From: Cristiano Giuffrida <giuffrida@cs.vu.nl>
Date: Tue, 11 Mar 2014 23:13:37 +0000 (+0100)
Subject: vm: Improve live update support.
X-Git-Url: http://zhaoyanbai.com/repos/%22http:/www.isc.org/icons/zpipe.c?a=commitdiff_plain;h=63483e02e6a5924979f1c177070e18c56323a45c;p=minix.git

vm: Improve live update support.

Change-Id: I02da3ea32cd05c4ed84a6e199236e5df6e25cb60
---

diff --git a/minix/lib/libsys/vm_memctl.c b/minix/lib/libsys/vm_memctl.c
index 0637476cc..35d9b171e 100644
--- a/minix/lib/libsys/vm_memctl.c
+++ b/minix/lib/libsys/vm_memctl.c
@@ -21,7 +21,7 @@ vm_memctl(endpoint_t ep, int req, void** addr, size_t *len)
 	if(addr) {
 		*addr = m.VM_RS_CTL_ADDR;
 	}
-	if(addr) {
+	if(len) {
 		*len = m.VM_RS_CTL_LEN;
 	}
 
diff --git a/minix/servers/vm/exit.c b/minix/servers/vm/exit.c
index a62a7d490..dd0ee6d63 100644
--- a/minix/servers/vm/exit.c
+++ b/minix/servers/vm/exit.c
@@ -75,6 +75,10 @@ SANITYCHECK(SCL_FUNCTIONS);
 		printf("VM: unannounced VM_EXIT %d\n", msg->VME_ENDPOINT);
 		return EINVAL;
 	}
+	if(vmp->vm_flags & VMF_VM_INSTANCE) {
+	    vmp->vm_flags &= ~VMF_VM_INSTANCE;
+	    num_vm_instances--;
+	}
 
 	{
 		/* Free pagetable and pages allocated by pt code. */
diff --git a/minix/servers/vm/glo.h b/minix/servers/vm/glo.h
index d4c1e71b1..d808aa213 100644
--- a/minix/servers/vm/glo.h
+++ b/minix/servers/vm/glo.h
@@ -40,3 +40,4 @@ EXTERN  mem_type_t mem_type_anon,       /* anonymous memory */
 
 /* total number of memory pages */
 EXTERN int total_pages;
+EXTERN int num_vm_instances;
diff --git a/minix/servers/vm/main.c b/minix/servers/vm/main.c
index 0fd120372..3e20a2da7 100644
--- a/minix/servers/vm/main.c
+++ b/minix/servers/vm/main.c
@@ -515,6 +515,10 @@ void init_vm(void)
 
 	/* Initialize the structures for queryexit */
 	init_query_exit();
+
+	/* Mark VM instances. */
+	num_vm_instances = 1;
+	vmproc[VM_PROC_NR].vm_flags |= VMF_VM_INSTANCE;
 }
 
 /*===========================================================================*
diff --git a/minix/servers/vm/pagetable.c b/minix/servers/vm/pagetable.c
index e0fe60af2..342774d37 100644
--- a/minix/servers/vm/pagetable.c
+++ b/minix/servers/vm/pagetable.c
@@ -83,8 +83,7 @@ static struct {
 	phys_bytes phys;
 } sparepagedirs[SPAREPAGEDIRS];
 
-extern char _end;	
-#define is_staticaddr(v) ((vir_bytes) (v) < (vir_bytes) &_end)
+#define is_staticaddr(v) ((vir_bytes) (v) < VM_OWN_HEAPSTART)
 
 #define MAX_KERNMAPPINGS 10
 static struct {
@@ -154,10 +153,8 @@ static u32_t findhole(int pages)
 	u32_t holev = NO_MEM;
 	int holesize = -1;
 
