Skip to contentSkip to conversation

ceph: fix statx AT_STATX_DONT_SYNC vs AT_STATX_FORCE_SYNC check

RFC PATCH In review [email protected] 3 participants 10 messages Apr 11, 2022 – Apr 19, 2022 Ceph client

lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 ceph: fix statx AT_STATX_DONT_SYNC vs AT_STATX_FORCE_SYNC check

From: Xiubo Li <[email protected]>

From the posix and the initial statx supporting commit comments, the AT_STATX_DONT_SYNC is a lightweight stat flag and the AT_STATX_FORCE_SYNC is a heaverweight one. And also checked all the other current usage about these two flags they are all doing the same, that is only when the AT_STATX_FORCE_SYNC is not set and the AT_STATX_DONT_SYNC is set will they skip sync retriving the attributes from storage.

---
 fs/ceph/inode.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/ceph/inode.c +1 -1
index 6788a1f88eb6..1ee6685def83 100644
@@ -2887,7 +2887,7 @@ int ceph_getattr(struct user_namespace *mnt_userns, const struct path *path,
 		return -ESTALE;
 
 	/* Skip the getattr altogether if we're asked not to sync */
-	if (!(flags & AT_STATX_DONT_SYNC)) {
+	if ((flags & AT_STATX_SYNC_TYPE) != AT_STATX_DONT_SYNC) {
 		err = ceph_do_getattr(inode,
 				statx_to_caps(request_mask, inode->i_mode),
 				flags & AT_STATX_FORCE_SYNC);
Signature · 1 line
2.27.0
#1 lore raw
Quoted 20 lines · On Mon, 2022-04-11 at 17:34 +0800, [email protected] wrote:

From: Xiubo Li <[email protected]>

From the posix and the initial statx supporting commit comments, the AT_STATX_DONT_SYNC is a lightweight stat flag and the AT_STATX_FORCE_SYNC is a heaverweight one. And also checked all the other current usage about these two flags they are all doing the same, that is only when the AT_STATX_FORCE_SYNC is not set and the AT_STATX_DONT_SYNC is set will they skip sync retriving the attributes from storage.

---
 fs/ceph/inode.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/ceph/inode.c +0 -0
index 6788a1f88eb6..1ee6685def83 100644

fs/ceph/inode.c int ceph_getattr(struct user_namespace *mnt_userns, const struct path *path, · line 2893

Quoted 8 lines
fs/ceph/inode.c +1 -1
@@ -2887,7 +2887,7 @@ int ceph_getattr(struct user_namespace *mnt_userns, const struct path *path,
 		return -ESTALE;
 
 	/* Skip the getattr altogether if we're asked not to sync */
-	if (!(flags & AT_STATX_DONT_SYNC)) {
+	if ((flags & AT_STATX_SYNC_TYPE) != AT_STATX_DONT_SYNC) {
 		err = ceph_do_getattr(inode,
 				statx_to_caps(request_mask, inode->i_mode),
 				flags & AT_STATX_FORCE_SYNC);

I don't get it.

The only way I can see that this is a problem is if someone sent down a mask with both DONT_SYNC and FORCE_SYNC set in it, and in that case I don't see that ignoring FORCE_SYNC would be wrong...

Signature · 1 line
Jeff Layton <[email protected]>
author
#2 lore raw
Quoted 31 lines · Jeff Layton
Quoted 20 lines · On Mon, 2022-04-11 at 17:34 +0800, [email protected] wrote:

From: Xiubo Li <[email protected]>

From the posix and the initial statx supporting commit comments,
the AT_STATX_DONT_SYNC is a lightweight stat flag and the AT_STATX_FORCE_SYNC is a heaverweight one. And also checked all the other current usage about these two flags they are all doing the same, that is only when the AT_STATX_FORCE_SYNC is not set and the AT_STATX_DONT_SYNC is set will they skip sync retriving the attributes from storage.

---
  fs/ceph/inode.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
fs/ceph/inode.c +0 -0
index 6788a1f88eb6..1ee6685def83 100644

fs/ceph/inode.c int ceph_getattr(struct user_namespace *mnt_userns, const struct path *path, · line 2893

Quoted 8 lines
fs/ceph/inode.c +1 -1
@@ -2887,7 +2887,7 @@ int ceph_getattr(struct user_namespace *mnt_userns, const struct path *path,
  		return -ESTALE;
 
  	/* Skip the getattr altogether if we're asked not to sync */
-	if (!(flags & AT_STATX_DONT_SYNC)) {
+	if ((flags & AT_STATX_SYNC_TYPE) != AT_STATX_DONT_SYNC) {
  		err = ceph_do_getattr(inode,
  				statx_to_caps(request_mask, inode->i_mode),
  				flags & AT_STATX_FORCE_SYNC);

I don't get it.

