char netlib_id[]="\ @(#)netlib.c (c) Copyright 1993-2012 Hewlett-Packard Company. Version 2.6.0"; /****************************************************************/ /* */ /* netlib.c */ /* */ /* the common utility routines available to all... */ /* */ /* establish_control() establish the control socket */ /* calibrate_local_cpu() do local cpu calibration */ /* calibrate_remote_cpu() do remote cpu calibration */ /* send_request() send a request to the remote */ /* recv_response() receive a response from remote */ /* send_response() send a response to the remote */ /* recv_request() recv a request from the remote */ /* dump_request() dump request contents */ /* dump_response() dump response contents */ /* cpu_start() start measuring cpu */ /* cpu_stop() stop measuring cpu */ /* calc_cpu_util() calculate the cpu utilization */ /* calc_service_demand() calculate the service demand */ /* calc_thruput() calulate the tput in units */ /* calibrate() really calibrate local cpu */ /* identify_local() print local host information */ /* identify_remote() print remote host information */ /* format_number() format the number (KB, MB,etc) */ /* format_units() return the format in english */ /* msec_sleep() sleep for some msecs */ /* start_timer() start a timer */ /* random_ip_address() select a random IP address from */ /* specified range */ /* */ /* the routines you get when WANT_DLPI is defined... */ /* ...all moved to src/nettest_dlpi.c */ /* */ /* dl_open() open a file descriptor and */ /* attach to the card */ /* dl_mtu() find the MTU of the card */ /* dl_bind() bind the sap do the card */ /* dl_connect() sender's have of connect */ /* dl_accpet() receiver's half of connect */ /* dl_set_window() set the window size */ /* dl_stats() retrieve statistics */ /* dl_send_disc() initiate disconnect (sender) */ /* dl_recv_disc() accept disconnect (receiver) */ /****************************************************************/ /****************************************************************/ /* */ /* Global include files */ /* */ /****************************************************************/ #ifdef HAVE_CONFIG_H #include #endif /* It would seem that most of the includes being done here from "sys/" actually have higher-level wrappers at just /usr/include. This is based on a spot-check of a couple systems at my disposal. If you have trouble compiling you may want to add "sys/" raj 10/95 */ #include #include #ifdef HAVE_SYSCALL_H #include #endif #ifdef MPE # define NSIG _NSIG #endif /* MPE */ #include #include #include #include #include #include #include #ifdef HAVE_ENDIAN_H #include #endif #ifndef WIN32 /* at some point, I would like to get rid of all these "sys/" includes where appropriate. if you have a system that requires/ them, speak now, or your system may not compile later revisions of netperf. raj 1/96 */ #include #include #include #ifndef MPE #include #include #endif /* MPE */ #include #include #include #include #include #include #if !defined(MPE) && !defined(__VMS) #include #endif /* MPE */ #else /* WIN32 */ #include #include #include #define netperf_socklen_t socklen_t #include #include /* the only time someone should need to define DONT_IPV6 in the "sources" file is if they are trying to compile on Windows 2000 or NT4 and I suspect this may not be their only problem :) */ #ifndef DONT_IPV6 #include #endif #include #define SIGALRM (14) #define sleep(x) Sleep((x)*1000) #endif /* WIN32 */ #ifdef HAVE_UNAME #include #endif #ifdef _AIX #include #include #include #define PRIORITY PRI_LOW #else/* _AIX */ #ifdef __sgi #include #include #define PRIORITY NDPLOMIN #endif /* __sgi */ #endif /* _AIX */ #ifdef HAVE_MPCTL #include #endif #if !defined(HAVE_GETADDRINFO) || !defined(HAVE_GETNAMEINFO) # include "missing/getaddrinfo.h" #endif #include "hist.h" /****************************************************************/ /* */ /* Local Include Files */ /* */ /****************************************************************/ #define NETLIB #include "netlib.h" #include "netsh.h" #include "netcpu.h" #include "netperf_version.h" /****************************************************************/ /* */ /* Global constants, macros and variables */ /* */ /****************************************************************/ #if defined(WIN32) || defined(__VMS) struct timezone { int dummy ; } ; #ifndef __VMS SOCKET win_kludge_socket = INVALID_SOCKET; SOCKET win_kludge_socket2 = INVALID_SOCKET; #endif /* __VMS */ #endif /* WIN32 || __VMS */ #ifndef LONG_LONG_MAX #define LONG_LONG_MAX 9223372036854775807LL #endif /* LONG_LONG_MAX */ /* older versions of netperf knew about the HP kernel IDLE counter. this is now obsolete - in favor of either pstat(), times, or a process-level looper process. we also now require support for the "long" integer type. raj 4/95. */ int lib_num_loc_cpus, /* the number of cpus in the system */ lib_num_rem_cpus; /* how many we think are in the remote */ struct cpu_stats_struct lib_local_cpu_stats, lib_remote_cpu_stats; #define PAGES_PER_CHILD 2 int lib_use_idle; int cpu_method; struct timeval time1, time2; struct timezone tz; float lib_elapsed, lib_local_maxrate, lib_remote_maxrate; float lib_local_per_cpu_util[MAXCPUS]; int lib_cpu_map[MAXCPUS]; int *request_array; int *response_array; /* INVALID_SOCKET == INVALID_HANDLE_VALUE == (unsigned int)(~0) == -1 */ SOCKET netlib_control = INVALID_SOCKET; SOCKET server_sock = INVALID_SOCKET; int control_family = AF_UNSPEC; /* global variables to hold the value for processor affinity */ int local_proc_affinity = -1,remote_proc_affinity = -1; /* these are to allow netperf to be run easily through those evil, end-to-end breaking things known as firewalls */ char local_data_port[10]; char remote_data_port[10]; char *local_data_address=NULL; char *remote_data_address=NULL; char *local_sysname, *remote_sysname; char *local_release, *remote_release; char *local_version, *remote_version; char *local_machine, *remote_machine; int local_data_family=AF_UNSPEC; int remote_data_family=AF_UNSPEC; char *netperf_version; enum netperf_output_modes netperf_output_mode = HUMAN; /* in the past, I was overlaying a structure on an array of ints. now I am going to have a "real" structure, and point an array of ints at it. the real structure will be forced to the same alignment as the type "double." this change will mean that pre-2.1 netperfs cannot be mixed with 2.1 and later. raj 11/95 */ union netperf_request_struct netperf_request; union netperf_response_struct netperf_response; FILE *where; char libfmt = '?'; #ifdef WIN32 HANDLE hAlarm = INVALID_HANDLE_VALUE; int timed_out=0; #endif int times_up; #ifdef WIN32 /* we use a getopt implementation from net.sources */ /* * get option letter from argument vector */ int opterr = 1, /* should error messages be printed? */ optind = 1, /* index into parent argv vector */ optopt; /* character checked for validity */ char *optarg; /* argument associated with option */ #define EMSG "" #endif /* WIN32 */ static int measuring_cpu; int netlib_get_page_size(void) { /* not all systems seem to have the sysconf for page size. for those which do not, we will assume that the page size is 8192 bytes. this should be more than enough to be sure that there is no page or cache thrashing by looper processes on MP systems. otherwise that's really just too bad - such systems should define _SC_PAGE_SIZE - raj 4/95 */ #ifndef _SC_PAGE_SIZE #ifdef WIN32 SYSTEM_INFO SystemInfo; GetSystemInfo(&SystemInfo); return SystemInfo.dwPageSize; #else return(8192L); #endif /* WIN32 */ #else return(sysconf(_SC_PAGE_SIZE)); #endif /* _SC_PAGE_SIZE */ } #ifdef WANT_INTERVALS #ifdef WIN32 HANDLE WinTimer; UINT timerRes; void stop_itimer() { CancelWaitableTimer(WinTimer); CloseHandle(WinTimer); timeEndPeriod(timerRes); } #else static unsigned int usec_per_itvl; void stop_itimer() { struct itimerval new_interval; struct itimerval old_interval; new_interval.it_interval.tv_sec = 0; new_interval.it_interval.tv_usec = 0; new_interval.it_value.tv_sec = 0; new_interval.it_value.tv_usec = 0; if (setitimer(ITIMER_REAL,&new_interval,&old_interval) != 0) { /* there was a problem arming the interval timer */ perror("netperf: setitimer"); exit(1); } return; } #endif /* WIN32 */ #endif /* WANT_INTERVALS */ #ifdef WIN32 static void error(char *pch) { if (!opterr) { return; /* without printing */ } fprintf(stderr, "%s: %s: %c\n", (NULL != program) ? program : "getopt", pch, optopt); } int getopt(int argc, char **argv, char *ostr) { static char *place = EMSG; /* option letter processing */ register char *oli; /* option letter list index */ if (!*place) { /* update scanning pointer */ if (optind >= argc || *(place = argv[optind]) != '-' || !*++place) { return EOF; } if (*place == '-') { /* found "--" */ ++optind; place = EMSG ; /* Added by shiva for Netperf */ return EOF; } } /* option letter okay? */ if ((optopt = (int)*place++) == (int)':' || !(oli = strchr(ostr, optopt))) { if (!*place) { ++optind; } error("illegal option"); return BADCH; } if (*++oli != ':') { /* don't need argument */ optarg = NULL; if (!