Merge master.kernel.org:/pub/scm/linux/kernel/git/paulus/powerpc
[sfrench/cifs-2.6.git] / arch / powerpc / mm / hash_utils_64.c
index 89b35c1813141c41172295325d8684c4d88c568b..c0d2a694fa3031e5d7ab8d6cce1a371f3a71d044 100644 (file)
@@ -21,7 +21,6 @@
 #undef DEBUG
 #undef DEBUG_LOW
 
-#include <linux/config.h>
 #include <linux/spinlock.h>
 #include <linux/errno.h>
 #include <linux/sched.h>
@@ -92,10 +91,15 @@ unsigned long htab_size_bytes;
 unsigned long htab_hash_mask;
 int mmu_linear_psize = MMU_PAGE_4K;
 int mmu_virtual_psize = MMU_PAGE_4K;
+int mmu_vmalloc_psize = MMU_PAGE_4K;
+int mmu_io_psize = MMU_PAGE_4K;
 #ifdef CONFIG_HUGETLB_PAGE
 int mmu_huge_psize = MMU_PAGE_16M;
 unsigned int HPAGE_SHIFT;
 #endif
+#ifdef CONFIG_PPC_64K_PAGES
+int mmu_ci_restrictions;
+#endif
 
 /* There are definitions of page sizes arrays to be used when none
  * is provided by the firmware.
@@ -162,34 +166,12 @@ int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
                hash = hpt_hash(va, shift);
                hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
 
-               /* The crap below can be cleaned once ppd_md.probe() can
-                * set up the hash callbacks, thus we can just used the
-                * normal insert callback here.
-                */
-#ifdef CONFIG_PPC_ISERIES
-               if (_machine == PLATFORM_ISERIES_LPAR)
-                       ret = iSeries_hpte_insert(hpteg, va,
-                                                 paddr,
-                                                 tmp_mode,
-                                                 HPTE_V_BOLTED,
-                                                 psize);
-               else
-#endif
-#ifdef CONFIG_PPC_PSERIES
-               if (_machine & PLATFORM_LPAR)
-                       ret = pSeries_lpar_hpte_insert(hpteg, va,
-                                                      paddr,
-                                                      tmp_mode,
-                                                      HPTE_V_BOLTED,
-                                                      psize);
-               else
-#endif
-#ifdef CONFIG_PPC_MULTIPLATFORM
-                       ret = native_hpte_insert(hpteg, va,
-                                                paddr,
-                                                tmp_mode, HPTE_V_BOLTED,
-                                                psize);
-#endif
+               DBG("htab_bolt_mapping: calling %p\n", ppc_md.hpte_insert);
+
+               BUG_ON(!ppc_md.hpte_insert);
+               ret = ppc_md.hpte_insert(hpteg, va, paddr,
+                               tmp_mode, HPTE_V_BOLTED, psize);
+
                if (ret < 0)
                        break;
        }
@@ -295,8 +277,7 @@ static void __init htab_init_page_sizes(void)
         * Not in the device-tree, let's fallback on known size
         * list for 16M capable GP & GR
         */
-       if ((_machine != PLATFORM_ISERIES_LPAR) &&
-           cpu_has_feature(CPU_FTR_16M_PAGE))
+       if (cpu_has_feature(CPU_FTR_16M_PAGE))
                memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
                       sizeof(mmu_psize_defaults_gp));
  found:
@@ -309,20 +290,31 @@ static void __init htab_init_page_sizes(void)
        else if (mmu_psize_defs[MMU_PAGE_1M].shift)
                mmu_linear_psize = MMU_PAGE_1M;
 
+#ifdef CONFIG_PPC_64K_PAGES
        /*
         * Pick a size for the ordinary pages. Default is 4K, we support
-        * 64K if cache inhibited large pages are supported by the
-        * processor
+        * 64K for user mappings and vmalloc if supported by the processor.
+        * We only use 64k for ioremap if the processor
+        * (and firmware) support cache-inhibited large pages.
+        * If not, we use 4k and set mmu_ci_restrictions so that
+        * hash_page knows to switch processes that use cache-inhibited
+        * mappings to 4k pages.
         */
-#ifdef CONFIG_PPC_64K_PAGES
-       if (mmu_psize_defs[MMU_PAGE_64K].shift &&
-           cpu_has_feature(CPU_FTR_CI_LARGE_PAGE))
+       if (mmu_psize_defs[MMU_PAGE_64K].shift) {
                mmu_virtual_psize = MMU_PAGE_64K;
+               mmu_vmalloc_psize = MMU_PAGE_64K;
+               if (cpu_has_feature(CPU_FTR_CI_LARGE_PAGE))
+                       mmu_io_psize = MMU_PAGE_64K;
+               else
+                       mmu_ci_restrictions = 1;
+       }
 #endif
 
