drbd: allow petabyte storage on 64bit arch

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

drivers/block/drbd/drbd_bitmap.c

@@ -37,10 +37,46 @@
  * convention:
  * function name drbd_bm_... => used elsewhere, "public".
  * function name      bm_... => internal to implementation, "private".
+ */
 
- * Note that since find_first_bit returns int, at the current granularity of
+
- * the bitmap (4KB per byte), this implementation "only" supports up to
+/*
- * 1<<(32+12) == 16 TB...
+ * LIMITATIONS:
+ * We want to support >= peta byte of backend storage, while for now still using
+ * a granularity of one bit per 4KiB of storage.
+ * 1 << 50		bytes backend storage (1 PiB)
+ * 1 << (50 - 12)	bits needed
+ *	38 --> we need u64 to index and count bits
+ * 1 << (38 - 3)	bitmap bytes needed
+ *	35 --> we still need u64 to index and count bytes
+ *			(that's 32 GiB of bitmap for 1 PiB storage)
+ * 1 << (35 - 2)	32bit longs needed
+ *	33 --> we'd even need u64 to index and count 32bit long words.
+ * 1 << (35 - 3)	64bit longs needed
+ *	32 --> we could get away with a 32bit unsigned int to index and count
+ *	64bit long words, but I rather stay with unsigned long for now.
+ *	We probably should neither count nor point to bytes or long words
+ *	directly, but either by bitnumber, or by page index and offset.
+ * 1 << (35 - 12)
+ *	22 --> we need that much 4KiB pages of bitmap.
+ *	1 << (22 + 3) --> on a 64bit arch,
+ *	we need 32 MiB to store the array of page pointers.
+ *
+ * Because I'm lazy, and because the resulting patch was too large, too ugly
+ * and still incomplete, on 32bit we still "only" support 16 TiB (minus some),
+ * (1 << 32) bits * 4k storage.
+ *
+
+ * bitmap storage and IO:
+ *	Bitmap is stored little endian on disk, and is kept little endian in
+ *	core memory. Currently we still hold the full bitmap in core as long
+ *	as we are "attached" to a local disk, which at 32 GiB for 1PiB storage
+ *	seems excessive.
+ *
+ *	We plan to reduce the amount of in-core bitmap pages by pageing them in
+ *	and out against their on-disk location as necessary, but need to make
+ *	sure we don't cause too much meta data IO, and must not deadlock in
+ *	tight memory situations. This needs some more work.
  */
 
 /*
@@ -56,13 +92,9 @@
 struct drbd_bitmap {
 	struct page **bm_pages;
 	spinlock_t bm_lock;
-	/* WARNING unsigned long bm_*:
+
-	 * 32bit number of bit offset is just enough for 512 MB bitmap.
+	/* see LIMITATIONS: above */
-	 * it will blow up if we make the bitmap bigger...
+
-	 * not that it makes much sense to have a bitmap that large,
-	 * rather change the granularity to 16k or 64k or something.
-	 * (that implies other problems, however...)
-	 */
 	unsigned long bm_set;       /* nr of set bits; THINK maybe atomic_t? */
 	unsigned long bm_bits;
 	size_t   bm_words;
@@ -517,43 +549,39 @@ static void bm_set_surplus(struct drbd_bitmap *b)
 	bm_unmap(p_addr);
 }
 
+/* you better not modify the bitmap while this is running,
+ * or its results will be stale */
 static unsigned long bm_count_bits(struct drbd_bitmap *b)
 {
-	unsigned long *p_addr, *bm, offset = 0;
+	unsigned long *p_addr;
 	unsigned long bits = 0;
-	unsigned long i, do_now;
+	unsigned long mask = (1UL << (b->bm_bits & BITS_PER_LONG_MASK)) -1;
-	unsigned long words;
+	int idx, last_page, i, last_word;
 
-	/* due to 64bit alignment, the last long on a 32bit arch
+	/* because of the "extra long to catch oob access" we allocate in
-	 * may be not used at all. The last used long will likely
+	 * drbd_bm_resize, bm_number_of_pages -1 is not necessarily the page
-	 * be only partially used, always. Don't count those bits,
+	 * containing the last _relevant_ bitmap word */
-	 * but mask them out. */
+	last_page = bm_bit_to_page_idx(b, b->bm_bits-1);
-	words = (b->bm_bits + BITS_PER_LONG - 1) >> LN2_BPL;
 
