On 2/26/20 6:48 AM, Brian Foster wrote:
On Tue, Feb 25, 2020 at 10:36:18PM -0700, Allison Collins wrote:
On 2/25/20 6:34 AM, Brian Foster wrote:
On Mon, Feb 24, 2020 at 04:14:48PM -0700, Allison Collins wrote:
On 2/24/20 8:25 AM, Brian Foster wrote:
On Sat, Feb 22, 2020 at 07:06:05PM -0700, Allison Collins wrote:
This patch modifies the attr remove routines to be delay ready. This means they no
longer roll or commit transactions, but instead return -EAGAIN to have the calling
routine roll and refresh the transaction. In this series, xfs_attr_remove_args has
become xfs_attr_remove_iter, which uses a sort of state machine like switch to keep
track of where it was when EAGAIN was returned. xfs_attr_node_removename has also
been modified to use the switch, and a new version of xfs_attr_remove_args
consists of a simple loop to refresh the transaction until the operation is
completed.
This patch also adds a new struct xfs_delattr_context, which we will use to keep
track of the current state of an attribute operation. The new xfs_delattr_state
enum is used to track various operations that are in progress so that we know not
to repeat them, and resume where we left off before EAGAIN was returned to cycle
out the transaction. Other members take the place of local variables that need
to retain their values across multiple function recalls.
Below is a state machine diagram for attr remove operations. The XFS_DAS_* states
indicate places where the function would return -EAGAIN, and then immediately
resume from after being recalled by the calling function. States marked as a
"subroutine state" indicate that they belong to a subroutine, and so the calling
function needs to pass them back to that subroutine to allow it to finish where
it left off. But they otherwise do not have a role in the calling function other
than just passing through.
xfs_attr_remove_iter()
XFS_DAS_RM_SHRINK ─┐
(subroutine state) │
│
XFS_DAS_RMTVAL_REMOVE ─┤
(subroutine state) │
└─>xfs_attr_node_removename()
│
v
need to remove
┌─n── rmt blocks?
│ │
│ y
│ │
│ v
│ ┌─>XFS_DAS_RMTVAL_REMOVE
│ │ │
│ │ v
│ └──y── more blks
│ to remove?
│ │
│ n
│ │
│ v
│ need to
└─────> shrink tree? ─n─┐
│ │
y │
│ │
v │
XFS_DAS_RM_SHRINK │
│ │
v │
done <─────┘
Wow. :P I guess I have nothing against verbose commit logs, but I wonder
how useful this level of documentation is for a patch that shouldn't
really change the existing flow of the operation.
Yes Darrick had requested a diagram in the last review, so I had put this
together. I wasnt sure where the best place to put it even was, so I put it
here at least for now. I have no idea if there is a limit on commit message
length, but if there is, I'm pretty sure I blew right past it in this patch
and the next. Maybe if anything it can just be here for now while we work
through things?
No problem.. if it's useful it's good to have a record of out around
somewhere until the end result is more stabilized and we can determine
whether this warrants a permanent home somewhere in the code.
Signed-off-by: Allison Collins <allison.henderson@xxxxxxxxxx>
---
fs/xfs/libxfs/xfs_attr.c | 114 +++++++++++++++++++++++++++++++++++++------
fs/xfs/libxfs/xfs_attr.h | 1 +
fs/xfs/libxfs/xfs_da_btree.h | 30 ++++++++++++
fs/xfs/scrub/common.c | 2 +
fs/xfs/xfs_acl.c | 2 +
fs/xfs/xfs_attr_list.c | 1 +
fs/xfs/xfs_ioctl.c | 2 +
fs/xfs/xfs_ioctl32.c | 2 +
fs/xfs/xfs_iops.c | 2 +
fs/xfs/xfs_xattr.c | 1 +
10 files changed, 141 insertions(+), 16 deletions(-)
diff --git a/fs/xfs/libxfs/xfs_attr.c b/fs/xfs/libxfs/xfs_attr.c
index 5d73bdf..cd3a3f7 100644
--- a/fs/xfs/libxfs/xfs_attr.c
+++ b/fs/xfs/libxfs/xfs_attr.c
@@ -368,11 +368,60 @@ xfs_has_attr(
*/
int
xfs_attr_remove_args(
+ struct xfs_da_args *args)
+{
+ int error = 0;
+ int err2 = 0;
+
+ do {
+ error = xfs_attr_remove_iter(args);
+ if (error && error != -EAGAIN)
+ goto out;
+
I'm a little confused on the logic of this loop given that the only
caller commits the transaction (which also finishes dfops). IOW, it
seems we shouldn't ever need to finish/roll when error != -EAGAIN. If
that is the case, this can be simplified to something like:
Well, we need to do it when error == -EAGAIN or 0, right? Which I think
better imitates the defer_finish routines. That's why a lot of the existing
code that just finishes off with a transaction just sort of gets sawed off
at the end. Otherwise they would need one more state just to return -EAGAIN
as the last thing they have to do. Did that make sense?