*place) ++optind; } else { /* need an argument */ if (*place) { optarg = place; /* no white space */ } else if (argc <= ++optind) { /* no arg */ place = EMSG; error("option requires an argument"); return BADCH; } else { optarg = argv[optind]; /* white space */ } place = EMSG; ++optind; } return optopt; /* return option letter */ } #endif /* WIN32 */ /*---------------------------------------------------------------------------- * WIN32 implementation of perror, does not deal very well with WSA errors * The stdlib.h version of perror only deals with the ancient XENIX error codes. * * +*+SAF Why can't all WSA errors go through GetLastError? Most seem to... *--------------------------------------------------------------------------*/ #ifdef WIN32 void PrintWin32Error(FILE *stream, LPSTR text) { LPSTR szTemp; DWORD dwResult; DWORD dwError; dwError = GetLastError(); dwResult = FormatMessage( FORMAT_MESSAGE_ALLOCATE_BUFFER|FORMAT_MESSAGE_FROM_SYSTEM |FORMAT_MESSAGE_ARGUMENT_ARRAY, NULL, dwError, LANG_NEUTRAL, (LPTSTR)&szTemp, 0, NULL ); if (dwResult) fprintf(stream, "%s: %s\n", text, szTemp); else fprintf(stream, "%s: error 0x%x\n", text, dwError); fflush(stream); if (szTemp) LocalFree((HLOCAL)szTemp); } #endif /* WIN32 */ char * nsec_enabled_to_str(int enabled) { switch (enabled) { case NSEC_UNKNOWN: return("Unknown"); case NSEC_DISABLED: return("Disabled"); case NSEC_PERMISSIVE: return("Permissive"); case NSEC_ENFORCING: return("Enforcing"); default: return("UNKNOWN MODE"); } } char * nsec_type_to_str(int type) { switch (type) { case NSEC_TYPE_UNKNOWN: return("Unknown"); case NSEC_TYPE_SELINUX: return("SELinux"); default: return("UNKNOWN TYPE"); } } char * inet_ttos(int type) { switch (type) { case SOCK_DGRAM: return("SOCK_DGRAM"); break; case SOCK_STREAM: return("SOCK_STREAM"); break; #ifdef SOCK_DCCP case SOCK_DCCP: return("SOCK_DCCP"); #endif #ifdef SOCK_SEQPACKET case SOCK_SEQPACKET: return("SOCK_SEQPACKET"); #endif default: return("SOCK_UNKNOWN"); } } char unknown[32]; char * inet_ptos(int protocol) { switch (protocol) { case IPPROTO_TCP: return("IPPROTO_TCP"); break; case IPPROTO_UDP: return("IPPROTO_UDP"); break; #if defined(IPPROTO_SCTP) case IPPROTO_SCTP: return("IPPROTO_SCTP"); break; #endif #if defined(IPPROTO_DCCP) case IPPROTO_DCCP: return "IPPROTO_DCCP"; break; #endif #if defined(IPPROTO_UDPLITE) case IPPROTO_UDPLITE: return "IPPROTO_UDPLITE"; break; #endif default: snprintf(unknown,sizeof(unknown),"IPPROTO_UNKNOWN(%d)",protocol); return(unknown); } } /* one of these days, this should not be required */ #ifndef AF_INET_SDP #define AF_INET_SDP 27 #define PF_INET_SDP AF_INET_SDP #endif char * inet_ftos(int family) { switch(family) { case AF_INET: return("AF_INET"); #if defined(AF_INET6) case AF_INET6: return("AF_INET6"); #endif #if defined(AF_INET_SDP) case AF_INET_SDP: return("AF_INET_SDP"); #endif #if defined(AF_RDS) case AF_RDS: return("AF_RDS"); #endif default: return("AF_UNSPEC"); } } int inet_nton(int af, const void *src, char *dst, int cnt) { switch (af) { case AF_INET: /* magic constants again... :) */ if (cnt >= 4) { memcpy(dst,src,4); return 4; } else { Set_errno(ENOSPC); return(-1); } break; #if defined(AF_INET6) case AF_INET6: if (cnt >= 16) { memcpy(dst,src,16); return(16); } else { Set_errno(ENOSPC); return(-1); } break; #endif #if defined(AF_RDS) case AF_RDS: if (cnt >= 4) { memcpy(dst,src,4); return 4; } #endif default: Set_errno(EAFNOSUPPORT); return(-1); } } double ntohd(double net_double) { /* we rely on things being nicely packed */ union { double whole_thing; unsigned int words[2]; unsigned char bytes[8]; } conv_rec; unsigned char scratch; int i; /* on those systems where ntohl is a no-op, we want to return the original value, unchanged */ if (ntohl(1L) == 1L) { return(net_double); } conv_rec.whole_thing = net_double; /* we know that in the message passing routines that ntohl will have been called on the 32 bit quantities. we need to put those back the way they belong before we swap */ conv_rec.words[0] = htonl(conv_rec.words[0]); conv_rec.words[1] = htonl(conv_rec.words[1]); /* now swap */ for (i=0; i<= 3; i++) { scratch = conv_rec.bytes[i]; conv_rec.bytes[i] = conv_rec.bytes[7-i]; conv_rec.bytes[7-i] = scratch; } #if defined(__FLOAT_WORD_ORDER) && defined(__BYTE_ORDER) if (__FLOAT_WORD_ORDER != __BYTE_ORDER) { /* Fixup mixed endian floating point machines */ unsigned int scratch = conv_rec.words[0]; conv_rec.words[0] = conv_rec.words[1]; conv_rec.words[1] = scratch; } #endif return(conv_rec.whole_thing); } double htond(double host_double) { /* we rely on things being nicely packed */ union { double whole_thing; unsigned int words[2]; unsigned char bytes[8]; } conv_rec; unsigned char scratch; int i; /* on those systems where ntohl is a no-op, we want to return the original value, unchanged */ if (ntohl(1L) == 1L) { return(host_double); } conv_rec.whole_thing = host_double; /* now swap */ for (i=0; i<= 3; i++) { scratch = conv_rec.bytes[i]; conv_rec.bytes[i] = conv_rec.bytes[7-i]; conv_rec.bytes[7-i] = scratch; } #if defined(__FLOAT_WORD_ORDER) && defined(__BYTE_ORDER) if (__FLOAT_WORD_ORDER != __BYTE_ORDER) { /* Fixup mixed endian floating point machines */ unsigned int scratch = conv_rec.words[0]; conv_rec.words[0] = conv_rec.words[1]; conv_rec.words[1] = scratch; } #endif /* we know that in the message passing routines htonl will be called on the 32 bit quantities. we need to set things up so that when this happens, the proper order will go out on the network */ conv_rec.words[0] = htonl(conv_rec.words[0]); conv_rec.words[1] = htonl(conv_rec.words[1]); return(conv_rec.whole_thing); } /* The original patch from Google used lrand48, but I have been informed that is not easily available under Windows. So, rather than have some #ifdefs here I'll just simplistically replace lrand48 with rand(), which should be "good enough" at some point it may be sufficient to just call rand() directly rather than call this raj 20101130 */ unsigned int rand32(){ return (unsigned int)rand() * 2 + rand() % 2; } /* this routine will set the ip address of the sockaddr in the addrinfo to a random number in range, based on the address family. for grins, we will sanity check the value of mask_len against the address family. initial version from google, enhancements by raj 20101129 */ void random_ip_address(struct addrinfo *res, int mask_len) { switch(res->ai_family) { case AF_INET: { struct sockaddr_in *foo = (struct sockaddr_in *)res->ai_addr; unsigned int addr = ntohl(foo->sin_addr.s_addr); unsigned int mask = ((unsigned int)1 << (32 - mask_len)) - 1; if ((mask_len < 0) || (mask_len > 32)) { fprintf(where, "Mask length must be between 0 and 32 inclusive for AF_INET\n"); fflush(where); exit(-1); } addr = ntohl(foo->sin_addr.s_addr); do { addr = (addr & ~mask) | (rand32() & mask); } while ((addr & 0xff) == 0xff); foo->sin_addr.s_addr = htonl(addr); break; } #if defined(AF_INET6) case AF_INET6: { struct sockaddr_in6 *foo = (struct sockaddr_in6 *)res->ai_addr; unsigned int i, len; unsigned int *addr = (unsigned int *)&(foo->sin6_addr.s6_addr); unsigned int mask; if ((mask_len < 0) || (mask_len > 128)) { fprintf(where, "Mask length must be between 0 and 128 inclusive for AF_INET\n"); fflush(where); exit(-1); } for (i = 0; i < 4; i ++){ addr[i] = ntohl(addr[i]); len = mask_len - i * 32; len = ((len < 32) ? len : 32); len = ((len > 0) ? len : 0); mask = ((unsigned int)1 << (32 - len)) - 1; addr[i] = (addr[i] & ~mask) | (rand32() & mask); addr[i] = htonl(addr[i]); } break; } #endif default: fprintf(where, "Unexpected Address Family of %u\n",res->ai_family); fflush(where); exit(-1); } } #if defined(HAVE_SENDFILE) int netperf_sendfile(SOCKET send_socket, struct ring_elt *send_ring) { int len; int ret = 0; #if defined(__linux) || defined(__sun) off_t scratch_offset; /* the linux sendfile() call will update the offset variable, which is something we do _not_ want to happen to the value in the send_ring! so, we have to use a scratch variable. */ #endif /* __linux || defined(__sun) */ #if defined (__sun) size_t scratch_len; /* the sun sendfilev() needs a place to tell us how many bytes were written, even though it also returns the value */ sendfilevec_t sv; #endif /* __sun */ /* you can look at netlib.h for a description of the fields we are passing to sendfile(). 08/2000 */ #if defined(__linux) scratch_offset = send_ring->offset; len=sendfile(send_socket, send_ring->fildes, &scratch_offset, /* modified after the call! */ send_ring->length); #elif defined (__sun) /* We must call with SFV_NOWAIT and a large file size (>= 16MB) to get zero-copy, as well as compiling with -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64 */ sv.sfv_fd = send_ring->fildes; sv.sfv_flag = SFV_NOWAIT; sv.sfv_off = send_ring->offset; sv.sfv_len = send_ring->length; len = sendfilev(send_socket, &sv, 1, &scratch_len); #elif defined(__FreeBSD__) /* so close to HP-UX and yet so far away... :) */ ret = sendfile(send_ring->fildes, send_socket, send_ring->offset, send_ring->length, NULL, (off_t *)&len, send_ring->flags); #elif defined(USE_OSX) len = send_ring->length; ret = sendfile(send_ring->fildes, send_socket, send_ring->offset, (off_t *)&len, NULL, send_ring->flags); #else /* original sendile HP-UX */ len=sendfile(send_socket, send_ring->fildes, send_ring->offset, send_ring->length, send_ring->hdtrl, send_ring->flags); #endif /* for OSX and FreeBSD, a non-zero ret means something failed. I would hope that the length fields are set to -1 or the like, but at the moment I do not know I can count on that. for other platforms, ret will be set to zero and we can rely directly on len. raj 2013-05-01 */ if (ret != 0) return -1; else return len; } #endif /* one of these days, this should be abstracted-out just like the CPU util stuff. raj 2005-01-27 */ int get_num_cpus() { /* on HP-UX, even when we use the looper procs we need the pstat */ /* call */ int temp_cpus; #ifdef __hpux #include struct pst_dynamic psd; if (pstat_getdynamic((struct pst_dynamic *)&psd, (size_t)sizeof(psd), (size_t)1, 0) != -1) { temp_cpus = psd.psd_proc_cnt; } else { temp_cpus = 1; } #else /* MW: was included for non-Windows systems above. */ /* Thus if _SC_NPROC_ONLN is defined, we should be able to use sysconf. */ #ifdef _SC_NPROCESSORS_ONLN temp_cpus = sysconf(_SC_NPROCESSORS_ONLN); #ifdef USE_PERFSTAT temp_cpus = perfstat_cpu(NULL,NULL, sizeof(perfstat_cpu_t), 0); #endif /* USE_PERFSTAT */ #else /* no _SC_NPROCESSORS_ONLN */ #ifdef WIN32 SYSTEM_INFO SystemInfo; GetSystemInfo(&SystemInfo); temp_cpus = SystemInfo.dwNumberOfProcessors; #else /* we need to know some other ways to do this, or just fall-back on a global command line option - raj 4/95 */ temp_cpus = shell_num_cpus; #endif /* WIN32 */ #endif /* _SC_NPROCESSORS_ONLN */ #endif /* __hpux */ if (temp_cpus > MAXCPUS) { fprintf(where, "Sorry, this system has more CPUs (%d) than I can handle (%d).