-	vmin = (vir_bytes) (&_end); /* marks end of VM BSS */
-	vmin += 1024*1024*1024;	/* reserve 1GB virtual address space for VM heap */
-	vmin &= ARCH_VM_ADDR_MASK; 
-	vmax = vmin + 100 * 1024 * 1024; /* allow 100MB of address space for VM */
+	vmin = VM_OWN_MMAPBASE;
+	vmax = VM_OWN_MMAPTOP;
 
 	/* Input sanity check. */
 	assert(vmin + VM_PAGE_SIZE >= vmin);
@@ -678,8 +675,7 @@ int pt_map_in_range(struct vmproc *src_vmp, struct vmproc *dst_vmp,
 int pt_ptmap(struct vmproc *src_vmp, struct vmproc *dst_vmp)
 {
 /* Transfer mappings to page dir and page tables from source process and
- * destination process. Make sure all the mappings are above the stack, not
- * to corrupt valid mappings in the data segment of the destination process.
+ * destination process.
  */
 	int pde, r;
 	phys_bytes physaddr;
diff --git a/minix/servers/vm/proto.h b/minix/servers/vm/proto.h
index b96c31e27..79d3ac4c6 100644
--- a/minix/servers/vm/proto.h
+++ b/minix/servers/vm/proto.h
@@ -47,7 +47,9 @@ int vm_isokendpt(endpoint_t ep, int *proc);
 int get_stack_ptr(int proc_nr, vir_bytes *sp);
 int do_info(message *);
 int swap_proc_slot(struct vmproc *src_vmp, struct vmproc *dst_vmp);
-int swap_proc_dyn_data(struct vmproc *src_vmp, struct vmproc *dst_vmp);
+int swap_proc_dyn_data(struct vmproc *src_vmp, struct vmproc *dst_vmp,
+    int sys_upd_flags);
+void adjust_proc_refs(void);
 int do_getrusage(message *m);
 
 /* exit.c */
@@ -149,8 +151,8 @@ int map_unmap_region(struct vmproc *vmp, struct vir_region *vr,
 int map_unmap_range(struct vmproc *vmp, vir_bytes, vir_bytes);
 int map_free_proc(struct vmproc *vmp);
 int map_proc_copy(struct vmproc *dst, struct vmproc *src);
-int map_proc_copy_from(struct vmproc *dst, struct vmproc *src, struct
-	vir_region *start_src_vr);
+int map_proc_copy_range(struct vmproc *dst, struct vmproc *src, struct
+	vir_region *start_src_vr, struct vir_region *end_src_vr);
 struct vir_region *map_lookup(struct vmproc *vmp, vir_bytes addr,
 	struct phys_region **pr);
 int map_pf(struct vmproc *vmp, struct vir_region *region, vir_bytes
@@ -172,10 +174,7 @@ void physblock_set(struct vir_region *region, vir_bytes offset,
 int map_ph_writept(struct vmproc *vmp, struct vir_region *vr,
         struct phys_region *pr);
 
-struct vir_region * map_region_lookup_tag(struct vmproc *vmp, u32_t
-	tag);
-void map_region_set_tag(struct vir_region *vr, u32_t tag);
-u32_t map_region_get_tag(struct vir_region *vr);
+struct vir_region* map_region_lookup_type(struct vmproc *vmp, u32_t flags);
 int map_get_phys(struct vmproc *vmp, vir_bytes addr, phys_bytes *r);
 int map_get_ref(struct vmproc *vmp, vir_bytes addr, u8_t *cnt);
 unsigned int physregions(struct vir_region *vr);
diff --git a/minix/servers/vm/region.c b/minix/servers/vm/region.c
index 84b308698..5816b747c 100644
--- a/minix/servers/vm/region.c
+++ b/minix/servers/vm/region.c
@@ -933,28 +933,30 @@ int map_proc_copy(struct vmproc *dst, struct vmproc *src)
 /* Copy all the memory regions from the src process to the dst process. */
 	region_init(&dst->vm_regions_avl);
 
-	return map_proc_copy_from(dst, src, NULL);
+	return map_proc_copy_range(dst, src, NULL, NULL);
 }
 
