[linux-2.6-block.git] / kernel / events / ring_buffer.c

/*
 * Performance events ring-buffer code:
 *
 *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
 *  Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
 *  Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
 *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
 *
 * For licensing details see kernel-base/COPYING
 */

#include <linux/perf_event.h>
#include <linux/vmalloc.h>
#include <linux/slab.h>
#include <linux/circ_buf.h>

#include "internal.h"

static void perf_output_wakeup(struct perf_output_handle *handle)
{
	atomic_set(&handle->rb->poll, POLL_IN);

	handle->event->pending_wakeup = 1;
	irq_work_queue(&handle->event->pending);
}

/*
 * We need to ensure a later event_id doesn't publish a head when a former
 * event isn't done writing. However since we need to deal with NMIs we
 * cannot fully serialize things.
 *
 * We only publish the head (and generate a wakeup) when the outer-most
 * event completes.
 */
static void perf_output_get_handle(struct perf_output_handle *handle)
{
	struct ring_buffer *rb = handle->rb;

	preempt_disable();
	local_inc(&rb->nest);
	handle->wakeup = local_read(&rb->wakeup);
}

static void perf_output_put_handle(struct perf_output_handle *handle)
{
	struct ring_buffer *rb = handle->rb;
	unsigned long head;

again:
	head = local_read(&rb->head);

	/*
	 * IRQ/NMI can happen here, which means we can miss a head update.
	 */

	if (!local_dec_and_test(&rb->nest))
		goto out;

	/*
	 * Since the mmap() consumer (userspace) can run on a different CPU:
	 *
	 *   kernel				user
	 *
	 *   READ ->data_tail			READ ->data_head
	 *   smp_mb()	(A)			smp_rmb()	(C)
	 *   WRITE $data			READ $data
	 *   smp_wmb()	(B)			smp_mb()	(D)
	 *   STORE ->data_head			WRITE ->data_tail
	 *
	 * Where A pairs with D, and B pairs with C.
	 *
	 * I don't think A needs to be a full barrier because we won't in fact
	 * write data until we see the store from userspace. So we simply don't
	 * issue the data WRITE until we observe it. Be conservative for now.
	 *
	 * OTOH, D needs to be a full barrier since it separates the data READ
	 * from the tail WRITE.
	 *
	 * For B a WMB is sufficient since it separates two WRITEs, and for C
	 * an RMB is sufficient since it separates two READs.
	 *
	 * See perf_output_begin().
	 */
	smp_wmb();
	rb->user_page->data_head = head;

	/*
	 * Now check if we missed an update, rely on the (compiler)
	 * barrier in atomic_dec_and_test() to re-read rb->head.
	 */
	if (unlikely(head != local_read(&rb->head))) {
		local_inc(&rb->nest);
		goto again;
	}

	if (handle->wakeup != local_read(&rb->wakeup))
		perf_output_wakeup(handle);

out:
	preempt_enable();
}

int perf_output_begin(struct perf_output_handle *handle,
		      struct perf_event *event, unsigned int size)
{
	struct ring_buffer *rb;
	unsigned long tail, offset, head;
	int have_lost;
	struct perf_sample_data sample_data;
	struct {
		struct perf_event_header header;
		u64			 id;
		u64			 lost;
	} lost_event;

	rcu_read_lock();
	/*
	 * For inherited events we send all the output towards the parent.
	 */
	if (event->parent)
		event = event->parent;

	rb = rcu_dereference(event->rb);
	if (!rb)
		goto out;

	handle->rb	= rb;
	handle->event	= event;

	if (!rb->nr_pages)
		goto out;

	have_lost = local_read(&rb->lost);
	if (have_lost) {
		lost_event.header.size = sizeof(lost_event);
		perf_event_header__init_id(&lost_event.header, &sample_data,
					   event);
		size += lost_event.header.size;
	}

	perf_output_get_handle(handle);

