arch/tile: support multiple huge page sizes dynamically

This change adds support for a new "super" bit in the PTE, using the new
arch_make_huge_pte() method.  The Tilera hypervisor sees the bit set at a
given level of the page table and gangs together 4, 16, or 64 consecutive
pages from that level of the hierarchy to create a larger TLB entry.

One extra "super" page size can be specified at each of the three levels
of the page table hierarchy on tilegx, using the "hugepagesz" argument
on the boot command line.  A new hypervisor API is added to allow Linux
to tell the hypervisor how many PTEs to gang together at each level of
the page table.

To allow pre-allocating huge pages larger than the buddy allocator can
handle, this change modifies the Tilera bootmem support to put all of
memory on tilegx platforms into bootmem.

As part of this change I eliminate the vestigial CONFIG_HIGHPTE support,
which never worked anyway, and eliminate the hv_page_size() API in favor
of the standard vma_kernel_pagesize() API.

Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>

4 files changed, 120 insertions, 56 deletions

diff --git a/arch/tile/kernel/hvglue.lds b/arch/tile/kernel/hvglue.lds
index 2b7cd0a..d44c5a6 100644
--- a/arch/tile/kernel/hvglue.lds
+++ b/arch/tile/kernel/hvglue.lds
@@ -55,4 +55,5 @@ hv_store_mapping = TEXT_OFFSET + 0x106a0;
 hv_inquire_realpa = TEXT_OFFSET + 0x106c0;
 hv_flush_all = TEXT_OFFSET + 0x106e0;
 hv_get_ipi_pte = TEXT_OFFSET + 0x10700;
-hv_glue_internals = TEXT_OFFSET + 0x10720;
+hv_set_pte_super_shift = TEXT_OFFSET + 0x10720;
+hv_glue_internals = TEXT_OFFSET + 0x10740;
diff --git a/arch/tile/kernel/proc.c b/arch/tile/kernel/proc.c
index 446a7f5..dafc447 100644
--- a/arch/tile/kernel/proc.c
+++ b/arch/tile/kernel/proc.c
@@ -22,6 +22,7 @@
 #include <linux/proc_fs.h>
 #include <linux/sysctl.h>
 #include <linux/hardirq.h>
+#include <linux/hugetlb.h>
 #include <linux/mman.h>
 #include <asm/unaligned.h>
 #include <asm/pgtable.h>
diff --git a/arch/tile/kernel/setup.c b/arch/tile/kernel/setup.c
index 32948e2..445c220 100644
--- a/arch/tile/kernel/setup.c
+++ b/arch/tile/kernel/setup.c
@@ -28,6 +28,7 @@
 #include <linux/highmem.h>
 #include <linux/smp.h>
 #include <linux/timex.h>
+#include <linux/hugetlb.h>
 #include <asm/setup.h>
 #include <asm/sections.h>
 #include <asm/cacheflush.h>
@@ -49,9 +50,6 @@ char chip_model[64] __write_once;
 struct pglist_data node_data[MAX_NUMNODES] __read_mostly;
 EXPORT_SYMBOL(node_data);
 
-/* We only create bootmem data on node 0. */
-static bootmem_data_t __initdata node0_bdata;
-
 /* Information on the NUMA nodes that we compute early */
 unsigned long __cpuinitdata node_start_pfn[MAX_NUMNODES];
 unsigned long __cpuinitdata node_end_pfn[MAX_NUMNODES];
@@ -518,37 +516,96 @@ static void __init setup_memory(void)
 #endif
 }
 
-static void __init setup_bootmem_allocator(void)
+/*
+ * On 32-bit machines, we only put bootmem on the low controller,
+ * since PAs > 4GB can't be used in bootmem.  In principle one could
+ * imagine, e.g., multiple 1 GB controllers all of which could support
+ * bootmem, but in practice using controllers this small isn't a
+ * particularly interesting scenario, so we just keep it simple and
+ * use only the first controller for bootmem on 32-bit machines.
+ */
+static inline int node_has_bootmem(int nid)
 {
-	unsigned long bootmap_size, first_alloc_pfn, last_alloc_pfn;
+#ifdef CONFIG_64BIT
+	return 1;
+#else
+	return nid == 0;
+#endif
+}
 
-	/* Provide a node 0 bdata. */
-	NODE_DATA(0)->bdata = &node0_bdata;
+static inline unsigned long alloc_bootmem_pfn(int nid,
+					      unsigned long size,
+					      unsigned long goal)
+{
+	void *kva = __alloc_bootmem_node(NODE_DATA(nid), size,
+					 PAGE_SIZE, goal);
+	unsigned long pfn = kaddr_to_pfn(kva);
+	BUG_ON(goal && PFN_PHYS(pfn) != goal);
+	return pfn;
+}
 
