xen: convert p2m to a 3 level tree

Make the p2m structure a 3 level tree which covers the full possible
physical space.

The p2m structure contains mappings from the domain's pfns to system-wide
mfns.  The structure has 3 levels and two roots.  The first root is for
the domain's own use, and is linked with virtual addresses.  The second
is all mfn references, and is used by Xen on save/restore to allow it to
update the p2m mapping for the domain.

At boot, the domain builder provides a simple flat p2m array for all the
initially present pages.  We construct the two levels above that using
the early_brk allocator.  After early boot time, set_phys_to_machine()
will allocate any missing levels using the normal kernel allocator
(at GFP_KERNEL, so it must be called in a normal blocking context).

Because the early_brk() API requires us to pre-reserve the maximum amount
of memory we could allocate, there is still a CONFIG_XEN_MAX_DOMAIN_MEMORY
config option, but its only negative side-effect is to increase the
kernel's apparent bss size.  However, since all unused brk memory is
returned to the heap, there's no real downside to making it large.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>

arch/x86/xen/Kconfig

@@ -19,15 +19,12 @@ config XEN_PVHVM
 	depends on X86_LOCAL_APIC
 
 config XEN_MAX_DOMAIN_MEMORY
-       int "Maximum allowed size of a domain in gigabytes"
-       default 8 if X86_32
-       default 32 if X86_64
+       int
+       default 128
        depends on XEN
        help
-         The pseudo-physical to machine address array is sized
-         according to the maximum possible memory size of a Xen
-         domain.  This array uses 1 page per gigabyte, so there's no
-         need to be too stingy here.
+         This only affects the sizing of some bss arrays, the unused
+         portions of which are freed.
 
 config XEN_SAVE_RESTORE
        bool

arch/x86/xen/mmu.c

@@ -170,51 +170,162 @@ DEFINE_PER_CPU(unsigned long, xen_current_cr3);	 /* actual vcpu cr3 */
  */
 #define USER_LIMIT	((STACK_TOP_MAX + PGDIR_SIZE - 1) & PGDIR_MASK)
 
-static unsigned long max_p2m_pfn __read_mostly = MAX_DOMAIN_PAGES;
+/*
+ * Xen leaves the responsibility for maintaining p2m mappings to the
+ * guests themselves, but it must also access and update the p2m array
+ * during suspend/resume when all the pages are reallocated.
+ *
+ * The p2m table is logically a flat array, but we implement it as a
+ * three-level tree to allow the address space to be sparse.
+ *
+ *                               Xen
+ *                                |
+ *     p2m_top              p2m_top_mfn
+ *       /  \                   /   \
+ * p2m_mid p2m_mid	p2m_mid_mfn p2m_mid_mfn
+ *    / \      / \         /           /
+ *  p2m p2m p2m p2m p2m p2m p2m ...
+ *
+ * The p2m_top and p2m_top_mfn levels are limited to 1 page, so the
+ * maximum representable pseudo-physical address space is:
+ *  P2M_TOP_PER_PAGE * P2M_MID_PER_PAGE * P2M_PER_PAGE pages
+ *
+ * P2M_PER_PAGE depends on the architecture, as a mfn is always
+ * unsigned long (8 bytes on 64-bit, 4 bytes on 32), leading to
+ * 512 and 1024 entries respectively. 
+ */
 
-#define P2M_ENTRIES_PER_PAGE		(PAGE_SIZE / sizeof(unsigned long))
-#define TOP_ENTRIES(pages)		((pages) / P2M_ENTRIES_PER_PAGE)
-#define MAX_TOP_ENTRIES			TOP_ENTRIES(MAX_DOMAIN_PAGES)
+static unsigned long max_p2m_pfn __read_mostly;
 
-/* Placeholder for holes in the address space */
-static RESERVE_BRK_ARRAY(unsigned long, p2m_missing, P2M_ENTRIES_PER_PAGE);
+#define P2M_PER_PAGE		(PAGE_SIZE / sizeof(unsigned long))
+#define P2M_MID_PER_PAGE	(PAGE_SIZE / sizeof(unsigned long *))
+#define P2M_TOP_PER_PAGE	(PAGE_SIZE / sizeof(unsigned long **))
 