	do {
		/*
		 * Userspace could choose to issue a mb() before updating the
		 * tail pointer. So that all reads will be completed before the
		 * write is issued.
		 *
		 * See perf_output_put_handle().
		 */
		tail = ACCESS_ONCE(rb->user_page->data_tail);
		smp_mb();
		offset = head = local_read(&rb->head);
		if (!rb->overwrite &&
		    unlikely(CIRC_SPACE(head, tail, perf_data_size(rb)) < size))
			goto fail;
		head += size;
	} while (local_cmpxchg(&rb->head, offset, head) != offset);

	if (head - local_read(&rb->wakeup) > rb->watermark)
		local_add(rb->watermark, &rb->wakeup);

	handle->page = offset >> (PAGE_SHIFT + page_order(rb));
	handle->page &= rb->nr_pages - 1;
	handle->size = offset & ((PAGE_SIZE << page_order(rb)) - 1);
	handle->addr = rb->data_pages[handle->page];
	handle->addr += handle->size;
	handle->size = (PAGE_SIZE << page_order(rb)) - handle->size;

	if (have_lost) {
		lost_event.header.type = PERF_RECORD_LOST;
		lost_event.header.misc = 0;
		lost_event.id          = event->id;
		lost_event.lost        = local_xchg(&rb->lost, 0);

		perf_output_put(handle, lost_event);
		perf_event__output_id_sample(event, handle, &sample_data);
	}

	return 0;

fail:
	local_inc(&rb->lost);
	perf_output_put_handle(handle);
out:
	rcu_read_unlock();

	return -ENOSPC;
}

unsigned int perf_output_copy(struct perf_output_handle *handle,
		      const void *buf, unsigned int len)
{
	return __output_copy(handle, buf, len);
}

unsigned int perf_output_skip(struct perf_output_handle *handle,
			      unsigned int len)
{
	return __output_skip(handle, NULL, len);
}

void perf_output_end(struct perf_output_handle *handle)
{
	perf_output_put_handle(handle);
	rcu_read_unlock();
}

static void
ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
{
	long max_size = perf_data_size(rb);

	if (watermark)
		rb->watermark = min(max_size, watermark);

	if (!rb->watermark)
		rb->watermark = max_size / 2;

	if (flags & RING_BUFFER_WRITABLE)
		rb->overwrite = 0;
	else
		rb->overwrite = 1;

	atomic_set(&rb->refcount, 1);

	INIT_LIST_HEAD(&rb->event_list);
	spin_lock_init(&rb->event_lock);
}

#ifndef CONFIG_PERF_USE_VMALLOC

/*
 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
 */

struct page *
perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
{
	if (pgoff > rb->nr_pages)
		return NULL;

	if (pgoff == 0)
		return virt_to_page(rb->user_page);

	return virt_to_page(rb->data_pages[pgoff - 1]);
}

static void *perf_mmap_alloc_page(int cpu)
{
	struct page *page;
	int node;

	node = (cpu == -1) ? cpu : cpu_to_node(cpu);
	page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
	if (!page)
		return NULL;

	return page_address(page);
}

struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
{
	struct ring_buffer *rb;
	unsigned long size;
	int i;

	size = sizeof(struct ring_buffer);
	size += nr_pages * sizeof(void *);

	rb = kzalloc(size, GFP_KERNEL);
	if (!rb)
		goto fail;

	rb->user_page = perf_mmap_alloc_page(cpu);
	if (!rb->user_page)
		goto fail_user_page;

	for (i = 0; i < nr_pages; i++) {
		rb->data_pages[i] = perf_mmap_alloc_page(cpu);
		if (!rb->data_pages[i])
			goto fail_data_pages;
	}

	rb->nr_pages = nr_pages;

	ring_buffer_init(rb, watermark, flags);

	return rb;

fail_data_pages:
	for (i--; i >= 0; i--)
		free_page((unsigned long)rb->data_pages[i]);

