Add code to initialize CPUs on sun4v.

3 files changed, 184 insertions, 28 deletions

diff --git a/sys/arch/sparc64/include/cpu.h b/sys/arch/sparc64/include/cpu.h
index 46e45da2b06..a80575ebd1b 100644
--- a/sys/arch/sparc64/include/cpu.h
+++ b/sys/arch/sparc64/include/cpu.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: cpu.h,v 1.57 2008/03/17 23:10:21 kettenis Exp $	*/
+/*	$OpenBSD: cpu.h,v 1.58 2008/03/23 23:46:21 kettenis Exp $	*/
 /*	$NetBSD: cpu.h,v 1.28 2001/06/14 22:56:58 thorpej Exp $ */
 
 /*
@@ -133,6 +133,18 @@ struct cpu_info {
 	void			(*ci_spinup)(void); /* spinup routine */
 	void			*ci_initstack;
 	paddr_t			ci_paddr;	/* Phys addr of this structure. */
+
+#ifdef SUN4V
+	struct rwindow64	ci_rw;
+	u_int64_t		ci_rwsp;
+
+	paddr_t			ci_mmfsa;
+	paddr_t			ci_cpumq;
+	paddr_t			ci_devmq;
+
+	paddr_t			ci_cpuset;
+	paddr_t			ci_mondo;
+#endif
 };
 
 #define CPUF_RUNNING	0x0001		/* CPU is running */
diff --git a/sys/arch/sparc64/sparc64/cpu.c b/sys/arch/sparc64/sparc64/cpu.c
index 545b146b1bf..7331682c048 100644
--- a/sys/arch/sparc64/sparc64/cpu.c
+++ b/sys/arch/sparc64/sparc64/cpu.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: cpu.c,v 1.35 2008/03/16 22:22:15 kettenis Exp $	*/
+/*	$OpenBSD: cpu.c,v 1.36 2008/03/23 23:46:21 kettenis Exp $	*/
 /*	$NetBSD: cpu.c,v 1.13 2001/05/26 21:27:15 chs Exp $ */
 
 /*
@@ -62,6 +62,7 @@
 #include <machine/cpu.h>
 #include <machine/reg.h>
 #include <machine/trap.h>
+#include <machine/hypervisor.h>
 #include <machine/openfirm.h>
 #include <machine/pmap.h>
 #include <machine/sparc64.h>
@@ -78,29 +79,28 @@ struct cacheinfo cacheinfo = {
 /* Linked list of all CPUs in system. */
 struct cpu_info *cpus = NULL;
 
-struct cpu_info *alloc_cpuinfo(int);
+struct cpu_info *alloc_cpuinfo(struct mainbus_attach_args *);
 
 /* The following are used externally (sysctl_hw). */
 char	machine[] = MACHINE;		/* from <machine/param.h> */
 char	cpu_model[100];
 
-void cpu_hatch(void);
-
 /* The CPU configuration driver. */
-static void cpu_attach(struct device *, struct device *, void *);
-int  cpu_match(struct device *, void *, void *);
+int cpu_match(struct device *, void *, void *);
+void cpu_attach(struct device *, struct device *, void *);
 
 struct cfattach cpu_ca = {
 	sizeof(struct device), cpu_match, cpu_attach
 };
 
-extern struct cfdriver cpu_cd;
+void cpu_init(struct cpu_info *ci);
+void cpu_hatch(void);
 
 #define	IU_IMPL(v)	((((u_int64_t)(v))&VER_IMPL) >> VER_IMPL_SHIFT)
 #define	IU_VERS(v)	((((u_int64_t)(v))&VER_MASK) >> VER_MASK_SHIFT)
 
 struct cpu_info *
-alloc_cpuinfo(int node)
+alloc_cpuinfo(struct mainbus_attach_args *ma)
 {
 	paddr_t pa0, pa;
 	vaddr_t va, va0;
@@ -109,9 +109,11 @@ alloc_cpuinfo(int node)
 	struct cpu_info *cpi, *ci;
 	extern paddr_t cpu0paddr;
 
-	portid = getpropint(node, "portid", -1);
+	portid = getpropint(ma->ma_node, "portid", -1);
 	if (portid == -1)
-		portid = getpropint(node, "upa-portid", -1);
+		portid = getpropint(ma->ma_node, "upa-portid", -1);
+	if (portid == -1 && ma->ma_nreg > 0)
+		portid = (ma->ma_reg[0].ur_paddr >> 32) & 0x0fffffff;
 	if (portid == -1)
 		panic("alloc_cpuinfo: portid");
 
