]> www.pilppa.org Git - linux-2.6-omap-h63xx.git/blobdiff - arch/x86/mm/pat.c
arch/x86/mm/pat.c: use boot_cpu_has()
[linux-2.6-omap-h63xx.git] / arch / x86 / mm / pat.c
index 72c0f6097402bfae75d7d658efa4dc6544994c78..e83b770676d5c93102a3ed3c409be41f7bfb6429 100644 (file)
 #include <linux/kernel.h>
 #include <linux/gfp.h>
 #include <linux/fs.h>
+#include <linux/bootmem.h>
 
 #include <asm/msr.h>
 #include <asm/tlbflush.h>
 #include <asm/processor.h>
+#include <asm/page.h>
 #include <asm/pgtable.h>
 #include <asm/pat.h>
 #include <asm/e820.h>
 #include <asm/cacheflush.h>
 #include <asm/fcntl.h>
 #include <asm/mtrr.h>
+#include <asm/io.h>
 
-int pat_wc_enabled = 1;
+#ifdef CONFIG_X86_PAT
+int __read_mostly pat_wc_enabled = 1;
 
-static u64 __read_mostly boot_pat_state;
-
-static int nopat(char *str)
+void __cpuinit pat_disable(char *reason)
 {
        pat_wc_enabled = 0;
-       printk(KERN_INFO "x86: PAT support disabled.\n");
+       printk(KERN_INFO "%s\n", reason);
+}
 
+static int nopat(char *str)
+{
+       pat_disable("PAT support disabled.");
        return 0;
 }
 early_param("nopat", nopat);
+#endif
 
-static int pat_known_cpu(void)
-{
-       if (!pat_wc_enabled)
-               return 0;
-
-       if (cpu_has_pat)
-               return 1;
 
-       pat_wc_enabled = 0;
-       printk(KERN_INFO "CPU and/or kernel does not support PAT.\n");
+static int debug_enable;
+static int __init pat_debug_setup(char *str)
+{
+       debug_enable = 1;
        return 0;
 }
+__setup("debugpat", pat_debug_setup);
+
+#define dprintk(fmt, arg...) \
+       do { if (debug_enable) printk(KERN_INFO fmt, ##arg); } while (0)
+
+
+static u64 __read_mostly boot_pat_state;
 
 enum {
        PAT_UC = 0,             /* uncached */
@@ -63,17 +72,19 @@ void pat_init(void)
 {
        u64 pat;
 
-#ifndef CONFIG_X86_PAT
-       nopat(NULL);
-#endif
-
-       /* Boot CPU enables PAT based on CPU feature */
-       if (!smp_processor_id() && !pat_known_cpu())
+       if (!pat_wc_enabled)
                return;
 
-       /* APs enable PAT iff boot CPU has enabled it before */
-       if (smp_processor_id() && !pat_wc_enabled)
-               return;
+       /* Paranoia check. */
+       if (!cpu_has_pat) {
+               printk(KERN_ERR "PAT enabled, but CPU feature cleared\n");
+               /*
+                * Panic if this happens on the secondary CPU, and we
+                * switched to PAT on the boot CPU. We have no way to
+                * undo PAT.
+               */
+               BUG_ON(boot_pat_state);
+       }
 
        /* Set PWT to Write-Combining. All other bits stay the same */
        /*
@@ -92,9 +103,8 @@ void pat_init(void)
              PAT(4,WB) | PAT(5,WC) | PAT(6,UC_MINUS) | PAT(7,UC);
 
        /* Boot CPU check */
-       if (!smp_processor_id()) {
+       if (!boot_pat_state)
                rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
-       }
 
        wrmsrl(MSR_IA32_CR_PAT, pat);
        printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n",
@@ -190,6 +200,21 @@ static int pat_x_mtrr_type(u64 start, u64 end, unsigned long prot,
        return 0;
 }
 
+/*
+ * req_type typically has one of the:
+ * - _PAGE_CACHE_WB
+ * - _PAGE_CACHE_WC
+ * - _PAGE_CACHE_UC_MINUS
+ * - _PAGE_CACHE_UC
+ *
+ * req_type will have a special case value '-1', when requester want to inherit
+ * the memory type from mtrr (if WB), existing PAT, defaulting to UC_MINUS.
+ *
+ * If ret_type is NULL, function will return an error if it cannot reserve the
+ * region with req_type. If ret_type is non-null, function will return
+ * available type in ret_type in case of no error. In case of any error
+ * it will return a negative return value.
+ */
 int reserve_memtype(u64 start, u64 end, unsigned long req_type,
                        unsigned long *ret_type)
 {
@@ -200,9 +225,14 @@ int reserve_memtype(u64 start, u64 end, unsigned long req_type,
 
