sh: reworked dynamic PMB mapping.

This implements a fairly significant overhaul of the dynamic PMB mapping
code. The primary change here is that the PMB gets its own VMA that
follows the uncached mapping and we attempt to be a bit more intelligent
with dynamic sizing, multi-entry mapping, and so forth.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>

arch/sh/include/asm/io.h

@@ -291,21 +291,21 @@ unsigned long long poke_real_address_q(unsigned long long addr,
  * doesn't exist, so everything must go through page tables.
  */
 #ifdef CONFIG_MMU
-void __iomem *__ioremap_caller(unsigned long offset, unsigned long size,
+void __iomem *__ioremap_caller(phys_addr_t offset, unsigned long size,
 			       pgprot_t prot, void *caller);
 void __iounmap(void __iomem *addr);
 
 static inline void __iomem *
-__ioremap(unsigned long offset, unsigned long size, pgprot_t prot)
+__ioremap(phys_addr_t offset, unsigned long size, pgprot_t prot)
 {
 	return __ioremap_caller(offset, size, prot, __builtin_return_address(0));
 }
 
 static inline void __iomem *
-__ioremap_29bit(unsigned long offset, unsigned long size, pgprot_t prot)
+__ioremap_29bit(phys_addr_t offset, unsigned long size, pgprot_t prot)
 {
 #ifdef CONFIG_29BIT
-	unsigned long last_addr = offset + size - 1;
+	phys_addr_t last_addr = offset + size - 1;
 
 	/*
 	 * For P1 and P2 space this is trivial, as everything is already
@@ -329,7 +329,7 @@ __ioremap_29bit(unsigned long offset, unsigned long size, pgprot_t prot)
 }
 
 static inline void __iomem *
-__ioremap_mode(unsigned long offset, unsigned long size, pgprot_t prot)
+__ioremap_mode(phys_addr_t offset, unsigned long size, pgprot_t prot)
 {
 	void __iomem *ret;
 
@@ -349,35 +349,32 @@ __ioremap_mode(unsigned long offset, unsigned long size, pgprot_t prot)
 #define __iounmap(addr)				do { } while (0)
 #endif /* CONFIG_MMU */
 
-static inline void __iomem *
-ioremap(unsigned long offset, unsigned long size)
+static inline void __iomem *ioremap(phys_addr_t offset, unsigned long size)
 {
 	return __ioremap_mode(offset, size, PAGE_KERNEL_NOCACHE);
 }
 
 static inline void __iomem *
-ioremap_cache(unsigned long offset, unsigned long size)
+ioremap_cache(phys_addr_t offset, unsigned long size)
 {
 	return __ioremap_mode(offset, size, PAGE_KERNEL);
 }
 
 #ifdef CONFIG_HAVE_IOREMAP_PROT
 static inline void __iomem *
-ioremap_prot(resource_size_t offset, unsigned long size, unsigned long flags)
+ioremap_prot(phys_addr_t offset, unsigned long size, unsigned long flags)
 {
 	return __ioremap_mode(offset, size, __pgprot(flags));
 }
 #endif
 
 #ifdef CONFIG_IOREMAP_FIXED
-extern void __iomem *ioremap_fixed(resource_size_t, unsigned long,
-				   unsigned long, pgprot_t);
+extern void __iomem *ioremap_fixed(phys_addr_t, unsigned long, pgprot_t);
 extern int iounmap_fixed(void __iomem *);
 extern void ioremap_fixed_init(void);
 #else
 static inline void __iomem *
-ioremap_fixed(resource_size_t phys_addr, unsigned long offset,
-	      unsigned long size, pgprot_t prot)
+ioremap_fixed(phys_addr_t phys_addr, unsigned long size, pgprot_t prot)
 {
 	BUG();
 	return NULL;

arch/sh/include/asm/mmu.h

@@ -55,19 +55,29 @@ typedef struct {
 
 #ifdef CONFIG_PMB
 /* arch/sh/mm/pmb.c */
-long pmb_remap(unsigned long virt, unsigned long phys,
-	       unsigned long size, pgprot_t prot);
-void pmb_unmap(unsigned long addr);
-void pmb_init(void);
 bool __in_29bit_mode(void);
+
+void pmb_init(void);
+int pmb_bolt_mapping(unsigned long virt, phys_addr_t phys,
+		     unsigned long size, pgprot_t prot);
+void __iomem *pmb_remap_caller(phys_addr_t phys, unsigned long size,
+			       pgprot_t prot, void *caller);
+int pmb_unmap(void __iomem *addr);
+
 #else
-static inline long pmb_remap(unsigned long virt, unsigned long phys,
-			     unsigned long size, pgprot_t prot)
+
+static inline void __iomem *
+pmb_remap_caller(phys_addr_t phys, unsigned long size,
+		 pgprot_t prot, void *caller)
+{
+	return NULL;
+}
+
+static inline int pmb_unmap(void __iomem *addr)
 {
 	return -EINVAL;
 }
 