 /*========================================================================*
- *			     map_proc_copy_from			     	  *
+ *			     map_proc_copy_range			     	  *
  *========================================================================*/
-int map_proc_copy_from(struct vmproc *dst, struct vmproc *src,
-	struct vir_region *start_src_vr)
+int map_proc_copy_range(struct vmproc *dst, struct vmproc *src,
+	struct vir_region *start_src_vr, struct vir_region *end_src_vr)
 {
 	struct vir_region *vr;
 	region_iter v_iter;
 
 	if(!start_src_vr)
 		start_src_vr = region_search_least(&src->vm_regions_avl);
+	if(!end_src_vr)
+		end_src_vr = region_search_greatest(&src->vm_regions_avl);
 
-	assert(start_src_vr);
+	assert(start_src_vr && end_src_vr);
 	assert(start_src_vr->parent == src);
 	region_start_iter(&src->vm_regions_avl, &v_iter,
 		start_src_vr->vaddr, AVL_EQUAL);
 	assert(region_get_iter(&v_iter) == start_src_vr);
 
-	/* Copy source regions after the destination's last region (if any). */
+	/* Copy source regions into the destination. */
 
 	SANITYCHECK(SCL_FUNCTIONS);
 
@@ -982,6 +984,9 @@ int map_proc_copy_from(struct vmproc *dst, struct vmproc *src,
 		}
 	}
 #endif
+		if(vr == end_src_vr) {
+			break;
+		}
 		region_incr_iter(&v_iter);
 	}
 
@@ -1290,6 +1295,26 @@ int map_unmap_range(struct vmproc *vmp, vir_bytes unmap_start, vir_bytes length)
 
 }
 
+/*========================================================================*
+ *			  map_region_lookup_type			  *
+ *========================================================================*/
+struct vir_region* map_region_lookup_type(struct vmproc *vmp, u32_t type)
+{
+	struct vir_region *vr;
+	struct phys_region *pr;
+	vir_bytes used = 0, weighted = 0;
+	region_iter v_iter;
+	region_start_iter_least(&vmp->vm_regions_avl, &v_iter);
+
+	while((vr = region_get_iter(&v_iter))) {
+		region_incr_iter(&v_iter);
+		if(vr->flags & type)
+			return vr;
+	}
+
+	return NULL;
+}
+
 /*========================================================================*
  *				map_get_phys				  *
  *========================================================================*/
diff --git a/minix/servers/vm/region.h b/minix/servers/vm/region.h
index 29982abbd..934ec79b6 100644
--- a/minix/servers/vm/region.h
+++ b/minix/servers/vm/region.h
@@ -76,9 +76,10 @@ typedef struct vir_region {
 /* Mapping type: */
 #define VR_ANON		0x100	/* Memory to be cleared and allocated */
 #define VR_DIRECT	0x200	/* Mapped, but not managed by VM */
+#define VR_PREALLOC_MAP	0x400   /* Preallocated map. */
 
-/* map_page_region flags */
-#define MF_PREALLOC	0x01
+/* map_page_region_flags */
+#define MF_PREALLOC    0x01
 
 #endif
 
diff --git a/minix/servers/vm/rs.c b/minix/servers/vm/rs.c
index 3587e8617..067eb9331 100644
--- a/minix/servers/vm/rs.c
+++ b/minix/servers/vm/rs.c
@@ -16,6 +16,8 @@
 #include <minix/bitmap.h>
 #include <minix/rs.h>
 
+#include <sys/mman.h>
+
 #include <errno.h>
 #include <string.h>
 #include <env.h>
@@ -77,6 +79,7 @@ int do_rs_update(message *m_ptr)
 	src_e = m_ptr->m_lsys_vm_update.src;
 	dst_e = m_ptr->m_lsys_vm_update.dst;
         sys_upd_flags = m_ptr->m_lsys_vm_update.flags;
+        reply_e = m_ptr->m_source;
 
