11:18 p.m.
Hello,
In mdcache_unlink, ganesha removes the dirent from its parent by
calling mdcache_dirent_remove(). It also clears MDCACHE_TRUST_ATTRS. So, if
a subsequent GETATTR comes for the parent (which can happen in the same
compound from nfs4_op_remove), it frees all entries at the end of
mdcache_refresh_attrs(). In this scenario, isn't it possible to keep the
rest of dirents intact?
I'm thinking if we update parents attrs in unlink itself, would it prevent
any further refreshes? The use case here is large directories with frequent
additions/deletions/enumerations.
Thanks,
Pradeep
8:02 a.m.
The danger in not invalidating is that something that adds an entry races with the
deletion such that we miss the addition. Or even a deletion external to Ganesha.
Frank
From: Pradeep [mailto:pradeepthomas@gmail.com]
Sent: Sunday, February 28, 2021 11:19 PM
To: Frank Filz <ffilzlnx(a)mindspring.com>; Daniel Gryniewicz <dang(a)redhat.com>;
Ganesha-devel <devel(a)lists.nfs-ganesha.org>
Subject: [NFS-Ganesha-Devel] unlink of one dirent could invalidate all.
Hello,
In mdcache_unlink, ganesha removes the dirent from its parent by calling
mdcache_dirent_remove(). It also clears MDCACHE_TRUST_ATTRS. So, if a subsequent GETATTR
comes for the parent (which can happen in the same compound from nfs4_op_remove), it frees
all entries at the end of mdcache_refresh_attrs(). In this scenario, isn't it possible
to keep the rest of dirents intact?
I'm thinking if we update parents attrs in unlink itself, would it prevent any further
refreshes? The use case here is large directories with frequent
additions/deletions/enumerations.
Thanks,
Pradeep
8:52 a.m.
Thanks Frank. If Ganesha is the only server, do you think additions can be
missed - I see that we do mdcache_dirent_add()
from mdcache_alloc_and_check_handle(). Wouldn't that be good enough to
cover additions or are there cases we might miss it? place_new_dirent()
also invalidates if it can't place new dirent.
Thanks,
Pradeep
On Mon, Mar 1, 2021 at 8:02 AM Frank Filz <ffilzlnx(a)mindspring.com> wrote:
The danger in not invalidating is that something that adds an entry
races
with the deletion such that we miss the addition. Or even a deletion
external to Ganesha.
Frank
*From:* Pradeep [mailto:pradeepthomas@gmail.com]
*Sent:* Sunday, February 28, 2021 11:19 PM
*To:* Frank Filz <ffilzlnx(a)mindspring.com>; Daniel Gryniewicz <
dang(a)redhat.com>; Ganesha-devel <devel(a)lists.nfs-ganesha.org>
*Subject:* [NFS-Ganesha-Devel] unlink of one dirent could invalidate all.
Hello,
In mdcache_unlink, ganesha removes the dirent from its parent by
calling mdcache_dirent_remove(). It also clears MDCACHE_TRUST_ATTRS. So, if
a subsequent GETATTR comes for the parent (which can happen in the same
compound from nfs4_op_remove), it frees all entries at the end of
mdcache_refresh_attrs(). In this scenario, isn't it possible to keep the
rest of dirents intact?
I'm thinking if we update parents attrs in unlink itself, would it prevent
any further refreshes? The use case here is large directories with frequent
additions/deletions/enumerations.
Thanks,
Pradeep
1:25 p.m.
Yea, if a single Ganesha process is the only process accessing the directory, then
mdcache_dirent_add would keep additions current. Note however, that unless you are using a
FSAL that can place new dirents into chunks because the FSAL can compute the entry cookie,
the chunk cache will be invalidated by any additions to the directory.
Frank
From: Pradeep [mailto:pradeepthomas@gmail.com]
Sent: Monday, March 1, 2021 8:53 AM
To: Frank Filz <ffilzlnx(a)mindspring.com>
Cc: Daniel Gryniewicz <dang(a)redhat.com>; Ganesha-devel
<devel(a)lists.nfs-ganesha.org>
Subject: [NFS-Ganesha-Devel] Re: unlink of one dirent could invalidate all.
Thanks Frank. If Ganesha is the only server, do you think additions can be missed - I see
that we do mdcache_dirent_add() from mdcache_alloc_and_check_handle(). Wouldn't that
be good enough to cover additions or are there cases we might miss it? place_new_dirent()
also invalidates if it can't place new dirent.
