Merge branch 'drm-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/airlied...
[pandora-kernel.git] / arch / sparc / kernel / smp_64.c
index b20f253..3691907 100644 (file)
@@ -20,7 +20,8 @@
 #include <linux/cache.h>
 #include <linux/jiffies.h>
 #include <linux/profile.h>
-#include <linux/lmb.h>
+#include <linux/bootmem.h>
+#include <linux/vmalloc.h>
 #include <linux/cpu.h>
 
 #include <asm/head.h>
@@ -47,6 +48,8 @@
 #include <asm/ldc.h>
 #include <asm/hypervisor.h>
 
+#include "cpumap.h"
+
 int sparc64_multi_core __read_mostly;
 
 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
@@ -1313,6 +1316,8 @@ int __cpu_disable(void)
        cpu_clear(cpu, cpu_online_map);
        ipi_call_unlock();
 
+       cpu_map_rebuild();
+
        return 0;
 }
 
@@ -1371,32 +1376,171 @@ void smp_send_stop(void)
 {
 }
 
-void __init real_setup_per_cpu_areas(void)
+/**
+ * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
+ * @cpu: cpu to allocate for
+ * @size: size allocation in bytes
+ * @align: alignment
+ *
+ * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
+ * does the right thing for NUMA regardless of the current
+ * configuration.
+ *
+ * RETURNS:
+ * Pointer to the allocated area on success, NULL on failure.
+ */
+static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size,
+                                       unsigned long align)
 {
-       unsigned long base, shift, paddr, goal, size, i;
-       char *ptr;
+       const unsigned long goal = __pa(MAX_DMA_ADDRESS);
+#ifdef CONFIG_NEED_MULTIPLE_NODES
+       int node = cpu_to_node(cpu);
+       void *ptr;
+
+       if (!node_online(node) || !NODE_DATA(node)) {
+               ptr = __alloc_bootmem(size, align, goal);
+               pr_info("cpu %d has no node %d or node-local memory\n",
+                       cpu, node);
+               pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
+                        cpu, size, __pa(ptr));
+       } else {
+               ptr = __alloc_bootmem_node(NODE_DATA(node),
+                                          size, align, goal);
+               pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
+                        "%016lx\n", cpu, size, node, __pa(ptr));
+       }
+       return ptr;
+#else
+       return __alloc_bootmem(size, align, goal);
+#endif
+}
+
+static size_t pcpur_size __initdata;
+static void **pcpur_ptrs __initdata;
 
-       /* Copy section for each CPU (we discard the original) */
-       goal = PERCPU_ENOUGH_ROOM;
+static struct page * __init pcpur_get_page(unsigned int cpu, int pageno)
+{
+       size_t off = (size_t)pageno << PAGE_SHIFT;
 
-       shift = PAGE_SHIFT;
-       for (size = PAGE_SIZE; size < goal; size <<= 1UL)
-               shift++;
+       if (off >= pcpur_size)
+               return NULL;
 