 	/* Lookup slots for source and destination process. */
 	if(vm_isokendpt(src_e, &src_p) != OK) {
@@ -90,6 +93,14 @@ int do_rs_update(message *m_ptr)
 	}
 	dst_vmp = &vmproc[dst_p];
 
+	/* Check flags. */
+	if((sys_upd_flags & (SF_VM_ROLLBACK|SF_VM_NOMMAP)) == 0) {
+	        /* Can't preallocate when transfering mmapped regions. */
+	        if(map_region_lookup_type(dst_vmp, VR_PREALLOC_MAP)) {
+			return ENOSYS;
+	        }
+	}
+
 	/* Let the kernel do the update first. */
 	r = sys_update(src_e, dst_e,
 	    sys_upd_flags & SF_VM_ROLLBACK ? SYS_UPD_ROLLBACK : 0);
@@ -102,21 +113,22 @@ int do_rs_update(message *m_ptr)
 	if(r != OK) {
 		return r;
 	}
-	r = swap_proc_dyn_data(src_vmp, dst_vmp);
+	r = swap_proc_dyn_data(src_vmp, dst_vmp, sys_upd_flags);
 	if(r != OK) {
 		return r;
 	}
 	pt_bind(&src_vmp->vm_pt, src_vmp);
 	pt_bind(&dst_vmp->vm_pt, dst_vmp);
 
-	/* Reply, update-aware. */
-	reply_e = m_ptr->m_source;
-	if(reply_e == src_e) reply_e = dst_e;
-	else if(reply_e == dst_e) reply_e = src_e;
-	m_ptr->m_type = OK;
-	r = ipc_send(reply_e, m_ptr);
-	if(r != OK) {
-		panic("ipc_send() error");
+	/* Reply in case of external request, update-aware. */
+	if(reply_e != VM_PROC_NR) {
+            if(reply_e == src_e) reply_e = dst_e;
+            else if(reply_e == dst_e) reply_e = src_e;
+            m_ptr->m_type = OK;
+            r = ipc_send(reply_e, m_ptr);
+            if(r != OK) {
+                    panic("ipc_send() error");
+            }
 	}
 
 	return SUSPEND;
@@ -134,6 +146,17 @@ static int rs_memctl_make_vm_instance(struct vmproc *new_vm_vmp)
 
 	this_vm_vmp = &vmproc[VM_PROC_NR];
 
+	/* Check if the operation is allowed. */
+	assert(num_vm_instances == 1 || num_vm_instances == 2);
+	if(num_vm_instances == 2) {
+		printf("VM can currently support no more than 2 VM instances at the time.");
+		return EPERM;
+	}
+
+	/* Copy settings from current VM. */
+	new_vm_vmp->vm_flags |= VMF_VM_INSTANCE;
+	num_vm_instances++;
+
 	/* Pin memory for the new VM instance. */
 	r = map_pin_memory(new_vm_vmp);
 	if(r != OK) {
@@ -187,20 +210,26 @@ static int rs_memctl_heap_prealloc(struct vmproc *vmp,
 static int rs_memctl_map_prealloc(struct vmproc *vmp,
 	vir_bytes *addr, size_t *len)
 {
-#if 0
 	struct vir_region *vr;
+	vir_bytes base, top;
+	int is_vm;
+
 	if(*len <= 0) {
 		return EINVAL;
 	}
 	*len = CLICK_CEIL(*len);
 
-	if(!(vr = map_page_region(vmp, VM_DATATOP - *len, VM_DATATOP, *len,
-		MAP_NONE, VR_ANON|VR_WRITABLE|VR_CONTIG, MF_PREALLOC))) {
+	is_vm = (vmp->vm_endpoint == VM_PROC_NR);
+	base = is_vm ? VM_OWN_MMAPBASE : VM_MMAPBASE;
+	top = is_vm ? VM_OWN_MMAPTOP : VM_MMAPTOP;
+
+	if (!(vr = map_page_region(vmp, base, top, *len,
+	    VR_ANON|VR_WRITABLE|VR_UNINITIALIZED, MF_PREALLOC,
+	    &mem_type_anon))) {
 		return ENOMEM;
 	}
-	map_region_set_tag(vr, VRT_PREALLOC_MAP);
-	*addr = arch_map2vir(vmp, vr->vaddr);
-#endif
+	vr->flags |= VR_PREALLOC_MAP;
+	*addr = vr->vaddr;
 	return OK;
 }
 
