block: add missing block_bio_complete() tracepoint
Subsystems (4)
PATCH 1/1 block: add missing block_bio_complete() tracepoint
The block_bio_complete() TP has been missed so long, so that bio-based drivers haven't been able to trace its IO behavior. Add it.
In some rare cases, such as loop_switch, @bio->bi_bdev can be NULL. Thus convert it to TRACE_EVENT_CONDITION() as Steven suggested.
From now on, request-based drivers will also get BLK_TA_COMPLETEs for all bio's in requests. This needs to be handled in userland properly.
Also remove external use of the TP in DM and unexport it.
---
* v2:
- Merge previous patches into one as suggested by Tejun.
- Drop BIO_COMPLETE_MASK. Now I think it should be handled in userland like other (maybe duplicated, in some sense) bio complete reports.
block/blk-core.c | 1 - drivers/md/dm.c | 1 - fs/bio.c | 2 ++ include/trace/events/block.h | 4 +++- 4 files changed, 5 insertions(+), 3 deletions(-)
block/blk-core.c +0 -1
@@ -37,7 +37,6 @@EXPORT_TRACEPOINT_SYMBOL_GPL (block_bio_remap);EXPORT_TRACEPOINT_SYMBOL_GPL (block_rq_remap); -EXPORT_TRACEPOINT_SYMBOL_GPL (block_bio_complete);DEFINE_IDA (blk_queue_ida);
drivers/md/dm.c +0 -1
@@ -648,7 +648,6 @@ static void dec_pending(struct dm_io *io, int error)queue_io (md, bio); }else {/* done with normal IO or empty flush */ -trace_block_bio_complete (md->queue, bio, io_error);bio_endio (bio, io_error); } }
fs/bio.c +2 -0
@@ -1447,6 +1447,8 @@ void bio_endio(struct bio *bio, int error)else if (!test_bit (BIO_UPTODATE , &bio->bi_flags)) error = -EIO ; +trace_block_bio_complete (bdev_get_queue (bio->bi_bdev), bio, error); +if (bio->bi_end_io) bio->bi_end_io (bio, error); }
include/trace/events/block.h +3 -1
@@ -213,12 +213,14 @@ TRACE_EVENT(block_bio_bounce,* This tracepoint indicates there is no further work to do on this * block IO operation @bio. */ -TRACE_EVENT(block_bio_complete, +TRACE_EVENT_CONDITION(block_bio_complete, TP_PROTO(struct request_queue *q, struct bio *bio, int error), TP_ARGS(q, bio, error), + TP_CONDITION(bio->bi_bdev != NULL), + TP_STRUCT__entry( __field( dev_t, dev ) __field( sector_t, sector )
Signature · 1 line
1.7.8.2
Quoted 9 lines · Namhyung Kim
The
block_bio_complete()TP has been missed so long, so that bio-based drivers haven't been able to trace its IO behavior. Add it.In some rare cases, such as
loop_switch, @bio->bi_bdev can be NULL. Thus convert it toTRACE_EVENT_CONDITION()as Steven suggested.From now on, request-based drivers will also get
BLK_TA_COMPLETEsfor all bio's in requests. This needs to be handled in userland properly.
Quoted 6 lines
Also remove external use of the TP in DM and unexport it.
Signed-off-by Namhyung Kim <[email protected]>Cc Tejun Heo <[email protected]>Cc Steven Rostedt <[email protected]>
I like the smiplicity change but do we know how we can filter this out from userland? Also, what's the reason not to do it from blktrace.c? What would be the downside of doing that?
Thanks.
Signature · 1 line
tejun
Hello,
Quoted 17 lines · Tejun Heo
Quoted 9 lines · Namhyung Kim
The
block_bio_complete()TP has been missed so long, so that bio-based drivers haven't been able to trace its IO behavior. Add it.In some rare cases, such as
loop_switch, @bio->bi_bdev can be NULL. Thus convert it toTRACE_EVENT_CONDITION()as Steven suggested.From now on, request-based drivers will also get
BLK_TA_COMPLETEsfor
all bio's in requests. This needs to be handled in userland properly.Quoted 6 lines
Also remove external use of the TP in DM and unexport it.
Signed-off-by Namhyung Kim <[email protected]>Cc Tejun Heo <[email protected]>Cc Steven Rostedt <[email protected]>
Quoted 6 lines
I like the smiplicity change but do we know how we can filter this out from userland? Also, what's the reason not to do it from
blktrace.c? What would be the downside of doing that?Thanks.
The userland tool cannot distinguish bounced bio from original one at completion TP, but it can expect there will be a duplicated BLK_TA_COMPLETE as it sees BLK_TA_BOUNCE for the bio before.
