qed: VFs to try utilizing the doorbell bar

VFs are currently not mapping their doorbell bar, instead relying
on the small doorbell window they have in their limited regview bar.

In order to increase the number of possible Tx connections [queues]
employeed by VF past 16, we need to start using the doorbell bar if
one such is exposed - VF would communicate this fact to PF which would
return the size-bar internally configured into chip, according to
which the VF would decide whether to actually utilize the doorbell
bar.
Signed-off-by: Yuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/qlogic/qed/qed.h

@@ -412,6 +412,11 @@ struct qed_fw_data {
 	u32			init_ops_size;
 };
 
+enum BAR_ID {
+	BAR_ID_0,		/* used for GRC */
+	BAR_ID_1		/* Used for doorbells */
+};
+
 #define DRV_MODULE_VERSION		      \
 	__stringify(QED_MAJOR_VERSION) "."    \
 	__stringify(QED_MINOR_VERSION) "."    \

drivers/net/ethernet/qlogic/qed/qed_dev.c

@@ -69,12 +69,6 @@ static DEFINE_SPINLOCK(qm_lock);
 #define QED_MIN_DPIS            (4)
 #define QED_MIN_PWM_REGION      (QED_WID_SIZE * QED_MIN_DPIS)
 
-/* API common to all protocols */
-enum BAR_ID {
-	BAR_ID_0,       /* used for GRC */
-	BAR_ID_1        /* Used for doorbells */
-};
-
 static u32 qed_hw_bar_size(struct qed_hwfn *p_hwfn,
 			   struct qed_ptt *p_ptt, enum BAR_ID bar_id)
 {
@@ -83,7 +77,7 @@ static u32 qed_hw_bar_size(struct qed_hwfn *p_hwfn,
 	u32 val;
 
 	if (IS_VF(p_hwfn->cdev))
-		return 1 << 17;
+		return qed_vf_hw_bar_size(p_hwfn, bar_id);
 
 	val = qed_rd(p_hwfn, p_ptt, bar_reg);
 	if (val)

drivers/net/ethernet/qlogic/qed/qed_main.c

@@ -122,7 +122,7 @@ static void qed_free_pci(struct qed_dev *cdev)
 {
 	struct pci_dev *pdev = cdev->pdev;
 
-	if (cdev->doorbells)
+	if (cdev->doorbells && cdev->db_size)
 		iounmap(cdev->doorbells);
 	if (cdev->regview)
 		iounmap(cdev->regview);
@@ -206,15 +206,23 @@ static int qed_init_pci(struct qed_dev *cdev, struct pci_dev *pdev)
 		goto err2;
 	}
 
-	if (IS_PF(cdev)) {
 	cdev->db_phys_addr = pci_resource_start(cdev->pdev, 2);
 	cdev->db_size = pci_resource_len(cdev->pdev, 2);
+	if (!cdev->db_size) {
+		if (IS_PF(cdev)) {
+			DP_NOTICE(cdev, "No Doorbell bar available\n");
+			return -EINVAL;
+		} else {
+			return 0;
+		}
+	}
+
 	cdev->doorbells = ioremap_wc(cdev->db_phys_addr, cdev->db_size);
+
 	if (!cdev->doorbells) {
 		DP_NOTICE(cdev, "Cannot map doorbell space\n");
 		return -ENOMEM;
 	}
-	}
 
 	return 0;
 

drivers/net/ethernet/qlogic/qed/qed_reg_addr.h

@@ -560,6 +560,7 @@
 	0x2aae60UL
 #define PGLUE_B_REG_PF_BAR1_SIZE \
 	0x2aae64UL
+#define PGLUE_B_REG_VF_BAR1_SIZE 0x2aae68UL
 #define PRS_REG_ENCAPSULATION_TYPE_EN	0x1f0730UL
 #define PRS_REG_GRE_PROTOCOL		0x1f0734UL
 #define PRS_REG_VXLAN_PORT		0x1f0738UL

drivers/net/ethernet/qlogic/qed/qed_sriov.c

@@ -1384,6 +1384,60 @@ static void qed_iov_vf_cleanup(struct qed_hwfn *p_hwfn,
 	qed_iov_clean_vf(p_hwfn, p_vf->relative_vf_id);
 }
 
