cxgb4vf: Add code to provision T4 PCI-E SR-IOV Virtual Functions with hardware resources

Add code to provision T4 PCI-E SR-IOV Virtual Functions with hardware resources.

Signed-off-by Casey Leedom

---
 drivers/net/cxgb4/cxgb4_main.c |  106 
++++++++++++++++++++++++++++++++++++++++
 1 files changed, 106 insertions(+), 0 deletions(-)

drivers/net/cxgb4/cxgb4_main.c +97 -0

index 27f65b5..6528167 100644
@@ -77,6 +77,76 @@
  */
 #define MAX_SGE_TIMERVAL 200U
 
+#ifdef CONFIG_PCI_IOV
+/*
+ * Virtual Function provisioning constants.  We need two extra Ingress Queues
+ * with Interrupt capability to serve as the VF's Firmware Event Queue and
+ * Forwarded Interrupt Queue (when using MSI mode) -- neither will have Free
+ * Lists associated with them).  For each Ethernet/Control Egress Queue and
+ * for each Free List, we need an Egress Context.
+ */
+enum {
+	VFRES_NPORTS = 1,		/* # of "ports" per VF */
+	VFRES_NQSETS = 2,		/* # of "Queue Sets" per VF */
+
+	VFRES_NVI = VFRES_NPORTS,	/* # of Virtual Interfaces */
+	VFRES_NETHCTRL = VFRES_NQSETS,	/* # of EQs used for ETH or CTRL Qs */
+	VFRES_NIQFLINT = VFRES_NQSETS+2,/* # of ingress Qs/w Free List(s)/intr */
+	VFRES_NIQ = 0,			/* # of non-fl/int ingress queues */
+	VFRES_NEQ = VFRES_NQSETS*2,	/* # of egress queues */
+	VFRES_TC = 0,			/* PCI-E traffic class */
+	VFRES_NEXACTF = 16,		/* # of exact MPS filters */
+
+	VFRES_R_CAPS = FW_CMD_CAP_DMAQ|FW_CMD_CAP_VF|FW_CMD_CAP_PORT,
+	VFRES_WX_CAPS = FW_CMD_CAP_DMAQ|FW_CMD_CAP_VF,
+};
+
+/*
+ * Provide a Port Access Rights Mask for the specified PF/VF.  This is very
+ * static and likely not to be useful in the long run.  We really need to
+ * implement some form of persistent configuration which the firmware
+ * controls.
+ */
+static unsigned int pfvfres_pmask(struct adapter *adapter,
+				  unsigned int pf, unsigned int vf)
+{
+	unsigned int portn, portvec;
+
+	/*
+	 * Give PF's access to all of the ports.
+	 */
+	if (vf == 0)
+		return FW_PFVF_CMD_PMASK_MASK;
+
+	/*
+	 * For VFs, we'll assign them access to the ports based purely on the
+	 * PF.  We assign active ports in order, wrapping around if there are
+	 * fewer active ports than PFs: e.g. active port[pf % nports].
+	 * Unfortunately the adapter's port_info structs haven't been
+	 * initialized yet so we have to compute this.
+	 */
+	if (adapter->params.nports == 0)
+		return 0;
+
+	portn = pf % adapter->params.nports;
+	portvec = adapter->params.portvec;
+	for (;;) {
+		/*
+		 * Isolate the lowest set bit in the port vector.  If we're at
+		 * the port number that we want, return that as the pmask.
+		 * otherwise mask that bit out of the port vector and
+		 * decrement our port number ...
+		 */
+		unsigned int pmask = portvec ^ (portvec & (portvec-1));
+		if (portn == 0)
+			return pmask;
+		portn--;
+		portvec &= ~pmask;
+	}
+	/*NOTREACHED*/
+}
+#endif
+
 enum {
 	MEMWIN0_APERTURE = 65536,
 	MEMWIN0_BASE     = 0x30000,

@@ -2925,6 +2995,42 @@ static int adap_init0(struct adapter *adap)
 	t4_read_mtu_tbl(adap, adap->params.mtus, NULL);
 	t4_load_mtus(adap, adap->params.mtus, adap->params.a_wnd,
 		     adap->params.b_wnd);
+
+#ifdef CONFIG_PCI_IOV
+	/*
+	 * Provision resource limits for Virtual Functions.  We currently
+	 * grant them all the same static resource limits except for the Port
+	 * Access Rights Mask which we're assigning based on the PF.  All of
+	 * the static provisioning stuff for both the PF and VF really needs
+	 * to be managed in a persistent manner for each device which the
+	 * firmware controls.
+	 */
+	{
+		int pf, vf;
+
+		for (pf = 0; pf < ARRAY_SIZE(num_vf); pf++) {
+			if (num_vf[pf] <= 0)
+				continue;
+
+			/* VF numbering starts at 1! */
+			for (vf = 1; vf <= num_vf[pf]; vf++) {
+				ret = t4_cfg_pfvf(adap, 0, pf, vf,
+						  VFRES_NEQ, VFRES_NETHCTRL,
+						  VFRES_NIQFLINT, VFRES_NIQ,
+						  VFRES_TC, VFRES_NVI,
+						  FW_PFVF_CMD_CMASK_MASK,
+						  pfvfres_pmask(adap, pf, vf),
+						  VFRES_NEXACTF,
+						  VFRES_R_CAPS, VFRES_WX_CAPS);

| From: Simon Horman <[email protected]>
| Date: Saturday, June 26, 2010 06:37 am
| 
| I wonder if it would be cleaner to move the guts of the last hunk
| into a function (e.g. adap_init_sriov()) and have that be a dummy
| function in the case that CONFIG_PCI_IOV in the first hunk is not set.

I have no objection to this. I'd like to do minor tuning work like
that as a follow on rather than respin the patch. But, as I said in my submission, I'm not familiar with this process so if making changes like the above is required I'll definitely jump on it. I don't know what issues are "critical show stoppers" and which ones are "nice to have" once things are in place.