Revert "ceph: fix dentry leak in splice_dentry()"
Subsystems (2)
PATCH 1/4 Revert "ceph: fix dentry leak in splice_dentry()"
This reverts commit 8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3.
splice_dentry() is used by three places. For two places, req->r_dentry is passed to splice_dentry(). In the case of error, req->r_dentry does not get updated. So splice_dentry() should not drop reference.
--- fs/ceph/inode.c | 1 - 1 file changed, 1 deletion(-)
fs/ceph/inode.c +0 -1
@@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); -dput (dn); dn = realdn;/* note realdn contains the error */ goto out; }else if (realdn) {
Signature · 1 line
2.17.1
Ilya Dryomov
maintainerQuoted 18 lines · Zheng
This reverts commit
8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3.
splice_dentry()is used by three places. For two places,req->r_dentryis passed tosplice_dentry(). In the case of error,req->r_dentrydoes not get updated. Sosplice_dentry()should not drop reference.Cc [email protected] #4.18Signed-off-by "Yan, Zheng" <[email protected]>--- fs/ceph/inode.c | 1 - 1 file changed, 1 deletion(-)fs/ceph/inode.c +0
-0
fs/ceph/inode.c static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) · line 1145
Quoted 7 lines
fs/ceph/inode.c +0
-1@@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); -dput (dn); dn = realdn;/* note realdn contains the error */ goto out; }else if (realdn) {
I think this needs a comment, at the very least. The implicit contract here is clearly way too confusing.
Thanks,
Ilya
Quoted 17 lines · Zheng
This reverts commit
8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3.
splice_dentry()is used by three places. For two places,req->r_dentryis passed tosplice_dentry(). In the case of error,req->r_dentrydoes not get updated. Sosplice_dentry()should not drop reference.Cc [email protected] #4.18Signed-off-by "Yan, Zheng" <[email protected]>--- fs/ceph/inode.c | 1 - 1 file changed, 1 deletion(-)fs/ceph/inode.c +0
-0
fs/ceph/inode.c static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) · line 1145
Quoted 7 lines
fs/ceph/inode.c +0
-1@@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); -dput (dn); dn = realdn;/* note realdn contains the error */ goto out; }else if (realdn) {
This might be ok, buI have some real concerns about splice_dentry and its callers -- particularly ceph_fill_trace:
We hold a reference to dn on entry to splice_dentry. We then call d_splice_alias and get back an error, and now we don't put the old dentry.
Fine -- we have to then expect the caller to do it. Unfortunately, the callers in ceph_fill_trace do this:
dn = splice_dentry(dn, in);
if (IS_ERR(dn)) {
err = PTR_ERR(dn);
goto done;
}
req->r_dentry = dn; /* may have spliced */The old value of dn gets clobbered once that comes back with an ERR_PTR. I guess we could claim that r_dentry will still be set to the old value at that point and that it would get cleaned up when it gets cleaned up.
But...I see this higher up in ceph_fill_trace at the end of the CEPH_MFS_OP_RENAME condition block:
dn = req->r_old_dentry; /* use old_dentry */
So now I'm worried about the case where the splice succeeds. ISTM that "dn" can represent either r_dentry or r_old_dentry at the point where splice_dentry gets called, but we only ever reset the value of r_dentry there.
If dn == r_old_dentry at the time that splice_dentry is called, and then that succeeds, we'll end up leaking the reference to r_dentry and then doing an overput on r_old_dentry.
I think it might help to establish clear "ownership" of the dentry references throughout that function. Consider zeroing out r_dentry and r_old_dentry at the time that you set the local variables? That might make this whole thing less fragile.
Signature · 1 line
Jeff Layton <[email protected]>
Quoted 56 lines · Jeff Layton
Quoted 17 lines · Zheng
This reverts commit
8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3.
splice_dentry()is used by three places. For two places,req->r_dentryis passed tosplice_dentry(). In the case of error,req->r_dentrydoes not get updated. Sosplice_dentry()should not drop reference.Cc [email protected] #4.18Signed-off-by "Yan, Zheng" <[email protected]>---
fs/ceph/inode.c| 1 -1 file changed, 1 deletion(-)fs/ceph/inode.c +0
-0fs/ceph/inode.c static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) · line 1145
Quoted 7 lines
fs/ceph/inode.c +0
-1@@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); -dput (dn); dn = realdn;/* note realdn contains the error */ goto out; }else if (realdn) {This might be ok, buI have some real concerns about
splice_dentryand its callers -- particularlyceph_fill_trace:We hold a reference to dn on entry to
splice_dentry. We then calld_splice_aliasand get back an error, and now we don't put the old dentry.Fine -- we have to then expect the caller to do it. Unfortunately, the callers in
ceph_fill_tracedo this:dn = splice_dentry(dn, in); if (IS_ERR(dn)) { err = PTR_ERR(dn); goto done; } req->r_dentry = dn; /* may have spliced */The old value of dn gets clobbered once that comes back with an
ERR_PTR. I guess we could claim thatr_dentrywill still be set to the old value at that point and that it would get cleaned up when it gets cleaned up.But...I see this higher up in
ceph_fill_traceat the end of theCEPH_MFS_OP_RENAMEcondition block:dn = req->r_old_dentry; /* use old_dentry */
Quoted 9 lines
So now I'm worried about the case where the splice succeeds. ISTM that "dn" can represent either
r_dentryorr_old_dentryat the point wheresplice_dentrygets called, but we only ever reset the value ofr_dentrythere.If dn ==
r_old_dentryat the time thatsplice_dentryis called, and then that succeeds, we'll end up leaking the reference tor_dentryand then doing an overput onr_old_dentry.
