FWARC 2008/540 Memory DR Domain Service

FWARC 2009/300 CPU MD node property for real address bits
6720954 add memory dr feature to ldoms
6742779 fsflush_do_pages() may incorrectly skip constituent file large pages

1 files changed, 184 insertions, 4 deletions

diff --git a/usr/src/uts/sun4v/promif/promif_emul.c b/usr/src/uts/sun4v/promif/promif_emul.c
index 192ba5e2b3..97268b4777 100644
--- a/usr/src/uts/sun4v/promif/promif_emul.c
+++ b/usr/src/uts/sun4v/promif/promif_emul.c
@@ -20,12 +20,10 @@
  */
 
 /*
- * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
+ * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
  * Use is subject to license terms.
  */
 
-#pragma ident	"%Z%%M%	%I%	%E% SMI"
-
 #include <sys/promif_impl.h>
 #include <sys/machsystm.h>
 #include <sys/lpad.h>
@@ -37,19 +35,39 @@
 #include <sys/hypervisor_api.h>
 #include <sys/mdesc.h>
 #include <sys/mach_descrip.h>
+#include <sys/cpu_module.h>
 
 #ifndef _KMDB
+#include <sys/pte.h>
+#include <vm/hat_sfmmu.h>
+#include <sys/memlist_impl.h>
+
 static processorid_t cif_cpu;
 static struct translation *cif_prom_trans;
 static size_t cif_prom_ntrans;
 
 int cif_cpu_mp_ready;
 int (*prom_cif_handler)(void *) = NULL;
+
+extern struct memlist *phys_avail;
+extern struct vnode prom_ppages;
+extern void kdi_tlb_page_unlock(caddr_t, int);
+
+#define	COMBINE(hi, lo) (((uint64_t)(uint32_t)(hi) << 32) | (uint32_t)(lo))
+#define	OFW_PT_START_ADDR	0xfffffffc00000000	/* OBP PT start */
+#define	OFW_PT_END_ADDR		0xffffffffffffffff	/* OBP PT end */
+
+#define	PROM_ADDR(a)	(((a) >= OFW_START_ADDR && (a) <= OFW_END_ADDR) || \
+			((a) >= OFW_PT_START_ADDR && (a) <= OFW_PT_END_ADDR))
 #endif
 
 #ifdef DEBUG
 uint_t cif_debug;
-#endif /* DEBUG */
+int prom_free_debug;
+#define	PMFREE_DEBUG(args...) if (prom_free_debug) printf(args)
+#else
+#define	PMFREE_DEBUG(args...)
+#endif
 
 extern int (*cif_handler)(void *);
 
@@ -171,6 +189,161 @@ cif_init(char *pgmname, caddr_t root, ihandle_t in, ihandle_t out,
 