-#define pmb_unmap(addr)		do { } while (0)
 #define pmb_init(addr)		do { } while (0)
 
 #ifdef CONFIG_29BIT
@@ -77,6 +87,13 @@ static inline long pmb_remap(unsigned long virt, unsigned long phys,
 #endif
 
 #endif /* CONFIG_PMB */
+
+static inline void __iomem *
+pmb_remap(phys_addr_t phys, unsigned long size, pgprot_t prot)
+{
+	return pmb_remap_caller(phys, size, prot, __builtin_return_address(0));
+}
+
 #endif /* __ASSEMBLY__ */
 
 #endif /* __MMU_H */

arch/sh/mm/ioremap.c

@@ -34,17 +34,32 @@
  * caller shouldn't need to know that small detail.
  */
 void __iomem * __init_refok
-__ioremap_caller(unsigned long phys_addr, unsigned long size,
+__ioremap_caller(phys_addr_t phys_addr, unsigned long size,
 		 pgprot_t pgprot, void *caller)
 {
 	struct vm_struct *area;
 	unsigned long offset, last_addr, addr, orig_addr;
+	void __iomem *mapped;
 
 	/* Don't allow wraparound or zero size */
 	last_addr = phys_addr + size - 1;
 	if (!size || last_addr < phys_addr)
 		return NULL;
 
+	/*
+	 * If we can't yet use the regular approach, go the fixmap route.
+	 */
+	if (!mem_init_done)
+		return ioremap_fixed(phys_addr, size, pgprot);
+
+	/*
+	 * First try to remap through the PMB.
+	 * PMB entries are all pre-faulted.
+	 */
+	mapped = pmb_remap_caller(phys_addr, size, pgprot, caller);
+	if (mapped && !IS_ERR(mapped))
+		return mapped;
+
 	/*
 	 * Mappings have to be page-aligned
 	 */
@@ -52,12 +67,6 @@ __ioremap_caller(unsigned long phys_addr, unsigned long size,
 	phys_addr &= PAGE_MASK;
 	size = PAGE_ALIGN(last_addr+1) - phys_addr;
 
-	/*
-	 * If we can't yet use the regular approach, go the fixmap route.
-	 */
-	if (!mem_init_done)
-		return ioremap_fixed(phys_addr, offset, size, pgprot);
-
 	/*
 	 * Ok, go for it..
 	 */
@@ -67,29 +76,6 @@ __ioremap_caller(unsigned long phys_addr, unsigned long size,
 	area->phys_addr = phys_addr;
 	orig_addr = addr = (unsigned long)area->addr;
 
-#ifdef CONFIG_PMB
-	/*
-	 * First try to remap through the PMB once a valid VMA has been
-	 * established. Smaller allocations (or the rest of the size
-	 * remaining after a PMB mapping due to the size not being
-	 * perfectly aligned on a PMB size boundary) are then mapped
-	 * through the UTLB using conventional page tables.
-	 *
-	 * PMB entries are all pre-faulted.
-	 */
-	if (unlikely(phys_addr >= P1SEG)) {
-		unsigned long mapped;
-
-		mapped = pmb_remap(addr, phys_addr, size, pgprot);
-		if (likely(mapped)) {
-			addr		+= mapped;
-			phys_addr	+= mapped;
-			size		-= mapped;
-		}
-	}
-#endif
-
-	if (likely(size))
 	if (ioremap_page_range(addr, addr + size, phys_addr, pgprot)) {
 		vunmap((void *)orig_addr);
 		return NULL;
@@ -133,23 +119,11 @@ void __iounmap(void __iomem *addr)
 	if (iounmap_fixed(addr) == 0)
 		return;
 