@@ -153,10 +155,13 @@ alloc_cpuinfo(int node)
 	cpi->ci_spinup = NULL;
 #endif
 
-	cpi->ci_initstack = (void *)EINTSTACK;
+	cpi->ci_initstack = cpi;
 	cpi->ci_paddr = pa0;
+#ifdef SUN4V
+	cpi->ci_mmfsa = pa0;
+#endif
 	cpi->ci_self = cpi;
-	cpi->ci_node = node;
+	cpi->ci_node = ma->ma_node;
 
 	sched_init_cpu(cpi);
 
@@ -195,6 +200,8 @@ cpu_match(parent, vcf, aux)
 		portid = getpropint(ma->ma_node, "portid", -1);
 	if (portid == -1)
 		portid = getpropint(ma->ma_node, "cpuid", -1);
+	if (portid == -1 && ma->ma_nreg > 0)
+		portid = (ma->ma_reg[0].ur_paddr >> 32) & 0xff;
 	if (portid == -1)
 		return (0);
 
@@ -210,7 +217,7 @@ cpu_match(parent, vcf, aux)
  * Discover interesting goop about the virtual address cache
  * (slightly funny place to do it, but this is where it is to be found).
  */
-static void
+void
 cpu_attach(parent, dev, aux)
 	struct device *parent;
 	struct device *dev;
@@ -223,10 +230,11 @@ cpu_attach(parent, dev, aux)
 	struct cpu_info *ci;
 	const char *sep;
 	register int i, l;
-	u_int64_t ver;
+	u_int64_t ver = 0;
 	extern u_int64_t cpu_clockrate[];
 
-	ver = getver();
+	if (CPU_ISSUN4U)
+		ver = getver();
 	impl = IU_IMPL(ver);
 	vers = IU_VERS(ver);
 
@@ -236,7 +244,7 @@ cpu_attach(parent, dev, aux)
 	/*
 	 * Allocate cpu_info structure if needed.
 	 */
-	ci = alloc_cpuinfo(node);
+	ci = alloc_cpuinfo(ma);
 
 	clk = getpropint(node, "clock-frequency", 0);
 	if (clk == 0) {
@@ -253,6 +261,9 @@ cpu_attach(parent, dev, aux)
 	    ma->ma_name, vers >> 4, vers & 0xf, clockfreq(clk));
 	printf(": %s\n", cpu_model);
 
+	if (ci->ci_upaid == cpu_myid())
+		cpu_init(ci);
+
 	cacheinfo.c_physical = 1; /* Dunno... */
 	cacheinfo.c_split = 1;
 	l = getpropint(node, "icache-line-size", 0);
@@ -370,6 +381,15 @@ cpu_myid(void)
 {
 	char buf[32];
 
+#ifdef SUN4V
+	if (CPU_ISSUN4V) {
+		uint64_t myid;
+
+		hv_cpu_myid(&myid);
+		return myid;
+	}
+#endif
+
 	if (OF_getprop(findroot(), "name", buf, sizeof(buf)) > 0 &&
 	    strcmp(buf, "SUNW,Ultra-Enterprise-10000") == 0)
 		return lduwa(0x1fff40000d0UL, ASI_PHYS_NON_CACHED);
@@ -377,6 +397,42 @@ cpu_myid(void)
 		return CPU_UPAID;
 }
 
+void
+cpu_init(struct cpu_info *ci)
+{
+#ifdef SUN4V
+	paddr_t pa = ci->ci_paddr;
+	int err;
+
+	if (CPU_ISSUN4U)
+		return;
+
+#define MONDO_QUEUE_SIZE	32
+#define QUEUE_ENTRY_SIZE	64
+
+	pa += CPUINFO_VA - INTSTACK;
+	pa += PAGE_SIZE;
+
+	ci->ci_cpumq = pa;
+	err = hv_cpu_qconf(CPU_MONDO_QUEUE, ci->ci_cpumq, MONDO_QUEUE_SIZE);
+	if (err != H_EOK)
+		panic("Unable to set cpu mondo queue: %d", err);
+	pa += MONDO_QUEUE_SIZE * QUEUE_ENTRY_SIZE;
+
+	ci->ci_devmq = pa;
+	err = hv_cpu_qconf(DEVICE_MONDO_QUEUE, ci->ci_devmq, MONDO_QUEUE_SIZE);
+	if (err != H_EOK)
+		panic("Unable to set device mondo queue: %d", err);
+	pa += MONDO_QUEUE_SIZE * QUEUE_ENTRY_SIZE;
+
+	ci->ci_mondo = pa;
+	pa += 64;
+
+	ci->ci_cpuset = pa;
+	pa += 64;
+#endif
+}
+
 struct cfdriver cpu_cd = {
 	NULL, "cpu", DV_DULL
 };
@@ -405,7 +461,7 @@ cpu_boot_secondary_processors(void)
 			continue;
 