\n" "Please alter MAXCPUS in netlib.h and recompile.\n", temp_cpus, MAXCPUS); fflush(where); exit(1); } return(temp_cpus); } #ifdef WIN32 #ifdef __GNUC__ #define S64_SUFFIX(x) x##LL #else #define S64_SUFFIX(x) x##i64 #endif /* * Number of 100 nanosecond units from 1/1/1601 to 1/1/1970 */ #define EPOCH_BIAS S64_SUFFIX(116444736000000000) /* * Union to facilitate converting from FILETIME to unsigned __int64 */ typedef union { unsigned __int64 ft_scalar; FILETIME ft_struct; } FT; void gettimeofday( struct timeval *tv , struct timezone *not_used ) { FT nt_time; __int64 UnixTime; /* microseconds since 1/1/1970 */ GetSystemTimeAsFileTime( &(nt_time.ft_struct) ); UnixTime = ((nt_time.ft_scalar - EPOCH_BIAS) / S64_SUFFIX(10)); tv->tv_sec = (long)(time_t)(UnixTime / S64_SUFFIX(1000000)); tv->tv_usec = (unsigned long)(UnixTime % S64_SUFFIX(1000000)); } #endif /* WIN32 */ /* this routine will disable any running timer */ void stop_timer() { #ifndef WIN32 alarm(0); #else /* at some point we may need some win32 equivalent */ if (hAlarm != (HANDLE) INVALID_HANDLE_VALUE) { SetEvent(hAlarm); } #endif /* WIN32 */ } /************************************************************************/ /* */ /* signal catcher */ /* */ /************************************************************************/ #ifndef WIN32 void #if defined(__hpux) catcher(sig, code, scp) int sig; int code; struct sigcontext *scp; #else catcher(int sig) #endif /* __hpux || __VMS */ { #ifdef __hpux if (debug > 2) { fprintf(where,"caught signal %d ",sig); if (scp) { fprintf(where,"while in syscall %d\n", scp->sc_syscall); } else { fprintf(where,"null scp\n"); } fflush(where); } #endif /* RAJ_DEBUG */ switch(sig) { case SIGINT: times_up = 1; break; case SIGALRM: if (--test_len_ticks == 0) { /* the test is over */ if (times_up != 0) { fprintf(where,"catcher: timer popped with times_up != 0\n"); fflush(where); } times_up = 1; #if defined(WANT_INTERVALS) && !defined(WANT_SPIN) stop_itimer(); /* we should also stop the normal test timer lest it fire at an inopportune moment - we do not know if we got here off the interval timer or the test timer... */ stop_timer(); #endif /* WANT_INTERVALS */ break; } else { #ifdef WANT_INTERVALS #ifdef __hpux /* the test is not over yet and we must have been using the interval timer. if we were in SYS_SIGSUSPEND we want to re-start the system call. Otherwise, we want to get out of the sigsuspend call. I NEED TO KNOW HOW TO DO THIS FOR OTHER OPERATING SYSTEMS. If you know how, please let me know. rick jones */ if (scp->sc_syscall != SYS_SIGSUSPEND) { if (debug > 2) { fprintf(where, "catcher: Time to send burst > interval!\n"); fflush(where); } scp->sc_syscall_action = SIG_RESTART; } #endif /* __hpux */ #else /* WANT_INTERVALS */ fprintf(where, "catcher: interval timer running unexpectedly!\n"); fflush(where); times_up = 1; #endif /* WANT_INTERVALS */ break; } } return; } #endif /* WIN32 */ void install_signal_catchers() { /* just a simple little routine to catch a bunch of signals */ #ifndef WIN32 struct sigaction action; int i; fprintf(where,"installing catcher for all signals\n"); fflush(where); sigemptyset(&(action.sa_mask)); action.sa_handler = catcher; #ifdef SA_INTERRUPT action.sa_flags = SA_INTERRUPT; #else /* SA_INTERRUPT */ action.sa_flags = 0; #endif /* SA_INTERRUPT */ for (i = 1; i <= NSIG; i++) { switch (i) { case SIGALRM: case SIGPROF: case SIGSTOP: case SIGKILL: break; default: if (sigaction(i,&action,NULL) != 0) { fprintf(where, "Could not install signal catcher for sig %d, errno %d\n", i, errno); fflush(where); } } } #else return; #endif /* WIN32 */ } #ifdef WIN32 #define SIGALRM (14) void emulate_alarm( int seconds ) { DWORD ErrorCode; DWORD HandlesClosedFlags = 0; /* Wait on this event for parm seconds. */ ErrorCode = WaitForSingleObject(hAlarm, seconds*1000); if (ErrorCode == WAIT_FAILED) { perror("WaitForSingleObject failed"); exit(1); } if (ErrorCode == WAIT_TIMEOUT) { /* WaitForSingleObject timed out; this means the timer wasn't canceled. */ times_up = 1; /* Give the other threads time to notice that times_up has changed state before taking the harsh step of closing the sockets. */ timed_out=0; if (WaitForSingleObject(hAlarm, PAD_TIME/2*1000) == WAIT_TIMEOUT) { timed_out=1; /* We have yet to find a good way to fully emulate the effects of signals and getting EINTR from system calls under winsock, so what we do here is close the socket out from under the other thread. It is rather kludgy, but should be sufficient to get this puppy shipped. The concept can be attributed/blamed :) on Robin raj 1/96 */ if (win_kludge_socket != INVALID_SOCKET) { HandlesClosedFlags |= 1; closesocket(win_kludge_socket); } if (win_kludge_socket2 != INVALID_SOCKET) { HandlesClosedFlags |= 2; closesocket(win_kludge_socket2); } } if(debug) { fprintf(where, "emulate_alarm - HandlesClosedFlags: %x\n", HandlesClosedFlags); fflush(where); } } } #endif /* WIN32 */ void start_timer(int time) { #ifdef WIN32 /*+*+SAF What if StartTimer is called twice without the first timer */ /*+*+SAF expiring? */ DWORD thread_id ; HANDLE tHandle; if (hAlarm == (HANDLE) INVALID_HANDLE_VALUE) { /* Create the Alarm event object */ hAlarm = CreateEvent( (LPSECURITY_ATTRIBUTES) NULL, /* no security */ FALSE, /* auto reset event */ FALSE, /* init. state = reset */ (void *)NULL); /* unnamed event object */ if (hAlarm == (HANDLE) INVALID_HANDLE_VALUE) { perror("CreateEvent failure"); exit(1); } } else { ResetEvent(hAlarm); } tHandle = CreateThread(0, 0, (LPTHREAD_START_ROUTINE)emulate_alarm, (LPVOID)(ULONG_PTR)time, 0, &thread_id ) ; CloseHandle(tHandle); #else /* not WIN32 */ struct sigaction action; int ret; if (debug) { fprintf(where,"About to start a timer for %d seconds.\n",time); fflush(where); } action.sa_handler = catcher; #ifdef SA_INTERRUPT /* on some systems (SunOS 4.blah), system calls are restarted. we do */ /* not want that */ action.sa_flags = SA_INTERRUPT; #else /* SA_INTERRUPT */ action.sa_flags = 0; #endif /* SA_INTERRUPT */ sigemptyset(&(action.sa_mask)); sigaddset(&(action.sa_mask),SIGALRM); if (sigaction(SIGALRM, &action, NULL) < 0) { fprintf(where, "start_timer: error installing alarm handler errno %d\n", errno); fflush(where); exit(-1); } sigemptyset(&(action.sa_mask)); sigaddset(&(action.sa_mask),SIGINT); if (sigaction(SIGINT, &action, NULL) < 0) { fprintf(where, "start_timer: error installing SIGINT handler errno %d\n", errno); fflush(where); exit(-1); } /* this is the easy case - just set the timer for so many seconds */ ret = alarm(time); if (ret != 0) { fprintf(where, "error starting alarm timer, ret %d errno %d\n", ret, errno); fflush(where); exit(-1); } #endif /* WIN32 */ test_len_ticks = 1; } #ifdef WANT_INTERVALS /* this routine will enable the interval timer and set things up so that for a timed test the test will end at the proper time. it should detect the presence of POSIX.4 timer_* routines one of these days */ void start_itimer(unsigned int interval_len_msec ) { #ifdef WIN32 LARGE_INTEGER liDueTime; TIMECAPS ptc; MMRESULT mmr; /* make sure timer resolution is at least as small as interval length */ timerRes=interval_len_msec; mmr=timeGetDevCaps(&ptc, sizeof (ptc)); if (mmr==TIMERR_NOERROR){ if (interval_len_msec 0) { /* this was a timed test */ test_len_ticks = (test_time * 1000000) / usec_per_itvl; } else { /* this was not a timed test, use MAXINT */ test_len_ticks = INT_MAX; } if (debug) { fprintf(where, "setting the interval timer to %d sec %d usec test len %d ticks\n", usec_per_itvl / 1000000, usec_per_itvl % 1000000, test_len_ticks); fflush(where); } /* if this was not a timed test, then we really aught to enable the signal catcher raj 2/95 */ new_interval.it_interval.tv_sec = usec_per_itvl / 1000000; new_interval.it_interval.tv_usec = usec_per_itvl % 1000000; new_interval.it_value.tv_sec = usec_per_itvl / 1000000; new_interval.it_value.tv_usec = usec_per_itvl % 1000000; if (setitimer(ITIMER_REAL,&new_interval,&old_interval) != 0) { /* there was a problem arming the interval timer */ perror("netperf: setitimer"); exit(1); } #endif /* WIN32*/ } #endif /* WANT_INTERVALS */ void netlib_init_cpu_map() { int i; #ifdef HAVE_MPCTL int num; i = 0; /* I go back and forth on whether this should be the system-wide set of calls, or if the processor set versions (sans the _SYS) should be used. at the moment I believe that the system-wide version should be used. raj 2006-04-03 */ num = mpctl(MPC_GETNUMSPUS_SYS,0,0); lib_cpu_map[i] = mpctl(MPC_GETFIRSTSPU_SYS,0,0); for (i = 1;((i < num) && (i < MAXCPUS)); i++) { lib_cpu_map[i] = mpctl(MPC_GETNEXTSPU_SYS,lib_cpu_map[i-1],0); } /* from here, we set them all to -1 because if we launch more loopers than actual CPUs, well, I'm not sure why :) */ for (; i < MAXCPUS; i++) { lib_cpu_map[i] = -1; } #else /* we assume that there is indeed a contiguous mapping */ for (i = 0; i < MAXCPUS; i++) { lib_cpu_map[i] = i; } #endif } /****************************************************************/ /* */ /* netlib_init() */ /* */ /* initialize the performance library... */ /* */ /****************************************************************/ void netlib_init() { int i; where = stdout; request_array = (int *)(&netperf_request); response_array = (int *)(&netperf_response); for (i = 0; i < MAXCPUS; i++) { lib_local_per_cpu_util[i] = -1.0; } lib_local_cpu_stats.peak_cpu_id = -1; lib_local_cpu_stats.peak_cpu_util = -1.0; lib_remote_cpu_stats.peak_cpu_id = -1; lib_remote_cpu_stats.peak_cpu_util = -1.0; netperf_version = strdup(NETPERF_VERSION); /* on those systems where we know that CPU numbers may not start at zero and be contiguous, we provide a way to map from a contiguous, starting from 0 CPU id space to the actual CPU ids. at present this is only used for the netcpu_looper stuff because we ass-u-me that someone setting processor affinity from the netperf commandline will provide a "proper" CPU identifier. raj 2006-04-03 */ netlib_init_cpu_map(); if (debug) { fprintf(where, "netlib_init: request_array at %p\n" "netlib_init: response_array at %p\n", request_array, response_array); fflush(where); } /* some functionality might want to use random numbers, so we should initialize the random number generator */ srand(getpid()); } /* this routine will conver the string into an unsigned integer. it is used primarily for the command-line options taking a number (such as the socket size) which could be rather large. If someone enters 32M, then the number will be converted to 32 * 1024 * 1024. If they inter 32m, the number will be converted to 32 * 1000 * 1000 */ unsigned int convert(char *string) { unsigned int base; base = atoi(string); if (strstr(string,"K")) { base *= 1024; } if (strstr(string,"M")) { base *= (1024 * 1024); } if (strstr(string,"G")) { base *= (1024 * 1024 * 1024); } if (strstr(string,"k")) { base *= (1000); } if (strstr(string,"m")) { base *= (1000 * 1000); } if (strstr(string,"g")) { base *= (1000 * 1000 * 1000); } return(base); } /* this routine is like convert, but it is used for an interval time specification instead of stuff like socket buffer or send sizes. it converts everything to microseconds for internal use. if there is an 'm' at the end it assumes the user provided milliseconds, s will imply seconds, u will imply microseconds. in the future n will imply nanoseconds but for now it will be ignored. if there is no suffix or an unrecognized suffix, it will be assumed the user provided milliseconds, which was the long-time netperf default. one of these days, we should probably revisit that nanosecond business wrt the return value being just an int rather than a uint64_t or something. raj 2006-02-06 */ unsigned int convert_timespec(char *string) { unsigned int base; base = atoi(string); if (strstr(string,"m")) { base *= 1000; } else if (strstr(string,"u")) { base *= (1); } else if (strstr(string,"s")) { base *= (1000 * 1000); } else { base *= (1000); } return(base); } /* this routine will allocate a circular list of buffers for either send or receive operations. each of these buffers will be aligned and offset as per the users request. the circumference of this ring will be controlled by the setting of width. the buffers will be filled with data from the file specified in fill_file. if fill_file is an empty string, the buffers will be filled from "default_fill" which will be "netperf" so anyone sniffing the traffic will have a better idea what this traffic happens to be. */ struct ring_elt * allocate_buffer_ring(int width, int buffer_size, int alignment, int offset) { struct ring_elt *first_link = NULL; struct ring_elt *temp_link = NULL; struct ring_elt *prev_link; int i; int malloc_size; int bytes_left; int bytes_read; int do_fill; FILE *fill_source; char default_fill[] = "netperf"; int fill_cursor = 0; malloc_size = buffer_size + alignment + offset; /* did the user wish to have the buffers pre-filled with data from a */ /* particular source? */ if (strcmp(local_fill_file,"") == 0) { do_fill = 0; fill_source = NULL; } else { do_fill = 1; fill_source = (FILE *)fopen(local_fill_file,"r"); if (fill_source == (FILE *)NULL) { fprintf(where,"Could not open requested fill file: %s\n", strerror(errno)); fflush(where); } } assert(width >= 1); prev_link = NULL; for (i = 1; i <= width; i++) { /* get the ring element */ temp_link = (struct ring_elt *)malloc(sizeof(struct ring_elt)); if (temp_link == NULL) { fprintf(where, "malloc(%u) failed!\n", (unsigned int)sizeof(struct ring_elt)); exit(-1); } temp_link->completion_ptr = NULL; /* remember the first one so we can close the ring at the end */ if (i == 1) { first_link = temp_link; } temp_link->buffer_base = (char *)malloc(malloc_size); if (temp_link->buffer_base == NULL) { fprintf(where, "malloc(%d) failed!\n", malloc_size); exit(-1); } #ifndef WIN32 temp_link->buffer_ptr = (char *)(( (long)(temp_link->buffer_base) + (long)alignment - 1) & ~((long)alignment - 1)); #else temp_link->buffer_ptr = (char *)(( (ULONG_PTR)(temp_link->buffer_base) + (ULONG_PTR)alignment - 1) & ~((ULONG_PTR)alignment - 1)); #endif temp_link->buffer_ptr += offset; /* is where the buffer fill code goes. */ if (do_fill) { char *bufptr = temp_link->buffer_ptr; bytes_left = buffer_size; while (bytes_left) { if (((bytes_read = (int)fread(bufptr, 1, bytes_left, fill_source)) == 0) && (feof(fill_source))){ rewind(fill_source); } bufptr += bytes_read; bytes_left -= bytes_read; } } else { /* use the default fill to ID our data traffic on the network. it ain't exactly pretty, but it should work */ int j; char *bufptr = temp_link->buffer_ptr; for (j = 0; j < buffer_size; j++) { bufptr[j] = default_fill[fill_cursor]; fill_cursor += 1; /* the Windows DDK compiler with an x86_64 target wants a cast here */ if (fill_cursor > (int)strlen(default_fill)) { fill_cursor = 0; } } } temp_link->next = prev_link; prev_link = temp_link; } if (first_link) { /* SAF Prefast made me do it... */ first_link->next = temp_link; } return(first_link); /* it's a circle, doesn't matter which we return */ } /* this routine will dirty the first dirty_count bytes of the specified buffer and/or read clean_count bytes from the buffer. it will go N bytes at a time, the only question is how large should N be and if we should be going continguously, or based on some assumption of cache line size */ void access_buffer(char *buffer_ptr,int length, int dirty_count, int clean_count) { char *temp_buffer; char *limit; int i, dirty_totals; temp_buffer = buffer_ptr; limit = temp_buffer + length; dirty_totals = 0; for (i = 0; ((i < dirty_count) && (temp_buffer < limit)); i++) { *temp_buffer += (char)i; dirty_totals += *temp_buffer; temp_buffer++; } for (i = 0; ((i < clean_count) && (temp_buffer < limit)); i++) { dirty_totals += *temp_buffer; temp_buffer++; } if (debug > 100) { fprintf(where, "This was here to try to avoid dead-code elimination %d\n", dirty_totals); fflush(where); } } #ifdef HAVE_ICSC_EXS #include #include /* this routine will allocate a circular list of buffers for either send or receive operations. each of these buffers will be aligned and offset as per the users request. the circumference of this ring will be controlled by the setting of send_width. the buffers will be filled with data from the file specified in local_fill_file. if local_fill_file is an empty string, the buffers will not be filled with any particular data */ struct ring_elt * allocate_exs_buffer_ring (int width, int buffer_size, int alignment, int offset, exs_mhandle_t *mhandlep) { struct ring_elt *first_link; struct ring_elt *temp_link; struct ring_elt *prev_link; int i; int malloc_size; int bytes_left; int bytes_read; int do_fill; FILE *fill_source; int mmap_size; char *mmap_buffer, *mmap_buffer_aligned; malloc_size = buffer_size + alignment + offset; /* did the user wish to have the buffers pre-filled with data from a */ /* particular source? */ if (strcmp (local_fill_file, "") == 0) { do_fill = 0; fill_source = NULL; } else { do_fill = 1; fill_source = (FILE *) fopen (local_fill_file, "r"); if (fill_source == (FILE *) NULL) { perror ("Could not open requested fill file"); exit (1); } } assert (width >= 1); if (debug) { fprintf (where, "allocate_exs_buffer_ring: " "width=%d buffer_size=%d alignment=%d offset=%d\n", width, buffer_size, alignment, offset); } /* allocate shared memory */ mmap_size = width * malloc_size; mmap_buffer = (char *) mmap ((caddr_t)NULL, mmap_size+NBPG-1, PROT_READ|PROT_WRITE, MAP_SHARED|MAP_ANONYMOUS, -1, 0); if (mmap_buffer == NULL) { perror ("allocate_exs_buffer_ring: mmap failed"); exit (1); } mmap_buffer_aligned = (char *) ((uintptr_t)mmap_buffer & ~(NBPG-1)); if (debug) { fprintf (where, "allocate_exs_buffer_ring: " "mmap buffer size=%d address=0x%p aligned=0x%p\n", mmap_size, mmap_buffer, mmap_buffer_aligned); } /* register shared memory */ *mhandlep = exs_mregister ((void *)mmap_buffer_aligned, (size_t)mmap_size, 0); if (*mhandlep == EXS_MHANDLE_INVALID) { perror ("allocate_exs_buffer_ring: exs_mregister failed"); exit (1); } if (debug) { fprintf (where, "allocate_exs_buffer_ring: mhandle=%d\n", *mhandlep); } /* allocate ring elements */ first_link = (struct ring_elt *) malloc (width * sizeof (struct ring_elt)); if (first_link == NULL) { printf ("malloc(%d) failed!