Quoted 4 lines

The only way I can see that this is a problem is if someone sent down a mask with both DONT_SYNC and FORCE_SYNC set in it, and in that case I don't see that ignoring FORCE_SYNC would be wrong...

There has 3 cases for the flags:

case1: flags & AT_STATX_SYNC_TYPE == 0

case2: flags & AT_STATX_SYNC_TYPE == AT_STATX_DONT_SYNC

case3: flags & AT_STATX_SYNC_TYPE == AT_STATX_DONT_SYNC | AT_STATX_FORCE_SYNC

Only in case2, which is only the DONT_SYNC bit is set, will ignore calling ceph_do_getattr() here. And for case3 it will ignore the DONT_SYNC bit.

-- Xiubo

#3 lore raw
Quoted 42 lines · Xiubo Li
Quoted 31 lines · Jeff Layton
Quoted 20 lines · On Mon, 2022-04-11 at 17:34 +0800, [email protected] wrote:

From: Xiubo Li <[email protected]>

From the posix and the initial statx supporting commit comments,
the AT_STATX_DONT_SYNC is a lightweight stat flag and the AT_STATX_FORCE_SYNC is a heaverweight one. And also checked all the other current usage about these two flags they are all doing the same, that is only when the AT_STATX_FORCE_SYNC is not set and the AT_STATX_DONT_SYNC is set will they skip sync retriving the attributes from storage.

---
  fs/ceph/inode.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
fs/ceph/inode.c +0 -0
index 6788a1f88eb6..1ee6685def83 100644

fs/ceph/inode.c int ceph_getattr(struct user_namespace *mnt_userns, const struct path *path, · line 2893

Quoted 8 lines
fs/ceph/inode.c +1 -1
@@ -2887,7 +2887,7 @@ int ceph_getattr(struct user_namespace *mnt_userns, const struct path *path,
  		return -ESTALE;
 
  	/* Skip the getattr altogether if we're asked not to sync */
-	if (!(flags & AT_STATX_DONT_SYNC)) {
+	if ((flags & AT_STATX_SYNC_TYPE) != AT_STATX_DONT_SYNC) {
  		err = ceph_do_getattr(inode,
  				statx_to_caps(request_mask, inode->i_mode),
  				flags & AT_STATX_FORCE_SYNC);

I don't get it.

Quoted 4 lines

The only way I can see that this is a problem is if someone sent down a mask with both DONT_SYNC and FORCE_SYNC set in it, and in that case I don't see that ignoring FORCE_SYNC would be wrong...

There has 3 cases for the flags:

case1: flags & AT_STATX_SYNC_TYPE == 0

case2: flags & AT_STATX_SYNC_TYPE == AT_STATX_DONT_SYNC

Quoted 8 lines

case3: flags & AT_STATX_SYNC_TYPE == AT_STATX_DONT_SYNC | AT_STATX_FORCE_SYNC

Only in case2, which is only the DONT_SYNC bit is set, will ignore calling ceph_do_getattr() here. And for case3 it will ignore the DONT_SYNC bit.

Sure, but the patch doesn't functionally change the behavior of the code. It may make the condition more idiomatic to read, but I don't think there is a bug here.

Signature · 1 line
Jeff Layton <[email protected]>
author
#4 lore raw
Quoted 47 lines · Jeff Layton
Quoted 46 lines · Xiubo Li
Quoted 31 lines · Jeff Layton
Quoted 20 lines · On Mon, 2022-04-11 at 17:34 +0800, [email protected] wrote:

From: Xiubo Li <[email protected]>

From the posix and the initial statx supporting commit comments,
the AT_STATX_DONT_SYNC is a lightweight stat flag and the AT_STATX_FORCE_SYNC is a heaverweight one. And also checked all the other current usage about these two flags they are all doing the same, that is only when the AT_STATX_FORCE_SYNC is not set and the AT_STATX_DONT_SYNC is set will they skip sync retriving the attributes from storage.

---
   fs/ceph/inode.c | 2 +-
   1 file changed, 1 insertion(+), 1 deletion(-)
fs/ceph/inode.c +0 -0
index 6788a1f88eb6..1ee6685def83 100644

fs/ceph/inode.c int ceph_getattr(struct user_namespace *mnt_userns, const struct path *path, · line 2893

Quoted 8 lines
fs/ceph/inode.c +1 -1
@@ -2887,7 +2887,7 @@ int ceph_getattr(struct user_namespace *mnt_userns, const struct path *path,
   		return -ESTALE;
 
   	/* Skip the getattr altogether if we're asked not to sync */
-	if (!(flags & AT_STATX_DONT_SYNC)) {
+	if ((flags & AT_STATX_SYNC_TYPE) != AT_STATX_DONT_SYNC) {
   		err = ceph_do_getattr(inode,
   				statx_to_caps(request_mask, inode->i_mode),
   				flags & AT_STATX_FORCE_SYNC);

I don't get it.

