Merge branch 'net-rds-RDMA-fixes'

Gerd Rausch says:

====================
net/rds: RDMA fixes

A number of net/rds fixes necessary to make "rds_rdma.ko"
pass some basic Oracle internal tests.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

net/rds/ib.h

@@ -156,6 +156,7 @@ struct rds_ib_connection {
 
 	/* To control the number of wrs from fastreg */
 	atomic_t		i_fastreg_wrs;
+	atomic_t		i_fastreg_inuse_count;
 
 	/* interrupt handling */
 	struct tasklet_struct	i_send_tasklet;

net/rds/ib_cm.c

@@ -40,6 +40,7 @@
 #include "rds_single_path.h"
 #include "rds.h"
 #include "ib.h"
+#include "ib_mr.h"
 
 /*
  * Set the selected protocol version
@@ -526,7 +527,6 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
 	attr.qp_type = IB_QPT_RC;
 	attr.send_cq = ic->i_send_cq;
 	attr.recv_cq = ic->i_recv_cq;
-	atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
 
 	/*
 	 * XXX this can fail if max_*_wr is too large?  Are we supposed
@@ -993,6 +993,11 @@ void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
 				ic->i_cm_id, err);
 		}
 
+		/* kick off "flush_worker" for all pools in order to reap
+		 * all FRMR registrations that are still marked "FRMR_IS_INUSE"
+		 */
+		rds_ib_flush_mrs();
+
 		/*
 		 * We want to wait for tx and rx completion to finish
 		 * before we tear down the connection, but we have to be
@@ -1005,6 +1010,7 @@ void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
 		wait_event(rds_ib_ring_empty_wait,
 			   rds_ib_ring_empty(&ic->i_recv_ring) &&
 			   (atomic_read(&ic->i_signaled_sends) == 0) &&
+			   (atomic_read(&ic->i_fastreg_inuse_count) == 0) &&
 			   (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
 		tasklet_kill(&ic->i_send_tasklet);
 		tasklet_kill(&ic->i_recv_tasklet);
@@ -1132,6 +1138,7 @@ int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 	spin_lock_init(&ic->i_ack_lock);
 #endif
 	atomic_set(&ic->i_signaled_sends, 0);
+	atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
 
 	/*
 	 * rds_ib_conn_shutdown() waits for these to be emptied so they

net/rds/ib_frmr.c

@@ -32,6 +32,24 @@
 
 #include "ib_mr.h"
 
+static inline void
+rds_transition_frwr_state(struct rds_ib_mr *ibmr,
+			  enum rds_ib_fr_state old_state,
+			  enum rds_ib_fr_state new_state)
+{
+	if (cmpxchg(&ibmr->u.frmr.fr_state,
+		    old_state, new_state) == old_state &&
+	    old_state == FRMR_IS_INUSE) {
+		/* enforce order of ibmr->u.frmr.fr_state update
+		 * before decrementing i_fastreg_inuse_count
+		 */
+		smp_mb__before_atomic();
+		atomic_dec(&ibmr->ic->i_fastreg_inuse_count);
+		if (waitqueue_active(&rds_ib_ring_empty_wait))
+			wake_up(&rds_ib_ring_empty_wait);
+	}
+}
+
 static struct rds_ib_mr *rds_ib_alloc_frmr(struct rds_ib_device *rds_ibdev,
 					   int npages)
 {
@@ -75,6 +93,8 @@ static struct rds_ib_mr *rds_ib_alloc_frmr(struct rds_ib_device *rds_ibdev,
 		pool->max_items_soft = pool->max_items;
 
 	frmr->fr_state = FRMR_IS_FREE;
+	init_waitqueue_head(&frmr->fr_inv_done);
+	init_waitqueue_head(&frmr->fr_reg_done);
 	return ibmr;
 
 out_no_cigar:
@@ -116,13 +136,19 @@ static int rds_ib_post_reg_frmr(struct rds_ib_mr *ibmr)
 	if (unlikely(ret != ibmr->sg_len))
 		return ret < 0 ? ret : -EINVAL;
 