\n", width * sizeof (struct ring_elt)); exit (1); } /* initialize buffer ring */ prev_link = first_link + width - 1; for (i = 0, temp_link = first_link; i < width; i++, temp_link++) { temp_link->buffer_base = (char *) mmap_buffer_aligned + (i*malloc_size); #ifndef WIN32 temp_link->buffer_ptr = (char *) (((long)temp_link->buffer_base + (long)alignment - 1) & ~((long)alignment - 1)); #else temp_link->buffer_ptr = (char *) (((ULONG_PTR)temp_link->buffer_base + (ULONG_PTR)alignment - 1) & ~((ULONG_PTR)alignment - 1)); #endif temp_link->buffer_ptr += offset; if (debug) { fprintf (where, "allocate_exs_buffer_ring: " "buffer: index=%d base=0x%p ptr=0x%p\n", i, temp_link->buffer_base, temp_link->buffer_ptr); } /* is where the buffer fill code goes. */ if (do_fill) { bytes_left = buffer_size; while (bytes_left) { if (((bytes_read = (int) fread (temp_link->buffer_ptr, 1, bytes_left, fill_source)) == 0) && (feof (fill_source))) { rewind (fill_source); } bytes_left -= bytes_read; } } /* do linking */ prev_link->next = temp_link; prev_link = temp_link; } return (first_link); /* it is a circle, doesn't matter which we return */ } #endif /* HAVE_ICSC_EXS */ #ifdef HAVE_SENDFILE /* this routine will construct a ring of sendfile_ring_elt structs that the routine sendfile_tcp_stream() will use to get parameters to its calls to sendfile(). It will setup the ring to point at the file specified in the global -F option that is already used to pre-fill buffers in the send() case. 08/2000 if there is no file specified in a global -F option, we will create a tempoarary file and fill it with random data and use that instead. raj 2007-08-09 */ struct ring_elt * alloc_sendfile_buf_ring(int width, int buffer_size, int alignment, int offset) { struct ring_elt *first_link = NULL; struct ring_elt *temp_link = NULL; struct ring_elt *prev_link; int i; int fildes; struct stat statbuf; /* if the user has not specified a file with the -F option, we will fail the test. otherwise, go ahead and try to open the file. 08/2000 */ if (strcmp(local_fill_file,"") == 0) { /* use an temp file for the fill file */ char temp_file[] = {"netperfXXXXXX\0"}; int *temp_buffer; /* make sure we have at least an ints worth, even if the user is using an insane buffer size for a sendfile test. we are ass-u-me-ing that malloc will return something at least aligned on an int boundary... */ temp_buffer = (int *) malloc(buffer_size + sizeof(int)); if (temp_buffer) { /* ok, we have the buffer we are going to write, lets get a temporary filename */ fildes = mkstemp(temp_file); /* no need to call open because mkstemp did it */ if (-1 != fildes) { int count; int *int_ptr; /* we initialize the random number generator in netlib_init() now. raj 20110111 */ /* unlink the file so it goes poof when we exit. unless/until shown to be a problem we will blissfully ignore the return value. raj 2007-08-09 */ unlink(temp_file); /* now fill-out the file with at least buffer_size * width bytes */ for (count = 0; count < width; count++) { /* fill the buffer with random data. it doesn't have to be really random, just "random enough" :) we do this here rather than up above because we want each write to the file to be different random data */ int_ptr = temp_buffer; for (i = 0; i <= buffer_size/sizeof(int); i++) { *int_ptr = rand(); int_ptr++; } if (write(fildes,temp_buffer,buffer_size+sizeof(int)) != buffer_size + sizeof(int)) { perror("allocate_sendfile_buf_ring: incomplete write"); exit(-1); } } } else { perror("alloc_sendfile_buf_ring: could not allocate temp name"); exit(-1); } } else { perror("alloc_sendfile_buf_ring: could not allocate buffer for file"); exit(-1); } } else { /* the user pointed us at a file, so try it */ fildes = open(local_fill_file , O_RDONLY); if (fildes == -1){ perror("alloc_sendfile_buf_ring: Could not open requested file"); exit(1); } /* make sure there is enough file there to allow us to make a complete ring. that way we do not need additional logic in the ring setup to deal with wrap-around issues. we might want that someday, but not just now. 08/2000 */ if (stat(local_fill_file,&statbuf) != 0) { perror("alloc_sendfile_buf_ring: could not stat file"); exit(1); } if (statbuf.st_size < (width * buffer_size)) { /* the file is too short */ fprintf(stderr, "alloc_sendfile_buf_ring: specified file too small.\n" "file must be larger than send_width * send_size\n"); fflush(stderr); exit(1); } } /* so, at this point we know that fildes is a descriptor which references a file of sufficient size for our nefarious porpoises. raj 2007-08-09 */ prev_link = NULL; for (i = 1; i <= width; i++) { /* get the ring element. we should probably make sure the malloc() was successful, but for now we'll just let the code bomb mysteriously. 08/2000 */ temp_link = (struct ring_elt *) malloc(sizeof(struct ring_elt)); if (temp_link == NULL) { fprintf(where, "malloc(%u) failed!\n", (unsigned int) sizeof(struct ring_elt)); exit(1); } /* remember the first one so we can close the ring at the end */ if (i == 1) { first_link = temp_link; } /* now fill-in the fields of the structure with the apropriate stuff. just how should we deal with alignment and offset I wonder? until something better comes-up, I think we will just ignore them. 08/2000 */ temp_link->fildes = fildes; /* from which file do we send? */ temp_link->offset = offset; /* starting at which offset? */ offset += buffer_size; /* get ready for the next elt */ temp_link->length = buffer_size; /* how many bytes to send */ temp_link->hdtrl = NULL; /* no header or trailer */ temp_link->flags = 0; /* no flags */ /* is where the buffer fill code went. */ temp_link->next = prev_link; prev_link = temp_link; } /* close the ring */ first_link->next = temp_link; return(first_link); /* it's a dummy ring */ } #endif /* HAVE_SENDFILE */ /***********************************************************************/ /* */ /* dump_request() */ /* */ /* display the contents of the request array to the user. it will */ /* display the contents in decimal, hex, and ascii, with four bytes */ /* per line. */ /* */ /***********************************************************************/ void dump_request() { int counter = 0; fprintf(where,"request contents:\n"); for (counter = 0; counter < ((sizeof(netperf_request)/4)-3); counter += 4) { fprintf(where,"%d:\t%8x %8x %8x %8x \t|%4.4s| |%4.4s| |%4.4s| |%4.4s|\n", counter, request_array[counter], request_array[counter+1], request_array[counter+2], request_array[counter+3], (char *)&request_array[counter], (char *)&request_array[counter+1], (char *)&request_array[counter+2], (char *)&request_array[counter+3]); } fflush(where); } /***********************************************************************/ /* */ /* dump_response() */ /* */ /* display the content of the response array to the user. it will */ /* display the contents in decimal, hex, and ascii, with four bytes */ /* per line. */ /* */ /***********************************************************************/ void dump_response() { int counter = 0; fprintf(where,"response contents\n"); for (counter = 0; counter < ((sizeof(netperf_response)/4)-3); counter += 4) { fprintf(where,"%d:\t%8x %8x %8x %8x \t>%4.4s< >%4.4s< >%4.4s< >%4.4s<\n", counter, response_array[counter], response_array[counter+1], response_array[counter+2], response_array[counter+3], (char *)&response_array[counter], (char *)&response_array[counter+1], (char *)&response_array[counter+2], (char *)&response_array[counter+3]); } fflush(where); } /* format_number() return a pointer to a formatted string containing the value passed translated into the units specified. It assumes that the base units are bytes. If the format calls for bits, it will use SI units (10^) if the format calls for bytes, it will use CS units (2^)... This routine should look familiar to uses of the latest ttcp... we would like to use "t" or "T" for transactions, but probably should leave those for terabits and terabytes respectively, so for transactions, we will use "x" which will, by default, do absolutely nothing to the result. why? so we don't have to special case code elsewhere such as in the TCP_RR-as-bidirectional test case. */ char * format_number(double number) { static char fmtbuf[64]; switch (libfmt) { case 'B': snprintf(fmtbuf, sizeof(fmtbuf), "%-7.2f" , number); break; case 'K': snprintf(fmtbuf, sizeof(fmtbuf), "%-7.2f" , number / 1024.0); break; case 'M': snprintf(fmtbuf, sizeof(fmtbuf), "%-7.2f", number / 1024.0 / 1024.0); break; case 'G': snprintf(fmtbuf, sizeof(fmtbuf), "%-7.2f", number / 1024.0 / 1024.0 / 1024.0); break; case 'b': snprintf(fmtbuf, sizeof(fmtbuf), "%-7.2f" , number * 8); break; case 'k': snprintf(fmtbuf, sizeof(fmtbuf), "%-7.2f", number * 8 / 1000.0); break; case 'm': snprintf(fmtbuf, sizeof(fmtbuf), "%-7.2f", number * 8 / 1000.0 / 1000.0); break; case 'g': snprintf(fmtbuf, sizeof(fmtbuf), "%-7.2f", number * 8 / 1000.0 / 1000.0 / 1000.0); break; case 'x': snprintf(fmtbuf, sizeof(fmtbuf), "%-7.2f", number); break; default: snprintf(fmtbuf, sizeof(fmtbuf), "%-7.2f", number / 1024.0); } return fmtbuf; } char format_cpu_method(int method) { char method_char; switch (method) { case CPU_UNKNOWN: method_char = 'U'; break; case HP_IDLE_COUNTER: method_char = 'I'; break; case PSTAT: method_char = 'P'; break; case KSTAT: method_char = 'K'; break; case KSTAT_10: method_char = 'M'; break; case PERFSTAT: method_char = 'E'; break; case TIMES: /* historical only, completely unsuitable for netperf's purposes */ method_char = 'T'; break; case GETRUSAGE: /* historical only, completely unsuitable for netperf;s purposes */ method_char = 'R'; break; case LOOPER: method_char = 'L'; break; case NT_METHOD: method_char = 'N'; break; case PROC_STAT: method_char = 'S'; break; case SYSCTL: method_char = 'C'; break; case OSX: method_char = 'O'; break; default: method_char = '?'; } return method_char; } char * format_units() { static char unitbuf[64]; switch (libfmt) { case 'B': strcpy(unitbuf, "Bytes"); break; case 'K': strcpy(unitbuf, "KBytes"); break; case 'M': strcpy(unitbuf, "MBytes"); break; case 'G': strcpy(unitbuf, "GBytes"); break; case 'b': strcpy(unitbuf, "10^0bits"); break; case 'k': strcpy(unitbuf, "10^3bits"); break; case 'm': strcpy(unitbuf, "10^6bits"); break; case 'g': strcpy(unitbuf, "10^9bits"); break; case 'x': strcpy(unitbuf, "Trans"); break; case 'u': strcpy(unitbuf,"Usec"); break; default: strcpy(unitbuf, "KBytes"); } return unitbuf; } /****************************************************************/ /* */ /* shutdown_control() */ /* */ /* tear-down the control connection between me and the server. */ /****************************************************************/ void shutdown_control() { char *buf = (char *)&netperf_response; int buflen = sizeof(netperf_response); /* stuff for select, use fd_set for better compliance */ fd_set readfds; struct timeval timeout; if (debug) { fprintf(where, "shutdown_control: shutdown of control connection requested.