Date:	Wed, 21 May 2008 01:23:09 +0530
From:	"K.Prasad" <prasad@...ux.vnet.ibm.com>
To:	"Andrew Morton" <akpm@...ux-foundation.org>
Cc:	"K.Prasad" <prasad@...ux.vnet.ibm.com>,
	linux-kernel@...r.kernel.org, dwilder@...ibm.com,
	"hch@...radead.org" <hch@...radead.org>
Subject: Re: [RFC Patch 1/1] trace_printk and trace_dump interface - v2

On Mon, May 19, 2008 at 04:21:49PM -0700, Andrew Morton wrote:
> On Sat, 17 May 2008 07:52:16 +0530
> "K.Prasad" <prasad@...ux.vnet.ibm.com> wrote:
>> Resending this patch due to style issues found in previous patch.
> >> This patch introduces two new interfaces called trace_printk and
> > trace_dump which can be used to print to the debugfs mount directly.
> > It uses the 'trace' infrastructure underneath and is a patch over it.
> > A sample file is also created to demonstrate its ease of use.
> >> Signed-off-by: K.Prasad <prasad@...ux.vnet.ibm.com>
> > ---
> >   Documentation/trace.txt        |   22 ++++
> >   include/linux/trace.h          |   57 +++++++++++
> >   lib/trace.c                    |  205 > ++++++++++++++++++++++++++++++++++++++++-
> >   samples/trace/Makefile         |    2
> >   samples/trace/fork_new_trace.c |   99 +++++++++++++++++++
> The patch is wordwrapped and space-stuffed.
Must be an act of my mail client. Will re-send with proper word-wrapping
and spacing during the next iteration.

>> ...
> >
> >   static void remove_tree(struct trace_info *trace)
> >   {
> > +	struct list_head *pos, *temp;
> > +	struct trace_dir *dr = NULL;
> > +
> >   	mutex_lock(&trace_mutex);
> >   	debugfs_remove(trace->dir);
> >> +	list_for_each_safe(pos, temp, &trace_dirs) {
> > +		dr = list_entry(pos, struct trace_dir, trace_dir_list);
> list_for_each_entry_safe()?
Ok. Will change.

>> +		if (dr->ti == trace) {
> > +			list_del(pos);
> > +			kfree(dr);
> > +		}
> > +	}
> >   	if (trace->root) {
> >   		if (--trace->root->users == 0)
> >   			remove_root(trace);
> > @@ -142,11 +153,17 @@ static struct trace_root *lookup_root(co
> >   static struct dentry *create_tree(struct trace_info *trace, const char > *root,
> >   				  const char *name)
> >   {
> > -	struct dentry *dir = NULL;
> > +	struct trace_dir *temp;
> >>   	if (root == NULL || name == NULL)
> >   		return ERR_PTR(-EINVAL);
> >> +	temp = kzalloc(sizeof(struct trace_dir), GFP_KERNEL);
> > +	if ((temp == NULL) || (strlen(name) > TRACE_NAME_SIZE))
> > +		return ERR_PTR(-ENOMEM);
> This can leak `temp'.
Yes, I missed it. Will modify to look like this:

	if (strlen(name) > TRACE_NAME_SIZE)
		return ERR_PTR(-ENOMEM);

	temp = kzalloc(sizeof(struct trace_dir), GFP_KERNEL);
	if (temp == NULL)
		return ERR_PTR(-ENOMEM);
	
> I suspect it has an off-by-one.
> You should check the incoming args before allocating any resources or
> doing anything which has side-effects.
>> +	strlcpy(temp->trace_dir_name, name, sizeof(temp->trace_dir_name));
> then use kstrdup().
Ok.

> Please don't call variables "temp" or "tmp".  Surely something more
> communcative can be thought up.
Ok. Will name this instance,
struct trace_dir *tdd;