+/* Returns either 0, or log(size) */
+static u32 qed_iov_vf_db_bar_size(struct qed_hwfn *p_hwfn,
+				  struct qed_ptt *p_ptt)
+{
+	u32 val = qed_rd(p_hwfn, p_ptt, PGLUE_B_REG_VF_BAR1_SIZE);
+
+	if (val)
+		return val + 11;
+	return 0;
+}
+
+static void
+qed_iov_vf_mbx_acquire_resc_cids(struct qed_hwfn *p_hwfn,
+				 struct qed_ptt *p_ptt,
+				 struct qed_vf_info *p_vf,
+				 struct vf_pf_resc_request *p_req,
+				 struct pf_vf_resc *p_resp)
+{
+	u8 num_vf_cons = p_hwfn->pf_params.eth_pf_params.num_vf_cons;
+	u8 db_size = qed_db_addr_vf(1, DQ_DEMS_LEGACY) -
+		     qed_db_addr_vf(0, DQ_DEMS_LEGACY);
+	u32 bar_size;
+
+	p_resp->num_cids = min_t(u8, p_req->num_cids, num_vf_cons);
+
+	/* If VF didn't bother asking for QIDs than don't bother limiting
+	 * number of CIDs. The VF doesn't care about the number, and this
+	 * has the likely result of causing an additional acquisition.
+	 */
+	if (!(p_vf->acquire.vfdev_info.capabilities &
+	      VFPF_ACQUIRE_CAP_QUEUE_QIDS))
+		return;
+
+	/* If doorbell bar was mapped by VF, limit the VF CIDs to an amount
+	 * that would make sure doorbells for all CIDs fall within the bar.
+	 * If it doesn't, make sure regview window is sufficient.
+	 */
+	if (p_vf->acquire.vfdev_info.capabilities &
+	    VFPF_ACQUIRE_CAP_PHYSICAL_BAR) {
+		bar_size = qed_iov_vf_db_bar_size(p_hwfn, p_ptt);
+		if (bar_size)
+			bar_size = 1 << bar_size;
+
+		if (p_hwfn->cdev->num_hwfns > 1)
+			bar_size /= 2;
+	} else {
+		bar_size = PXP_VF_BAR0_DQ_LENGTH;
+	}
+
+	if (bar_size / db_size < 256)
+		p_resp->num_cids = min_t(u8, p_resp->num_cids,
+					 (u8)(bar_size / db_size));
+}
+
 static u8 qed_iov_vf_mbx_acquire_resc(struct qed_hwfn *p_hwfn,
 				      struct qed_ptt *p_ptt,
 				      struct qed_vf_info *p_vf,
@@ -1417,9 +1471,7 @@ static u8 qed_iov_vf_mbx_acquire_resc(struct qed_hwfn *p_hwfn,
 	p_resp->num_vlan_filters = min_t(u8, p_vf->num_vlan_filters,
 					 p_req->num_vlan_filters);
 
-	p_resp->num_cids =
-	    min_t(u8, p_req->num_cids,
-		  p_hwfn->pf_params.eth_pf_params.num_vf_cons);
+	qed_iov_vf_mbx_acquire_resc_cids(p_hwfn, p_ptt, p_vf, p_req, p_resp);
 
 	/* This isn't really needed/enforced, but some legacy VFs might depend
 	 * on the correct filling of this field.
@@ -1572,6 +1624,9 @@ static void qed_iov_vf_mbx_acquire(struct qed_hwfn *p_hwfn,
 	if (req->vfdev_info.capabilities & VFPF_ACQUIRE_CAP_QUEUE_QIDS)
 		pfdev_info->capabilities |= PFVF_ACQUIRE_CAP_QUEUE_QIDS;
 
+	/* Share the sizes of the bars with VF */
+	resp->pfdev_info.bar_size = qed_iov_vf_db_bar_size(p_hwfn, p_ptt);
+
 	qed_iov_vf_mbx_acquire_stats(p_hwfn, &pfdev_info->stats_info);
 
 	memcpy(pfdev_info->port_mac, p_hwfn->hw_info.hw_mac_addr, ETH_ALEN);

drivers/net/ethernet/qlogic/qed/qed_vf.c

@@ -169,6 +169,61 @@ static void qed_vf_pf_add_qid(struct qed_hwfn *p_hwfn,
 	p_qid_tlv->qid = p_cid->qid_usage_idx;
 }
 
