* [for-next][PATCH 0/2] tracing: Remove bottle neck mutex in splicing from ring buffer
@ 2014-12-23 16:13 Steven Rostedt
2014-12-23 16:13 ` [for-next][PATCH 1/2] tracing: Add ref count to tracer for when they are being read by pipe Steven Rostedt
2014-12-23 16:13 ` [for-next][PATCH 2/2] tracing: Remove taking of trace_types_lock in pipe files Steven Rostedt
0 siblings, 2 replies; 3+ messages in thread
From: Steven Rostedt @ 2014-12-23 16:13 UTC (permalink / raw)
To: linux-kernel; +Cc: Ingo Molnar, Andrew Morton
While profiling with trace-cmd, I found that the there was a global
mutex being taken in the per-cpu splice code. This was slowing down
the recording and causing more missed events than necessary.
By adding a ref count to the tracer, and preventing the tracer from
changing when the ref count is non-zero (when a trace file is opened)
we can remove the taking of the trace_types_lock mutex from the
splice page and speed things up a bit.
git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace.git
for-next
Head SHA1: d716ff71dd12bc6328f84a9ec1c3647daf01c827
Steven Rostedt (Red Hat) (2):
tracing: Add ref count to tracer for when they are being read by pipe
tracing: Remove taking of trace_types_lock in pipe files
----
kernel/trace/trace.c | 126 ++++++++++++++++++---------------------------------
kernel/trace/trace.h | 1 +
2 files changed, 44 insertions(+), 83 deletions(-)
^ permalink raw reply [flat|nested] 3+ messages in thread
* [for-next][PATCH 1/2] tracing: Add ref count to tracer for when they are being read by pipe
2014-12-23 16:13 [for-next][PATCH 0/2] tracing: Remove bottle neck mutex in splicing from ring buffer Steven Rostedt
@ 2014-12-23 16:13 ` Steven Rostedt
2014-12-23 16:13 ` [for-next][PATCH 2/2] tracing: Remove taking of trace_types_lock in pipe files Steven Rostedt
1 sibling, 0 replies; 3+ messages in thread
From: Steven Rostedt @ 2014-12-23 16:13 UTC (permalink / raw)
To: linux-kernel; +Cc: Ingo Molnar, Andrew Morton
[-- Attachment #1: 0001-tracing-Add-ref-count-to-tracer-for-when-they-are-be.patch --]
[-- Type: text/plain, Size: 2608 bytes --]
From: "Steven Rostedt (Red Hat)" <rostedt@goodmis.org>
When one of the trace pipe files are being read (by either the trace_pipe
or trace_pipe_raw), do not allow the current_trace to change. By adding
a ref count that is incremented when the pipe files are opened, will
prevent the current_trace from being changed.
This will allow for the removal of the global trace_types_lock from
reading the pipe buffers (which is currently a bottle neck).
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
---
kernel/trace/trace.c | 16 +++++++++++++++-
kernel/trace/trace.h | 1 +
2 files changed, 16 insertions(+), 1 deletion(-)
diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
index 2e767972e99c..ed3fba1d6570 100644
--- a/kernel/trace/trace.c
+++ b/kernel/trace/trace.c
@@ -4140,6 +4140,12 @@ static int tracing_set_tracer(struct trace_array *tr, const char *buf)
goto out;
}
+ /* If trace pipe files are being read, we can't change the tracer */
+ if (tr->current_trace->ref) {
+ ret = -EBUSY;
+ goto out;
+ }
+
trace_branch_disable();
tr->current_trace->enabled--;
@@ -4363,6 +4369,8 @@ static int tracing_open_pipe(struct inode *inode, struct file *filp)
iter->trace->pipe_open(iter);
nonseekable_open(inode, filp);
+
+ tr->current_trace->ref++;
out:
mutex_unlock(&trace_types_lock);
return ret;
@@ -4382,6 +4390,8 @@ static int tracing_release_pipe(struct inode *inode, struct file *file)
mutex_lock(&trace_types_lock);
+ tr->current_trace->ref--;
+
if (iter->trace->pipe_close)
iter->trace->pipe_close(iter);
@@ -5331,6 +5341,8 @@ static int tracing_buffers_open(struct inode *inode, struct file *filp)
filp->private_data = info;
+ tr->current_trace->ref++;
+
mutex_unlock(&trace_types_lock);
ret = nonseekable_open(inode, filp);
@@ -5437,6 +5449,8 @@ static int tracing_buffers_release(struct inode *inode, struct file *file)
mutex_lock(&trace_types_lock);
+ iter->tr->current_trace->ref--;