-#ifdef CONFIG_PMB
 	/*
-	 * Purge any PMB entries that may have been established for this
-	 * mapping, then proceed with conventional VMA teardown.
-	 *
-	 * XXX: Note that due to the way that remove_vm_area() does
-	 * matching of the resultant VMA, we aren't able to fast-forward
-	 * the address past the PMB space until the end of the VMA where
-	 * the page tables reside. As such, unmap_vm_area() will be
-	 * forced to linearly scan over the area until it finds the page
-	 * tables where PTEs that need to be unmapped actually reside,
-	 * which is far from optimal. Perhaps we need to use a separate
-	 * VMA for the PMB mappings?
-	 *					-- PFM.
+	 * If the PMB handled it, there's nothing else to do.
 	 */
-	pmb_unmap(vaddr);
-#endif
+	if (pmb_unmap(addr) == 0)
+		return;
 
 	p = remove_vm_area((void *)(vaddr & PAGE_MASK));
 	if (!p) {

arch/sh/mm/ioremap_fixed.c

@@ -45,14 +45,21 @@ void __init ioremap_fixed_init(void)
 }
 
 void __init __iomem *
-ioremap_fixed(resource_size_t phys_addr, unsigned long offset,
-	      unsigned long size, pgprot_t prot)
+ioremap_fixed(phys_addr_t phys_addr, unsigned long size, pgprot_t prot)
 {
 	enum fixed_addresses idx0, idx;
 	struct ioremap_map *map;
 	unsigned int nrpages;
+	unsigned long offset;
 	int i, slot;
 
+	/*
+	 * Mappings have to be page-aligned
+	 */
+	offset = phys_addr & ~PAGE_MASK;
+	phys_addr &= PAGE_MASK;
+	size = PAGE_ALIGN(phys_addr + size) - phys_addr;
+
 	slot = -1;
 	for (i = 0; i < FIX_N_IOREMAPS; i++) {
 		map = &ioremap_maps[i];

arch/sh/mm/pmb.c

@@ -23,6 +23,7 @@
 #include <linux/err.h>
 #include <linux/io.h>
 #include <linux/spinlock.h>
+#include <linux/vmalloc.h>
 #include <asm/sizes.h>
 #include <asm/system.h>
 #include <asm/uaccess.h>
@@ -51,6 +52,16 @@ struct pmb_entry {
 	struct pmb_entry *link;
 };
 
+static struct {
+	unsigned long size;
+	int flag;
+} pmb_sizes[] = {
+	{ .size	= SZ_512M, .flag = PMB_SZ_512M, },
+	{ .size = SZ_128M, .flag = PMB_SZ_128M, },
+	{ .size = SZ_64M,  .flag = PMB_SZ_64M,  },
+	{ .size = SZ_16M,  .flag = PMB_SZ_16M,  },
+};
+
 static void pmb_unmap_entry(struct pmb_entry *, int depth);
 
 static DEFINE_RWLOCK(pmb_rwlock);
@@ -72,6 +83,88 @@ static __always_inline unsigned long mk_pmb_data(unsigned int entry)
 	return mk_pmb_entry(entry) | PMB_DATA;
 }
 
+static __always_inline unsigned int pmb_ppn_in_range(unsigned long ppn)
+{
+	return ppn >= __pa(memory_start) && ppn < __pa(memory_end);
+}
+
+/*
+ * Ensure that the PMB entries match our cache configuration.
+ *
+ * When we are in 32-bit address extended mode, CCR.CB becomes
+ * invalid, so care must be taken to manually adjust cacheable
+ * translations.
+ */
+static __always_inline unsigned long pmb_cache_flags(void)
+{
+	unsigned long flags = 0;
+
+#if defined(CONFIG_CACHE_OFF)
+	flags |= PMB_WT | PMB_UB;
+#elif defined(CONFIG_CACHE_WRITETHROUGH)
+	flags |= PMB_C | PMB_WT | PMB_UB;
+#elif defined(CONFIG_CACHE_WRITEBACK)
+	flags |= PMB_C;
+#endif
+
+	return flags;
+}
+
+/*
+ * Convert typical pgprot value to the PMB equivalent
+ */
+static inline unsigned long pgprot_to_pmb_flags(pgprot_t prot)
+{
+	unsigned long pmb_flags = 0;
+	u64 flags = pgprot_val(prot);
+
+	if (flags & _PAGE_CACHABLE)
+		pmb_flags |= PMB_C;
+	if (flags & _PAGE_WT)
+		pmb_flags |= PMB_WT | PMB_UB;
+
+	return pmb_flags;
+}
+
+static bool pmb_can_merge(struct pmb_entry *a, struct pmb_entry *b)
+{
+	return (b->vpn == (a->vpn + a->size)) &&
+	       (b->ppn == (a->ppn + a->size)) &&
+	       (b->flags == a->flags);
+}
+
+static bool pmb_size_valid(unsigned long size)
+{
+	int i;
+
+	for (i = 0; i < ARRAY_SIZE(pmb_sizes); i++)
+		if (pmb_sizes[i].size == size)
+			return true;
+
+	return false;
+}
+
+static inline bool pmb_addr_valid(unsigned long addr, unsigned long size)
+{
+	return (addr >= P1SEG && (addr + size - 1) < P3SEG);
+}
+
+static inline bool pmb_prot_valid(pgprot_t prot)
+{
+	return (pgprot_val(prot) & _PAGE_USER) == 0;
+}
+
+static int pmb_size_to_flags(unsigned long size)
+{
+	int i;
+
+	for (i = 0; i < ARRAY_SIZE(pmb_sizes); i++)
+		if (pmb_sizes[i].size == size)
+			return pmb_sizes[i].flag;
+
+	return 0;
+}
+
 static int pmb_alloc_entry(void)
 {
 	int pos;
@@ -138,34 +231,14 @@ static void pmb_free(struct pmb_entry *pmbe)
 	pmbe->link	= NULL;
 }
 