+int _qed_vf_pf_release(struct qed_hwfn *p_hwfn, bool b_final)
+{
+	struct qed_vf_iov *p_iov = p_hwfn->vf_iov_info;
+	struct pfvf_def_resp_tlv *resp;
+	struct vfpf_first_tlv *req;
+	u32 size;
+	int rc;
+
+	/* clear mailbox and prep first tlv */
+	req = qed_vf_pf_prep(p_hwfn, CHANNEL_TLV_RELEASE, sizeof(*req));
+
+	/* add list termination tlv */
+	qed_add_tlv(p_hwfn, &p_iov->offset,
+		    CHANNEL_TLV_LIST_END, sizeof(struct channel_list_end_tlv));
+
+	resp = &p_iov->pf2vf_reply->default_resp;
+	rc = qed_send_msg2pf(p_hwfn, &resp->hdr.status, sizeof(*resp));
+
+	if (!rc && resp->hdr.status != PFVF_STATUS_SUCCESS)
+		rc = -EAGAIN;
+
+	qed_vf_pf_req_end(p_hwfn, rc);
+	if (!b_final)
+		return rc;
+
+	p_hwfn->b_int_enabled = 0;
+
+	if (p_iov->vf2pf_request)
+		dma_free_coherent(&p_hwfn->cdev->pdev->dev,
+				  sizeof(union vfpf_tlvs),
+				  p_iov->vf2pf_request,
+				  p_iov->vf2pf_request_phys);
+	if (p_iov->pf2vf_reply)
+		dma_free_coherent(&p_hwfn->cdev->pdev->dev,
+				  sizeof(union pfvf_tlvs),
+				  p_iov->pf2vf_reply, p_iov->pf2vf_reply_phys);
+
+	if (p_iov->bulletin.p_virt) {
+		size = sizeof(struct qed_bulletin_content);
+		dma_free_coherent(&p_hwfn->cdev->pdev->dev,
+				  size,
+				  p_iov->bulletin.p_virt, p_iov->bulletin.phys);
+	}
+
+	kfree(p_hwfn->vf_iov_info);
+	p_hwfn->vf_iov_info = NULL;
+
+	return rc;
+}
+
+int qed_vf_pf_release(struct qed_hwfn *p_hwfn)
+{
+	return _qed_vf_pf_release(p_hwfn, true);
+}
+
 #define VF_ACQUIRE_THRESH 3
 static void qed_vf_pf_acquire_reduce_resc(struct qed_hwfn *p_hwfn,
 					  struct vf_pf_resc_request *p_req,
@@ -235,6 +290,11 @@ static int qed_vf_pf_acquire(struct qed_hwfn *p_hwfn)
 	/* Fill capability field with any non-deprecated config we support */
 	req->vfdev_info.capabilities |= VFPF_ACQUIRE_CAP_100G;
 
+	/* If we've mapped the doorbell bar, try using queue qids */
+	if (p_iov->b_doorbell_bar)
+		req->vfdev_info.capabilities |= VFPF_ACQUIRE_CAP_PHYSICAL_BAR |
+						VFPF_ACQUIRE_CAP_QUEUE_QIDS;
+
 	/* pf 2 vf bulletin board address */
 	req->bulletin_addr = p_iov->bulletin.phys;
 	req->bulletin_size = p_iov->bulletin.size;
@@ -364,10 +424,27 @@ static int qed_vf_pf_acquire(struct qed_hwfn *p_hwfn)
 	return rc;
 }
 
+u32 qed_vf_hw_bar_size(struct qed_hwfn *p_hwfn, enum BAR_ID bar_id)
+{
+	u32 bar_size;
+
+	/* Regview size is fixed */
+	if (bar_id == BAR_ID_0)
+		return 1 << 17;
+
+	/* Doorbell is received from PF */
+	bar_size = p_hwfn->vf_iov_info->acquire_resp.pfdev_info.bar_size;
+	if (bar_size)
+		return 1 << bar_size;
+	return 0;
+}
+
 int qed_vf_hw_prepare(struct qed_hwfn *p_hwfn)
 {
+	struct qed_hwfn *p_lead = QED_LEADING_HWFN(p_hwfn->cdev);
 	struct qed_vf_iov *p_iov;
 	u32 reg;
+	int rc;
 
