[PATCH 0/9] NTB Update

[PATCH 0/9] NTB Update

[Jon Mason]

The following patches are intended for inclusion in the 3.13 kernel.
They include a number of bug fixes, some cleanups, and a couple of
features.

Please let me know if there are any issues.  Otherwise, I'll send a pull
request to Linus on Friday for these patches.

Thanks,
Jon

---

Alexander Gordeev (1):
      ntb: Fix missed call to pci_enable_msix()

Jon Mason (7):
      NTB: Xeon Doorbell errata workaround
      NTB: Fix NTB-RP Link Up
      NTB: Fix ntb_transport link down race
      NTB: correct dmaengine_get/put usage
      NTB: Document HW errata
      NTB: Enable Snoop on Primary Side
      NTB: Disable interrupts and poll under high load

Michael Opdenacker (1):
      NTB: remove duplicate defines

[PATCH 1/9] NTB: Xeon Doorbell errata workaround

[Jon Mason]

Modifications to the 14th bit of the B2BDOORBELL register will not be
mirrored to the remote system due to a hardware issue.  To get around
the issue, shrink the number of available doorbell bits by 1.  The max
number of doorbells was being used as a way to referencing the Link
Doorbell bit.  Since this would no longer work, the driver must now
explicitly reference that bit.

This does not affect the xeon_errata_workaround case, as it is not using
the b2bdoorbell register.

Signed-off-by: Jon Mason <jon.mason@intel.com>
---
 drivers/ntb/ntb_hw.c   |   20 +++++++++++++++-----
 drivers/ntb/ntb_regs.h |    1 +
 2 files changed, 16 insertions(+), 5 deletions(-)

diff --git a/drivers/ntb/ntb_hw.c b/drivers/ntb/ntb_hw.c
index 1cb6e51..6c9d733 100644
--- a/drivers/ntb/ntb_hw.c
+++ b/drivers/ntb/ntb_hw.c
@@ -678,6 +678,7 @@ static int ntb_xeon_setup(struct ntb_device *ndev)
 				return -EINVAL;
 
 			ndev->limits.max_mw = SNB_ERRATA_MAX_MW;
+			ndev->limits.max_db_bits = SNB_MAX_DB_BITS;
 			ndev->reg_ofs.spad_write = ndev->mw[1].vbase +
 						   SNB_SPAD_OFFSET;
 			ndev->reg_ofs.rdb = ndev->mw[1].vbase +
@@ -690,6 +691,13 @@ static int ntb_xeon_setup(struct ntb_device *ndev)
 			       SNB_PBAR4LMT_OFFSET);
 		} else {
 			ndev->limits.max_mw = SNB_MAX_MW;
+
+			/* HW Errata on bit 14 of b2bdoorbell register.  Writes
+			 * will not be mirrored to the remote system.  Shrink
+			 * the number of bits by one, since bit 14 is the last
+			 * bit.
+			 */
+			ndev->limits.max_db_bits = SNB_MAX_DB_BITS - 1;
 			ndev->reg_ofs.spad_write = ndev->reg_base +
 						   SNB_B2B_SPAD_OFFSET;
 			ndev->reg_ofs.rdb = ndev->reg_base +
@@ -769,6 +777,7 @@ static int ntb_xeon_setup(struct ntb_device *ndev)
 		 * have an equal amount.
 		 */
 		ndev->limits.max_spads = SNB_MAX_COMPAT_SPADS / 2;
+		ndev->limits.max_db_bits = SNB_MAX_DB_BITS;
 		/* Note: The SDOORBELL is the cause of the errata.  You REALLY
 		 * don't want to touch it.
 		 */
@@ -793,6 +802,7 @@ static int ntb_xeon_setup(struct ntb_device *ndev)
 		 * have an equal amount.
 		 */
 		ndev->limits.max_spads = SNB_MAX_COMPAT_SPADS / 2;
+		ndev->limits.max_db_bits = SNB_MAX_DB_BITS;
 		ndev->reg_ofs.rdb = ndev->reg_base + SNB_PDOORBELL_OFFSET;
 		ndev->reg_ofs.ldb = ndev->reg_base + SNB_SDOORBELL_OFFSET;
 		ndev->reg_ofs.ldb_mask = ndev->reg_base + SNB_SDBMSK_OFFSET;
@@ -819,7 +829,6 @@ static int ntb_xeon_setup(struct ntb_device *ndev)
 	ndev->reg_ofs.lnk_stat = ndev->reg_base + SNB_SLINK_STATUS_OFFSET;
 	ndev->reg_ofs.spci_cmd = ndev->reg_base + SNB_PCICMD_OFFSET;
 
