1 files changed, 26 insertions, 233 deletions

diff --git a/drivers/infiniband/hw/hfi1/verbs.c b/drivers/infiniband/hw/hfi1/verbs.c
index 13374c727b14..48e11e510358 100644
--- a/drivers/infiniband/hw/hfi1/verbs.c
+++ b/drivers/infiniband/hw/hfi1/verbs.c
@@ -129,8 +129,6 @@ unsigned short piothreshold = 256;
 module_param(piothreshold, ushort, S_IRUGO);
 MODULE_PARM_DESC(piothreshold, "size used to determine sdma vs. pio");
 
-#define COPY_CACHELESS 1
-#define COPY_ADAPTIVE  2
 static unsigned int sge_copy_mode;
 module_param(sge_copy_mode, uint, S_IRUGO);
 MODULE_PARM_DESC(sge_copy_mode,
@@ -151,159 +149,13 @@ static int pio_wait(struct rvt_qp *qp,
 /* 16B trailing buffer */
 static const u8 trail_buf[MAX_16B_PADDING];
 
-static uint wss_threshold;
+static uint wss_threshold = 80;
 module_param(wss_threshold, uint, S_IRUGO);
 MODULE_PARM_DESC(wss_threshold, "Percentage (1-100) of LLC to use as a threshold for a cacheless copy");
 static uint wss_clean_period = 256;
 module_param(wss_clean_period, uint, S_IRUGO);
 MODULE_PARM_DESC(wss_clean_period, "Count of verbs copies before an entry in the page copy table is cleaned");
 
-/* memory working set size */
-struct hfi1_wss {
-	unsigned long *entries;
-	atomic_t total_count;
-	atomic_t clean_counter;
-	atomic_t clean_entry;
-
-	int threshold;
-	int num_entries;
-	long pages_mask;
-};
-
-static struct hfi1_wss wss;
-
-int hfi1_wss_init(void)
-{
-	long llc_size;
-	long llc_bits;
-	long table_size;
-	long table_bits;
-
-	/* check for a valid percent range - default to 80 if none or invalid */
-	if (wss_threshold < 1 || wss_threshold > 100)
-		wss_threshold = 80;
-	/* reject a wildly large period */
-	if (wss_clean_period > 1000000)
-		wss_clean_period = 256;
-	/* reject a zero period */
-	if (wss_clean_period == 0)
-		wss_clean_period = 1;
-
-	/*
-	 * Calculate the table size - the next power of 2 larger than the
-	 * LLC size.  LLC size is in KiB.
-	 */
-	llc_size = wss_llc_size() * 1024;
-	table_size = roundup_pow_of_two(llc_size);
-
-	/* one bit per page in rounded up table */
-	llc_bits = llc_size / PAGE_SIZE;
-	table_bits = table_size / PAGE_SIZE;
-	wss.pages_mask = table_bits - 1;
-	wss.num_entries = table_bits / BITS_PER_LONG;
-
-	wss.threshold = (llc_bits * wss_threshold) / 100;
-	if (wss.threshold == 0)
-		wss.threshold = 1;
-
-	atomic_set(&wss.clean_counter, wss_clean_period);
-
-	wss.entries = kcalloc(wss.num_entries, sizeof(*wss.entries),
-			      GFP_KERNEL);
-	if (!wss.entries) {
-		hfi1_wss_exit();
-		return -ENOMEM;
-	}
-
-	return 0;
-}
-
-void hfi1_wss_exit(void)
-{
-	/* coded to handle partially initialized and repeat callers */
-	kfree(wss.entries);
-	wss.entries = NULL;
-}
-
-/*
- * Advance the clean counter.  When the clean period has expired,
- * clean an entry.
- *
- * This is implemented in atomics to avoid locking.  Because multiple
- * variables are involved, it can be racy which can lead to slightly
- * inaccurate information.  Since this is only a heuristic, this is
- * OK.  Any innaccuracies will clean themselves out as the counter
- * advances.  That said, it is unlikely the entry clean operation will
- * race - the next possible racer will not start until the next clean
- * period.
- *
- * The clean counter is implemented as a decrement to zero.  When zero
- * is reached an entry is cleaned.
- */
-static void wss_advance_clean_counter(void)
-{
-	int entry;
-	int weight;
-	unsigned long bits;
-
-	/* become the cleaner if we decrement the counter to zero */
-	if (atomic_dec_and_test(&wss.clean_counter)) {
-		/*
-		 * Set, not add, the clean period.  This avoids an issue
-		 * where the counter could decrement below the clean period.
-		 * Doing a set can result in lost decrements, slowing the
-		 * clean advance.  Since this a heuristic, this possible
-		 * slowdown is OK.
-		 *
-		 * An alternative is to loop, advancing the counter by a
-		 * clean period until the result is > 0. However, this could
-		 * lead to several threads keeping another in the clean loop.
-		 * This could be mitigated by limiting the number of times
-		 * we stay in the loop.
-		 */
-		atomic_set(&wss.clean_counter, wss_clean_period);
-
-		/*
-		 * Uniquely grab the entry to clean and move to next.
-		 * The current entry is always the lower bits of
-		 * wss.clean_entry.  The table size, wss.num_entries,
-		 * is always a power-of-2.
-		 */
-		entry = (atomic_inc_return(&wss.clean_entry) - 1)
-			& (wss.num_entries - 1);
-
-		/* clear the entry and count the bits */
-		bits = xchg(&wss.entries[entry], 0);
-		weight = hweight64((u64)bits);
-		/* only adjust the contended total count if needed */
-		if (weight)
-			atomic_sub(weight, &wss.total_count);
-	}
-}
-
-/*
- * Insert the given address into the working set array.
- */
-static void wss_insert(void *address)
-{
-	u32 page = ((unsigned long)address >> PAGE_SHIFT) & wss.pages_mask;
-	u32 entry = page / BITS_PER_LONG; /* assumes this ends up a shift */
-	u32 nr = page & (BITS_PER_LONG - 1);
-
-	if (!test_and_set_bit(nr, &wss.entries[entry]))
-		atomic_inc(&wss.total_count);
-
-	wss_advance_clean_counter();
-}
-
-/*
- * Is the working set larger than the threshold?
- */
-static inline bool wss_exceeds_threshold(void)
-{
-	return atomic_read(&wss.total_count) >= wss.threshold;
-}
-
 /*
  * Translate ib_wr_opcode into ib_wc_opcode.
  */
@@ -438,79 +290,6 @@ static const u32 pio_opmask[BIT(3)] = {
  */
 __be64 ib_hfi1_sys_image_guid;
 
-/**
- * hfi1_copy_sge - copy data to SGE memory
- * @ss: the SGE state
- * @data: the data to copy
- * @length: the length of the data
- * @release: boolean to release MR
- * @copy_last: do a separate copy of the last 8 bytes
- */
-void hfi1_copy_sge(
-	struct rvt_sge_state *ss,
-	void *data, u32 length,
-	bool release,
-	bool copy_last)
-{
-	struct rvt_sge *sge = &ss->sge;
-	int i;
-	bool in_last = false;
-	bool cacheless_copy = false;
-
-	if (sge_copy_mode == COPY_CACHELESS) {
-		cacheless_copy = length >= PAGE_SIZE;
-	} else if (sge_copy_mode == COPY_ADAPTIVE) {
-		if (length >= PAGE_SIZE) {
-			/*
-			 * NOTE: this *assumes*:
-			 * o The first vaddr is the dest.
-			 * o If multiple pages, then vaddr is sequential.
-			 */
-			wss_insert(sge->vaddr);
-			if (length >= (2 * PAGE_SIZE))
-				wss_insert(sge->vaddr + PAGE_SIZE);
-
-			cacheless_copy = wss_exceeds_threshold();
-		} else {
-			wss_advance_clean_counter();
-		}
-	}
-	if (copy_last) {
-		if (length > 8) {
-			length -= 8;
-		} else {
-			copy_last = false;
-			in_last = true;
-		}
-	}
-
-again:
-	while (length) {
-		u32 len = rvt_get_sge_length(sge, length);
-
-		WARN_ON_ONCE(len == 0);
-		if (unlikely(in_last)) {
-			/* enforce byte transfer ordering */
-			for (i = 0; i < len; i++)
-				((u8 *)sge->vaddr)[i] = ((u8 *)data)[i];
-		} else if (cacheless_copy) {
-			cacheless_memcpy(sge->vaddr, data, len);
-		} else {
-			memcpy(sge->vaddr, data, len);
-		}
-		rvt_update_sge(ss, len, release);
-		data += len;
-		length -= len;
-	}
-
-	if (copy_last) {
-		copy_last = false;
-		in_last = true;
-		length = 8;
-		goto again;
-	}
-}
-
 /*
  * Make sure the QP is ready and able to accept the given opcode.
  */
@@ -713,7 +492,7 @@ static void verbs_sdma_complete(
 
 	spin_lock(&qp->s_lock);
 	if (tx->wqe) {
-		hfi1_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
+		rvt_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
 		struct hfi1_opa_header *hdr;
 
@@ -737,7 +516,7 @@ static int wait_kmem(struct hfi1_ibdev *dev,
 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
 		write_seqlock(&dev->iowait_lock);
 		list_add_tail(&ps->s_txreq->txreq.list,
-			      &priv->s_iowait.tx_head);
+			      &ps->wait->tx_head);
 		if (list_empty(&priv->s_iowait.list)) {
 			if (list_empty(&dev->memwait))
 				mod_timer(&dev->mem_timer, jiffies + 1);
@@ -748,7 +527,7 @@ static int wait_kmem(struct hfi1_ibdev *dev,
 			rvt_get_qp(qp);
 		}
 		write_sequnlock(&dev->iowait_lock);
-		qp->s_flags &= ~RVT_S_BUSY;
+		hfi1_qp_unbusy(qp, ps->wait);
 		ret = -EBUSY;
 	}
 	spin_unlock_irqrestore(&qp->s_lock, flags);
@@ -950,8 +729,7 @@ int hfi1_verbs_send_dma(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
 		if (unlikely(ret))
 			goto bail_build;
 	}
-	ret =  sdma_send_txreq(tx->sde, &priv->s_iowait, &tx->txreq,
-			       ps->pkts_sent);
+	ret =  sdma_send_txreq(tx->sde, ps->wait, &tx->txreq, ps->pkts_sent);
 	if (unlikely(ret < 0)) {
 		if (ret == -ECOMM)
 			goto bail_ecomm;
@@ -1001,7 +779,7 @@ static int pio_wait(struct rvt_qp *qp,
 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
 		write_seqlock(&dev->iowait_lock);
 		list_add_tail(&ps->s_txreq->txreq.list,
-			      &priv->s_iowait.tx_head);
+			      &ps->wait->tx_head);
 		if (list_empty(&priv->s_iowait.list)) {
 			struct hfi1_ibdev *dev = &dd->verbs_dev;
 			int was_empty;
@@ -1020,7 +798,7 @@ static int pio_wait(struct rvt_qp *qp,
 				hfi1_sc_wantpiobuf_intr(sc, 1);
 		}
 		write_sequnlock(&dev->iowait_lock);
-		qp->s_flags &= ~RVT_S_BUSY;
+		hfi1_qp_unbusy(qp, ps->wait);
 		ret = -EBUSY;
 	}
 	spin_unlock_irqrestore(&qp->s_lock, flags);
@@ -1160,7 +938,7 @@ int hfi1_verbs_send_pio(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
 pio_bail:
 	if (qp->s_wqe) {
 		spin_lock_irqsave(&qp->s_lock, flags);
-		hfi1_send_complete(qp, qp->s_wqe, wc_status);
+		rvt_send_complete(qp, qp->s_wqe, wc_status);
 		spin_unlock_irqrestore(&qp->s_lock, flags);
 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
 		spin_lock_irqsave(&qp->s_lock, flags);
@@ -1367,7 +1145,7 @@ int hfi1_verbs_send(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 			hfi1_cdbg(PIO, "%s() Failed. Completing with err",
 				  __func__);
 			spin_lock_irqsave(&qp->s_lock, flags);
-			hfi1_send_complete(qp, qp->s_wqe, IB_WC_GENERAL_ERR);
+			rvt_send_complete(qp, qp->s_wqe, IB_WC_GENERAL_ERR);
 			spin_unlock_irqrestore(&qp->s_lock, flags);
 		}
 		return -EINVAL;
@@ -1582,6 +1360,7 @@ static int hfi1_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr)
 	struct hfi1_pportdata *ppd;
 	struct hfi1_devdata *dd;
 	u8 sc5;
+	u8 sl;
 
 	if (hfi1_check_mcast(rdma_ah_get_dlid(ah_attr)) &&
 	    !(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH))
@@ -1590,8 +1369,13 @@ static int hfi1_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr)
 	/* test the mapping for validity */
 	ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
 	ppd = ppd_from_ibp(ibp);
-	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
 	dd = dd_from_ppd(ppd);
+
+	sl = rdma_ah_get_sl(ah_attr);
+	if (sl >= ARRAY_SIZE(ibp->sl_to_sc))
+		return -EINVAL;
+
+	sc5 = ibp->sl_to_sc[sl];
 	if (sc_to_vlt(dd, sc5) > num_vls && sc_to_vlt(dd, sc5) != 0xf)
 		return -EINVAL;
 	return 0;
@@ -1937,7 +1721,7 @@ int hfi1_register_ib_device(struct hfi1_devdata *dd)
 	dd->verbs_dev.rdi.driver_f.check_modify_qp = hfi1_check_modify_qp;
 	dd->verbs_dev.rdi.driver_f.modify_qp = hfi1_modify_qp;
 	dd->verbs_dev.rdi.driver_f.notify_restart_rc = hfi1_restart_rc;
-	dd->verbs_dev.rdi.driver_f.check_send_wqe = hfi1_check_send_wqe;
+	dd->verbs_dev.rdi.driver_f.setup_wqe = hfi1_setup_wqe;
 	dd->verbs_dev.rdi.driver_f.comp_vect_cpu_lookup =
 						hfi1_comp_vect_mappings_lookup;
 
@@ -1950,10 +1734,16 @@ int hfi1_register_ib_device(struct hfi1_devdata *dd)
 	dd->verbs_dev.rdi.dparms.lkey_table_size = hfi1_lkey_table_size;
 	dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
 	dd->verbs_dev.rdi.dparms.npkeys = hfi1_get_npkeys(dd);
+	dd->verbs_dev.rdi.dparms.sge_copy_mode = sge_copy_mode;
+	dd->verbs_dev.rdi.dparms.wss_threshold = wss_threshold;
+	dd->verbs_dev.rdi.dparms.wss_clean_period = wss_clean_period;
 
 	/* post send table */
 	dd->verbs_dev.rdi.post_parms = hfi1_post_parms;
 
+	/* opcode translation table */
+	dd->verbs_dev.rdi.wc_opcode = ib_hfi1_wc_opcode;
+
 	ppd = dd->pport;
 	for (i = 0; i < dd->num_pports; i++, ppd++)
 		rvt_init_port(&dd->verbs_dev.rdi,
@@ -1961,6 +1751,9 @@ int hfi1_register_ib_device(struct hfi1_devdata *dd)
 			      i,
 			      ppd->pkeys);
 
+	rdma_set_device_sysfs_group(&dd->verbs_dev.rdi.ibdev,
+				    &ib_hfi1_attr_group);
+
 	ret = rvt_register_device(&dd->verbs_dev.rdi, RDMA_DRIVER_HFI1);
 	if (ret)
 		goto err_verbs_txreq;