Thanks,
Pradeep
On Mon, Mar 1, 2021 at 8:02 AM Frank Filz <ffilzlnx(a)mindspring.com
<mailto:ffilzlnx@mindspring.com> > wrote:
The danger in not invalidating is that something that adds an entry races with the
deletion such that we miss the addition. Or even a deletion external to Ganesha.
Frank
From: Pradeep [mailto:pradeepthomas@gmail.com <mailto:pradeepthomas@gmail.com> ]
Sent: Sunday, February 28, 2021 11:19 PM
To: Frank Filz <ffilzlnx(a)mindspring.com <mailto:ffilzlnx@mindspring.com> >;
Daniel Gryniewicz <dang(a)redhat.com <mailto:dang@redhat.com> >; Ganesha-devel
<devel(a)lists.nfs-ganesha.org <mailto:devel@lists.nfs-ganesha.org> >
Subject: [NFS-Ganesha-Devel] unlink of one dirent could invalidate all.
Hello,
In mdcache_unlink, ganesha removes the dirent from its parent by calling
mdcache_dirent_remove(). It also clears MDCACHE_TRUST_ATTRS. So, if a subsequent GETATTR
comes for the parent (which can happen in the same compound from nfs4_op_remove), it frees
all entries at the end of mdcache_refresh_attrs(). In this scenario, isn't it possible
to keep the rest of dirents intact?
I'm thinking if we update parents attrs in unlink itself, would it prevent any further
refreshes? The use case here is large directories with frequent
additions/deletions/enumerations.
Thanks,
Pradeep
4:24 p.m.
Got it. Thanks Frank.
On Mon, Mar 1, 2021 at 1:25 PM Frank Filz <ffilzlnx(a)mindspring.com> wrote:
Yea, if a single Ganesha process is the only process accessing the
directory, then mdcache_dirent_add would keep additions current. Note
however, that unless you are using a FSAL that can place new dirents into
chunks because the FSAL can compute the entry cookie, the chunk cache will
be invalidated by any additions to the directory.
Frank
*From:* Pradeep [mailto:pradeepthomas@gmail.com]
*Sent:* Monday, March 1, 2021 8:53 AM
*To:* Frank Filz <ffilzlnx(a)mindspring.com>
*Cc:* Daniel Gryniewicz <dang(a)redhat.com>; Ganesha-devel <
devel(a)lists.nfs-ganesha.org>
*Subject:* [NFS-Ganesha-Devel] Re: unlink of one dirent could invalidate
all.
Thanks Frank. If Ganesha is the only server, do you think additions can be
missed - I see that we do mdcache_dirent_add()
from mdcache_alloc_and_check_handle(). Wouldn't that be good enough to
cover additions or are there cases we might miss it? place_new_dirent()
also invalidates if it can't place new dirent.
Thanks,
Pradeep
On Mon, Mar 1, 2021 at 8:02 AM Frank Filz <ffilzlnx(a)mindspring.com> wrote:
The danger in not invalidating is that something that adds an entry races
with the deletion such that we miss the addition. Or even a deletion
external to Ganesha.
Frank
*From:* Pradeep [mailto:pradeepthomas@gmail.com]
*Sent:* Sunday, February 28, 2021 11:19 PM
*To:* Frank Filz <ffilzlnx(a)mindspring.com>; Daniel Gryniewicz <
dang(a)redhat.com>; Ganesha-devel <devel(a)lists.nfs-ganesha.org>
*Subject:* [NFS-Ganesha-Devel] unlink of one dirent could invalidate all.
Hello,
In mdcache_unlink, ganesha removes the dirent from its parent by
calling mdcache_dirent_remove(). It also clears MDCACHE_TRUST_ATTRS. So, if
a subsequent GETATTR comes for the parent (which can happen in the same
compound from nfs4_op_remove), it frees all entries at the end of
mdcache_refresh_attrs(). In this scenario, isn't it possible to keep the
rest of dirents intact?
I'm thinking if we update parents attrs in unlink itself, would it prevent
any further refreshes? The use case here is large directories with frequent
additions/deletions/enumerations.
Thanks,
Pradeep
1173
days inactive
1174
days old
4 comments
2 participants
participants (2)
-
Frank Filz -
Pradeep