- /* Array of pointers to pages containing p2m entries */
-static RESERVE_BRK_ARRAY(unsigned long *, p2m_top, MAX_TOP_ENTRIES);
+#define MAX_P2M_PFN		(P2M_TOP_PER_PAGE * P2M_MID_PER_PAGE * P2M_PER_PAGE)
 
-/* Arrays of p2m arrays expressed in mfns used for save/restore */
-static RESERVE_BRK_ARRAY(unsigned long, p2m_top_mfn, MAX_TOP_ENTRIES);
+/* Placeholders for holes in the address space */
+static RESERVE_BRK_ARRAY(unsigned long, p2m_missing, P2M_PER_PAGE);
+static RESERVE_BRK_ARRAY(unsigned long *, p2m_mid_missing, P2M_MID_PER_PAGE);
+static RESERVE_BRK_ARRAY(unsigned long, p2m_mid_missing_mfn, P2M_MID_PER_PAGE);
 
-static RESERVE_BRK_ARRAY(unsigned long, p2m_top_mfn_list,
-			 (MAX_TOP_ENTRIES / P2M_ENTRIES_PER_PAGE));
+static RESERVE_BRK_ARRAY(unsigned long **, p2m_top, P2M_TOP_PER_PAGE);
+static RESERVE_BRK_ARRAY(unsigned long, p2m_top_mfn, P2M_TOP_PER_PAGE);
+
+RESERVE_BRK(p2m_mid, PAGE_SIZE * (MAX_DOMAIN_PAGES / (P2M_PER_PAGE * P2M_MID_PER_PAGE)));
+RESERVE_BRK(p2m_mid_mfn, PAGE_SIZE * (MAX_DOMAIN_PAGES / (P2M_PER_PAGE * P2M_MID_PER_PAGE)));
 
 static inline unsigned p2m_top_index(unsigned long pfn)
 {
-	BUG_ON(pfn >= max_p2m_pfn);
-	return pfn / P2M_ENTRIES_PER_PAGE;
+	BUG_ON(pfn >= MAX_P2M_PFN);
+	return pfn / (P2M_MID_PER_PAGE * P2M_PER_PAGE);
+}
+
+static inline unsigned p2m_mid_index(unsigned long pfn)
+{
+	return (pfn / P2M_PER_PAGE) % P2M_MID_PER_PAGE;
 }
 
 static inline unsigned p2m_index(unsigned long pfn)
 {
-	return pfn % P2M_ENTRIES_PER_PAGE;
+	return pfn % P2M_PER_PAGE;
 }
 
-/* Build the parallel p2m_top_mfn structures */
+static void p2m_top_init(unsigned long ***top)
+{
+	unsigned i;
+
+	for (i = 0; i < P2M_TOP_PER_PAGE; i++)
+		top[i] = p2m_mid_missing;
+}
+
+static void p2m_top_mfn_init(unsigned long *top)
+{
+	unsigned i;
+
+	for (i = 0; i < P2M_TOP_PER_PAGE; i++)
+		top[i] = virt_to_mfn(p2m_mid_missing_mfn);
+}
+
+static void p2m_mid_init(unsigned long **mid)
+{
+	unsigned i;
+
+	for (i = 0; i < P2M_MID_PER_PAGE; i++)
+		mid[i] = p2m_missing;
+}
+
+static void p2m_mid_mfn_init(unsigned long *mid)
+{
+	unsigned i;
+
+	for (i = 0; i < P2M_MID_PER_PAGE; i++)
+		mid[i] = virt_to_mfn(p2m_missing);
+}
+
+static void p2m_init(unsigned long *p2m)
+{
+	unsigned i;
+
+	for (i = 0; i < P2M_MID_PER_PAGE; i++)
+		p2m[i] = INVALID_P2M_ENTRY;
+}
+
+/*
+ * Build the parallel p2m_top_mfn and p2m_mid_mfn structures
+ *
+ * This is called both at boot time, and after resuming from suspend:
+ * - At boot time we're called very early, and must use extend_brk()
+ *   to allocate memory.
+ *
+ * - After resume we're called from within stop_machine, but the mfn
+ *   tree should alreay be completely allocated.
+ */
 void xen_build_mfn_list_list(void)
 {
-	unsigned pfn, idx;
+	unsigned pfn, i;
 