-	ndev->limits.max_db_bits = SNB_MAX_DB_BITS;
 	ndev->limits.msix_cnt = SNB_MSIX_CNT;
 	ndev->bits_per_vector = SNB_DB_BITS_PER_VEC;
 
@@ -986,7 +995,7 @@ static irqreturn_t xeon_event_msix_irq(int irq, void *dev)
 		dev_err(&ndev->pdev->dev, "Error determining link status\n");
 
 	/* bit 15 is always the link bit */
-	writew(1 << ndev->limits.max_db_bits, ndev->reg_ofs.ldb);
+	writew(1 << SNB_LINK_DB, ndev->reg_ofs.ldb);
 
 	return IRQ_HANDLED;
 }
@@ -1176,9 +1185,10 @@ static int ntb_setup_interrupts(struct ntb_device *ndev)
 	 */
 	if (ndev->hw_type == BWD_HW)
 		writeq(~0, ndev->reg_ofs.ldb_mask);
-	else
-		writew(~(1 << ndev->limits.max_db_bits),
-		       ndev->reg_ofs.ldb_mask);
+	else {
+		u16 var = 1 << SNB_LINK_DB;
+		writew(~var, ndev->reg_ofs.ldb_mask);
+	}
 
 	rc = ntb_setup_msix(ndev);
 	if (!rc)
diff --git a/drivers/ntb/ntb_regs.h b/drivers/ntb/ntb_regs.h
index aa4bdd3..5381b65 100644
--- a/drivers/ntb/ntb_regs.h
+++ b/drivers/ntb/ntb_regs.h
@@ -55,6 +55,7 @@
 #define SNB_MAX_COMPAT_SPADS	16
 /* Reserve the uppermost bit for link interrupt */
 #define SNB_MAX_DB_BITS		15
+#define SNB_LINK_DB		15
 #define SNB_DB_BITS_PER_VEC	5
 #define SNB_MAX_MW		2
 #define SNB_ERRATA_MAX_MW	1
-- 
1.7.9.5

[PATCH 2/9] NTB: Fix NTB-RP Link Up

[Jon Mason]

The Xeon NTB-RP setup, the transparent side does not get a link up/down
interrupt.  Since the presence of a NTB device on the transparent side
means that we have a NTB link up, we can work around the lack of an
interrupt by simply calling the link up function to notify the upper
layers.

Signed-off-by: Jon Mason <jon.mason@intel.com>
---
 drivers/ntb/ntb_hw.c |   36 ++++++++++++++++++++++++++++--------
 1 file changed, 28 insertions(+), 8 deletions(-)

diff --git a/drivers/ntb/ntb_hw.c b/drivers/ntb/ntb_hw.c
index 6c9d733..a9b065f 100644
--- a/drivers/ntb/ntb_hw.c
+++ b/drivers/ntb/ntb_hw.c
@@ -1296,6 +1296,32 @@ static void ntb_free_debugfs(struct ntb_device *ndev)
 	}
 }
 
+static void ntb_hw_link_up(struct ntb_device *ndev)
+{
+	if (ndev->conn_type == NTB_CONN_TRANSPARENT)
+		ntb_link_event(ndev, NTB_LINK_UP);
+	else
+		/* Let's bring the NTB link up */
+		writel(NTB_CNTL_BAR23_SNOOP | NTB_CNTL_BAR45_SNOOP,
+		       ndev->reg_ofs.lnk_cntl);
+}
+
+static void ntb_hw_link_down(struct ntb_device *ndev)
+{
+	u32 ntb_cntl;
+
+	if (ndev->conn_type == NTB_CONN_TRANSPARENT) {
+		ntb_link_event(ndev, NTB_LINK_DOWN);
+		return;
+	}
+
+	/* Bring NTB link down */
+	ntb_cntl = readl(ndev->reg_ofs.lnk_cntl);
+	ntb_cntl &= ~(NTB_CNTL_BAR23_SNOOP | NTB_CNTL_BAR45_SNOOP);
+	ntb_cntl |= NTB_CNTL_LINK_DISABLE;
+	writel(ntb_cntl, ndev->reg_ofs.lnk_cntl);
+}
+
 static int ntb_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 {
 	struct ntb_device *ndev;
@@ -1384,9 +1410,7 @@ static int ntb_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 	if (rc)
 		goto err6;
 
