Skip to contentSkip to conversation

perf tool: load data variable symbols

RFC Lin Ming 3 participants 3 messages Feb 10, 2011

lore
Share & tools
lore

Lin Ming

author
#0 lore raw

Hi, all

Currently, perf tool only load function symbols when parsing perf.data.

But it is also helpful if variable symbols can be loaded.

For example, Intel load latency monitoring facility records data linear address of the load operation. It's useful if the data linear address is resolved into symbol, just like functions.

enum map_type {
        MAP__FUNCTION = 0,
        MAP__VARIABLE,
};

We already have MAP__VARIABLE defined, although it's not used now.

For both kernel and userspace applications, we can load the variable symbols from .data and .bss section.

What do you think?

Thanks,
Lin Ming

#1 lore raw

Em Thu, Feb 10, 2011 at 10:15:33PM +0800, Lin Ming escreveu:

Quoted 15 lines

Hi, all

Currently, perf tool only load function symbols when parsing perf.data.

But it is also helpful if variable symbols can be loaded.

For example, Intel load latency monitoring facility records data linear address of the load operation. It's useful if the data linear address is resolved into symbol, just like functions.

enum map_type {
        MAP__FUNCTION = 0,
        MAP__VARIABLE,
};
Quoted 6 lines

We already have MAP__VARIABLE defined, although it's not used now.

For both kernel and userspace applications, we can load the variable symbols from .data and .bss section.

What do you think?

It suposedly already works, as you noticed there are no users in the tree, IIRC I wrote that because somebody at IBM asked for it for some tool and since then I haven't heard about such tool.

It will load global variable locations from the symtab, like:

[acme@felicio linux]$ readelf -s ~/bin/perf | grep OBJECT | grep nr_

  216: 000000000076b9b0  8 OBJECT  LOCAL  DEFAULT   29 nr_tasks
  222: 00000000007eba48  8 OBJECT  LOCAL  DEFAULT   29 nr_run_events
  223: 00000000007eba50  8 OBJECT  LOCAL  DEFAULT   29 nr_sleep_events
  224: 00000000007eba58  8 OBJECT  LOCAL  DEFAULT   29 nr_wakeup_events
  225: 00000000007eba60  8 OBJECT  LOCAL  DEFAULT   29 nr_sleep_corrections
  226: 00000000007eba68  8 OBJECT  LOCAL  DEFAULT   29 nr_run_events_optimized
  233: 00000000007ebaa0  8 OBJECT  LOCAL  DEFAULT   29 nr_runs
  238: 00000000007ebac0  8 OBJECT  LOCAL  DEFAULT   29 nr_timestamps
  239: 00000000007ebac8  8 OBJECT  LOCAL  DEFAULT   29 nr_unordered_timestamps
  240: 00000000007ebad0  8 OBJECT  LOCAL  DEFAULT   29 nr_state_machine_bugs
  241: 00000000007ebad8  8 OBJECT  LOCAL  DEFAULT   29 nr_context_switch_bugs
  242: 00000000007ebae0  8 OBJECT  LOCAL  DEFAULT   29 nr_events
  243: 00000000007ebae8  8 OBJECT  LOCAL  DEFAULT   29 nr_lost_chunks
  244: 00000000007ebaf0  8 OBJECT  LOCAL  DEFAULT   29 nr_lost_events
  624: 0000000000817d28  8 OBJECT  LOCAL  DEFAULT   29 nr_unordered
  700: 000000000081a200  8 OBJECT  LOCAL  DEFAULT   29 nr_allocs
  701: 000000000081a208  8 OBJECT  LOCAL  DEFAULT   29 nr_cross_allocs
 1145: 0000000000831180  4 OBJECT  LOCAL  DEFAULT   29 vmlinux_path__nr_entries

[acme@felicio linux]$

But for things to start to get interesting we need to extend what we have in 'perf probe', i.e. the DWARF location expressions, to do reverse lookup based on data address from sample, register file snapshots, and those DWARF location expressions.

- Arnaldo

#2 lore raw
Quoted 48 lines · Arnaldo Carvalho de Melo

Em Thu, Feb 10, 2011 at 10:15:33PM +0800, Lin Ming escreveu:

Quoted 15 lines

Hi, all

Currently, perf tool only load function symbols when parsing perf.data.

But it is also helpful if variable symbols can be loaded.

For example, Intel load latency monitoring facility records data linear address of the load operation. It's useful if the data linear address is resolved into symbol, just like functions.

enum map_type {
        MAP__FUNCTION = 0,
        MAP__VARIABLE,
};
Quoted 6 lines

We already have MAP__VARIABLE defined, although it's not used now.

For both kernel and userspace applications, we can load the variable symbols from .data and .bss section.

What do you think?

It suposedly already works, as you noticed there are no users in the tree, IIRC I wrote that because somebody at IBM asked for it for some tool and since then I haven't heard about such tool.

It will load global variable locations from the symtab, like:

[acme@felicio linux]$ readelf -s ~/bin/perf | grep OBJECT | grep nr_

  216: 000000000076b9b0  8 OBJECT  LOCAL  DEFAULT   29 nr_tasks
  222: 00000000007eba48  8 OBJECT  LOCAL  DEFAULT   29 nr_run_events
  223: 00000000007eba50  8 OBJECT  LOCAL  DEFAULT   29 nr_sleep_events
  224: 00000000007eba58  8 OBJECT  LOCAL  DEFAULT   29 nr_wakeup_events
  225: 00000000007eba60  8 OBJECT  LOCAL  DEFAULT   29 nr_sleep_corrections
  226: 00000000007eba68  8 OBJECT  LOCAL  DEFAULT   29 nr_run_events_optimized
  233: 00000000007ebaa0  8 OBJECT  LOCAL  DEFAULT   29 nr_runs
  238: 00000000007ebac0  8 OBJECT  LOCAL  DEFAULT   29 nr_timestamps
  239: 00000000007ebac8  8 OBJECT  LOCAL  DEFAULT   29 nr_unordered_timestamps
  240: 00000000007ebad0  8 OBJECT  LOCAL  DEFAULT   29 nr_state_machine_bugs
  241: 00000000007ebad8  8 OBJECT  LOCAL  DEFAULT   29 nr_context_switch_bugs
  242: 00000000007ebae0  8 OBJECT  LOCAL  DEFAULT   29 nr_events
  243: 00000000007ebae8  8 OBJECT  LOCAL  DEFAULT   29 nr_lost_chunks
  244: 00000000007ebaf0  8 OBJECT  LOCAL  DEFAULT   29 nr_lost_events
  624: 0000000000817d28  8 OBJECT  LOCAL  DEFAULT   29 nr_unordered
  700: 000000000081a200  8 OBJECT  LOCAL  DEFAULT   29 nr_allocs
  701: 000000000081a208  8 OBJECT  LOCAL  DEFAULT   29 nr_cross_allocs
 1145: 0000000000831180  4 OBJECT  LOCAL  DEFAULT   29 vmlinux_path__nr_entries
Quoted 6 lines

[acme@felicio linux]$

But for things to start to get interesting we need to extend what we have in 'perf probe', i.e. the DWARF location expressions, to do reverse lookup based on data address from sample, register file snapshots, and those DWARF location expressions.

Another way is to reverse map the reported IP (PEBS + fixup) to a data access using the dwarf info. That would also work for dynamically allocated data structures.

(clearly you'd loose variable things like which entry in an array, but you should still be able to identify the structure members)