Good catch
Quoted 4 lines
I think it might help to establish clear "ownership" of the dentry references throughout that function. Consider zeroing out
r_dentryandr_old_dentryat the time that you set the local variables? That might make this whole thing less fragile.
Before the function return, we need to set req->r_dentry again. This will introduce duplicated code because there are several ‘goto’ in the function. How about following change.
- dn = req->r_old_dentry; /* use old_dentry */ + /* swap r_dentry and r_old_dentry */ + req->r_dentry = req->r_old_dentry; + req->r_old_dentry = dn; + dn = req->r_dentry;
Regards
Yan, Zheng
Quoted 2 lines
Signature · 1 line
Jeff Layton <[email protected]>
Quoted 75 lines · Zheng
Quoted 54 lines · Jeff Layton
Quoted 17 lines · Zheng
This reverts commit
8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3.
splice_dentry()is used by three places. For two places,req->r_dentryis passed tosplice_dentry(). In the case of error,req->r_dentrydoes not get updated. Sosplice_dentry()should not drop reference.Cc [email protected] #4.18Signed-off-by "Yan, Zheng" <[email protected]>---
fs/ceph/inode.c| 1 -1 file changed, 1 deletion(-)fs/ceph/inode.c +0
-0fs/ceph/inode.c static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) · line 1145
Quoted 7 lines
fs/ceph/inode.c +0
-1@@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); -dput (dn); dn = realdn;/* note realdn contains the error */ goto out; }else if (realdn) {This might be ok, buI have some real concerns about
splice_dentryand its callers -- particularlyceph_fill_trace:We hold a reference to dn on entry to
splice_dentry. We then calld_splice_aliasand get back an error, and now we don't put the old dentry.Fine -- we have to then expect the caller to do it. Unfortunately, the callers in
ceph_fill_tracedo this:dn = splice_dentry(dn, in); if (IS_ERR(dn)) { err = PTR_ERR(dn); goto done; } req->r_dentry = dn; /* may have spliced */The old value of dn gets clobbered once that comes back with an
ERR_PTR. I guess we could claim thatr_dentrywill still be set to the old value at that point and that it would get cleaned up when it gets cleaned up.But...I see this higher up in
ceph_fill_traceat the end of theCEPH_MFS_OP_RENAMEcondition block:dn = req->r_old_dentry; /* use old_dentry */Quoted 9 lines
So now I'm worried about the case where the splice succeeds. ISTM that "dn" can represent either
r_dentryorr_old_dentryat the point wheresplice_dentrygets called, but we only ever reset the value ofr_dentrythere.If dn ==
r_old_dentryat the time thatsplice_dentryis called, and then that succeeds, we'll end up leaking the reference tor_dentryand then doing an overput onr_old_dentry.Good catch
Quoted 4 lines
I think it might help to establish clear "ownership" of the dentry references throughout that function. Consider zeroing out
r_dentryandr_old_dentryat the time that you set the local variables? That might make this whole thing less fragile.Before the function return, we need to set
req->r_dentryagain. This will introduce duplicated code because there are several ‘goto’ in the function. How about following change.
Review
Quoted 5 lines
(quoted diff) +4
-1- dn = req->r_old_dentry; /* use old_dentry */ + /* swap r_dentry and r_old_dentry */ + req->r_dentry = req->r_old_dentry; + req->r_old_dentry = dn; + dn = req->r_dentry;
Why do we need to reset r_dentry? I don't see where it gets used after ceph_fill_trace returns, so I gather we're just resetting it to clean up the references? Your proposed fix looks like it would fix that bug. It will mean that the values in the req will change but maybe that's ok.
Another idea might be to just keep an extra reference to "dn", and not reset r_dentry and r_old_dentry? It's a bit more code, but it might be a cleaner fix, long-term.
Signature · 1 line
Jeff Layton <[email protected]>
Quoted 86 lines · Jeff Layton
Quoted 75 lines · Zheng
Quoted 54 lines · Jeff Layton
Quoted 17 lines · Zheng
This reverts commit
8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3.
