Hello Salvatore,
On Fri, Nov 20, 2020 at 7:30 PM Salvatore Bonaccorso carnil@debian.org wrote:
Hi Andrey,
On Fri, Nov 20, 2020 at 05:31:59PM +0100, Andrey Zhizhikin wrote:
Hello Salvatore,
On Fri, Nov 20, 2020 at 4:53 PM Salvatore Bonaccorso carnil@debian.org wrote:
Hi Andrey,
On Fri, Nov 20, 2020 at 03:29:39PM +0100, Andrey Zhizhikin wrote:
Hello Salvatore,
On Fri, Nov 20, 2020 at 2:34 PM Salvatore Bonaccorso carnil@debian.org wrote:
Hi Andrey,
On Fri, Nov 20, 2020 at 10:54:22AM +0100, Andrey Zhizhikin wrote:
On Fri, Nov 20, 2020 at 8:39 AM Salvatore Bonaccorso carnil@debian.org wrote: > > This reverts commit 168200b6d6ea0cb5765943ec5da5b8149701f36a upstream. > (but only from 4.19.y)
This revert would fail the build of 4.19.y with gcc10, I believe the original commit was introduced to address exactly this case. If this is intended behavior that 4.19.y is not compiled with newer gcc versions - then this revert is OK.
TTBOMK, this would not regress the build for newer gcc (specifically gcc10) as 4.19.158 is failing perf tool builds there as well (without the above commit reverted). Just as an example v4.19.y does not have cff20b3151cc ("perf tests bp_account: Make global variable static") which is there in v5.6-rc6 to fix build failures with 10.0.1.
But it did regress builds with older gcc's as for instance used in Debian buster (gcc 8.3.0) since 4.19.152.
Do I possibly miss something? If there is a solution to make it build with newer GCCs and *not* regress previously working GCC versions then this is surely the best outcome though.
I guess (and from what I understand in Leo's reply), porting of 95c6fe970a01 ("perf cs-etm: Change tuple from traceID-CPU# to traceID-metadata") should solve the issue for both older and newer gcc versions.
The breakage is now in [tools/perf/util/cs-etm-decoder/cs-etm-decoder.c] file (which uses traceid_list inside). This is solved with the above commit, which concealed traceid_list internally inside [tools/perf/util/cs-etm.c] file and exposed to [tools/perf/util/cs-etm-decoder/cs-etm-decoder.c] via cs_etm__get_cpu() call.
Can you try out to port that commit to see if that would solve your regression?
So something like the following will compile as well with the older gcc version.
I realize: I mainline the order of the commits was:
95c6fe970a01 ("perf cs-etm: Change tuple from traceID-CPU# to traceID-metadata") 168200b6d6ea ("perf cs-etm: Move definition of 'traceid_list' global variable from header f ile")
But to v4.19.y only 168200b6d6ea was backported, and while that was done I now realize the comment was also changed including the change fom 95c6fe970a01.
Thus the proposed backported patch would drop the change in tools/perf/util/cs-etm.c to the comment as this was already done. Thecnically currently the comment would be wrong, because it reads:
/* RB tree for quick conversion between traceID and metadata pointers */
but backport of 95c6fe970a01 is not included.
Would the right thing to do thus be:
- Revert b801d568c7d8 "perf cs-etm: Move definition of 'traceid_list' global variable from header file"
- Backport 95c6fe970a01 ("perf cs-etm: Change tuple from traceID-CPU# to traceID-metadata")
- Backport 168200b6d6ea ("perf cs-etm: Move definition of 'traceid_list' global variable from header file")
Yes, I believe this would be the correct course of action here; this should cover the regression you've encountered and should ensure that perf builds on both the "old" and "new" gcc versions.
Although perf tools in v4.19.y won't compile with recent GCCs.
Greg did already queued up the first part of it, so the revert. I think we can pick the later two commits again up after the v4.19.159 release?
Sounds reasonable to me.
Regards, Salvatore