vserver 1.9.5.x5
[linux-2.6.git] / arch / parisc / kernel / cache.c
index 27ae25f..f46a07a 100644 (file)
@@ -17,6 +17,7 @@
 #include <linux/mm.h>
 #include <linux/module.h>
 #include <linux/seq_file.h>
+#include <linux/pagemap.h>
 
 #include <asm/pdc.h>
 #include <asm/cache.h>
@@ -32,6 +33,17 @@ int dcache_stride;
 int icache_stride;
 EXPORT_SYMBOL(dcache_stride);
 
+
+#if defined(CONFIG_SMP)
+/* On some machines (e.g. ones with the Merced bus), there can be
+ * only a single PxTLB broadcast at a time; this must be guaranteed
+ * by software.  We put a spinlock around all TLB flushes  to
+ * ensure this.
+ */
+DEFINE_SPINLOCK(pa_tlb_lock);
+EXPORT_SYMBOL(pa_tlb_lock);
+#endif
+
 struct pdc_cache_info cache_info;
 #ifndef CONFIG_PA20
 static struct pdc_btlb_info btlb_info;
@@ -43,6 +55,11 @@ flush_data_cache(void)
 {
        on_each_cpu((void (*)(void *))flush_data_cache_local, NULL, 1, 1);
 }
+void 
+flush_instruction_cache(void)
+{
+       on_each_cpu((void (*)(void *))flush_instruction_cache_local, NULL, 1, 1);
+}
 #endif
 
 void
@@ -68,7 +85,7 @@ update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t pte)
 {
        struct page *page = pte_page(pte);
 
-       if (VALID_PAGE(page) && page_mapping(page) &&
+       if (pfn_valid(page_to_pfn(page)) && page_mapping(page) &&
            test_bit(PG_dcache_dirty, &page->flags)) {
 
                flush_kernel_dcache_page(page_address(page));
@@ -81,10 +98,11 @@ show_cache_info(struct seq_file *m)
 {
        seq_printf(m, "I-cache\t\t: %ld KB\n", 
                cache_info.ic_size/1024 );
-       seq_printf(m, "D-cache\t\t: %ld KB (%s)%s\n", 
+       seq_printf(m, "D-cache\t\t: %ld KB (%s%s, %d-way associative)\n", 
                cache_info.dc_size/1024,
                (cache_info.dc_conf.cc_wt ? "WT":"WB"),
-               (cache_info.dc_conf.cc_sh ? " - shared I/D":"")
+               (cache_info.dc_conf.cc_sh ? ", shared I/D":""),
+               (cache_info.dc_conf.cc_assoc)
        );
 
        seq_printf(m, "ITLB entries\t: %ld\n" "DTLB entries\t: %ld%s\n",
@@ -122,51 +140,60 @@ parisc_cache_init(void)
                panic("parisc_cache_init: pdc_cache_info failed");
 