Filtering it out from kernel side seems to hide a real information that (paranoid?) user might want to get, and it looks like providing "polcy not mechanism" IMHO. That's why I changed my mind finally.
I cannot think of the downside, anyway it's not a big deal, if you think it's wrong choice, I'm OK to change it again.
Thanks,
Namhyung
Quoted 10 lines · Namhyung Kim
The userland tool cannot distinguish bounced bio from original one at completion TP, but it can expect there will be a duplicated
BLK_TA_COMPLETEas it seesBLK_TA_BOUNCEfor the bio before.Filtering it out from kernel side seems to hide a real information that (paranoid?) user might want to get, and it looks like providing "polcy not mechanism" IMHO. That's why I changed my mind finally.
I cannot think of the downside, anyway it's not a big deal, if you think it's wrong choice, I'm OK to change it again.
It's just that this patch as it stands will break the existing tools and is likely to cause some amount of head scratching for blktrace users upgrading to new kernel with existing userland, so yeah, I think it should be filtered in kernel from blktrace.c.
Thanks.
Signature · 1 line
tejun
Quoted 12 lines · Tejun Heo
Quoted 10 lines · Namhyung Kim
The userland tool cannot distinguish bounced bio from original one at completion TP, but it can expect there will be a duplicated
BLK_TA_COMPLETEas it seesBLK_TA_BOUNCEfor the bio before.Filtering it out from kernel side seems to hide a real information that (paranoid?) user might want to get, and it looks like providing "polcy not mechanism" IMHO. That's why I changed my mind finally.
I cannot think of the downside, anyway it's not a big deal, if you think it's wrong choice, I'm OK to change it again.
Quoted 7 lines
It's just that this patch as it stands will break the existing tools and is likely to cause some amount of head scratching for blktrace users upgrading to new kernel with existing userland, so yeah, I think it should be filtered in kernel from
blktrace.c.Thanks.
Will it break blktrace? Looking at the code, not tested though, it will just add one more 'C' line for bounced bio, and that's it? The blktrace will get way more 'C' lines for normal request based devices and it needs to be handled anyway. Am I missing something?
Thanks,
Namhyung
Quoted 9 lines · Namhyung Kim
Quoted 4 lines
It's just that this patch as it stands will break the existing tools and is likely to cause some amount of head scratching for blktrace users upgrading to new kernel with existing userland, so yeah, I think it should be filtered in kernel from
blktrace.c.Will it break blktrace? Looking at the code, not tested though, it will just add one more 'C' line for bounced bio, and that's it? The blktrace will get way more 'C' lines for normal request based devices and it needs to be handled anyway. Am I missing something?
Hmmm... maybe I'm confused. Wouldn't it cause duplicate completions for all rq based drivers too?
Signature · 1 line
tejun
Quoted 6 lines · Tejun Heo
Quoted 5 lines · Namhyung Kim
Quoted 4 lines
It's just that this patch as it stands will break the existing tools and is likely to cause some amount of head scratching for blktrace users upgrading to new kernel with existing userland, so yeah, I think it should be filtered in kernel from
blktrace.c.
Quoted 8 lines
Quoted 4 lines
Will it break blktrace? Looking at the code, not tested though, it will just add one more 'C' line for bounced bio, and that's it? The blktrace will get way more 'C' lines for normal request based devices and it needs to be handled anyway. Am I missing something?
Hmmm... maybe I'm confused. Wouldn't it cause duplicate completions for all rq based drivers too?
Yes, blktrace will see both of completions for rq and bio's it contains from now on, and I thought it's your intention? For the bounced bio's, it'll see one more completion per bio.
The users of older blktrace might get confused by this, thus changing the tool is inevitable regardless of the bounced bio's.
Thanks,
Namhyung
Hello,
Quoted 3 lines · Namhyung Kim
Yes, blktrace will see both of completions for rq and bio's it contains from now on, and I thought it's your intention? For the bounced bio's, it'll see one more completion per bio.
Umm... no, blktrace users shouldn't see confusing outputs. My point was that we should leave raw tracepoints as raw monitor points so that they can provide raw data to different users and filter from blktrace.c so that the existing users aren't disturbed.
Thanks.
Signature · 1 line
tejun
Hi,
Quoted 12 lines · Tejun Heo
Hello,
Quoted 3 lines · Namhyung Kim
Yes, blktrace will see both of completions for rq and bio's it contains from now on, and I thought it's your intention? For the bounced bio's, it'll see one more completion per bio.