-	while (offset < words) {
+	/* all but last page */
-		i = do_now = min_t(size_t, words-offset, LWPP);
+	for (idx = 0; idx < last_page; idx++) {
-		p_addr = __bm_map_pidx(b, bm_word_to_page_idx(b, offset), KM_USER0);
+		p_addr = __bm_map_pidx(b, idx, KM_USER0);
-		bm = p_addr + MLPP(offset);
+		for (i = 0; i < LWPP; i++)
-		while (i--) {
+			bits += hweight_long(p_addr[i]);
-			bits += hweight_long(*bm++);
-		}
-		offset += do_now;
-		if (offset == words) {
-			/* last word may only be partially used,
-			 * see also bm_clear_surplus. */
-			i = (1UL << (b->bm_bits & (BITS_PER_LONG-1))) -1;
-			if (i) {
-				bits -= hweight_long(p_addr[do_now-1] & ~i);
-				p_addr[do_now-1] &= i;
-			}
-			/* 32bit arch, may have an unused padding long */
-			if (words != b->bm_words)
-				p_addr[do_now] = 0;
-		}
 		__bm_unmap(p_addr, KM_USER0);
 		cond_resched();
 	}
-
+	/* last (or only) page */
+	last_word = ((b->bm_bits - 1) & BITS_PER_PAGE_MASK) >> LN2_BPL;
+	p_addr = __bm_map_pidx(b, idx, KM_USER0);
+	for (i = 0; i < last_word; i++)
+		bits += hweight_long(p_addr[i]);
+	p_addr[last_word] &= cpu_to_lel(mask);
+	bits += hweight_long(p_addr[last_word]);
+	/* 32bit arch, may have an unused padding long */
+	if (BITS_PER_LONG == 32 && (last_word & 1) == 0)
+		p_addr[last_word+1] = 0;
+	__bm_unmap(p_addr, KM_USER0);
 	return bits;
 }
 
@@ -564,8 +592,6 @@ static void bm_memset(struct drbd_bitmap *b, size_t offset, int c, size_t len)
 	unsigned int idx;
 	size_t do_now, end;
 
-#define BM_SECTORS_PER_BIT (BM_BLOCK_SIZE/512)
-
 	end = offset + len;
 
 	if (end > b->bm_words) {
@@ -645,8 +671,14 @@ int drbd_bm_resize(struct drbd_conf *mdev, sector_t capacity, int set_new_bits)
 	words = ALIGN(bits, 64) >> LN2_BPL;
 
 	if (get_ldev(mdev)) {
-		D_ASSERT((u64)bits <= (((u64)mdev->ldev->md.md_size_sect-MD_BM_OFFSET) << 12));
+		u64 bits_on_disk = ((u64)mdev->ldev->md.md_size_sect-MD_BM_OFFSET) << 12;
 		put_ldev(mdev);
+		if (bits > bits_on_disk) {
+			dev_info(DEV, "bits = %lu\n", bits);
+			dev_info(DEV, "bits_on_disk = %llu\n", bits_on_disk);
+			err = -ENOSPC;
+			goto out;
+		}
 	}
 
 	/* one extra long to catch off by one errors */
@@ -1113,9 +1145,12 @@ int drbd_bm_write_lazy(struct drbd_conf *mdev, unsigned upper_idx) __must_hold(l
  * @mdev:	DRBD device.
  * @idx:	bitmap page index
  *
- * We don't want to special case on logical_block_size of the underlaying
+ * We don't want to special case on logical_block_size of the backend device,
- * device, so we submit PAGE_SIZE aligned pieces containing the requested enr.
+ * so we submit PAGE_SIZE aligned pieces.
  * Note that on "most" systems, PAGE_SIZE is 4k.
+ *
+ * In case this becomes an issue on systems with larger PAGE_SIZE,
+ * we may want to change this again to write 4k aligned 4k pieces.
  */
 int drbd_bm_write_page(struct drbd_conf *mdev, unsigned int idx) __must_hold(local)
 {
@@ -1144,52 +1179,57 @@ int drbd_bm_write_page(struct drbd_conf *mdev, unsigned int idx) __must_hold(loc
 