 #if 0
-       printk(KERN_DEBUG "ic_size %lx dc_size %lx it_size %lx pdc_cache_info %d*long pdc_cache_cf %d\n",
-           cache_info.ic_size,
-           cache_info.dc_size,
-           cache_info.it_size,
-           sizeof (struct pdc_cache_info) / sizeof (long),
-           sizeof (struct pdc_cache_cf)
-       );
-
-       printk(KERN_DEBUG "dc base %x dc stride %x dc count %x dc loop %d\n",
-           cache_info.dc_base,
-           cache_info.dc_stride,
-           cache_info.dc_count,
-           cache_info.dc_loop);
-
-       printk(KERN_DEBUG "dc conf: alias %d block %d line %d wt %d sh %d cst %d assoc %d\n",
-           cache_info.dc_conf.cc_alias,
-           cache_info.dc_conf.cc_block,
-           cache_info.dc_conf.cc_line,
-           cache_info.dc_conf.cc_wt,
-           cache_info.dc_conf.cc_sh,
-           cache_info.dc_conf.cc_cst,
-           cache_info.dc_conf.cc_assoc);
-
-       printk(KERN_DEBUG "ic conf: alias %d block %d line %d wt %d sh %d cst %d assoc %d\n",
-           cache_info.ic_conf.cc_alias,
-           cache_info.ic_conf.cc_block,
-           cache_info.ic_conf.cc_line,
-           cache_info.ic_conf.cc_wt,
-           cache_info.ic_conf.cc_sh,
-           cache_info.ic_conf.cc_cst,
-           cache_info.ic_conf.cc_assoc);
-
-       printk(KERN_DEBUG "dt conf: sh %d page %d cst %d aid %d pad1 %d \n",
-           cache_info.dt_conf.tc_sh,
-           cache_info.dt_conf.tc_page,
-           cache_info.dt_conf.tc_cst,
-           cache_info.dt_conf.tc_aid,
-           cache_info.dt_conf.tc_pad1);
-
-       printk(KERN_DEBUG "it conf: sh %d page %d cst %d aid %d pad1 %d \n",
-           cache_info.it_conf.tc_sh,
-           cache_info.it_conf.tc_page,
-           cache_info.it_conf.tc_cst,
-           cache_info.it_conf.tc_aid,
-           cache_info.it_conf.tc_pad1);
+       printk("ic_size %lx dc_size %lx it_size %lx\n",
+               cache_info.ic_size,
+               cache_info.dc_size,
+               cache_info.it_size);
+
+       printk("DC  base 0x%lx stride 0x%lx count 0x%lx loop 0x%lx\n",
+               cache_info.dc_base,
+               cache_info.dc_stride,
+               cache_info.dc_count,
+               cache_info.dc_loop);
+
+       printk("dc_conf = 0x%lx  alias %d blk %d line %d shift %d\n",
+               *(unsigned long *) (&cache_info.dc_conf),
+               cache_info.dc_conf.cc_alias,
+               cache_info.dc_conf.cc_block,
+               cache_info.dc_conf.cc_line,
+               cache_info.dc_conf.cc_shift);
+       printk("        wt %d sh %d cst %d assoc %d\n",
+               cache_info.dc_conf.cc_wt,
+               cache_info.dc_conf.cc_sh,
+               cache_info.dc_conf.cc_cst,
+               cache_info.dc_conf.cc_assoc);
+
+       printk("IC  base 0x%lx stride 0x%lx count 0x%lx loop 0x%lx\n",
+               cache_info.ic_base,
+               cache_info.ic_stride,
+               cache_info.ic_count,
+               cache_info.ic_loop);
+
+       printk("ic_conf = 0x%lx  alias %d blk %d line %d shift %d\n",
+               *(unsigned long *) (&cache_info.ic_conf),
+               cache_info.ic_conf.cc_alias,
+               cache_info.ic_conf.cc_block,
+               cache_info.ic_conf.cc_line,
+               cache_info.ic_conf.cc_shift);
+       printk("        wt %d sh %d cst %d assoc %d\n",
+               cache_info.ic_conf.cc_wt,
+               cache_info.ic_conf.cc_sh,
+               cache_info.ic_conf.cc_cst,
+               cache_info.ic_conf.cc_assoc);
+
+       printk("D-TLB conf: sh %d page %d cst %d aid %d pad1 %d \n",
+               cache_info.dt_conf.tc_sh,
+               cache_info.dt_conf.tc_page,
+               cache_info.dt_conf.tc_cst,
+               cache_info.dt_conf.tc_aid,
+               cache_info.dt_conf.tc_pad1);
+
+       printk("I-TLB conf: sh %d page %d cst %d aid %d pad1 %d \n",
+               cache_info.it_conf.tc_sh,
+               cache_info.it_conf.tc_page,
+               cache_info.it_conf.tc_cst,
+               cache_info.it_conf.tc_aid,
+               cache_info.it_conf.tc_pad1);
 #endif
 
        split_tlb = 0;
@@ -178,10 +205,14 @@ parisc_cache_init(void)
                split_tlb = 1;
        }
 
