| Message ID | 1453398364-22537-5-git-send-email-ross.zwisler@linux.intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu 21-01-16 10:46:03, Ross Zwisler wrote: > Fix the way that DAX PMD radix tree entries are handled. With this patch > we now check to see if a PMD entry exists in the radix tree on write, even > if we are just trying to insert a PTE. If it exists, we dirty that instead > of inserting our own PTE entry. > > Fix a bug in the PMD path in dax_writeback_mapping_range() where we were > previously passing a loff_t into radix_tree_lookup instead of a pgoff_t. Ah, good catch! > Account for the fact that multiple fsync/msync operations may be happening > at the same time and don't flush entries that are beyond end_index. > > Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Just one nit below. You can add: Reviewed-by: Jan Kara <jack@suse.cz> > --- > fs/dax.c | 39 +++++++++++++++++++++++++++------------ > 1 file changed, 27 insertions(+), 12 deletions(-) > > diff --git a/fs/dax.c b/fs/dax.c > index 55ae394..afacc30 100644 > --- a/fs/dax.c > +++ b/fs/dax.c ... > @@ -460,31 +468,33 @@ int dax_writeback_mapping_range(struct address_space *mapping, loff_t start, > { > struct inode *inode = mapping->host; > struct block_device *bdev = inode->i_sb->s_bdev; > + pgoff_t start_index, end_index, pmd_index; > pgoff_t indices[PAGEVEC_SIZE]; > - pgoff_t start_page, end_page; > struct pagevec pvec; > - void *entry; > + bool done = false; > int i, ret = 0; > + void *entry; > > if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) > return -EIO; > > + start_index = start >> PAGE_CACHE_SHIFT; > + end_index = end >> PAGE_CACHE_SHIFT; > + pmd_index = DAX_PMD_INDEX(start_index); > + > rcu_read_lock(); > - entry = radix_tree_lookup(&mapping->page_tree, start & PMD_MASK); > + entry = radix_tree_lookup(&mapping->page_tree, pmd_index); > rcu_read_unlock(); > > /* see if the start of our range is covered by a PMD entry */ > - if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) > - start &= PMD_MASK; > - > - start_page = start >> PAGE_CACHE_SHIFT; > - end_page = end >> PAGE_CACHE_SHIFT; > + if (RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) Don't you miss a check that entry != NULL? I agree that RADIX_DAX_TYPE(NULL) is != from RADIX_DAX_PMD so it works as desired but it looks a bit dangerous. Honza
On Fri, Jan 22, 2016 at 04:11:41PM +0100, Jan Kara wrote: > On Thu 21-01-16 10:46:03, Ross Zwisler wrote: > > Fix the way that DAX PMD radix tree entries are handled. With this patch > > we now check to see if a PMD entry exists in the radix tree on write, even > > if we are just trying to insert a PTE. If it exists, we dirty that instead > > of inserting our own PTE entry. > > > > Fix a bug in the PMD path in dax_writeback_mapping_range() where we were > > previously passing a loff_t into radix_tree_lookup instead of a pgoff_t. > > Ah, good catch! > > > Account for the fact that multiple fsync/msync operations may be happening > > at the same time and don't flush entries that are beyond end_index. > > > > Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> > > Just one nit below. You can add: > > Reviewed-by: Jan Kara <jack@suse.cz> > > > --- > > fs/dax.c | 39 +++++++++++++++++++++++++++------------ > > 1 file changed, 27 insertions(+), 12 deletions(-) > > > > diff --git a/fs/dax.c b/fs/dax.c > > index 55ae394..afacc30 100644 > > --- a/fs/dax.c > > +++ b/fs/dax.c > ... > > @@ -460,31 +468,33 @@ int dax_writeback_mapping_range(struct address_space *mapping, loff_t start, > > { > > struct inode *inode = mapping->host; > > struct block_device *bdev = inode->i_sb->s_bdev; > > + pgoff_t start_index, end_index, pmd_index; > > pgoff_t indices[PAGEVEC_SIZE]; > > - pgoff_t start_page, end_page; > > struct pagevec pvec; > > - void *entry; > > + bool done = false; > > int i, ret = 0; > > + void *entry; > > > > if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) > > return -EIO; > > > > + start_index = start >> PAGE_CACHE_SHIFT; > > + end_index = end >> PAGE_CACHE_SHIFT; > > + pmd_index = DAX_PMD_INDEX(start_index); > > + > > rcu_read_lock(); > > - entry = radix_tree_lookup(&mapping->page_tree, start & PMD_MASK); > > + entry = radix_tree_lookup(&mapping->page_tree, pmd_index); > > rcu_read_unlock(); > > > > /* see if the start of our range is covered by a PMD entry */ > > - if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) > > - start &= PMD_MASK; > > - > > - start_page = start >> PAGE_CACHE_SHIFT; > > - end_page = end >> PAGE_CACHE_SHIFT; > > + if (RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) > > Don't you miss a check that entry != NULL? I agree that RADIX_DAX_TYPE(NULL) > is != from RADIX_DAX_PMD so it works as desired but it looks a bit > dangerous. Yea, the initial version had a NULL check first, but I realized I could optimize it out and make things simpler since RADIX_DAX_TYPE() just looked at the value of 'entry' and never treated it like a pointer (because to us, it isn't a pointer). I will add it back if you think that not having it may confuse the reader.
diff --git a/fs/dax.c b/fs/dax.c index 55ae394..afacc30 100644 --- a/fs/dax.c +++ b/fs/dax.c @@ -327,19 +327,27 @@ static int copy_user_bh(struct page *to, struct inode *inode, } #define NO_SECTOR -1 +#define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_CACHE_SHIFT)) static int dax_radix_entry(struct address_space *mapping, pgoff_t index, sector_t sector, bool pmd_entry, bool dirty) { struct radix_tree_root *page_tree = &mapping->page_tree; + pgoff_t pmd_index = DAX_PMD_INDEX(index); int type, error = 0; void *entry; __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); spin_lock_irq(&mapping->tree_lock); - entry = radix_tree_lookup(page_tree, index); + entry = radix_tree_lookup(page_tree, pmd_index); + if (RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) { + index = pmd_index; + goto dirty; + } + + entry = radix_tree_lookup(page_tree, index); if (entry) { type = RADIX_DAX_TYPE(entry); if (WARN_ON_ONCE(type != RADIX_DAX_PTE && @@ -460,31 +468,33 @@ int dax_writeback_mapping_range(struct address_space *mapping, loff_t start, { struct inode *inode = mapping->host; struct block_device *bdev = inode->i_sb->s_bdev; + pgoff_t start_index, end_index, pmd_index; pgoff_t indices[PAGEVEC_SIZE]; - pgoff_t start_page, end_page; struct pagevec pvec; - void *entry; + bool done = false; int i, ret = 0; + void *entry; if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) return -EIO; + start_index = start >> PAGE_CACHE_SHIFT; + end_index = end >> PAGE_CACHE_SHIFT; + pmd_index = DAX_PMD_INDEX(start_index); + rcu_read_lock(); - entry = radix_tree_lookup(&mapping->page_tree, start & PMD_MASK); + entry = radix_tree_lookup(&mapping->page_tree, pmd_index); rcu_read_unlock(); /* see if the start of our range is covered by a PMD entry */ - if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) - start &= PMD_MASK; - - start_page = start >> PAGE_CACHE_SHIFT; - end_page = end >> PAGE_CACHE_SHIFT; + if (RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) + start_index = pmd_index; - tag_pages_for_writeback(mapping, start_page, end_page); + tag_pages_for_writeback(mapping, start_index, end_index); pagevec_init(&pvec, 0); - while (1) { - pvec.nr = find_get_entries_tag(mapping, start_page, + while (!done) { + pvec.nr = find_get_entries_tag(mapping, start_index, PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE, pvec.pages, indices); @@ -492,6 +502,11 @@ int dax_writeback_mapping_range(struct address_space *mapping, loff_t start, break; for (i = 0; i < pvec.nr; i++) { + if (indices[i] > end_index) { + done = true; + break; + } + ret = dax_writeback_one(bdev, mapping, indices[i], pvec.pages[i]); if (ret < 0)
Fix the way that DAX PMD radix tree entries are handled. With this patch we now check to see if a PMD entry exists in the radix tree on write, even if we are just trying to insert a PTE. If it exists, we dirty that instead of inserting our own PTE entry. Fix a bug in the PMD path in dax_writeback_mapping_range() where we were previously passing a loff_t into radix_tree_lookup instead of a pgoff_t. Account for the fact that multiple fsync/msync operations may be happening at the same time and don't flush entries that are beyond end_index. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> --- fs/dax.c | 39 +++++++++++++++++++++++++++------------ 1 file changed, 27 insertions(+), 12 deletions(-)