+
__trace_array_put(iter->tr);
if (info->spare)
@@ -6416,7 +6430,7 @@ static int instance_delete(const char *name)
goto out_unlock;
ret = -EBUSY;
- if (tr->ref)
+ if (tr->ref || (tr->current_trace && tr->current_trace->ref))
goto out_unlock;
list_del(&tr->list);
diff --git a/kernel/trace/trace.h b/kernel/trace/trace.h
index 8de48bac1ce2..0eddfeb05fee 100644
--- a/kernel/trace/trace.h
+++ b/kernel/trace/trace.h
@@ -388,6 +388,7 @@ struct tracer {
struct tracer *next;
struct tracer_flags *flags;
int enabled;
+ int ref;
bool print_max;
bool allow_instances;
#ifdef CONFIG_TRACER_MAX_TRACE
--
2.1.3
^ permalink raw reply [flat|nested] 3+ messages in thread
* [for-next][PATCH 2/2] tracing: Remove taking of trace_types_lock in pipe files
2014-12-23 16:13 [for-next][PATCH 0/2] tracing: Remove bottle neck mutex in splicing from ring buffer Steven Rostedt
2014-12-23 16:13 ` [for-next][PATCH 1/2] tracing: Add ref count to tracer for when they are being read by pipe Steven Rostedt
@ 2014-12-23 16:13 ` Steven Rostedt
1 sibling, 0 replies; 3+ messages in thread
From: Steven Rostedt @ 2014-12-23 16:13 UTC (permalink / raw)
To: linux-kernel; +Cc: Ingo Molnar, Andrew Morton
[-- Attachment #1: 0002-tracing-Remove-taking-of-trace_types_lock-in-pipe-fi.patch --]
[-- Type: text/plain, Size: 6429 bytes --]
From: "Steven Rostedt (Red Hat)" <rostedt@goodmis.org>
Taking the global mutex "trace_types_lock" in the trace_pipe files
causes a bottle neck as most the pipe files can be read per cpu
and there's no reason to serialize them.
The current_trace variable was given a ref count and it can not
change when the ref count is not zero. Opening the trace_pipe
files will up the ref count (and decremented on close), so that
the lock no longer needs to be taken when accessing the
current_trace variable.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
---
kernel/trace/trace.c | 110 +++++++++++++--------------------------------------
1 file changed, 28 insertions(+), 82 deletions(-)
diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
index ed3fba1d6570..7669b1f3234e 100644
--- a/kernel/trace/trace.c
+++ b/kernel/trace/trace.c
@@ -4332,17 +4332,7 @@ static int tracing_open_pipe(struct inode *inode, struct file *filp)
}
trace_seq_init(&iter->seq);
-
- /*
- * We make a copy of the current tracer to avoid concurrent
- * changes on it while we are reading.
- */
- iter->trace = kmalloc(sizeof(*iter->trace), GFP_KERNEL);
- if (!iter->trace) {
- ret = -ENOMEM;
- goto fail;
- }
- *iter->trace = *tr->current_trace;
+ iter->trace = tr->current_trace;
if (!alloc_cpumask_var(&iter->started, GFP_KERNEL)) {
ret = -ENOMEM;
@@ -4399,7 +4389,6 @@ static int tracing_release_pipe(struct inode *inode, struct file *file)
free_cpumask_var(iter->started);
mutex_destroy(&iter->mutex);
- kfree(iter->trace);
kfree(iter);
trace_array_put(tr);
@@ -4432,7 +4421,7 @@ tracing_poll_pipe(struct file *filp, poll_table *poll_table)
return trace_poll(iter, filp, poll_table);
}
-/* Must be called with trace_types_lock mutex held. */
+/* Must be called with iter->mutex held. */
static int tracing_wait_pipe(struct file *filp)
{
struct trace_iterator *iter = filp->private_data;
@@ -4477,7 +4466,6 @@ tracing_read_pipe(struct file *filp, char __user *ubuf,
size_t cnt, loff_t *ppos)
{
struct trace_iterator *iter = filp->private_data;
- struct trace_array *tr = iter->tr;
ssize_t sret;
/* return any leftover data */
@@ -4487,12 +4475,6 @@ tracing_read_pipe(struct file *filp, char __user *ubuf,
trace_seq_init(&iter->seq);
- /* copy the tracer to avoid using a global lock all around */
- mutex_lock(&trace_types_lock);
- if (unlikely(iter->trace->name != tr->current_trace->name))
- *iter->trace = *tr->current_trace;
- mutex_unlock(&trace_types_lock);
-
/*
* Avoid more than one consumer on a single file descriptor
* This is just a matter of traces coherency, the ring buffer itself
@@ -4652,7 +4634,6 @@ static ssize_t tracing_splice_read_pipe(struct file *filp,
