Btrfs, raid56: support parity scrub on raid56

The implementation is:
- Read and check all the data with checksum in the same stripe.
  All the data which has checksum is COW data, and we are sure
  that it is not changed though we don't lock the stripe. because
  the space of that data just can be reclaimed after the current
  transction is committed, and then the fs can use it to store the
  other data, but when doing scrub, we hold the current transaction,
  that is that data can not be recovered, it is safe that read and check
  it out of the stripe lock.
- Lock the stripe
- Read out all the data without checksum and parity
  The data without checksum and the parity may be changed if we don't
  lock the stripe, so we need read it in the stripe lock context.
- Check the parity
- Re-calculate the new parity and write back it if the old parity
  is not right
- Unlock the stripe

If we can not read out the data or the data we read is corrupted,
we will try to repair it. If the repair fails. we will mark the
horizontal sub-stripe(pages on the same horizontal) as corrupted
sub-stripe, and we will skip the parity check and repair of that
horizontal sub-stripe.

And in order to skip the horizontal sub-stripe that has no data, we
introduce a bitmap. If there is some data on the horizontal sub-stripe,
we will the relative bit to 1, and when we check and repair the
parity, we will skip those horizontal sub-stripes that the relative
bits is 0.
Signed-off-by: NMiao Xie <miaox@cn.fujitsu.com>

fs/btrfs/raid56.c

@@ -72,6 +72,7 @@
 enum btrfs_rbio_ops {
 	BTRFS_RBIO_WRITE	= 0,
 	BTRFS_RBIO_READ_REBUILD	= 1,
+	BTRFS_RBIO_PARITY_SCRUB	= 2,
 };
 
 struct btrfs_raid_bio {
@@ -130,6 +131,7 @@ struct btrfs_raid_bio {
 	/* number of data stripes (no p/q) */
 	int nr_data;
 
+	int stripe_npages;
 	/*
 	 * set if we're doing a parity rebuild
 	 * for a read from higher up, which is handled
@@ -144,6 +146,7 @@ struct btrfs_raid_bio {
 	/* second bad stripe (for raid6 use) */
 	int failb;
 
+	int scrubp;
 	/*
 	 * number of pages needed to represent the full
 	 * stripe
@@ -178,6 +181,11 @@ struct btrfs_raid_bio {
 	 * here for faster lookup
 	 */
 	struct page **bio_pages;
+
+	/*
+	 * bitmap to record which horizontal stripe has data
+	 */
+	unsigned long *dbitmap;
 };
 
 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio);
@@ -192,6 +200,10 @@ static void __free_raid_bio(struct btrfs_raid_bio *rbio);
 static void index_rbio_pages(struct btrfs_raid_bio *rbio);
 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio);
 
+static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
+					 int need_check);
+static void async_scrub_parity(struct btrfs_raid_bio *rbio);
+
 /*
  * the stripe hash table is used for locking, and to collect
  * bios in hopes of making a full stripe
@@ -593,10 +605,20 @@ static int rbio_can_merge(struct btrfs_raid_bio *last,
 	    cur->raid_map[0])
 		return 0;
 
-	/* reads can't merge with writes */
-	if (last->operation != cur->operation) {
+	/* we can't merge with different operations */
+	if (last->operation != cur->operation)
+		return 0;
+	/*
+	 * We've need read the full stripe from the drive.
+	 * check and repair the parity and write the new results.
+	 *
+	 * We're not allowed to add any new bios to the
+	 * bio list here, anyone else that wants to
+	 * change this stripe needs to do their own rmw.
+	 */
+	if (last->operation == BTRFS_RBIO_PARITY_SCRUB ||
+	    cur->operation == BTRFS_RBIO_PARITY_SCRUB)
 		return 0;
-	}
 
 	return 1;
 }
@@ -789,9 +811,12 @@ static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
 
 			if (next->operation == BTRFS_RBIO_READ_REBUILD)
 				async_read_rebuild(next);
-			else if (next->operation == BTRFS_RBIO_WRITE){
+			else if (next->operation == BTRFS_RBIO_WRITE) {
 				steal_rbio(rbio, next);
 				async_rmw_stripe(next);
+			} else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
+				steal_rbio(rbio, next);
+				async_scrub_parity(next);
 			}
 
 			goto done_nolock;
@@ -957,9 +982,11 @@ static struct btrfs_raid_bio *alloc_rbio(struct btrfs_root *root,
 	struct btrfs_raid_bio *rbio;
 	int nr_data = 0;
 	int num_pages = rbio_nr_pages(stripe_len, bbio->num_stripes);
+	int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE);
 	void *p;
 
