Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland/infiniband

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland/infiniband:
  IB/ehca: Wait for async events to finish before destroying QP
  IB/ipath: Fix SDMA error recovery in absence of link status change
  IB/ipath: Need to always request and handle PIO avail interrupts
  IB/ipath: Fix count of packets received by kernel
  IB/ipath: Return the correct opcode for RDMA WRITE with immediate
  IB/ipath: Fix bug that can leave sends disabled after freeze recovery
  IB/ipath: Only increment SSN if WQE is put on send queue
  IB/ipath: Only warn about prototype chip during init
  RDMA/cxgb3: Fix severe limit on userspace memory registration size
  RDMA/cxgb3: Don't add PBL memory to gen_pool in chunks

19 files changed, 458 insertions, 316 deletions

diff --git a/drivers/infiniband/hw/cxgb3/cxio_hal.c b/drivers/infiniband/hw/cxgb3/cxio_hal.c
index 5fd8506a865..ebf9d3043f8 100644
--- a/drivers/infiniband/hw/cxgb3/cxio_hal.c
+++ b/drivers/infiniband/hw/cxgb3/cxio_hal.c
@@ -588,7 +588,7 @@ static int cxio_hal_destroy_ctrl_qp(struct cxio_rdev *rdev_p)
  * caller aquires the ctrl_qp lock before the call
  */
 static int cxio_hal_ctrl_qp_write_mem(struct cxio_rdev *rdev_p, u32 addr,
-				      u32 len, void *data, int completion)
+				      u32 len, void *data)
 {
 	u32 i, nr_wqe, copy_len;
 	u8 *copy_data;
@@ -624,7 +624,7 @@ static int cxio_hal_ctrl_qp_write_mem(struct cxio_rdev *rdev_p, u32 addr,
 		flag = 0;
 		if (i == (nr_wqe - 1)) {
 			/* last WQE */
-			flag = completion ? T3_COMPLETION_FLAG : 0;
+			flag = T3_COMPLETION_FLAG;
 			if (len % 32)
 				utx_len = len / 32 + 1;
 			else
@@ -683,21 +683,20 @@ static int cxio_hal_ctrl_qp_write_mem(struct cxio_rdev *rdev_p, u32 addr,
 	return 0;
 }
 
-/* IN: stag key, pdid, perm, zbva, to, len, page_size, pbl, and pbl_size
- * OUT: stag index, actual pbl_size, pbl_addr allocated.
+/* IN: stag key, pdid, perm, zbva, to, len, page_size, pbl_size and pbl_addr
+ * OUT: stag index
  * TBD: shared memory region support
  */
 static int __cxio_tpt_op(struct cxio_rdev *rdev_p, u32 reset_tpt_entry,
 			 u32 *stag, u8 stag_state, u32 pdid,
 			 enum tpt_mem_type type, enum tpt_mem_perm perm,
-			 u32 zbva, u64 to, u32 len, u8 page_size, __be64 *pbl,
-			 u32 *pbl_size, u32 *pbl_addr)
+			 u32 zbva, u64 to, u32 len, u8 page_size,
+			 u32 pbl_size, u32 pbl_addr)
 {
 	int err;
 	struct tpt_entry tpt;
 	u32 stag_idx;
 	u32 wptr;
-	int rereg = (*stag != T3_STAG_UNSET);
 
 	stag_state = stag_state > 0;
 	stag_idx = (*stag) >> 8;
@@ -711,30 +710,8 @@ static int __cxio_tpt_op(struct cxio_rdev *rdev_p, u32 reset_tpt_entry,
 	PDBG("%s stag_state 0x%0x type 0x%0x pdid 0x%0x, stag_idx 0x%x\n",
 	     __func__, stag_state, type, pdid, stag_idx);
 
-	if (reset_tpt_entry)
-		cxio_hal_pblpool_free(rdev_p, *pbl_addr, *pbl_size << 3);
-	else if (!rereg) {
-		*pbl_addr = cxio_hal_pblpool_alloc(rdev_p, *pbl_size << 3);
-		if (!*pbl_addr) {
-			return -ENOMEM;
-		}
-	}
-
 	mutex_lock(&rdev_p->ctrl_qp.lock);
 