>>   	mutex_lock(&trace_mutex);
> >>   	trace->root = lookup_root(root);
> > @@ -155,17 +172,49 @@ static struct dentry *create_tree(struct
> >   		goto err;
> >   	}
> >> -	dir = debugfs_create_dir(name, trace->root->root);
> > -	if (IS_ERR(dir))
> > +	temp->trace_root = trace->root->root;
> > +	temp->trace_dir = debugfs_create_dir(name, trace->root->root);
> > +
> > +	if (IS_ERR(temp->trace_dir))
> >   		remove_root(trace);
> > -	else
> > +	else {
> >   		trace->root->users++;
> > +		temp->ti = trace;
> > +		list_add_tail(&temp->trace_dir_list, &trace_dirs);
> > +	}
> >>   err:
> >   	mutex_unlock(&trace_mutex);
> > -	return dir;
> > +	return temp->trace_dir;
> >   }
> >> +int trace_exists(const char *parent_dir, const char *dir,
> > +					struct trace_info **ti)
> > +{
> > +	struct list_head *pos;
> > +	struct trace_root *r;
> > +	struct trace_dir *temp;
> > +
> > +	list_for_each(pos, &trace_roots) {
> > +		r = list_entry(pos, struct trace_root, list);
> list_for_each_entry()?
>> +		if (!strcmp(parent_dir, r->name))
> > +			goto search_dir;
> > +	}
> > +	return TRACE_PARENT_DIR_ABSENT;
> > +
> > + search_dir:
> > +	list_for_each(pos, &trace_dirs) {
> > +		temp = list_entry(pos, struct trace_dir, trace_dir_list);
> > +
> > +		if (!strcmp(dir, temp->trace_dir_name)) {
> > +			*ti = temp->ti;
> > +			return TRACE_DIR_EXISTS;
> > +		}
> > +	}
> > +	return TRACE_PARENT_DIR_EXISTS;
> > +}
> > +EXPORT_SYMBOL_GPL(trace_exists);
> I wonder if the whole "trace_*" namespace was a good choice.  There are
> other trace patches in-kernel, out-of-kernel and presumably in our
> future.

The name 'trace' (previously GTSC), I gather that it was the chosen after
much deliberation (http://tinyurl.com/6odoh4), however I'm open to the
idea of changing the name (say dbg_printk/dbg_dump?).

Kindly let me know of your suggestions for this, and I will change them
during the next version.

>>   static int dropped_open(struct inode *inode, struct file *filp)
> >   {
> >   	filp->private_data = inode->i_private;
> > @@ -561,3 +610,149 @@ void trace_cleanup(struct trace_info *tr
> >   	kfree(trace);
> >   }
> >   EXPORT_SYMBOL_GPL(trace_cleanup);
> > +
> > +/**
> > + *	trace_cleanup_all - Removes all trace directories under a parent_dir
> > + *	@parent_dir: Name of the parent directory
> > + */
> > +void trace_cleanup_all(const char *parent_dir)
> > +{
> > +	struct list_head *pos, *pos_temp;
> > +	struct trace_dir *temp;
> > +
> > +	list_for_each_safe(pos, pos_temp, &trace_dirs) {
> > +		temp = list_entry(pos, struct trace_dir, trace_dir_list);
> list_for_each_entry_safe()?
>> +		if (!strncmp(parent_dir, temp->trace_root->d_iname, \
> > +							strlen(parent_dir)))
> > +			trace_cleanup(temp->ti);
> > +	}
> > +}
> > +EXPORT_SYMBOL_GPL(trace_cleanup_all);
> >
> > ...
> >
> > +static inline int init_trace_interface(struct trace_printk_data *tpk)
> > +{
> > +	int ret = 0;
> > +	tpk->exists = trace_exists(tpk->parent_dir, tpk->dir, &tpk->ti);
> > +
> > +	switch (tpk->exists) {
> > +
> > +	case TRACE_PARENT_DIR_EXISTS:
> > +	case TRACE_PARENT_DIR_ABSENT:
> > +		if (!tpk->buf_size)
> > +			tpk->buf_size = DEFAULT_TRACE_BUF_SIZE;
> > +		if (!tpk->sub_buf_size)
> > +			tpk->sub_buf_size = DEFAULT_TRACE_SUB_BUF_NR;
> > +		tpk->ti = trace_setup(tpk->parent_dir, tpk->dir,
> > +				tpk->buf_size, tpk->sub_buf_size, tpk->flags);
> > +		printk(KERN_INFO "Trace interface %s setup\n",
> > +							tpk->ti->dir->d_iname);
> > +		if (IS_ERR(tpk->ti)) {
> > +			printk(KERN_ERR "Error initialising %s interface\n",
> > +							tpk->ti->dir->d_iname);
> > +			return -EPERM;
> > +		}
> > +	/* Fall through */
> > +	case TRACE_DIR_EXISTS:
> > +		if (tpk->ti->state == TRACE_SETUP)
> > +			ret = trace_start(tpk->ti);
> > +		else
> > +			ret = -EPERM;
> > +	}
> > +
> > +	return 0;
> > +}
> Two callsites, far too large to be inlined.
> Please just don't use inline at all, except in exceptional
> circumstances.  The compiler will work it out for normal usage.
That 'inline' was unintentional. I will remove it.