-	/* Let's bring the NTB link up */
-	writel(NTB_CNTL_BAR23_SNOOP | NTB_CNTL_BAR45_SNOOP,
-	       ndev->reg_ofs.lnk_cntl);
+	ntb_hw_link_up(ndev);
 
 	return 0;
 
@@ -1416,12 +1440,8 @@ static void ntb_pci_remove(struct pci_dev *pdev)
 {
 	struct ntb_device *ndev = pci_get_drvdata(pdev);
 	int i;
-	u32 ntb_cntl;
 
-	/* Bring NTB link down */
-	ntb_cntl = readl(ndev->reg_ofs.lnk_cntl);
-	ntb_cntl |= NTB_CNTL_LINK_DISABLE;
-	writel(ntb_cntl, ndev->reg_ofs.lnk_cntl);
+	ntb_hw_link_down(ndev);
 
 	ntb_transport_free(ndev->ntb_transport);
 
-- 
1.7.9.5

[PATCH 3/9] ntb: Fix missed call to pci_enable_msix()

[Jon Mason]

From: Alexander Gordeev <agordeev@redhat.com>

Current MSI-X enablement code assumes MSI-Xs were successfully
allocated in case less than requested vectors were available.
That assumption is wrong, since MSI-Xs should be enabled with
a repeated call to pci_enable_msix(). This update fixes this.

Signed-off-by: Alexander Gordeev <agordeev@redhat.com>
Signed-off-by: Jon Mason <jon.mason@intel.com>
---
 drivers/ntb/ntb_hw.c |    4 ++++
 1 file changed, 4 insertions(+)

diff --git a/drivers/ntb/ntb_hw.c b/drivers/ntb/ntb_hw.c
index a9b065f..fa97948 100644
--- a/drivers/ntb/ntb_hw.c
+++ b/drivers/ntb/ntb_hw.c
@@ -1084,6 +1084,10 @@ static int ntb_setup_msix(struct ntb_device *ndev)
 			 "Only %d MSI-X vectors.  Limiting the number of queues to that number.\n",
 			 rc);
 		msix_entries = rc;
+
+		rc = pci_enable_msix(pdev, ndev->msix_entries, msix_entries);
+		if (rc)
+			goto err1;
 	}
 
 	for (i = 0; i < msix_entries; i++) {
-- 
1.7.9.5

[PATCH 4/9] NTB: Fix ntb_transport link down race

[Jon Mason]

A WARN_ON is being hit in ntb_qp_link_work due to the NTB transport link
being down while the ntb qp link is still active.  This is caused by the
transport link being brought down prior to the qp link worker thread
being terminated.  To correct this, shutdown the qp's prior to bringing
the transport link down.  Also, only call the qp worker thread if it is
in interrupt context, otherwise call the function directly.

Signed-off-by: Jon Mason <jon.mason@intel.com>
---
 drivers/ntb/ntb_transport.c |   43 ++++++++++++++++++++++++++++---------------
 1 file changed, 28 insertions(+), 15 deletions(-)

diff --git a/drivers/ntb/ntb_transport.c b/drivers/ntb/ntb_transport.c
index 12a9e83..7010f23 100644
--- a/drivers/ntb/ntb_transport.c
+++ b/drivers/ntb/ntb_transport.c
@@ -584,11 +584,8 @@ static int ntb_set_mw(struct ntb_transport *nt, int num_mw, unsigned int size)
 	return 0;
 }
 
-static void ntb_qp_link_cleanup(struct work_struct *work)
+static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 {
-	struct ntb_transport_qp *qp = container_of(work,
-						   struct ntb_transport_qp,
-						   link_cleanup);
 	struct ntb_transport *nt = qp->transport;
 	struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 
@@ -602,6 +599,16 @@ static void ntb_qp_link_cleanup(struct work_struct *work)
 
 	dev_info(&pdev->dev, "qp %d: Link Down\n", qp->qp_num);
 	qp->qp_link = NTB_LINK_DOWN;
+}
+
+static void ntb_qp_link_cleanup_work(struct work_struct *work)
+{
+	struct ntb_transport_qp *qp = container_of(work,
+						   struct ntb_transport_qp,
+						   link_cleanup);
+	struct ntb_transport *nt = qp->transport;
+
+	ntb_qp_link_cleanup(qp);
 
 	if (nt->transport_link == NTB_LINK_UP)
 		schedule_delayed_work(&qp->link_work,
@@ -613,22 +620,20 @@ static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 	schedule_work(&qp->link_cleanup);
 }
 