Hmm.. I could just be missing something or not far along enough in the
series. Can you point me at an example of where we need to finish/roll
before the caller of xfs_attr_remove_args() commits the transaction?
Ok, in looking for an example, realized all such examples appear in the next
patch ;-) So maybe we can get away with simplifying it in this patch.
Ah, Ok. Yeah, I think that would be best so long as it is correct, since
right now at least we have separate xfs_attr_[set|remove]_args() loop
functions and I didn't see any code that warranted the extra roll in the
remove path.
For the next patch though, it's any place the roll/finish disappears, and an
"return -EAGAIN" does not. For example, at the end of
xfs_attr_leaf_addname.
I see, thanks. Hmmm... so I think that particular example is basically a
programming pattern thing moreso than a functional requirement. I.e.,
the current _clearflag() function clears the flag and rolls the
transaction perhaps simply so it can be reliably used in different
contexts. The use in the _addname() case is functionally spurious afaict
because we roll the transaction only to make no further changes and then
commit the final transaction in the higher level code.
I could see leaving the loop as is if this were the case for every exit
path back to xfs_attr_set_args(), but is that really the case? If not,
haven't we introduced a spurious roll for any zero return back to the
_args() function? I think it might be best to fix up the loop to not
roll on error == 0, explicitly plumb in the -EAGAIN in those spurious
cases like _addname() where we currently roll, and then come up with a
follow up patch to remove the ones that end up as spurious. That way
we're not conflating too much refactoring with functional change and can
review/document the functional change independently (i.e., if removing
one of those rolls ends up introducing a bug, we don't have to revert an
entire refactoring patch to restore original behavior).
Now that I think of it, the better option is probably to remove the
xfs_trans_roll_inode() call from _addname() first, before these patches
introduce the delay ready infrastructure, since it's already isolated as
spurious at that point. That should be a simple patch with a
clear/obvious explanation.
Alrighty then, got it. I will add these suggestions into the next
version. Thanks again for all the reviews!
Allison
Brian
int
xfs_attr_remove_args(
struct xfs_da_args *args)
{
int error;
do {
error = xfs_attr_remove_iter(args);
if (error != -EAGAIN)
break;
if (args->dac.flags & XFS_DAC_FINISH_TRANS) {
args->dac.flags &= ~XFS_DAC_FINISH_TRANS;
error = xfs_defer_finish(&args->trans);
if (error)
break;
}
error = xfs_trans_roll_inode(&args->trans, args->dp);
if (error)
break;
} while (true);
return error;
}
That has the added benefit of eliminating the whole err2 pattern, which
always strikes me as a landmine.
+ if (args->dac.flags & XFS_DAC_FINISH_TRANS) {
BTW, _FINISH_TRANS also seems misnamed given that we finish deferred
operations, not necessarily the transaction. XFS_DAC_DEFER_FINISH?
Sure, will update
+ args->dac.flags &= ~XFS_DAC_FINISH_TRANS;
+
+ err2 = xfs_defer_finish(&args->trans);
+ if (err2) {
+ error = err2;
+ goto out;
+ }
+ }
+
+ err2 = xfs_trans_roll_inode(&args->trans, args->dp);
+ if (err2) {
+ error = err2;
+ goto out;
+ }
+
+ } while (error == -EAGAIN);
+out:
+ return error;
+}
+
+/*
+ * Remove the attribute specified in @args.
+ *
+ * This function may return -EAGAIN to signal that the transaction needs to be
+ * rolled. Callers should continue calling this function until they receive a
+ * return value other than -EAGAIN.