-/*
- * Ensure that the PMB entries match our cache configuration.
- *
- * When we are in 32-bit address extended mode, CCR.CB becomes
- * invalid, so care must be taken to manually adjust cacheable
- * translations.
- */
-static __always_inline unsigned long pmb_cache_flags(void)
-{
-	unsigned long flags = 0;
-
-#if defined(CONFIG_CACHE_WRITETHROUGH)
-	flags |= PMB_C | PMB_WT | PMB_UB;
-#elif defined(CONFIG_CACHE_WRITEBACK)
-	flags |= PMB_C;
-#endif
-
-	return flags;
-}
-
 /*
  * Must be run uncached.
  */
 static void __set_pmb_entry(struct pmb_entry *pmbe)
 {
-	writel_uncached(pmbe->vpn | PMB_V, mk_pmb_addr(pmbe->entry));
-	writel_uncached(pmbe->ppn | pmbe->flags | PMB_V,
-			mk_pmb_data(pmbe->entry));
+	/* Set V-bit */
+	__raw_writel(pmbe->ppn | pmbe->flags | PMB_V, mk_pmb_data(pmbe->entry));
+	__raw_writel(pmbe->vpn | PMB_V, mk_pmb_addr(pmbe->entry));
 }
 
 static void __clear_pmb_entry(struct pmb_entry *pmbe)
@@ -193,39 +266,56 @@ static void set_pmb_entry(struct pmb_entry *pmbe)
 	spin_unlock_irqrestore(&pmbe->lock, flags);
 }
 
-static struct {
-	unsigned long size;
-	int flag;
-} pmb_sizes[] = {
-	{ .size	= SZ_512M, .flag = PMB_SZ_512M, },
-	{ .size = SZ_128M, .flag = PMB_SZ_128M, },
-	{ .size = SZ_64M,  .flag = PMB_SZ_64M,  },
-	{ .size = SZ_16M,  .flag = PMB_SZ_16M,  },
-};
-
-long pmb_remap(unsigned long vaddr, unsigned long phys,
+int pmb_bolt_mapping(unsigned long vaddr, phys_addr_t phys,
 		     unsigned long size, pgprot_t prot)
+{
+	return 0;
+}
+
+void __iomem *pmb_remap_caller(phys_addr_t phys, unsigned long size,
+			       pgprot_t prot, void *caller)
 {
 	struct pmb_entry *pmbp, *pmbe;
-	unsigned long wanted;
-	int pmb_flags, i;
-	long err;
-	u64 flags;
+	unsigned long pmb_flags;
+	int i, mapped;
+	unsigned long orig_addr, vaddr;
+	phys_addr_t offset, last_addr;
+	phys_addr_t align_mask;
+	unsigned long aligned;
+	struct vm_struct *area;
 
-	flags = pgprot_val(prot);
+	/*
+	 * Small mappings need to go through the TLB.
+	 */
+	if (size < SZ_16M)
+		return ERR_PTR(-EINVAL);
+	if (!pmb_prot_valid(prot))
+		return ERR_PTR(-EINVAL);
 
-	pmb_flags = PMB_WT | PMB_UB;
+	pmbp = NULL;
+	pmb_flags = pgprot_to_pmb_flags(prot);
+	mapped = 0;
 
-	/* Convert typical pgprot value to the PMB equivalent */
-	if (flags & _PAGE_CACHABLE) {
-		pmb_flags |= PMB_C;
+	for (i = 0; i < ARRAY_SIZE(pmb_sizes); i++)
+		if (size >= pmb_sizes[i].size)
+			break;
 