-static void ntb_transport_link_cleanup(struct work_struct *work)
+static void ntb_transport_link_cleanup(struct ntb_transport *nt)
 {
-	struct ntb_transport *nt = container_of(work, struct ntb_transport,
-						link_cleanup);
 	int i;
 
+	/* Pass along the info to any clients */
+	for (i = 0; i < nt->max_qps; i++)
+		if (!test_bit(i, &nt->qp_bitmap))
+			ntb_qp_link_cleanup(&nt->qps[i]);
+
 	if (nt->transport_link == NTB_LINK_DOWN)
 		cancel_delayed_work_sync(&nt->link_work);
 	else
 		nt->transport_link = NTB_LINK_DOWN;
 
-	/* Pass along the info to any clients */
-	for (i = 0; i < nt->max_qps; i++)
-		if (!test_bit(i, &nt->qp_bitmap))
-			ntb_qp_link_down(&nt->qps[i]);
-
 	/* The scratchpad registers keep the values if the remote side
 	 * goes down, blast them now to give them a sane value the next
 	 * time they are accessed
@@ -637,6 +642,14 @@ static void ntb_transport_link_cleanup(struct work_struct *work)
 		ntb_write_local_spad(nt->ndev, i, 0);
 }
 
+static void ntb_transport_link_cleanup_work(struct work_struct *work)
+{
+	struct ntb_transport *nt = container_of(work, struct ntb_transport,
+						link_cleanup);
+
+	ntb_transport_link_cleanup(nt);
+}
+
 static void ntb_transport_event_callback(void *data, enum ntb_hw_event event)
 {
 	struct ntb_transport *nt = data;
@@ -880,7 +893,7 @@ static int ntb_transport_init_queue(struct ntb_transport *nt,
 	}
 
 	INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
-	INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup);
+	INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
 
 	spin_lock_init(&qp->ntb_rx_pend_q_lock);
 	spin_lock_init(&qp->ntb_rx_free_q_lock);
@@ -936,7 +949,7 @@ int ntb_transport_init(struct pci_dev *pdev)
 	}
 
 	INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
-	INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup);
+	INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
 
 	rc = ntb_register_event_callback(nt->ndev,
 					 ntb_transport_event_callback);
@@ -972,7 +985,7 @@ void ntb_transport_free(void *transport)
 	struct ntb_device *ndev = nt->ndev;
 	int i;
 
-	nt->transport_link = NTB_LINK_DOWN;
+	ntb_transport_link_cleanup(nt);
 
 	/* verify that all the qp's are freed */
 	for (i = 0; i < nt->max_qps; i++) {
-- 
1.7.9.5

[PATCH 5/9] NTB: correct dmaengine_get/put usage

[Jon Mason]

dmaengine_get() causes the initialization of the per-cpu channel tables.
It needs to be called prior to dma_find_channel().

Initial version by Dan Williams <dan.j.williams@intel.com>

Signed-off-by: Jon Mason <jon.mason@intel.com>
---
 drivers/ntb/ntb_transport.c |    9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/drivers/ntb/ntb_transport.c b/drivers/ntb/ntb_transport.c
index 7010f23..172501d 100644
--- a/drivers/ntb/ntb_transport.c
+++ b/drivers/ntb/ntb_transport.c
@@ -1419,11 +1419,12 @@ ntb_transport_create_queue(void *data, struct pci_dev *pdev,
 	qp->tx_handler = handlers->tx_handler;
 	qp->event_handler = handlers->event_handler;
 
+	dmaengine_get();
 	qp->dma_chan = dma_find_channel(DMA_MEMCPY);
-	if (!qp->dma_chan)
+	if (!qp->dma_chan) {
+		dmaengine_put();
 		dev_info(&pdev->dev, "Unable to allocate DMA channel, using CPU instead\n");
-	else
-		dmaengine_get();
+	}
 