Quoted 4 lines

The only way I can see that this is a problem is if someone sent down a mask with both DONT_SYNC and FORCE_SYNC set in it, and in that case I don't see that ignoring FORCE_SYNC would be wrong...

There has 3 cases for the flags:

case1: flags & AT_STATX_SYNC_TYPE == 0

case2: flags & AT_STATX_SYNC_TYPE == AT_STATX_DONT_SYNC

case3: flags & AT_STATX_SYNC_TYPE == AT_STATX_DONT_SYNC | AT_STATX_FORCE_SYNC

Quoted 7 lines
Quoted 4 lines

Only in case2, which is only the DONT_SYNC bit is set, will ignore calling ceph_do_getattr() here. And for case3 it will ignore the DONT_SYNC bit.

Sure, but the patch doesn't functionally change the behavior of the code. It may make the condition more idiomatic to read, but I don't think there is a bug here.

- if (!(flags & AT_STATX_DONT_SYNC)) {

For example, in both case2 and case3 the above condition is false, right ? That means for case3 it will ignore the FORCE_SYNC always.

+ if ((flags & AT_STATX_SYNC_TYPE) != AT_STATX_DONT_SYNC) {

For exmaple in case2 the above condition is false and then it will skip calling the ceph_do_getattr(). And in case3 the above condition is true, it will call the ceph_do_getattr(..., flags & FORCE_SYNC).

The logic changed, right ?

-- Xiubo

#5 lore raw
Quoted 64 lines · Xiubo Li
Quoted 54 lines · Jeff Layton
Quoted 42 lines · Xiubo Li
Quoted 31 lines · Jeff Layton
Quoted 20 lines · On Mon, 2022-04-11 at 17:34 +0800, [email protected] wrote:

From: Xiubo Li <[email protected]>

From the posix and the initial statx supporting commit comments,
the AT_STATX_DONT_SYNC is a lightweight stat flag and the AT_STATX_FORCE_SYNC is a heaverweight one. And also checked all the other current usage about these two flags they are all doing the same, that is only when the AT_STATX_FORCE_SYNC is not set and the AT_STATX_DONT_SYNC is set will they skip sync retriving the attributes from storage.

---
   fs/ceph/inode.c | 2 +-
   1 file changed, 1 insertion(+), 1 deletion(-)
fs/ceph/inode.c +0 -0
index 6788a1f88eb6..1ee6685def83 100644

fs/ceph/inode.c int ceph_getattr(struct user_namespace *mnt_userns, const struct path *path, · line 2893

Quoted 8 lines
fs/ceph/inode.c +1 -1
@@ -2887,7 +2887,7 @@ int ceph_getattr(struct user_namespace *mnt_userns, const struct path *path,
   		return -ESTALE;
 
   	/* Skip the getattr altogether if we're asked not to sync */
-	if (!(flags & AT_STATX_DONT_SYNC)) {
+	if ((flags & AT_STATX_SYNC_TYPE) != AT_STATX_DONT_SYNC) {
   		err = ceph_do_getattr(inode,
   				statx_to_caps(request_mask, inode->i_mode),
   				flags & AT_STATX_FORCE_SYNC);

I don't get it.

Quoted 4 lines

The only way I can see that this is a problem is if someone sent down a mask with both DONT_SYNC and FORCE_SYNC set in it, and in that case I don't see that ignoring FORCE_SYNC would be wrong...

There has 3 cases for the flags:

case1: flags & AT_STATX_SYNC_TYPE == 0

case2: flags & AT_STATX_SYNC_TYPE == AT_STATX_DONT_SYNC

Quoted 8 lines

case3: flags & AT_STATX_SYNC_TYPE == AT_STATX_DONT_SYNC | AT_STATX_FORCE_SYNC

Only in case2, which is only the DONT_SYNC bit is set, will ignore calling ceph_do_getattr() here. And for case3 it will ignore the DONT_SYNC bit.

Sure, but the patch doesn't functionally change the behavior of the code. It may make the condition more idiomatic to read, but I don't think there is a bug here.

fs/ceph/inode.c +0 -1
-	if (!(flags & AT_STATX_DONT_SYNC)) {

For example, in both case2 and case3 the above condition is false, right ? That means for case3 it will ignore the FORCE_SYNC always.

fs/ceph/inode.c +1 -0
+	if ((flags & AT_STATX_SYNC_TYPE) != AT_STATX_DONT_SYNC) {
Quoted 6 lines

For exmaple in case2 the above condition is false and then it will skip calling the ceph_do_getattr(). And in case3 the above condition is true, it will call the ceph_do_getattr(..., flags & FORCE_SYNC).