-       dcache_stride = (1 << (cache_info.dc_conf.cc_block + 3)) *
-                                               cache_info.dc_conf.cc_line;
-       icache_stride = (1 << (cache_info.ic_conf.cc_block + 3)) *
-                                               cache_info.ic_conf.cc_line;
+       /* "New and Improved" version from Jim Hull 
+        *      (1 << (cc_block-1)) * (cc_line << (4 + cnf.cc_shift))
+        */
+#define CAFL_STRIDE(cnf) (cnf.cc_line << (3 + cnf.cc_block + cnf.cc_shift))
+       dcache_stride = CAFL_STRIDE(cache_info.dc_conf);
+       icache_stride = CAFL_STRIDE(cache_info.ic_conf);
+#undef CAFL_STRIDE
+
 #ifndef CONFIG_PA20
        if (pdc_btlb_info(&btlb_info) < 0) {
                memset(&btlb_info, 0, sizeof btlb_info);
@@ -190,8 +221,8 @@ parisc_cache_init(void)
 
        if ((boot_cpu_data.pdc.capabilities & PDC_MODEL_NVA_MASK) ==
                                                PDC_MODEL_NVA_UNSUPPORTED) {
-               printk(KERN_WARNING "Only equivalent aliasing supported\n");
-#ifndef CONFIG_SMP
+               printk(KERN_WARNING "parisc_cache_init: Only equivalent aliasing supported!\n");
+#if 0
                panic("SMP kernel required to avoid non-equivalent aliasing");
 #endif
        }
@@ -227,76 +258,109 @@ void disable_sr_hashing(void)
        disable_sr_hashing_asm(srhash_type);
 }
 
-void __flush_dcache_page(struct page *page)
+void flush_dcache_page(struct page *page)
 {
        struct address_space *mapping = page_mapping(page);
-       struct mm_struct *mm = current->active_mm;
-       struct list_head *l;
+       struct vm_area_struct *mpnt;
+       struct prio_tree_iter iter;
+       unsigned long offset;
+       unsigned long addr;
+       pgoff_t pgoff;
+       pte_t *pte;
+       unsigned long pfn = page_to_pfn(page);
 
-       flush_kernel_dcache_page(page_address(page));
 
-       if (!mapping)
+       if (mapping && !mapping_mapped(mapping)) {
+               set_bit(PG_dcache_dirty, &page->flags);
                return;
-       /* check shared list first if it's not empty...it's usually
-        * the shortest */
-       list_for_each(l, &mapping->i_mmap_shared) {
-               struct vm_area_struct *mpnt;
-               unsigned long off;
-
-               mpnt = list_entry(l, struct vm_area_struct, shared);
-
-               /*
-                * If this VMA is not in our MM, we can ignore it.
-                */
-               if (mpnt->vm_mm != mm)
-                       continue;
-
-               if (page->index < mpnt->vm_pgoff)
-                       continue;
-
-               off = page->index - mpnt->vm_pgoff;
-               if (off >= (mpnt->vm_end - mpnt->vm_start) >> PAGE_SHIFT)
-                       continue;
+       }
 
-               flush_cache_page(mpnt, mpnt->vm_start + (off << PAGE_SHIFT));
+       flush_kernel_dcache_page(page_address(page));
 
-               /* All user shared mappings should be equivalently mapped,
-                * so once we've flushed one we should be ok
-                */
+       if (!mapping)
                return;
-       }
 
-       /* then check private mapping list for read only shared mappings
-        * which are flagged by VM_MAYSHARE */
-       list_for_each(l, &mapping->i_mmap) {
-               struct vm_area_struct *mpnt;
-               unsigned long off;
+       pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 
-               mpnt = list_entry(l, struct vm_area_struct, shared);
+       /* We have carefully arranged in arch_get_unmapped_area() that
+        * *any* mappings of a file are always congruently mapped (whether
+        * declared as MAP_PRIVATE or MAP_SHARED), so we only need
+        * to flush one address here for them all to become coherent */
 