	free_page((unsigned long)rb->user_page);

fail_user_page:
	kfree(rb);

fail:
	return NULL;
}

static void perf_mmap_free_page(unsigned long addr)
{
	struct page *page = virt_to_page((void *)addr);

	page->mapping = NULL;
	__free_page(page);
}

void rb_free(struct ring_buffer *rb)
{
	int i;

	perf_mmap_free_page((unsigned long)rb->user_page);
	for (i = 0; i < rb->nr_pages; i++)
		perf_mmap_free_page((unsigned long)rb->data_pages[i]);
	kfree(rb);
}

#else
static int data_page_nr(struct ring_buffer *rb)
{
	return rb->nr_pages << page_order(rb);
}

struct page *
perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
{
	/* The '>' counts in the user page. */
	if (pgoff > data_page_nr(rb))
		return NULL;

	return vmalloc_to_page((void *)rb->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 rb_free_work(struct work_struct *work)
{
	struct ring_buffer *rb;
	void *base;
	int i, nr;

	rb = container_of(work, struct ring_buffer, work);
	nr = data_page_nr(rb);

	base = rb->user_page;
	/* The '<=' counts in the user page. */
	for (i = 0; i <= nr; i++)
		perf_mmap_unmark_page(base + (i * PAGE_SIZE));

	vfree(base);
	kfree(rb);
}

void rb_free(struct ring_buffer *rb)
{
	schedule_work(&rb->work);
}

struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
{
	struct ring_buffer *rb;
	unsigned long size;
	void *all_buf;

	size = sizeof(struct ring_buffer);
	size += sizeof(void *);

	rb = kzalloc(size, GFP_KERNEL);
	if (!rb)
		goto fail;

	INIT_WORK(&rb->work, rb_free_work);

	all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
	if (!all_buf)
		goto fail_all_buf;

	rb->user_page = all_buf;
	rb->data_pages[0] = all_buf + PAGE_SIZE;
	rb->page_order = ilog2(nr_pages);
	rb->nr_pages = !!nr_pages;

	ring_buffer_init(rb, watermark, flags);

	return rb;

fail_all_buf:
	kfree(rb);