@@ -210,19 +239,17 @@ static int rs_memctl_map_prealloc(struct vmproc *vmp,
 static int rs_memctl_get_prealloc_map(struct vmproc *vmp,
 	vir_bytes *addr, size_t *len)
 {
-#if 0
 	struct vir_region *vr;
 
-	vr = map_region_lookup_tag(vmp, VRT_PREALLOC_MAP);
+	vr = map_region_lookup_type(vmp, VR_PREALLOC_MAP);
 	if(!vr) {
 		*addr = 0;
 		*len = 0;
 	}
 	else {
-		*addr = arch_map2vir(vmp, vr->vaddr);
+		*addr = vr->vaddr;
 		*len = vr->length;
 	}
-#endif
 	return OK;
 }
 
diff --git a/minix/servers/vm/utility.c b/minix/servers/vm/utility.c
index 17edbe2c1..63a6217ec 100644
--- a/minix/servers/vm/utility.c
+++ b/minix/servers/vm/utility.c
@@ -18,6 +18,7 @@
 #include <minix/syslib.h>
 #include <minix/type.h>
 #include <minix/bitmap.h>
+#include <minix/rs.h>
 #include <string.h>
 #include <errno.h>
 #include <env.h>
@@ -219,20 +220,22 @@ int swap_proc_slot(struct vmproc *src_vmp, struct vmproc *dst_vmp)
 /*===========================================================================*
  *			      swap_proc_dyn_data	     		     *
  *===========================================================================*/
-int swap_proc_dyn_data(struct vmproc *src_vmp, struct vmproc *dst_vmp)
+int swap_proc_dyn_data(struct vmproc *src_vmp, struct vmproc *dst_vmp,
+	int sys_upd_flags)
 {
 	int is_vm;
 	int r;
+	struct vir_region *start_vr, *end_vr;
 
 	is_vm = (dst_vmp->vm_endpoint == VM_PROC_NR);
 
-        /* For VM, transfer memory regions above the stack first. */
+        /* For VM, transfer memory mapped regions first. */
         if(is_vm) {
 #if LU_DEBUG
-		printf("VM: swap_proc_dyn_data: tranferring regions above the stack from old VM (%d) to new VM (%d)\n",
+		printf("VM: swap_proc_dyn_data: tranferring memory mapped regions from old (%d) to new VM (%d)\n",
 			src_vmp->vm_endpoint, dst_vmp->vm_endpoint);
 #endif
-		r = pt_map_in_range(src_vmp, dst_vmp, VM_STACKTOP, 0);
+		r = pt_map_in_range(src_vmp, dst_vmp, VM_OWN_HEAPBASE, VM_OWN_MMAPTOP);
 		if(r != OK) {
 			printf("swap_proc_dyn_data: pt_map_in_range failed\n");
 			return r;
@@ -249,23 +252,29 @@ int swap_proc_dyn_data(struct vmproc *src_vmp, struct vmproc *dst_vmp)
 	map_setparent(src_vmp);
 	map_setparent(dst_vmp);
 