 		cpuid = getpropint(ci->ci_node, "cpuid", -1);
-		if (cpuid == -1) {
+		if (CPU_ISSUN4U && cpuid == -1) {
 			prom_start_cpu(ci->ci_node,
 			    (void *)cpu_mp_startup, ci->ci_paddr);
 		} else {
@@ -425,13 +481,16 @@ cpu_boot_secondary_processors(void)
 void
 cpu_hatch(void)
 {
+	struct cpu_info *ci = curcpu();
 	int s;
 
-	curcpu()->ci_flags |= CPUF_RUNNING;
+	cpu_init(ci);
+
+	ci->ci_flags |= CPUF_RUNNING;
 	sparc_membar(Sync);
 
 	s = splhigh();
-	microuptime(&curcpu()->ci_schedstate.spc_runtime);
+	microuptime(&ci->ci_schedstate.spc_runtime);
 	splx(s);
 
 	tick_start();
diff --git a/sys/arch/sparc64/sparc64/pmap.c b/sys/arch/sparc64/sparc64/pmap.c
index a45bd55dd2b..fa3dfa9dab1 100644
--- a/sys/arch/sparc64/sparc64/pmap.c
+++ b/sys/arch/sparc64/sparc64/pmap.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: pmap.c,v 1.54 2008/03/22 16:01:32 kettenis Exp $	*/
+/*	$OpenBSD: pmap.c,v 1.55 2008/03/23 23:46:21 kettenis Exp $	*/
 /*	$NetBSD: pmap.c,v 1.107 2001/08/31 16:47:41 eeh Exp $	*/
 #undef	NO_VCACHE /* Don't forget the locked TLB in dostart */
 /*
@@ -43,6 +43,7 @@
 #include <machine/pcb.h>
 #include <machine/sparc64.h>
 #include <machine/ctlreg.h>
+#include <machine/hypervisor.h>
 #include <machine/openfirm.h>
 #include <machine/kcore.h>
 
@@ -235,6 +236,10 @@ int tsbsize;		/* tsbents = 512 * 2^^tsbsize */
  */
 #define TSB_TAG_INVALID (~0LL)
 
+#ifdef SUN4V
+struct tsb_desc *tsb_desc;
+#endif
+
 struct pmap kernel_pmap_;
 
 extern int physmem;
@@ -1044,6 +1049,19 @@ remap_data:
 	BDPRINTF(PDB_BOOT1, ("TSB allocated at %p size %08x\r\n", (void *)tsb_dmmu,
 	    (int)TSBSIZE));
 
+#ifdef SUN4V
+	if (CPU_ISSUN4V) {
+		valloc(tsb_desc, struct tsb_desc, sizeof(struct tsb_desc));
+		bzero(tsb_desc, sizeof(struct tsb_desc));
+		tsb_desc->td_idxpgsz = 0;
+		tsb_desc->td_assoc = 1;
+		tsb_desc->td_size = TSBENTS;
+		tsb_desc->td_ctxidx = -1;
+		tsb_desc->td_pgsz = 0xf;
+		tsb_desc->td_pa = (paddr_t)tsb_dmmu + kdatap - kdata;
+	}
+#endif
+
 	first_phys_addr = mem->start;
 	BDPRINTF(PDB_BOOT1, ("firstaddr after pmap=%08lx\r\n", 
 		(u_long)firstaddr));
@@ -1385,6 +1403,9 @@ remap_data:
 		cpus->ci_spinup = main; /* Call main when we're running. */
 		cpus->ci_initstack = (void *)u0[1];
 		cpus->ci_paddr = cpu0paddr;
+#ifdef SUN4V
+		cpus->ci_mmfsa = cpu0paddr;
+#endif
 		proc0paddr = cpus->ci_cpcb;
 
 		cpu0paddr += 64 * KB;
@@ -1407,11 +1428,23 @@ remap_data:
 	pmap_bootstrap_cpu(cpus->ci_paddr);
 }
 