The logic changed, right ?

Yes, but my argument is that sending down a mask that has AT_STATX_DONT_SYNC|AT_STATX_FORCE_SYNC makes no sense whatsoever. You're simultaneously asking it to not fetch attributes and to forcibly fetch attributes. We're within our rights to just ignore FORCE_SYNC at that point.

Arguably, this ought to be caught in vfs_statx before we ever call down into the fs driver with something like:

    if ((flags & AT_STATX_SYNC_TYPE) == (AT_STATX_DONT_SYNC|AT_STATX_FORCE_SYNC))
	    return -EINVAL;

David, should we add something like this, or is this too risky a change?

Signature · 1 line
Jeff Layton <[email protected]>
#6 lore raw
Quoted 1 line · Jeff Layton
    if ((flags & AT_STATX_SYNC_TYPE) == (AT_STATX_DONT_SYNC|AT_STATX_FORCE_SYNC))

You can't do that. DONT_SYNC and FORCE_SYNC aren't bit flags - they're an enumeration in a bit field. There's a reserved value at 0x6000 that doesn't have a symbol assigned.

David

#7 lore raw
Quoted 8 lines · David Howells
Quoted 1 line · Jeff Layton
    if ((flags & AT_STATX_SYNC_TYPE) == (AT_STATX_DONT_SYNC|AT_STATX_FORCE_SYNC))

You can't do that. DONT_SYNC and FORCE_SYNC aren't bit flags - they're an enumeration in a bit field. There's a reserved value at 0x6000 that doesn't have a symbol assigned.

Right, but nothing prevents you from setting both flags at the same time. How should we interpret such a request?

Signature · 1 line
Jeff Layton <[email protected]>
#8 lore raw
Quoted 5 lines · Jeff Layton
Quoted 4 lines · David Howells
Quoted 1 line · Jeff Layton
    if ((flags & AT_STATX_SYNC_TYPE) == (AT_STATX_DONT_SYNC|AT_STATX_FORCE_SYNC))
Quoted 7 lines
Quoted 4 lines

You can't do that. DONT_SYNC and FORCE_SYNC aren't bit flags - they're an enumeration in a bit field. There's a reserved value at 0x6000 that doesn't have a symbol assigned.

Right, but nothing prevents you from setting both flags at the same time. How should we interpret such a request?

A good question without a necessarily right answer.

Possibly we should do:

#define AT_STATX_SYNC_TYPE 0x6000 /* Type of synchronisation required from statx() */
#define AT_STATX_SYNC_AS_STAT 0x0000 /* - Do whatever stat() does */
#define AT_STATX_FORCE_SYNC 0x2000 /* - Force the attributes to be sync'd with the server */
#define AT_STATX_DONT_SYNC 0x4000 /* - Don't sync attributes with the server */
+#define AT_STATX_SYNC_RESERVED 0x6000

and give EINVAL if we see the reserved value. But also these values can be considered a hint, so possibly we should just ignore the reserved value. Oh for fsinfo()...

David

#9 lore raw
Quoted 25 lines · David Howells
Quoted 5 lines · Jeff Layton
Quoted 4 lines · David Howells
Quoted 1 line · Jeff Layton
    if ((flags & AT_STATX_SYNC_TYPE) == (AT_STATX_DONT_SYNC|AT_STATX_FORCE_SYNC))
Quoted 7 lines
Quoted 4 lines

You can't do that. DONT_SYNC and FORCE_SYNC aren't bit flags - they're an enumeration in a bit field. There's a reserved value at 0x6000 that doesn't have a symbol assigned.

Right, but nothing prevents you from setting both flags at the same time. How should we interpret such a request?

A good question without a necessarily right answer.

Possibly we should do:

#define AT_STATX_SYNC_TYPE 0x6000 /* Type of synchronisation required from statx() */
#define AT_STATX_SYNC_AS_STAT 0x0000 /* - Do whatever stat() does */
#define AT_STATX_FORCE_SYNC 0x2000 /* - Force the attributes to be sync'd with the server */
#define AT_STATX_DONT_SYNC 0x4000 /* - Don't sync attributes with the server */
+#define AT_STATX_SYNC_RESERVED 0x6000

Quoted 6 lines

and give EINVAL if we see the reserved value. But also these values can be considered a hint, so possibly we should just ignore the reserved value. Oh for fsinfo()...

David

That was what the code (pre-patch) did. If someone set DONT_SYNC|FORCE_SYNC, it would just ignore FORCE_SYNC. It's not ideal, but I suppose we're within our rights to prefer either behaviour in that case if someone sets both flags.

In hindsight, setting both should have probably caused the syscall to throw back -EINVAL, but changing that now is probably a bit dangerous.

Signature · 1 line
Jeff Layton <[email protected]>