+	if (cmpxchg(&frmr->fr_state,
+		    FRMR_IS_FREE, FRMR_IS_INUSE) != FRMR_IS_FREE)
+		return -EBUSY;
+
+	atomic_inc(&ibmr->ic->i_fastreg_inuse_count);
+
 	/* Perform a WR for the fast_reg_mr. Each individual page
 	 * in the sg list is added to the fast reg page list and placed
 	 * inside the fast_reg_mr WR.  The key used is a rolling 8bit
 	 * counter, which should guarantee uniqueness.
 	 */
 	ib_update_fast_reg_key(frmr->mr, ibmr->remap_count++);
-	frmr->fr_state = FRMR_IS_INUSE;
+	frmr->fr_reg = true;
 
 	memset(&reg_wr, 0, sizeof(reg_wr));
 	reg_wr.wr.wr_id = (unsigned long)(void *)ibmr;
@@ -138,12 +164,23 @@ static int rds_ib_post_reg_frmr(struct rds_ib_mr *ibmr)
 	ret = ib_post_send(ibmr->ic->i_cm_id->qp, &reg_wr.wr, NULL);
 	if (unlikely(ret)) {
 		/* Failure here can be because of -ENOMEM as well */
-		frmr->fr_state = FRMR_IS_STALE;
+		rds_transition_frwr_state(ibmr, FRMR_IS_INUSE, FRMR_IS_STALE);
+
 		atomic_inc(&ibmr->ic->i_fastreg_wrs);
 		if (printk_ratelimit())
 			pr_warn("RDS/IB: %s returned error(%d)\n",
 				__func__, ret);
+		goto out;
 	}
+
+	/* Wait for the registration to complete in order to prevent an invalid
+	 * access error resulting from a race between the memory region already
+	 * being accessed while registration is still pending.
+	 */
+	wait_event(frmr->fr_reg_done, !frmr->fr_reg);
+
+out:
+
 	return ret;
 }
 
@@ -255,12 +292,29 @@ static int rds_ib_post_inv(struct rds_ib_mr *ibmr)
 
 	ret = ib_post_send(i_cm_id->qp, s_wr, NULL);
 	if (unlikely(ret)) {
-		frmr->fr_state = FRMR_IS_STALE;
+		rds_transition_frwr_state(ibmr, FRMR_IS_INUSE, FRMR_IS_STALE);
 		frmr->fr_inv = false;
+		/* enforce order of frmr->fr_inv update
+		 * before incrementing i_fastreg_wrs
+		 */
+		smp_mb__before_atomic();
 		atomic_inc(&ibmr->ic->i_fastreg_wrs);
 		pr_err("RDS/IB: %s returned error(%d)\n", __func__, ret);
 		goto out;
 	}
+
+	/* Wait for the FRMR_IS_FREE (or FRMR_IS_STALE) transition in order to
+	 * 1) avoid a silly bouncing between "clean_list" and "drop_list"
+	 *    triggered by function "rds_ib_reg_frmr" as it is releases frmr
+	 *    regions whose state is not "FRMR_IS_FREE" right away.
+	 * 2) prevents an invalid access error in a race
+	 *    from a pending "IB_WR_LOCAL_INV" operation
+	 *    with a teardown ("dma_unmap_sg", "put_page")
+	 *    and de-registration ("ib_dereg_mr") of the corresponding
+	 *    memory region.
+	 */
+	wait_event(frmr->fr_inv_done, frmr->fr_state != FRMR_IS_INUSE);
+
 out:
 	return ret;
 }
@@ -271,7 +325,7 @@ void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
 	struct rds_ib_frmr *frmr = &ibmr->u.frmr;
 
 	if (wc->status != IB_WC_SUCCESS) {
-		frmr->fr_state = FRMR_IS_STALE;
+		rds_transition_frwr_state(ibmr, FRMR_IS_INUSE, FRMR_IS_STALE);
 		if (rds_conn_up(ic->conn))
 			rds_ib_conn_error(ic->conn,
 					  "frmr completion <%pI4,%pI4> status %u(%s), vendor_err 0x%x, disconnecting and reconnecting\n",
@@ -283,10 +337,20 @@ void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
 	}
 
 	if (frmr->fr_inv) {
-		frmr->fr_state = FRMR_IS_FREE;
+		rds_transition_frwr_state(ibmr, FRMR_IS_INUSE, FRMR_IS_FREE);
 		frmr->fr_inv = false;
+		wake_up(&frmr->fr_inv_done);
 	}
 
+	if (frmr->fr_reg) {
+		frmr->fr_reg = false;
+		wake_up(&frmr->fr_reg_done);
+	}
+
+	/* enforce order of frmr->{fr_reg,fr_inv} update
+	 * before incrementing i_fastreg_wrs
+	 */
+	smp_mb__before_atomic();
 	atomic_inc(&ic->i_fastreg_wrs);
 }
 
@@ -295,14 +359,18 @@ void rds_ib_unreg_frmr(struct list_head *list, unsigned int *nfreed,
 {
 	struct rds_ib_mr *ibmr, *next;
 	struct rds_ib_frmr *frmr;
-	int ret = 0;
+	int ret = 0, ret2;
 	unsigned int freed = *nfreed;
 
 	/* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
 	list_for_each_entry(ibmr, list, unmap_list) {
-		if (ibmr->sg_dma_len)
-			ret |= rds_ib_post_inv(ibmr);
+		if (ibmr->sg_dma_len) {
+			ret2 = rds_ib_post_inv(ibmr);
+			if (ret2 && !ret)
+				ret = ret2;
+		}
 	}
+
 	if (ret)
 		pr_warn("RDS/IB: %s failed (err=%d)\n", __func__, ret);
 

net/rds/ib_mr.h

@@ -57,6 +57,9 @@ struct rds_ib_frmr {
 	struct ib_mr		*mr;
 	enum rds_ib_fr_state	fr_state;
 	bool			fr_inv;
+	wait_queue_head_t	fr_inv_done;
+	bool			fr_reg;
+	wait_queue_head_t	fr_reg_done;
 	struct ib_send_wr	fr_wr;
 	unsigned int		dma_npages;
 	unsigned int		sg_byte_len;
@@ -97,6 +100,7 @@ struct rds_ib_mr_pool {
 	struct llist_head	free_list;	/* unused MRs */
 	struct llist_head	clean_list;	/* unused & unmapped MRs */
 	wait_queue_head_t	flush_wait;
+	spinlock_t		clean_lock;	/* "clean_list" concurrency */
 
 	atomic_t		free_pinned;	/* memory pinned by free MRs */
 	unsigned long		max_items;

net/rds/ib_rdma.c

@@ -40,9 +40,6 @@
 
 struct workqueue_struct *rds_ib_mr_wq;
 
-static DEFINE_PER_CPU(unsigned long, clean_list_grace);
-#define CLEAN_LIST_BUSY_BIT 0
-
 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
 {
 	struct rds_ib_device *rds_ibdev;
@@ -195,12 +192,11 @@ struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
 {
 	struct rds_ib_mr *ibmr = NULL;
 	struct llist_node *ret;
-	unsigned long *flag;
+	unsigned long flags;
 
-	preempt_disable();
-	flag = this_cpu_ptr(&clean_list_grace);
-	set_bit(CLEAN_LIST_BUSY_BIT, flag);
+	spin_lock_irqsave(&pool->clean_lock, flags);
 	ret = llist_del_first(&pool->clean_list);
+	spin_unlock_irqrestore(&pool->clean_lock, flags);
 	if (ret) {
 		ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
 		if (pool->pool_type == RDS_IB_MR_8K_POOL)
@@ -209,23 +205,9 @@ struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
 			rds_ib_stats_inc(s_ib_rdma_mr_1m_reused);
 	}
 