splice_dentry()is used by three places. For two places,req->r_dentryis passed tosplice_dentry(). In the case of error,req->r_dentrydoes not get updated. Sosplice_dentry()should not drop reference.Cc [email protected] #4.18Signed-off-by "Yan, Zheng" <[email protected]>---
fs/ceph/inode.c| 1 -1 file changed, 1 deletion(-)fs/ceph/inode.c +0
-0fs/ceph/inode.c static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) · line 1145
Quoted 7 lines
fs/ceph/inode.c +0
-1@@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); -dput (dn); dn = realdn;/* note realdn contains the error */ goto out; }else if (realdn) {This might be ok, buI have some real concerns about
splice_dentryand its callers -- particularlyceph_fill_trace:We hold a reference to dn on entry to
splice_dentry. We then calld_splice_aliasand get back an error, and now we don't put the old dentry.Fine -- we have to then expect the caller to do it. Unfortunately, the callers in
ceph_fill_tracedo this:dn = splice_dentry(dn, in); if (IS_ERR(dn)) { err = PTR_ERR(dn); goto done; } req->r_dentry = dn; /* may have spliced */The old value of dn gets clobbered once that comes back with an
ERR_PTR. I guess we could claim thatr_dentrywill still be set to the old value at that point and that it would get cleaned up when it gets cleaned up.But...I see this higher up in
ceph_fill_traceat the end of theCEPH_MFS_OP_RENAMEcondition block:dn = req->r_old_dentry; /* use old_dentry */Quoted 9 lines
So now I'm worried about the case where the splice succeeds. ISTM that "dn" can represent either
r_dentryorr_old_dentryat the point wheresplice_dentrygets called, but we only ever reset the value ofr_dentrythere.If dn ==
r_old_dentryat the time thatsplice_dentryis called, and then that succeeds, we'll end up leaking the reference tor_dentryand then doing an overput onr_old_dentry.Good catch
Quoted 4 lines
I think it might help to establish clear "ownership" of the dentry references throughout that function. Consider zeroing out
r_dentryandr_old_dentryat the time that you set the local variables? That might make this whole thing less fragile.Before the function return, we need to set
req->r_dentryagain. This will introduce duplicated code because there are several ‘goto’ in the function. How about following change.Review
Quoted 5 lines
(quoted diff) +4
-1- dn = req->r_old_dentry; /* use old_dentry */ + /* swap r_dentry and r_old_dentry */ + req->r_dentry = req->r_old_dentry; + req->r_old_dentry = dn; + dn = req->r_dentry;
Quoted 5 lines
Why do we need to reset
r_dentry? I don't see where it gets used afterceph_fill_tracereturns, so I gather we're just resetting it to clean up the references? Your proposed fix looks like it would fix that bug. It will mean that the values in the req will change but maybe that's ok.
It’s used by ceph_finish_lookup
Quoted 3 lines
Another idea might be to just keep an extra reference to "dn", and not reset
r_dentryandr_old_dentry? It's a bit more code, but it might be a cleaner fix, long-term.
I will check, Thanks
Yan, Zheng
Quoted 2 lines
Signature · 1 line
Jeff Layton <[email protected]>
Quoted 84 lines · Zheng
Quoted 84 lines · Jeff Layton
Quoted 80 lines · Zheng
Quoted 54 lines · Jeff Layton
Quoted 17 lines · Zheng
This reverts commit
8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3.
splice_dentry()is used by three places. For two places,req->r_dentryis passed tosplice_dentry(). In the case of error,req->r_dentrydoes not get updated. Sosplice_dentry()should not drop reference.Cc [email protected] #4.18Signed-off-by "Yan, Zheng" <[email protected]>---
fs/ceph/inode.c| 1 -1 file changed, 1 deletion(-)fs/ceph/inode.c +0
-0fs/ceph/inode.c static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) · line 1145
Quoted 7 lines
fs/ceph/inode.c +0
-1@@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); -dput (dn); dn = realdn;/* note realdn contains the error */ goto out; }else if (realdn) {This might be ok, buI have some real concerns about
splice_dentryand its callers -- particularlyceph_fill_trace:We hold a reference to dn on entry to
splice_dentry. We then calld_splice_aliasand get back an error, and now we don't put the old dentry.Fine -- we have to then expect the caller to do it. Unfortunately, the callers in
ceph_fill_tracedo this:dn = splice_dentry(dn, in); if (IS_ERR(dn)) { err = PTR_ERR(dn); goto done; } req->r_dentry = dn; /* may have spliced */The old value of dn gets clobbered once that comes back with an
ERR_PTR. I guess we could claim thatr_dentrywill still be set to the old value at that point and that it would get cleaned up when it gets cleaned up.But...I see this higher up in
ceph_fill_traceat the end of theCEPH_MFS_OP_RENAMEcondition block:dn = req->r_old_dentry; /* use old_dentry */Quoted 9 lines
So now I'm worried about the case where the splice succeeds. ISTM that "dn" can represent either
r_dentryorr_old_dentryat the point wheresplice_dentrygets called, but we only ever reset the value ofr_dentrythere.If dn ==
r_old_dentryat the time thatsplice_dentryis called, and then that succeeds, we'll end up leaking the reference tor_dentryand then doing an overput onr_old_dentry.Good catch
Quoted 4 lines
I think it might help to establish clear "ownership" of the dentry references throughout that function. Consider zeroing out
r_dentryandr_old_dentryat the time that you set the local variables? That might make this whole thing less fragile.Before the function return, we need to set
req->r_dentryagain. This will introduce duplicated code because there are several ‘goto’ in the function. How about following change.(quoted diff) +4
-1- dn = req->r_old_dentry; /* use old_dentry */ + /* swap r_dentry and r_old_dentry */ + req->r_dentry = req->r_old_dentry; + req->r_old_dentry = dn; + dn = req->r_dentry;
Quoted 8 lines
Quoted 5 lines
Why do we need to reset
r_dentry? I don't see where it gets used afterceph_fill_tracereturns, so I gather we're just resetting it to clean up the references? Your proposed fix looks like it would fix that bug. It will mean that the values in the req will change but maybe that's ok.It’s used by
ceph_finish_lookup
In that case, it might be bad idea to overwrite the value of r_dentry with the r_old_dentry pointer?