fail:
	return NULL;
}

#endif
Commit	Line	Data
76369139 FW	1	/*
	2	* Performance events ring-buffer code:
	3	*
	4	* Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
	5	* Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
	6	* Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
d36b6910	7	* Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
76369139 FW	8	*
	9	* For licensing details see kernel-base/COPYING
	10	*/
	11
	12	#include <linux/perf_event.h>
	13	#include <linux/vmalloc.h>
	14	#include <linux/slab.h>
26c86da8	15	#include <linux/circ_buf.h>
76369139 FW	16
	17	#include "internal.h"
	18
76369139 FW	19	static void perf_output_wakeup(struct perf_output_handle *handle)
	20	{
	21	atomic_set(&handle->rb->poll, POLL_IN);
	22
a8b0ca17 PZ	23	handle->event->pending_wakeup = 1;
a8b0ca17 PZ	24	irq_work_queue(&handle->event->pending);
76369139 FW	25	}
	26
	27	/*
	28	* We need to ensure a later event_id doesn't publish a head when a former
	29	* event isn't done writing. However since we need to deal with NMIs we
	30	* cannot fully serialize things.
	31	*
	32	* We only publish the head (and generate a wakeup) when the outer-most
	33	* event completes.
	34	*/
	35	static void perf_output_get_handle(struct perf_output_handle *handle)
	36	{
	37	struct ring_buffer *rb = handle->rb;
	38
	39	preempt_disable();
	40	local_inc(&rb->nest);
	41	handle->wakeup = local_read(&rb->wakeup);
	42	}
	43
	44	static void perf_output_put_handle(struct perf_output_handle *handle)
	45	{
	46	struct ring_buffer *rb = handle->rb;
	47	unsigned long head;
	48
	49	again:
	50	head = local_read(&rb->head);
	51
	52	/*
	53	* IRQ/NMI can happen here, which means we can miss a head update.
	54	*/
	55
	56	if (!local_dec_and_test(&rb->nest))
	57	goto out;
	58
	59	/*
bf378d34 PZ	60	* Since the mmap() consumer (userspace) can run on a different CPU:
	61	*
	62	* kernel user
	63	*
	64	* READ ->data_tail READ ->data_head
	65	* smp_mb() (A) smp_rmb() (C)
	66	* WRITE $data READ $data
	67	* smp_wmb() (B) smp_mb() (D)
	68	* STORE ->data_head WRITE ->data_tail
	69	*
	70	* Where A pairs with D, and B pairs with C.
	71	*
	72	* I don't think A needs to be a full barrier because we won't in fact
	73	* write data until we see the store from userspace. So we simply don't
	74	* issue the data WRITE until we observe it. Be conservative for now.
	75	*
	76	* OTOH, D needs to be a full barrier since it separates the data READ
	77	* from the tail WRITE.
	78	*
	79	* For B a WMB is sufficient since it separates two WRITEs, and for C
	80	* an RMB is sufficient since it separates two READs.
	81	*
	82	* See perf_output_begin().
76369139	83	*/
bf378d34	84	smp_wmb();
76369139 FW	85	rb->user_page->data_head = head;
	86
	87	/*
	88	* Now check if we missed an update, rely on the (compiler)
	89	* barrier in atomic_dec_and_test() to re-read rb->head.
	90	*/
	91	if (unlikely(head != local_read(&rb->head))) {
	92	local_inc(&rb->nest);
	93	goto again;
	94	}
	95
	96	if (handle->wakeup != local_read(&rb->wakeup))
	97	perf_output_wakeup(handle);
	98
	99	out:
	100	preempt_enable();
	101	}
	102
	103	int perf_output_begin(struct perf_output_handle *handle,
a7ac67ea	104	struct perf_event *event, unsigned int size)
76369139 FW	105	{
	106	struct ring_buffer *rb;
	107	unsigned long tail, offset, head;
	108	int have_lost;
	109	struct perf_sample_data sample_data;
	110	struct {
	111	struct perf_event_header header;
	112	u64 id;
	113	u64 lost;
	114	} lost_event;
	115
	116	rcu_read_lock();
	117	/*
	118	* For inherited events we send all the output towards the parent.
	119	*/
	120	if (event->parent)