-	rbio = kzalloc(sizeof(*rbio) + num_pages * sizeof(struct page *) * 2,
+	rbio = kzalloc(sizeof(*rbio) + num_pages * sizeof(struct page *) * 2 +
+		       DIV_ROUND_UP(stripe_npages, BITS_PER_LONG / 8),
 			GFP_NOFS);
 	if (!rbio)
 		return ERR_PTR(-ENOMEM);
@@ -974,6 +1001,7 @@ static struct btrfs_raid_bio *alloc_rbio(struct btrfs_root *root,
 	rbio->fs_info = root->fs_info;
 	rbio->stripe_len = stripe_len;
 	rbio->nr_pages = num_pages;
+	rbio->stripe_npages = stripe_npages;
 	rbio->faila = -1;
 	rbio->failb = -1;
 	atomic_set(&rbio->refs, 1);
@@ -987,6 +1015,7 @@ static struct btrfs_raid_bio *alloc_rbio(struct btrfs_root *root,
 	p = rbio + 1;
 	rbio->stripe_pages = p;
 	rbio->bio_pages = p + sizeof(struct page *) * num_pages;
+	rbio->dbitmap = p + sizeof(struct page *) * num_pages * 2;
 
 	if (raid_map[bbio->num_stripes - 1] == RAID6_Q_STRIPE)
 		nr_data = bbio->num_stripes - 2;
@@ -1781,6 +1810,14 @@ static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
 	index_rbio_pages(rbio);
 