Quoted 13 lines
Quoted 4 lines
Another idea might be to just keep an extra reference to "dn", and not reset
r_dentryandr_old_dentry? It's a bit more code, but it might be a cleaner fix, long-term.I will check, Thanks
Yan, Zheng
Quoted 2 lines
Signature · 1 line
Jeff Layton <[email protected]>
Signature · 1 line
Jeff Layton <[email protected]>
Quoted 95 lines · Jeff Layton
Quoted 90 lines · Zheng
Quoted 89 lines · Jeff Layton
Quoted 75 lines · Zheng
Quoted 54 lines · Jeff Layton
Quoted 17 lines · Zheng
This reverts commit
8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3.
splice_dentry()is used by three places. For two places,req->r_dentryis passed tosplice_dentry(). In the case of error,req->r_dentrydoes not get updated. Sosplice_dentry()should not drop reference.Cc [email protected] #4.18Signed-off-by "Yan, Zheng" <[email protected]>---
fs/ceph/inode.c| 1 -1 file changed, 1 deletion(-)fs/ceph/inode.c +0
-0fs/ceph/inode.c static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) · line 1145
Quoted 7 lines
fs/ceph/inode.c +0
-1@@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); -dput (dn); dn = realdn;/* note realdn contains the error */ goto out; }else if (realdn) {This might be ok, buI have some real concerns about
splice_dentryand its callers -- particularlyceph_fill_trace:We hold a reference to dn on entry to
splice_dentry. We then calld_splice_aliasand get back an error, and now we don't put the old dentry.Fine -- we have to then expect the caller to do it. Unfortunately, the callers in
ceph_fill_tracedo this:dn = splice_dentry(dn, in); if (IS_ERR(dn)) { err = PTR_ERR(dn); goto done; } req->r_dentry = dn; /* may have spliced */The old value of dn gets clobbered once that comes back with an
ERR_PTR. I guess we could claim thatr_dentrywill still be set to the old value at that point and that it would get cleaned up when it gets cleaned up.But...I see this higher up in
ceph_fill_traceat the end of theCEPH_MFS_OP_RENAMEcondition block:dn = req->r_old_dentry; /* use old_dentry */Quoted 9 lines
So now I'm worried about the case where the splice succeeds. ISTM that "dn" can represent either
r_dentryorr_old_dentryat the point wheresplice_dentrygets called, but we only ever reset the value ofr_dentrythere.If dn ==
r_old_dentryat the time thatsplice_dentryis called, and then that succeeds, we'll end up leaking the reference tor_dentryand then doing an overput onr_old_dentry.Good catch
Quoted 4 lines
I think it might help to establish clear "ownership" of the dentry references throughout that function. Consider zeroing out
r_dentryandr_old_dentryat the time that you set the local variables? That might make this whole thing less fragile.Before the function return, we need to set
req->r_dentryagain. This will introduce duplicated code because there are several ‘goto’ in the function. How about following change.Review
Quoted 5 lines
(quoted diff) +4
-1- dn = req->r_old_dentry; /* use old_dentry */ + /* swap r_dentry and r_old_dentry */ + req->r_dentry = req->r_old_dentry; + req->r_old_dentry = dn; + dn = req->r_dentry;Why do we need to reset
r_dentry? I don't see where it gets used afterceph_fill_tracereturns, so I gather we're just resetting it to clean up the references? Your proposed fix looks like it would fix that bug. It will mean that the values in the req will change but maybe that's ok.
Quoted 5 lines
Quoted 2 lines
It’s used by
ceph_finish_lookupIn that case, it might be bad idea to overwrite the value of
r_dentrywith ther_old_dentrypointer?
ceph_finish_lookup() is not used in the rename case ;)
Rgeards
Yan, Zheng
Quoted 18 lines
Quoted 14 lines
Quoted 5 lines
Another idea might be to just keep an extra reference to "dn", and not reset
r_dentryandr_old_dentry? It's a bit more code, but it might be a cleaner fix, long-term.I will check, Thanks
Yan, Zheng
Quoted 2 lines
Signature · 1 line
Jeff Layton <[email protected]>Signature · 1 line
Jeff Layton <[email protected]>
Quoted 93 lines · Zheng
Quoted 93 lines · Jeff Layton
Quoted 90 lines · Zheng
Quoted 89 lines · Jeff Layton
Quoted 75 lines · Zheng
Quoted 54 lines · Jeff Layton
Quoted 17 lines · Zheng
This reverts commit
8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3.