\n"); fflush(where); } /* first, we say that we will be sending no more data on the */ /* connection */ if (shutdown(netlib_control,1) == SOCKET_ERROR) { Print_errno(where, "shutdown_control: error in shutdown"); fflush(where); exit(1); } /* Now, we hang on a select waiting for the socket to become readable to receive the shutdown indication from the remote. this will be "just" like the recv_response() code we only select once. it is assumed that if the response is split (which should not be happening, that we will receive the whole thing and not have a problem ;-) */ FD_ZERO(&readfds); FD_SET(netlib_control,&readfds); timeout.tv_sec = 60; /* wait one minute then punt */ timeout.tv_usec = 0; /* select had better return one, or there was either a problem or a timeout... */ if (select(FD_SETSIZE, &readfds, 0, 0, &timeout) != 1) { Print_errno(where, "shutdown_control: no response received"); fflush(where); exit(1); } /* we now assume that the socket has come ready for reading */ recv(netlib_control, buf, buflen,0); } /* bind_to_specific_processor will bind the calling process to the processor in "processor" It has lots of ugly ifdefs to deal with all the different ways systems do processor affinity. this is a generalization of work initially done by stephen burger. raj 2004/12/13 */ void bind_to_specific_processor(int processor_affinity, int use_cpu_map) { int mapped_affinity; /* this is in place because the netcpu_looper processor affinity ass-u-me-s a contiguous CPU id space starting with 0. for the regular netperf/netserver affinity, we ass-u-me the user has used a suitable CPU id even when the space is not contiguous and starting from zero */ if (use_cpu_map) { mapped_affinity = lib_cpu_map[processor_affinity]; } else { mapped_affinity = processor_affinity; } #ifdef HAVE_MPCTL /* indeed, at some point it would be a good idea to check the return status and pass-along notification of error... raj 2004/12/13 */ mpctl(MPC_SETPROCESS_FORCE, mapped_affinity, getpid()); #elif HAVE_PROCESSOR_BIND #include #include #include processor_bind(P_PID,P_MYID,mapped_affinity,NULL); #elif HAVE_BINDPROCESSOR #include /* this is the call on AIX. It takes a "what" of BINDPROCESS or BINDTHRAD, then "who" and finally "where" which is a CPU number or it seems PROCESSOR_CLASS_ANY there also seems to be a mycpu() call to return the current CPU assignment. this is all based on the sys/processor.h include file. from empirical testing, it would seem that the my_cpu() call returns the current CPU on which we are running rather than the CPU binding, so it's return value will not tell you if you are bound vs unbound. */ bindprocessor(BINDPROCESS,getpid(),(cpu_t)mapped_affinity); #elif HAVE_SCHED_SETAFFINITY #include /* in theory this should cover systems with more CPUs than bits in a long, without having to specify __USE_GNU. we "cheat" by taking defines from /usr/include/bits/sched.h, which we ass-u-me is included by . If they are not there we will just fall-back on what we had before, which is to use just the size of an unsigned long. raj 2006-09-14 */ #if defined(__CPU_SETSIZE) #define NETPERF_CPU_SETSIZE __CPU_SETSIZE #if defined(__CPU_SET_S) #define NETPERF_CPU_SET(cpu, cpusetp) __CPU_SET_S(cpu, sizeof (cpu_set_t), cpusetp) #define NETPERF_CPU_ZERO(cpusetp) __CPU_ZERO_S (sizeof (cpu_set_t), cpusetp) #else #define NETPERF_CPU_SET(cpu, cpusetp) __CPU_SET(cpu, cpusetp) #define NETPERF_CPU_ZERO(cpusetp) __CPU_ZERO (cpusetp) #endif typedef cpu_set_t netperf_cpu_set_t; #else #define NETPERF_CPU_SETSIZE sizeof(unsigned long) #define NETPERF_CPU_SET(cpu, cpusetp) *cpusetp = 1 << cpu #define NETPERF_CPU_ZERO(cpusetp) *cpusetp = (unsigned long)0 typedef unsigned long netperf_cpu_set_t; #endif netperf_cpu_set_t netperf_cpu_set; unsigned int len = sizeof(netperf_cpu_set); if (mapped_affinity < 8*sizeof(netperf_cpu_set)) { NETPERF_CPU_ZERO(&netperf_cpu_set); NETPERF_CPU_SET(mapped_affinity,&netperf_cpu_set); if (sched_setaffinity(getpid(), len, &netperf_cpu_set)) { if (debug) { fprintf(stderr, "failed to set PID %d's CPU affinity errno %d\n", getpid(),errno); fflush(stderr); } } } else { if (debug) { fprintf(stderr, "CPU number larger than pre-compiled limits. Consider a recompile.\n"); fflush(stderr); } } #elif HAVE_BIND_TO_CPU_ID /* this is the one for Tru64 */ #include #include #include /* really should be checking a return code one of these days. raj 2005/08/31 */ bind_to_cpu_id(getpid(), mapped_affinity,0); #elif WIN32 { ULONG_PTR AffinityMask; ULONG_PTR ProcessAffinityMask; ULONG_PTR SystemAffinityMask; if ((mapped_affinity < 0) || (mapped_affinity > MAXIMUM_PROCESSORS)) { fprintf(where, "Invalid processor_affinity specified: %d\n", mapped_affinity); fflush(where); return; } if (!GetProcessAffinityMask( GetCurrentProcess(), &ProcessAffinityMask, &SystemAffinityMask)) { perror("GetProcessAffinityMask failed"); fflush(stderr); exit(1); } AffinityMask = (ULONG_PTR)1 << mapped_affinity; if (AffinityMask & ProcessAffinityMask) { if (!SetThreadAffinityMask( GetCurrentThread(), AffinityMask)) { perror("SetThreadAffinityMask failed"); fflush(stderr); } } else if (debug) { fprintf(where, "Processor affinity set to CPU# %d\n", mapped_affinity); fflush(where); } } #else if (debug) { fprintf(where, "Processor affinity not available for this platform!\n"); fflush(where); } #endif } /* * Sets a socket to non-blocking operation. */ int set_nonblock (SOCKET sock) { #ifdef WIN32 unsigned long flags = 1; return (ioctlsocket(sock, FIONBIO, &flags) != SOCKET_ERROR); #else return (fcntl(sock, F_SETFL, O_NONBLOCK) != -1); #endif } /* send a request, only converting the first n ints-worth of the test-specific data via htonl() before sending on the connection. the first two ints, which are before the test-specific portion are always converted. raj 2008-02-05 */ void send_request_n(int n) { int counter,count; if (n < 0) count = sizeof(netperf_request)/4; else count = 2 + n; /* silently truncate if the caller called for more than we have */ if (count > sizeof(netperf_request)/4) { if (debug > 1) { fprintf(where, "WARNING, htonl conversion count of %d was larger than netperf_request\n", count - 2); fflush(where); } count = sizeof(netperf_request)/4; } /* display the contents of the request if the debug level is high enough. otherwise, just send the darned thing ;-) */ if (debug > 1) { fprintf(where, "entered send_request_n...contents before %d htonls:\n", count); dump_request(); } /* pass the processor affinity request value to netserver this is a kludge and I know it. sgb 8/11/04. we keep this here to deal with there being two paths to this place - direct and via send_request() */ netperf_request.content.dummy = remote_proc_affinity; /* put the entire request array into network order. We do this arbitrarily rather than trying to figure-out just how much of the request array contains real information. this should be simpler, and at any rate, the performance of sending control messages for this benchmark is not of any real concern. */ for (counter = 0; counter < count; counter++) { request_array[counter] = htonl(request_array[counter]); } if (debug > 1) { fprintf(where,"send_request_n...contents after %d htonls:\n", count); dump_request(); fprintf(where, "\nsend_request: about to send %u bytes from %p\n", (unsigned int) sizeof(netperf_request), &netperf_request); fflush(where); } if (send(netlib_control, (char *)&netperf_request, sizeof(netperf_request), 0) != sizeof(netperf_request)) { perror("send_request: send call failure"); exit(1); } } /***********************************************************************/ /* */ /* send_request() */ /* */ /* send a netperf request on the control socket to the remote half of */ /* the connection. to get us closer to intervendor interoperability, */ /* we will call htonl on each of the int that compose the message to */ /* be sent. the server-half of the connection will call the ntohl */ /* routine to undo any changes that may have been made... */ /* */ /***********************************************************************/ void send_request() { /* pass the processor affinity request value to netserver this is a kludge and I know it. sgb 8/11/04 */ netperf_request.content.dummy = remote_proc_affinity; /* call send_request_n telling it to convert everything */ send_request_n(-1); } /* send a response, only converting the first n ints-worth of the test-specific data via htonl() before sending on the connection. the first two ints, which are before the test-specific portion are always converted. raj 2008-02-05 */ void send_response_n(int n) { int counter, count; int bytes_sent; if (n < 0) count = sizeof(netperf_request)/4; else count = 2 + n; /* silently truncate if the caller called for more than we have */ if (count > sizeof(netperf_request)/4) { if (debug > 1) { fprintf(where, "WARNING, htonl conversion count of %d was larger than netperf_request\n", count - 2); fflush(where); } count = sizeof(netperf_request)/4; } /* display the contents of the request if the debug level is high */ /* enough. otherwise, just send the darned thing ;-) */ if (debug > 1) { fprintf(where, "send_response_n: contents of %u ints before %d htonl,\n", (unsigned int) sizeof(netperf_response)/4, count); dump_response(); } /* put the entire response_array into network order. We do this arbitrarily rather than trying to figure-out just how much of the request array contains real information. this should be simpler, and at any rate, the performance of sending control messages for this benchmark is not of any real concern. */ for (counter = 0; counter < count; counter++) { response_array[counter] = htonl(response_array[counter]); } if (debug > 1) { fprintf(where, "send_response_n: contents after htonl\n"); dump_response(); fprintf(where, "about to send %u bytes