	121	event = event->parent;
	122
	123	rb = rcu_dereference(event->rb);
	124	if (!rb)
	125	goto out;
	126
	127	handle->rb = rb;
	128	handle->event = event;
76369139 FW	129
	130	if (!rb->nr_pages)
	131	goto out;
	132
	133	have_lost = local_read(&rb->lost);
	134	if (have_lost) {
	135	lost_event.header.size = sizeof(lost_event);
	136	perf_event_header__init_id(&lost_event.header, &sample_data,
	137	event);
	138	size += lost_event.header.size;
	139	}
	140
	141	perf_output_get_handle(handle);
	142
	143	do {
	144	/*
	145	* Userspace could choose to issue a mb() before updating the
	146	* tail pointer. So that all reads will be completed before the
	147	* write is issued.
bf378d34 PZ	148	*
bf378d34 PZ	149	* See perf_output_put_handle().
76369139 FW	150	*/
76369139 FW	151	tail = ACCESS_ONCE(rb->user_page->data_tail);
bf378d34	152	smp_mb();
76369139	153	offset = head = local_read(&rb->head);
26c86da8 PZ	154	if (!rb->overwrite &&
26c86da8 PZ	155	unlikely(CIRC_SPACE(head, tail, perf_data_size(rb)) < size))
76369139	156	goto fail;
26c86da8	157	head += size;
76369139 FW	158	} while (local_cmpxchg(&rb->head, offset, head) != offset);
	159
	160	if (head - local_read(&rb->wakeup) > rb->watermark)
	161	local_add(rb->watermark, &rb->wakeup);
	162
	163	handle->page = offset >> (PAGE_SHIFT + page_order(rb));
	164	handle->page &= rb->nr_pages - 1;
	165	handle->size = offset & ((PAGE_SIZE << page_order(rb)) - 1);
	166	handle->addr = rb->data_pages[handle->page];
	167	handle->addr += handle->size;
	168	handle->size = (PAGE_SIZE << page_order(rb)) - handle->size;
	169
	170	if (have_lost) {
	171	lost_event.header.type = PERF_RECORD_LOST;
	172	lost_event.header.misc = 0;
	173	lost_event.id = event->id;
	174	lost_event.lost = local_xchg(&rb->lost, 0);
	175
	176	perf_output_put(handle, lost_event);
	177	perf_event__output_id_sample(event, handle, &sample_data);
	178	}
	179
	180	return 0;
	181
	182	fail:
	183	local_inc(&rb->lost);
	184	perf_output_put_handle(handle);
	185	out:
	186	rcu_read_unlock();
	187
	188	return -ENOSPC;
	189	}
	190
91d7753a	191	unsigned int perf_output_copy(struct perf_output_handle *handle,
76369139 FW	192	const void *buf, unsigned int len)
76369139 FW	193	{
91d7753a	194	return __output_copy(handle, buf, len);
76369139 FW	195	}
76369139 FW	196
5685e0ff JO	197	unsigned int perf_output_skip(struct perf_output_handle *handle,
	198	unsigned int len)
	199	{
	200	return __output_skip(handle, NULL, len);
	201	}
	202
76369139 FW	203	void perf_output_end(struct perf_output_handle *handle)
76369139 FW	204	{
76369139 FW	205	perf_output_put_handle(handle);
	206	rcu_read_unlock();
	207	}
	208
	209	static void
	210	ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
	211	{
	212	long max_size = perf_data_size(rb);
	213
	214	if (watermark)
	215	rb->watermark = min(max_size, watermark);
	216
	217	if (!rb->watermark)
	218	rb->watermark = max_size / 2;
	219
	220	if (flags & RING_BUFFER_WRITABLE)
dd9c086d SE	221	rb->overwrite = 0;
	222	else
	223	rb->overwrite = 1;
76369139 FW	224
76369139 FW	225	atomic_set(&rb->refcount, 1);
10c6db11 PZ	226
	227	INIT_LIST_HEAD(&rb->event_list);
	228	spin_lock_init(&rb->event_lock);
76369139 FW	229	}
	230
	231	#ifndef CONFIG_PERF_USE_VMALLOC
	232
	233	/*
	234	* Back perf_mmap() with regular GFP_KERNEL-0 pages.
	235	*/
	236
	237	struct page *
	238	perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
	239	{
	240	if (pgoff > rb->nr_pages)
	241	return NULL;
	242
	243	if (pgoff == 0)
	244	return virt_to_page(rb->user_page);
	245
	246	return virt_to_page(rb->data_pages[pgoff - 1]);
	247	}
	248