        /* Only track when pat_wc_enabled */
        if (!pat_wc_enabled) {
-               if (ret_type)
-                       *ret_type = req_type;
-
+               /* This is identical to page table setting without PAT */
+               if (ret_type) {
+                       if (req_type == -1) {
+                               *ret_type = _PAGE_CACHE_WB;
+                       } else {
+                               *ret_type = req_type;
+                       }
+               }
                return 0;
        }
 
@@ -214,8 +244,29 @@ int reserve_memtype(u64 start, u64 end, unsigned long req_type,
                return 0;
        }
 
-       req_type &= _PAGE_CACHE_MASK;
-       err = pat_x_mtrr_type(start, end, req_type, &actual_type);
+       if (req_type == -1) {
+               /*
+                * Special case where caller wants to inherit from mtrr or
+                * existing pat mapping, defaulting to UC_MINUS in case of
+                * no match.
+                */
+               u8 mtrr_type = mtrr_type_lookup(start, end);
+               if (mtrr_type == 0xFE) { /* MTRR match error */
+                       err = -1;
+               }
+
+               if (mtrr_type == MTRR_TYPE_WRBACK) {
+                       req_type = _PAGE_CACHE_WB;
+                       actual_type = _PAGE_CACHE_WB;
+               } else {
+                       req_type = _PAGE_CACHE_UC_MINUS;
+                       actual_type = _PAGE_CACHE_UC_MINUS;
+               }
+       } else {
+               req_type &= _PAGE_CACHE_MASK;
+               err = pat_x_mtrr_type(start, end, req_type, &actual_type);
+       }
+
        if (err) {
                if (ret_type)
                        *ret_type = actual_type;
@@ -241,7 +292,7 @@ int reserve_memtype(u64 start, u64 end, unsigned long req_type,
                struct memtype *saved_ptr;
 
                if (parse->start >= end) {
-                       printk("New Entry\n");
+                       dprintk("New Entry\n");
                        list_add(&new_entry->nd, parse->nd.prev);
                        new_entry = NULL;
                        break;
@@ -291,7 +342,7 @@ int reserve_memtype(u64 start, u64 end, unsigned long req_type,
                                break;
                        }
 
-                       printk("Overlap at 0x%Lx-0x%Lx\n",
+                       dprintk("Overlap at 0x%Lx-0x%Lx\n",
                               saved_ptr->start, saved_ptr->end);
                        /* No conflict. Go ahead and add this new entry */
                        list_add(&new_entry->nd, saved_ptr->nd.prev);
@@ -343,8 +394,8 @@ int reserve_memtype(u64 start, u64 end, unsigned long req_type,
                                break;
                        }
 
-                       printk("Overlap at 0x%Lx-0x%Lx\n",
-                              saved_ptr->start, saved_ptr->end);
+                       dprintk("Overlap at 0x%Lx-0x%Lx\n",
+                                saved_ptr->start, saved_ptr->end);
                        /* No conflict. Go ahead and add this new entry */
                        list_add(&new_entry->nd, &saved_ptr->nd);
                        new_entry = NULL;
@@ -353,7 +404,7 @@ int reserve_memtype(u64 start, u64 end, unsigned long req_type,
        }
 
        if (err) {
-               printk(
+               printk(KERN_INFO
        "reserve_memtype failed 0x%Lx-0x%Lx, track %s, req %s\n",
                        start, end, cattr_name(new_entry->type),
                        cattr_name(req_type));
@@ -365,16 +416,16 @@ int reserve_memtype(u64 start, u64 end, unsigned long req_type,
        if (new_entry) {
                /* No conflict. Not yet added to the list. Add to the tail */
                list_add_tail(&new_entry->nd, &memtype_list);
-               printk("New Entry\n");
-       }
+               dprintk("New Entry\n");
+       }
 
        if (ret_type) {
-               printk(
+               dprintk(
        "reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n",
                        start, end, cattr_name(actual_type),
                        cattr_name(req_type), cattr_name(*ret_type));
        } else {
-               printk(
+               dprintk(
        "reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s\n",
                        start, end, cattr_name(actual_type),
                        cattr_name(req_type));
@@ -411,11 +462,142 @@ int free_memtype(u64 start, u64 end)
        spin_unlock(&memtype_lock);
 
        if (err) {
-               printk(KERN_DEBUG "%s:%d freeing invalid memtype %Lx-%Lx\n",
+               printk(KERN_INFO "%s:%d freeing invalid memtype %Lx-%Lx\n",
                        current->comm, current->pid, start, end);
        }
 
-       printk( "free_memtype request 0x%Lx-0x%Lx\n", start, end);
+       dprintk("free_memtype request 0x%Lx-0x%Lx\n", start, end);
        return err;
 }
 