+ */
+int
+xfs_attr_remove_iter(
struct xfs_da_args *args)
{
struct xfs_inode *dp = args->dp;
int error;
+ /* State machine switch */
+ switch (args->dac.dela_state) {
+ case XFS_DAS_RM_SHRINK:
+ case XFS_DAS_RMTVAL_REMOVE:
+ goto node;
+ default:
+ break;
+ }
+
if (!xfs_inode_hasattr(dp)) {
error = -ENOATTR;
} else if (dp->i_d.di_aformat == XFS_DINODE_FMT_LOCAL) {
@@ -381,6 +430,7 @@ xfs_attr_remove_args(
} else if (xfs_bmap_one_block(dp, XFS_ATTR_FORK)) {
error = xfs_attr_leaf_removename(args);
} else {
+node:
error = xfs_attr_node_removename(args);
}
@@ -895,9 +945,8 @@ xfs_attr_leaf_removename(
/* bp is gone due to xfs_da_shrink_inode */
if (error)
return error;
- error = xfs_defer_finish(&args->trans);
- if (error)
- return error;
+
+ args->dac.flags |= XFS_DAC_FINISH_TRANS;
}
return 0;
}
@@ -1218,6 +1267,11 @@ xfs_attr_node_addname(
* This will involve walking down the Btree, and may involve joining
* leaf nodes and even joining intermediate nodes up to and including
* the root node (a special case of an intermediate node).
+ *
+ * This routine is meant to function as either an inline or delayed operation,
+ * and may return -EAGAIN when the transaction needs to be rolled. Calling
+ * functions will need to handle this, and recall the function until a
+ * successful error code is returned.
*/
STATIC int
xfs_attr_node_removename(
@@ -1230,10 +1284,24 @@ xfs_attr_node_removename(
struct xfs_inode *dp = args->dp;
trace_xfs_attr_node_removename(args);
+ state = args->dac.da_state;
+ blk = args->dac.blk;
+
+ /* State machine switch */
+ switch (args->dac.dela_state) {
+ case XFS_DAS_RMTVAL_REMOVE:
+ goto rm_node_blks;
+ case XFS_DAS_RM_SHRINK:
+ goto rm_shrink;
+ default:
+ break;
+ }
error = xfs_attr_node_hasname(args, &state);
if (error != -EEXIST)
goto out;
+ else
+ error = 0;
This doesn't look necessary.
Well, at this point error has to be -EEXIST. Which is great because we need
the attr to exist, but we dont want to return that as error for this
function. Which can happen if error is not otherwise set.
AFAICT every codepath after this assigns error one way or another before
it's returned. There's another error = 0 assignment just before the out:
label.
Ok, I see it. Will remove.
/*
* If there is an out-of-line value, de-allocate the blocks.
@@ -1243,6 +1311,14 @@ xfs_attr_node_removename(
blk = &state->path.blk[ state->path.active-1 ];
ASSERT(blk->bp != NULL);
ASSERT(blk->magic == XFS_ATTR_LEAF_MAGIC);
+
+ /*
+ * Store blk and state in the context incase we need to cycle out the
+ * transaction
+ */
+ args->dac.blk = blk;
+ args->dac.da_state = state;
+
if (args->rmtblkno > 0) {
/*
* Fill in disk block numbers in the state structure
@@ -1261,13 +1337,21 @@ xfs_attr_node_removename(
if (error)
goto out;
- error = xfs_trans_roll_inode(&args->trans, args->dp);
+ error = xfs_attr_rmtval_invalidate(args);
Remind me why we lose the above trans roll? I vaguely recall that this
was intentional, but I could be mistaken...
I think we removed it in v5. We used to have a XFS_DAS_RM_INVALIDATE
state, but then we reasoned that because these are just in-core changes, we
didnt need it, so we eliminated this state entirely.
Maybe i just add a comment here? Just as a reminder
Ah, Ok. Normally I'd say document things like this in the commit log so
we don't lose track, though I don't know how much space we have there.
;)
Ok, I'll see if I can squeeze in a few more lines :-)
if (error)
goto out;
+ }
- error = xfs_attr_rmtval_remove(args);
- if (error)
- goto out;
+rm_node_blks:
+
+ if (args->rmtblkno > 0) {
+ error = xfs_attr_rmtval_unmap(args);
+
+ if (error) {
+ if (error == -EAGAIN)
+ args->dac.dela_state = XFS_DAS_RMTVAL_REMOVE;
Might be helpful for the code labels to match the state names. I.e., use
das_rmtval_remove: for the label above.
Sure, I can update add the das prefix.
+ return error;
+ }
/*
* Refill the state structure with buffers, the prior calls
@@ -1293,17 +1377,15 @@ xfs_attr_node_removename(
error = xfs_da3_join(state);
if (error)
goto out;
- error = xfs_defer_finish(&args->trans);
- if (error)
- goto out;
- /*
- * Commit the Btree join operation and start a new trans.
- */
- error = xfs_trans_roll_inode(&args->trans, dp);
- if (error)
- goto out;
+
+ args->dac.flags |= XFS_DAC_FINISH_TRANS;
+ args->dac.dela_state = XFS_DAS_RM_SHRINK;
+ return -EAGAIN;
}
+rm_shrink:
+ args->dac.dela_state = XFS_DAS_RM_SHRINK;
+
There's an xfs_defer_finish() call further down this function. Should
that be replaced with the flag?
