staging: zsmalloc: remove x86 dependency
authorSeth Jennings <sjenning@linux.vnet.ibm.com>
Mon, 2 Jul 2012 21:15:49 +0000 (16:15 -0500)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Mon, 9 Jul 2012 18:34:59 +0000 (11:34 -0700)
This patch replaces the page table assisted object mapping
method, which has x86 dependencies, with a arch-independent
method that does a simple copy into a temporary per-cpu
buffer.

While a copy seems like it would be worse than mapping the pages,
tests demonstrate the copying is always faster and, in the case of
running inside a KVM guest, roughly 4x faster.

Signed-off-by: Seth Jennings <sjenning@linux.vnet.ibm.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/staging/zsmalloc/Kconfig
drivers/staging/zsmalloc/zsmalloc-main.c
drivers/staging/zsmalloc/zsmalloc_int.h

index a5ab720..9084565 100644 (file)
@@ -1,9 +1,5 @@
 config ZSMALLOC
        tristate "Memory allocator for compressed pages"
-       # X86 dependency is because of the use of __flush_tlb_one and set_pte
-       # in zsmalloc-main.c.
-       # TODO: convert these to portable functions
-       depends on X86
        default n
        help
          zsmalloc is a slab-based memory allocator designed to store
index 10b0d60..a7a6f22 100644 (file)
@@ -470,6 +470,57 @@ static struct page *find_get_zspage(struct size_class *class)
        return page;
 }
 
+static void zs_copy_map_object(char *buf, struct page *firstpage,
+                               int off, int size)
+{
+       struct page *pages[2];
+       int sizes[2];
+       void *addr;
+
+       pages[0] = firstpage;
+       pages[1] = get_next_page(firstpage);
+       BUG_ON(!pages[1]);
+
+       sizes[0] = PAGE_SIZE - off;
+       sizes[1] = size - sizes[0];
+
+       /* disable page faults to match kmap_atomic() return conditions */
+       pagefault_disable();
+
+       /* copy object to per-cpu buffer */
+       addr = kmap_atomic(pages[0]);
+       memcpy(buf, addr + off, sizes[0]);
+       kunmap_atomic(addr);
+       addr = kmap_atomic(pages[1]);
+       memcpy(buf + sizes[0], addr, sizes[1]);
+       kunmap_atomic(addr);
+}
+
+static void zs_copy_unmap_object(char *buf, struct page *firstpage,
+                               int off, int size)
+{
+       struct page *pages[2];
+       int sizes[2];
+       void *addr;
+
+       pages[0] = firstpage;
+       pages[1] = get_next_page(firstpage);
+       BUG_ON(!pages[1]);
+
+       sizes[0] = PAGE_SIZE - off;
+       sizes[1] = size - sizes[0];
+
+       /* copy per-cpu buffer to object */
+       addr = kmap_atomic(pages[0]);
+       memcpy(addr + off, buf, sizes[0]);
+       kunmap_atomic(addr);
+       addr = kmap_atomic(pages[1]);
+       memcpy(addr, buf + sizes[0], sizes[1]);
+       kunmap_atomic(addr);
+
+       /* enable page faults to match kunmap_atomic() return conditions */
+       pagefault_enable();
+}
 
 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
                                void *pcpu)
@@ -480,18 +531,23 @@ static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
        switch (action) {
        case CPU_UP_PREPARE:
                area = &per_cpu(zs_map_area, cpu);
-               if (area->vm)
-                       break;
-               area->vm = alloc_vm_area(2 * PAGE_SIZE, area->vm_ptes);
-               if (!area->vm)
-                       return notifier_from_errno(-ENOMEM);
+               /*
+                * Make sure we don't leak memory if a cpu UP notification
+                * and zs_init() race and both call zs_cpu_up() on the same cpu
+                */
+               if (area->vm_buf)
+                       return 0;
+               area->vm_buf = (char *)__get_free_page(GFP_KERNEL);
+               if (!area->vm_buf)
+                       return -ENOMEM;
+               return 0;
                break;
        case CPU_DEAD:
        case CPU_UP_CANCELED:
                area = &per_cpu(zs_map_area, cpu);
-               if (area->vm)
-                       free_vm_area(area->vm);
-               area->vm = NULL;
+               if (area->vm_buf)
+                       free_page((unsigned long)area->vm_buf);
+               area->vm_buf = NULL;
                break;
        }
 
@@ -714,22 +770,11 @@ void *zs_map_object(struct zs_pool *pool, unsigned long handle)
        if (off + class->size <= PAGE_SIZE) {
                /* this object is contained entirely within a page */
                area->vm_addr = kmap_atomic(page);
-       } else {
-               /* this object spans two pages */
-               struct page *nextp;
-
-               nextp = get_next_page(page);
-               BUG_ON(!nextp);
-
-
-               set_pte(area->vm_ptes[0], mk_pte(page, PAGE_KERNEL));
-               set_pte(area->vm_ptes[1], mk_pte(nextp, PAGE_KERNEL));
-
-               /* We pre-allocated VM area so mapping can never fail */
-               area->vm_addr = area->vm->addr;
+               return area->vm_addr + off;
        }
 
-       return area->vm_addr + off;
+       zs_copy_map_object(area->vm_buf, page, off, class->size);
+       return area->vm_buf;
 }
 EXPORT_SYMBOL_GPL(zs_map_object);
 
@@ -751,14 +796,10 @@ void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
        off = obj_idx_to_offset(page, obj_idx, class->size);
 
        area = &__get_cpu_var(zs_map_area);
-       if (off + class->size <= PAGE_SIZE) {
+       if (off + class->size <= PAGE_SIZE)
                kunmap_atomic(area->vm_addr);
-       } else {
-               set_pte(area->vm_ptes[0], __pte(0));
-               set_pte(area->vm_ptes[1], __pte(0));
-               __flush_tlb_one((unsigned long)area->vm_addr);
-               __flush_tlb_one((unsigned long)area->vm_addr + PAGE_SIZE);
-       }
+       else
+               zs_copy_unmap_object(area->vm_buf, page, off, class->size);
        put_cpu_var(zs_map_area);
 }
 EXPORT_SYMBOL_GPL(zs_unmap_object);
index 6fd32a9..f760dae 100644 (file)
@@ -110,9 +110,8 @@ enum fullness_group {
 static const int fullness_threshold_frac = 4;
 
 struct mapping_area {
-       struct vm_struct *vm;
-       pte_t *vm_ptes[2];
-       char *vm_addr;
+       char *vm_buf; /* copy buffer for objects that span pages */
+       char *vm_addr; /* address of kmap_atomic()'ed pages */
 };
 
 struct size_class {