-	/* For regular processes, transfer regions above the stack now.
-	 * In case of rollback, we need to skip this step. To sandbox the
-	 * new instance and prevent state corruption on rollback, we share all
-	 * the regions between the two instances as COW.
+	/* Don't transfer mmapped regions if not required. */
+	if(is_vm || (sys_upd_flags & (SF_VM_ROLLBACK|SF_VM_NOMMAP))) {
+		return OK;
+	}
+
+	/* Make sure regions are consistent. */
+	assert(region_search_root(&src_vmp->vm_regions_avl) && region_search_root(&dst_vmp->vm_regions_avl));
+
+	/* Transfer memory mapped regions now. To sandbox the new instance and
+	 * prevent state corruption on rollback, we share all the regions
+	 * between the two instances as COW.
 	 */
-	if(!is_vm) {
-		struct vir_region *vr;
-		vr = map_lookup(dst_vmp, VM_STACKTOP, NULL);
-		if(vr && !map_lookup(src_vmp, VM_STACKTOP, NULL)) {
+	start_vr = region_search(&dst_vmp->vm_regions_avl, VM_MMAPBASE, AVL_GREATER_EQUAL);
+	end_vr = region_search(&dst_vmp->vm_regions_avl, VM_MMAPTOP, AVL_LESS);
+	if(start_vr) {
 #if LU_DEBUG
-			printf("VM: swap_proc_dyn_data: tranferring regions above the stack from %d to %d\n",
-				src_vmp->vm_endpoint, dst_vmp->vm_endpoint);
+		printf("VM: swap_proc_dyn_data: tranferring memory mapped regions from %d to %d\n",
+			dst_vmp->vm_endpoint, src_vmp->vm_endpoint);
 #endif
-			r = map_proc_copy_from(src_vmp, dst_vmp, vr);
-			if(r != OK) {
-				return r;
-			}
+		assert(end_vr);
+		r = map_proc_copy_range(src_vmp, dst_vmp, start_vr, end_vr);
+		if(r != OK) {
+			return r;
 		}
 	}
 
@@ -355,3 +364,25 @@ int do_getrusage(message *m)
 	return sys_datacopy(SELF, (vir_bytes) &r_usage, m->m_source,
 		m->m_lc_vm_rusage.addr, (vir_bytes) sizeof(r_usage));
 }
+
+/*===========================================================================*
+ *                            adjust_proc_refs                              *
+ *===========================================================================*/
+void adjust_proc_refs()
+{
+       struct vmproc *vmp;
+       region_iter iter;
+
+       /* Fix up region parents. */
+       for(vmp = vmproc; vmp < &vmproc[VMP_NR]; vmp++) {
+               struct vir_region *vr;
+               if(!(vmp->vm_flags & VMF_INUSE))
+                       continue;
+               region_start_iter_least(&vmp->vm_regions_avl, &iter);
+               while((vr = region_get_iter(&iter))) {
+                       USE(vr, vr->parent = vmp;);
+                       region_incr_iter(&iter);
+               }
+       }
+}
+
diff --git a/minix/servers/vm/vm.h b/minix/servers/vm/vm.h
index 3a6e56acc..751e7a819 100644
--- a/minix/servers/vm/vm.h
+++ b/minix/servers/vm/vm.h
@@ -79,4 +79,10 @@
 #define VM_MMAPBASE	VM_PAGE_SIZE
 #endif
 
+extern char _end;
+#define VM_OWN_HEAPSTART ((vir_bytes) (&_end))
+#define VM_OWN_HEAPBASE   roundup(VM_OWN_HEAPSTART, VM_PAGE_SIZE)
+#define VM_OWN_MMAPBASE (VM_OWN_HEAPBASE+1024*1024*1024)
+#define VM_OWN_MMAPTOP   (VM_OWN_MMAPBASE+100 * 1024 * 1024)
+
 #endif
diff --git a/minix/servers/vm/vmproc.h b/minix/servers/vm/vmproc.h
index a174714a4..0ec2a8219 100644
--- a/minix/servers/vm/vmproc.h
+++ b/minix/servers/vm/vmproc.h
@@ -35,5 +35,6 @@ struct vmproc {
 #define VMF_INUSE	0x001	/* slot contains a process */
 #define VMF_EXITING	0x002	/* PM is cleaning up this process */
 #define VMF_WATCHEXIT	0x008	/* Store in queryexit table */
+#define VMF_VM_INSTANCE 0x010   /* This is a VM process instance */
 
 #endif