 /* NOTE
  * find_first_bit returns int, we return unsigned long.
- * should not make much difference anyways, but ...
+ * For this to work on 32bit arch with bitnumbers > (1<<32),
+ * we'd need to return u64, and get a whole lot of other places
+ * fixed where we still use unsigned long.
  *
  * this returns a bit number, NOT a sector!
  */
-#define BPP_MASK ((1UL << (PAGE_SHIFT+3)) - 1)
 static unsigned long __bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo,
 	const int find_zero_bit, const enum km_type km)
 {
 	struct drbd_bitmap *b = mdev->bitmap;
-	unsigned long i = -1UL;
 	unsigned long *p_addr;
-	unsigned long bit_offset; /* bit offset of the mapped page. */
+	unsigned long bit_offset;
+	unsigned i;
+
 
 	if (bm_fo > b->bm_bits) {
 		dev_err(DEV, "bm_fo=%lu bm_bits=%lu\n", bm_fo, b->bm_bits);
+		bm_fo = DRBD_END_OF_BITMAP;
 	} else {
 		while (bm_fo < b->bm_bits) {
 			/* bit offset of the first bit in the page */
-			bit_offset = bm_fo & ~BPP_MASK;
+			bit_offset = bm_fo & ~BITS_PER_PAGE_MASK;
 			p_addr = __bm_map_pidx(b, bm_bit_to_page_idx(b, bm_fo), km);
 
 			if (find_zero_bit)
-				i = generic_find_next_zero_le_bit(p_addr, PAGE_SIZE*8, bm_fo & BPP_MASK);
+				i = generic_find_next_zero_le_bit(p_addr,
+						PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK);
 			else
-				i = generic_find_next_le_bit(p_addr, PAGE_SIZE*8, bm_fo & BPP_MASK);
+				i = generic_find_next_le_bit(p_addr,
+						PAGE_SIZE*8, bm_fo & BITS_PER_PAGE_MASK);
 
 			__bm_unmap(p_addr, km);
 			if (i < PAGE_SIZE*8) {
-				i = bit_offset + i;
+				bm_fo = bit_offset + i;
-				if (i >= b->bm_bits)
+				if (bm_fo >= b->bm_bits)
 					break;
 				goto found;
 			}
 			bm_fo = bit_offset + PAGE_SIZE*8;
 		}
-		i = -1UL;
+		bm_fo = DRBD_END_OF_BITMAP;
 	}
  found:
-	return i;
+	return bm_fo;
 }
 
 static unsigned long bm_find_next(struct drbd_conf *mdev,
 	unsigned long bm_fo, const int find_zero_bit)
 {
 	struct drbd_bitmap *b = mdev->bitmap;
-	unsigned long i = -1UL;
+	unsigned long i = DRBD_END_OF_BITMAP;
 
 	ERR_IF(!b) return i;
 	ERR_IF(!b->bm_pages) return i;
@@ -1267,9 +1307,9 @@ static int __bm_change_bits_to(struct drbd_conf *mdev, const unsigned long s,
 			last_page_nr = page_nr;
 		}
 		if (val)
-			c += (0 == generic___test_and_set_le_bit(bitnr & BPP_MASK, p_addr));
+			c += (0 == generic___test_and_set_le_bit(bitnr & BITS_PER_PAGE_MASK, p_addr));
 		else
-			c -= (0 != generic___test_and_clear_le_bit(bitnr & BPP_MASK, p_addr));
+			c -= (0 != generic___test_and_clear_le_bit(bitnr & BITS_PER_PAGE_MASK, p_addr));
 	}
 	if (p_addr)
 		__bm_unmap(p_addr, km);
@@ -1418,7 +1458,7 @@ int drbd_bm_test_bit(struct drbd_conf *mdev, const unsigned long bitnr)
 		bm_print_lock_info(mdev);
 	if (bitnr < b->bm_bits) {
 		p_addr = bm_map_pidx(b, bm_bit_to_page_idx(b, bitnr));
-		i = generic_test_le_bit(bitnr & BPP_MASK, p_addr) ? 1 : 0;
+		i = generic_test_le_bit(bitnr & BITS_PER_PAGE_MASK, p_addr) ? 1 : 0;
 		bm_unmap(p_addr);
 	} else if (bitnr == b->bm_bits) {
 		i = -1;
@@ -1517,13 +1557,15 @@ int drbd_bm_e_weight(struct drbd_conf *mdev, unsigned long enr)
 	return count;
 }
 