-		if ((flags & _PAGE_WT) == 0)
-			pmb_flags &= ~(PMB_WT | PMB_UB);
-	}
+	last_addr = phys + size;
+	align_mask = ~(pmb_sizes[i].size - 1);
+	offset = phys & ~align_mask;
+	phys &= align_mask;
+	aligned = ALIGN(last_addr, pmb_sizes[i].size) - phys;
 
-	pmbp = NULL;
-	wanted = size;
+	area = __get_vm_area_caller(aligned, VM_IOREMAP, uncached_end,
+				    P3SEG, caller);
+	if (!area)
+		return NULL;
+
+	area->phys_addr = phys;
+	orig_addr = vaddr = (unsigned long)area->addr;
+
+	if (!pmb_addr_valid(vaddr, aligned))
+		return ERR_PTR(-EFAULT);
 
 again:
 	for (i = 0; i < ARRAY_SIZE(pmb_sizes); i++) {
@@ -237,19 +327,19 @@ again:
 		pmbe = pmb_alloc(vaddr, phys, pmb_flags | pmb_sizes[i].flag,
 				 PMB_NO_ENTRY);
 		if (IS_ERR(pmbe)) {
-			err = PTR_ERR(pmbe);
-			goto out;
+			pmb_unmap_entry(pmbp, mapped);
+			return pmbe;
 		}
 
 		spin_lock_irqsave(&pmbe->lock, flags);
 
+		pmbe->size = pmb_sizes[i].size;
+
 		__set_pmb_entry(pmbe);
 
-		phys	+= pmb_sizes[i].size;
-		vaddr	+= pmb_sizes[i].size;
-		size	-= pmb_sizes[i].size;
-
-		pmbe->size = pmb_sizes[i].size;
+		phys	+= pmbe->size;
+		vaddr	+= pmbe->size;
+		size	-= pmbe->size;
 
 		/*
 		 * Link adjacent entries that span multiple PMB entries
@@ -269,6 +359,7 @@ again:
 		 * pmb_sizes[i].size again.
 		 */
 		i--;
+		mapped++;
 
 		spin_unlock_irqrestore(&pmbe->lock, flags);
 	}
@@ -276,61 +367,35 @@ again:
 	if (size >= SZ_16M)
 		goto again;
 
-	return wanted - size;
-
-out:
-	pmb_unmap_entry(pmbp, NR_PMB_ENTRIES);
-
-	return err;
+	return (void __iomem *)(offset + (char *)orig_addr);
 }
 
-void pmb_unmap(unsigned long addr)
+int pmb_unmap(void __iomem *addr)
 {
 	struct pmb_entry *pmbe = NULL;
-	int i;
+	unsigned long vaddr = (unsigned long __force)addr;
+	int i, found = 0;
 
 	read_lock(&pmb_rwlock);
 
 	for (i = 0; i < ARRAY_SIZE(pmb_entry_list); i++) {
 		if (test_bit(i, pmb_map)) {
 			pmbe = &pmb_entry_list[i];
-			if (pmbe->vpn == addr)
+			if (pmbe->vpn == vaddr) {
+				found = 1;
 				break;
 			}
 		}
+	}
 
 	read_unlock(&pmb_rwlock);
 
+	if (found) {
 		pmb_unmap_entry(pmbe, NR_PMB_ENTRIES);
-}
-
-static bool pmb_can_merge(struct pmb_entry *a, struct pmb_entry *b)
-{
-	return (b->vpn == (a->vpn + a->size)) &&
-	       (b->ppn == (a->ppn + a->size)) &&
-	       (b->flags == a->flags);
-}
-
-static bool pmb_size_valid(unsigned long size)
-{
-	int i;
-
-	for (i = 0; i < ARRAY_SIZE(pmb_sizes); i++)
-		if (pmb_sizes[i].size == size)
-			return true;
-
-	return false;
-}
-
-static int pmb_size_to_flags(unsigned long size)
-{
-	int i;
-
-	for (i = 0; i < ARRAY_SIZE(pmb_sizes); i++)
-		if (pmb_sizes[i].size == size)
-			return pmb_sizes[i].flag;
-
 		return 0;
+	}
+
+	return -EINVAL;
 }
 
 static void __pmb_unmap_entry(struct pmb_entry *pmbe, int depth)
@@ -368,11 +433,6 @@ static void pmb_unmap_entry(struct pmb_entry *pmbe, int depth)
 	write_unlock_irqrestore(&pmb_rwlock, flags);
 }
 
-static __always_inline unsigned int pmb_ppn_in_range(unsigned long ppn)
-{
-	return ppn >= __pa(memory_start) && ppn < __pa(memory_end);
-}
-
 static void __init pmb_notify(void)
 {
 	int i;