-extern void sun4u_set_tsbs(void);
+void sun4u_bootstrap_cpu(paddr_t);
+void sun4v_bootstrap_cpu(paddr_t);
 
 void
 pmap_bootstrap_cpu(paddr_t intstack)
 {
+	if (CPU_ISSUN4V)
+		sun4v_bootstrap_cpu(intstack);
+	else
+		sun4u_bootstrap_cpu(intstack);
+}
+
+extern void sun4u_set_tsbs(void);
+
+void
+sun4u_bootstrap_cpu(paddr_t intstack)
+{
 	u_int64_t data;
 	paddr_t pa;
 	vaddr_t va;
@@ -1426,16 +1459,16 @@ pmap_bootstrap_cpu(paddr_t intstack)
 
 	index = 15; /* XXX */
 	for (va = ktext, pa = ktextp; va < ektext; va += 4*MEG, pa += 4*MEG) {
-		data = TSB_DATA(0, PGSZ_4M, pa, 1, 0, 1, FORCE_ALIAS, 1, 0);
-		data |= TLB_L;
+		data = SUN4U_TSB_DATA(0, PGSZ_4M, pa, 1, 0, 1, FORCE_ALIAS, 1, 0);
+		data |= SUN4U_TLB_L;
 		prom_itlb_load(index, data, va);
 		prom_dtlb_load(index, data, va);
 		index--;
 	}
 
 	for (va = kdata, pa = kdatap; va < ekdata; va += 4*MEG, pa += 4*MEG) {
-		data = TSB_DATA(0, PGSZ_4M, pa, 1, 1, 1, FORCE_ALIAS, 1, 0);
-		data |= TLB_L;
+		data = SUN4U_TSB_DATA(0, PGSZ_4M, pa, 1, 1, 1, FORCE_ALIAS, 1, 0);
+		data |= SUN4U_TLB_L;
 		prom_dtlb_load(index, data, va);
 		index--;
 	}
@@ -1444,13 +1477,65 @@ pmap_bootstrap_cpu(paddr_t intstack)
 	 * Establish the 64KB locked mapping for the interrupt stack.
 	 */
 
-	data = TSB_DATA(0, PGSZ_64K, intstack, 1, 1, 1, FORCE_ALIAS, 1, 0);
-	data |= TLB_L;
+	data = SUN4U_TSB_DATA(0, PGSZ_64K, intstack, 1, 1, 1, FORCE_ALIAS, 1, 0);
+	data |= SUN4U_TLB_L;
 	prom_dtlb_load(index, data, INTSTACK);
 
 	sun4u_set_tsbs();
 }
 
+void
+sun4v_bootstrap_cpu(paddr_t intstack)
+{
+#ifdef SUN4V
+	u_int64_t data;
+	paddr_t pa;
+	vaddr_t va;
+	int err;
+
+	/*
+	 * Establish the 4MB locked mappings for kernel data and text.
+	 *
+	 * The text segment needs to be mapped into the DTLB too,
+	 * because of .rodata.
+	 */
+
+	for (va = ktext, pa = ktextp; va < ektext; va += 4*MEG, pa += 4*MEG) {
+		data = SUN4V_TSB_DATA(0, PGSZ_4M, pa, 1, 0, 1, 0, 1, 0);
+		data |= SUN4V_TLB_X;
+		err = hv_mmu_map_perm_addr(va, data, MAP_ITLB|MAP_DTLB);
+		if (err != H_EOK)
+			prom_printf("err: %d\r\n", err);
+	}
+
+	for (va = kdata, pa = kdatap; va < ekdata; va += 4*MEG, pa += 4*MEG) {
+		data = SUN4V_TSB_DATA(0, PGSZ_4M, pa, 1, 1, 1, 0, 1, 0);
+		err = hv_mmu_map_perm_addr(va, data, MAP_DTLB);
+		if (err != H_EOK)
+			prom_printf("err: %d\r\n", err);
+	}
+
+#ifndef MULTIPROCESSOR
+	/*
+	 * Establish the 64KB locked mapping for the interrupt stack.
+	 */
+	data = SUN4V_TSB_DATA(0, PGSZ_64K, intstack, 1, 1, 1, 0, 1, 0);
+	err = hv_mmu_map_perm_addr(INTSTACK, data, MAP_DTLB);
+	if (err != H_EOK)
+		prom_printf("err: %d\r\n", err);
+#endif
+
+	stxa(0, ASI_SCRATCHPAD, intstack + (CPUINFO_VA - INTSTACK));
+
+	err = hv_mmu_tsb_ctx0(1, (paddr_t)tsb_desc + kdatap - kdata);
+	if (err != H_EOK)
+		prom_printf("err: %d\r\n", err);
+	err = hv_mmu_tsb_ctxnon0(1, (paddr_t)tsb_desc + kdatap - kdata);
+	if (err != H_EOK)
+		prom_printf("err: %d\r\n", err);
+#endif
+}
+
 /*
  * Initialize anything else for pmap handling.
  * Called during vm_init().