-	/* write PBL first if any - update pbl only if pbl list exist */
-	if (pbl) {
-
-		PDBG("%s *pdb_addr 0x%x, pbl_base 0x%x, pbl_size %d\n",
-		     __func__, *pbl_addr, rdev_p->rnic_info.pbl_base,
-		     *pbl_size);
-		err = cxio_hal_ctrl_qp_write_mem(rdev_p,
-				(*pbl_addr >> 5),
-				(*pbl_size << 3), pbl, 0);
-		if (err)
-			goto ret;
-	}
-
 	/* write TPT entry */
 	if (reset_tpt_entry)
 		memset(&tpt, 0, sizeof(tpt));
@@ -749,23 +726,23 @@ static int __cxio_tpt_op(struct cxio_rdev *rdev_p, u32 reset_tpt_entry,
 				V_TPT_ADDR_TYPE((zbva ? TPT_ZBTO : TPT_VATO)) |
 				V_TPT_PAGE_SIZE(page_size));
 		tpt.rsvd_pbl_addr = reset_tpt_entry ? 0 :
-				    cpu_to_be32(V_TPT_PBL_ADDR(PBL_OFF(rdev_p, *pbl_addr)>>3));
+				    cpu_to_be32(V_TPT_PBL_ADDR(PBL_OFF(rdev_p, pbl_addr)>>3));
 		tpt.len = cpu_to_be32(len);
 		tpt.va_hi = cpu_to_be32((u32) (to >> 32));
 		tpt.va_low_or_fbo = cpu_to_be32((u32) (to & 0xFFFFFFFFULL));
 		tpt.rsvd_bind_cnt_or_pstag = 0;
 		tpt.rsvd_pbl_size = reset_tpt_entry ? 0 :
-				  cpu_to_be32(V_TPT_PBL_SIZE((*pbl_size) >> 2));
+				  cpu_to_be32(V_TPT_PBL_SIZE(pbl_size >> 2));
 	}
 	err = cxio_hal_ctrl_qp_write_mem(rdev_p,
 				       stag_idx +
 				       (rdev_p->rnic_info.tpt_base >> 5),
-				       sizeof(tpt), &tpt, 1);
+				       sizeof(tpt), &tpt);
 
 	/* release the stag index to free pool */
 	if (reset_tpt_entry)
 		cxio_hal_put_stag(rdev_p->rscp, stag_idx);
-ret:
+
 	wptr = rdev_p->ctrl_qp.wptr;
 	mutex_unlock(&rdev_p->ctrl_qp.lock);
 	if (!err)
@@ -776,44 +753,67 @@ ret:
 	return err;
 }
 
+int cxio_write_pbl(struct cxio_rdev *rdev_p, __be64 *pbl,
+		   u32 pbl_addr, u32 pbl_size)
+{
+	u32 wptr;
+	int err;
+
+	PDBG("%s *pdb_addr 0x%x, pbl_base 0x%x, pbl_size %d\n",
+	     __func__, pbl_addr, rdev_p->rnic_info.pbl_base,
+	     pbl_size);
+
+	mutex_lock(&rdev_p->ctrl_qp.lock);
+	err = cxio_hal_ctrl_qp_write_mem(rdev_p, pbl_addr >> 5, pbl_size << 3,
+					 pbl);
+	wptr = rdev_p->ctrl_qp.wptr;
+	mutex_unlock(&rdev_p->ctrl_qp.lock);
+	if (err)
+		return err;
+
+	if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
+				     SEQ32_GE(rdev_p->ctrl_qp.rptr,
+					      wptr)))
+		return -ERESTARTSYS;
+
+	return 0;
+}
+
 int cxio_register_phys_mem(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid,
 			   enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
-			   u8 page_size, __be64 *pbl, u32 *pbl_size,
-			   u32 *pbl_addr)
+			   u8 page_size, u32 pbl_size, u32 pbl_addr)
 {
 	*stag = T3_STAG_UNSET;
 	return __cxio_tpt_op(rdev_p, 0, stag, 1, pdid, TPT_NON_SHARED_MR, perm,
-			     zbva, to, len, page_size, pbl, pbl_size, pbl_addr);
+			     zbva, to, len, page_size, pbl_size, pbl_addr);
 }
 
 int cxio_reregister_phys_mem(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid,
 			   enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
-			   u8 page_size, __be64 *pbl, u32 *pbl_size,
-			   u32 *pbl_addr)
+			   u8 page_size, u32 pbl_size, u32 pbl_addr)
 {
 	return __cxio_tpt_op(rdev_p, 0, stag, 1, pdid, TPT_NON_SHARED_MR, perm,
-			     zbva, to, len, page_size, pbl, pbl_size, pbl_addr);
+			     zbva, to, len, page_size, pbl_size, pbl_addr);
 }
 
 int cxio_dereg_mem(struct cxio_rdev *rdev_p, u32 stag, u32 pbl_size,
 		   u32 pbl_addr)
 {
-	return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0, NULL,
-			     &pbl_size, &pbl_addr);
+	return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0,
+			     pbl_size, pbl_addr);
 }
 
 int cxio_allocate_window(struct cxio_rdev *rdev_p, u32 * stag, u32 pdid)
 {
-	u32 pbl_size = 0;
 	*stag = T3_STAG_UNSET;
 	return __cxio_tpt_op(rdev_p, 0, stag, 0, pdid, TPT_MW, 0, 0, 0ULL, 0, 0,
-			     NULL, &pbl_size, NULL);
+			     0, 0);
 }
 
 int cxio_deallocate_window(struct cxio_rdev *rdev_p, u32 stag)
 {
-	return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0, NULL,
-			     NULL, NULL);
+	return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0,
+			     0, 0);
 }
 