-#ifdef CONFIG_PCI
-	/* Don't let boot memory alias the PCI region. */
-	last_alloc_pfn = min(max_low_pfn, pci_reserve_start_pfn);
+static void __init setup_bootmem_allocator_node(int i)
+{
+	unsigned long start, end, mapsize, mapstart;
+
+	if (node_has_bootmem(i)) {
+		NODE_DATA(i)->bdata = &bootmem_node_data[i];
+	} else {
+		/* Share controller zero's bdata for now. */
+		NODE_DATA(i)->bdata = &bootmem_node_data[0];
+		return;
+	}
+
+	/* Skip up to after the bss in node 0. */
+	start = (i == 0) ? min_low_pfn : node_start_pfn[i];
+
+	/* Only lowmem, if we're a HIGHMEM build. */
+#ifdef CONFIG_HIGHMEM
+	end = node_lowmem_end_pfn[i];
 #else
-	last_alloc_pfn = max_low_pfn;
+	end = node_end_pfn[i];
 #endif
 
-	/*
-	 * Initialize the boot-time allocator (with low memory only):
-	 * The first argument says where to put the bitmap, and the
-	 * second says where the end of allocatable memory is.
-	 */
-	bootmap_size = init_bootmem(min_low_pfn, last_alloc_pfn);
+	/* No memory here. */
+	if (end == start)
+		return;
+
+	/* Figure out where the bootmem bitmap is located. */
+	mapsize = bootmem_bootmap_pages(end - start);
+	if (i == 0) {
+		/* Use some space right before the heap on node 0. */
+		mapstart = start;
+		start += mapsize;
+	} else {
+		/* Allocate bitmap on node 0 to avoid page table issues. */
+		mapstart = alloc_bootmem_pfn(0, PFN_PHYS(mapsize), 0);
+	}
 
+	/* Initialize a node. */
+	init_bootmem_node(NODE_DATA(i), mapstart, start, end);
+
+	/* Free all the space back into the allocator. */
+	free_bootmem(PFN_PHYS(start), PFN_PHYS(end - start));
+
+#if defined(CONFIG_PCI)
 	/*
-	 * Let the bootmem allocator use all the space we've given it
-	 * except for its own bitmap.
+	 * Throw away any memory aliased by the PCI region.  FIXME: this
+	 * is a temporary hack to work around bug 10502, and needs to be
+	 * fixed properly.
 	 */
-	first_alloc_pfn = min_low_pfn + PFN_UP(bootmap_size);
-	if (first_alloc_pfn >= last_alloc_pfn)
-		early_panic("Not enough memory on controller 0 for bootmem\n");
+	if (pci_reserve_start_pfn < end && pci_reserve_end_pfn > start)
+		reserve_bootmem(PFN_PHYS(pci_reserve_start_pfn),
+				PFN_PHYS(pci_reserve_end_pfn -
+					 pci_reserve_start_pfn),
+				BOOTMEM_EXCLUSIVE);
+#endif
+}
 
-	free_bootmem(PFN_PHYS(first_alloc_pfn),
-		     PFN_PHYS(last_alloc_pfn - first_alloc_pfn));
+static void __init setup_bootmem_allocator(void)
+{
+	int i;
+	for (i = 0; i < MAX_NUMNODES; ++i)
+		setup_bootmem_allocator_node(i);
 
 #ifdef CONFIG_KEXEC
 	if (crashk_res.start != crashk_res.end)
@@ -579,14 +636,6 @@ static int __init percpu_size(void)
 	return size;
 }
 
-static inline unsigned long alloc_bootmem_pfn(int size, unsigned long goal)
-{
-	void *kva = __alloc_bootmem(size, PAGE_SIZE, goal);
-	unsigned long pfn = kaddr_to_pfn(kva);
-	BUG_ON(goal && PFN_PHYS(pfn) != goal);
-	return pfn;
-}
-
 static void __init zone_sizes_init(void)
 {
 	unsigned long zones_size[MAX_NR_ZONES] = { 0 };
@@ -624,21 +673,22 @@ static void __init zone_sizes_init(void)
 		 * though, there'll be no lowmem, so we just alloc_bootmem
 		 * the memmap.  There will be no percpu memory either.
 		 */
-		if (__pfn_to_highbits(start) == 0) {
-			/* In low PAs, allocate via bootmem. */
+		if (i != 0 && cpu_isset(i, isolnodes)) {
+			node_memmap_pfn[i] =
+				alloc_bootmem_pfn(0, memmap_size, 0);
+			BUG_ON(node_percpu[i] != 0);
+		} else if (node_has_bootmem(start)) {
 			unsigned long goal = 0;
 			node_memmap_pfn[i] =
-				alloc_bootmem_pfn(memmap_size, goal);
+				alloc_bootmem_pfn(i, memmap_size, 0);
 			if (kdata_huge)
 				goal = PFN_PHYS(lowmem_end) - node_percpu[i];
 			if (node_percpu[i])
 				node_percpu_pfn[i] =
-				    alloc_bootmem_pfn(node_percpu[i], goal);
-		} else if (cpu_isset(i, isolnodes)) {
-			node_memmap_pfn[i] = alloc_bootmem_pfn(memmap_size, 0);
-			BUG_ON(node_percpu[i] != 0);
+					alloc_bootmem_pfn(i, node_percpu[i],
+							  goal);
 		} else {
-			/* In high PAs, just reserve some pages. */
+			/* In non-bootmem zones, just reserve some pages. */
 			node_memmap_pfn[i] = node_free_pfn[i];
 			node_free_pfn[i] += PFN_UP(memmap_size);
 			if (!kdata_huge) {
@@ -662,16 +712,9 @@ static void __init zone_sizes_init(void)
 		zones_size[ZONE_NORMAL] = end - start;
 #endif
 
