| Message ID | 20241114155822.898466-1-alan.maguire@oracle.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Check DW_OP_[GNU_]entry_value for possible parameter matching | expand |
On Thu, Nov 14, 2024 at 03:58:20PM +0000, Alan Maguire wrote: > Currently, pahole relies on DWARF to find whether a particular func > has its parameter mismatched with standard or optimized away. > In both these cases, the func will not be put in BTF and this > will prevent fentry/fexit tracing for these functions. > > The current parameter checking focuses on the first location/expression > to match intended parameter register. But in some cases, the first > location/expression does not have expected matching information, > but further location like DW_OP_[GNU_]entry_value can provide > information which matches the expected parameter register. > > Patch 1 supports this; patch 2 adds locking around dwarf_getlocation* > as it is unsafe in a multithreaded environment. > > Run ~4000 times without observing a segmentation fault (as compared > to without patch 2, where a segmentation fault is observed approximately > every 200 invokations). > > Changes since v1: > > - used Eduard's approach of using a __dwarf_getlocations() > internal wrapper (Eduard, patch 1). > - renamed function to parameter__reg(); did not rename > __dwarf_getlocations() since its functionality is based around > retrieving DWARF location info rather than parameter register > indices (Yonghong, patch 2) > - added locking around dwarf_getlocation*() usage in dwarf_loader > to avoid segmentation faults reported by Eduard (Jiri, Arnaldo, > patch 2) looks good, I got 95 more functions in clang build including perf_event_read and there's no change in generated functions with gcc build Acked-by: Jiri Olsa <jolsa@kernel.org> thanks, jirka > > Alan Maguire (1): > dwarf_loader: use libdw__lock for dwarf_getlocation(s) > > Eduard Zingerman (1): > dwarf_loader: Check DW_OP_[GNU_]entry_value for possible parameter > matching > > dwarf_loader.c | 121 +++++++++++++++++++++++++++++++++++++++---------- > 1 file changed, 96 insertions(+), 25 deletions(-) > > -- > 2.31.1 >
On Fri, Nov 15, 2024 at 10:34:35AM +0100, Jiri Olsa wrote: > On Thu, Nov 14, 2024 at 03:58:20PM +0000, Alan Maguire wrote: > > Currently, pahole relies on DWARF to find whether a particular func > > has its parameter mismatched with standard or optimized away. > > In both these cases, the func will not be put in BTF and this > > will prevent fentry/fexit tracing for these functions. > > > > The current parameter checking focuses on the first location/expression > > to match intended parameter register. But in some cases, the first > > location/expression does not have expected matching information, > > but further location like DW_OP_[GNU_]entry_value can provide > > information which matches the expected parameter register. > > > > Patch 1 supports this; patch 2 adds locking around dwarf_getlocation* > > as it is unsafe in a multithreaded environment. > > > > Run ~4000 times without observing a segmentation fault (as compared > > to without patch 2, where a segmentation fault is observed approximately > > every 200 invokations). > > > > Changes since v1: > > > > - used Eduard's approach of using a __dwarf_getlocations() > > internal wrapper (Eduard, patch 1). > > - renamed function to parameter__reg(); did not rename > > __dwarf_getlocations() since its functionality is based around > > retrieving DWARF location info rather than parameter register > > indices (Yonghong, patch 2) > > - added locking around dwarf_getlocation*() usage in dwarf_loader > > to avoid segmentation faults reported by Eduard (Jiri, Arnaldo, > > patch 2) > > looks good, I got 95 more functions in clang build including perf_event_read > and there's no change in generated functions with gcc build > > Acked-by: Jiri Olsa <jolsa@kernel.org> I think since you tested it we can stick this, right? Tested-by: Jiri Olsa <jolsa@kernel.org> > thanks, > jirka > > > > > Alan Maguire (1): > > dwarf_loader: use libdw__lock for dwarf_getlocation(s) > > > > Eduard Zingerman (1): > > dwarf_loader: Check DW_OP_[GNU_]entry_value for possible parameter > > matching > > > > dwarf_loader.c | 121 +++++++++++++++++++++++++++++++++++++++---------- > > 1 file changed, 96 insertions(+), 25 deletions(-) > > > > -- > > 2.31.1 > >