 int cxio_rdma_init(struct cxio_rdev *rdev_p, struct t3_rdma_init_attr *attr)
diff --git a/drivers/infiniband/hw/cxgb3/cxio_hal.h b/drivers/infiniband/hw/cxgb3/cxio_hal.h
index 69ab08ebc68..6e128f6bab0 100644
--- a/drivers/infiniband/hw/cxgb3/cxio_hal.h
+++ b/drivers/infiniband/hw/cxgb3/cxio_hal.h
@@ -154,14 +154,14 @@ int cxio_create_qp(struct cxio_rdev *rdev, u32 kernel_domain, struct t3_wq *wq,
 int cxio_destroy_qp(struct cxio_rdev *rdev, struct t3_wq *wq,
 		    struct cxio_ucontext *uctx);
 int cxio_peek_cq(struct t3_wq *wr, struct t3_cq *cq, int opcode);
+int cxio_write_pbl(struct cxio_rdev *rdev_p, __be64 *pbl,
+		   u32 pbl_addr, u32 pbl_size);
 int cxio_register_phys_mem(struct cxio_rdev *rdev, u32 * stag, u32 pdid,
 			   enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
-			   u8 page_size, __be64 *pbl, u32 *pbl_size,
-			   u32 *pbl_addr);
+			   u8 page_size, u32 pbl_size, u32 pbl_addr);
 int cxio_reregister_phys_mem(struct cxio_rdev *rdev, u32 * stag, u32 pdid,
 			   enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
-			   u8 page_size, __be64 *pbl, u32 *pbl_size,
-			   u32 *pbl_addr);
+			   u8 page_size, u32 pbl_size, u32 pbl_addr);
 int cxio_dereg_mem(struct cxio_rdev *rdev, u32 stag, u32 pbl_size,
 		   u32 pbl_addr);
 int cxio_allocate_window(struct cxio_rdev *rdev, u32 * stag, u32 pdid);
diff --git a/drivers/infiniband/hw/cxgb3/cxio_resource.c b/drivers/infiniband/hw/cxgb3/cxio_resource.c
index 45ed4f25ef7..bd233c08765 100644
--- a/drivers/infiniband/hw/cxgb3/cxio_resource.c
+++ b/drivers/infiniband/hw/cxgb3/cxio_resource.c
@@ -250,7 +250,6 @@ void cxio_hal_destroy_resource(struct cxio_hal_resource *rscp)
  */
 
 #define MIN_PBL_SHIFT 8			/* 256B == min PBL size (32 entries) */
-#define PBL_CHUNK 2*1024*1024
 
 u32 cxio_hal_pblpool_alloc(struct cxio_rdev *rdev_p, int size)
 {
@@ -267,14 +266,35 @@ void cxio_hal_pblpool_free(struct cxio_rdev *rdev_p, u32 addr, int size)
 
 int cxio_hal_pblpool_create(struct cxio_rdev *rdev_p)
 {
-	unsigned long i;
+	unsigned pbl_start, pbl_chunk;
+
 	rdev_p->pbl_pool = gen_pool_create(MIN_PBL_SHIFT, -1);
-	if (rdev_p->pbl_pool)
-		for (i = rdev_p->rnic_info.pbl_base;
-		     i <= rdev_p->rnic_info.pbl_top - PBL_CHUNK + 1;
-		     i += PBL_CHUNK)
-			gen_pool_add(rdev_p->pbl_pool, i, PBL_CHUNK, -1);
-	return rdev_p->pbl_pool ? 0 : -ENOMEM;
+	if (!rdev_p->pbl_pool)
+		return -ENOMEM;
+
+	pbl_start = rdev_p->rnic_info.pbl_base;
+	pbl_chunk = rdev_p->rnic_info.pbl_top - pbl_start + 1;
+
+	while (pbl_start < rdev_p->rnic_info.pbl_top) {
+		pbl_chunk = min(rdev_p->rnic_info.pbl_top - pbl_start + 1,
+				pbl_chunk);
+		if (gen_pool_add(rdev_p->pbl_pool, pbl_start, pbl_chunk, -1)) {
+			PDBG("%s failed to add PBL chunk (%x/%x)\n",
+			     __func__, pbl_start, pbl_chunk);
+			if (pbl_chunk <= 1024 << MIN_PBL_SHIFT) {
+				printk(KERN_WARNING MOD "%s: Failed to add all PBL chunks (%x/%x)\n",
+				       __func__, pbl_start, rdev_p->rnic_info.pbl_top - pbl_start);
+				return 0;
+			}
+			pbl_chunk >>= 1;
+		} else {
+			PDBG("%s added PBL chunk (%x/%x)\n",
+			     __func__, pbl_start, pbl_chunk);
+			pbl_start += pbl_chunk;
+		}
+	}
+
+	return 0;
 }
 
 void cxio_hal_pblpool_destroy(struct cxio_rdev *rdev_p)
diff --git a/drivers/infiniband/hw/cxgb3/iwch_mem.c b/drivers/infiniband/hw/cxgb3/iwch_mem.c
index 58c3d61bcd1..ec49a5cbdeb 100644
--- a/drivers/infiniband/hw/cxgb3/iwch_mem.c
+++ b/drivers/infiniband/hw/cxgb3/iwch_mem.c
@@ -35,17 +35,26 @@
 #include <rdma/ib_verbs.h>
 
 #include "cxio_hal.h"
+#include "cxio_resource.h"
 #include "iwch.h"
 #include "iwch_provider.h"
 
