Message ID | 20180911154246.6844-3-toshi.kani@hpe.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | fix sync to flush processor cache for ext4 DAX files | expand |
On Tue, Sep 11, 2018 at 8:42 AM, Toshi Kani <toshi.kani@hpe.com> wrote: > Sync syscall to an existing DAX file needs to flush processor cache, > but it does not currently. This is because 'ext4_da_aops' is set to > address_space_operations of existing DAX files, instead of 'ext4_dax_aops', > since S_DAX flag is set after ext4_set_aops() in the open path. > > New file > -------- > lookup_open > ext4_create > __ext4_new_inode > ext4_set_inode_flags // Set S_DAX flag > ext4_set_aops // Set aops to ext4_dax_aops > > Existing file > ------------- > lookup_open > ext4_lookup > ext4_iget > ext4_set_aops // Set aops to ext4_da_aops > ext4_set_inode_flags // Set S_DAX flag > > Change ext4_iget() to call ext4_set_inode_flags() before ext4_set_aops(). > > Fixes: 5f0663bb4a64f588f0a2dd6d1be68d40f9af0086 Same format nit: Fixes: 5f0663bb4a64 ("ext4, dax: introduce ext4_dax_aops") Cc: <stable@vger.kernel.org> > Signed-off-by: Toshi Kani <toshi.kani@hpe.com> > Cc: Jan Kara <jack@suse.cz> > Cc: Dan Williams <dan.j.williams@intel.com> > Cc: "Theodore Ts'o" <tytso@mit.edu> > Cc: Andreas Dilger <adilger.kernel@dilger.ca> > --- > fs/ext4/inode.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c > index 775cd9b4af55..93cbbb859c40 100644 > --- a/fs/ext4/inode.c > +++ b/fs/ext4/inode.c > @@ -4998,6 +4998,8 @@ struct inode *ext4_iget(struct super_block *sb, unsigned long ino) > if (ret) > goto bad_inode; > > + ext4_set_inode_flags(inode); > + Hmm, does this have unintended behavior changes? I notice that there are some checks for flags "IS_APPEND(inode) || IS_IMMUTABLE(inode)" *before* the call to ext4_set_inode_flags(). I didn't look too much deeper at whether those checks are bogus, but it would seem safer to do something like this for a lower risk fix. Thoughts? diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c index d0dd585add6a..1e9ab445c777 100644 --- a/fs/ext4/inode.c +++ b/fs/ext4/inode.c @@ -4999,7 +4999,6 @@ struct inode *ext4_iget(struct super_block *sb, unsigned long ino) if (S_ISREG(inode->i_mode)) { inode->i_op = &ext4_file_inode_operations; inode->i_fop = &ext4_file_operations; - ext4_set_aops(inode); } else if (S_ISDIR(inode->i_mode)) { inode->i_op = &ext4_dir_inode_operations; inode->i_fop = &ext4_dir_operations; @@ -5042,6 +5041,12 @@ struct inode *ext4_iget(struct super_block *sb, unsigned long ino) } brelse(iloc.bh); ext4_set_inode_flags(inode); + /* + * Now that we have determined whether DAX is enabled, set the + * proper address spaces operations + */ + if (S_ISREG(inode->i_mode)) + ext4_set_aops(inode); unlock_new_inode(inode); return inode;
On Tue, 2018-09-11 at 11:15 -0700, Dan Williams wrote: > On Tue, Sep 11, 2018 at 8:42 AM, Toshi Kani <toshi.kani@hpe.com> wrote: > > Sync syscall to an existing DAX file needs to flush processor cache, > > but it does not currently. This is because 'ext4_da_aops' is set to > > address_space_operations of existing DAX files, instead of 'ext4_dax_aops', > > since S_DAX flag is set after ext4_set_aops() in the open path. > > > > New file > > -------- > > lookup_open > > ext4_create > > __ext4_new_inode > > ext4_set_inode_flags // Set S_DAX flag > > ext4_set_aops // Set aops to ext4_dax_aops > > > > Existing file > > ------------- > > lookup_open > > ext4_lookup > > ext4_iget > > ext4_set_aops // Set aops to ext4_da_aops > > ext4_set_inode_flags // Set S_DAX flag > > > > Change ext4_iget() to call ext4_set_inode_flags() before ext4_set_aops(). > > > > Fixes: 5f0663bb4a64f588f0a2dd6d1be68d40f9af0086 > > Same format nit: > > Fixes: 5f0663bb4a64 ("ext4, dax: introduce ext4_dax_aops") > Cc: <stable@vger.kernel.org> Will do. > > Signed-off-by: Toshi Kani <toshi.kani@hpe.com> > > Cc: Jan Kara <jack@suse.cz> > > Cc: Dan Williams <dan.j.williams@intel.com> > > Cc: "Theodore Ts'o" <tytso@mit.edu> > > Cc: Andreas Dilger <adilger.kernel@dilger.ca> > > --- > > fs/ext4/inode.c | 3 ++- > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c > > index 775cd9b4af55..93cbbb859c40 100644 > > --- a/fs/ext4/inode.c > > +++ b/fs/ext4/inode.c > > @@ -4998,6 +4998,8 @@ struct inode *ext4_iget(struct super_block *sb, unsigned long ino) > > if (ret) > > goto bad_inode; > > > > + ext4_set_inode_flags(inode); > > + > > Hmm, does this have unintended behavior changes? > > I notice that there are some checks for flags "IS_APPEND(inode) || > IS_IMMUTABLE(inode)" *before* the call to ext4_set_inode_flags(). I > didn't look too much deeper at whether those checks are bogus, but it > would seem safer to do something like this for a lower risk fix. > > Thoughts? Good catch! Agreed. Thanks! -Toshi
On Tue 11-09-18 11:15:18, Dan Williams wrote: > On Tue, Sep 11, 2018 at 8:42 AM, Toshi Kani <toshi.kani@hpe.com> wrote: > > Sync syscall to an existing DAX file needs to flush processor cache, > > but it does not currently. This is because 'ext4_da_aops' is set to > > address_space_operations of existing DAX files, instead of 'ext4_dax_aops', > > since S_DAX flag is set after ext4_set_aops() in the open path. > > > > New file > > -------- > > lookup_open > > ext4_create > > __ext4_new_inode > > ext4_set_inode_flags // Set S_DAX flag > > ext4_set_aops // Set aops to ext4_dax_aops > > > > Existing file > > ------------- > > lookup_open > > ext4_lookup > > ext4_iget > > ext4_set_aops // Set aops to ext4_da_aops > > ext4_set_inode_flags // Set S_DAX flag > > > > Change ext4_iget() to call ext4_set_inode_flags() before ext4_set_aops(). > > > > Fixes: 5f0663bb4a64f588f0a2dd6d1be68d40f9af0086 > > Same format nit: > > Fixes: 5f0663bb4a64 ("ext4, dax: introduce ext4_dax_aops") > Cc: <stable@vger.kernel.org> > > > > Signed-off-by: Toshi Kani <toshi.kani@hpe.com> > > Cc: Jan Kara <jack@suse.cz> > > Cc: Dan Williams <dan.j.williams@intel.com> > > Cc: "Theodore Ts'o" <tytso@mit.edu> > > Cc: Andreas Dilger <adilger.kernel@dilger.ca> > > --- > > fs/ext4/inode.c | 3 ++- > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c > > index 775cd9b4af55..93cbbb859c40 100644 > > --- a/fs/ext4/inode.c > > +++ b/fs/ext4/inode.c > > @@ -4998,6 +4998,8 @@ struct inode *ext4_iget(struct super_block *sb, unsigned long ino) > > if (ret) > > goto bad_inode; > > > > + ext4_set_inode_flags(inode); > > + > > Hmm, does this have unintended behavior changes? > > I notice that there are some checks for flags "IS_APPEND(inode) || > IS_IMMUTABLE(inode)" *before* the call to ext4_set_inode_flags(). I > didn't look too much deeper at whether those checks are bogus, but it > would seem safer to do something like this for a lower risk fix. > > Thoughts? Well, safer but it would leave the landmine around for others to hit. Toshi, please move the ext4_set_inode_flags() call to be just after the assignment: ei->i_flags = le32_to_cpu(raw_inode->i_flags); in ext4_iget(). That way people won't introduce checks for i_flags that can never hit... And yes, it fixes also other bugs (mostly in sanity checks AFAICS) than the DAX issue. Honza > > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c > index d0dd585add6a..1e9ab445c777 100644 > --- a/fs/ext4/inode.c > +++ b/fs/ext4/inode.c > @@ -4999,7 +4999,6 @@ struct inode *ext4_iget(struct super_block *sb, > unsigned long ino) > if (S_ISREG(inode->i_mode)) { > inode->i_op = &ext4_file_inode_operations; > inode->i_fop = &ext4_file_operations; > - ext4_set_aops(inode); > } else if (S_ISDIR(inode->i_mode)) { > inode->i_op = &ext4_dir_inode_operations; > inode->i_fop = &ext4_dir_operations; > @@ -5042,6 +5041,12 @@ struct inode *ext4_iget(struct super_block *sb, > unsigned long ino) > } > brelse(iloc.bh); > ext4_set_inode_flags(inode); > + /* > + * Now that we have determined whether DAX is enabled, set the > + * proper address spaces operations > + */ > + if (S_ISREG(inode->i_mode)) > + ext4_set_aops(inode); > > unlock_new_inode(inode); > return inode;