splice_dentry()is used by three places. For two places,req->r_dentryis passed tosplice_dentry(). In the case of error,req->r_dentrydoes not get updated. Sosplice_dentry()should not drop reference.Cc [email protected] #4.18Signed-off-by "Yan, Zheng" <[email protected]>---
fs/ceph/inode.c| 1 -1 file changed, 1 deletion(-)fs/ceph/inode.c +0
-0fs/ceph/inode.c static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) · line 1145
Quoted 7 lines
fs/ceph/inode.c +0
-1@@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); -dput (dn); dn = realdn;/* note realdn contains the error */ goto out; }else if (realdn) {This might be ok, buI have some real concerns about
splice_dentryand its callers -- particularlyceph_fill_trace:We hold a reference to dn on entry to
splice_dentry. We then calld_splice_aliasand get back an error, and now we don't put the old dentry.Fine -- we have to then expect the caller to do it. Unfortunately, the callers in
ceph_fill_tracedo this:dn = splice_dentry(dn, in); if (IS_ERR(dn)) { err = PTR_ERR(dn); goto done; } req->r_dentry = dn; /* may have spliced */The old value of dn gets clobbered once that comes back with an
ERR_PTR. I guess we could claim thatr_dentrywill still be set to the old value at that point and that it would get cleaned up when it gets cleaned up.But...I see this higher up in
ceph_fill_traceat the end of theCEPH_MFS_OP_RENAMEcondition block:dn = req->r_old_dentry; /* use old_dentry */Quoted 9 lines
So now I'm worried about the case where the splice succeeds. ISTM that "dn" can represent either
r_dentryorr_old_dentryat the point wheresplice_dentrygets called, but we only ever reset the value ofr_dentrythere.If dn ==
r_old_dentryat the time thatsplice_dentryis called, and then that succeeds, we'll end up leaking the reference tor_dentryand then doing an overput onr_old_dentry.Good catch
Quoted 4 lines
I think it might help to establish clear "ownership" of the dentry references throughout that function. Consider zeroing out
r_dentryandr_old_dentryat the time that you set the local variables? That might make this whole thing less fragile.Before the function return, we need to set
req->r_dentryagain. This will introduce duplicated code because there are several ‘goto’ in the function. How about following change.Review
Quoted 5 lines
(quoted diff) +4
-1- dn = req->r_old_dentry; /* use old_dentry */ + /* swap r_dentry and r_old_dentry */ + req->r_dentry = req->r_old_dentry; + req->r_old_dentry = dn; + dn = req->r_dentry;Why do we need to reset
r_dentry? I don't see where it gets used afterceph_fill_tracereturns, so I gather we're just resetting it to clean up the references? Your proposed fix looks like it would fix that bug. It will mean that the values in the req will change but maybe that's ok.
Quoted 9 lines
Quoted 5 lines
Quoted 2 lines
It’s used by
ceph_finish_lookupIn that case, it might be bad idea to overwrite the value of
r_dentrywith ther_old_dentrypointer?
ceph_finish_lookup()is not used in the rename case ;)
Ok, but still...I worry about this sort of special-casing, as it can lead to people making incorrect assumptions later. I think it would be best to just take a dentry ref for "dn" and just put it before it goes out of scope.
At the very least, if you are going to override r_dentry like this, please add a comment explaining why it's safe to do this in the rename case.
Quoted 20 lines
Quoted 18 lines
Quoted 14 lines
Quoted 5 lines
Another idea might be to just keep an extra reference to "dn", and not reset
r_dentryandr_old_dentry? It's a bit more code, but it might be a cleaner fix, long-term.I will check, Thanks
Yan, Zheng
Quoted 2 lines
Signature · 1 line
Jeff Layton <[email protected]>Signature · 1 line
Jeff Layton <[email protected]>
Signature · 1 line
Jeff Layton <[email protected]>
Ilya Dryomov
maintainerQuoted 105 lines · Jeff Layton
Quoted 93 lines · Zheng
Quoted 93 lines · Jeff Layton
Quoted 90 lines · Zheng
Quoted 89 lines · Jeff Layton
Quoted 75 lines · Zheng
Quoted 54 lines · Jeff Layton
Quoted 17 lines · Zheng
This reverts commit
8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3.
splice_dentry()is used by three places. For two places,req->r_dentryis passed tosplice_dentry(). In the case of error,req->r_dentrydoes not get updated. Sosplice_dentry()should not drop reference.Cc [email protected] #4.18Signed-off-by "Yan, Zheng" <[email protected]>---
fs/ceph/inode.c| 1 -1 file changed, 1 deletion(-)fs/ceph/inode.c +0
-0fs/ceph/inode.c static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) · line 1145
Quoted 7 lines
fs/ceph/inode.c +0
-1@@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); -dput (dn); dn = realdn;/* note realdn contains the error */ goto out; }else if (realdn) {This might be ok, buI have some real concerns about
splice_dentryand its callers -- particularlyceph_fill_trace:We hold a reference to dn on entry to
splice_dentry. We then calld_splice_aliasand get back an error, and now we don't put the old dentry.Fine -- we have to then expect the caller to do it. Unfortunately, the callers in
ceph_fill_tracedo this:dn = splice_dentry(dn, in); if (IS_ERR(dn)) { err = PTR_ERR(dn); goto done; } req->r_dentry = dn; /* may have spliced */The old value of dn gets clobbered once that comes back with an
ERR_PTR. I guess we could claim thatr_dentrywill still be set to the old value at that point and that it would get cleaned up when it gets cleaned up.But...I see this higher up in
ceph_fill_traceat the end of theCEPH_MFS_OP_RENAMEcondition block:dn = req->r_old_dentry; /* use old_dentry */Quoted 9 lines
So now I'm worried about the case where the splice succeeds. ISTM that "dn" can represent either
r_dentryorr_old_dentryat the point wheresplice_dentrygets called, but we only ever reset the value ofr_dentrythere.If dn ==
r_old_dentryat the time thatsplice_dentryis called, and then that succeeds, we'll end up leaking the reference tor_dentryand then doing an overput onr_old_dentry.Good catch
Quoted 4 lines
I think it might help to establish clear "ownership" of the dentry references throughout that function. Consider zeroing out
r_dentryandr_old_dentryat the time that you set the local variables? That might make this whole thing less fragile.Before the function return, we need to set
req->r_dentryagain. This will introduce duplicated code because there are several ‘goto’ in the function. How about following change.Review
Quoted 5 lines
(quoted diff) +4
-1- dn = req->r_old_dentry; /* use old_dentry */ + /* swap r_dentry and r_old_dentry */ + req->r_dentry = req->r_old_dentry; + req->r_old_dentry = dn; + dn = req->r_dentry;Why do we need to reset
r_dentry? I don't see where it gets used afterceph_fill_tracereturns, so I gather we're just resetting it to clean up the references? Your proposed fix looks like it would fix that bug. It will mean that the values in the req will change but maybe that's ok.Quoted 9 lines
Quoted 5 lines
Quoted 2 lines
It’s used by
ceph_finish_lookupIn that case, it might be bad idea to overwrite the value of
r_dentrywith ther_old_dentrypointer?