-int iwch_register_mem(struct iwch_dev *rhp, struct iwch_pd *php,
-					struct iwch_mr *mhp,
-					int shift,
-					__be64 *page_list)
+static void iwch_finish_mem_reg(struct iwch_mr *mhp, u32 stag)
 {
-	u32 stag;
 	u32 mmid;
 
+	mhp->attr.state = 1;
+	mhp->attr.stag = stag;
+	mmid = stag >> 8;
+	mhp->ibmr.rkey = mhp->ibmr.lkey = stag;
+	insert_handle(mhp->rhp, &mhp->rhp->mmidr, mhp, mmid);
+	PDBG("%s mmid 0x%x mhp %p\n", __func__, mmid, mhp);
+}
+
+int iwch_register_mem(struct iwch_dev *rhp, struct iwch_pd *php,
+		      struct iwch_mr *mhp, int shift)
+{
+	u32 stag;
 
 	if (cxio_register_phys_mem(&rhp->rdev,
 				   &stag, mhp->attr.pdid,
@@ -53,28 +62,21 @@ int iwch_register_mem(struct iwch_dev *rhp, struct iwch_pd *php,
 				   mhp->attr.zbva,
 				   mhp->attr.va_fbo,
 				   mhp->attr.len,
-				   shift-12,
-				   page_list,
-				   &mhp->attr.pbl_size, &mhp->attr.pbl_addr))
+				   shift - 12,
+				   mhp->attr.pbl_size, mhp->attr.pbl_addr))
 		return -ENOMEM;
-	mhp->attr.state = 1;
-	mhp->attr.stag = stag;
-	mmid = stag >> 8;
-	mhp->ibmr.rkey = mhp->ibmr.lkey = stag;
-	insert_handle(rhp, &rhp->mmidr, mhp, mmid);
-	PDBG("%s mmid 0x%x mhp %p\n", __func__, mmid, mhp);
+
+	iwch_finish_mem_reg(mhp, stag);
+
 	return 0;
 }
 
 int iwch_reregister_mem(struct iwch_dev *rhp, struct iwch_pd *php,
 					struct iwch_mr *mhp,
 					int shift,
-					__be64 *page_list,
 					int npages)
 {
 	u32 stag;
-	u32 mmid;
-
 
 	/* We could support this... */
 	if (npages > mhp->attr.pbl_size)
@@ -87,19 +89,40 @@ int iwch_reregister_mem(struct iwch_dev *rhp, struct iwch_pd *php,
 				   mhp->attr.zbva,
 				   mhp->attr.va_fbo,
 				   mhp->attr.len,
-				   shift-12,
-				   page_list,
-				   &mhp->attr.pbl_size, &mhp->attr.pbl_addr))
+				   shift - 12,
+				   mhp->attr.pbl_size, mhp->attr.pbl_addr))
 		return -ENOMEM;
-	mhp->attr.state = 1;
-	mhp->attr.stag = stag;
-	mmid = stag >> 8;
-	mhp->ibmr.rkey = mhp->ibmr.lkey = stag;
-	insert_handle(rhp, &rhp->mmidr, mhp, mmid);
-	PDBG("%s mmid 0x%x mhp %p\n", __func__, mmid, mhp);
+
+	iwch_finish_mem_reg(mhp, stag);
+
+	return 0;
+}
+
+int iwch_alloc_pbl(struct iwch_mr *mhp, int npages)
+{
+	mhp->attr.pbl_addr = cxio_hal_pblpool_alloc(&mhp->rhp->rdev,
+						    npages << 3);
+
+	if (!mhp->attr.pbl_addr)
+		return -ENOMEM;
+
+	mhp->attr.pbl_size = npages;
+
 	return 0;
 }
 
+void iwch_free_pbl(struct iwch_mr *mhp)
+{
+	cxio_hal_pblpool_free(&mhp->rhp->rdev, mhp->attr.pbl_addr,
+			      mhp->attr.pbl_size << 3);
+}
+
+int iwch_write_pbl(struct iwch_mr *mhp, __be64 *pages, int npages, int offset)
+{
+	return cxio_write_pbl(&mhp->rhp->rdev, pages,
+			      mhp->attr.pbl_addr + (offset << 3), npages);
+}
+
 int build_phys_page_list(struct ib_phys_buf *buffer_list,
 					int num_phys_buf,
 					u64 *iova_start,
diff --git a/drivers/infiniband/hw/cxgb3/iwch_provider.c b/drivers/infiniband/hw/cxgb3/iwch_provider.c
index d07d3a377b5..8934178a23e 100644
--- a/drivers/infiniband/hw/cxgb3/iwch_provider.c
+++ b/drivers/infiniband/hw/cxgb3/iwch_provider.c
@@ -442,6 +442,7 @@ static int iwch_dereg_mr(struct ib_mr *ib_mr)
 	mmid = mhp->attr.stag >> 8;
 	cxio_dereg_mem(&rhp->rdev, mhp->attr.stag, mhp->attr.pbl_size,
 		       mhp->attr.pbl_addr);
+	iwch_free_pbl(mhp);
 	remove_handle(rhp, &rhp->mmidr, mmid);
 	if (mhp->kva)
 		kfree((void *) (unsigned long) mhp->kva);
@@ -475,6 +476,8 @@ static struct ib_mr *iwch_register_phys_mem(struct ib_pd *pd,
 	if (!mhp)
 		return ERR_PTR(-ENOMEM);
 