On Wed, 2018-09-12 at 11:31 +0200, Jan Kara wrote: > On Tue 11-09-18 11:15:18, Dan Williams wrote: > > On Tue, Sep 11, 2018 at 8:42 AM, Toshi Kani <toshi.kani@hpe.com> wrote: > > > Sync syscall to an existing DAX file needs to flush processor cache, > > > but it does not currently. This is because 'ext4_da_aops' is set to > > > address_space_operations of existing DAX files, instead of 'ext4_dax_aops', > > > since S_DAX flag is set after ext4_set_aops() in the open path. > > > > > > New file > > > -------- > > > lookup_open > > > ext4_create > > > __ext4_new_inode > > > ext4_set_inode_flags // Set S_DAX flag > > > ext4_set_aops // Set aops to ext4_dax_aops > > > > > > Existing file > > > ------------- > > > lookup_open > > > ext4_lookup > > > ext4_iget > > > ext4_set_aops // Set aops to ext4_da_aops > > > ext4_set_inode_flags // Set S_DAX flag > > > > > > Change ext4_iget() to call ext4_set_inode_flags() before ext4_set_aops(). > > > > > > Fixes: 5f0663bb4a64f588f0a2dd6d1be68d40f9af0086 > > > > Same format nit: > > > > Fixes: 5f0663bb4a64 ("ext4, dax: introduce ext4_dax_aops") > > Cc: <stable@vger.kernel.org> > > > > > > > Signed-off-by: Toshi Kani <toshi.kani@hpe.com> > > > Cc: Jan Kara <jack@suse.cz> > > > Cc: Dan Williams <dan.j.williams@intel.com> > > > Cc: "Theodore Ts'o" <tytso@mit.edu> > > > Cc: Andreas Dilger <adilger.kernel@dilger.ca> > > > --- > > > fs/ext4/inode.c | 3 ++- > > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > > > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c > > > index 775cd9b4af55..93cbbb859c40 100644 > > > --- a/fs/ext4/inode.c > > > +++ b/fs/ext4/inode.c > > > @@ -4998,6 +4998,8 @@ struct inode *ext4_iget(struct super_block *sb, unsigned long ino) > > > if (ret) > > > goto bad_inode; > > > > > > + ext4_set_inode_flags(inode); > > > + > > > > Hmm, does this have unintended behavior changes? > > > > I notice that there are some checks for flags "IS_APPEND(inode) || > > IS_IMMUTABLE(inode)" *before* the call to ext4_set_inode_flags(). I > > didn't look too much deeper at whether those checks are bogus, but it > > would seem safer to do something like this for a lower risk fix. > > > > Thoughts? > > Well, safer but it would leave the landmine around for others to hit. > Toshi, please move the ext4_set_inode_flags() call to be just after the > assignment: > > ei->i_flags = le32_to_cpu(raw_inode->i_flags); > > in ext4_iget(). That way people won't introduce checks for i_flags that can > never hit... And yes, it fixes also other bugs (mostly in sanity checks > AFAICS) than the DAX issue. Sure. Assuming you think the implicit change Dan pointed out is not a problem, yes, I will go with this cleaner approach. Thanks! -Toshi
On Wed, Sep 12, 2018 at 9:08 AM, Kani, Toshi <toshi.kani@hpe.com> wrote: > On Wed, 2018-09-12 at 11:31 +0200, Jan Kara wrote: >> On Tue 11-09-18 11:15:18, Dan Williams wrote: >> > On Tue, Sep 11, 2018 at 8:42 AM, Toshi Kani <toshi.kani@hpe.com> wrote: >> > > Sync syscall to an existing DAX file needs to flush processor cache, >> > > but it does not currently. This is because 'ext4_da_aops' is set to >> > > address_space_operations of existing DAX files, instead of 'ext4_dax_aops', >> > > since S_DAX flag is set after ext4_set_aops() in the open path. >> > > >> > > New file >> > > -------- >> > > lookup_open >> > > ext4_create >> > > __ext4_new_inode >> > > ext4_set_inode_flags // Set S_DAX flag >> > > ext4_set_aops // Set aops to ext4_dax_aops >> > > >> > > Existing file >> > > ------------- >> > > lookup_open >> > > ext4_lookup >> > > ext4_iget >> > > ext4_set_aops // Set aops to ext4_da_aops >> > > ext4_set_inode_flags // Set S_DAX flag >> > > >> > > Change ext4_iget() to call ext4_set_inode_flags() before ext4_set_aops(). >> > > >> > > Fixes: 5f0663bb4a64f588f0a2dd6d1be68d40f9af0086 >> > >> > Same format nit: >> > >> > Fixes: 5f0663bb4a64 ("ext4, dax: introduce ext4_dax_aops") >> > Cc: <stable@vger.kernel.org> >> > >> > >> > > Signed-off-by: Toshi Kani <toshi.kani@hpe.com> >> > > Cc: Jan Kara <jack@suse.cz> >> > > Cc: Dan Williams <dan.j.williams@intel.com> >> > > Cc: "Theodore Ts'o" <tytso@mit.edu> >> > > Cc: Andreas Dilger <adilger.kernel@dilger.ca> >> > > --- >> > > fs/ext4/inode.c | 3 ++- >> > > 1 file changed, 2 insertions(+), 1 deletion(-) >> > > >> > > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c >> > > index 775cd9b4af55..93cbbb859c40 100644 >> > > --- a/fs/ext4/inode.c >> > > +++ b/fs/ext4/inode.c >> > > @@ -4998,6 +4998,8 @@ struct inode *ext4_iget(struct super_block *sb, unsigned long ino) >> > > if (ret) >> > > goto bad_inode; >> > > >> > > + ext4_set_inode_flags(inode); >> > > + >> > >> > Hmm, does this have unintended behavior changes? >> > >> > I notice that there are some checks for flags "IS_APPEND(inode) || >> > IS_IMMUTABLE(inode)" *before* the call to ext4_set_inode_flags(). I >> > didn't look too much deeper at whether those checks are bogus, but it >> > would seem safer to do something like this for a lower risk fix. >> > >> > Thoughts? >> >> Well, safer but it would leave the landmine around for others to hit. >> Toshi, please move the ext4_set_inode_flags() call to be just after the >> assignment: >> >> ei->i_flags = le32_to_cpu(raw_inode->i_flags); >> >> in ext4_iget(). That way people won't introduce checks for i_flags that can >> never hit... And yes, it fixes also other bugs (mostly in sanity checks >> AFAICS) than the DAX issue. > > Sure. Assuming you think the implicit change Dan pointed out is not a > problem, yes, I will go with this cleaner approach. > Yes, Jan's proposal looks best to me.
diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c index 775cd9b4af55..93cbbb859c40 100644 --- a/fs/ext4/inode.c +++ b/fs/ext4/inode.c @@ -4998,6 +4998,8 @@ struct inode *ext4_iget(struct super_block *sb, unsigned long ino) if (ret) goto bad_inode; + ext4_set_inode_flags(inode); + if (S_ISREG(inode->i_mode)) { inode->i_op = &ext4_file_inode_operations; inode->i_fop = &ext4_file_operations; @@ -5043,7 +5045,6 @@ struct inode *ext4_iget(struct super_block *sb, unsigned long ino) goto bad_inode; } brelse(iloc.bh); - ext4_set_inode_flags(inode); unlock_new_inode(inode); return inode;
Sync syscall to an existing DAX file needs to flush processor cache, but it does not currently. This is because 'ext4_da_aops' is set to address_space_operations of existing DAX files, instead of 'ext4_dax_aops', since S_DAX flag is set after ext4_set_aops() in the open path. New file -------- lookup_open ext4_create __ext4_new_inode ext4_set_inode_flags // Set S_DAX flag ext4_set_aops // Set aops to ext4_dax_aops Existing file ------------- lookup_open ext4_lookup ext4_iget ext4_set_aops // Set aops to ext4_da_aops ext4_set_inode_flags // Set S_DAX flag Change ext4_iget() to call ext4_set_inode_flags() before ext4_set_aops(). Fixes: 5f0663bb4a64f588f0a2dd6d1be68d40f9af0086 Signed-off-by: Toshi Kani <toshi.kani@hpe.com> Cc: Jan Kara <jack@suse.cz> Cc: Dan Williams <dan.j.williams@intel.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> --- fs/ext4/inode.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)