 	for (pagenr = 0; pagenr < nr_pages; pagenr++) {
+		/*
+		 * Now we just use bitmap to mark the horizontal stripes in
+		 * which we have data when doing parity scrub.
+		 */
+		if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
+		    !test_bit(pagenr, rbio->dbitmap))
+			continue;
+
 		/* setup our array of pointers with pages
 		 * from each stripe
 		 */
@@ -1925,7 +1962,13 @@ static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
 	} else if (err == 0) {
 		rbio->faila = -1;
 		rbio->failb = -1;
-		finish_rmw(rbio);
+
+		if (rbio->operation == BTRFS_RBIO_WRITE)
+			finish_rmw(rbio);
+		else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
+			finish_parity_scrub(rbio, 0);
+		else
+			BUG();
 	} else {
 		rbio_orig_end_io(rbio, err, 0);
 	}
@@ -2133,3 +2176,462 @@ static void read_rebuild_work(struct btrfs_work *work)
 	rbio = container_of(work, struct btrfs_raid_bio, work);
 	__raid56_parity_recover(rbio);
 }
+
+/*
+ * The following code is used to scrub/replace the parity stripe
+ *
+ * Note: We need make sure all the pages that add into the scrub/replace
+ * raid bio are correct and not be changed during the scrub/replace. That
+ * is those pages just hold metadata or file data with checksum.
+ */
+
+struct btrfs_raid_bio *
+raid56_parity_alloc_scrub_rbio(struct btrfs_root *root, struct bio *bio,
+			       struct btrfs_bio *bbio, u64 *raid_map,
+			       u64 stripe_len, struct btrfs_device *scrub_dev,
+			       unsigned long *dbitmap, int stripe_nsectors)
+{
+	struct btrfs_raid_bio *rbio;
+	int i;
+
+	rbio = alloc_rbio(root, bbio, raid_map, stripe_len);
+	if (IS_ERR(rbio))
+		return NULL;
+	bio_list_add(&rbio->bio_list, bio);
+	/*
+	 * This is a special bio which is used to hold the completion handler
+	 * and make the scrub rbio is similar to the other types
+	 */
+	ASSERT(!bio->bi_iter.bi_size);
+	rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
+
+	for (i = 0; i < bbio->num_stripes; i++) {
+		if (bbio->stripes[i].dev == scrub_dev) {
+			rbio->scrubp = i;
+			break;
+		}
+	}
+
+	/* Now we just support the sectorsize equals to page size */
+	ASSERT(root->sectorsize == PAGE_SIZE);
+	ASSERT(rbio->stripe_npages == stripe_nsectors);
+	bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
+
+	return rbio;
+}
+
+void raid56_parity_add_scrub_pages(struct btrfs_raid_bio *rbio,
+				   struct page *page, u64 logical)
+{
+	int stripe_offset;
+	int index;
+
+	ASSERT(logical >= rbio->raid_map[0]);
+	ASSERT(logical + PAGE_SIZE <= rbio->raid_map[0] +
+				rbio->stripe_len * rbio->nr_data);
+	stripe_offset = (int)(logical - rbio->raid_map[0]);
+	index = stripe_offset >> PAGE_CACHE_SHIFT;
+	rbio->bio_pages[index] = page;
+}
+
+/*
+ * We just scrub the parity that we have correct data on the same horizontal,
+ * so we needn't allocate all pages for all the stripes.
+ */
+static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
+{
+	int i;
+	int bit;
+	int index;
+	struct page *page;
+
+	for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
+		for (i = 0; i < rbio->bbio->num_stripes; i++) {
+			index = i * rbio->stripe_npages + bit;
+			if (rbio->stripe_pages[index])
+				continue;
+
+			page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
+			if (!page)
+				return -ENOMEM;
+			rbio->stripe_pages[index] = page;
+			ClearPageUptodate(page);
+		}
+	}
+	return 0;
+}
+
+/*
+ * end io function used by finish_rmw.  When we finally
+ * get here, we've written a full stripe
+ */
+static void raid_write_parity_end_io(struct bio *bio, int err)
+{
+	struct btrfs_raid_bio *rbio = bio->bi_private;
+
+	if (err)
+		fail_bio_stripe(rbio, bio);
+
+	bio_put(bio);
+
+	if (!atomic_dec_and_test(&rbio->stripes_pending))
+		return;
+
+	err = 0;
+
+	if (atomic_read(&rbio->error))
+		err = -EIO;
+
+	rbio_orig_end_io(rbio, err, 0);
+}
+
+static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
+					 int need_check)
+{
+	struct btrfs_bio *bbio = rbio->bbio;
+	void *pointers[bbio->num_stripes];
+	int nr_data = rbio->nr_data;
+	int stripe;
+	int pagenr;
+	int p_stripe = -1;
+	int q_stripe = -1;
+	struct page *p_page = NULL;
+	struct page *q_page = NULL;
+	struct bio_list bio_list;
+	struct bio *bio;
+	int ret;
+
+	bio_list_init(&bio_list);
+
+	if (bbio->num_stripes - rbio->nr_data == 1) {
+		p_stripe = bbio->num_stripes - 1;
+	} else if (bbio->num_stripes - rbio->nr_data == 2) {
+		p_stripe = bbio->num_stripes - 2;
+		q_stripe = bbio->num_stripes - 1;
+	} else {
+		BUG();
+	}
+
+	/*
+	 * Because the higher layers(scrubber) are unlikely to
+	 * use this area of the disk again soon, so don't cache
+	 * it.
+	 */
+	clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
+
+	if (!need_check)
+		goto writeback;
+
+	p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
+	if (!p_page)
+		goto cleanup;
+	SetPageUptodate(p_page);
+
+	if (q_stripe != -1) {
+		q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
+		if (!q_page) {
+			__free_page(p_page);
+			goto cleanup;
+		}
+		SetPageUptodate(q_page);
+	}
+
+	atomic_set(&rbio->error, 0);
+
+	for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
+		struct page *p;
+		void *parity;
+		/* first collect one page from each data stripe */
+		for (stripe = 0; stripe < nr_data; stripe++) {
+			p = page_in_rbio(rbio, stripe, pagenr, 0);
+			pointers[stripe] = kmap(p);
+		}
+
+		/* then add the parity stripe */
+		pointers[stripe++] = kmap(p_page);
+
+		if (q_stripe != -1) {
+
+			/*
+			 * raid6, add the qstripe and call the
+			 * library function to fill in our p/q
+			 */
+			pointers[stripe++] = kmap(q_page);
+
+			raid6_call.gen_syndrome(bbio->num_stripes, PAGE_SIZE,
+						pointers);
+		} else {
+			/* raid5 */
+			memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
+			run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE);
+		}
+
+		/* Check scrubbing pairty and repair it */
+		p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
+		parity = kmap(p);
+		if (memcmp(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE))
+			memcpy(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE);
+		else
+			/* Parity is right, needn't writeback */
+			bitmap_clear(rbio->dbitmap, pagenr, 1);
+		kunmap(p);
+
+		for (stripe = 0; stripe < bbio->num_stripes; stripe++)
+			kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
+	}
+