+	mhp->rhp = rhp;
+
 	/* First check that we have enough alignment */
 	if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK)) {
 		ret = -EINVAL;
@@ -492,7 +495,17 @@ static struct ib_mr *iwch_register_phys_mem(struct ib_pd *pd,
 	if (ret)
 		goto err;
 
-	mhp->rhp = rhp;
+	ret = iwch_alloc_pbl(mhp, npages);
+	if (ret) {
+		kfree(page_list);
+		goto err_pbl;
+	}
+
+	ret = iwch_write_pbl(mhp, page_list, npages, 0);
+	kfree(page_list);
+	if (ret)
+		goto err_pbl;
+
 	mhp->attr.pdid = php->pdid;
 	mhp->attr.zbva = 0;
 
@@ -502,12 +515,15 @@ static struct ib_mr *iwch_register_phys_mem(struct ib_pd *pd,
 
 	mhp->attr.len = (u32) total_size;
 	mhp->attr.pbl_size = npages;
-	ret = iwch_register_mem(rhp, php, mhp, shift, page_list);
-	kfree(page_list);
-	if (ret) {
-		goto err;
-	}
+	ret = iwch_register_mem(rhp, php, mhp, shift);
+	if (ret)
+		goto err_pbl;
+
 	return &mhp->ibmr;
+
+err_pbl:
+	iwch_free_pbl(mhp);
+
 err:
 	kfree(mhp);
 	return ERR_PTR(ret);
@@ -560,7 +576,7 @@ static int iwch_reregister_phys_mem(struct ib_mr *mr,
 			return ret;
 	}
 
-	ret = iwch_reregister_mem(rhp, php, &mh, shift, page_list, npages);
+	ret = iwch_reregister_mem(rhp, php, &mh, shift, npages);
 	kfree(page_list);
 	if (ret) {
 		return ret;
@@ -602,6 +618,8 @@ static struct ib_mr *iwch_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 	if (!mhp)
 		return ERR_PTR(-ENOMEM);
 
+	mhp->rhp = rhp;
+
 	mhp->umem = ib_umem_get(pd->uobject->context, start, length, acc, 0);
 	if (IS_ERR(mhp->umem)) {
 		err = PTR_ERR(mhp->umem);
@@ -615,10 +633,14 @@ static struct ib_mr *iwch_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 	list_for_each_entry(chunk, &mhp->umem->chunk_list, list)
 		n += chunk->nents;
 
-	pages = kmalloc(n * sizeof(u64), GFP_KERNEL);
+	err = iwch_alloc_pbl(mhp, n);
+	if (err)
+		goto err;
+
+	pages = (__be64 *) __get_free_page(GFP_KERNEL);
 	if (!pages) {
 		err = -ENOMEM;
-		goto err;
+		goto err_pbl;
 	}
 
 	i = n = 0;
@@ -630,25 +652,38 @@ static struct ib_mr *iwch_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 				pages[i++] = cpu_to_be64(sg_dma_address(
 					&chunk->page_list[j]) +
 					mhp->umem->page_size * k);
+				if (i == PAGE_SIZE / sizeof *pages) {
+					err = iwch_write_pbl(mhp, pages, i, n);
+					if (err)
+						goto pbl_done;
+					n += i;
+					i = 0;
+				}
 			}
 		}
 
-	mhp->rhp = rhp;
+	if (i)
+		err = iwch_write_pbl(mhp, pages, i, n);
+
+pbl_done:
+	free_page((unsigned long) pages);
+	if (err)
+		goto err_pbl;
+
 	mhp->attr.pdid = php->pdid;
 	mhp->attr.zbva = 0;
 	mhp->attr.perms = iwch_ib_to_tpt_access(acc);
 	mhp->attr.va_fbo = virt;
 	mhp->attr.page_size = shift - 12;
 	mhp->attr.len = (u32) length;
-	mhp->attr.pbl_size = i;
-	err = iwch_register_mem(rhp, php, mhp, shift, pages);
-	kfree(pages);
+
+	err = iwch_register_mem(rhp, php, mhp, shift);
 	if (err)
-		goto err;
+		goto err_pbl;
 
 	if (udata && !t3a_device(rhp)) {
 		uresp.pbl_addr = (mhp->attr.pbl_addr -
-	                         rhp->rdev.rnic_info.pbl_base) >> 3;
+				 rhp->rdev.rnic_info.pbl_base) >> 3;
 		PDBG("%s user resp pbl_addr 0x%x\n", __func__,
 		     uresp.pbl_addr);
 
@@ -661,6 +696,9 @@ static struct ib_mr *iwch_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 
 	return &mhp->ibmr;
 