-       paddr = lmb_alloc(size * NR_CPUS, PAGE_SIZE);
-       if (!paddr) {
-               prom_printf("Cannot allocate per-cpu memory.\n");
-               prom_halt();
+       return virt_to_page(pcpur_ptrs[cpu] + off);
+}
+
+#define PCPU_CHUNK_SIZE (4UL * 1024UL * 1024UL)
+
+static void __init pcpu_map_range(unsigned long start, unsigned long end,
+                                 struct page *page)
+{
+       unsigned long pfn = page_to_pfn(page);
+       unsigned long pte_base;
+
+       BUG_ON((pfn<<PAGE_SHIFT)&(PCPU_CHUNK_SIZE - 1UL));
+
+       pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
+                   _PAGE_CP_4U | _PAGE_CV_4U |
+                   _PAGE_P_4U | _PAGE_W_4U);
+       if (tlb_type == hypervisor)
+               pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
+                           _PAGE_CP_4V | _PAGE_CV_4V |
+                           _PAGE_P_4V | _PAGE_W_4V);
+
+       while (start < end) {
+               pgd_t *pgd = pgd_offset_k(start);
+               unsigned long this_end;
+               pud_t *pud;
+               pmd_t *pmd;
+               pte_t *pte;
+
+               pud = pud_offset(pgd, start);
+               if (pud_none(*pud)) {
+                       pmd_t *new;
+
+                       new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
+                       pud_populate(&init_mm, pud, new);
+               }
+
+               pmd = pmd_offset(pud, start);
+               if (!pmd_present(*pmd)) {
+                       pte_t *new;
+
+                       new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
+                       pmd_populate_kernel(&init_mm, pmd, new);
+               }
+
+               pte = pte_offset_kernel(pmd, start);
+               this_end = (start + PMD_SIZE) & PMD_MASK;
+               if (this_end > end)
+                       this_end = end;
+
+               while (start < this_end) {
+                       unsigned long paddr = pfn << PAGE_SHIFT;
+
+                       pte_val(*pte) = (paddr | pte_base);
+
+                       start += PAGE_SIZE;
+                       pte++;
+                       pfn++;
+               }
+       }
+}
+
+void __init setup_per_cpu_areas(void)
+{
+       size_t dyn_size, static_size = __per_cpu_end - __per_cpu_start;
+       static struct vm_struct vm;
+       unsigned long delta, cpu;
+       size_t pcpu_unit_size;
+       size_t ptrs_size;
+
+       pcpur_size = PFN_ALIGN(static_size + PERCPU_MODULE_RESERVE +
+                              PERCPU_DYNAMIC_RESERVE);
+       dyn_size = pcpur_size - static_size - PERCPU_MODULE_RESERVE;
+
+
+       ptrs_size = PFN_ALIGN(nr_cpu_ids * sizeof(pcpur_ptrs[0]));
+       pcpur_ptrs = alloc_bootmem(ptrs_size);
+
+       for_each_possible_cpu(cpu) {
+               pcpur_ptrs[cpu] = pcpu_alloc_bootmem(cpu, PCPU_CHUNK_SIZE,
+                                                    PCPU_CHUNK_SIZE);
+
+               free_bootmem(__pa(pcpur_ptrs[cpu] + pcpur_size),
+                            PCPU_CHUNK_SIZE - pcpur_size);
+
+               memcpy(pcpur_ptrs[cpu], __per_cpu_load, static_size);
        }
 
-       ptr = __va(paddr);
-       base = ptr - __per_cpu_start;
+       /* allocate address and map */
+       vm.flags = VM_ALLOC;
+       vm.size = nr_cpu_ids * PCPU_CHUNK_SIZE;
+       vm_area_register_early(&vm, PCPU_CHUNK_SIZE);
+
+       for_each_possible_cpu(cpu) {
+               unsigned long start = (unsigned long) vm.addr;
+               unsigned long end;
 
-       for (i = 0; i < NR_CPUS; i++, ptr += size) {
-               __per_cpu_offset(i) = base + (i * size);
-               memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
+               start += cpu * PCPU_CHUNK_SIZE;
+               end = start + PCPU_CHUNK_SIZE;
+               pcpu_map_range(start, end, virt_to_page(pcpur_ptrs[cpu]));
+       }
+
+       pcpu_unit_size = pcpu_setup_first_chunk(pcpur_get_page, static_size,
+                                               PERCPU_MODULE_RESERVE, dyn_size,
+                                               PCPU_CHUNK_SIZE, vm.addr, NULL);
+
+       free_bootmem(__pa(pcpur_ptrs), ptrs_size);
+
+       delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
+       for_each_possible_cpu(cpu) {
+               __per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size;
        }
 
        /* Setup %g5 for the boot cpu.  */
        __local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
+
+       of_fill_in_cpu_data();
+       if (tlb_type == hypervisor)
+               mdesc_fill_in_cpu_data(cpu_all_mask);
 }