perf_event: Provide vmalloc() based mmap() backing
Peter Zijlstra [Mon, 21 Sep 2009 14:08:49 +0000 (16:08 +0200)]
Some architectures such as Sparc, ARM and MIPS (basically
everything with flush_dcache_page()) need to deal with dcache
aliases by carefully placing pages in both kernel and user maps.

These architectures typically have to use vmalloc_user() for this.

However, on other architectures, vmalloc() is not needed and has
the downsides of being more restricted and slower than regular
allocations.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: David Miller <davem@davemloft.net>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <1254830228.21044.272.camel@laptop>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

arch/sparc/Kconfig
include/linux/perf_event.h
init/Kconfig
kernel/perf_event.c
tools/perf/design.txt

index 97fca46..9b70a2f 100644 (file)
@@ -26,6 +26,7 @@ config SPARC
        select RTC_CLASS
        select RTC_DRV_M48T59
        select HAVE_PERF_EVENTS
+       select PERF_USE_VMALLOC
        select HAVE_DMA_ATTRS
        select HAVE_DMA_API_DEBUG
 
@@ -48,6 +49,7 @@ config SPARC64
        select RTC_DRV_SUN4V
        select RTC_DRV_STARFIRE
        select HAVE_PERF_EVENTS
+       select PERF_USE_VMALLOC
 
 config ARCH_DEFCONFIG
        string
index 3a9d36d..2e6d95f 100644 (file)
@@ -442,6 +442,7 @@ enum perf_callchain_context {
 #include <linux/hrtimer.h>
 #include <linux/fs.h>
 #include <linux/pid_namespace.h>
+#include <linux/workqueue.h>
 #include <asm/atomic.h>
 
 #define PERF_MAX_STACK_DEPTH           255
@@ -513,6 +514,10 @@ struct file;
 
 struct perf_mmap_data {
        struct rcu_head                 rcu_head;
+#ifdef CONFIG_PERF_USE_VMALLOC
+       struct work_struct              work;
+#endif
+       int                             data_order;
        int                             nr_pages;       /* nr of data pages  */
        int                             writable;       /* are we writable   */
        int                             nr_locked;      /* nr pages mlocked  */
index c7bac39..09c5c64 100644 (file)
@@ -921,6 +921,11 @@ config HAVE_PERF_EVENTS
        help
          See tools/perf/design.txt for details.
 
+config PERF_USE_VMALLOC
+       bool
+       help
+         See tools/perf/design.txt for details
+
 menu "Kernel Performance Events And Counters"
 
 config PERF_EVENTS
@@ -976,6 +981,19 @@ config PERF_COUNTERS
 
          Say N if unsure.
 
+config DEBUG_PERF_USE_VMALLOC
+       default n
+       bool "Debug: use vmalloc to back perf mmap() buffers"
+       depends on PERF_EVENTS && DEBUG_KERNEL
+       select PERF_USE_VMALLOC
+       help
+        Use vmalloc memory to back perf mmap() buffers.
+
+        Mostly useful for debugging the vmalloc code on platforms
+        that don't require it.
+
+        Say N if unsure.
+
 endmenu
 
 config VM_EVENT_COUNTERS
index e491fb0..9d0b5c6 100644 (file)
@@ -20,6 +20,7 @@
 #include <linux/percpu.h>
 #include <linux/ptrace.h>
 #include <linux/vmstat.h>
+#include <linux/vmalloc.h>
 #include <linux/hardirq.h>
 #include <linux/rculist.h>
 #include <linux/uaccess.h>
@@ -2091,49 +2092,31 @@ unlock:
        rcu_read_unlock();
 }
 
-static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
+static unsigned long perf_data_size(struct perf_mmap_data *data)
 {
-       struct perf_event *event = vma->vm_file->private_data;
-       struct perf_mmap_data *data;
-       int ret = VM_FAULT_SIGBUS;
-
-       if (vmf->flags & FAULT_FLAG_MKWRITE) {
-               if (vmf->pgoff == 0)
-                       ret = 0;
-               return ret;
-       }
-
-       rcu_read_lock();
-       data = rcu_dereference(event->data);
-       if (!data)
-               goto unlock;
-
-       if (vmf->pgoff == 0) {
-               vmf->page = virt_to_page(data->user_page);
-       } else {
-               int nr = vmf->pgoff - 1;
-
-               if ((unsigned)nr > data->nr_pages)
-                       goto unlock;
+       return data->nr_pages << (PAGE_SHIFT + data->data_order);
+}
 
-               if (vmf->flags & FAULT_FLAG_WRITE)
-                       goto unlock;
+#ifndef CONFIG_PERF_USE_VMALLOC
 