+err_pbl:
+	iwch_free_pbl(mhp);
+
 err:
 	ib_umem_release(mhp->umem);
 	kfree(mhp);
diff --git a/drivers/infiniband/hw/cxgb3/iwch_provider.h b/drivers/infiniband/hw/cxgb3/iwch_provider.h
index db5100d27ca..836163fc542 100644
--- a/drivers/infiniband/hw/cxgb3/iwch_provider.h
+++ b/drivers/infiniband/hw/cxgb3/iwch_provider.h
@@ -340,14 +340,14 @@ int iwch_quiesce_qps(struct iwch_cq *chp);
 int iwch_resume_qps(struct iwch_cq *chp);
 void stop_read_rep_timer(struct iwch_qp *qhp);
 int iwch_register_mem(struct iwch_dev *rhp, struct iwch_pd *php,
-					struct iwch_mr *mhp,
-					int shift,
-					__be64 *page_list);
+		      struct iwch_mr *mhp, int shift);
 int iwch_reregister_mem(struct iwch_dev *rhp, struct iwch_pd *php,
 					struct iwch_mr *mhp,
 					int shift,
-					__be64 *page_list,
 					int npages);
+int iwch_alloc_pbl(struct iwch_mr *mhp, int npages);
+void iwch_free_pbl(struct iwch_mr *mhp);
+int iwch_write_pbl(struct iwch_mr *mhp, __be64 *pages, int npages, int offset);
 int build_phys_page_list(struct ib_phys_buf *buffer_list,
 					int num_phys_buf,
 					u64 *iova_start,
diff --git a/drivers/infiniband/hw/ehca/ehca_classes.h b/drivers/infiniband/hw/ehca/ehca_classes.h
index 00bab60f6de..1e9e99a1393 100644
--- a/drivers/infiniband/hw/ehca/ehca_classes.h
+++ b/drivers/infiniband/hw/ehca/ehca_classes.h
@@ -192,6 +192,8 @@ struct ehca_qp {
 	int mtu_shift;
 	u32 message_count;
 	u32 packet_count;
+	atomic_t nr_events; /* events seen */
+	wait_queue_head_t wait_completion;
 };
 
 #define IS_SRQ(qp) (qp->ext_type == EQPT_SRQ)
diff --git a/drivers/infiniband/hw/ehca/ehca_irq.c b/drivers/infiniband/hw/ehca/ehca_irq.c
index ca5eb0cb628..ce1ab0571be 100644
--- a/drivers/infiniband/hw/ehca/ehca_irq.c
+++ b/drivers/infiniband/hw/ehca/ehca_irq.c
@@ -204,6 +204,8 @@ static void qp_event_callback(struct ehca_shca *shca, u64 eqe,
 
 	read_lock(&ehca_qp_idr_lock);
 	qp = idr_find(&ehca_qp_idr, token);
+	if (qp)
+		atomic_inc(&qp->nr_events);
 	read_unlock(&ehca_qp_idr_lock);
 
 	if (!qp)
@@ -223,6 +225,8 @@ static void qp_event_callback(struct ehca_shca *shca, u64 eqe,
 	if (fatal && qp->ext_type == EQPT_SRQBASE)
 		dispatch_qp_event(shca, qp, IB_EVENT_QP_LAST_WQE_REACHED);
 
+	if (atomic_dec_and_test(&qp->nr_events))
+		wake_up(&qp->wait_completion);
 	return;
 }
 
diff --git a/drivers/infiniband/hw/ehca/ehca_qp.c b/drivers/infiniband/hw/ehca/ehca_qp.c
index 18fba92fa7a..3f59587338e 100644
--- a/drivers/infiniband/hw/ehca/ehca_qp.c
+++ b/drivers/infiniband/hw/ehca/ehca_qp.c
@@ -566,6 +566,8 @@ static struct ehca_qp *internal_create_qp(
 		return ERR_PTR(-ENOMEM);
 	}
 
+	atomic_set(&my_qp->nr_events, 0);
+	init_waitqueue_head(&my_qp->wait_completion);
 	spin_lock_init(&my_qp->spinlock_s);
 	spin_lock_init(&my_qp->spinlock_r);
 	my_qp->qp_type = qp_type;
@@ -1934,6 +1936,9 @@ static int internal_destroy_qp(struct ib_device *dev, struct ehca_qp *my_qp,
 	idr_remove(&ehca_qp_idr, my_qp->token);
 	write_unlock_irqrestore(&ehca_qp_idr_lock, flags);
 
+	/* now wait until all pending events have completed */
+	wait_event(my_qp->wait_completion, !atomic_read(&my_qp->nr_events));
+
 	h_ret = hipz_h_destroy_qp(shca->ipz_hca_handle, my_qp);
 	if (h_ret != H_SUCCESS) {
 		ehca_err(dev, "hipz_h_destroy_qp() failed h_ret=%li "
diff --git a/drivers/infiniband/hw/ipath/ipath_driver.c b/drivers/infiniband/hw/ipath/ipath_driver.c
index acf30c06a0c..ce7b7c34360 100644
--- a/drivers/infiniband/hw/ipath/ipath_driver.c
+++ b/drivers/infiniband/hw/ipath/ipath_driver.c
@@ -1197,7 +1197,7 @@ void ipath_kreceive(struct ipath_portdata *pd)
 	}
 
 reloop:
-	for (last = 0, i = 1; !last; i++) {
+	for (last = 0, i = 1; !last; i += !last) {
 		hdr = dd->ipath_f_get_msgheader(dd, rhf_addr);
 		eflags = ipath_hdrget_err_flags(rhf_addr);
 		etype = ipath_hdrget_rcv_type(rhf_addr);
@@ -1428,6 +1428,40 @@ static void ipath_update_pio_bufs(struct ipath_devdata *dd)
 	spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
 }
 