+	__free_page(p_page);
+	if (q_page)
+		__free_page(q_page);
+
+writeback:
+	/*
+	 * time to start writing.  Make bios for everything from the
+	 * higher layers (the bio_list in our rbio) and our p/q.  Ignore
+	 * everything else.
+	 */
+	for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
+		struct page *page;
+
+		page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
+		ret = rbio_add_io_page(rbio, &bio_list,
+			       page, rbio->scrubp, pagenr, rbio->stripe_len);
+		if (ret)
+			goto cleanup;
+	}
+
+	nr_data = bio_list_size(&bio_list);
+	if (!nr_data) {
+		/* Every parity is right */
+		rbio_orig_end_io(rbio, 0, 0);
+		return;
+	}
+
+	atomic_set(&rbio->stripes_pending, nr_data);
+
+	while (1) {
+		bio = bio_list_pop(&bio_list);
+		if (!bio)
+			break;
+
+		bio->bi_private = rbio;
+		bio->bi_end_io = raid_write_parity_end_io;
+		BUG_ON(!test_bit(BIO_UPTODATE, &bio->bi_flags));
+		submit_bio(WRITE, bio);
+	}
+	return;
+
+cleanup:
+	rbio_orig_end_io(rbio, -EIO, 0);
+}
+
+static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
+{
+	if (stripe >= 0 && stripe < rbio->nr_data)
+		return 1;
+	return 0;
+}
+
+/*
+ * While we're doing the parity check and repair, we could have errors
+ * in reading pages off the disk.  This checks for errors and if we're
+ * not able to read the page it'll trigger parity reconstruction.  The
+ * parity scrub will be finished after we've reconstructed the failed
+ * stripes
+ */
+static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio)
+{
+	if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
+		goto cleanup;
+
+	if (rbio->faila >= 0 || rbio->failb >= 0) {
+		int dfail = 0, failp = -1;
+
+		if (is_data_stripe(rbio, rbio->faila))
+			dfail++;
+		else if (is_parity_stripe(rbio->faila))
+			failp = rbio->faila;
+
+		if (is_data_stripe(rbio, rbio->failb))
+			dfail++;
+		else if (is_parity_stripe(rbio->failb))
+			failp = rbio->failb;
+
+		/*
+		 * Because we can not use a scrubbing parity to repair
+		 * the data, so the capability of the repair is declined.
+		 * (In the case of RAID5, we can not repair anything)
+		 */
+		if (dfail > rbio->bbio->max_errors - 1)
+			goto cleanup;
+
+		/*
+		 * If all data is good, only parity is correctly, just
+		 * repair the parity.
+		 */
+		if (dfail == 0) {
+			finish_parity_scrub(rbio, 0);
+			return;
+		}
+
+		/*
+		 * Here means we got one corrupted data stripe and one
+		 * corrupted parity on RAID6, if the corrupted parity
+		 * is scrubbing parity, luckly, use the other one to repair
+		 * the data, or we can not repair the data stripe.
+		 */
+		if (failp != rbio->scrubp)
+			goto cleanup;
+
+		__raid_recover_end_io(rbio);
+	} else {
+		finish_parity_scrub(rbio, 1);
+	}
+	return;
+
+cleanup:
+	rbio_orig_end_io(rbio, -EIO, 0);
+}
+
+/*
+ * end io for the read phase of the rmw cycle.  All the bios here are physical
+ * stripe bios we've read from the disk so we can recalculate the parity of the
+ * stripe.
+ *
+ * This will usually kick off finish_rmw once all the bios are read in, but it
+ * may trigger parity reconstruction if we had any errors along the way
+ */
+static void raid56_parity_scrub_end_io(struct bio *bio, int err)
+{
+	struct btrfs_raid_bio *rbio = bio->bi_private;
+
+	if (err)
+		fail_bio_stripe(rbio, bio);
+	else
+		set_bio_pages_uptodate(bio);
+
+	bio_put(bio);
+
+	if (!atomic_dec_and_test(&rbio->stripes_pending))
+		return;
+
+	/*
+	 * this will normally call finish_rmw to start our write
+	 * but if there are any failed stripes we'll reconstruct
+	 * from parity first
+	 */
+	validate_rbio_for_parity_scrub(rbio);
+}
+
+static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio)
+{
+	int bios_to_read = 0;
+	struct btrfs_bio *bbio = rbio->bbio;
+	struct bio_list bio_list;
+	int ret;
+	int pagenr;
+	int stripe;
+	struct bio *bio;
+
+	ret = alloc_rbio_essential_pages(rbio);
+	if (ret)
+		goto cleanup;
+
+	bio_list_init(&bio_list);
+
+	atomic_set(&rbio->error, 0);
+	/*
+	 * build a list of bios to read all the missing parts of this
+	 * stripe
+	 */
+	for (stripe = 0; stripe < bbio->num_stripes; stripe++) {
+		for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
+			struct page *page;
+			/*
+			 * we want to find all the pages missing from
+			 * the rbio and read them from the disk.  If
+			 * page_in_rbio finds a page in the bio list
+			 * we don't need to read it off the stripe.
+			 */
+			page = page_in_rbio(rbio, stripe, pagenr, 1);
+			if (page)
+				continue;
+
+			page = rbio_stripe_page(rbio, stripe, pagenr);
+			/*
+			 * the bio cache may have handed us an uptodate
+			 * page.  If so, be happy and use it
+			 */
+			if (PageUptodate(page))
+				continue;
+
+			ret = rbio_add_io_page(rbio, &bio_list, page,
+				       stripe, pagenr, rbio->stripe_len);
+			if (ret)
+				goto cleanup;
+		}
+	}
+
+	bios_to_read = bio_list_size(&bio_list);
+	if (!bios_to_read) {
+		/*
+		 * this can happen if others have merged with
+		 * us, it means there is nothing left to read.
+		 * But if there are missing devices it may not be
+		 * safe to do the full stripe write yet.
+		 */
+		goto finish;
+	}
+
+	/*
+	 * the bbio may be freed once we submit the last bio.  Make sure
+	 * not to touch it after that
+	 */
+	atomic_set(&rbio->stripes_pending, bios_to_read);
+	while (1) {
+		bio = bio_list_pop(&bio_list);
+		if (!bio)
+			break;
+
+		bio->bi_private = rbio;
+		bio->bi_end_io = raid56_parity_scrub_end_io;
+
+		btrfs_bio_wq_end_io(rbio->fs_info, bio,
+				    BTRFS_WQ_ENDIO_RAID56);
+
+		BUG_ON(!test_bit(BIO_UPTODATE, &bio->bi_flags));
+		submit_bio(READ, bio);
+	}
+	/* the actual write will happen once the reads are done */
+	return;
+
+cleanup:
+	rbio_orig_end_io(rbio, -EIO, 0);
+	return;
+
+finish:
+	validate_rbio_for_parity_scrub(rbio);
+}
+
+static void scrub_parity_work(struct btrfs_work *work)
+{
+	struct btrfs_raid_bio *rbio;
+
+	rbio = container_of(work, struct btrfs_raid_bio, work);
+	raid56_parity_scrub_stripe(rbio);
+}
+
+static void async_scrub_parity(struct btrfs_raid_bio *rbio)
+{
+	btrfs_init_work(&rbio->work, btrfs_rmw_helper,
+			scrub_parity_work, NULL, NULL);
+
+	btrfs_queue_work(rbio->fs_info->rmw_workers,
+			 &rbio->work);
+}
+
+void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
+{
+	if (!lock_stripe_add(rbio))
+		async_scrub_parity(rbio);
+}