-               vmf->page = virt_to_page(data->data_pages[nr]);
-       }
+/*
+ * Back perf_mmap() with regular GFP_KERNEL-0 pages.
+ */
 
-       get_page(vmf->page);
-       vmf->page->mapping = vma->vm_file->f_mapping;
-       vmf->page->index   = vmf->pgoff;
+static struct page *
+perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
+{
+       if (pgoff > data->nr_pages)
+               return NULL;
 
-       ret = 0;
-unlock:
-       rcu_read_unlock();
+       if (pgoff == 0)
+               return virt_to_page(data->user_page);
 
-       return ret;
+       return virt_to_page(data->data_pages[pgoff - 1]);
 }
 
-static int perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
+static struct perf_mmap_data *
+perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
 {
        struct perf_mmap_data *data;
        unsigned long size;
@@ -2158,19 +2141,10 @@ static int perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
                        goto fail_data_pages;
        }
 
+       data->data_order = 0;
        data->nr_pages = nr_pages;
-       atomic_set(&data->lock, -1);
-
-       if (event->attr.watermark) {
-               data->watermark = min_t(long, PAGE_SIZE * nr_pages,
-                                     event->attr.wakeup_watermark);
-       }
-       if (!data->watermark)
-               data->watermark = max(PAGE_SIZE, PAGE_SIZE * nr_pages / 4);
 
-       rcu_assign_pointer(event->data, data);
-
-       return 0;
+       return data;
 
 fail_data_pages:
        for (i--; i >= 0; i--)
@@ -2182,7 +2156,7 @@ fail_user_page:
        kfree(data);
 
 fail:
-       return -ENOMEM;
+       return NULL;
 }
 
 static void perf_mmap_free_page(unsigned long addr)
@@ -2193,28 +2167,169 @@ static void perf_mmap_free_page(unsigned long addr)
        __free_page(page);
 }
 
-static void __perf_mmap_data_free(struct rcu_head *rcu_head)
+static void perf_mmap_data_free(struct perf_mmap_data *data)
 {
-       struct perf_mmap_data *data;
        int i;
 
-       data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
-
        perf_mmap_free_page((unsigned long)data->user_page);
        for (i = 0; i < data->nr_pages; i++)
                perf_mmap_free_page((unsigned long)data->data_pages[i]);
+}
+
+#else
+
+/*
+ * Back perf_mmap() with vmalloc memory.
+ *
+ * Required for architectures that have d-cache aliasing issues.
+ */
+
+static struct page *
+perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
+{
+       if (pgoff > (1UL << data->data_order))
+               return NULL;
+
+       return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
+}
+
+static void perf_mmap_unmark_page(void *addr)
+{
+       struct page *page = vmalloc_to_page(addr);
+
+       page->mapping = NULL;
+}
+
+static void perf_mmap_data_free_work(struct work_struct *work)
+{
+       struct perf_mmap_data *data;
+       void *base;
+       int i, nr;
+
+       data = container_of(work, struct perf_mmap_data, work);
+       nr = 1 << data->data_order;
+
+       base = data->user_page;
+       for (i = 0; i < nr + 1; i++)
+               perf_mmap_unmark_page(base + (i * PAGE_SIZE));
+
+       vfree(base);
+}
+
+static void perf_mmap_data_free(struct perf_mmap_data *data)
+{
+       schedule_work(&data->work);
+}
+
+static struct perf_mmap_data *
+perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
+{
+       struct perf_mmap_data *data;
+       unsigned long size;
+       void *all_buf;
 
+       WARN_ON(atomic_read(&event->mmap_count));
+
+       size = sizeof(struct perf_mmap_data);
+       size += sizeof(void *);
+
+       data = kzalloc(size, GFP_KERNEL);
+       if (!data)
+               goto fail;
+
+       INIT_WORK(&data->work, perf_mmap_data_free_work);
+
+       all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
+       if (!all_buf)
+               goto fail_all_buf;
+
+       data->user_page = all_buf;
+       data->data_pages[0] = all_buf + PAGE_SIZE;
+       data->data_order = ilog2(nr_pages);
+       data->nr_pages = 1;
+
+       return data;
+
+fail_all_buf:
+       kfree(data);
+
+fail:
+       return NULL;
+}
+
+#endif
+
+static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
+{
+       struct perf_event *event = vma->vm_file->private_data;
+       struct perf_mmap_data *data;
+       int ret = VM_FAULT_SIGBUS;
+
+       if (vmf->flags & FAULT_FLAG_MKWRITE) {
+               if (vmf->pgoff == 0)
+                       ret = 0;
+               return ret;
+       }
+
+       rcu_read_lock();
+       data = rcu_dereference(event->data);
+       if (!data)
+               goto unlock;
+
+       if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
+               goto unlock;
+
+       vmf->page = perf_mmap_to_page(data, vmf->pgoff);
+       if (!vmf->page)
+               goto unlock;
+
+       get_page(vmf->page);
+       vmf->page->mapping = vma->vm_file->f_mapping;
+       vmf->page->index   = vmf->pgoff;
+
+       ret = 0;
+unlock:
+       rcu_read_unlock();
+
+       return ret;
+}
+
+static void
+perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
+{
+       long max_size = perf_data_size(data);
+
+       atomic_set(&data->lock, -1);
+
+       if (event->attr.watermark) {
+               data->watermark = min_t(long, max_size,
+                                       event->attr.wakeup_watermark);
+       }
+
+       if (!data->watermark)
+               data->watermark = max_t(long, PAGE_SIZE, max_size / 2);
+
+
+       rcu_assign_pointer(event->data, data);
+}
+
+static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
+{
+       struct perf_mmap_data *data;
+
+       data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
+       perf_mmap_data_free(data);
        kfree(data);
 }
 
-static void perf_mmap_data_free(struct perf_event *event)
+static void perf_mmap_data_release(struct perf_event *event)
 {
        struct perf_mmap_data *data = event->data;
 
        WARN_ON(atomic_read(&event->mmap_count));
 
        rcu_assign_pointer(event->data, NULL);
-       call_rcu(&data->rcu_head, __perf_mmap_data_free);
+       call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
 }
 
 static void perf_mmap_open(struct vm_area_struct *vma)
@@ -2230,11 +2345,12 @@ static void perf_mmap_close(struct vm_area_struct *vma)
 
        WARN_ON_ONCE(event->ctx->parent_ctx);
        if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
+               unsigned long size = perf_data_size(event->data);
                struct user_struct *user = current_user();
 
-               atomic_long_sub(event->data->nr_pages + 1, &user->locked_vm);
+               atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
                vma->vm_mm->locked_vm -= event->data->nr_locked;
-               perf_mmap_data_free(event);
+               perf_mmap_data_release(event);
                mutex_unlock(&event->mmap_mutex);
        }
 }
@@ -2252,6 +2368,7 @@ static int perf_mmap(struct file *file, struct vm_area_struct *vma)
        unsigned long user_locked, user_lock_limit;
        struct user_struct *user = current_user();
        unsigned long locked, lock_limit;
+       struct perf_mmap_data *data;
        unsigned long vma_size;
        unsigned long nr_pages;
        long user_extra, extra;
@@ -2314,10 +2431,15 @@ static int perf_mmap(struct file *file, struct vm_area_struct *vma)
        }
 
        WARN_ON(event->data);
-       ret = perf_mmap_data_alloc(event, nr_pages);
-       if (ret)
+
+       data = perf_mmap_data_alloc(event, nr_pages);
+       ret = -ENOMEM;
+       if (!data)
                goto unlock;
 
+       ret = 0;
+       perf_mmap_data_init(event, data);
+
        atomic_set(&event->mmap_count, 1);
        atomic_long_add(user_extra, &user->locked_vm);
        vma->vm_mm->locked_vm += extra;
@@ -2505,7 +2627,7 @@ static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
        if (!data->writable)
                return true;
 
-       mask = (data->nr_pages << PAGE_SHIFT) - 1;
+       mask = perf_data_size(data) - 1;
 
        offset = (offset - tail) & mask;
        head   = (head   - tail) & mask;
@@ -2610,7 +2732,7 @@ void perf_output_copy(struct perf_output_handle *handle,
                      const void *buf, unsigned int len)
 {
        unsigned int pages_mask;
-       unsigned int offset;
+       unsigned long offset;
        unsigned int size;
        void **pages;
 
@@ -2619,12 +2741,14 @@ void perf_output_copy(struct perf_output_handle *handle,
        pages           = handle->data->data_pages;
 
        do {
-               unsigned int page_offset;
+               unsigned long page_offset;
+               unsigned long page_size;
                int nr;
 
                nr          = (offset >> PAGE_SHIFT) & pages_mask;
-               page_offset = offset & (PAGE_SIZE - 1);
-               size        = min_t(unsigned int, PAGE_SIZE - page_offset, len);
+               page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
+               page_offset = offset & (page_size - 1);
+               size        = min_t(unsigned int, page_size - page_offset, len);
 
                memcpy(pages[nr] + page_offset, buf, size);
 
index f1946d1..fdd42a8 100644 (file)
@@ -455,3 +455,6 @@ will need at least this:
 
 If your architecture does have hardware capabilities, you can override the
 weak stub hw_perf_event_init() to register hardware counters.
+
+Architectures that have d-cache aliassing issues, such as Sparc and ARM,
+should select PERF_USE_VMALLOC in order to avoid these for perf mmap().