From: "Abhishek Sagar" <sagar.abhishek@gmail.com>
To: "Jim Keniston" <jkenisto@us.ibm.com>
Cc: "Srinivasa Ds" <srinivasa@in.ibm.com>,
linux-kernel@vger.kernel.org, prasanna@in.ibm.com,
davem@davemloft.net, anil.s.keshavamurthy@intel.com,
"Ananth N Mavinakayanahalli" <ananth@in.ibm.com>
Subject: Re: [PATCH][RFC] kprobes: Add user entry-handler in kretprobes
Date: Sat, 17 Nov 2007 00:23:57 +0530 [thread overview]
Message-ID: <863e9df20711161053q7f6ae3d5vb19b23d2ce6b20f7@mail.gmail.com> (raw)
In-Reply-To: <1195171644.3804.124.camel@dyn9047018096.beaverton.ibm.com>
On Nov 16, 2007 5:37 AM, Jim Keniston <jkenisto@us.ibm.com> wrote:
> On Thu, 2007-11-15 at 20:30 +0530, Abhishek Sagar wrote:
> > On Nov 15, 2007 4:21 AM, Jim Keniston <jkenisto@us.ibm.com> wrote:
> > > 2. Simplify the task of correlating data (e.g., timestamps) between
> > > function entry and function return.
> >
> > Would adding of data and len fields in ri help? Instead of "pouching"
> > data in one go at registration time, this would let user handlers do
> > the allocation
>
> Yes and no. Adding just a data field -- void*, or maybe unsigned long
> long so it's big enought to accommodate big timestamps -- would be a big
> improvement on your current proposal. That would save the user the
> drudgery of mapping the ri pointer to his/her per-instance data.
> There's plenty of precedent for passing "private_data" values to
> callbacks.
>
> I don't think a len field would help much. If such info were needed, it
> could be stored in the data structure pointed to by the data field.
>
> I still don't think "letting [i.e., requiring that] user handlers do the
> allocation" is a win. I'm still interested to see how this plays out in
> real examples.
>
> > and allow them to use different kinds of data
> > structures per-instance.
>
> I haven't been able to think of any scenarios where this would be
> useful. A "data pouch" could always contain a union, FWIW.
I'm inlining a sample module which uses a data pointer in ri. I didn't
go for a timestamp because it's not reliable. Some platforms might
simply not have any h/w timestamp counters. For the same reason a lot
of platforms (on ARM, say) have their sched_clock() mapped on jiffies.
This may prevent timestamps from being distinct across function entry
and exit. Plus a data pointer looks pretty harmless.
--- test module ---
#include <linux/kernel.h>
#include <linux/version.h>
#include <linux/module.h>
#include <linux/kprobes.h>
#include <linux/ktime.h>
#define PRINT_DELAY (10 * HZ)
/* This module calculates the total time and instances of func being called
* across all cpu's. An average is calculated every 10 seconds and displayed.
* Only one function instance per-task is monitored. This cuts out the
* possibility of measuring time for recursive and nested function
* invocations.