+
+/*
+ * /dev/mem mmap interface. The memtype used for mapping varies:
+ * - Use UC for mappings with O_SYNC flag
+ * - Without O_SYNC flag, if there is any conflict in reserve_memtype,
+ *   inherit the memtype from existing mapping.
+ * - Else use UC_MINUS memtype (for backward compatibility with existing
+ *   X drivers.
+ */
+pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
+                               unsigned long size, pgprot_t vma_prot)
+{
+       return vma_prot;
+}
+
+#ifdef CONFIG_NONPROMISC_DEVMEM
+/* This check is done in drivers/char/mem.c in case of NONPROMISC_DEVMEM*/
+static inline int range_is_allowed(unsigned long pfn, unsigned long size)
+{
+       return 1;
+}
+#else
+static inline int range_is_allowed(unsigned long pfn, unsigned long size)
+{
+       u64 from = ((u64)pfn) << PAGE_SHIFT;
+       u64 to = from + size;
+       u64 cursor = from;
+
+       while (cursor < to) {
+               if (!devmem_is_allowed(pfn)) {
+                       printk(KERN_INFO
+               "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
+                               current->comm, from, to);
+                       return 0;
+               }
+               cursor += PAGE_SIZE;
+               pfn++;
+       }
+       return 1;
+}
+#endif /* CONFIG_NONPROMISC_DEVMEM */
+
+int phys_mem_access_prot_allowed(struct file *file, unsigned long pfn,
+                               unsigned long size, pgprot_t *vma_prot)
+{
+       u64 offset = ((u64) pfn) << PAGE_SHIFT;
+       unsigned long flags = _PAGE_CACHE_UC_MINUS;
+       int retval;
+
+       if (!range_is_allowed(pfn, size))
+               return 0;
+
+       if (file->f_flags & O_SYNC) {
+               flags = _PAGE_CACHE_UC;
+       }
+
+#ifdef CONFIG_X86_32
+       /*
+        * On the PPro and successors, the MTRRs are used to set
+        * memory types for physical addresses outside main memory,
+        * so blindly setting UC or PWT on those pages is wrong.
+        * For Pentiums and earlier, the surround logic should disable
+        * caching for the high addresses through the KEN pin, but
+        * we maintain the tradition of paranoia in this code.
+        */
+       if (!pat_wc_enabled &&
+           ! ( boot_cpu_has(X86_FEATURE_MTRR) ||
+               boot_cpu_has(X86_FEATURE_K6_MTRR) ||
+               boot_cpu_has(X86_FEATURE_CYRIX_ARR) ||
+               boot_cpu_has(X86_FEATURE_CENTAUR_MCR)) &&
+          (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
+               flags = _PAGE_CACHE_UC;
+       }
+#endif
+
+       /*
+        * With O_SYNC, we can only take UC mapping. Fail if we cannot.
+        * Without O_SYNC, we want to get
+        * - WB for WB-able memory and no other conflicting mappings
+        * - UC_MINUS for non-WB-able memory with no other conflicting mappings
+        * - Inherit from confliting mappings otherwise
+        */
+       if (flags != _PAGE_CACHE_UC_MINUS) {
+               retval = reserve_memtype(offset, offset + size, flags, NULL);
+       } else {
+               retval = reserve_memtype(offset, offset + size, -1, &flags);
+       }
+
+       if (retval < 0)
+               return 0;
+
+       if (pfn <= max_pfn_mapped &&
+            ioremap_change_attr((unsigned long)__va(offset), size, flags) < 0) {
+               free_memtype(offset, offset + size);
+               printk(KERN_INFO
+               "%s:%d /dev/mem ioremap_change_attr failed %s for %Lx-%Lx\n",
+                       current->comm, current->pid,
+                       cattr_name(flags),
+                       offset, (unsigned long long)(offset + size));
+               return 0;
+       }
+
+       *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
+                            flags);
+       return 1;
+}
+
+void map_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
+{
+       u64 addr = (u64)pfn << PAGE_SHIFT;
+       unsigned long flags;
+       unsigned long want_flags = (pgprot_val(vma_prot) & _PAGE_CACHE_MASK);
+
+       reserve_memtype(addr, addr + size, want_flags, &flags);
+       if (flags != want_flags) {
+               printk(KERN_INFO
+               "%s:%d /dev/mem expected mapping type %s for %Lx-%Lx, got %s\n",
+                       current->comm, current->pid,
+                       cattr_name(want_flags),
+                       addr, (unsigned long long)(addr + size),
+                       cattr_name(flags));
+       }
+}
+
+void unmap_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
+{
+       u64 addr = (u64)pfn << PAGE_SHIFT;
+
+       free_memtype(addr, addr + size);
+}
+