >
> >> > And if CPU1 needed to do use upper dentry only, then it will do flags=WRITE
> >> > and that will take oi->lock and make sure OVL_UPPER_DATA is set.
> >> >
> >> > So only *additional* smp_rmb()/smp_wmb() we require for the case when
> >> > data is copied up later and we need to make sure OVL_UPPER_DATA is
> >> > visible only after the full data copy up is done and stable.
> >> >
> >> >
> >>
> >> Right. forgot about that wmb.
> >>
> >> >>
> >> >> So IMO you may only need to add smp_rmb() before
> >> >> ovl_test_flag(OVL_UPPER_DATA in ovl_d_real() and the smp_wmb()
> >> >> in ovl_inode_update() should be sufficient.
> >> >> Change the comment in ovl_inode_update() to mention that wmb also
> >> >> matches rmb in ovl_d_real() w.r.t OVL_UPPER_DATA flag.
> >> >
> >> > Hmm..., I agree that we require smp_rmb() here but it will pair with
> >> > smp_wmb() in ovl_copy_meta_data_inode() and not the one in
> >> > ovl_inode_update(), right? Something like.
> >>
> >> Right. my bad.
> >>
> >> >
> >> > ovl_d_real() {
> >> > bool has_upper_data;
> >> >
> >> > has_upper_data = ovl_test_flag(OVL_UPPER_DATA, d_inode(dentry));
> >> > /* Pairs with smp_wmb() in ovl_copy_up_meta_inode_data() */
> >> > smp_rmb();
> >> > if (!has_upper_data)
> >> > goto lower;
> >>
> >> Just put smp_rmb() here. no need for the bool variable.
> >> rmb does matter if you goto lower...
> >
> > I thought smp_rmb() has to be put *only* after LOAD of oi->flags.
> > Something like.
> >
> > LOAD oi->flags
> > smp_rmb()
> > Look at results of oi->flags and take action.
> >
> > So that means I need to store results of oi->flags load in variable
> > temporarily so that I can analyze it after smp_rmb(). IOW, I am not
> > sure how would I get rid of boolean here. I need some kind of temp
> > variable.
> >
>
> One of us is very confused.
>
> Remember you are not synchronizing the value of OVL_UPPER_DATA between CPUs
> You don't care if user gets lower or upper dentry.
> You only care about the upper case so you can put smb_rmb() after goto
> lower line
> which will make sure CPU cannot read inconsistent upper inode state
> from before smp_wmb()
> in ovl_copy_up_meta_inode_data() after CPU read positive
> OVL_UPPER_DATA before smp_rmb().
> That's the way I understand it.
ok, I think I get it now. You are suggesting following structure.
if (!ovl_test_flag(OVL_UPPER_DATA, d_inode(dentry)))
goto lower;
smp_rmb();
return real;
So if we are returning lower, we don't have to do smp_rmb(). But if we
saw OVL_UPPER_DATA, set, then we need to do smp_rmb() to make sure upper
is consistent (just in case it was data copied up just now).
In fact, I should probably put is outside if condition block. That is.
real = ovl_dentry_upper(dentry);
if (real && (!inode || inode == d_inode(real))) {
if (!inode) {
err = ovl_check_append_only(d_inode(real), open_flags);
if (err)
return ERR_PTR(err);
if (!ovl_test_flag(OVL_UPPER_DATA, d_inode(dentry)))
goto lower;
}
smp_rmb();
return real;
}
Vivek
next prev parent reply other threads:[~2017-10-19 15:59 UTC|newest]
Thread overview: 48+ messages / expand[flat|nested] mbox.gz Atom feed top
2017-10-17 21:05 [RFC PATCH 00/11][V4] overlayfs: Delayed copy up of data Vivek Goyal
2017-10-17 21:05 ` [PATCH 01/11] ovl: Create origin xattr on copy up for all files Vivek Goyal
2017-10-18 4:09 ` Amir Goldstein
2017-10-18 12:55 ` Vivek Goyal
2017-10-18 13:56 ` Amir Goldstein
2017-10-17 21:05 ` [PATCH 02/11] ovl: ovl_check_setxattr() get rid of redundant -EOPNOTSUPP check Vivek Goyal
2017-10-18 4:11 ` Amir Goldstein
2017-10-17 21:05 ` [PATCH 03/11] ovl: During copy up, first copy up metadata and then data Vivek Goyal
2017-10-18 4:13 ` Amir Goldstein
2017-10-18 4:39 ` Amir Goldstein
2017-10-17 21:05 ` [PATCH 04/11] ovl: Provide a mount option metacopy=on/off for metadata copyup Vivek Goyal
2017-10-18 4:31 ` Amir Goldstein
2017-10-18 13:03 ` Vivek Goyal
2017-10-18 14:09 ` Amir Goldstein
2017-10-18 14:26 ` Vivek Goyal
2017-10-18 14:38 ` Amir Goldstein
2017-10-18 14:10 ` Vivek Goyal
2017-10-18 14:26 ` Amir Goldstein
2017-10-17 21:05 ` [PATCH 05/11] ovl: Copy up only metadata during copy up where it makes sense Vivek Goyal
2017-10-18 4:46 ` Amir Goldstein
2017-10-17 21:05 ` [PATCH 06/11] ovl: Set xattr OVL_XATTR_METACOPY on upper file Vivek Goyal
2017-10-18 4:57 ` Amir Goldstein
2017-10-18 13:30 ` Vivek Goyal
2017-10-17 21:05 ` [PATCH 07/11] ovl: Fix ovl_getattr() to get number of blocks from lower Vivek Goyal
2017-10-18 5:01 ` Amir Goldstein
2017-10-18 13:39 ` Vivek Goyal
2017-10-17 21:05 ` [PATCH 08/11] ovl: Set OVL_METACOPY flag during ovl_lookup() Vivek Goyal
2017-10-18 5:06 ` Amir Goldstein
2017-10-18 13:53 ` Vivek Goyal
2017-10-17 21:05 ` [PATCH 09/11] ovl: Return lower dentry if only metadata copy up took place Vivek Goyal
2017-10-18 5:07 ` Amir Goldstein
2017-10-17 21:05 ` [PATCH 10/11] ovl: Introduce read/write barriers around metacopy flag update Vivek Goyal
2017-10-18 5:19 ` Amir Goldstein
2017-10-18 15:32 ` Vivek Goyal
2017-10-18 16:05 ` Amir Goldstein
2017-10-17 21:05 ` [PATCH 11/11] ovl: Put barriers to order oi->__upperdentry and OVL_METACOPY update Vivek Goyal
2017-10-18 5:40 ` Amir Goldstein
2017-10-19 13:00 ` Vivek Goyal
2017-10-19 13:21 ` Amir Goldstein
2017-10-19 14:58 ` Vivek Goyal
2017-10-19 15:08 ` Amir Goldstein
2017-10-19 15:22 ` Vivek Goyal
2017-10-19 15:39 ` Amir Goldstein
2017-10-19 15:59 ` Vivek Goyal [this message]
2017-10-19 16:33 ` Amir Goldstein
2017-10-19 20:33 ` Vivek Goyal
2017-10-20 4:09 ` Amir Goldstein
2017-10-20 15:41 ` Vivek Goyal
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20171019155932.GA24029@redhat.com \
--to=vgoyal@redhat.com \
--cc=amir73il@gmail.com \
--cc=linux-unionfs@vger.kernel.org \
--cc=miklos@szeredi.hu \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).