*
* Note: If compiling as a standalone module, make sure sched_clock() is
* exported in the kernel. */
/* per-task data */
struct prof_data {
struct task_struct *task;
struct list_head list;
unsigned long long entry_stamp;
};
static const char *func = "sys_open";
static spinlock_t time_lock;
static ktime_t total_time;
static unsigned long hits;
static LIST_HEAD(data_nodes); /* list of per-task data */
static struct delayed_work work_print;
static struct prof_data *get_per_task_data(struct task_struct *tsk)
{
struct prof_data *p;
/* lookup prof_data corresponding to tsk */
list_for_each_entry(p, &data_nodes, list) {
if (p->task == tsk)
return p;
}
return NULL;
}
/* called with kretprobe_lock held */
static int entry_handler(struct kretprobe_instance *ri, struct pt_regs *regs)
{
struct prof_data *stats;
stats = get_per_task_data(current);
if (stats)
return 1; /* recursive/nested call */
stats = kmalloc(sizeof(struct prof_data), GFP_ATOMIC);
if (!stats)
return 1;
stats->entry_stamp = sched_clock();
stats->task = current;
INIT_LIST_HEAD(&stats->list);
list_add(&stats->list, &data_nodes);
ri->data = stats;
return 0;
}
/* called with kretprobe_lock held */
static int return_handler(struct kretprobe_instance *ri, struct pt_regs *regs)
{
unsigned long flags;
struct prof_data *stats = (struct prof_data *)ri->data;
u64 elapsed;
BUG_ON(ri->data == NULL);
elapsed = (long long)sched_clock() - (long long)stats->entry_stamp;
/* update stats */
spin_lock_irqsave(&time_lock, flags);
++hits;
total_time = ktime_add_ns(total_time, elapsed);
spin_unlock_irqrestore(&time_lock, flags);
list_del(&stats->list);
kfree(stats);
return 0;
}
static struct kretprobe my_kretprobe = {
.handler = return_handler,
.entry_handler = entry_handler,
};
/* called after every PRINT_DELAY seconds */
static void print_time(struct work_struct *work)
{
unsigned long flags;
s64 time_ns;
struct timespec ts;
BUG_ON(work != &work_print.work);
spin_lock_irqsave(&time_lock, flags);
time_ns = ktime_to_ns(total_time);
do_div(time_ns, hits);
spin_unlock_irqrestore(&time_lock, flags);
ts = ns_to_timespec(time_ns);
printk(KERN_DEBUG "Avg. running time of %s = %ld sec, %ld nsec\n",
func, ts.tv_sec, ts.tv_nsec);
schedule_delayed_work(&work_print, PRINT_DELAY);
}
static int __init test_module_init(void)
{
int ret;
my_kretprobe.kp.symbol_name = (char *)func;
spin_lock_init(&time_lock);
if ((ret = register_kretprobe(&my_kretprobe)) < 0) {
printk("Failed to register test kretprobe!\n");
return -1;
}
printk("Kretprobe active on %s\n", my_kretprobe.kp.symbol_name);
INIT_DELAYED_WORK(&work_print, print_time);
schedule_delayed_work(&work_print, PRINT_DELAY);
return 0;
}
static void __exit test_module_exit(void)
{
unregister_kretprobe(&my_kretprobe);
printk("kretprobe unregistered\n");
printk("Missed probing %d instances of %s\n",
my_kretprobe.nmissed, func);
}
module_init(test_module_init)
module_exit(test_module_exit)
MODULE_LICENSE("GPL");
--
Abhishek Sagar
next prev parent reply other threads:[~2007-11-16 18:54 UTC|newest]
Thread overview: 21+ messages / expand[flat|nested] mbox.gz Atom feed top
[not found] <47389BEB.1000901@gmail.com>
2007-11-12 18:39 ` Abhishek Sagar
2007-11-13 10:47 ` Abhishek Sagar
2007-11-14 7:57 ` Srinivasa Ds
2007-11-14 8:49 ` Abhishek Sagar
2007-11-14 10:23 ` Srinivasa Ds
2007-11-14 13:30 ` Abhishek Sagar
2007-11-14 22:51 ` Jim Keniston
2007-11-15 13:16 ` Abhishek Sagar
2007-11-15 21:16 ` Jim Keniston
2007-11-16 17:50 ` Abhishek Sagar
2007-11-17 0:54 ` Jim Keniston
2007-11-17 18:15 ` Abhishek Sagar
2007-11-19 12:26 ` Abhishek Sagar
2007-11-21 5:55 ` Jim Keniston
2007-11-21 10:20 ` Abhishek Sagar
2007-11-27 0:54 ` Jim Keniston
2007-11-15 15:00 ` Abhishek Sagar
2007-11-16 0:07 ` Jim Keniston
2007-11-16 18:53 ` Abhishek Sagar [this message]
2007-11-16 23:09 ` Jim Keniston
2007-11-17 17:09 ` Abhishek Sagar
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=863e9df20711161053q7f6ae3d5vb19b23d2ce6b20f7@mail.gmail.com \
--to=sagar.abhishek@gmail.com \
--cc=ananth@in.ibm.com \
--cc=anil.s.keshavamurthy@intel.com \
--cc=davem@davemloft.net \
--cc=jkenisto@us.ibm.com \
--cc=linux-kernel@vger.kernel.org \
--cc=prasanna@in.ibm.com \
--cc=srinivasa@in.ibm.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®