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>

5 files changed, 102 insertions, 23 deletions

diff --git a/arch/tile/include/asm/hugetlb.h b/arch/tile/include/asm/hugetlb.h
index d396d180516..b2042380a5a 100644
--- a/arch/tile/include/asm/hugetlb.h
+++ b/arch/tile/include/asm/hugetlb.h
@@ -106,4 +106,25 @@ static inline void arch_release_hugepage(struct page *page)
 {
 }
 
+#ifdef CONFIG_HUGETLB_SUPER_PAGES
+static inline pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma,
+				       struct page *page, int writable)
+{
+	size_t pagesize = huge_page_size(hstate_vma(vma));
+	if (pagesize != PUD_SIZE && pagesize != PMD_SIZE)
+		entry = pte_mksuper(entry);
+	return entry;
+}
+#define arch_make_huge_pte arch_make_huge_pte
+
+/* Sizes to scale up page size for PTEs with HV_PTE_SUPER bit. */
+enum {
+	HUGE_SHIFT_PGDIR = 0,
+	HUGE_SHIFT_PMD = 1,
+	HUGE_SHIFT_PAGE = 2,
+	HUGE_SHIFT_ENTRIES
+};
+extern int huge_shift[HUGE_SHIFT_ENTRIES];
+#endif
+
 #endif /* _ASM_TILE_HUGETLB_H */
diff --git a/arch/tile/include/asm/page.h b/arch/tile/include/asm/page.h
index c750943f961..9d9131e5c55 100644
--- a/arch/tile/include/asm/page.h
+++ b/arch/tile/include/asm/page.h
@@ -87,8 +87,7 @@ typedef HV_PTE pgprot_t;
 /*
  * User L2 page tables are managed as one L2 page table per page,
  * because we use the page allocator for them.  This keeps the allocation
- * simple and makes it potentially useful to implement HIGHPTE at some point.
- * However, it's also inefficient, since L2 page tables are much smaller
+ * simple, but it's also inefficient, since L2 page tables are much smaller
  * than pages (currently 2KB vs 64KB).  So we should revisit this.
  */
 typedef struct page *pgtable_t;
@@ -137,7 +136,7 @@ static inline __attribute_const__ int get_order(unsigned long size)
 
 #define HUGETLB_PAGE_ORDER	(HPAGE_SHIFT - PAGE_SHIFT)
 
-#define HUGE_MAX_HSTATE		2
+#define HUGE_MAX_HSTATE		6
 
 #ifdef CONFIG_HUGETLB_PAGE
 #define HAVE_ARCH_HUGETLB_UNMAPPED_AREA
diff --git a/arch/tile/include/asm/pgtable.h b/arch/tile/include/asm/pgtable.h
index 319f4826d97..73b1a4c9ad0 100644
--- a/arch/tile/include/asm/pgtable.h
+++ b/arch/tile/include/asm/pgtable.h
@@ -71,6 +71,7 @@ extern void set_page_homes(void);
 
 #define _PAGE_PRESENT           HV_PTE_PRESENT
 #define _PAGE_HUGE_PAGE         HV_PTE_PAGE
+#define _PAGE_SUPER_PAGE        HV_PTE_SUPER
 #define _PAGE_READABLE          HV_PTE_READABLE
 #define _PAGE_WRITABLE          HV_PTE_WRITABLE
 #define _PAGE_EXECUTABLE        HV_PTE_EXECUTABLE
@@ -87,6 +88,7 @@ extern void set_page_homes(void);
 #define _PAGE_ALL (\
   _PAGE_PRESENT | \
   _PAGE_HUGE_PAGE | \
+  _PAGE_SUPER_PAGE | \
   _PAGE_READABLE | \
   _PAGE_WRITABLE | \
   _PAGE_EXECUTABLE | \
@@ -197,6 +199,7 @@ static inline void __pte_clear(pte_t *ptep)
 #define pte_write hv_pte_get_writable
 #define pte_exec hv_pte_get_executable
 #define pte_huge hv_pte_get_page
+#define pte_super hv_pte_get_super
 #define pte_rdprotect hv_pte_clear_readable
 #define pte_exprotect hv_pte_clear_executable
 #define pte_mkclean hv_pte_clear_dirty
@@ -209,6 +212,7 @@ static inline void __pte_clear(pte_t *ptep)
 #define pte_mkyoung hv_pte_set_accessed
 #define pte_mkwrite hv_pte_set_writable
 #define pte_mkhuge hv_pte_set_page
+#define pte_mksuper hv_pte_set_super
 
 #define pte_special(pte) 0
 #define pte_mkspecial(pte) (pte)
@@ -338,13 +342,8 @@ static inline pte_t pte_modify(pte_t pte, pgprot_t newprot)
  */
 #define pgd_offset_k(address) pgd_offset(&init_mm, address)
 
