drbd: Remove the open-coded page pool

If the network stack keeps a reference for too long, DRBD keeps
references on a higher number of pages as a consequence.

Fix all that by no longer relying on page reference counts dropping to
an expected value. Instead, DRBD gives up its reference and lets the
system handle everything else. While at it, remove the open-coded
custom page pool mechanism and use the page_pool included in the
kernel.

Signed-off-by: Philipp Reisner <[email protected]>
Signed-off-by: Christoph Böhmwalder <[email protected]>

Author: Philipp-Reisner

drivers/block/drbd/drbd_int.h

@@ -380,6 +380,9 @@ enum {
 	/* this is/was a write request */
 	__EE_WRITE,
 
+	/* hand back using mempool_free(e, drbd_buffer_page_pool) */
+	__EE_RELEASE_TO_MEMPOOL,
+
 	/* this is/was a write same request */
 	__EE_WRITE_SAME,
 
@@ -402,6 +405,7 @@ enum {
 #define EE_IN_INTERVAL_TREE	(1<<__EE_IN_INTERVAL_TREE)
 #define EE_SUBMITTED		(1<<__EE_SUBMITTED)
 #define EE_WRITE		(1<<__EE_WRITE)
+#define EE_RELEASE_TO_MEMPOOL	(1<<__EE_RELEASE_TO_MEMPOOL)
 #define EE_WRITE_SAME		(1<<__EE_WRITE_SAME)
 #define EE_APPLICATION		(1<<__EE_APPLICATION)
 #define EE_RS_THIN_REQ		(1<<__EE_RS_THIN_REQ)
@@ -858,7 +862,6 @@ struct drbd_device {
 	struct list_head sync_ee;   /* IO in progress (P_RS_DATA_REPLY gets written to disk) */
 	struct list_head done_ee;   /* need to send P_WRITE_ACK */
 	struct list_head read_ee;   /* [RS]P_DATA_REQUEST being read */
-	struct list_head net_ee;    /* zero-copy network send in progress */
 
 	struct list_head resync_reads;
 	atomic_t pp_in_use;		/* allocated from page pool */
@@ -1329,31 +1332,14 @@ extern struct kmem_cache *drbd_al_ext_cache;	/* activity log extents */
 extern mempool_t drbd_request_mempool;
 extern mempool_t drbd_ee_mempool;
 
-/* drbd's page pool, used to buffer data received from the peer,
- * or data requested by the peer.
- *
- * This does not have an emergency reserve.
- *
- * When allocating from this pool, it first takes pages from the pool.
- * Only if the pool is depleted will try to allocate from the system.
- *
- * The assumption is that pages taken from this pool will be processed,
- * and given back, "quickly", and then can be recycled, so we can avoid
- * frequent calls to alloc_page(), and still will be able to make progress even
- * under memory pressure.
- */
-extern struct page *drbd_pp_pool;
-extern spinlock_t   drbd_pp_lock;
-extern int	    drbd_pp_vacant;
-extern wait_queue_head_t drbd_pp_wait;
-
 /* We also need a standard (emergency-reserve backed) page pool
  * for meta data IO (activity log, bitmap).
  * We can keep it global, as long as it is used as "N pages at a time".
  * 128 should be plenty, currently we probably can get away with as few as 1.
  */
 #define DRBD_MIN_POOL_PAGES	128
 extern mempool_t drbd_md_io_page_pool;
+extern mempool_t drbd_buffer_page_pool;
 
 /* We also need to make sure we get a bio
  * when we need it for housekeeping purposes */
@@ -1488,10 +1474,7 @@ extern struct drbd_peer_request *drbd_alloc_peer_req(struct drbd_peer_device *,
 						     sector_t, unsigned int,
 						     unsigned int,
 						     gfp_t) __must_hold(local);
-extern void __drbd_free_peer_req(struct drbd_device *, struct drbd_peer_request *,
-				 int);
-#define drbd_free_peer_req(m,e) __drbd_free_peer_req(m, e, 0)
-#define drbd_free_net_peer_req(m,e) __drbd_free_peer_req(m, e, 1)
+extern void drbd_free_peer_req(struct drbd_device *device, struct drbd_peer_request *req);
 extern struct page *drbd_alloc_pages(struct drbd_peer_device *, unsigned int, bool);
 extern void _drbd_clear_done_ee(struct drbd_device *device, struct list_head *to_be_freed);
 extern int drbd_connected(struct drbd_peer_device *);
@@ -1610,16 +1593,6 @@ static inline struct page *page_chain_next(struct page *page)
 	for (; page && ({ n = page_chain_next(page); 1; }); page = n)
 
 
-static inline int drbd_peer_req_has_active_page(struct drbd_peer_request *peer_req)
-{
-	struct page *page = peer_req->pages;
-	page_chain_for_each(page) {
-		if (page_count(page) > 1)
-			return 1;
-	}
-	return 0;
-}
-
 static inline union drbd_state drbd_read_state(struct drbd_device *device)
 {
 	struct drbd_resource *resource = device->resource;

drivers/block/drbd/drbd_main.c

@@ -114,20 +114,10 @@ struct kmem_cache *drbd_al_ext_cache;	/* activity log extents */
 mempool_t drbd_request_mempool;
 mempool_t drbd_ee_mempool;
 mempool_t drbd_md_io_page_pool;
+mempool_t drbd_buffer_page_pool;
 struct bio_set drbd_md_io_bio_set;
 struct bio_set drbd_io_bio_set;
 