	249	static void *perf_mmap_alloc_page(int cpu)
	250	{
	251	struct page *page;
	252	int node;
	253
	254	node = (cpu == -1) ? cpu : cpu_to_node(cpu);
	255	page = alloc_pages_node(node, GFP_KERNEL \| __GFP_ZERO, 0);
	256	if (!page)
	257	return NULL;
	258
	259	return page_address(page);
	260	}
	261
	262	struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
	263	{
	264	struct ring_buffer *rb;
	265	unsigned long size;
	266	int i;
	267
	268	size = sizeof(struct ring_buffer);
	269	size += nr_pages * sizeof(void *);
	270
	271	rb = kzalloc(size, GFP_KERNEL);
	272	if (!rb)
	273	goto fail;
	274
	275	rb->user_page = perf_mmap_alloc_page(cpu);
	276	if (!rb->user_page)
	277	goto fail_user_page;
	278
	279	for (i = 0; i < nr_pages; i++) {
	280	rb->data_pages[i] = perf_mmap_alloc_page(cpu);
	281	if (!rb->data_pages[i])
	282	goto fail_data_pages;
	283	}
	284
	285	rb->nr_pages = nr_pages;
	286
	287	ring_buffer_init(rb, watermark, flags);
	288
	289	return rb;
	290
	291	fail_data_pages:
	292	for (i--; i >= 0; i--)
293	free_page((unsigned long)rb->data_pages[i]);
294
295	free_page((unsigned long)rb->user_page);
296
297	fail_user_page:
298	kfree(rb);
299
300	fail:
301	return NULL;
302	}
303
304	static void perf_mmap_free_page(unsigned long addr)
305	{
306	struct page page = virt_to_page((void )addr);
307
308	page->mapping = NULL;
309	__free_page(page);
310	}
311
312	void rb_free(struct ring_buffer *rb)
313	{
314	int i;
315
316	perf_mmap_free_page((unsigned long)rb->user_page);
317	for (i = 0; i < rb->nr_pages; i++)
318	perf_mmap_free_page((unsigned long)rb->data_pages[i]);
319	kfree(rb);
320	}
321
322	#else
5919b309 JO	323	static int data_page_nr(struct ring_buffer *rb)
	324	{
	325	return rb->nr_pages << page_order(rb);
	326	}
76369139 FW	327
	328	struct page *
	329	perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
	330	{
5919b309 JO	331	/* The '>' counts in the user page. */
5919b309 JO	332	if (pgoff > data_page_nr(rb))
76369139 FW	333	return NULL;
	334
	335	return vmalloc_to_page((void )rb->user_page + pgoff PAGE_SIZE);
	336	}
	337
	338	static void perf_mmap_unmark_page(void *addr)
	339	{
	340	struct page *page = vmalloc_to_page(addr);
	341
	342	page->mapping = NULL;
	343	}
	344
	345	static void rb_free_work(struct work_struct *work)
	346	{
	347	struct ring_buffer *rb;
	348	void *base;
	349	int i, nr;
	350
	351	rb = container_of(work, struct ring_buffer, work);
5919b309	352	nr = data_page_nr(rb);
76369139 FW	353
76369139 FW	354	base = rb->user_page;
5919b309 JO	355	/* The '<=' counts in the user page. */
5919b309 JO	356	for (i = 0; i <= nr; i++)
76369139 FW	357	perf_mmap_unmark_page(base + (i * PAGE_SIZE));
	358
	359	vfree(base);
	360	kfree(rb);
	361	}
	362
	363	void rb_free(struct ring_buffer *rb)
	364	{
	365	schedule_work(&rb->work);
	366	}
	367
	368	struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
	369	{
	370	struct ring_buffer *rb;
	371	unsigned long size;
	372	void *all_buf;
	373
	374	size = sizeof(struct ring_buffer);
	375	size += sizeof(void *);
	376
	377	rb = kzalloc(size, GFP_KERNEL);
	378	if (!rb)
	379	goto fail;
	380
	381	INIT_WORK(&rb->work, rb_free_work);
	382
	383	all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
	384	if (!all_buf)
	385	goto fail_all_buf;
	386
	387	rb->user_page = all_buf;
	388	rb->data_pages[0] = all_buf + PAGE_SIZE;
	389	rb->page_order = ilog2(nr_pages);
5919b309	390	rb->nr_pages = !!nr_pages;
76369139 FW	391
	392	ring_buffer_init(rb, watermark, flags);
	393
	394	return rb;
	395
	396	fail_all_buf:
	397	kfree(rb);
	398
	399	fail:
	400	return NULL;
	401	}
	402
	403	#endif