Mathieu Desnoyers -- November 2005 This is a complete rework of genevent. The 'genevent' program parses event descriptions and generates the inline functions to record events in the kernel. There are several files in the directory: genevent.c, genevent.h, crc32.tab, parser.c and parser.h In fact, crc32.tab, parser.c and parser.h are the same files as those in LTT library. Important notes : * Do not put "-" symbols in facilities name. * Use the exact same name for facility xml file and for facility name. Note about strings : There are three methods to write strings in genevent, each suitable and efficient for a particular case. They are explained here from the fastest to the slowest. 1 - The C code presents a fixed size string. For example, you find : char mystring[10]; as string definition. you must then define it as an array of char : Note, however, that you might not want to declare a fixed size for trace size and unnecessary copy matters. For instance, on a 32 bits architecture, copying a n bytes array takes approximately* n/4 memory read and write, for n/2 memory operations. Using the slower method described in (3), with a strlen and memcpy, where "u" is the number of used caracters, takes u+1 reads for the strlen, and approximately* (u+1)/4 read and write for the memcpy, for a total of : (3/2)*(u+1) memory access. So, if (n/2) > (3/2)*(u+1), or : n > 3*u+3 where n is the size of the array u is the average number of used caracters (excluding the \0) it becomes faster to use the method number 3 with strlen. 2 - The C code presents a variable size string together with its size. A typical use for this case is filenames in the Linux kernel. The dentry strucure has a d_name member, which is a struct qstr containing a unsigned int len and const unsigned char *name. you must use a sequence to declare this efficiently : 3 - The C code presents a \0 terminated string. This is the slowest, but most convenient way to declare a string. You are discouraged to use it when options 1 or 2 are available. It will dynamically calculate the string length (byte by byte read) and only afterward do a memcpy. Note that, as explained in 1, if n > 3*u+3, it becomes faster to use this method instead of copying the whole fixed size array. Declare like this : Here is a brief description of how to use genevent. make make install * Add new events to the kernel with genevent su - cd /usr/local/share/LinuxTraceToolkitViewer/facilities cp process.xml yourfacility.xml * edit yourfacility.xml to fit your needs. cd /tmp /usr/local/bin/genevent /usr/local/share/LinuxTraceToolkitViewer/yourfacility.xml cp ltt-facility-yourfacility.h ltt-facility-id-yourfacility.h \ /usr/src/linux-2.6.12-rc4-mm2-lttng-0.2/include/linux/ltt cp ltt-facility-loader-yourfacility.c ltt-facility-loader-yourfacility.h \ /usr/src/linux-2.6.12-rc4-mm2-lttng-0.2/ltt * edit the kernel file you want to instrument - Add #include at the beginning of the file. - Add a call to the tracing functions. See their names and parameters in /usr/src/linux-2.6.12-rc4-mm2-lttng-0.2/include/linux/ltt/ltt-facility-yourfacility.h * The approximation comes from the fact that copies of number of caracters non multiple of the architecture size takes more operations (maximum of : (architecture size (in bytes) - 1) operations).