+       flush_dcache_mmap_lock(mapping);
+       vma_prio_tree_foreach(mpnt, &iter, &mapping->i_mmap, pgoff, pgoff) {
+               offset = (pgoff - mpnt->vm_pgoff) << PAGE_SHIFT;
+               addr = mpnt->vm_start + offset;
 
-               if (mpnt->vm_mm != mm || !(mpnt->vm_flags & VM_MAYSHARE))
-                       continue;
+               /* Flush instructions produce non access tlb misses.
+                * On PA, we nullify these instructions rather than
+                * taking a page fault if the pte doesn't exist.
+                * This is just for speed.  If the page translation
+                * isn't there, there's no point exciting the
+                * nadtlb handler into a nullification frenzy */
 
-               if (page->index < mpnt->vm_pgoff)
-                       continue;
 
-               off = page->index - mpnt->vm_pgoff;
-               if (off >= (mpnt->vm_end - mpnt->vm_start) >> PAGE_SHIFT)
+               if(!(pte = translation_exists(mpnt, addr)))
                        continue;
 
-               flush_cache_page(mpnt, mpnt->vm_start + (off << PAGE_SHIFT));
+               /* make sure we really have this page: the private
+                * mappings may cover this area but have COW'd this
+                * particular page */
+               if(pte_pfn(*pte) != pfn)
+                       continue;
+
+               __flush_cache_page(mpnt, addr);
 
-               /* All user shared mappings should be equivalently mapped,
-                * so once we've flushed one we should be ok
-                */
                break;
        }
+       flush_dcache_mmap_unlock(mapping);
 }
-EXPORT_SYMBOL(__flush_dcache_page);
+EXPORT_SYMBOL(flush_dcache_page);
 
 /* Defined in arch/parisc/kernel/pacache.S */
 EXPORT_SYMBOL(flush_kernel_dcache_range_asm);
 EXPORT_SYMBOL(flush_kernel_dcache_page);
 EXPORT_SYMBOL(flush_data_cache_local);
 EXPORT_SYMBOL(flush_kernel_icache_range_asm);
+
+void clear_user_page_asm(void *page, unsigned long vaddr)
+{
+       /* This function is implemented in assembly in pacache.S */
+       extern void __clear_user_page_asm(void *page, unsigned long vaddr);
+
+       purge_tlb_start();
+       __clear_user_page_asm(page, vaddr);
+       purge_tlb_end();
+}
+
+#define FLUSH_THRESHOLD 0x80000 /* 0.5MB */
+int parisc_cache_flush_threshold = FLUSH_THRESHOLD;
+
+void parisc_setup_cache_timing(void)
+{
+       unsigned long rangetime, alltime;
+       extern char _text;      /* start of kernel code, defined by linker */
+       extern char _end;       /* end of BSS, defined by linker */
+       unsigned long size;
+
+       alltime = mfctl(16);
+       flush_data_cache();
+       alltime = mfctl(16) - alltime;
+
+       size = (unsigned long)(&_end - _text);
+       rangetime = mfctl(16);
+       flush_kernel_dcache_range((unsigned long)&_text, size);
+       rangetime = mfctl(16) - rangetime;
+
+       printk(KERN_DEBUG "Whole cache flush %lu cycles, flushing %lu bytes %lu cycles\n",
+               alltime, size, rangetime);
+
+       /* Racy, but if we see an intermediate value, it's ok too... */
+       parisc_cache_flush_threshold = size * alltime / rangetime;
+
+       parisc_cache_flush_threshold = (parisc_cache_flush_threshold + L1_CACHE_BYTES - 1) &~ (L1_CACHE_BYTES - 1); 
+       if (!parisc_cache_flush_threshold)
+               parisc_cache_flush_threshold = FLUSH_THRESHOLD;
+
+       printk("Setting cache flush threshold to %x (%d CPUs online)\n", parisc_cache_flush_threshold, num_online_cpus());
+}