ceph_finish_lookup()is not used in the rename case ;)
Quoted 7 lines
Ok, but still...I worry about this sort of special-casing, as it can lead to people making incorrect assumptions later. I think it would be best to just take a dentry ref for "dn" and just put it before it goes out of scope.
At the very least, if you are going to override
r_dentrylike this, please add a comment explaining why it's safe to do this in the rename case.
Hi Zheng,
What is the status here? This patch has missed 4.19, do you have a new version pending somewhere?
Thanks,
Ilya
Quoted 115 lines · Ilya Dryomov
Quoted 105 lines · Jeff Layton
Quoted 93 lines · Zheng
Quoted 93 lines · Jeff Layton
Quoted 90 lines · Zheng
Quoted 89 lines · Jeff Layton
Quoted 75 lines · Zheng
Quoted 54 lines · Jeff Layton
Quoted 17 lines · Zheng
This reverts commit
8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3.
splice_dentry()is used by three places. For two places,req->r_dentryis passed tosplice_dentry(). In the case of error,req->r_dentrydoes not get updated. Sosplice_dentry()should not drop reference.Cc [email protected] #4.18Signed-off-by "Yan, Zheng" <[email protected]>---
fs/ceph/inode.c| 1 -1 file changed, 1 deletion(-)fs/ceph/inode.c +0
-0fs/ceph/inode.c static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) · line 1145
Quoted 7 lines
fs/ceph/inode.c +0
-1@@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); -dput (dn); dn = realdn;/* note realdn contains the error */ goto out; }else if (realdn) {This might be ok, buI have some real concerns about
splice_dentryand its callers -- particularlyceph_fill_trace:We hold a reference to dn on entry to
splice_dentry. We then calld_splice_aliasand get back an error, and now we don't put the old dentry.Fine -- we have to then expect the caller to do it. Unfortunately, the callers in
ceph_fill_tracedo this:dn = splice_dentry(dn, in); if (IS_ERR(dn)) { err = PTR_ERR(dn); goto done; } req->r_dentry = dn; /* may have spliced */The old value of dn gets clobbered once that comes back with an
ERR_PTR. I guess we could claim thatr_dentrywill still be set to the old value at that point and that it would get cleaned up when it gets cleaned up.But...I see this higher up in
ceph_fill_traceat the end of theCEPH_MFS_OP_RENAMEcondition block:dn = req->r_old_dentry; /* use old_dentry */Quoted 9 lines
So now I'm worried about the case where the splice succeeds. ISTM that "dn" can represent either
r_dentryorr_old_dentryat the point wheresplice_dentrygets called, but we only ever reset the value ofr_dentrythere.If dn ==
r_old_dentryat the time thatsplice_dentryis called, and then that succeeds, we'll end up leaking the reference tor_dentryand then doing an overput onr_old_dentry.Good catch
Quoted 4 lines
I think it might help to establish clear "ownership" of the dentry references throughout that function. Consider zeroing out
r_dentryandr_old_dentryat the time that you set the local variables? That might make this whole thing less fragile.Before the function return, we need to set
req->r_dentryagain. This will introduce duplicated code because there are several ‘goto’ in the function. How about following change.Review
Quoted 5 lines
(quoted diff) +4
-1- dn = req->r_old_dentry; /* use old_dentry */ + /* swap r_dentry and r_old_dentry */ + req->r_dentry = req->r_old_dentry; + req->r_old_dentry = dn; + dn = req->r_dentry;Why do we need to reset
r_dentry? I don't see where it gets used afterceph_fill_tracereturns, so I gather we're just resetting it to clean up the references? Your proposed fix looks like it would fix that bug. It will mean that the values in the req will change but maybe that's ok.Quoted 9 lines
Quoted 5 lines
Quoted 2 lines
It’s used by
ceph_finish_lookupIn that case, it might be bad idea to overwrite the value of
r_dentrywith ther_old_dentrypointer?
ceph_finish_lookup()is not used in the rename case ;)Quoted 7 lines
Ok, but still...I worry about this sort of special-casing, as it can lead to people making incorrect assumptions later. I think it would be best to just take a dentry ref for "dn" and just put it before it goes out of scope.
