Hello everyone,
Over the last few months, I have discussed with Andew Morton and Andi Kleen how we could merge perfmon2 with the mainline Linux kernel. From those discussions, he became clear the current code base could not be merged as is. It was too big and complex for LKML people to get a handle on. As you may have seen in 2.6.25, I have started major restructuring of the code to make it more modular and readable. That X86 support has been changed heavily to pushed all model-specific code into its corresponding PMU description module. The restructuring effort will continue in 2.6.26 along with some simplifications wherever possible. However, it became clear to me that it would be quite hard to split the code into manageable pieces to feed to LKML. So I have decided to take another approach, suggested by akpm and ak. I have started more or less from scratch, importing only selected bits and pieces from our current development tree. The first piece I want to focus on is per-thread monitoring, and in particular per-thread counting. Per-thread monitoring is a key value-add of perfmon. I think this needs to be in first. To that extent, I have created a separate tree which only contains enough code to support per-thread counting. This is limited to X86 processors only, both AMD and Intel. On the Intel side, so far only Intel Core-based processors are supported. The code has been stripped down to the bare minimum. Yet existing perfmon application and tools doing self-monitoring or monitoring of another thread can run unmodified assuming you use the current libpfm. The following functionalities are missing/changed for this lightweight version: * no system-wide monitoring * no kernel support for sampling * no sampling buffer formats * no PMU description modules. Mapping tables are compiled in. * no event set multiplexing * The following syscalls have been removed: - pfm_restart() - pfm_getinfo_evtsets() - pfm_create_evtsets() - pfm_delete_evtsets() * pfm_start() command has lost its (optional) argument (pfarg_start_t). * the context file descriptor is not readable, nor pollable. No pfarg_msg_t * No signal is sent on 64-bit counter overflow. Yet 64-bit virtualization of the counters is maintained * Mutual exclusion with Oprofile is maintained. * PMU sharing with NMI watchdog is maintained * /sys/kernel/perfmon is maintained, though with fewer entries * debugfs statistics are maintained as compile time option * Itanium, MIPS, Power, SPARC supprt has been removed As a results the patch against 2.6.26-rc3 is 200KB and about 1900 lines of C. I think I can squeeze it some more. As I said, certain tools and test programs still work because the shared data structures have not changed int size, though some fields have disappeared. Once I have something that works decently, I will probably create another GIT tree on kernel.org. Note that 200KB is too big for LKML, so the patch would have to be broken down some more. Hopefully with this codebase we will increase our chances of getting some pieces merged. Note that the current GIT tree is not abandonned, far from it. Instead it will remain the bleeding edge tree where all new developments must take place. Functionalities will then migrate from this tree to the merge tree and eventually to mainline. I would like to get something in good shape for Ottawa Linux Symposium (OLS) which is mid-July. Andi Kleen will deliver a paper/presentation on 'submitted patches for the kernel'. I'd like to show him we understand how to do this and are actively working on it Thanks. ------------------------------------------------------------------------- This SF.net email is sponsored by: Microsoft Defy all challenges. Microsoft(R) Visual Studio 2008. http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/ _______________________________________________ perfmon2-devel mailing list perfmon2-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
