[PATCH] perf_event: Provide vmalloc() based mmap() backing

[Peter Zijlstra <peterz@infradead.org>]

OK, so I rewrote it a few more times, does the below make everybody feel
warm and fuzzy?

Dave, I preserved your ACK, since you don't seem to care about the
particular form, as long as the problem gets solved. Feel free to kick
me if I got that wrong ;-)

---
Subject: perf_event: Provide vmalloc() based mmap() backing
From: Peter Zijlstra <a.p.zijlstra@chello.nl>
Date: Mon, 21 Sep 2009 16:08:49 +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.

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>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 arch/sparc/Kconfig         |    2 
 include/linux/perf_event.h |    5 
 init/Kconfig               |   18 +++
 kernel/perf_event.c        |  248 +++++++++++++++++++++++++++++++++------------
 tools/perf/design.txt      |    3 
 5 files changed, 214 insertions(+), 62 deletions(-)

Index: linux-2.6/include/linux/perf_event.h
===================================================================
--- linux-2.6.orig/include/linux/perf_event.h
+++ linux-2.6/include/linux/perf_event.h
@@ -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: linux-2.6/kernel/perf_event.c
===================================================================
--- linux-2.6.orig/kernel/perf_event.c
+++ linux-2.6/kernel/perf_event.c
@@ -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 p
 			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
 	__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_ar
 
 	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, 
 	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, 
 	}
 
 	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 per
 	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
 		      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
 	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: linux-2.6/arch/sparc/Kconfig
===================================================================
--- linux-2.6.orig/arch/sparc/Kconfig
+++ linux-2.6/arch/sparc/Kconfig
@@ -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: linux-2.6/init/Kconfig
===================================================================
--- linux-2.6.orig/init/Kconfig
+++ linux-2.6/init/Kconfig
@@ -930,6 +930,11 @@ 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
@@ -985,6 +990,19 @@ 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: linux-2.6/tools/perf/design.txt
===================================================================
--- linux-2.6.orig/tools/perf/design.txt
+++ linux-2.6/tools/perf/design.txt
@@ -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().