-	for (pfn = 0; pfn < max_p2m_pfn; pfn += P2M_ENTRIES_PER_PAGE) {
-		unsigned topidx = p2m_top_index(pfn);
+	/* Pre-initialize p2m_top_mfn to be completely missing */
+	if (p2m_top_mfn == NULL) {
+		p2m_mid_missing_mfn = extend_brk(PAGE_SIZE, PAGE_SIZE);
+		p2m_mid_mfn_init(p2m_mid_missing_mfn);
 
-		p2m_top_mfn[topidx] = virt_to_mfn(p2m_top[topidx]);
+		p2m_top_mfn = extend_brk(PAGE_SIZE, PAGE_SIZE);
+		p2m_top_mfn_init(p2m_top_mfn);
 	}
 
-	for (idx = 0;
-	     idx < TOP_ENTRIES(max_p2m_pfn)/P2M_ENTRIES_PER_PAGE;
-	     idx++) {
-		unsigned topidx = idx * P2M_ENTRIES_PER_PAGE;
-		p2m_top_mfn_list[idx] = virt_to_mfn(&p2m_top_mfn[topidx]);
+	for (pfn = 0; pfn < max_p2m_pfn; pfn += P2M_PER_PAGE) {
+		unsigned topidx = p2m_top_index(pfn);
+		unsigned mididx = p2m_mid_index(pfn);
+		unsigned long **mid;
+		unsigned long mid_mfn;
+		unsigned long *mid_mfn_p;
+
+		mid = p2m_top[topidx];
+
+		/* Don't bother allocating any mfn mid levels if
+		   they're just missing */
+		if (mid[mididx] == p2m_missing)
+			continue;
+
+		mid_mfn = p2m_top_mfn[topidx];
+		mid_mfn_p = mfn_to_virt(mid_mfn);
+
+		if (mid_mfn_p == p2m_mid_missing_mfn) {
+			/*
+			 * XXX boot-time only!  We should never find
+			 * missing parts of the mfn tree after
+			 * runtime.  extend_brk() will BUG if we call
+			 * it too late.
+			 */
+			mid_mfn_p = extend_brk(PAGE_SIZE, PAGE_SIZE);
+			p2m_mid_mfn_init(mid_mfn_p);
+
+			mid_mfn = virt_to_mfn(mid_mfn_p);
+			
+			p2m_top_mfn[topidx] = mid_mfn;
+		}
+
+		mid_mfn_p[mididx] = virt_to_mfn(mid[mididx]);
 	}
 }
 
@@ -223,7 +334,7 @@ void xen_setup_mfn_list_list(void)
 	BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
 
 	HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
-		virt_to_mfn(p2m_top_mfn_list);
+		virt_to_mfn(p2m_top_mfn);
 	HYPERVISOR_shared_info->arch.max_pfn = max_p2m_pfn;
 }
 
@@ -233,99 +344,154 @@ void __init xen_build_dynamic_phys_to_machine(void)
 	unsigned long *mfn_list = (unsigned long *)xen_start_info->mfn_list;
 	unsigned long max_pfn = min(MAX_DOMAIN_PAGES, xen_start_info->nr_pages);
 	unsigned pfn;
-	unsigned i;
 
 	max_p2m_pfn = max_pfn;
 
-	p2m_missing = extend_brk(sizeof(*p2m_missing) * P2M_ENTRIES_PER_PAGE,
-				 PAGE_SIZE);
-	for (i = 0; i < P2M_ENTRIES_PER_PAGE; i++)
-		p2m_missing[i] = ~0UL;
+	p2m_missing = extend_brk(PAGE_SIZE, PAGE_SIZE);
+	p2m_init(p2m_missing);
 