-/* set all bits covered by the AL-extent al_enr */
+/* Set all bits covered by the AL-extent al_enr.
+ * Returns number of bits changed. */
 unsigned long drbd_bm_ALe_set_all(struct drbd_conf *mdev, unsigned long al_enr)
 {
 	struct drbd_bitmap *b = mdev->bitmap;
 	unsigned long *p_addr, *bm;
 	unsigned long weight;
-	int count, s, e, i, do_now;
+	unsigned long s, e;
+	int count, i, do_now;
 	ERR_IF(!b) return 0;
 	ERR_IF(!b->bm_pages) return 0;
 
@@ -1552,7 +1594,7 @@ unsigned long drbd_bm_ALe_set_all(struct drbd_conf *mdev, unsigned long al_enr)
 		if (e == b->bm_words)
 			b->bm_set -= bm_clear_surplus(b);
 	} else {
-		dev_err(DEV, "start offset (%d) too large in drbd_bm_ALe_set_all\n", s);
+		dev_err(DEV, "start offset (%lu) too large in drbd_bm_ALe_set_all\n", s);
 	}
 	weight = b->bm_set - weight;
 	spin_unlock_irq(&b->bm_lock);

drivers/block/drbd/drbd_int.h

@@ -1003,9 +1003,9 @@ struct drbd_conf {
 	struct hlist_head *tl_hash;
 	unsigned int tl_hash_s;
 
-	/* blocks to sync in this run [unit BM_BLOCK_SIZE] */
+	/* blocks to resync in this run [unit BM_BLOCK_SIZE] */
 	unsigned long rs_total;
-	/* number of sync IOs that failed in this run */
+	/* number of resync blocks that failed in this run */
 	unsigned long rs_failed;
 	/* Syncer's start time [unit jiffies] */
 	unsigned long rs_start;
@@ -1399,7 +1399,9 @@ struct bm_extent {
  * you should use 64bit OS for that much storage, anyways. */
 #define DRBD_MAX_SECTORS_FLEX BM_BIT_TO_SECT(0xffff7fff)
 #else
-#define DRBD_MAX_SECTORS_FLEX BM_BIT_TO_SECT(0x1LU << 32)
+/* we allow up to 1 PiB now on 64bit architecture with "flexible" meta data */
+#define DRBD_MAX_SECTORS_FLEX (1UL << 51)
+/* corresponds to (1UL << 38) bits right now. */
 #endif
 #endif
 
@@ -1419,11 +1421,15 @@ extern int  drbd_bm_resize(struct drbd_conf *mdev, sector_t sectors, int set_new
 extern void drbd_bm_cleanup(struct drbd_conf *mdev);
 extern void drbd_bm_set_all(struct drbd_conf *mdev);
 extern void drbd_bm_clear_all(struct drbd_conf *mdev);
+/* set/clear/test only a few bits at a time */
 extern int  drbd_bm_set_bits(
 		struct drbd_conf *mdev, unsigned long s, unsigned long e);
 extern int  drbd_bm_clear_bits(
 		struct drbd_conf *mdev, unsigned long s, unsigned long e);
-/* bm_set_bits variant for use while holding drbd_bm_lock */
+extern int drbd_bm_count_bits(
+	struct drbd_conf *mdev, const unsigned long s, const unsigned long e);
+/* bm_set_bits variant for use while holding drbd_bm_lock,
+ * may process the whole bitmap in one go */
 extern void _drbd_bm_set_bits(struct drbd_conf *mdev,
 		const unsigned long s, const unsigned long e);
 extern int  drbd_bm_test_bit(struct drbd_conf *mdev, unsigned long bitnr);
@@ -1436,6 +1442,8 @@ extern unsigned long drbd_bm_ALe_set_all(struct drbd_conf *mdev,
 extern size_t	     drbd_bm_words(struct drbd_conf *mdev);
 extern unsigned long drbd_bm_bits(struct drbd_conf *mdev);
 extern sector_t      drbd_bm_capacity(struct drbd_conf *mdev);
+
+#define DRBD_END_OF_BITMAP	(~(unsigned long)0)
 extern unsigned long drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo);
 /* bm_find_next variants for use while you hold drbd_bm_lock() */
 extern unsigned long _drbd_bm_find_next(struct drbd_conf *mdev, unsigned long bm_fo);
@@ -1452,8 +1460,6 @@ extern void drbd_bm_get_lel(struct drbd_conf *mdev, size_t offset,
 
 extern void drbd_bm_lock(struct drbd_conf *mdev, char *why);
 extern void drbd_bm_unlock(struct drbd_conf *mdev);
-
-extern int drbd_bm_count_bits(struct drbd_conf *mdev, const unsigned long s, const unsigned long e);
 /* drbd_main.c */
 