On Fri, Nov 15, 2024 at 11:52:38AM -0300, Arnaldo Carvalho de Melo wrote: > On Fri, Nov 15, 2024 at 10:34:35AM +0100, Jiri Olsa wrote: > > On Thu, Nov 14, 2024 at 03:58:20PM +0000, Alan Maguire wrote: > > > Currently, pahole relies on DWARF to find whether a particular func > > > has its parameter mismatched with standard or optimized away. > > > In both these cases, the func will not be put in BTF and this > > > will prevent fentry/fexit tracing for these functions. > > > > > > The current parameter checking focuses on the first location/expression > > > to match intended parameter register. But in some cases, the first > > > location/expression does not have expected matching information, > > > but further location like DW_OP_[GNU_]entry_value can provide > > > information which matches the expected parameter register. > > > > > > Patch 1 supports this; patch 2 adds locking around dwarf_getlocation* > > > as it is unsafe in a multithreaded environment. > > > > > > Run ~4000 times without observing a segmentation fault (as compared > > > to without patch 2, where a segmentation fault is observed approximately > > > every 200 invokations). > > > > > > Changes since v1: > > > > > > - used Eduard's approach of using a __dwarf_getlocations() > > > internal wrapper (Eduard, patch 1). > > > - renamed function to parameter__reg(); did not rename > > > __dwarf_getlocations() since its functionality is based around > > > retrieving DWARF location info rather than parameter register > > > indices (Yonghong, patch 2) > > > - added locking around dwarf_getlocation*() usage in dwarf_loader > > > to avoid segmentation faults reported by Eduard (Jiri, Arnaldo, > > > patch 2) > > > > looks good, I got 95 more functions in clang build including perf_event_read > > and there's no change in generated functions with gcc build > > > > Acked-by: Jiri Olsa <jolsa@kernel.org> > > I think since you tested it we can stick this, right? > > Tested-by: Jiri Olsa <jolsa@kernel.org> yep, sure jirka > > > thanks, > > jirka > > > > > > > > Alan Maguire (1): > > > dwarf_loader: use libdw__lock for dwarf_getlocation(s) > > > > > > Eduard Zingerman (1): > > > dwarf_loader: Check DW_OP_[GNU_]entry_value for possible parameter > > > matching > > > > > > dwarf_loader.c | 121 +++++++++++++++++++++++++++++++++++++++---------- > > > 1 file changed, 96 insertions(+), 25 deletions(-) > > > > > > -- > > > 2.31.1 > > >
Currently, pahole relies on DWARF to find whether a particular func has its parameter mismatched with standard or optimized away. In both these cases, the func will not be put in BTF and this will prevent fentry/fexit tracing for these functions. The current parameter checking focuses on the first location/expression to match intended parameter register. But in some cases, the first location/expression does not have expected matching information, but further location like DW_OP_[GNU_]entry_value can provide information which matches the expected parameter register. Patch 1 supports this; patch 2 adds locking around dwarf_getlocation* as it is unsafe in a multithreaded environment. Run ~4000 times without observing a segmentation fault (as compared to without patch 2, where a segmentation fault is observed approximately every 200 invokations). Changes since v1: - used Eduard's approach of using a __dwarf_getlocations() internal wrapper (Eduard, patch 1). - renamed function to parameter__reg(); did not rename __dwarf_getlocations() since its functionality is based around retrieving DWARF location info rather than parameter register indices (Yonghong, patch 2) - added locking around dwarf_getlocation*() usage in dwarf_loader to avoid segmentation faults reported by Eduard (Jiri, Arnaldo, patch 2) Alan Maguire (1): dwarf_loader: use libdw__lock for dwarf_getlocation(s) Eduard Zingerman (1): dwarf_loader: Check DW_OP_[GNU_]entry_value for possible parameter matching dwarf_loader.c | 121 +++++++++++++++++++++++++++++++++++++++---------- 1 file changed, 96 insertions(+), 25 deletions(-)