 	/* Set number of hwfns - might be overriden once leading hwfn learns
 	 * actual configuration from PF.
@@ -375,10 +452,6 @@ int qed_vf_hw_prepare(struct qed_hwfn *p_hwfn)
 	if (IS_LEAD_HWFN(p_hwfn))
 		p_hwfn->cdev->num_hwfns = 1;
 
-	/* Set the doorbell bar. Assumption: regview is set */
-	p_hwfn->doorbells = (u8 __iomem *)p_hwfn->regview +
-					  PXP_VF_BAR0_START_DQ;
-
 	reg = PXP_VF_BAR0_ME_OPAQUE_ADDRESS;
 	p_hwfn->hw_info.opaque_fid = (u16)REG_RD(p_hwfn, reg);
 
@@ -390,6 +463,30 @@ int qed_vf_hw_prepare(struct qed_hwfn *p_hwfn)
 	if (!p_iov)
 		return -ENOMEM;
 
+	/* Doorbells are tricky; Upper-layer has alreday set the hwfn doorbell
+	 * value, but there are several incompatibily scenarios where that
+	 * would be incorrect and we'd need to override it.
+	 */
+	if (!p_hwfn->doorbells) {
+		p_hwfn->doorbells = (u8 __iomem *)p_hwfn->regview +
+						  PXP_VF_BAR0_START_DQ;
+	} else if (p_hwfn == p_lead) {
+		/* For leading hw-function, value is always correct, but need
+		 * to handle scenario where legacy PF would not support 100g
+		 * mapped bars later.
+		 */
+		p_iov->b_doorbell_bar = true;
+	} else {
+		/* here, value would be correct ONLY if the leading hwfn
+		 * received indication that mapped-bars are supported.
+		 */
+		if (p_lead->vf_iov_info->b_doorbell_bar)
+			p_iov->b_doorbell_bar = true;
+		else
+			p_hwfn->doorbells = (u8 __iomem *)
+			    p_hwfn->regview + PXP_VF_BAR0_START_DQ;
+	}
+
 	/* Allocate vf2pf msg */
 	p_iov->vf2pf_request = dma_alloc_coherent(&p_hwfn->cdev->pdev->dev,
 						  sizeof(union vfpf_tlvs),
@@ -429,7 +526,33 @@ int qed_vf_hw_prepare(struct qed_hwfn *p_hwfn)
 
 	p_hwfn->hw_info.personality = QED_PCI_ETH;
 
-	return qed_vf_pf_acquire(p_hwfn);
+	rc = qed_vf_pf_acquire(p_hwfn);
+
+	/* If VF is 100g using a mapped bar and PF is too old to support that,
+	 * acquisition would succeed - but the VF would have no way knowing
+	 * the size of the doorbell bar configured in HW and thus will not
+	 * know how to split it for 2nd hw-function.
+	 * In this case we re-try without the indication of the mapped
+	 * doorbell.
+	 */
+	if (!rc && p_iov->b_doorbell_bar &&
+	    !qed_vf_hw_bar_size(p_hwfn, BAR_ID_1) &&
+	    (p_hwfn->cdev->num_hwfns > 1)) {
+		rc = _qed_vf_pf_release(p_hwfn, false);
+		if (rc)
+			return rc;
+
+		p_iov->b_doorbell_bar = false;
+		p_hwfn->doorbells = (u8 __iomem *)p_hwfn->regview +
+						  PXP_VF_BAR0_START_DQ;
+		rc = qed_vf_pf_acquire(p_hwfn);
+	}
+
+	DP_VERBOSE(p_hwfn, QED_MSG_IOV,
+		   "Regview [%p], Doorbell [%p], Device-doorbell [%p]\n",
+		   p_hwfn->regview, p_hwfn->doorbells, p_hwfn->cdev->doorbells);
+
+	return rc;
 
 free_vf2pf_request:
 	dma_free_coherent(&p_hwfn->cdev->pdev->dev,
@@ -1133,54 +1256,6 @@ int qed_vf_pf_reset(struct qed_hwfn *p_hwfn)
 	return rc;
 }
 