Umm... no, blktrace users shouldn't see confusing outputs. My point was that we should leave raw tracepoints as raw monitor points so that they can provide raw data to different users and filter from
blktrace.cso that the existing users aren't disturbed.
That means blktrace should not be changed, and not see duplicated completions for rq based drivers at all?
Thanks,
Namhyung
Quoted 10 lines · Namhyung Kim
Hi,
Quoted 7 lines · Tejun Heo
Hello,
Quoted 3 lines · Namhyung Kim
Yes, blktrace will see both of completions for rq and bio's it contains from now on, and I thought it's your intention? For the bounced bio's, it'll see one more completion per bio.
Quoted 8 lines
Quoted 5 lines
Umm... no, blktrace users shouldn't see confusing outputs. My point was that we should leave raw tracepoints as raw monitor points so that they can provide raw data to different users and filter from
blktrace.cso that the existing users aren't disturbed.That means blktrace should not be changed, and not see duplicated completions for rq based drivers at all?
Yeap
Signature · 1 line
tejun
Quoted 17 lines · Tejun Heo
Quoted 16 lines · Namhyung Kim
Hi,
Quoted 12 lines · Tejun Heo
Hello,
Quoted 3 lines · Namhyung Kim
Yes, blktrace will see both of completions for rq and bio's it contains from now on, and I thought it's your intention? For the bounced bio's, it'll see one more completion per bio.
Umm... no, blktrace users shouldn't see confusing outputs. My point was that we should leave raw tracepoints as raw monitor points so that they can provide raw data to different users and filter from
blktrace.cso that the existing users aren't disturbed.
Quoted 5 lines
Quoted 2 lines
That means blktrace should not be changed, and not see duplicated completions for rq based drivers at all?
Yeap
I see. I'll prepare the patch soon.
Thanks,
Namhyung
Quoted 6 lines · Namhyung Kim
The
block_bio_complete()TP has been missed so long, so that bio-based drivers haven't been able to trace its IO behavior. Add it.In some rare cases, such as
loop_switch, @bio->bi_bdev can be NULL. Thus convert it toTRACE_EVENT_CONDITION()as Steven suggested.
Now I see that it seems TRACE_EVENT_CONDITION() can protect event tracing from such condition, but what about other users of the TP like blktrace? I think it'll still get NULL pointer dereference on bdev_get_queue() after the change, right? If so, convert to T_E_C() looks meaningless IMHO. Do I miss something?
Thanks,
Namhyung
Quoted 78 lines
From now on, request-based drivers will also get
BLK_TA_COMPLETEsfor
all bio's in requests. This needs to be handled in userland properly.Also remove external use of the TP in DM and unexport it.
Signed-off-by Namhyung Kim <[email protected]>Cc Tejun Heo <[email protected]>Cc Steven Rostedt <[email protected]>---
* v2:
- Merge previous patches into one as suggested by Tejun.
- DropBIO_COMPLETE_MASK. Now I think it should be handled in userland like other (maybe duplicated, in some sense) bio complete reports.block/blk-core.c | 1 - drivers/md/dm.c | 1 - fs/bio.c | 2 ++ include/trace/events/block.h | 4 +++- 4 files changed, 5 insertions(+), 3 deletions(-)block/blk-core.c +0
-1@@ -37,7 +37,6 @@EXPORT_TRACEPOINT_SYMBOL_GPL (block_bio_remap);EXPORT_TRACEPOINT_SYMBOL_GPL (block_rq_remap); -EXPORT_TRACEPOINT_SYMBOL_GPL (block_bio_complete);DEFINE_IDA (blk_queue_ida);drivers/md/dm.c +0
-1@@ -648,7 +648,6 @@ static void dec_pending(struct dm_io *io, int error)queue_io (md, bio); }else {/* done with normal IO or empty flush */ -trace_block_bio_complete (md->queue, bio, io_error);bio_endio (bio, io_error); } }fs/bio.c +2
-0@@ -1447,6 +1447,8 @@ void bio_endio(struct bio *bio, int error)else if (!test_bit (BIO_UPTODATE ,&bio->bi_flags)) error = -EIO ; +trace_block_bio_complete (bdev_get_queue (bio->bi_bdev), bio, error); +if (bio->bi_end_io) bio->bi_end_io (bio, error); }include/trace/events/block.h +3
-1@@ -213,12 +213,14 @@ TRACE_EVENT(block_bio_bounce,* This tracepoint indicates there is no further work to do on this * block IO operation @bio. */ -TRACE_EVENT(block_bio_complete, +TRACE_EVENT_CONDITION(block_bio_complete, TP_PROTO(struct request_queue *q, struct bio *bio, int error), TP_ARGS(q, bio, error), + TP_CONDITION(bio->bi_bdev != NULL), + TP_STRUCT__entry( __field( dev_t, dev ) __field( sector_t, sector )
Quoted 8 lines · Namhyung Kim
Quoted 6 lines · Namhyung Kim
The
block_bio_complete()TP has been missed so long, so that bio-based drivers haven't been able to trace its IO behavior. Add it.In some rare cases, such as
loop_switch, @bio->bi_bdev can be NULL. Thus convert it toTRACE_EVENT_CONDITION()as Steven suggested.
Quoted 5 lines
Now I see that it seems
TRACE_EVENT_CONDITION()can protect event tracing from such condition, but what about other users of the TP like blktrace? I think it'll still get NULL pointer dereference onbdev_get_queue()after the change, right? If so, convert toT_E_C()looks meaningless IMHO. Do I miss something?
Not really following, but the whole point of using TRACE_EVENT_CONDITION() is avoiding the conditional jump when the TP is disabled. Whether the TP users need to test again / more isn't too important.
Thanks.
Signature · 1 line
tejun
Hi,
Quoted 15 lines · Tejun Heo
Quoted 8 lines · Namhyung Kim
Quoted 6 lines · Namhyung Kim
The
block_bio_complete()TP has been missed so long, so that bio-based drivers haven't been able to trace its IO behavior. Add it.In some rare cases, such as
loop_switch, @bio->bi_bdev can be NULL. Thus convert it toTRACE_EVENT_CONDITION()as Steven suggested.Quoted 5 lines
Now I see that it seems
TRACE_EVENT_CONDITION()can protect event tracing from such condition, but what about other users of the TP like blktrace? I think it'll still get NULL pointer dereference onbdev_get_queue()after the change, right? If so, convert toT_E_C()looks meaningless IMHO. Do I miss something?
Quoted 7 lines
Not really following, but the whole point of using
TRACE_EVENT_CONDITION()is avoiding the conditional jump when the TP is disabled. Whether the TP users need to test again / more isn't too important.Thanks.
Right, but the point is it could make a NULL pointer dereference during evaluation of the argument of the TP AFAICS. I'm not sure about the TP implementation though, I think I was wrong - T_E_C() cannot protect us from it because it happens just before jumping to the TP, right?
So I think we need a conditional jump (with the "likely" annotation) for this even when the TP is disabled.
Thanks,
Namhyung
Hello,
Quoted 7 lines · Namhyung Kim
Right, but the point is it could make a NULL pointer dereference during evaluation of the argument of the TP AFAICS. I'm not sure about the TP implementation though, I think I was wrong -
T_E_C()cannot protect us from it because it happens just before jumping to the TP, right?So I think we need a conditional jump (with the "likely" annotation) for this even when the TP is disabled.
Hmmm... still not following. Where the said NULL dereference happen? TEC conditional is equivalent to "if (COND) TP;". If you don't use TEC, it'll be "if (COND) if (TP enabled) TP;". With TEC, it will be "if (TP enabled) if (COND) TP;". There's no other difference.
Thanks.
Signature · 1 line
tejun
Hi,
Quoted 11 lines · Tejun Heo
Hello,
Quoted 7 lines · Namhyung Kim
Right, but the point is it could make a NULL pointer dereference during evaluation of the argument of the TP AFAICS. I'm not sure about the TP implementation though, I think I was wrong -
T_E_C()cannot protect us from it because it happens just before jumping to the TP, right?So I think we need a conditional jump (with the "likely" annotation) for this even when the TP is disabled.
Quoted 7 lines
Hmmm... still not following. Where the said NULL dereference happen? TEC conditional is equivalent to "if (COND) TP;". If you don't use TEC, it'll be "if (COND) if (TP enabled) TP;". With TEC, it will be "if (TP enabled) if (COND) TP;". There's no other difference.
Thanks.
I've made a quick investigation on TP implementation, and finally figured out what I was wrong - I thought the COND would be checkd in a probe, but it's not. Thanks for pointing it out.
However, for some reason, it seems gcc generated code that evaluates the arguments - bdev_get_queue() in this case - before checking the COND. Simple test module below caused a NULL pointer dereference when I used TRACE_EVENT_CONDITION(), but not for conditional jump:
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/bio.h>
static int __init init_mod(void)
{
struct bio *bio = bio_alloc(GFP_KERNEL, 0);
bio_endio(bio, 0);
bio_put(bio);
return 0;
}static void __exit exit_mod(void)
{
}
module_init(init_mod);module_exit(exit_mod);
Thanks,
Namhyung