-	clear_bit(CLEAN_LIST_BUSY_BIT, flag);
-	preempt_enable();
 	return ibmr;
 }
 
-static inline void wait_clean_list_grace(void)
-{
-	int cpu;
-	unsigned long *flag;
-
-	for_each_online_cpu(cpu) {
-		flag = &per_cpu(clean_list_grace, cpu);
-		while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
-			cpu_relax();
-	}
-}
-
 void rds_ib_sync_mr(void *trans_private, int direction)
 {
 	struct rds_ib_mr *ibmr = trans_private;
@@ -324,8 +306,7 @@ static unsigned int llist_append_to_list(struct llist_head *llist,
  * of clusters.  Each cluster has linked llist nodes of
  * MR_CLUSTER_SIZE mrs that are ready for reuse.
  */
-static void list_to_llist_nodes(struct rds_ib_mr_pool *pool,
-				struct list_head *list,
+static void list_to_llist_nodes(struct list_head *list,
 				struct llist_node **nodes_head,
 				struct llist_node **nodes_tail)
 {
@@ -402,8 +383,13 @@ int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 	 */
 	dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
 	dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
-	if (free_all)
+	if (free_all) {
+		unsigned long flags;
+
+		spin_lock_irqsave(&pool->clean_lock, flags);
 		llist_append_to_list(&pool->clean_list, &unmap_list);
+		spin_unlock_irqrestore(&pool->clean_lock, flags);
+	}
 
 	free_goal = rds_ib_flush_goal(pool, free_all);
 
@@ -416,27 +402,20 @@ int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 		rds_ib_unreg_fmr(&unmap_list, &nfreed, &unpinned, free_goal);
 
 	if (!list_empty(&unmap_list)) {
-		/* we have to make sure that none of the things we're about
-		 * to put on the clean list would race with other cpus trying
-		 * to pull items off.  The llist would explode if we managed to
-		 * remove something from the clean list and then add it back again
-		 * while another CPU was spinning on that same item in llist_del_first.
-		 *
-		 * This is pretty unlikely, but just in case  wait for an llist grace period
-		 * here before adding anything back into the clean list.
-		 */
-		wait_clean_list_grace();
+		unsigned long flags;
 
-		list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail);
+		list_to_llist_nodes(&unmap_list, &clean_nodes, &clean_tail);
 		if (ibmr_ret) {
 			*ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
 			clean_nodes = clean_nodes->next;
 		}
 		/* more than one entry in llist nodes */
-		if (clean_nodes)
+		if (clean_nodes) {
+			spin_lock_irqsave(&pool->clean_lock, flags);
 			llist_add_batch(clean_nodes, clean_tail,
 					&pool->clean_list);
-
+			spin_unlock_irqrestore(&pool->clean_lock, flags);
+		}
 	}
 
 	atomic_sub(unpinned, &pool->free_pinned);
@@ -471,7 +450,7 @@ struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool)
 				rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
 			else
 				rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
-			return ERR_PTR(-EAGAIN);
+			break;
 		}
 
 		/* We do have some empty MRs. Flush them out. */
@@ -485,7 +464,7 @@ struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool)
 			return ibmr;
 	}
 
-	return ibmr;
+	return NULL;
 }
 
 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
@@ -610,6 +589,7 @@ struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
 	init_llist_head(&pool->free_list);
 	init_llist_head(&pool->drop_list);
 	init_llist_head(&pool->clean_list);
+	spin_lock_init(&pool->clean_lock);
 	mutex_init(&pool->flush_lock);
 	init_waitqueue_head(&pool->flush_wait);
 	INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);