>> +/**
> > + * trace_printk - Output a string to debugfs mount 'directly' using > 'trace'
> > + * @tpk: Structure containing info such as parent_dir and directory
> > + * @format: String containing format string specifiers
> > + * @ap: List of arguments
> > + */
> > +int trace_printk(struct trace_printk_data *tpk, char *format, ...)
> > +{
> > +	int ret = 0;
> > +	va_list(ap);
> > +	unsigned long flags = 0;
> So I look at this and wonder "why did that need initialising?"
>> +	va_start(ap, format);
> > +
> > +	ret = init_trace_interface(tpk);
> > +	if (unlikely(ret))
> > +		return ret;
> > +
> > +	/* Now do the actual printing */
> > +	/* Take an RCU Lock over the trace_info state */
> > +	rcu_read_lock();
> > +	/* Take a spinlock for the global buffer used by relay */
> > +	if (tpk->flags & TRACE_GLOBAL_CHANNEL)
> > +		spin_lock_irqsave(&tpk->ti->trace_lock, flags);
> ah, because you got a compiler warning.
> That's why we have "uninitialized_var()": so the reader can understand
> what is happening.  Plus uninitialized_var() generates no code, whereas
> that assignment generates additional text.
> uninitialized_var() is fairly sucky, but adding a mysterious,
> seemingly-unneeded and code-generating "= 0" is suckier.
Ok. Will modify it:
	uninitialized_var(flags);
>> +	ret = trace_printf(tpk->ti, format, ap);
> > +	if (tpk->flags & TRACE_GLOBAL_CHANNEL)
> > +		spin_unlock_irqrestore(&tpk->ti->trace_lock, flags);
> > +	rcu_read_unlock();
> > +
> > +	va_end(ap);
> > +	return ret;
> > +}
> > +EXPORT_SYMBOL(trace_printk);
> > +
> > +/**
> > + * trace_dump - Output binary into debugfs mount 'directly' using 'trace'
> > + * @tpk: Structure containing info such as parent_dir and directory
> > + * @output: Data that needs to be output
> > + * @output_len: Length of the output data
> > + */
> > +int trace_dump(struct trace_printk_data *tpk, const void *output,
> > +							const int output_len)
> > +{
> > +	char *record;
> > +	unsigned long flags = 0;
> there too.
>> +	int ret = 0;
> > +
> > +	ret = init_trace_interface(tpk);
> > +	if (unlikely(ret))
> > +		return ret;
> > +
> > +	/* Now do the actual printing */
> > +	rcu_read_lock();
> > +	/* Take a spinlock for the global buffer used by relay */
> > +	if (tpk->flags & TRACE_GLOBAL_CHANNEL)
> > +		spin_lock_irqsave(&tpk->ti->trace_lock, flags);
> Making a callee's locking behaviour dependent upon an incoming argument
> is considered poor style and earns nastygrams from Linus.  Is there no
> sane alternative?
Will something like this look better?

static int trace_dump_per_cpu_channel(struct trace_printk_data *tpk,
				const void *output, const int output_len)
{
	char *record;
	int ret = 0;
	
	record = relay_reserve(tpk->ti->rchan, output_len);

	if (record && trace_running(tpk->ti))
		memcpy(record, output, output_len);
	else {
		if (record)
			ret = -EPERM;
		else
			ret = -ENOMEM;
	}
	return ret;
}

static int trace_dump_global_channel(struct trace_printk_data *tpk,
				const void *output, const int output_len)
{
	int ret;
	uninitialized_var(flags);
	
	spin_lock_irqsave(&tpk->ti->trace_lock, flags);
	ret = trace_dump_per_cpu_channel(tpk, output, output_len);
	spin_unlock_irqrestore(&tpk->ti->trace_lock, flags);

	return ret;
}

int trace_dump(struct trace_printk_data *tpk, const void *output,
							const int output_len)
{
	unsigned long flags = 0;
	int ret = 0;

	ret = init_trace_interface(tpk);
	if (unlikely(ret))
		return ret;

	/* Now do the actual printing */
	rcu_read_lock();
	/* Take a spinlock for the global buffer used by relay */
	if (tpk->flags & TRACE_GLOBAL_CHANNEL)
		ret = trace_dump_global_channel(tpk, output, output_len);
	else
		ret = trace_dump_per_cpu_channel(tpk, output, output_len);
	
	rcu_read_unlock();