Finally, I mentioned in a previous review that this function should
probably be further broken down before fitting in the state management
stuff. It doesn't look like that happened so I've attached a diff that
is just intended to give an idea of what I mean by sectioning off the
hunks that might be able to break down into helpers. The helpers
wouldn't contain any state management, so we create a clear separation
between the state code and functional components.
Yes, it's xfs_attr_node_shrink in patch 15. I moved it to another patch to
try and keep the activity in this one to a minimum. Apologies if it
surprised you! And then i mistakenly had taken the XFS_DAC_FINISH_TRANS
flag with it. I meant to keep all the state machine stuff here. Will fix!
Ok, I might have just not got there yet.
I think this initial
refactoring would make the introduction of state much more simple
I guess I didn't think people would be partial to introducing helpers before
or after the state logic. I put them after in this set because the states
are visible now, so I though it would make the goal of modularizing code
between the states more clear to folks. Do you think I should move it back
behind the state machine patches?
I do think the refactoring should be done first. This does make it more
challenging for the developer (IMO) because I know I'd probably have to
hack around with the state bits to have a better idea of how to refactor
things in some cases, and then go back and retrofit the refactoring.
The advantage is that the heavy lifting in this series becomes agnostic
to the state bits. Refactoring patches are easier to review and we can
make progress because there's less of a need to carry those out of tree
through however many versions of the state code we'll need before
getting it merged. Once the code is sufficiently factored, the state
code should be much simpler to introduce and review since we hopefully
won't be jumping around into the middle of functions, multiple branches
of logic deep, etc.
(I see Dave commented similarly on a couple of the subsequent patches. I
100% agree with the approach he describes there and that is similar to
what I was trying to describe with the diff I attached in my earlier
mail...)
Brian
Alrighty then, will move back. Thanks, and thanks again for the reviews!!
Allison
(and
perhaps alleviate the need for the huge diagram).
Well, I get the impression that people find the series sort of scary and
maybe the diagrams help them a bit. Maybe we can take them out later after
people feel like they are comfortable with things?
It might also be
interesting to see how much of the result could be folded up further
into _removename_iter()...
Yes, I think that is the goal we're reaching for. I will add the other
helpers I see in your diff too.
Thanks for the reviews!
Allison
Brian
/*
* If the result is small enough, push it all into the inode.
*/
diff --git a/fs/xfs/libxfs/xfs_attr.h b/fs/xfs/libxfs/xfs_attr.h
index ce7b039..ea873a5 100644
--- a/fs/xfs/libxfs/xfs_attr.h
+++ b/fs/xfs/libxfs/xfs_attr.h
@@ -155,6 +155,7 @@ int xfs_attr_set_args(struct xfs_da_args *args);
int xfs_attr_remove(struct xfs_inode *dp, struct xfs_name *name, int flags);
int xfs_has_attr(struct xfs_da_args *args);
int xfs_attr_remove_args(struct xfs_da_args *args);
+int xfs_attr_remove_iter(struct xfs_da_args *args);
int xfs_attr_list(struct xfs_inode *dp, char *buffer, int bufsize,
int flags, struct attrlist_cursor_kern *cursor);
bool xfs_attr_namecheck(const void *name, size_t length);
diff --git a/fs/xfs/libxfs/xfs_da_btree.h b/fs/xfs/libxfs/xfs_da_btree.h
index 14f1be3..3c78498 100644
--- a/fs/xfs/libxfs/xfs_da_btree.h
+++ b/fs/xfs/libxfs/xfs_da_btree.h
@@ -50,9 +50,39 @@ enum xfs_dacmp {
};
/*
+ * Enum values for xfs_delattr_context.da_state
+ *
+ * These values are used by delayed attribute operations to keep track of where
+ * they were before they returned -EAGAIN. A return code of -EAGAIN signals the
+ * calling function to roll the transaction, and then recall the subroutine to
+ * finish the operation. The enum is then used by the subroutine to jump back
+ * to where it was and resume executing where it left off.
+ */
+enum xfs_delattr_state {
+ XFS_DAS_RM_SHRINK, /* We are shrinking the tree */
+ XFS_DAS_RMTVAL_REMOVE, /* We are removing remote value blocks */
+};
+
+/*
+ * Defines for xfs_delattr_context.flags
+ */
+#define XFS_DAC_FINISH_TRANS 0x1 /* indicates to finish the transaction */
+
+/*
+ * Context used for keeping track of delayed attribute operations
+ */
+struct xfs_delattr_context {
+ struct xfs_da_state *da_state;
+ struct xfs_da_state_blk *blk;
+ unsigned int flags;
+ enum xfs_delattr_state dela_state;
+};
+
+/*
* Structure to ease passing around component names.