 extern struct kmem_cache *drbd_request_cache;
@@ -2158,10 +2164,8 @@ extern int _get_ldev_if_state(struct drbd_conf *mdev, enum drbd_disk_state mins)
 static inline void drbd_get_syncer_progress(struct drbd_conf *mdev,
 		unsigned long *bits_left, unsigned int *per_mil_done)
 {
-	/*
+	/* this is to break it at compile time when we change that, in case we
-	 * this is to break it at compile time when we change that
+	 * want to support more than (1<<32) bits on a 32bit arch. */
-	 * (we may feel 4TB maximum storage per drbd is not enough)
-	 */
 	typecheck(unsigned long, mdev->rs_total);
 
 	/* note: both rs_total and rs_left are in bits, i.e. in
@@ -2186,10 +2190,19 @@ static inline void drbd_get_syncer_progress(struct drbd_conf *mdev,
 				*bits_left, mdev->rs_total, mdev->rs_failed);
 		*per_mil_done = 0;
 	} else {
-		/* make sure the calculation happens in long context */
+		/* Make sure the division happens in long context.
-		unsigned long tmp = 1000UL -
+		 * We allow up to one petabyte storage right now,
-				(*bits_left >> 10)*1000UL
+		 * at a granularity of 4k per bit that is 2**38 bits.
-				/ ((mdev->rs_total >> 10) + 1UL);
+		 * After shift right and multiplication by 1000,
+		 * this should still fit easily into a 32bit long,
+		 * so we don't need a 64bit division on 32bit arch.
+		 * Note: currently we don't support such large bitmaps on 32bit
+		 * arch anyways, but no harm done to be prepared for it here.
+		 */
+		unsigned int shift = mdev->rs_total >= (1ULL << 32) ? 16 : 10;
+		unsigned long left = *bits_left >> shift;
+		unsigned long total = 1UL + (mdev->rs_total >> shift);
+		unsigned long tmp = 1000UL - left * 1000UL/total;
 		*per_mil_done = tmp;
 	}
 }

drivers/block/drbd/drbd_nl.c

@@ -527,17 +527,19 @@ static void drbd_md_set_sector_offsets(struct drbd_conf *mdev,
 	}
 }
 
+/* input size is expected to be in KB */
 char *ppsize(char *buf, unsigned long long size)
 {
-	/* Needs 9 bytes at max. */
+	/* Needs 9 bytes at max including trailing NUL:
+	 * -1ULL ==> "16384 EB" */
 	static char units[] = { 'K', 'M', 'G', 'T', 'P', 'E' };
 	int base = 0;
-	while (size >= 10000) {
+	while (size >= 10000 && base < sizeof(units)-1) {
 		/* shift + round */
 		size = (size >> 10) + !!(size & (1<<9));
 		base++;
 	}
-	sprintf(buf, "%lu %cB", (long)size, units[base]);
+	sprintf(buf, "%u %cB", (unsigned)size, units[base]);
 
 	return buf;
 }

drivers/block/drbd/drbd_proc.c

@@ -91,9 +91,9 @@ static void drbd_syncer_progress(struct drbd_conf *mdev, struct seq_file *seq)
 		seq_printf(seq, "sync'ed:");
 	seq_printf(seq, "%3u.%u%% ", res / 10, res % 10);
 
-	/* if more than 1 GB display in MB */
+	/* if more than a few GB, display in MB */
-	if (mdev->rs_total > 0x100000L)
+	if (mdev->rs_total > (4UL << (30 - BM_BLOCK_SHIFT)))
-		seq_printf(seq, "(%lu/%lu)M\n\t",
+		seq_printf(seq, "(%lu/%lu)M",
 			    (unsigned long) Bit2KB(rs_left >> 10),
 			    (unsigned long) Bit2KB(mdev->rs_total >> 10));
 	else

drivers/block/drbd/drbd_worker.c

@@ -577,7 +577,7 @@ next_sector:
 		size = BM_BLOCK_SIZE;
 		bit  = drbd_bm_find_next(mdev, mdev->bm_resync_fo);
 
-		if (bit == -1UL) {
+		if (bit == DRBD_END_OF_BITMAP) {
 			mdev->bm_resync_fo = drbd_bm_bits(mdev);
 			mdev->resync_work.cb = w_resync_inactive;
 			put_ldev(mdev);