fs/btrfs/raid56.h

@@ -39,6 +39,9 @@ static inline int nr_data_stripes(struct map_lookup *map)
 #define is_parity_stripe(x) (((x) == RAID5_P_STRIPE) ||		\
 			     ((x) == RAID6_Q_STRIPE))
 
+struct btrfs_raid_bio;
+struct btrfs_device;
+
 int raid56_parity_recover(struct btrfs_root *root, struct bio *bio,
 				 struct btrfs_bio *bbio, u64 *raid_map,
 				 u64 stripe_len, int mirror_num, int hold_bbio);
@@ -46,6 +49,15 @@ int raid56_parity_write(struct btrfs_root *root, struct bio *bio,
 			       struct btrfs_bio *bbio, u64 *raid_map,
 			       u64 stripe_len);
 
+struct btrfs_raid_bio *
+raid56_parity_alloc_scrub_rbio(struct btrfs_root *root, struct bio *bio,
+			       struct btrfs_bio *bbio, u64 *raid_map,
+			       u64 stripe_len, struct btrfs_device *scrub_dev,
+			       unsigned long *dbitmap, int stripe_nsectors);
+void raid56_parity_add_scrub_pages(struct btrfs_raid_bio *rbio,
+				   struct page *page, u64 logical);
+void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio);
+
 int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info);
 void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info);
 #endif