On Fri, May 18, 2018 at 06:35:13PM -0700, Dan Williams wrote: > Background: > > get_user_pages() in the filesystem pins file backed memory pages for > access by devices performing dma. However, it only pins the memory pages > not the page-to-file offset association. If a file is truncated the > pages are mapped out of the file and dma may continue indefinitely into > a page that is owned by a device driver. This breaks coherency of the > file vs dma, but the assumption is that if userspace wants the > file-space truncated it does not matter what data is inbound from the > device, it is not relevant anymore. The only expectation is that dma can > safely continue while the filesystem reallocates the block(s). > > Problem: > > This expectation that dma can safely continue while the filesystem > changes the block map is broken by dax. With dax the target dma page > *is* the filesystem block. The model of leaving the page pinned for dma, > but truncating the file block out of the file, means that the filesytem > is free to reallocate a block under active dma to another file and now > the expected data-incoherency situation has turned into active > data-corruption. > > Solution: > > Defer all filesystem operations (fallocate(), truncate()) on a dax mode > file while any page/block in the file is under active dma. This solution > assumes that dma is transient. Cases where dma operations are known to > not be transient, like RDMA, have been explicitly disabled via > commits like 5f1d43de5416 "IB/core: disable memory registration of > filesystem-dax vmas". > > The dax_layout_busy_page() routine is called by filesystems with a lock > held against mm faults (i_mmap_lock) to find pinned / busy dax pages. > The process of looking up a busy page invalidates all mappings > to trigger any subsequent get_user_pages() to block on i_mmap_lock. > The filesystem continues to call dax_layout_busy_page() until it finally > returns no more active pages. This approach assumes that the page > pinning is transient, if that assumption is violated the system would > have likely hung from the uncompleted I/O. > > Cc: Jeff Moyer <jmoyer@xxxxxxxxxx> > Cc: Dave Chinner <david@xxxxxxxxxxxxx> > Cc: Matthew Wilcox <mawilcox@xxxxxxxxxxxxx> > Cc: Alexander Viro <viro@xxxxxxxxxxxxxxxxxx> > Cc: "Darrick J. Wong" <darrick.wong@xxxxxxxxxx> > Cc: Ross Zwisler <ross.zwisler@xxxxxxxxxxxxxxx> > Cc: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx> > Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> > Reported-by: Christoph Hellwig <hch@xxxxxx> > Reviewed-by: Christoph Hellwig <hch@xxxxxx> > Reviewed-by: Jan Kara <jack@xxxxxxx> > Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx> > --- <> > @@ -492,6 +505,90 @@ static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index, > return entry; > } > > +/** > + * dax_layout_busy_page - find first pinned page in @mapping > + * @mapping: address space to scan for a page with ref count > 1 > + * > + * DAX requires ZONE_DEVICE mapped pages. These pages are never > + * 'onlined' to the page allocator so they are considered idle when > + * page->count == 1. A filesystem uses this interface to determine if > + * any page in the mapping is busy, i.e. for DMA, or other > + * get_user_pages() usages. > + * > + * It is expected that the filesystem is holding locks to block the > + * establishment of new mappings in this address_space. I.e. it expects > + * to be able to run unmap_mapping_range() and subsequently not race > + * mapping_mapped() becoming true. > + */ > +struct page *dax_layout_busy_page(struct address_space *mapping) > +{ > + pgoff_t indices[PAGEVEC_SIZE]; > + struct page *page = NULL; > + struct pagevec pvec; > + pgoff_t index, end; > + unsigned i; > + > + /* > + * In the 'limited' case get_user_pages() for dax is disabled. > + */ > + if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) > + return NULL; > + > + if (!dax_mapping(mapping) || !mapping_mapped(mapping)) > + return NULL; > + > + pagevec_init(&pvec); > + index = 0; > + end = -1; > + > + /* > + * If we race get_user_pages_fast() here either we'll see the > + * elevated page count in the pagevec_lookup and wait, or > + * get_user_pages_fast() will see that the page it took a reference > + * against is no longer mapped in the page tables and bail to the > + * get_user_pages() slow path. The slow path is protected by > + * pte_lock() and pmd_lock(). New references are not taken without > + * holding those locks, and unmap_mapping_range() will not zero the > + * pte or pmd without holding the respective lock, so we are > + * guaranteed to either see new references or prevent new > + * references from being established. > + */ > + unmap_mapping_range(mapping, 0, 0, 1); > + > + while (index < end && pagevec_lookup_entries(&pvec, mapping, index, > + min(end - index, (pgoff_t)PAGEVEC_SIZE), > + indices)) { > + for (i = 0; i < pagevec_count(&pvec); i++) { > + struct page *pvec_ent = pvec.pages[i]; > + void *entry; > + > + index = indices[i]; > + if (index >= end) > + break; > + > + if (!radix_tree_exceptional_entry(pvec_ent)) > + continue; > + > + xa_lock_irq(&mapping->i_pages); > + entry = get_unlocked_mapping_entry(mapping, index, NULL); > + if (entry) > + page = dax_busy_page(entry); > + put_unlocked_mapping_entry(mapping, index, entry); > + xa_unlock_irq(&mapping->i_pages); > + if (page) > + break; > + } > + pagevec_remove_exceptionals(&pvec); > + pagevec_release(&pvec); I must be missing something - now that we're using the common 4k zero page, we should only ever have exceptional entries in the DAX radix tree, right? If so, it seems like these two pagevec_* calls could/should go away, and the !radix_tree_exceptional_entry() check in the for loop above should be surrounded by a WARN_ON_ONCE()? Or has something changed that I'm overlooking?