-		/*
-		 * Everyone shares node 0's bootmem allocator, but
-		 * we use alloc_remap(), above, to put the actual
-		 * struct page array on the individual controllers,
-		 * which is most of the data that we actually care about.
-		 * We can't place bootmem allocators on the other
-		 * controllers since the bootmem allocator can only
-		 * operate on 32-bit physical addresses.
-		 */
-		NODE_DATA(i)->bdata = NODE_DATA(0)->bdata;
+		/* Take zone metadata from controller 0 if we're isolnode. */
+		if (node_isset(i, isolnodes))
+			NODE_DATA(i)->bdata = &bootmem_node_data[0];
 
 		free_area_init_node(i, zones_size, start, NULL);
 		printk(KERN_DEBUG "  Normal zone: %ld per-cpu pages\n",
@@ -854,6 +897,22 @@ subsys_initcall(topology_init);
 
 #endif /* CONFIG_NUMA */
 
+/*
+ * Initialize hugepage support on this cpu.  We do this on all cores
+ * early in boot: before argument parsing for the boot cpu, and after
+ * argument parsing but before the init functions run on the secondaries.
+ * So the values we set up here in the hypervisor may be overridden on
+ * the boot cpu as arguments are parsed.
+ */
+static __cpuinit void init_super_pages(void)
+{
+#ifdef CONFIG_HUGETLB_SUPER_PAGES
+	int i;
+	for (i = 0; i < HUGE_SHIFT_ENTRIES; ++i)
+		hv_set_pte_super_shift(i, huge_shift[i]);
+#endif
+}
+
 /**
  * setup_cpu() - Do all necessary per-cpu, tile-specific initialization.
  * @boot: Is this the boot cpu?
@@ -908,6 +967,8 @@ void __cpuinit setup_cpu(int boot)
 	/* Reset the network state on this cpu. */
 	reset_network_state();
 #endif
+
+	init_super_pages();
 }
 
 #ifdef CONFIG_BLK_DEV_INITRD
diff --git a/arch/tile/kernel/tlb.c b/arch/tile/kernel/tlb.c
index a5f241c..3fd54d5 100644
--- a/arch/tile/kernel/tlb.c
+++ b/arch/tile/kernel/tlb.c
@@ -15,6 +15,7 @@
 
 #include <linux/cpumask.h>
 #include <linux/module.h>
+#include <linux/hugetlb.h>
 #include <asm/tlbflush.h>
 #include <asm/homecache.h>
 #include <hv/hypervisor.h>
@@ -49,25 +50,25 @@ void flush_tlb_current_task(void)
 	flush_tlb_mm(current->mm);
 }
 
-void flush_tlb_page_mm(const struct vm_area_struct *vma, struct mm_struct *mm,
+void flush_tlb_page_mm(struct vm_area_struct *vma, struct mm_struct *mm,
 		       unsigned long va)
 {
-	unsigned long size = hv_page_size(vma);
+	unsigned long size = vma_kernel_pagesize(vma);
 	int cache = (vma->vm_flags & VM_EXEC) ? HV_FLUSH_EVICT_L1I : 0;
 	flush_remote(0, cache, mm_cpumask(mm),
 		     va, size, size, mm_cpumask(mm), NULL, 0);
 }
 
-void flush_tlb_page(const struct vm_area_struct *vma, unsigned long va)
+void flush_tlb_page(struct vm_area_struct *vma, unsigned long va)
 {
 	flush_tlb_page_mm(vma, vma->vm_mm, va);
 }
 EXPORT_SYMBOL(flush_tlb_page);
 
-void flush_tlb_range(const struct vm_area_struct *vma,
+void flush_tlb_range(struct vm_area_struct *vma,
 		     unsigned long start, unsigned long end)
 {
-	unsigned long size = hv_page_size(vma);
+	unsigned long size = vma_kernel_pagesize(vma);
 	struct mm_struct *mm = vma->vm_mm;
 	int cache = (vma->vm_flags & VM_EXEC) ? HV_FLUSH_EVICT_L1I : 0;
 	flush_remote(0, cache, mm_cpumask(mm), start, end - start, size,