record failure with 3.14-rc4
Hi guys,
If I try to run perf record as a non-root user, I end up with the following (unhelpful) error:
$ perf record -e cycles ls [...] Not enough memory for reading perf file header
This is because the addresses in /proc/kallsyms always read as 0x0 when viewed by a non-privileged user, causing kallsyms__get_function_start to return 0x0 in args.start. machine__create_kernel_maps then treats this as an error an returns -1 to perf_session__create_kernel_maps, causing perf_session__new to fail and perf to exit.
The perf tool code in 3.13 is perfectly happy creating maps at 0x0, but I can see this has changed quite substantially in the recent merge window.
Any ideas? We could fix kallsyms__get_function_start to return < 0 on failure then fix the checks in the callers. Adrian?
Will
Quoted 15 lines · Will Deacon
Hi guys,
If I try to run perf record as a non-root user, I end up with the following (unhelpful) error:
$ perf record -e cycles ls [...] Not enough memory for reading perf file headerThis is because the addresses in
/proc/kallsymsalways read as0x0when viewed by a non-privileged user, causingkallsyms__get_function_startto return0x0in args.start.machine__create_kernel_mapsthen treats this as an error an returns -1 toperf_session__create_kernel_maps, causingperf_session__newto fail and perf to exit.
Quoted 5 lines
The perf tool code in 3.13 is perfectly happy creating maps at
0x0, but I can see this has changed quite substantially in the recent merge window.Any ideas? We could fix
kallsyms__get_function_startto return < 0 on failure then fix the checks in the callers. Adrian?
hi,
not sure it's related, but i sent out rfc some time ago for similar bad error messages:
http://marc.info/?l=linux-kernel&m=138572558210216&w=2
maybe it could be updated for your case as well,
I'll try to check deeply later
jirka