-int qed_vf_pf_release(struct qed_hwfn *p_hwfn)
-{
-	struct qed_vf_iov *p_iov = p_hwfn->vf_iov_info;
-	struct pfvf_def_resp_tlv *resp;
-	struct vfpf_first_tlv *req;
-	u32 size;
-	int rc;
-
-	/* clear mailbox and prep first tlv */
-	req = qed_vf_pf_prep(p_hwfn, CHANNEL_TLV_RELEASE, sizeof(*req));
-
-	/* add list termination tlv */
-	qed_add_tlv(p_hwfn, &p_iov->offset,
-		    CHANNEL_TLV_LIST_END, sizeof(struct channel_list_end_tlv));
-
-	resp = &p_iov->pf2vf_reply->default_resp;
-	rc = qed_send_msg2pf(p_hwfn, &resp->hdr.status, sizeof(*resp));
-
-	if (!rc && resp->hdr.status != PFVF_STATUS_SUCCESS)
-		rc = -EAGAIN;
-
-	qed_vf_pf_req_end(p_hwfn, rc);
-
-	p_hwfn->b_int_enabled = 0;
-
-	if (p_iov->vf2pf_request)
-		dma_free_coherent(&p_hwfn->cdev->pdev->dev,
-				  sizeof(union vfpf_tlvs),
-				  p_iov->vf2pf_request,
-				  p_iov->vf2pf_request_phys);
-	if (p_iov->pf2vf_reply)
-		dma_free_coherent(&p_hwfn->cdev->pdev->dev,
-				  sizeof(union pfvf_tlvs),
-				  p_iov->pf2vf_reply, p_iov->pf2vf_reply_phys);
-
-	if (p_iov->bulletin.p_virt) {
-		size = sizeof(struct qed_bulletin_content);
-		dma_free_coherent(&p_hwfn->cdev->pdev->dev,
-				  size,
-				  p_iov->bulletin.p_virt, p_iov->bulletin.phys);
-	}
-
-	kfree(p_hwfn->vf_iov_info);
-	p_hwfn->vf_iov_info = NULL;
-
-	return rc;
-}
-
 void qed_vf_pf_filter_mcast(struct qed_hwfn *p_hwfn,
 			    struct qed_filter_mcast *p_filter_cmd)
 {

drivers/net/ethernet/qlogic/qed/qed_vf.h

@@ -119,6 +119,12 @@ struct vfpf_acquire_tlv {
 	 * references.
 	 */
 #define VFPF_ACQUIRE_CAP_QUEUE_QIDS     BIT(2)
+
+	/* The VF is using the physical bar. While this is mostly internal
+	 * to the VF, might affect the number of CIDs supported assuming
+	 * QUEUE_QIDS is set.
+	 */
+#define VFPF_ACQUIRE_CAP_PHYSICAL_BAR   BIT(3)
 		u64 capabilities;
 		u8 fw_major;
 		u8 fw_minor;
@@ -202,7 +208,8 @@ struct pfvf_acquire_resp_tlv {
 		u16 chip_rev;
 		u8 dev_type;
 
-		u8 padding;
+		/* Doorbell bar size configured in HW: log(size) or 0 */
+		u8 bar_size;
 
 		struct pfvf_stats_info stats_info;
 
@@ -663,6 +670,11 @@ struct qed_vf_iov {
 	 * compatibility [with older PFs] we'd still need to store these.
 	 */
 	struct qed_sb_info *sbs_info[PFVF_MAX_SBS_PER_VF];
+
+	/* Determines whether VF utilizes doorbells via limited register
+	 * bar or via the doorbell bar.
+	 */
+	bool b_doorbell_bar;
 };
 
 #ifdef CONFIG_QED_SRIOV
@@ -971,6 +983,8 @@ void qed_iov_vf_task(struct work_struct *work);
 void qed_vf_set_vf_start_tunn_update_param(struct qed_tunnel_info *p_tun);
 int qed_vf_pf_tunnel_param_update(struct qed_hwfn *p_hwfn,
 				  struct qed_tunnel_info *p_tunn);
+
+u32 qed_vf_hw_bar_size(struct qed_hwfn *p_hwfn, enum BAR_ID bar_id);
 #else
 static inline void qed_vf_get_link_params(struct qed_hwfn *p_hwfn,
 					  struct qed_mcp_link_params *params)
@@ -1147,6 +1161,13 @@ static inline int qed_vf_pf_tunnel_param_update(struct qed_hwfn *p_hwfn,
 {
 	return -EINVAL;
 }
+
+static inline u32
+qed_vf_hw_bar_size(struct qed_hwfn  *p_hwfn,
+		   enum BAR_ID bar_id)
+{
+	return 0;
+}
 #endif
 
 #endif