 	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
 		entry = kzalloc(sizeof(struct ntb_queue_entry), GFP_ATOMIC);
@@ -1464,6 +1465,8 @@ err2:
 err1:
 	while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
 		kfree(entry);
+	if (qp->dma_chan)
+		dmaengine_put();
 	set_bit(free_queue, &nt->qp_bitmap);
 err:
 	return NULL;
-- 
1.7.9.5

[PATCH 6/9] NTB: remove duplicate defines

[Jon Mason]

From: Michael Opdenacker <michael.opdenacker@free-electrons.com>

Remove duplicate defines in drivers/ntb/ntb_regs.h

Signed-off-by: Michael Opdenacker <michael.opdenacker@free-electrons.com>
Signed-off-by: Jon Mason <jon.mason@intel.com>
---
 drivers/ntb/ntb_regs.h |    3 ---
 1 file changed, 3 deletions(-)

diff --git a/drivers/ntb/ntb_regs.h b/drivers/ntb/ntb_regs.h
index 5381b65..a689f1c 100644
--- a/drivers/ntb/ntb_regs.h
+++ b/drivers/ntb/ntb_regs.h
@@ -76,9 +76,6 @@
 #define SNB_SBAR2XLAT_OFFSET	0x0030
 #define SNB_SBAR4XLAT_OFFSET	0x0038
 #define SNB_SBAR0BASE_OFFSET	0x0040
-#define SNB_SBAR0BASE_OFFSET	0x0040
-#define SNB_SBAR2BASE_OFFSET	0x0048
-#define SNB_SBAR4BASE_OFFSET	0x0050
 #define SNB_SBAR2BASE_OFFSET	0x0048
 #define SNB_SBAR4BASE_OFFSET	0x0050
 #define SNB_NTBCNTL_OFFSET	0x0058
-- 
1.7.9.5

[PATCH 7/9] NTB: Document HW errata

[Jon Mason]

Add a comment describing the necessary ordering of modifications to the
NTB Limit and Base registers.

Signed-off-by: Jon Mason <jon.mason@intel.com>
---
 drivers/ntb/ntb_hw.c |   12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/drivers/ntb/ntb_hw.c b/drivers/ntb/ntb_hw.c
index fa97948..391c377 100644
--- a/drivers/ntb/ntb_hw.c
+++ b/drivers/ntb/ntb_hw.c
@@ -689,6 +689,12 @@ static int ntb_xeon_setup(struct ntb_device *ndev)
 			 */
 			writeq(ndev->mw[1].bar_sz + 0x1000, ndev->reg_base +
 			       SNB_PBAR4LMT_OFFSET);
+			/* HW errata on the Limit registers.  They can only be
+			 * written when the base register is 4GB aligned and
+			 * < 32bit.  This should already be the case based on the
+			 * driver defaults, but write the Limit registers first
+			 * just in case.
+			 */
 		} else {
 			ndev->limits.max_mw = SNB_MAX_MW;
 
@@ -707,6 +713,12 @@ static int ntb_xeon_setup(struct ntb_device *ndev)
 			 * something silly
 			 */
 			writeq(0, ndev->reg_base + SNB_PBAR4LMT_OFFSET);
+			/* HW errata on the Limit registers.  They can only be
+			 * written when the base register is 4GB aligned and
+			 * < 32bit.  This should already be the case based on the
+			 * driver defaults, but write the Limit registers first
+			 * just in case.
+			 */
 		}
 
 		/* The Xeon errata workaround requires setting SBAR Base
-- 
1.7.9.5

[PATCH 8/9] NTB: Enable Snoop on Primary Side

[Jon Mason]

Enable Snoop from Primary to Secondary side on BAR23 and BAR45 on all
TLPs.  Previously, Snoop was only enabled from Secondary to Primary
side.  This can have a performance improvement on some workloads.

Also, make the code more obvious about how the link is being enabled.