+/*
+ * used to force update of pioavailshadow if we can't get a pio buffer.
+ * Needed primarily due to exitting freeze mode after recovering
+ * from errors.  Done lazily, because it's safer (known to not
+ * be writing pio buffers).
+ */
+static void ipath_reset_availshadow(struct ipath_devdata *dd)
+{
+	int i, im;
+	unsigned long flags;
+
+	spin_lock_irqsave(&ipath_pioavail_lock, flags);
+	for (i = 0; i < dd->ipath_pioavregs; i++) {
+		u64 val, oldval;
+		/* deal with 6110 chip bug on high register #s */
+		im = (i > 3 && (dd->ipath_flags & IPATH_SWAP_PIOBUFS)) ?
+			i ^ 1 : i;
+		val = le64_to_cpu(dd->ipath_pioavailregs_dma[im]);
+		/*
+		 * busy out the buffers not in the kernel avail list,
+		 * without changing the generation bits.
+		 */
+		oldval = dd->ipath_pioavailshadow[i];
+		dd->ipath_pioavailshadow[i] = val |
+			((~dd->ipath_pioavailkernel[i] <<
+			INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT) &
+			0xaaaaaaaaaaaaaaaaULL); /* All BUSY bits in qword */
+		if (oldval != dd->ipath_pioavailshadow[i])
+			ipath_dbg("shadow[%d] was %Lx, now %lx\n",
+				i, oldval, dd->ipath_pioavailshadow[i]);
+	}
+	spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
+}
+
 /**
  * ipath_setrcvhdrsize - set the receive header size
  * @dd: the infinipath device
@@ -1482,9 +1516,12 @@ static noinline void no_pio_bufs(struct ipath_devdata *dd)
 	 */
 	ipath_stats.sps_nopiobufs++;
 	if (!(++dd->ipath_consec_nopiobuf % 100000)) {
-		ipath_dbg("%u pio sends with no bufavail; dmacopy: "
-			"%llx %llx %llx %llx; shadow:  %lx %lx %lx %lx\n",
+		ipath_force_pio_avail_update(dd); /* at start */
+		ipath_dbg("%u tries no piobufavail ts%lx; dmacopy: "
+			"%llx %llx %llx %llx\n"
+			"ipath  shadow:  %lx %lx %lx %lx\n",
 			dd->ipath_consec_nopiobuf,
+			(unsigned long)get_cycles(),
 			(unsigned long long) le64_to_cpu(dma[0]),
 			(unsigned long long) le64_to_cpu(dma[1]),
 			(unsigned long long) le64_to_cpu(dma[2]),
@@ -1496,14 +1533,17 @@ static noinline void no_pio_bufs(struct ipath_devdata *dd)
 		 */
 		if ((dd->ipath_piobcnt2k + dd->ipath_piobcnt4k) >
 		    (sizeof(shadow[0]) * 4 * 4))
-			ipath_dbg("2nd group: dmacopy: %llx %llx "
-				  "%llx %llx; shadow: %lx %lx %lx %lx\n",
+			ipath_dbg("2nd group: dmacopy: "
+				  "%llx %llx %llx %llx\n"
+				  "ipath  shadow:  %lx %lx %lx %lx\n",
 				  (unsigned long long)le64_to_cpu(dma[4]),
 				  (unsigned long long)le64_to_cpu(dma[5]),
 				  (unsigned long long)le64_to_cpu(dma[6]),
 				  (unsigned long long)le64_to_cpu(dma[7]),
-				  shadow[4], shadow[5], shadow[6],
-				  shadow[7]);
+				  shadow[4], shadow[5], shadow[6], shadow[7]);
+
+		/* at end, so update likely happened */
+		ipath_reset_availshadow(dd);
 	}
 }
 
@@ -1652,19 +1692,46 @@ void ipath_chg_pioavailkernel(struct ipath_devdata *dd, unsigned start,
 			      unsigned len, int avail)
 {
 	unsigned long flags;
-	unsigned end;
+	unsigned end, cnt = 0, next;
 
 	/* There are two bits per send buffer (busy and generation) */
 	start *= 2;
-	len *= 2;
-	end = start + len;
+	end = start + len * 2;
 