-	p2m_top = extend_brk(sizeof(*p2m_top) * TOP_ENTRIES(max_pfn),
-			     PAGE_SIZE);
-	for (i = 0; i < TOP_ENTRIES(max_pfn); i++)
-		p2m_top[i] = p2m_missing;
+	p2m_mid_missing = extend_brk(PAGE_SIZE, PAGE_SIZE);
+	p2m_mid_init(p2m_mid_missing);
 
-	p2m_top_mfn = extend_brk(sizeof(*p2m_top_mfn) * TOP_ENTRIES(max_pfn),
-				 PAGE_SIZE);
-	p2m_top_mfn_list = extend_brk(sizeof(*p2m_top_mfn_list) *
-				      (TOP_ENTRIES(max_pfn) / P2M_ENTRIES_PER_PAGE),
-				      PAGE_SIZE);
+	p2m_top = extend_brk(PAGE_SIZE, PAGE_SIZE);
+	p2m_top_init(p2m_top);
 
-	for (pfn = 0; pfn < max_pfn; pfn += P2M_ENTRIES_PER_PAGE) {
+	/*
+	 * The domain builder gives us a pre-constructed p2m array in
+	 * mfn_list for all the pages initially given to us, so we just
+	 * need to graft that into our tree structure.
+	 */
+	for (pfn = 0; pfn < max_pfn; pfn += P2M_PER_PAGE) {
 		unsigned topidx = p2m_top_index(pfn);
+		unsigned mididx = p2m_mid_index(pfn);
+
+		if (p2m_top[topidx] == p2m_mid_missing) {
+			unsigned long **mid = extend_brk(PAGE_SIZE, PAGE_SIZE);
+			p2m_mid_init(mid);
 
-		p2m_top[topidx] = &mfn_list[pfn];
+			p2m_top[topidx] = mid;
+		}
+
+		p2m_top[topidx][mididx] = &mfn_list[pfn];
 	}
 
+	/* Allocate and initialize top and mid mfn levels */
 	xen_build_mfn_list_list();
 }
 
 unsigned long get_phys_to_machine(unsigned long pfn)
 {
-	unsigned topidx, idx;
+	unsigned topidx, mididx, idx;
 
-	if (unlikely(pfn >= max_p2m_pfn))
+	if (unlikely(pfn >= MAX_P2M_PFN))
 		return INVALID_P2M_ENTRY;
 
 	topidx = p2m_top_index(pfn);
+	mididx = p2m_mid_index(pfn);
 	idx = p2m_index(pfn);
-	return p2m_top[topidx][idx];
+
+	return p2m_top[topidx][mididx][idx];
 }
 EXPORT_SYMBOL_GPL(get_phys_to_machine);
 
-/* install a new p2m_top page */
-static bool install_p2mtop_page(unsigned long pfn, unsigned long *p)
+static void *alloc_p2m_page(void)
 {
-	unsigned topidx = p2m_top_index(pfn);
-	unsigned long **pfnp, *mfnp;
-	unsigned i;
+	return (void *)__get_free_page(GFP_KERNEL | __GFP_REPEAT);
+}
 
-	pfnp = &p2m_top[topidx];
-	mfnp = &p2m_top_mfn[topidx];
+static void free_p2m_page(void *p)
+{
+	free_page((unsigned long)p);
+}
 
-	for (i = 0; i < P2M_ENTRIES_PER_PAGE; i++)
-		p[i] = INVALID_P2M_ENTRY;
+/* 
+ * Fully allocate the p2m structure for a given pfn.  We need to check
+ * that both the top and mid levels are allocated, and make sure the
+ * parallel mfn tree is kept in sync.  We may race with other cpus, so
+ * the new pages are installed with cmpxchg; if we lose the race then
+ * simply free the page we allocated and use the one that's there.
+ */
+static bool alloc_p2m(unsigned long pfn)
+{
+	unsigned topidx, mididx;
+	unsigned long ***top_p, **mid;
+	unsigned long *top_mfn_p, *mid_mfn;
 