-#if defined(CONFIG_HIGHPTE)
-extern pte_t *pte_offset_map(pmd_t *, unsigned long address);
-#define pte_unmap(pte) kunmap_atomic(pte)
-#else
 #define pte_offset_map(dir, address) pte_offset_kernel(dir, address)
 #define pte_unmap(pte) do { } while (0)
-#endif
 
 /* Clear a non-executable kernel PTE and flush it from the TLB. */
 #define kpte_clear_flush(ptep, vaddr)		\
@@ -537,7 +536,8 @@ static inline pte_t *pte_offset_kernel(pmd_t *pmd, unsigned long address)
 /* Support /proc/NN/pgtable API. */
 struct seq_file;
 int arch_proc_pgtable_show(struct seq_file *m, struct mm_struct *mm,
-			   unsigned long vaddr, pte_t *ptep, void **datap);
+			   unsigned long vaddr, unsigned long pagesize,
+			   pte_t *ptep, void **datap);
 
 #endif /* !__ASSEMBLY__ */
 
diff --git a/arch/tile/include/asm/tlbflush.h b/arch/tile/include/asm/tlbflush.h
index 96199d214fb..dcf91b25a1e 100644
--- a/arch/tile/include/asm/tlbflush.h
+++ b/arch/tile/include/asm/tlbflush.h
@@ -38,16 +38,11 @@ DECLARE_PER_CPU(int, current_asid);
 /* The hypervisor tells us what ASIDs are available to us. */
 extern int min_asid, max_asid;
 
-static inline unsigned long hv_page_size(const struct vm_area_struct *vma)
-{
-	return (vma->vm_flags & VM_HUGETLB) ? HPAGE_SIZE : PAGE_SIZE;
-}
-
 /* Pass as vma pointer for non-executable mapping, if no vma available. */
-#define FLUSH_NONEXEC ((const struct vm_area_struct *)-1UL)
+#define FLUSH_NONEXEC ((struct vm_area_struct *)-1UL)
 
 /* Flush a single user page on this cpu. */
-static inline void local_flush_tlb_page(const struct vm_area_struct *vma,
+static inline void local_flush_tlb_page(struct vm_area_struct *vma,
 					unsigned long addr,
 					unsigned long page_size)
 {
@@ -60,7 +55,7 @@ static inline void local_flush_tlb_page(const struct vm_area_struct *vma,
 }
 
 /* Flush range of user pages on this cpu. */
-static inline void local_flush_tlb_pages(const struct vm_area_struct *vma,
+static inline void local_flush_tlb_pages(struct vm_area_struct *vma,
 					 unsigned long addr,
 					 unsigned long page_size,
 					 unsigned long len)
@@ -117,10 +112,10 @@ extern void flush_tlb_all(void);
 extern void flush_tlb_kernel_range(unsigned long start, unsigned long end);
 extern void flush_tlb_current_task(void);
 extern void flush_tlb_mm(struct mm_struct *);
-extern void flush_tlb_page(const struct vm_area_struct *, unsigned long);
-extern void flush_tlb_page_mm(const struct vm_area_struct *,
+extern void flush_tlb_page(struct vm_area_struct *, unsigned long);
+extern void flush_tlb_page_mm(struct vm_area_struct *,
 			      struct mm_struct *, unsigned long);
-extern void flush_tlb_range(const struct vm_area_struct *,
+extern void flush_tlb_range(struct vm_area_struct *,
 			    unsigned long start, unsigned long end);
 
 #define flush_tlb()     flush_tlb_current_task()
diff --git a/arch/tile/include/hv/hypervisor.h b/arch/tile/include/hv/hypervisor.h
index f27871775b7..85e5cab4c2f 100644
--- a/arch/tile/include/hv/hypervisor.h
+++ b/arch/tile/include/hv/hypervisor.h
@@ -66,6 +66,22 @@
 #define HV_DEFAULT_PAGE_SIZE_LARGE \
   (__HV_SIZE_ONE << HV_LOG2_DEFAULT_PAGE_SIZE_LARGE)
 
+#if CHIP_VA_WIDTH() > 32
+
+/** The log2 of the initial size of jumbo pages, in bytes.
+ * See HV_DEFAULT_PAGE_SIZE_JUMBO.
+ */
+#define HV_LOG2_DEFAULT_PAGE_SIZE_JUMBO 32
+
+/** The initial size of jumbo pages, in bytes. This value should
+ * be verified at runtime by calling hv_sysconf(HV_SYSCONF_PAGE_SIZE_JUMBO).
+ * It may also be modified when installing a new context.
+ */
+#define HV_DEFAULT_PAGE_SIZE_JUMBO \
+  (__HV_SIZE_ONE << HV_LOG2_DEFAULT_PAGE_SIZE_JUMBO)
+
+#endif
+
 /** The log2 of the granularity at which page tables must be aligned;
  *  in other words, the CPA for a page table must have this many zero
  *  bits at the bottom of the address.
@@ -284,8 +300,11 @@
 #define HV_DISPATCH_GET_IPI_PTE                   56
 #endif
 
