Am 2019-05-27 21:04, schrieb Sven Schnelle:
This patch implements dynamic ftrace for PA-RISC. The required mcount
call sequences can get pretty long, so instead of patching the
whole call sequence out of the functions, we are using
-fpatchable-function-entry from gcc. This puts a configurable amount of
NOPS before/at the start of the function. Taking do_sys_open() as
example,
which would look like this when the call is patched out:
1036b248: 08 00 02 40 nop
1036b24c: 08 00 02 40 nop
1036b250: 08 00 02 40 nop
1036b254: 08 00 02 40 nop
1036b258 <do_sys_open>:
1036b258: 08 00 02 40 nop
1036b25c: 08 03 02 41 copy r3,r1
1036b260: 6b c2 3f d9 stw rp,-14(sp)
1036b264: 08 1e 02 43 copy sp,r3
1036b268: 6f c1 01 00 stw,ma r1,80(sp)
When ftrace gets enabled for this function the kernel will patch these
NOPs to:
1036b248: 10 19 57 20 <address of ftrace>
1036b24c: 6f c1 00 80 stw,ma r1,40(sp)
1036b250: 48 21 3f d1 ldw -18(r1),r1
1036b254: e8 20 c0 02 bv,n r0(r1)
1036b258 <do_sys_open>:
1036b258: e8 3f 1f df b,l,n .-c,r1
1036b25c: 08 03 02 41 copy r3,r1
1036b260: 6b c2 3f d9 stw rp,-14(sp)
1036b264: 08 1e 02 43 copy sp,r3
1036b268: 6f c1 01 00 stw,ma r1,80(sp)
So the first NOP in do_sys_open() will be patched to jump backwards
into
some minimal trampoline code which pushes a stackframe, saves r1 which
holds the return address, loads the address of the real ftrace
function,
and branches to that location. For 64 Bit things are getting a bit more
complicated (and longer) because we must make sure that the address of
ftrace location is 8 byte aligned, and the offset passed to ldd for
fetching the address is 8 byte aligned as well.
Note that gcc has a bug which misplaces the function label, and needs a
patch to make dynamic ftrace work.
This changelog entry _really_ needs a link to the gcc bugtracker so
tracking
this when doing downstream toolchain patching isn't driving other people
mad
trying to find out what you are talking about.
Eike