-	if (cmpxchg(pfnp, p2m_missing, p) == p2m_missing) {
-		*mfnp = virt_to_mfn(p);
-		return true;
+	topidx = p2m_top_index(pfn);
+	mididx = p2m_mid_index(pfn);
+
+	top_p = &p2m_top[topidx];
+	mid = *top_p;
+
+	if (mid == p2m_mid_missing) {
+		/* Mid level is missing, allocate a new one */
+		mid = alloc_p2m_page();
+		if (!mid)
+			return false;
+
+		p2m_mid_init(mid);
+
+		if (cmpxchg(top_p, p2m_mid_missing, mid) != p2m_mid_missing)
+			free_p2m_page(mid);
 	}
 
-	return false;
-}
+	top_mfn_p = &p2m_top_mfn[topidx];
+	mid_mfn = mfn_to_virt(*top_mfn_p);
 
-static void alloc_p2m(unsigned long pfn)
-{
-	unsigned long *p;
+	if (mid_mfn == p2m_mid_missing_mfn) {
+		/* Separately check the mid mfn level */
+		unsigned long missing_mfn;
+		unsigned long mid_mfn_mfn;
+
+		mid_mfn = alloc_p2m_page();
+		if (!mid_mfn)
+			return false;
+
+		p2m_mid_mfn_init(mid_mfn);
+		
+		missing_mfn = virt_to_mfn(p2m_mid_missing_mfn);
+		mid_mfn_mfn = virt_to_mfn(mid_mfn);
+		if (cmpxchg(top_mfn_p, missing_mfn, mid_mfn_mfn) != missing_mfn)
+			free_p2m_page(mid_mfn);
+	}
 
-	p = (void *)__get_free_page(GFP_KERNEL | __GFP_NOFAIL);
-	BUG_ON(p == NULL);
+	if (p2m_top[topidx][mididx] == p2m_missing) {
+		/* p2m leaf page is missing */
+		unsigned long *p2m;
 
-	if (!install_p2mtop_page(pfn, p))
-		free_page((unsigned long)p);
+		p2m = alloc_p2m_page();
+		if (!p2m)
+			return false;
+
+		p2m_init(p2m);
+
+		if (cmpxchg(&mid[mididx], p2m_missing, p2m) != p2m_missing)
+			free_p2m_page(p2m);
+		else
+			mid_mfn[mididx] = virt_to_mfn(p2m);
+	}
+
+	return true;
 }
 
 /* Try to install p2m mapping; fail if intermediate bits missing */
 bool __set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 {
-	unsigned topidx, idx;
+	unsigned topidx, mididx, idx;
 
-	if (unlikely(pfn >= max_p2m_pfn)) {
+	if (unlikely(pfn >= MAX_P2M_PFN)) {
 		BUG_ON(mfn != INVALID_P2M_ENTRY);
 		return true;
 	}
 
 	topidx = p2m_top_index(pfn);
-	if (p2m_top[topidx] == p2m_missing) {
-		if (mfn == INVALID_P2M_ENTRY)
-			return true;
-		return false;
-	}
-
+	mididx = p2m_mid_index(pfn);
 	idx = p2m_index(pfn);
-	p2m_top[topidx][idx] = mfn;
+
+	if (p2m_top[topidx][mididx] == p2m_missing)
+		return mfn == INVALID_P2M_ENTRY;
+
+	p2m_top[topidx][mididx][idx] = mfn;
 
 	return true;
 }
@@ -338,7 +504,7 @@ void set_phys_to_machine(unsigned long pfn, unsigned long mfn)
 	}
 
 	if (unlikely(!__set_phys_to_machine(pfn, mfn)))  {
-		alloc_p2m(pfn);
+		WARN(!alloc_p2m(pfn), "Can't allocate p2m for %lx, %lx", pfn, mfn);
 
 		if (!__set_phys_to_machine(pfn, mfn))
 			BUG();