At the very least, if you are going to override
r_dentrylike this, please add a comment explaining why it's safe to do this in the rename case.
Quoted 6 lines
Hi Zheng,
What is the status here? This patch has missed 4.19, do you have a new version pending somewhere?
Thanks,
I'd like to get this patch and patch "ceph: fix dentry leak in ceph_readdir_prepopulate" upstreamed. (use other patch to do cleanup)
Regards
Yan, Zheng
Quoted 2 lines
Ilya
A message that is not in the archive (sent off-list or not yet received): CAAM7YAnKvoEA+QGYZuvcK6W93ZBqLbFLnX3YN8zV2Pgs4S0ZLA@mail.gmail.com
Quoted 165 lines · Zheng
Quoted 101 lines · Jeff Layton
Quoted 99 lines · Zheng
Quoted 98 lines · Jeff Layton
Quoted 84 lines · Zheng
Quoted 84 lines · Jeff Layton
Quoted 80 lines · Zheng
Quoted 54 lines · Jeff Layton
Quoted 17 lines · Zheng
This reverts commit
8b8f53af1ed9df88a4c0fbfdf3db58f62060edf3.
splice_dentry()is used by three places. For two places,req->r_dentryis passed tosplice_dentry(). In the case of error,req->r_dentrydoes not get updated. Sosplice_dentry()should not drop reference.Cc [email protected] #4.18Signed-off-by "Yan, Zheng" <[email protected]>---
fs/ceph/inode.c| 1 -1 file changed, 1 deletion(-)fs/ceph/inode.c +0
-0fs/ceph/inode.c static struct dentry *splice_dentry(struct dentry *dn, struct inode *in) · line 1145
Quoted 7 lines
fs/ceph/inode.c +0
-1@@ -1140,7 +1140,6 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); -dput (dn); dn = realdn;/* note realdn contains the error */ goto out; }else if (realdn) {This might be ok, buI have some real concerns about
splice_dentryand its callers -- particularlyceph_fill_trace:We hold a reference to dn on entry to
splice_dentry. We then calld_splice_aliasand get back an error, and now we don't put the old dentry.Fine -- we have to then expect the caller to do it. Unfortunately, the callers in
ceph_fill_tracedo this:dn = splice_dentry(dn, in); if (IS_ERR(dn)) { err = PTR_ERR(dn); goto done; } req->r_dentry = dn; /* may have spliced */The old value of dn gets clobbered once that comes back with an
ERR_PTR. I guess we could claim thatr_dentrywill still be set to the old value at that point and that it would get cleaned up when it gets cleaned up.But...I see this higher up in
ceph_fill_traceat the end of theCEPH_MFS_OP_RENAMEcondition block:dn = req->r_old_dentry; /* use old_dentry */Quoted 9 lines
So now I'm worried about the case where the splice succeeds. ISTM that "dn" can represent either
r_dentryorr_old_dentryat the point wheresplice_dentrygets called, but we only ever reset the value ofr_dentrythere.If dn ==
r_old_dentryat the time thatsplice_dentryis called, and then that succeeds, we'll end up leaking the reference tor_dentryand then doing an overput onr_old_dentry.Good catch
Quoted 4 lines
I think it might help to establish clear "ownership" of the dentry references throughout that function. Consider zeroing out
r_dentryandr_old_dentryat the time that you set the local variables? That might make this whole thing less fragile.Before the function return, we need to set
req->r_dentryagain. This will introduce duplicated code because there are several ‘goto’ in the function. How about following change.(quoted diff) +4
-1- dn = req->r_old_dentry; /* use old_dentry */ + /* swap r_dentry and r_old_dentry */ + req->r_dentry = req->r_old_dentry; + req->r_old_dentry = dn; + dn = req->r_dentry;Quoted 8 lines
Quoted 5 lines
Why do we need to reset
r_dentry? I don't see where it gets used afterceph_fill_tracereturns, so I gather we're just resetting it to clean up the references? Your proposed fix looks like it would fix that bug. It will mean that the values in the req will change but maybe that's ok.It’s used by
ceph_finish_lookupIn that case, it might be bad idea to overwrite the value of
r_dentrywith ther_old_dentrypointer?Quoted 8 lines
Quoted 3 lines
ceph_finish_lookup()is not used in the rename case ;)Ok, but still...I worry about this sort of special-casing, as it can lead to people making incorrect assumptions later. I think it would be best to just take a dentry ref for "dn" and just put it before it goes out of scope.