 #else
 
+static struct translation *
+read_prom_mappings(size_t *ntransp)
+{
+	char *prop = "translations";
+	pnode_t node;
+	size_t translen;
+	ihandle_t immu;
+	struct translation *transroot;
+
+	*ntransp = 0;
+
+	/*
+	 * the "translations" property is associated with the mmu node
+	 */
+	if ((immu = prom_mmu_ihandle()) == (ihandle_t)-1) {
+		PMFREE_DEBUG("no mmu ihandle");
+		return (NULL);
+	}
+	node = (pnode_t)prom_getphandle(immu);
+	if (node == OBP_NONODE || node == OBP_BADNODE) {
+		PMFREE_DEBUG("no mmu node");
+		return (NULL);
+	}
+
+	if ((translen = prom_getproplen(node, prop)) == -1) {
+		PMFREE_DEBUG("no translations property");
+		return (NULL);
+	}
+	transroot = (struct translation *)kmem_zalloc(translen, KM_SLEEP);
+
+	if (prom_getprop(node, prop, (caddr_t)transroot) == -1) {
+		PMFREE_DEBUG("translations getprop failed");
+		kmem_free(transroot, translen);
+		return (NULL);
+	}
+	*ntransp = translen / sizeof (*transroot);
+
+	return (transroot);
+}
+
+static void
+unmap_prom_mappings(struct translation *transroot, size_t ntransroot)
+{
+	int i, j, rv;
+	int npgs, nunmapped, nfreed, nskipped;
+	char *p;
+	tte_t tte;
+	pfn_t pfn;
+	page_t *pp;
+	uint64_t vaddr;
+	struct translation *promt;
+	cpuset_t other_cpus;
+
+	/*
+	 * During startup isa_list is allocated in OBP address space
+	 * so it needs to be re-allocated in kernel address space
+	 * before OBP memory is unmapped.
+	 *
+	 * see cpu_setup_common().
+	 */
+	p = kmem_zalloc(strlen(isa_list) + 1, KM_SLEEP);
+	(void) strcpy(p, isa_list);
+	isa_list = p;
+
+	nfreed = 0;
+	nunmapped = 0;
+	nskipped = 0;
+
+	for (i = 0, promt = transroot; i < ntransroot; i++, promt++) {
+		ASSERT(promt->tte_hi != 0);
+		ASSERT32(promt->virt_hi == 0 && promt->size_hi == 0);
+
+		vaddr = COMBINE(promt->virt_hi, promt->virt_lo);
+
+		if (!PROM_ADDR(vaddr)) {
+			nskipped++;
+			continue;
+		}
+
+		npgs = mmu_btopr(COMBINE(promt->size_hi, promt->size_lo));
+
+		if (npgs > 1) {
+			PMFREE_DEBUG("large trans vaddr=0x%lx, npgs=%d\n",
+			    vaddr, npgs);
+		}
+		for (j = 0; j < npgs; j++) {
+
+			pfn = sfmmu_vatopfn((caddr_t)vaddr, KHATID, &tte);
+
+			if (pfn == PFN_INVALID) {
+				tte.tte_inthi = promt->tte_hi;
+				tte.tte_intlo = promt->tte_lo;
+				pfn = TTE_TO_PFN((caddr_t)COMBINE(
+				    promt->virt_hi, promt->virt_lo), &tte);
+				PMFREE_DEBUG(
+				    "no mapping for vaddr=0x%lx (opfn=0x%lx)\n",
+				    vaddr, pfn);
+				break;
+			}
+			ASSERT(!TTE_IS_LOCKED(&tte));
+			ASSERT(TTE_IS_8K(&tte));
+
+			/*
+			 * Unload the current mapping for the page and
+			 * if it is the last mapping, free the page.
+			 */
+			pp = page_numtopp_nolock(pfn);
+			PMFREE_DEBUG("unmap vaddr=0x%lx pfn=0x%lx pp=0x%p",
+			    vaddr, pfn, (void *)pp);
+			ASSERT(pp);
+			ASSERT(PAGE_EXCL(pp));
+			ASSERT(PP_ISNORELOC(pp));
+			ASSERT(!PP_ISFREE(pp));
+			ASSERT(page_find(&prom_ppages, pfn));
+			ASSERT(page_get_pagecnt(pp->p_szc) == 1);
+
+			hat_unload(kas.a_hat, (caddr_t)vaddr, PAGESIZE,
+			    HAT_UNLOAD_UNLOCK);
+
+			if (pp->p_mapping) {
+				PMFREE_DEBUG(" skip\n");
+			} else {
+				PP_CLRNORELOC(pp);
+				page_destroy(pp, 0);
+				memlist_write_lock();
+				rv = memlist_add_span(pfn << PAGESHIFT,
+				    PAGESIZE, &phys_avail);
+				ASSERT(rv == MEML_SPANOP_OK);
+				memlist_write_unlock();
+				PMFREE_DEBUG(" free\n");
+				nfreed++;
+			}
+			nunmapped++;
+			vaddr += PAGESIZE;
+		}
+	}
+
+	if (transroot) {
+		PMFREE_DEBUG("nunmapped=%d nfreed=%d nskipped=%d\n",
+		    nunmapped, nfreed, nskipped);
+		kmem_free(transroot, ntransroot * sizeof (*transroot));
+	}
+
+	/*
+	 * Unload OBP permanent mappings.
+	 */
+	kdi_tlb_page_unlock((caddr_t)OFW_START_ADDR, 1);
+	kpreempt_disable();
+	other_cpus = cpu_ready_set;
+	CPUSET_DEL(other_cpus, CPU->cpu_id);
+	xt_some(other_cpus, vtag_unmap_perm_tl1, (uint64_t)OFW_START_ADDR,
+	    KCONTEXT);
+	kpreempt_enable();
+}
+
 static void cache_prom_data(void);
 
 /*
@@ -252,6 +425,8 @@ cif_init(void)
 	void (*kmdb_cb)(void);
 	uint64_t rtba;
 	uint64_t rv;
+	size_t ntransroot;
+	struct translation *transroot;
 
 	/*
 	 * Check if domaining is enabled. If not, do not
@@ -260,6 +435,8 @@ cif_init(void)
 	if (!domaining_enabled())
 		return;
 
+	transroot = read_prom_mappings(&ntransroot);
+
 	/*
 	 * Cache PROM data that is needed later, e.g. a shadow
 	 * copy of the device tree, IO mappings, etc.
@@ -301,6 +478,9 @@ cif_init(void)
 	}
 
 	cif_check_cpus();
+
+	if (transroot != NULL)
+		unmap_prom_mappings(transroot, ntransroot);
 }
 
 static void