+/** hv_set_pte_super_shift */
+#define HV_DISPATCH_SET_PTE_SUPER_SHIFT           57
+
 /** One more than the largest dispatch value */
-#define _HV_DISPATCH_END                          57
+#define _HV_DISPATCH_END                          58
 
 
 #ifndef __ASSEMBLER__
@@ -413,6 +432,11 @@ typedef enum {
    */
   HV_SYSCONF_VALID_PAGE_SIZES = 7,
 
+  /** The size of jumbo pages, in bytes.
+   * If no jumbo pages are available, zero will be returned.
+   */
+  HV_SYSCONF_PAGE_SIZE_JUMBO = 8,
+
 } HV_SysconfQuery;
 
 /** Offset to subtract from returned Kelvin temperature to get degrees
@@ -695,6 +719,29 @@ int hv_install_context(HV_PhysAddr page_table, HV_PTE access, HV_ASID asid,
 
 #ifndef __ASSEMBLER__
 
+
+/** Set the number of pages ganged together by HV_PTE_SUPER at a
+ * particular level of the page table.
+ *
+ * The current TILE-Gx hardware only supports powers of four
+ * (i.e. log2_count must be a multiple of two), and the requested
+ * "super" page size must be less than the span of the next level in
+ * the page table.  The largest size that can be requested is 64GB.
+ *
+ * The shift value is initially "0" for all page table levels,
+ * indicating that the HV_PTE_SUPER bit is effectively ignored.
+ *
+ * If you change the count from one non-zero value to another, the
+ * hypervisor will flush the entire TLB and TSB to avoid confusion.
+ *
+ * @param level Page table level (0, 1, or 2)
+ * @param log2_count Base-2 log of the number of pages to gang together,
+ * i.e. how much to shift left the base page size for the super page size.
+ * @return Zero on success, or a hypervisor error code on failure.
+ */
+int hv_set_pte_super_shift(int level, int log2_count);
+
+
 /** Value returned from hv_inquire_context(). */
 typedef struct
 {
@@ -1891,8 +1938,9 @@ int hv_flush_remote(HV_PhysAddr cache_pa, unsigned long cache_control,
 #define HV_PTE_INDEX_USER            10  /**< Page is user-accessible */
 #define HV_PTE_INDEX_ACCESSED        11  /**< Page has been accessed */
 #define HV_PTE_INDEX_DIRTY           12  /**< Page has been written */
-                                         /*   Bits 13-15 are reserved for
+                                         /*   Bits 13-14 are reserved for
                                               future use. */
+#define HV_PTE_INDEX_SUPER           15  /**< Pages ganged together for TLB */
 #define HV_PTE_INDEX_MODE            16  /**< Page mode; see HV_PTE_MODE_xxx */
 #define HV_PTE_MODE_BITS              3  /**< Number of bits in mode */
 #define HV_PTE_INDEX_CLIENT2         19  /**< Page client state 2 */
@@ -1987,7 +2035,10 @@ int hv_flush_remote(HV_PhysAddr cache_pa, unsigned long cache_control,
 
 /** Does this PTE map a page?
  *
- * If this bit is set in the level-1 page table, the entry should be
+ * If this bit is set in a level-0 page table, the entry should be
+ * interpreted as a level-2 page table entry mapping a jumbo page.
+ *
+ * If this bit is set in a level-1 page table, the entry should be
  * interpreted as a level-2 page table entry mapping a large page.
  *
  * This bit should not be modified by the client while PRESENT is set, as
@@ -1997,6 +2048,18 @@ int hv_flush_remote(HV_PhysAddr cache_pa, unsigned long cache_control,
  */
 #define HV_PTE_PAGE                  (__HV_PTE_ONE << HV_PTE_INDEX_PAGE)
 
+/** Does this PTE implicitly reference multiple pages?
+ *
+ * If this bit is set in the page table (either in the level-2 page table,
+ * or in a higher level page table in conjunction with the PAGE bit)
+ * then the PTE specifies a range of contiguous pages, not a single page.
+ * The hv_set_pte_super_shift() allows you to specify the count for
+ * each level of the page table.
+ *
+ * Note: this bit is not supported on TILEPro systems.
+ */
+#define HV_PTE_SUPER                 (__HV_PTE_ONE << HV_PTE_INDEX_SUPER)
+
 /** Is this a global (non-ASID) mapping?
  *
  * If this bit is set, the translations established by this PTE will
@@ -2215,6 +2278,7 @@ hv_pte_clear_##name(HV_PTE pte)                                 \
  */
 _HV_BIT(present,         PRESENT)
 _HV_BIT(page,            PAGE)
+_HV_BIT(super,           SUPER)
 _HV_BIT(client0,         CLIENT0)
 _HV_BIT(client1,         CLIENT1)
 _HV_BIT(client2,         CLIENT2)