Signed-off-by: Jon Mason <jon.mason@intel.com>
---
 drivers/ntb/ntb_hw.c   |   17 ++++++++++++-----
 drivers/ntb/ntb_regs.h |   12 +++++++-----
 2 files changed, 19 insertions(+), 10 deletions(-)

diff --git a/drivers/ntb/ntb_hw.c b/drivers/ntb/ntb_hw.c
index 391c377..80505ae 100644
--- a/drivers/ntb/ntb_hw.c
+++ b/drivers/ntb/ntb_hw.c
@@ -1316,10 +1316,16 @@ static void ntb_hw_link_up(struct ntb_device *ndev)
 {
 	if (ndev->conn_type == NTB_CONN_TRANSPARENT)
 		ntb_link_event(ndev, NTB_LINK_UP);
-	else
+	else {
+		u32 ntb_cntl;
+
 		/* Let's bring the NTB link up */
-		writel(NTB_CNTL_BAR23_SNOOP | NTB_CNTL_BAR45_SNOOP,
-		       ndev->reg_ofs.lnk_cntl);
+		ntb_cntl = readl(ndev->reg_ofs.lnk_cntl);
+		ntb_cntl &= ~(NTB_CNTL_LINK_DISABLE | NTB_CNTL_CFG_LOCK);
+		ntb_cntl |= NTB_CNTL_P2S_BAR23_SNOOP | NTB_CNTL_S2P_BAR23_SNOOP;
+		ntb_cntl |= NTB_CNTL_P2S_BAR45_SNOOP | NTB_CNTL_S2P_BAR45_SNOOP;
+		writel(ntb_cntl, ndev->reg_ofs.lnk_cntl);
+	}
 }
 
 static void ntb_hw_link_down(struct ntb_device *ndev)
@@ -1333,8 +1339,9 @@ static void ntb_hw_link_down(struct ntb_device *ndev)
 
 	/* Bring NTB link down */
 	ntb_cntl = readl(ndev->reg_ofs.lnk_cntl);
-	ntb_cntl &= ~(NTB_CNTL_BAR23_SNOOP | NTB_CNTL_BAR45_SNOOP);
-	ntb_cntl |= NTB_CNTL_LINK_DISABLE;
+	ntb_cntl &= ~(NTB_CNTL_P2S_BAR23_SNOOP | NTB_CNTL_S2P_BAR23_SNOOP);
+	ntb_cntl &= ~(NTB_CNTL_P2S_BAR45_SNOOP | NTB_CNTL_S2P_BAR45_SNOOP);
+	ntb_cntl |= NTB_CNTL_LINK_DISABLE | NTB_CNTL_CFG_LOCK;
 	writel(ntb_cntl, ndev->reg_ofs.lnk_cntl);
 }
 
diff --git a/drivers/ntb/ntb_regs.h b/drivers/ntb/ntb_regs.h
index a689f1c..9774506 100644
--- a/drivers/ntb/ntb_regs.h
+++ b/drivers/ntb/ntb_regs.h
@@ -143,11 +143,13 @@
 #define BWD_LTSSMSTATEJMP_FORCEDETECT	(1 << 2)
 #define BWD_IBIST_ERR_OFLOW	0x7FFF7FFF
 
-#define NTB_CNTL_CFG_LOCK	(1 << 0)
-#define NTB_CNTL_LINK_DISABLE	(1 << 1)
-#define NTB_CNTL_BAR23_SNOOP	(1 << 2)
-#define NTB_CNTL_BAR45_SNOOP	(1 << 6)
-#define BWD_CNTL_LINK_DOWN	(1 << 16)
+#define NTB_CNTL_CFG_LOCK		(1 << 0)
+#define NTB_CNTL_LINK_DISABLE		(1 << 1)
+#define NTB_CNTL_S2P_BAR23_SNOOP	(1 << 2)
+#define NTB_CNTL_P2S_BAR23_SNOOP	(1 << 4)
+#define NTB_CNTL_S2P_BAR45_SNOOP	(1 << 6)
+#define NTB_CNTL_P2S_BAR45_SNOOP	(1 << 8)
+#define BWD_CNTL_LINK_DOWN		(1 << 16)
 
 #define NTB_PPD_OFFSET		0x00D4
 #define SNB_PPD_CONN_TYPE	0x0003
-- 
1.7.9.5

[PATCH 9/9] NTB: Disable interrupts and poll under high load

[Jon Mason]