	return ret;
}

>> +	record = relay_reserve(tpk->ti->rchan, output_len);
> > +
> > +	if (record && trace_running(tpk->ti))
> > +		memcpy(record, output, output_len);
> > +	else {
> > +		if (record)
> > +			ret = -EPERM;
> > +		else
> > +			ret = -ENOMEM;
> > +	}
> > +	if (tpk->flags & TRACE_GLOBAL_CHANNEL)
> > +		spin_unlock_irqrestore(&tpk->ti->trace_lock, flags);
> > +	rcu_read_unlock();
> > +
> > +	return ret;
> > +}
> > +EXPORT_SYMBOL(trace_dump);
> >
> > ...
> >
> > +int init_module(void)
> > +{
> > +	int ret = 0;
> > +	int len_parent_dir, len_dir;
> > +
> > +	/* setup the kprobe */
> > +	kp.pre_handler = handler_pre;
> > +	kp.post_handler = NULL;
> > +	kp.fault_handler = NULL;
> > +	kp.symbol_name = PROBE_POINT;
> > +	ret = register_kprobe(&kp);
> > +	if (ret) {
> > +		printk(KERN_ERR "fork_trace: register_kprobe failed\n");
> > +		return ret;
> > +	}
> > +
> > +	len_parent_dir = strlen(SAMPLE_PARENT_DIR) + 1;
> > +	/* Initialising len_dir to the larger of the two dir names */
> > +	len_dir = strlen("kprobe_struct") + 1;
> > +
> > +	tpk = kzalloc(sizeof(*tpk), GFP_KERNEL);
> > +	if (!tpk)
> > +		ret = 1;
> > +
> > +	tpk->parent_dir = SAMPLE_PARENT_DIR;
> > +
> > +	/* Let's do a binary dump of struct kprobe using trace_dump */
> > +	tpk->dir = "kprobes_struct";
> > +	tpk->flags = TRACE_GLOBAL_CHANNEL;
> > +	trace_dump(tpk, &kp, sizeof(kp));
> > +
> > +	/* Now change the directory to collect fork pid data */
> > +	tpk->dir = PROBE_POINT;
> > +
> > +	if (ret)
> > +		printk(KERN_ERR "Unable to find required free memory. "
> > +				"Trace new sample module loading aborted");
> > +	return ret;
> > +}
> > +
> > +void cleanup_module(void)
> > +{
> > +	unregister_kprobe(&kp);
> > +
> > +	/* Just a single cleanup call passing the parent dir string */
> > +	trace_cleanup_all(SAMPLE_PARENT_DIR);
> > +}
> > +MODULE_LICENSE("GPL");
> Do init_module() and cleanup_module() actually get called?  We _used_
> to have magic code which automatically calls those functions but I'd
> have thought it got removed years ago?
> Anyway, we should be using module_init() and module_exit() here, and
> those functions should have static scope.
Ok. This is adapted from legacy code in samples/trace/fork_trace.c.
I will modify the code in both files.

>> Index: linux-blktrace-many/samples/trace/Makefile
> > ===================================================================
> > --- linux-blktrace-many.orig/samples/trace/Makefile
> > +++ linux-blktrace-many/samples/trace/Makefile
> > @@ -1,4 +1,4 @@
> >   # builds the trace example kernel modules;
> >   # then to use (as root):  insmod <fork_trace.ko>
> >> -obj-$(CONFIG_SAMPLE_TRACE) := fork_trace.o
> > +obj-$(CONFIG_SAMPLE_TRACE) := fork_trace.o fork_new_trace.o
> > Index: linux-blktrace-many/Documentation/trace.txt
> > ===================================================================
> > --- linux-blktrace-many.orig/Documentation/trace.txt
> > +++ linux-blktrace-many/Documentation/trace.txt
> > @@ -150,6 +150,28 @@ The steps a kernel data provider takes t
> >   5) Destroy the trace channel and underlying relay channel -
> >      trace_cleanup().
> >> +Alternatively the user may choose to make use of two new interfaces --
> > +trace_printk() and trace_dump() -- to setup trace interface and
> > +trace_cleanup_all() to tear-down the same.
> > +
> > +Steps to use:
> > +1) Create and populate an instance of trace_printk_data structure. The > fields
> > +   parent_dir and dir are mandatory. The fields buf_size, sub_buf_size > and flags
> > +   are optional and will take default values if not populated. The field
> > +   'exists' and ti are for the trace infrastructure to use. The pointer > to the
> > +   'struct trace_info' i.e. ti may be used to perform fine granular > operations
> > +   such as determine the state of the 'trace', stop individual traces, > etc.
> > +2) Default values for buf_size and sub_buf_size are 4096, 40 respectively.
> > +3) Use trace_dump() to output binary data which may be acted upon by a
> > +   high-level program (say dumping a structure). trace_printk() can be > used
> > +   for string output. Pass a pointer to the instance of trace_printk_data
> > +   structure to these functions along with other parameters. The output > from
> > +   these functions can be found at
> > +   <debugfs_mount>/<parent_dir>/<dir>/trace<0..n>.
> > +4) trace_cleanup_all() for a given parent directory will cleanup and > remove all
> > +   trace directories created under the specified directory.
> > +5) Sample code for the same can be found in samples/trace/fork_new_trace.c
> wordwrapping...
> I often fix it, but I think we need v3 on this one.

Thanks for looking into the patches. I will send out the next version of
the patch after hearing from you on the above.

Thanks,
K.Prasad
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists