tracing/irq: use softirq_to_name instead of __print_symbolic
PATCH 1/1 tracing/irq: use softirq_to_name instead of __print_symbolic
| commit c2adae0970ca1db8adb92fb56ae3bcabd916e8bd | Author: Steven Rostedt <[email protected]> | Date: Wed May 20 19:56:19 2009 -0400 | | tracing: convert irq events to use __print_symbolic | | The recording of the names at trace time is inefficient. This patch | implements the softirq event recording to only record the vector | and then use the __print_symbolic interface to print out the names. | | [ Impact: faster recording of softirq events ]
It's great to boost recording of softirq events, but why not simply use softirq_to_name in TP_printk()?
The above commit has 2 flaws:
- we waste memory defining local static struct trace_print_flags array in each softirq TRACE_EVENT
- if someone adds/removes a softirq, he may not know show_softirq_name() needs to be updated
Another issue with the above commit, that the output of softirq events becomes:
X-1701 [000] 1595.220739: softirq_entry: softirq=1 action=TIMER_SOFTIRQ
Compared to the original output:
X-1701 [000] 1595.220739: softirq_entry: softirq=1 action=TIMER
[ Impact: reserve the original output format of softirq events ]
--- include/trace/events/irq.h | 17 ++--------------- 1 files changed, 2 insertions(+), 15 deletions(-)
include/trace/events/irq.h +2 -15
@@ -7,19 +7,6 @@#undef TRACE_SYSTEM#define TRACE_SYSTEM irq -#define softirq_name (sirq) { sirq, #sirq } -#define show_softirq_name (val) \ - __print_symbolic(val, \ - softirq_name(HI_SOFTIRQ), \ - softirq_name(TIMER_SOFTIRQ), \ - softirq_name(NET_TX_SOFTIRQ), \ - softirq_name(NET_RX_SOFTIRQ), \ - softirq_name(BLOCK_SOFTIRQ), \ - softirq_name(TASKLET_SOFTIRQ), \ - softirq_name(SCHED_SOFTIRQ), \ - softirq_name(HRTIMER_SOFTIRQ), \ - softirq_name(RCU_SOFTIRQ)) -/** * irq_handler_entry - called immediately before the irq action handler * @irq: irq number
@@ -107,7 +94,7 @@ TRACE_EVENT(softirq_entry, ),TP_printk ("softirq=%d action=%s" , __entry->vec, -show_softirq_name (__entry->vec)) + softirq_to_name[__entry->vec]) );/**
@@ -136,7 +123,7 @@ TRACE_EVENT(softirq_exit, ),TP_printk ("softirq=%d action=%s" , __entry->vec, -show_softirq_name (__entry->vec)) + softirq_to_name[__entry->vec]) );#endif /* _TRACE_IRQ_H */
Signature · 1 line
1.5.4.rc3
Quoted 2 lines · Li Zefan
It's great to boost recording of softirq events, but why not simply use
softirq_to_nameinTP_printk()?
Because userspace programs using the binary trace buffer have no chance to retrieve the values from softirq_to_name.
Quoted 4 lines
The above commit has 2 flaws:
- we waste memory defining local static struct
trace_print_flagsarray in each softirqTRACE_EVENT
That could be solved by declaring the array manually and just passing the address to __print_symbolic. Steve, would that work? (also for __print_flags)
Quoted 2 lines
- if someone adds/removes a softirq, he may not know
show_softirq_name()needs to be updated
Just make sure you have the translation defined next to the actual flags. This is what I do in the XFS tracer:
typedef enum xfs_alloctype
{
XFS_ALLOCTYPE_ANY_AG, /* allocate anywhere, use rotor */ XFS_ALLOCTYPE_FIRST_AG, /* ... start at ag 0 */ XFS_ALLOCTYPE_START_AG, /* anywhere, start in this a.g. */ XFS_ALLOCTYPE_THIS_AG, /* anywhere in this a.g. */
XFS_ALLOCTYPE_START_BNO, /* near this block else anywhere */
XFS_ALLOCTYPE_NEAR_BNO, /* in this a.g. and near this block */
XFS_ALLOCTYPE_THIS_BNO /* at exactly this block */
} xfs_alloctype_t;
#define XFS_ALLOC_TYPES \
{ XFS_ALLOCTYPE_ANY_AG, "ANY_AG" }, \
{ XFS_ALLOCTYPE_FIRST_AG, "FIRST_AG" }, \
{ XFS_ALLOCTYPE_START_AG, "START_AG" }, \
{ XFS_ALLOCTYPE_THIS_AG, "THIS_AG" }, \
{ XFS_ALLOCTYPE_START_BNO, "START_BNO" }, \
{ XFS_ALLOCTYPE_NEAR_BNO, "NEAR_BNO" }, \
{ XFS_ALLOCTYPE_THIS_BNO, "THIS_BNO" }Quoted 8 lines
Another issue with the above commit, that the output of softirq events becomes:
X-1701 [000] 1595.220739:
softirq_entry: softirq=1 action=TIMER_SOFTIRQCompared to the original output:
X-1701 [000] 1595.220739:
softirq_entry: softirq=1 action=TIMER
Which is trivially fixable, see above :)
Steven Rostedt
maintainerQuoted 7 lines · Christoph Hellwig
Quoted 2 lines · Li Zefan
It's great to boost recording of softirq events, but why not simply use
softirq_to_nameinTP_printk()?Because userspace programs using the binary trace buffer have no chance to retrieve the values from
softirq_to_name.
Quoted 8 lines
Quoted 4 lines
The above commit has 2 flaws:
- we waste memory defining local static struct
trace_print_flagsarray in each softirqTRACE_EVENTThat could be solved by declaring the array manually and just passing the address to
__print_symbolic. Steve, would that work? (also for__print_flags)
I added the code to make it an array, but we still need to allocate it for every trace call. Otherwise, how do we export the information to userspace. One trace event (currently) has no way of depending on another trace event. If two trace events want the same array, currently they both must allocate it. I haven't looked too deeply into seeing how to manage that. It's not that much memory duplicated. I hate to make the code even more complex just to save a couple of hundred of bytes.
Quoted 30 lines
Quoted 3 lines
- if someone adds/removes a softirq, he may not know
show_softirq_name()needs to be updatedJust make sure you have the translation defined next to the actual flags. This is what I do in the XFS tracer:
typedef enum
xfs_alloctype
{XFS_ALLOCTYPE_ANY_AG, /* allocate anywhere, use rotor */ XFS_ALLOCTYPE_FIRST_AG, /* ... start at ag 0 */ XFS_ALLOCTYPE_START_AG, /* anywhere, start in this a.g. */ XFS_ALLOCTYPE_THIS_AG, /* anywhere in this a.g. */
XFS_ALLOCTYPE_START_BNO, /* near this block else anywhere */
XFS_ALLOCTYPE_NEAR_BNO, /* in this a.g. and near this block */
XFS_ALLOCTYPE_THIS_BNO/* at exactly this block */
}xfs_alloctype_t;#define XFS_ALLOC_TYPES \ { XFS_ALLOCTYPE_ANY_AG, "ANY_AG" }, \ { XFS_ALLOCTYPE_FIRST_AG, "FIRST_AG" }, \ { XFS_ALLOCTYPE_START_AG, "START_AG" }, \ { XFS_ALLOCTYPE_THIS_AG, "THIS_AG" }, \ { XFS_ALLOCTYPE_START_BNO, "START_BNO" }, \ { XFS_ALLOCTYPE_NEAR_BNO, "NEAR_BNO" }, \ { XFS_ALLOCTYPE_THIS_BNO, "THIS_BNO" }Quoted 3 lines
Another issue with the above commit, that the output of softirq events becomes:
Quoted 7 lines
Quoted 5 lines
X-1701 [000] 1595.220739:
softirq_entry: softirq=1 action=TIMER_SOFTIRQCompared to the original output:
X-1701 [000] 1595.220739:
softirq_entry: softirq=1 action=TIMERWhich is trivially fixable, see above :)
Hmm, I thought I fixed that already. Perhaps it is in the code that Ingo has not pulled.
-- Steve
Quoted 8 lines · Steven Rostedt
Quoted 6 lines · Christoph Hellwig
Quoted 2 lines · Li Zefan
It's great to boost recording of softirq events, but why not simply use
softirq_to_nameinTP_printk()?Because userspace programs using the binary trace buffer have no chance to retrieve the values from
softirq_to_name.
Do you mean, by parsing the format file?
# cat events/irq/softirq_entry/format
...
print fmt: "softirq=%d action=%s", REC->vec, ({ static const struct trace_print_flags
symbols[] = { { HI_SOFTIRQ, "HI_SOFTIRQ" }, { TIMER_SOFTIRQ, "TIMER_SOFTIRQ" },
{ NET_TX_SOFTIRQ, "NET_TX_SOFTIRQ" }, { NET_RX_SOFTIRQ, "NET_RX_SOFTIRQ" },
{ BLOCK_SOFTIRQ, "BLOCK_SOFTIRQ" }, { TASKLET_SOFTIRQ, "TASKLET_SOFTIRQ" },
{ SCHED_SOFTIRQ, "SCHED_SOFTIRQ" }, { HRTIMER_SOFTIRQ, "HRTIMER_SOFTIRQ" },
{ RCU_SOFTIRQ, "RCU_SOFTIRQ" }, { -1, ((void *)0) }};
ftrace_print_symbols_seq(p, REC->vec, symbols); })Quoted 46 lines
Quoted 7 lines
Quoted 4 lines
The above commit has 2 flaws:
- we waste memory defining local static struct
trace_print_flagsarray in each softirqTRACE_EVENTThat could be solved by declaring the array manually and just passing the address to
__print_symbolic. Steve, would that work? (also for__print_flags)I added the code to make it an array, but we still need to allocate it for every trace call. Otherwise, how do we export the information to userspace. One trace event (currently) has no way of depending on another trace event. If two trace events want the same array, currently they both must allocate it. I haven't looked too deeply into seeing how to manage that. It's not that much memory duplicated. I hate to make the code even more complex just to save a couple of hundred of bytes.
Quoted 30 lines
Quoted 2 lines
- if someone adds/removes a softirq, he may not know
show_softirq_name()needs to be updatedJust make sure you have the translation defined next to the actual flags. This is what I do in the XFS tracer:
typedef enum
xfs_alloctype
{XFS_ALLOCTYPE_ANY_AG, /* allocate anywhere, use rotor */ XFS_ALLOCTYPE_FIRST_AG, /* ... start at ag 0 */ XFS_ALLOCTYPE_START_AG, /* anywhere, start in this a.g. */ XFS_ALLOCTYPE_THIS_AG, /* anywhere in this a.g. */
XFS_ALLOCTYPE_START_BNO, /* near this block else anywhere */
XFS_ALLOCTYPE_NEAR_BNO, /* in this a.g. and near this block */
XFS_ALLOCTYPE_THIS_BNO/* at exactly this block */
}xfs_alloctype_t;#define XFS_ALLOC_TYPES \ { XFS_ALLOCTYPE_ANY_AG, "ANY_AG" }, \ { XFS_ALLOCTYPE_FIRST_AG, "FIRST_AG" }, \ { XFS_ALLOCTYPE_START_AG, "START_AG" }, \ { XFS_ALLOCTYPE_THIS_AG, "THIS_AG" }, \ { XFS_ALLOCTYPE_START_BNO, "START_BNO" }, \ { XFS_ALLOCTYPE_NEAR_BNO, "NEAR_BNO" }, \ { XFS_ALLOCTYPE_THIS_BNO, "THIS_BNO" }Quoted 5 lines
Another issue with the above commit, that the output of softirq events becomes:
X-1701 [000] 1595.220739:
softirq_entry: softirq=1 action=TIMER_SOFTIRQ
Quoted 8 lines
Quoted 4 lines
Quoted 3 lines
Compared to the original output:
X-1701 [000] 1595.220739:
softirq_entry: softirq=1 action=TIMERWhich is trivially fixable, see above :)
Hmm, I thought I fixed that already. Perhaps it is in the code that Ingo has not pulled.
I guess you forgot to fix it. ;)
I made a comment on that patch, but seems you didn't update the patch:
http://lkml.org/lkml/2009/5/21/9
I'll make a patch for this.