.ops = &tracing_pipe_buf_ops,
.spd_release = tracing_spd_release_pipe,
};
- struct trace_array *tr = iter->tr;
ssize_t ret;
size_t rem;
unsigned int i;
@@ -4660,12 +4641,6 @@ static ssize_t tracing_splice_read_pipe(struct file *filp,
if (splice_grow_spd(pipe, &spd))
return -ENOMEM;
- /* copy the tracer to avoid using a global lock all around */
- mutex_lock(&trace_types_lock);
- if (unlikely(iter->trace->name != tr->current_trace->name))
- *iter->trace = *tr->current_trace;
- mutex_unlock(&trace_types_lock);
-
mutex_lock(&iter->mutex);
if (iter->trace->splice_read) {
@@ -5373,21 +5348,16 @@ tracing_buffers_read(struct file *filp, char __user *ubuf,
if (!count)
return 0;
- mutex_lock(&trace_types_lock);
-
#ifdef CONFIG_TRACER_MAX_TRACE
- if (iter->snapshot && iter->tr->current_trace->use_max_tr) {
- size = -EBUSY;
- goto out_unlock;
- }
+ if (iter->snapshot && iter->tr->current_trace->use_max_tr)
+ return -EBUSY;
#endif
if (!info->spare)
info->spare = ring_buffer_alloc_read_page(iter->trace_buffer->buffer,
iter->cpu_file);
- size = -ENOMEM;
if (!info->spare)
- goto out_unlock;
+ return -ENOMEM;
/* Do we have previous read data to read? */
if (info->read < PAGE_SIZE)
@@ -5403,21 +5373,16 @@ tracing_buffers_read(struct file *filp, char __user *ubuf,
if (ret < 0) {
if (trace_empty(iter)) {
- if ((filp->f_flags & O_NONBLOCK)) {
- size = -EAGAIN;
- goto out_unlock;
- }
- mutex_unlock(&trace_types_lock);
+ if ((filp->f_flags & O_NONBLOCK))
+ return -EAGAIN;
+
ret = wait_on_pipe(iter, false);
- mutex_lock(&trace_types_lock);
- if (ret) {
- size = ret;
- goto out_unlock;
- }
+ if (ret)
+ return ret;
+
goto again;
}
- size = 0;
- goto out_unlock;
+ return 0;
}
info->read = 0;
@@ -5427,18 +5392,14 @@ tracing_buffers_read(struct file *filp, char __user *ubuf,
size = count;
ret = copy_to_user(ubuf, info->spare + info->read, size);
- if (ret == size) {
- size = -EFAULT;
- goto out_unlock;
- }
+ if (ret == size)
+ return -EFAULT;
+
size -= ret;
*ppos += size;
info->read += size;
- out_unlock:
- mutex_unlock(&trace_types_lock);
-
return size;
}
@@ -5536,30 +5497,20 @@ tracing_buffers_splice_read(struct file *file, loff_t *ppos,
int entries, size, i;
ssize_t ret = 0;
- mutex_lock(&trace_types_lock);
-
#ifdef CONFIG_TRACER_MAX_TRACE
- if (iter->snapshot && iter->tr->current_trace->use_max_tr) {
- ret = -EBUSY;
- goto out;
- }
+ if (iter->snapshot && iter->tr->current_trace->use_max_tr)
+ return -EBUSY;
#endif
- if (splice_grow_spd(pipe, &spd)) {
- ret = -ENOMEM;
- goto out;
- }
+ if (splice_grow_spd(pipe, &spd))
+ return -ENOMEM;
- if (*ppos & (PAGE_SIZE - 1)) {
- ret = -EINVAL;
- goto out;
- }
+ if (*ppos & (PAGE_SIZE - 1))
+ return -EINVAL;
if (len & (PAGE_SIZE - 1)) {
- if (len < PAGE_SIZE) {
- ret = -EINVAL;
- goto out;
- }
+ if (len < PAGE_SIZE)
+ return -EINVAL;
len &= PAGE_MASK;
}
@@ -5620,25 +5571,20 @@ tracing_buffers_splice_read(struct file *file, loff_t *ppos,
/* did we read anything? */
if (!spd.nr_pages) {
if (ret)
- goto out;
+ return ret;
+
+ if ((file->f_flags & O_NONBLOCK) || (flags & SPLICE_F_NONBLOCK))
+ return -EAGAIN;
- if ((file->f_flags & O_NONBLOCK) || (flags & SPLICE_F_NONBLOCK)) {
- ret = -EAGAIN;
- goto out;
- }
- mutex_unlock(&trace_types_lock);
ret = wait_on_pipe(iter, true);
- mutex_lock(&trace_types_lock);
if (ret)
- goto out;
+ return ret;
goto again;
}
ret = splice_to_pipe(pipe, &spd);
splice_shrink_spd(&spd);
-out:
- mutex_unlock(&trace_types_lock);
return ret;
}
--
2.1.3
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2014-12-23 16:17 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2014-12-23 16:13 [for-next][PATCH 0/2] tracing: Remove bottle neck mutex in splicing from ring buffer Steven Rostedt
2014-12-23 16:13 ` [for-next][PATCH 1/2] tracing: Add ref count to tracer for when they are being read by pipe Steven Rostedt
2014-12-23 16:13 ` [for-next][PATCH 2/2] tracing: Remove taking of trace_types_lock in pipe files Steven Rostedt
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
Powered by JetHome