-/* I do not use a standard mempool, because:
-   1) I want to hand out the pre-allocated objects first.
-   2) I want to be able to interrupt sleeping allocation with a signal.
-   Note: This is a single linked list, the next pointer is the private
-	 member of struct page.
- */
-struct page *drbd_pp_pool;
-DEFINE_SPINLOCK(drbd_pp_lock);
-int          drbd_pp_vacant;
-wait_queue_head_t drbd_pp_wait;
-
 DEFINE_RATELIMIT_STATE(drbd_ratelimit_state, 5 * HZ, 5);
 
 static const struct block_device_operations drbd_ops = {
@@ -1611,6 +1601,7 @@ static int _drbd_send_zc_bio(struct drbd_peer_device *peer_device, struct bio *b
 static int _drbd_send_zc_ee(struct drbd_peer_device *peer_device,
 			    struct drbd_peer_request *peer_req)
 {
+	bool use_sendpage = !(peer_req->flags & EE_RELEASE_TO_MEMPOOL);
 	struct page *page = peer_req->pages;
 	unsigned len = peer_req->i.size;
 	int err;
@@ -1619,8 +1610,13 @@ static int _drbd_send_zc_ee(struct drbd_peer_device *peer_device,
 	page_chain_for_each(page) {
 		unsigned l = min_t(unsigned, len, PAGE_SIZE);
 
-		err = _drbd_send_page(peer_device, page, 0, l,
-				      page_chain_next(page) ? MSG_MORE : 0);
+		if (likely(use_sendpage))
+			err = _drbd_send_page(peer_device, page, 0, l,
+					      page_chain_next(page) ? MSG_MORE : 0);
+		else
+			err = _drbd_no_send_page(peer_device, page, 0, l,
+						 page_chain_next(page) ? MSG_MORE : 0);
+
 		if (err)
 			return err;
 		len -= l;
@@ -1962,7 +1958,6 @@ void drbd_init_set_defaults(struct drbd_device *device)
 	INIT_LIST_HEAD(&device->sync_ee);
 	INIT_LIST_HEAD(&device->done_ee);
 	INIT_LIST_HEAD(&device->read_ee);
-	INIT_LIST_HEAD(&device->net_ee);
 	INIT_LIST_HEAD(&device->resync_reads);
 	INIT_LIST_HEAD(&device->resync_work.list);
 	INIT_LIST_HEAD(&device->unplug_work.list);
@@ -2043,7 +2038,6 @@ void drbd_device_cleanup(struct drbd_device *device)
 	D_ASSERT(device, list_empty(&device->sync_ee));
 	D_ASSERT(device, list_empty(&device->done_ee));
 	D_ASSERT(device, list_empty(&device->read_ee));
-	D_ASSERT(device, list_empty(&device->net_ee));
 	D_ASSERT(device, list_empty(&device->resync_reads));
 	D_ASSERT(device, list_empty(&first_peer_device(device)->connection->sender_work.q));
 	D_ASSERT(device, list_empty(&device->resync_work.list));
@@ -2055,19 +2049,11 @@ void drbd_device_cleanup(struct drbd_device *device)
 
 static void drbd_destroy_mempools(void)
 {
-	struct page *page;
-
-	while (drbd_pp_pool) {
-		page = drbd_pp_pool;
-		drbd_pp_pool = (struct page *)page_private(page);
-		__free_page(page);
-		drbd_pp_vacant--;
-	}
-
 	/* D_ASSERT(device, atomic_read(&drbd_pp_vacant)==0); */
 
 	bioset_exit(&drbd_io_bio_set);
 	bioset_exit(&drbd_md_io_bio_set);
+	mempool_exit(&drbd_buffer_page_pool);
 	mempool_exit(&drbd_md_io_page_pool);
 	mempool_exit(&drbd_ee_mempool);
 	mempool_exit(&drbd_request_mempool);
@@ -2086,9 +2072,8 @@ static void drbd_destroy_mempools(void)
 
 static int drbd_create_mempools(void)
 {
-	struct page *page;
 	const int number = (DRBD_MAX_BIO_SIZE/PAGE_SIZE) * drbd_minor_count;
-	int i, ret;
+	int ret;
 
 	/* caches */
 	drbd_request_cache = kmem_cache_create(
@@ -2125,6 +2110,10 @@ static int drbd_create_mempools(void)
 	if (ret)
 		goto Enomem;
 
+	ret = mempool_init_page_pool(&drbd_buffer_page_pool, number, 0);
+	if (ret)
+		goto Enomem;
+
 	ret = mempool_init_slab_pool(&drbd_request_mempool, number,
 				     drbd_request_cache);
 	if (ret)
@@ -2134,15 +2123,6 @@ static int drbd_create_mempools(void)
 	if (ret)
 		goto Enomem;
 
-	for (i = 0; i < number; i++) {
-		page = alloc_page(GFP_HIGHUSER);
-		if (!page)
-			goto Enomem;
-		set_page_private(page, (unsigned long)drbd_pp_pool);
-		drbd_pp_pool = page;
-	}
-	drbd_pp_vacant = number;
-
 	return 0;
 
 Enomem:
@@ -2169,10 +2149,6 @@ static void drbd_release_all_peer_reqs(struct drbd_device *device)
 	rr = drbd_free_peer_reqs(device, &device->done_ee);
 	if (rr)
 		drbd_err(device, "%d EEs in done list found!\n", rr);
-
-	rr = drbd_free_peer_reqs(device, &device->net_ee);
-	if (rr)
-		drbd_err(device, "%d EEs in net list found!\n", rr);
 }
 
 /* caution. no locking. */
@@ -2863,11 +2839,6 @@ static int __init drbd_init(void)
 		return err;
 	}
 
-	/*
-	 * allocate all necessary structs
-	 */
-	init_waitqueue_head(&drbd_pp_wait);
-
 	drbd_proc = NULL; /* play safe for drbd_cleanup */
 	idr_init(&drbd_devices);