Disable interrupts and poll under high load

Signed-off-by: Jon Mason <jon.mason@intel.com>
---
 drivers/ntb/ntb_hw.c        |   42 +++++++++++++++++++++++++++++++++++++-----
 drivers/ntb/ntb_hw.h        |    7 ++++---
 drivers/ntb/ntb_transport.c |   25 +++++++------------------
 3 files changed, 48 insertions(+), 26 deletions(-)

diff --git a/drivers/ntb/ntb_hw.c b/drivers/ntb/ntb_hw.c
index 80505ae..170e8e6 100644
--- a/drivers/ntb/ntb_hw.c
+++ b/drivers/ntb/ntb_hw.c
@@ -141,6 +141,24 @@ void ntb_unregister_event_callback(struct ntb_device *ndev)
 	ndev->event_cb = NULL;
 }
 
+static void ntb_irq_work(unsigned long data)
+{
+	struct ntb_db_cb *db_cb = (struct ntb_db_cb *)data;
+	int rc;
+
+	rc = db_cb->callback(db_cb->data, db_cb->db_num);
+	if (rc)
+		tasklet_schedule(&db_cb->irq_work);
+	else {
+		struct ntb_device *ndev = db_cb->ndev;
+		unsigned long mask;
+
+		mask = readw(ndev->reg_ofs.ldb_mask);
+		clear_bit(db_cb->db_num * ndev->bits_per_vector, &mask);
+		writew(mask, ndev->reg_ofs.ldb_mask);
+	}
+}
+
 /**
  * ntb_register_db_callback() - register a callback for doorbell interrupt
  * @ndev: pointer to ntb_device instance
@@ -155,7 +173,7 @@ void ntb_unregister_event_callback(struct ntb_device *ndev)
  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
  */
 int ntb_register_db_callback(struct ntb_device *ndev, unsigned int idx,
-			     void *data, void (*func)(void *data, int db_num))
+			     void *data, int (*func)(void *data, int db_num))
 {
 	unsigned long mask;
 
@@ -166,6 +184,10 @@ int ntb_register_db_callback(struct ntb_device *ndev, unsigned int idx,
 
 	ndev->db_cb[idx].callback = func;
 	ndev->db_cb[idx].data = data;
+	ndev->db_cb[idx].ndev = ndev;
+
+	tasklet_init(&ndev->db_cb[idx].irq_work, ntb_irq_work,
+		     (unsigned long) &ndev->db_cb[idx]);
 
 	/* unmask interrupt */
 	mask = readw(ndev->reg_ofs.ldb_mask);
@@ -194,6 +216,8 @@ void ntb_unregister_db_callback(struct ntb_device *ndev, unsigned int idx)
 	set_bit(idx * ndev->bits_per_vector, &mask);
 	writew(mask, ndev->reg_ofs.ldb_mask);
 
+	tasklet_disable(&ndev->db_cb[idx].irq_work);
+
 	ndev->db_cb[idx].callback = NULL;
 }
 
@@ -955,12 +979,16 @@ static irqreturn_t bwd_callback_msix_irq(int irq, void *data)
 {
 	struct ntb_db_cb *db_cb = data;
 	struct ntb_device *ndev = db_cb->ndev;
+	unsigned long mask;
 
 	dev_dbg(&ndev->pdev->dev, "MSI-X irq %d received for DB %d\n", irq,
 		db_cb->db_num);
 
-	if (db_cb->callback)
-		db_cb->callback(db_cb->data, db_cb->db_num);
+	mask = readw(ndev->reg_ofs.ldb_mask);
+	set_bit(db_cb->db_num * ndev->bits_per_vector, &mask);
+	writew(mask, ndev->reg_ofs.ldb_mask);
+
+	tasklet_schedule(&db_cb->irq_work);
 
 	/* No need to check for the specific HB irq, any interrupt means
 	 * we're connected.
@@ -976,12 +1004,16 @@ static irqreturn_t xeon_callback_msix_irq(int irq, void *data)
 {
 	struct ntb_db_cb *db_cb = data;
 	struct ntb_device *ndev = db_cb->ndev;
+	unsigned long mask;
 
 	dev_dbg(&ndev->pdev->dev, "MSI-X irq %d received for DB %d\n", irq,
 		db_cb->db_num);
 
-	if (db_cb->callback)
-		db_cb->callback(db_cb->data, db_cb->db_num);
+	mask = readw(ndev->reg_ofs.ldb_mask);
+	set_bit(db_cb->db_num * ndev->bits_per_vector, &mask);
+	writew(mask, ndev->reg_ofs.ldb_mask);
+
+	tasklet_schedule(&db_cb->irq_work);
 