how about below patch. let
splice_dentry()to updatereq->r_dentry.fs/ceph/inode.c +8
-12@@ -1106,8 +1106,9 @@ static void update_dentry_lease(struct dentry *dentry,* splice a dentry to an inode. * caller must hold directory i_mutex for this to be safe. */ -static struct dentry *splice_dentry (struct dentry *dn,struct inode *in) +static int splice_dentry (struct dentry **pdn,struct inode *in) { +struct dentry *dn = *pdn;struct dentry *realdn;BUG_ON (d_inode (dn));@@ -1140,28 +1141,23 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)if (IS_ERR (realdn)) {pr_err ("splice_dentry error %ld %p inode %p ino %llx.%llx\n" ,PTR_ERR (realdn), dn, in,ceph_vinop (in)); - dn = realdn; -/* -* Caller should release 'dn' in the case of error. -* If 'req->r_dentry' is passed to this function, -* caller should leave 'req->r_dentry' untouched. -*/ - goto out; - }else if (realdn) { +return PTR_ERR (realdn); + } + +if (realdn) {dout ("dn %p (%d) spliced with %p (%d) " "inode %p ino %llx.%llx\n" , dn,d_count (dn), realdn,d_count (realdn),d_inode (realdn),ceph_vinop (d_inode (realdn)));dput (dn); - dn = realdn; + *pdn = realdn; }else {BUG_ON (!ceph_dentry (dn));dout ("dn %p attached to %p ino %llx.%llx\n" , dn,d_inode (dn),ceph_vinop (d_inode (dn))); } -out: -return dn; +return 0 ; }/*
fs/ceph/inode.c int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req) · line 1350
Quoted 8 lines
fs/ceph/inode.c +4
-1@@ -1348,7 +1344,10 @@ int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req)dout ("dn %p gets new offset %lld\n" , req->r_old_dentry,ceph_dentry (req->r_old_dentry)->offset); - dn = req->r_old_dentry;/* use old_dentry */ +/* swap r_dentry and r_old_dentry */ + req->r_dentry = req->r_old_dentry; + req->r_old_dentry = dn; + dn = req->r_dentry;
I think the above deserves a comment explaining why this is safe to do for rename ops.
Quoted 73 lines
fs/ceph/inode.c +9
-21}/* null dentry? */ @@ -1373,12 +1372,10 @@ int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req)if (d_really_is_negative (dn)) {ceph_dir_clear_ordered (dir);ihold (in); - dn =splice_dentry (dn, in); -if (IS_ERR (dn)) { - err =PTR_ERR (dn); + err =splice_dentry (&req->r_dentry, in); +if (err <0 ) goto done; - } - req->r_dentry = dn;/* may have spliced */ + dn = req->r_dentry;/* may have spliced */ }else if (d_really_is_positive (dn) &&d_inode (dn) != in) { dout(" %p links to %p %llx.%llx, not %llx.%llx\n", dn, d_inode(dn), ceph_vinop(d_inode(dn)),@@ -1398,22 +1395,18 @@ int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req) }else if ((req->r_op ==CEPH_MDS_OP_LOOKUPSNAP || req->r_op ==CEPH_MDS_OP_MKSNAP ) && !test_bit (CEPH_MDS_R_ABORTED , &req->r_req_flags)) { -struct dentry *dn = req->r_dentry;struct inode *dir = req->r_parent;/* fill out a snapdir LOOKUPSNAP dentry */ -BUG_ON (!dn);BUG_ON (!dir);BUG_ON (ceph_snap (dir) !=CEPH_SNAPDIR ); -dout (" linking snapped dir %p to dn %p\n" , in, dn); +BUG_ON (!req->r_dentry); +dout (" linking snapped dir %p to dn %p\n" , in, req->r_dentry);ceph_dir_clear_ordered (dir);ihold (in); - dn =splice_dentry (dn, in); -if (IS_ERR (dn)) { - err =PTR_ERR (dn); + err =splice_dentry (&req->r_dentry, in); +if (err <0 ) goto done; - } - req->r_dentry = dn;/* may have spliced */ }else if (rinfo->head->is_dentry) {struct ceph_vino *ptvino =NULL ;@@ -1677,8 +1670,6 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req, }if (d_really_is_negative (dn)) { -struct dentry *realdn; -if (ceph_security_xattr_deadlock (in)) {dout (" skip splicing dn %p to inode %p" " (security xattr deadlock)\n" , dn, in);@@ -1687,13 +1678,9 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req, goto next_item; } - realdn =splice_dentry (dn, in); -if (IS_ERR (realdn)) { - err =PTR_ERR (realdn); -d_drop (dn); + err =splice_dentry (&dn, in); +if (err <0 ) goto next_item; - } - dn = realdn; }ceph_dentry (dn)->offset = rde->offset;
fs/ceph/inode.c int ceph_readdir_prepopulate(struct ceph_mds_request *req, · line 1702
Quoted 9 lines
fs/ceph/inode.c +1
-2@@ -1709,8 +1696,7 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req, err = ret; } next_item: -if (dn) -dput (dn); +dput (dn); } out:if (err ==0 && skipped ==0 ) {
The rest seems to be ok.
Quoted 29 lines
Quoted 29 lines
At the very least, if you are going to override
r_dentrylike this, please add a comment explaining why it's safe to do this in the rename case.Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 5 lines
Another idea might be to just keep an extra reference to "dn", and not reset
r_dentryandr_old_dentry? It's a bit more code, but it might be a cleaner fix, long-term.I will check, Thanks
Yan, Zheng
Quoted 10 lines
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
--
Jeff Layton <[email protected]>--
Jeff Layton <[email protected]>--
Jeff Layton <[email protected]>
Signature · 1 line
Jeff Layton <[email protected]>