Skip to contentSkip to conversation

Trapping Block I/O

DISCUSSION Block Device 3 participants 9 messages Sep 23, 2005 – Sep 27, 2005

lore
Share & tools
lore
#0 lore raw

Hi,

I need to trap _all_ the I/O going to each and every block device
in the system. I used jprobes to trap calls to generic_make_request. Is this the correct/only place to do such a thing ?
Or do I have to monitor the q->make_request_fn for every device ?

Thanks & Regards
-BD.

#1 lore raw
Quoted 6 lines · Block Device

I need to trap _all_ the I/O going to each and every block device
in the system. I used jprobes to trap calls to generic_make_request. Is this the correct/only place to do such a thing ?
Or do I have to monitor the q->make_request_fn for every device ?

Yes, generic_make_request or monitoring q->make_request_fn can trap the _all_ I/O tpo block devices but other approach might be a little bit odd/difficult but through that you can get every request to block device .... the approach is you create a block device and then create that block device as a wrapper on your device, now use your block device and in its request function (can alter the data and sectors etc) and calls generic_make_request for the original device on which you created wrapper .... So by doing this you can easily monitor requests (similar to this approach is used in LVM/RAID) ......

--
Fawad Lateef

#2 lore raw
Quoted 2 lines · Fawad Lateef

you created wrapper .... So by doing this you can easily monitor requests (similar to this approach is used in LVM/RAID) ......

Or just use btrace, pull it from:

git://brick.kernel.dk/data/git/blktrace.git

Signature · 1 line
Jens Axboe
#3 lore raw
Quoted 8 lines · Jens Axboe
Quoted 2 lines · Fawad Lateef

you created wrapper .... So by doing this you can easily monitor requests (similar to this approach is used in LVM/RAID) ......

Or just use btrace, pull it from:

git://brick.kernel.dk/data/git/blktrace.git

Thnx for telling about btrace .... I havn't tried/looked at it before !!!!

--
Fawad Lateef

#4 lore raw
Quoted 11 lines · Fawad Lateef
Quoted 8 lines · Jens Axboe
Quoted 2 lines · Fawad Lateef

you created wrapper .... So by doing this you can easily monitor requests (similar to this approach is used in LVM/RAID) ......

Or just use btrace, pull it from:

git://brick.kernel.dk/data/git/blktrace.git

Thnx for telling about btrace .... I havn't tried/looked at it before !!!!

Well it's pretty new, so no wonder. But it should do everything you want and lots more. There's a list for it here:

[email protected]

I'm a little pressed for time these days, but I'll do a proper announce / demo of all the features starting next week since it's basically feature complete now.

If you don't use git, there are also snapshots available on kernel.org, more precisely here:

kernel.org/pub/linux/kernel/people/axboe/blktrace/

but kernel.org is pretty slow these days, so pulling from the git repo above is greatly recommended.

Signature · 1 line
Jens Axboe
#5 lore raw
Quoted 9 lines · Jens Axboe

Well it's pretty new, so no wonder. But it should do everything you want and lots more. There's a list for it here:

[email protected]

I'm a little pressed for time these days, but I'll do a proper announce / demo of all the features starting next week since it's basically feature complete now.

Quoted 8 lines

If you don't use git, there are also snapshots available on kernel.org, more precisely here:

kernel.org/pub/linux/kernel/people/axboe/blktrace/

but kernel.org is pretty slow these days, so pulling from the git repo above is greatly recommended.

Ya, I looked at it and its looking very good tool to tracing block I/O layer, but this tracing requires recompilation of the kernel and have to use on kernel directly from kernel.org but its not a big deal, I hope it will get into the main kernel soon ....

By the way my approach about creating wrapper and getting the device requests without modification into the kernel and can be easily used on any block device ...... ;)

--
Fawad Lateef

#6 lore raw
Quoted 15 lines · Fawad Lateef
Quoted 14 lines · Jens Axboe

Well it's pretty new, so no wonder. But it should do everything you want and lots more. There's a list for it here:

[email protected]

I'm a little pressed for time these days, but I'll do a proper announce / demo of all the features starting next week since it's basically feature complete now.

If you don't use git, there are also snapshots available on kernel.org, more precisely here:

kernel.org/pub/linux/kernel/people/axboe/blktrace/

Quoted 8 lines
Quoted 3 lines

but kernel.org is pretty slow these days, so pulling from the git repo above is greatly recommended.

Ya, I looked at it and its looking very good tool to tracing block I/O layer, but this tracing requires recompilation of the kernel and have to use on kernel directly from kernel.org but its not a big deal, I hope it will get into the main kernel soon ....

That is true, I plan on submitting it for 2.6.15. The goal was to get relayfs pushed in first and that did happen for 2.6.14.

Quoted 3 lines

By the way my approach about creating wrapper and getting the device requests without modification into the kernel and can be easily used on any block device ...... ;)

There are certainly a lot of ways to get the data out to user space, by far the bulk of the code is in the monitoring application. blktrace should be pretty fast though, one of the goals was to make sure it would be very light weight on the kernel side (which it is) and very fast on getting the data out (also achieved, relayfs works very well). The xprobe approach does have certain advantages, the main one being that you can easily modify it.

Signature · 1 line
Jens Axboe
#7 lore raw
Quoted 8 lines · Jens Axboe
Quoted 4 lines

Ya, I looked at it and its looking very good tool to tracing block I/O layer, but this tracing requires recompilation of the kernel and have to use on kernel directly from kernel.org but its not a big deal, I hope it will get into the main kernel soon ....

That is true, I plan on submitting it for 2.6.15. The goal was to get relayfs pushed in first and that did happen for 2.6.14.

That will be nice ....

Quoted 9 lines

There are certainly a lot of ways to get the data out to user space, by far the bulk of the code is in the monitoring application. blktrace should be pretty fast though, one of the goals was to make sure it would be very light weight on the kernel side (which it is) and very fast on getting the data out (also achieved, relayfs works very well). The xprobe approach does have certain advantages, the main one being that you can easily modify it.

I will test/use it soon and will let you know if find any bug/problem :)

--
Fawad Lateef

#8 lore raw
In reply to Fawad Lateef

The problem to creating a wrapper in the manner sugegsted is that any filesystems mounted or other stuff which uses the original device must be stopped
and changed to start using the new pseudo device. I want something which can be as
non intrusive as possible. Am looking @ Jens btrace :) ..

Thanks &Regards
-BD

Quoted 22 lines · Fawad Lateef
Quoted 6 lines · Block Device

I need to trap _all_ the I/O going to each and every block device
in the system. I used jprobes to trap calls to generic_make_request. Is this the correct/only place to do such a thing ?
Or do I have to monitor the q->make_request_fn for every device ?

Yes, generic_make_request or monitoring q->make_request_fn can trap the _all_ I/O tpo block devices but other approach might be a little bit odd/difficult but through that you can get every request to block device .... the approach is you create a block device and then create that block device as a wrapper on your device, now use your block device and in its request function (can alter the data and sectors etc) and calls generic_make_request for the original device on which you created wrapper .... So by doing this you can easily monitor requests (similar to this approach is used in LVM/RAID) ......

--
Fawad Lateef