 	/* On Sandybridge, there are 16 bits in the interrupt register
 	 * but only 4 vectors.  So, 5 bits are assigned to the first 3
diff --git a/drivers/ntb/ntb_hw.h b/drivers/ntb/ntb_hw.h
index 0a31ced..bbdb7ed 100644
--- a/drivers/ntb/ntb_hw.h
+++ b/drivers/ntb/ntb_hw.h
@@ -106,10 +106,11 @@ struct ntb_mw {
 };
 
 struct ntb_db_cb {
-	void (*callback) (void *data, int db_num);
+	int (*callback)(void *data, int db_num);
 	unsigned int db_num;
 	void *data;
 	struct ntb_device *ndev;
+	struct tasklet_struct irq_work;
 };
 
 struct ntb_device {
@@ -228,8 +229,8 @@ struct ntb_device *ntb_register_transport(struct pci_dev *pdev,
 void ntb_unregister_transport(struct ntb_device *ndev);
 void ntb_set_mw_addr(struct ntb_device *ndev, unsigned int mw, u64 addr);
 int ntb_register_db_callback(struct ntb_device *ndev, unsigned int idx,
-			     void *data, void (*db_cb_func) (void *data,
-							     int db_num));
+			     void *data, int (*db_cb_func)(void *data,
+							   int db_num));
 void ntb_unregister_db_callback(struct ntb_device *ndev, unsigned int idx);
 int ntb_register_event_callback(struct ntb_device *ndev,
 				void (*event_cb_func) (void *handle,
diff --git a/drivers/ntb/ntb_transport.c b/drivers/ntb/ntb_transport.c
index 172501d..676937b 100644
--- a/drivers/ntb/ntb_transport.c
+++ b/drivers/ntb/ntb_transport.c
@@ -119,7 +119,6 @@ struct ntb_transport_qp {
 
 	void (*rx_handler) (struct ntb_transport_qp *qp, void *qp_data,
 			    void *data, int len);
-	struct tasklet_struct rx_work;
 	struct list_head rx_pend_q;
 	struct list_head rx_free_q;
 	spinlock_t ntb_rx_pend_q_lock;
@@ -1188,11 +1187,14 @@ err:
 	goto out;
 }
 
-static void ntb_transport_rx(unsigned long data)
+static int ntb_transport_rxc_db(void *data, int db_num)
 {
-	struct ntb_transport_qp *qp = (struct ntb_transport_qp *)data;
+	struct ntb_transport_qp *qp = data;
 	int rc, i;
 
+	dev_dbg(&ntb_query_pdev(qp->ndev)->dev, "%s: doorbell %d received\n",
+		__func__, db_num);
+
 	/* Limit the number of packets processed in a single interrupt to
 	 * provide fairness to others
 	 */
@@ -1204,16 +1206,8 @@ static void ntb_transport_rx(unsigned long data)
 
 	if (qp->dma_chan)
 		dma_async_issue_pending(qp->dma_chan);
-}
-
-static void ntb_transport_rxc_db(void *data, int db_num)
-{
-	struct ntb_transport_qp *qp = data;
 
-	dev_dbg(&ntb_query_pdev(qp->ndev)->dev, "%s: doorbell %d received\n",
-		__func__, db_num);
-
-	tasklet_schedule(&qp->rx_work);
+	return i;
 }
 
 static void ntb_tx_copy_callback(void *data)
@@ -1446,19 +1440,15 @@ ntb_transport_create_queue(void *data, struct pci_dev *pdev,
 			     &qp->tx_free_q);
 	}
 
-	tasklet_init(&qp->rx_work, ntb_transport_rx, (unsigned long) qp);
-
 	rc = ntb_register_db_callback(qp->ndev, free_queue, qp,
 				      ntb_transport_rxc_db);
 	if (rc)
-		goto err3;
+		goto err2;
 
 	dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
 
 	return qp;
 
-err3:
-	tasklet_disable(&qp->rx_work);
 err2:
 	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
 		kfree(entry);
@@ -1505,7 +1495,6 @@ void ntb_transport_free_queue(struct ntb_transport_qp *qp)
 	}
 
 	ntb_unregister_db_callback(qp->ndev, qp->qp_num);
-	tasklet_disable(&qp->rx_work);
 
 	cancel_delayed_work_sync(&qp->link_work);
 
-- 
1.7.9.5