When writing probes, you can do the following: ==== probe process("python").library("libpython2.7.so*").function("PyEval_EvalFrameEx") { # ... probe contents here ... } ==== Notice that I'm just probing "python" as the "process" argument, a non-absolute path. The translator will look up "python" in my PATH and find it. Notice I'm using a wildcarded and non-absolute path for the "library" argument. The above works fine and is very user friendly. So, when using @cast(), I'd like to specify the symbol search path the same way: ==== # cat pytest1.stp function foo:long(object:long) { printf("%d\n", @cast(object, "PyFrameObject", "python:libpython2.7.so*")->f_lineno) } probe process("python").library("libpython2.7.so*").function("PyEval_EvalFrameEx") { foo($f) exit() } # stap -v ./pytest1.stp -c "python -c 'import sys; sys.exit(0)'" Pass 1: parsed user script and 117 library scripts using 242808virt/39472res/7568shr/32212data kb, in 190usr/30sys/234real ms. semantic error: while processing function foo semantic error: type definition 'PyFrameObject' not found in 'python:libpython2.7.so*': operator '@cast' at ./pytest1.stp:4:3 source: @cast(object, "PyFrameObject", "python:libpython2.7.so*")->f_lineno) ^ Pass 2: analyzed script: 1 probe, 1 function, 0 embeds, 0 globals using 250612virt/48776res/8880shr/40016data kb, in 120usr/380sys/498real ms. Pass 2: analysis failed. [man error::pass2] ==== So, that didn't work. Let's try without the wildcard: ==== # cat pytest2.stp function foo:long(object:long) { printf("%d\n", @cast(object, "PyFrameObject", "python:libpython2.7.so")->f_lineno) } probe process("python").library("libpython2.7.so*").function("PyEval_EvalFrameEx") { foo($f) exit() } # stap -v ./pytest2.stp -c "python -c 'import sys; sys.exit(0)'" Pass 1: parsed user script and 117 library scripts using 242808virt/39644res/7740shr/32212data kb, in 200usr/40sys/240real ms. semantic error: while processing function foo semantic error: type definition 'PyFrameObject' not found in 'python:libpython2.7.so': operator '@cast' at ./pytest2.stp:4:3 source: @cast(object, "PyFrameObject", "python:libpython2.7.so")->f_lineno) ^ Pass 2: analyzed script: 1 probe, 3 functions, 0 embeds, 0 globals using 250608virt/48792res/8892shr/40012data kb, in 100usr/400sys/511real ms. Pass 2: analysis failed. [man error::pass2] ==== OK, on this system /usr/lib64/libpython2.7.so is a symbolic link to /usr/lib64/libpython2.7.so.1.0. Let's give that a shot: ==== # cat pytest3.stp function foo:long(object:long) { printf("%d\n", @cast(object, "PyFrameObject", "python:libpython2.7.so.1.0")->f_lineno) } probe process("python").library("libpython2.7.so*").function("PyEval_EvalFrameEx") { foo($f) exit() } # stap -v ./pytest3.stp -c "python -c 'import sys; sys.exit(0)'" Pass 1: parsed user script and 117 library scripts using 242804virt/39504res/7604shr/32208data kb, in 210usr/30sys/242real ms. semantic error: while processing function foo semantic error: type definition 'PyFrameObject' not found in 'python:libpython2.7.so.1.0': operator '@cast' at ./pytest3.stp:4:3 source: @cast(object, "PyFrameObject", "python:libpython2.7.so.1.0")->f_lineno) ^ Pass 2: analyzed script: 1 probe, 1 function, 0 embeds, 0 globals using 250608virt/48804res/8912shr/40012data kb, in 100usr/370sys/492real ms. Pass 2: analysis failed. [man error::pass2] ==== Nope. The only thing I can seem to get to work for @cast() is a full absolute path to the library. One reason why this is annoying is that to handle both 32-bit distros and 64-bit distros, I'll have to specify the library in both /usr/lib/ and /usr/lib64: ==== # cat pytest4.stp function foo:long(object:long) { printf("%d\n", @cast(object, "PyFrameObject", "python:/usr/lib64/libpython2.7.so:/usr/lib/libpython2.7.so")->f_lineno) } probe process("python").library("libpython2.7.so*").function("PyEval_EvalFrameEx") { foo($f) exit() } # stap -v ./pytest4.stp -c "python -c 'import sys; sys.exit(0)'" Pass 1: parsed user script and 117 library scripts using 242808virt/39712res/7812shr/32212data kb, in 210usr/30sys/238real ms. Pass 2: analyzed script: 1 probe, 4 functions, 0 embeds, 0 globals using 251388virt/49532res/8828shr/40792data kb, in 120usr/190sys/317real ms. Pass 3: translated to C into "/tmp/stapYGfpK4/stap_d4c98f72a2d5b7cef816271d6bd3e397_3084_src.c" using 251388virt/49724res/9020shr/40792data kb, in 20usr/190sys/206real ms. Pass 4: compiled C into "stap_d4c98f72a2d5b7cef816271d6bd3e397_3084.ko" in 4730usr/1000sys/5459real ms. Pass 5: starting run. 59 Pass 5: run completed in 30usr/70sys/392real ms. ====
Having a search path in @cast in a function() makes it possible that a mismatch will exist between the probe (whose $context variable is being passed) and the @cast (who interprets that pointer). What you seem to really want is to make @cast() in a function refer to - match - the context of the probe that invoked it. This is kind of what we do already for @cast()s in probe handler bodies. It's as though we want to specialize the functions invoked from a probe handler by the probe handler, almost as if they were inlined. Then syntactic elements inside the functions (even $vars) could conceivably be resolved in the context of each probe where those functions are actually called from. What do you think?
(In reply to Frank Ch. Eigler from comment #1) > Having a search path in @cast in a function() makes it possible that a > mismatch will exist between the probe (whose $context variable is being > passed) and the @cast (who interprets that pointer). What you seem to > really want is to make @cast() in a function refer to - match - the context > of the probe that invoked it. This is kind of what we do already for > @cast()s in probe handler bodies. > > It's as though we want to specialize the functions invoked from a probe > handler by the probe handler, almost as if they were inlined. Then > syntactic elements inside the functions (even $vars) could conceivably be > resolved in the context of each probe where those functions are actually > called from. > > What do you think? That's an interesting idea. If it is doable, I'd say it would be quite handy.
more general solution