-       printk(KERN_INFO "Page orders: linear mapping = %d, others = %d\n",
+       printk(KERN_DEBUG "Page orders: linear mapping = %d, "
+              "virtual = %d, io = %d\n",
               mmu_psize_defs[mmu_linear_psize].shift,
-              mmu_psize_defs[mmu_virtual_psize].shift);
+              mmu_psize_defs[mmu_virtual_psize].shift,
+              mmu_psize_defs[mmu_io_psize].shift);
 
 #ifdef CONFIG_HUGETLB_PAGE
        /* Init large page size. Currently, we pick 16M or 1M depending
@@ -398,6 +390,41 @@ void create_section_mapping(unsigned long start, unsigned long end)
 }
 #endif /* CONFIG_MEMORY_HOTPLUG */
 
+static inline void make_bl(unsigned int *insn_addr, void *func)
+{
+       unsigned long funcp = *((unsigned long *)func);
+       int offset = funcp - (unsigned long)insn_addr;
+
+       *insn_addr = (unsigned int)(0x48000001 | (offset & 0x03fffffc));
+       flush_icache_range((unsigned long)insn_addr, 4+
+                          (unsigned long)insn_addr);
+}
+
+static void __init htab_finish_init(void)
+{
+       extern unsigned int *htab_call_hpte_insert1;
+       extern unsigned int *htab_call_hpte_insert2;
+       extern unsigned int *htab_call_hpte_remove;
+       extern unsigned int *htab_call_hpte_updatepp;
+
+#ifdef CONFIG_PPC_64K_PAGES
+       extern unsigned int *ht64_call_hpte_insert1;
+       extern unsigned int *ht64_call_hpte_insert2;
+       extern unsigned int *ht64_call_hpte_remove;
+       extern unsigned int *ht64_call_hpte_updatepp;
+
+       make_bl(ht64_call_hpte_insert1, ppc_md.hpte_insert);
+       make_bl(ht64_call_hpte_insert2, ppc_md.hpte_insert);
+       make_bl(ht64_call_hpte_remove, ppc_md.hpte_remove);
+       make_bl(ht64_call_hpte_updatepp, ppc_md.hpte_updatepp);
+#endif /* CONFIG_PPC_64K_PAGES */
+
+       make_bl(htab_call_hpte_insert1, ppc_md.hpte_insert);
+       make_bl(htab_call_hpte_insert2, ppc_md.hpte_insert);
+       make_bl(htab_call_hpte_remove, ppc_md.hpte_remove);
+       make_bl(htab_call_hpte_updatepp, ppc_md.hpte_updatepp);
+}
+
 void __init htab_initialize(void)
 {
        unsigned long table;
@@ -510,6 +537,8 @@ void __init htab_initialize(void)
                                         mmu_linear_psize));
        }
 
+       htab_finish_init();
+
        DBG(" <- htab_initialize()\n");
 }
 #undef KB
@@ -557,6 +586,7 @@ int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
        pte_t *ptep;
        cpumask_t tmp;
        int rc, user_region = 0, local = 0;
+       int psize;
 
        DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
                ea, access, trap);
@@ -576,10 +606,15 @@ int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
                        return 1;
                }
                vsid = get_vsid(mm->context.id, ea);
+               psize = mm->context.user_psize;
                break;
        case VMALLOC_REGION_ID:
                mm = &init_mm;
                vsid = get_kernel_vsid(ea);
+               if (ea < VMALLOC_END)
+                       psize = mmu_vmalloc_psize;
+               else
+                       psize = mmu_io_psize;
                break;
        default:
                /* Not a valid range
@@ -630,7 +665,40 @@ int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
 #ifndef CONFIG_PPC_64K_PAGES
        rc = __hash_page_4K(ea, access, vsid, ptep, trap, local);
 #else
-       if (mmu_virtual_psize == MMU_PAGE_64K)
+       if (mmu_ci_restrictions) {
+               /* If this PTE is non-cacheable, switch to 4k */
+               if (psize == MMU_PAGE_64K &&
+                   (pte_val(*ptep) & _PAGE_NO_CACHE)) {
+                       if (user_region) {
+                               psize = MMU_PAGE_4K;
+                               mm->context.user_psize = MMU_PAGE_4K;
+                               mm->context.sllp = SLB_VSID_USER |
+                                       mmu_psize_defs[MMU_PAGE_4K].sllp;
+                       } else if (ea < VMALLOC_END) {
+                               /*
+                                * some driver did a non-cacheable mapping
+                                * in vmalloc space, so switch vmalloc
+                                * to 4k pages
+                                */
+                               printk(KERN_ALERT "Reducing vmalloc segment "
+                                      "to 4kB pages because of "
+                                      "non-cacheable mapping\n");
+                               psize = mmu_vmalloc_psize = MMU_PAGE_4K;
+                       }
+               }
+               if (user_region) {
+                       if (psize != get_paca()->context.user_psize) {
+                               get_paca()->context = mm->context;
+                               slb_flush_and_rebolt();
+                       }
+               } else if (get_paca()->vmalloc_sllp !=
+                          mmu_psize_defs[mmu_vmalloc_psize].sllp) {
+                       get_paca()->vmalloc_sllp =
+                               mmu_psize_defs[mmu_vmalloc_psize].sllp;
+                       slb_flush_and_rebolt();
+               }
+       }
+       if (psize == MMU_PAGE_64K)
                rc = __hash_page_64K(ea, access, vsid, ptep, trap, local);
        else
                rc = __hash_page_4K(ea, access, vsid, ptep, trap, local);
@@ -682,7 +750,18 @@ void hash_preload(struct mm_struct *mm, unsigned long ea,
 #ifndef CONFIG_PPC_64K_PAGES
        __hash_page_4K(ea, access, vsid, ptep, trap, local);
 #else
-       if (mmu_virtual_psize == MMU_PAGE_64K)
+       if (mmu_ci_restrictions) {
+               /* If this PTE is non-cacheable, switch to 4k */
+               if (mm->context.user_psize == MMU_PAGE_64K &&
+                   (pte_val(*ptep) & _PAGE_NO_CACHE)) {
+                       mm->context.user_psize = MMU_PAGE_4K;
+                       mm->context.sllp = SLB_VSID_USER |
+                               mmu_psize_defs[MMU_PAGE_4K].sllp;
+                       get_paca()->context = mm->context;
+                       slb_flush_and_rebolt();
+               }
+       }
+       if (mm->context.user_psize == MMU_PAGE_64K)
                __hash_page_64K(ea, access, vsid, ptep, trap, local);
        else
                __hash_page_4K(ea, access, vsid, ptep, trap, local);
@@ -722,16 +801,6 @@ void flush_hash_range(unsigned long number, int local)
        }
 }
 
-static inline void make_bl(unsigned int *insn_addr, void *func)
-{
-       unsigned long funcp = *((unsigned long *)func);
-       int offset = funcp - (unsigned long)insn_addr;
-
-       *insn_addr = (unsigned int)(0x48000001 | (offset & 0x03fffffc));
-       flush_icache_range((unsigned long)insn_addr, 4+
-                          (unsigned long)insn_addr);
-}
-
 /*
  * low_hash_fault is called when we the low level hash code failed
  * to instert a PTE due to an hypervisor error
@@ -750,28 +819,3 @@ void low_hash_fault(struct pt_regs *regs, unsigned long address)
        }
        bad_page_fault(regs, address, SIGBUS);
 }
-
-void __init htab_finish_init(void)
-{
-       extern unsigned int *htab_call_hpte_insert1;
-       extern unsigned int *htab_call_hpte_insert2;
-       extern unsigned int *htab_call_hpte_remove;
-       extern unsigned int *htab_call_hpte_updatepp;
-
-#ifdef CONFIG_PPC_64K_PAGES
-       extern unsigned int *ht64_call_hpte_insert1;
-       extern unsigned int *ht64_call_hpte_insert2;
-       extern unsigned int *ht64_call_hpte_remove;
-       extern unsigned int *ht64_call_hpte_updatepp;
-
-       make_bl(ht64_call_hpte_insert1, ppc_md.hpte_insert);
-       make_bl(ht64_call_hpte_insert2, ppc_md.hpte_insert);
-       make_bl(ht64_call_hpte_remove, ppc_md.hpte_remove);
-       make_bl(ht64_call_hpte_updatepp, ppc_md.hpte_updatepp);
-#endif /* CONFIG_PPC_64K_PAGES */
-
-       make_bl(htab_call_hpte_insert1, ppc_md.hpte_insert);
-       make_bl(htab_call_hpte_insert2, ppc_md.hpte_insert);
-       make_bl(htab_call_hpte_remove, ppc_md.hpte_remove);
-       make_bl(htab_call_hpte_updatepp, ppc_md.hpte_updatepp);
-}