*/
typedef struct xfs_da_args {
+ struct xfs_delattr_context dac; /* context used for delay attr ops */
struct xfs_da_geometry *geo; /* da block geometry */
struct xfs_name name; /* name, length and argument flags*/
uint8_t filetype; /* filetype of inode for directories */
diff --git a/fs/xfs/scrub/common.c b/fs/xfs/scrub/common.c
index 1887605..9a649d1 100644
--- a/fs/xfs/scrub/common.c
+++ b/fs/xfs/scrub/common.c
@@ -24,6 +24,8 @@
#include "xfs_rmap_btree.h"
#include "xfs_log.h"
#include "xfs_trans_priv.h"
+#include "xfs_da_format.h"
+#include "xfs_da_btree.h"
#include "xfs_attr.h"
#include "xfs_reflink.h"
#include "scrub/scrub.h"
diff --git a/fs/xfs/xfs_acl.c b/fs/xfs/xfs_acl.c
index 42ac847..d65e6d8 100644
--- a/fs/xfs/xfs_acl.c
+++ b/fs/xfs/xfs_acl.c
@@ -10,6 +10,8 @@
#include "xfs_trans_resv.h"
#include "xfs_mount.h"
#include "xfs_inode.h"
+#include "xfs_da_format.h"
+#include "xfs_da_btree.h"
#include "xfs_attr.h"
#include "xfs_trace.h"
#include "xfs_error.h"
diff --git a/fs/xfs/xfs_attr_list.c b/fs/xfs/xfs_attr_list.c
index d37743b..881b9a4 100644
--- a/fs/xfs/xfs_attr_list.c
+++ b/fs/xfs/xfs_attr_list.c
@@ -12,6 +12,7 @@
#include "xfs_trans_resv.h"
#include "xfs_mount.h"
#include "xfs_da_format.h"
+#include "xfs_da_btree.h"
#include "xfs_inode.h"
#include "xfs_trans.h"
#include "xfs_bmap.h"
diff --git a/fs/xfs/xfs_ioctl.c b/fs/xfs/xfs_ioctl.c
index 28c07c9..7c1d9da 100644
--- a/fs/xfs/xfs_ioctl.c
+++ b/fs/xfs/xfs_ioctl.c
@@ -15,6 +15,8 @@
#include "xfs_iwalk.h"
#include "xfs_itable.h"
#include "xfs_error.h"
+#include "xfs_da_format.h"
+#include "xfs_da_btree.h"
#include "xfs_attr.h"
#include "xfs_bmap.h"
#include "xfs_bmap_util.h"
diff --git a/fs/xfs/xfs_ioctl32.c b/fs/xfs/xfs_ioctl32.c
index 769581a..d504f8f 100644
--- a/fs/xfs/xfs_ioctl32.c
+++ b/fs/xfs/xfs_ioctl32.c
@@ -17,6 +17,8 @@
#include "xfs_itable.h"
#include "xfs_fsops.h"
#include "xfs_rtalloc.h"
+#include "xfs_da_format.h"
+#include "xfs_da_btree.h"
#include "xfs_attr.h"
#include "xfs_ioctl.h"
#include "xfs_ioctl32.h"
diff --git a/fs/xfs/xfs_iops.c b/fs/xfs/xfs_iops.c
index e85bbf5..a2d299f 100644
--- a/fs/xfs/xfs_iops.c
+++ b/fs/xfs/xfs_iops.c
@@ -13,6 +13,8 @@
#include "xfs_inode.h"
#include "xfs_acl.h"
#include "xfs_quota.h"
+#include "xfs_da_format.h"
+#include "xfs_da_btree.h"
#include "xfs_attr.h"
#include "xfs_trans.h"
#include "xfs_trace.h"
diff --git a/fs/xfs/xfs_xattr.c b/fs/xfs/xfs_xattr.c
index 74133a5..d8dc72d 100644
--- a/fs/xfs/xfs_xattr.c
+++ b/fs/xfs/xfs_xattr.c
@@ -10,6 +10,7 @@
#include "xfs_log_format.h"
#include "xfs_da_format.h"
#include "xfs_inode.h"
+#include "xfs_da_btree.h"
#include "xfs_attr.h"
#include "xfs_acl.h"
--
2.7.4