-	/* Set or clear the generation bits. */
 	spin_lock_irqsave(&ipath_pioavail_lock, flags);
+	/* Set or clear the busy bit in the shadow. */
 	while (start < end) {
 		if (avail) {
-			__clear_bit(start + INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT,
-				dd->ipath_pioavailshadow);
+			unsigned long dma;
+			int i, im;
+			/*
+			 * the BUSY bit will never be set, because we disarm
+			 * the user buffers before we hand them back to the
+			 * kernel.  We do have to make sure the generation
+			 * bit is set correctly in shadow, since it could
+			 * have changed many times while allocated to user.
+			 * We can't use the bitmap functions on the full
+			 * dma array because it is always little-endian, so
+			 * we have to flip to host-order first.
+			 * BITS_PER_LONG is slightly wrong, since it's
+			 * always 64 bits per register in chip...
+			 * We only work on 64 bit kernels, so that's OK.
+			 */
+			/* deal with 6110 chip bug on high register #s */
+			i = start / BITS_PER_LONG;
+			im = (i > 3 && (dd->ipath_flags & IPATH_SWAP_PIOBUFS)) ?
+				i ^ 1 : i;
+			__clear_bit(INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT
+				+ start, dd->ipath_pioavailshadow);
+			dma = (unsigned long) le64_to_cpu(
+				dd->ipath_pioavailregs_dma[im]);
+			if (test_bit((INFINIPATH_SENDPIOAVAIL_CHECK_SHIFT
+				+ start) % BITS_PER_LONG, &dma))
+				__set_bit(INFINIPATH_SENDPIOAVAIL_CHECK_SHIFT
+					+ start, dd->ipath_pioavailshadow);
+			else
+				__clear_bit(INFINIPATH_SENDPIOAVAIL_CHECK_SHIFT
+					+ start, dd->ipath_pioavailshadow);
 			__set_bit(start, dd->ipath_pioavailkernel);
 		} else {
 			__set_bit(start + INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT,
@@ -1673,7 +1740,44 @@ void ipath_chg_pioavailkernel(struct ipath_devdata *dd, unsigned start,
 		}
 		start += 2;
 	}
+
+	if (dd->ipath_pioupd_thresh) {
+		end = 2 * (dd->ipath_piobcnt2k + dd->ipath_piobcnt4k);
+		next = find_first_bit(dd->ipath_pioavailkernel, end);
+		while (next < end) {
+			cnt++;
+			next = find_next_bit(dd->ipath_pioavailkernel, end,
+					next + 1);
+		}
+	}
 	spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
+
+	/*
+	 * When moving buffers from kernel to user, if number assigned to
+	 * the user is less than the pio update threshold, and threshold
+	 * is supported (cnt was computed > 0), drop the update threshold
+	 * so we update at least once per allocated number of buffers.
+	 * In any case, if the kernel buffers are less than the threshold,
+	 * drop the threshold.  We don't bother increasing it, having once
+	 * decreased it, since it would typically just cycle back and forth.
+	 * If we don't decrease below buffers in use, we can wait a long
+	 * time for an update, until some other context uses PIO buffers.
+	 */
+	if (!avail && len < cnt)
+		cnt = len;
+	if (cnt < dd->ipath_pioupd_thresh) {
+		dd->ipath_pioupd_thresh = cnt;
+		ipath_dbg("Decreased pio update threshold to %u\n",
+			dd->ipath_pioupd_thresh);
+		spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
+		dd->ipath_sendctrl &= ~(INFINIPATH_S_UPDTHRESH_MASK
+			<< INFINIPATH_S_UPDTHRESH_SHIFT);
+		dd->ipath_sendctrl |= dd->ipath_pioupd_thresh
+			<< INFINIPATH_S_UPDTHRESH_SHIFT;
+		ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
+			dd->ipath_sendctrl);
+		spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
+	}
 }
 
 /**
@@ -1794,8 +1898,8 @@ void ipath_cancel_sends(struct ipath_devdata *dd, int restore_sendctrl)
 
 		spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
 		skip_cancel =
-			!test_bit(IPATH_SDMA_DISABLED, statp) &&
-			test_and_set_bit(IPATH_SDMA_ABORTING, statp);
+			test_and_set_bit(IPATH_SDMA_ABORTING, statp)
+			&& !test_bit(IPATH_SDMA_DISABLED, statp);
 		spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
 		if (skip_cancel)
 			goto bail;
@@ -1826,6 +1930,9 @@ void ipath_cancel_sends(struct ipath_devdata *dd, int restore_sendctrl)
 	ipath_disarm_piobufs(dd, 0,
 		dd->ipath_piobcnt2k + dd->ipath_piobcnt4k);
 
+	if (dd->ipath_flags & IPATH_HAS_SEND_DMA)
+		set_bit(IPATH_SDMA_DISARMED, &dd->ipath_sdma_status);
+
 	if (restore_sendctrl) {
 		/* else done by caller later if needed */
 		spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
@@ -1845,7 +1952,6 @@ void ipath_cancel_sends(struct ipath_devdata *dd, int restore_sendctrl)
 		/* only wait so long for intr */
 		dd->ipath_sdma_abort_intr_timeout = jiffies + HZ;
 		dd->ipath_sdma_reset_wait = 200;
-		__set_bit(IPATH_SDMA_DISARMED, &dd->ipath_sdma_status);
 		if (!test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
 			tasklet_hi_schedule(&dd->ipath_sdma_abort_task);
 		spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
diff --git a/drivers/infiniband/hw/ipath/ipath_file_ops.c b/drivers/infiniband/hw/ipath/ipath_file_ops.c
index 8b1752202e7..3295177c937 100644
--- a/drivers/infiniband/hw/ipath/ipath_file_ops.c
+++ b/