from %p\n", (unsigned int) sizeof(netperf_response), &netperf_response); fflush(where); } /*KC*/ if ((bytes_sent = send(server_sock, (char *)&netperf_response, sizeof(netperf_response), 0)) != sizeof(netperf_response)) { perror("send_response_n: send call failure"); fprintf(where, "BytesSent: %d\n", bytes_sent); exit(1); } } /***********************************************************************/ /* */ /* send_response() */ /* */ /* send a netperf response on the control socket to the remote half of */ /* the connection. to get us closer to intervendor interoperability, */ /* we will call htonl on each of the int that compose the message to */ /* be sent. the other half of the connection will call the ntohl */ /* routine to undo any changes that may have been made... */ /* */ /***********************************************************************/ void send_response() { send_response_n(-1); } /* go back and "undo" the ntohl that recv_request() did, starting with the specified point and going to the end of the request array */ void fixup_request_n(int n) { int i; int limit; limit = sizeof(netperf_request) / 4; /* we must remember that the request_array also contains two ints of "other" stuff, so we start the fixup two in - at least I think we should. raj 2012-04-02 */ for (i = n + 2; i < limit; i++) { request_array[i] = htonl(request_array[i]); } if (debug > 1) { fprintf(where, "%s: request contents after fixup at the %d th int\n", __FUNCTION__, n); dump_request(); fflush(where); } } /* receive a request, only converting the first n ints-worth of the test-specific data via htonl() before sending on the connection. the first two ints, which are before the test-specific portion are always converted. raj 2008-02-05 */ int recv_request_timed_n(int n, int seconds) { int tot_bytes_recvd, bytes_recvd, bytes_left; char *buf = (char *)&netperf_request; int buflen = sizeof(netperf_request); int counter,count; fd_set readfds; struct timeval timeout; if (n < 0) count = sizeof(netperf_request)/4; else count = 2 + n; /* silently truncate if the caller called for more than we have */ if (count > sizeof(netperf_request)/4) { if (debug > 1) { fprintf(where, "WARNING, htonl conversion count of %d was larger than netperf_request\n", count - 2); fflush(where); } count = sizeof(netperf_request)/4; } /* for the time being, we rather rely on select decrementing timeout each time to preclude someone with nefarious intent from just dribbling data to us piecemeal. of course, who knows what someone with nefarious intent might come-up with. raj 2012-01-23 */ tot_bytes_recvd = 0; bytes_recvd = 0; /* nt_lint; bytes_recvd uninitialized if buflen == 0 */ bytes_left = buflen; timeout.tv_sec = seconds; timeout.tv_usec = 0; do { FD_ZERO(&readfds); FD_SET(server_sock,&readfds); if (select(FD_SETSIZE, &readfds, 0, 0, (seconds > 0) ? &timeout : NULL) != 1) { fprintf(where, "Issue receiving request on control connection. Errno %d (%s)\n", errno, strerror(errno)); fflush(where); close(server_sock); return -1; } if ((bytes_recvd = recv(server_sock, buf, bytes_left, 0)) > 0) { tot_bytes_recvd += bytes_recvd; buf += bytes_recvd; bytes_left -= bytes_recvd; } } while ((tot_bytes_recvd != buflen) && (bytes_recvd > 0 )); /* put the request into host order */ for (counter = 0; counter < count; counter++) { request_array[counter] = ntohl(request_array[counter]); } if (debug) { fprintf(where, "recv_request: received %d bytes of request.\n", tot_bytes_recvd); fflush(where); } if (bytes_recvd == SOCKET_ERROR) { Print_errno(where, "recv_request: error on recv"); fflush(where); close(server_sock); return -1; } if (bytes_recvd == 0) { /* the remote has shutdown the control connection, we should shut it down as well and return */ if (debug) { fprintf(where, "recv_request: remote requested shutdown of control\n"); fflush(where); } close(server_sock); return 0; } if (tot_bytes_recvd < buflen) { if (debug > 1) dump_request(); fprintf(where, "recv_request: partial request received of %d bytes\n", tot_bytes_recvd); fflush(where); close(server_sock); return -1; } if (debug > 1) { dump_request(); } /* get the processor affinity request value from netperf this is a kludge and I know it. sgb 8/11/04 */ local_proc_affinity = netperf_request.content.dummy; if (local_proc_affinity != -1) { bind_to_specific_processor(local_proc_affinity,0); } return buflen; } /* receive a request, only converting the first n ints-worth of the test-specific data via htonl() before sending on the connection. the first two ints, which are before the test-specific portion are always converted. raj 2008-02-05 */ int recv_request_n(int n) { return recv_request_timed_n(n,0); } /***********************************************************************/ /* */ /* recv_request() */ /* */ /* receive the remote's request on the control socket. we will put */ /* the entire response into host order before giving it to the */ /* calling routine. hopefully, this will go most of the way to */ /* insuring intervendor interoperability. if there are any problems, */ /* we will just punt the entire situation. */ /* */ /***********************************************************************/ int recv_request() { return recv_request_n(-1); } void recv_response_timed_n(int addl_time, int n) { int tot_bytes_recvd, bytes_recvd = 0, bytes_left; char *buf = (char *)&netperf_response; int buflen = sizeof(netperf_response); int counter,count; /* stuff for select, use fd_set for better compliance */ fd_set readfds; struct timeval timeout; tot_bytes_recvd = 0; bytes_left = buflen; if (n < 0) count = sizeof(netperf_request)/4; else count = 2 + n; /* silently truncate if the caller called for more than we have */ if (count > sizeof(netperf_request)/4) { if (debug > 1) { fprintf(where, "WARNING, htonl conversion count of %d was larger than netperf_response\n", count - 2); fflush(where); } count = sizeof(netperf_request)/4; } /* zero out the response structure */ /* BUG FIX SJB 2/4/93 - should be < not <= */ for (counter = 0; counter < sizeof(netperf_response)/sizeof(int); counter++) { response_array[counter] = 0; } /* we only select once. it is assumed that if the response is split (which should not be happening, that we will receive the whole thing and not have a problem ;-) */ FD_ZERO(&readfds); FD_SET(netlib_control,&readfds); timeout.tv_sec = 120 + addl_time; /* wait at least two minutes before punting - the USE_LOOPER CPU stuff may cause remote's to have a bit longer time of it than 60 seconds would allow. triggered by fix from Jeff Dwork. */ timeout.tv_usec = 0; /* select had better return one, or there was either a problem or a */ /* timeout... */ if ((counter = select(FD_SETSIZE, &readfds, 0, 0, &timeout)) != 1) { fprintf(where, "%s: no response received. errno %d counter %d\n", __FUNCTION__, errno, counter); exit(1); } while ((tot_bytes_recvd != buflen) && ((bytes_recvd = recv(netlib_control, buf, bytes_left,0)) > 0 )) { tot_bytes_recvd += bytes_recvd; buf += bytes_recvd; bytes_left -= bytes_recvd; } if (debug) { fprintf(where,"recv_response: received a %d byte response\n", tot_bytes_recvd); fflush(where); } /* put the desired quantity of the response into host order */ for (counter = 0; counter < count; counter++) { response_array[counter] = ntohl(response_array[counter]); } if (bytes_recvd == SOCKET_ERROR) { perror("recv_response"); exit(1); } if (tot_bytes_recvd < buflen) { fprintf(stderr, "recv_response: partial response received: %d bytes\n", tot_bytes_recvd); fflush(stderr); if (debug > 1) dump_response(); exit(1); } if (debug > 1) { dump_response(); } } /* recv_response_timed() receive the remote's response on the control socket. we will put the entire response into host order before giving it to the calling routine. hopefully, this will go most of the way to insuring intervendor interoperability. if there are any problems, we will just punt the entire situation. The call to select at the beginning is to get us out of hang situations where the remote gives-up but we don't find-out about it. This seems to happen only rarely, but it would be nice to be somewhat robust ;-) The "_timed" part is to allow the caller to add (or I suppose subtract) from the length of timeout on the select call. this was added since not all the CPU utilization mechanisms require a 40 second calibration, and we used to have an aribtrary 40 second sleep in "calibrate_remote_cpu" - since we don't _always_ need that, we want to simply add 40 seconds to the select() timeout from that call, but don't want to change all the "recv_response" calls in the code right away. sooo, we push the functionality of the old recv_response() into a new recv_response_timed(addl_timout) call, and have recv_response() call recv_response_timed(0). raj 2005-05-16 */ void recv_response_timed(int addl_time) { /* -1 => convert all the test-specific data via ntohl */ recv_response_timed_n(addl_time,-1); } void recv_response() { /* 0 => no additional time, -1 => convert all test-specific data */ recv_response_timed_n(0,-1); } void recv_response_n(int n) { recv_response_timed_n(0,n); } #if defined(USE_PSTAT) || defined (USE_SYSCTL) int hi_32(big_int) long long *big_int; { union overlay_u { long long dword; long words[2]; } *overlay; overlay = (union overlay_u *)big_int; /* on those systems which are byte swapped, we really wish to return words[1] - at least I think so - raj 4/95 */ if (htonl(1L) == 1L) { /* we are a "normal" :) machine */ return(overlay->words[0]); } else { return(overlay->words[1]); } } int lo_32(big_int) long long *big_int; { union overlay_u { long long dword; long words[2]; } *overlay; overlay = (union overlay_u *)big_int; /* on those systems which are byte swapped, we really wish to return words[0] - at least I think so - raj 4/95 */ if (htonl(1L) == 1L) { /* we are a "normal" :) machine */ return(overlay->words[1]); } else { return(overlay->words[0]); } } #endif /* USE_PSTAT || USE_SYSCTL */ void libmain() { fprintf(where,"hello world\n"); fprintf(where,"debug: %d\n",debug); } void get_sock_buffer (SOCKET sd, enum sock_buffer which, int *effective_sizep) { #ifdef SO_SNDBUF int optname = (which == SEND_BUFFER) ? SO_SNDBUF : SO_RCVBUF; netperf_socklen_t sock_opt_len; sock_opt_len = sizeof(*effective_sizep); if (getsockopt(sd, SOL_SOCKET, optname, (char *)effective_sizep, &sock_opt_len) < 0) { fprintf(where, "netperf: get_sock_buffer: getsockopt %s: errno %d\n", (which == SEND_BUFFER) ? "SO_SNDBUF" : "SO_RCVBUF", errno); fflush(where); *effective_sizep = -1; } if (debug) { fprintf(where, "netperf: get_sock_buffer: " "%s socket size determined to be %d\n", (which == SEND_BUFFER) ? "send" : "receive", *effective_sizep); fflush(where); } #else *effective_sizep = -1; #endif } void set_sock_buffer (SOCKET sd, enum sock_buffer which, int requested_size, int *effective_sizep) { #ifdef SO_SNDBUF int optname = (which == SEND_BUFFER) ? SO_SNDBUF : SO_RCVBUF; /* seems that under Windows, setting a value of zero is how one tells the stack you wish to enable copy-avoidance. Knuth only knows what it will do on other stacks, but it might be interesting to find-out, so we won't bother #ifdef'ing the change to allow asking for 0 bytes. Courtesy of SAF, 2007-05 raj 2007-05-31 */ if (requested_size >= 0) { if (setsockopt(sd, SOL_SOCKET, optname, (char *)&requested_size, sizeof(int)) < 0) { fprintf(where, "netperf: set_sock_buffer: %s option: errno %d (%s)\n", (which == SEND_BUFFER) ? "SO_SNDBUF" : "SO_RCVBUF", errno, strerror(errno)); fflush(where); exit(1); } if (debug > 1) { fprintf(where, "netperf: set_sock_buffer: %s of %d requested.\n", (which == SEND_BUFFER) ? "SO_SNDBUF" : "SO_RCVBUF", requested_size); fflush(where); } } /* the getsockopt() call that used to be here has been hoisted into its own routine to be used on those platforms where the socket buffer sizes might change from the beginning to the end of the run. raj 2008-01-15 */ get_sock_buffer(sd, which, effective_sizep); #else /* SO_SNDBUF */ *effective_sizep = -1; #endif /* SO_SNDBUF */ } void dump_addrinfo(FILE *dumploc, struct addrinfo *info, const char *host, char *port, int family) { struct sockaddr *ai_addr; struct addrinfo *temp; temp=info; fprintf(dumploc, "getaddrinfo returned the following for host '%s' port '%s' " " family %s\n", host, port, inet_ftos(family)); while (temp) { /* seems that Solaris 10 GA bits will not give a canonical name for ::0 or 0.0.0.0, and their fprintf() cannot deal with a null pointer, so we have to check for a null pointer. probably a safe thing to do anyway, eventhough it was not necessary on linux or hp-ux. raj 2005-02-09 */ fprintf(dumploc, "\tcannonical name: '%s'\n" "\tflags: %x family: %s: socktype: %s protocol %s addrlen %d\n", (temp->ai_canonname) ? temp->ai_canonname : "(nil)", temp->ai_flags, inet_ftos(temp->ai_family), inet_ttos(temp->ai_socktype), inet_ptos(temp->ai_protocol), temp->ai_addrlen); ai_addr = temp->ai_addr; if (ai_addr != NULL) { int i; fprintf(dumploc, "\tsa_family: %s sadata:", inet_ftos(ai_addr->sa_family)); for (i = 0; i < (int) temp->ai_addrlen; i++) { fprintf(dumploc, (temp->ai_family == AF_INET) ? " %d" : " %.2x", (u_char)ai_addr->sa_data[i]); } fprintf(dumploc,"\n"); } temp = temp->ai_next; } fflush(dumploc); } struct addrinfo * resolve_host(char *hostname, char *port, int family) { struct addrinfo hints; struct addrinfo *ai; int count; int error; if (debug) { fprintf(where, "resolve_host called with host '%s' port '%s' family %s\n", hostname, port, inet_ftos(family)); fflush(where); } memset(&hints, 0, sizeof(hints)); hints.ai_family = family; hints.ai_socktype = SOCK_STREAM; hints.ai_protocol = IPPROTO_TCP; hints.ai_flags = AI_CANONNAME | AI_ADDRCONFIG; count = 0; do { error = getaddrinfo((char *)hostname, (char *)port, &hints, &ai); count += 1; if (error == EAI_AGAIN) { if (debug) { fprintf(where,"Sleeping on getaddrinfo EAI_AGAIN\n"); fflush(where); } sleep(1); } } while ((error == EAI_AGAIN) && (count <= 5)); if (error) { printf("%s: could not resolve host '%s' port '%s' af %s" "\n\tgetaddrinfo returned %d %s\n", __FUNCTION__, hostname, port, inet_ftos(family), error, gai_strerror(error)); return(NULL); } if (debug) { dump_addrinfo(where, ai, hostname, port, family); } return (ai); } /* establish_control() set-up the control connection between netperf and the netserver so we can actually run some tests. if we cannot establish the control connection, that may or may not be a good thing, so we will let the caller decide what to do. to assist with pesky end-to-end-unfriendly things like firewalls, we allow the caller to specify both the remote hostname and port, and the local addressing info. i believe that in theory it is possible another, but for the time being, we are only going to take-in one requested address family parameter. this means that the only way (iirc) that we might get a mixed-mode connection would be if the address family is specified as AF_UNSPEC, and getaddrinfo() returns different families for the local and server names. the "names" can also be IP addresses in ASCII string form. raj 2003-02-27 */ SOCKET establish_control_internal(char *hostname, char *port, int remfam, char *localhost, char *localport, int locfam) { int not_connected; SOCKET control_sock; struct addrinfo *local_res; struct addrinfo *remote_res; struct addrinfo *local_res_temp; struct addrinfo *remote_res_temp; remote_res = resolve_host(hostname, port, remfam); if (!remote_res) return(INVALID_SOCKET); local_res = resolve_host(localhost, localport, locfam); if (!local_res) return(INVALID_SOCKET); if (debug) { fprintf(where, "establish_control called with host '%s' port '%s' remfam %s\n" "\t\tlocal '%s' port '%s' locfam %s\n", hostname, port, inet_ftos(remfam), localhost, localport, inet_ftos(locfam)); fflush(where); } not_connected = 1; local_res_temp = local_res; remote_res_temp = remote_res; /* we want to loop through all the possibilities. looping on the local addresses will be handled within the while loop. I suppose these is some more "C-expert" way to code this, but it has not lept to mind just yet :) raj 2003-02024 */ while (remote_res_temp != NULL) { /* I am guessing that we should use the address family of the local endpoint, and we will not worry about mixed family types - presumeably the stack or other transition mechanisms will be able to deal with that for us. famous last words :) raj 2003-02-26 */ control_sock = socket(local_res_temp->ai_family, SOCK_STREAM, 0); if (control_sock == INVALID_SOCKET) { /* at some point we'll need a more generic "display error" message for when/if we use GUIs and the like. unlike a bind or connect failure, failure to allocate a socket is "immediately fatal" and so we return to the caller. raj 2003-02-24 */ if (debug) { perror("establish_control: unable to allocate control socket"); } return(INVALID_SOCKET); } /* if we are going to control the local enpoint addressing, we need to call bind. of course, we should probably be setting one of the SO_REUSEmumble socket options? raj 2005-02-04 */ if (bind(control_sock, local_res_temp->ai_addr, local_res_temp->ai_addrlen) == 0) { if (debug) { fprintf(where, "bound control socket to %s and %s\n", localhost, localport); } if (connect(control_sock, remote_res_temp->ai_addr, remote_res_temp->ai_addrlen) == 0) { /* we have successfully connected to the remote netserver */ if (debug) { fprintf(where, "successful connection to remote netserver at %s and %s\n", hostname, port); } not_connected = 0; /* this should get us out of the while loop */ break; } else { /* the connect call failed */ if (debug) { fprintf(where, "establish_control: connect failed, errno %d %s\n" " trying next address combination\n", errno, strerror(errno)); fflush(where); } } } else { /* the bind failed */ if (debug) { fprintf(where, "establish_control: bind failed, errno %d %s\n" " trying next address combination\n", errno, strerror(errno)); fflush(where); } } if ((local_res_temp = local_res_temp->ai_next) == NULL) { /* wrap the local and move to the next server, don't forget to close the current control socket. raj 2003-02-24 */ local_res_temp = local_res; /* the outer while conditions will deal with the case when we get to the end of all the possible remote addresses. */ remote_res_temp = remote_res_temp->ai_next; /* it is simplest here to just close the control sock. since this is not a performance critical section of code, we don't worry about overheads for socket allocation or close. raj 2003-02-24 */ } close(control_sock); } control_family = local_res_temp->ai_family; /* we no longer need the addrinfo stuff */ freeaddrinfo(local_res); freeaddrinfo(remote_res); /* so, we are either connected or not */ if (not_connected) { fprintf(where, "establish control: are you sure there is a netserver " "listening on %s at port %s?\n", hostname, port); fflush(where); control_family = AF_UNSPEC; return(INVALID_SOCKET); } /* at this point, we are connected. we probably want some sort of version check with the remote at some point. raj 2003-02-24 */ return(control_sock); } void establish_control(char *hostname, char *port, int remfam, char *localhost, char *localport, int locfam) { netlib_control = establish_control_internal(hostname, port, remfam, localhost, localport, locfam); if (netlib_control == INVALID_SOCKET) { fprintf(where, "establish_control could not establish the control" " connection from %s port %s address family %s to %s" " port %s address family %s\n", localhost,localport,inet_ftos(locfam), hostname,port,inet_ftos(remfam)); fflush(where); exit(INVALID_SOCKET); } } /***********************************************************************/ /* */ /* get_id() */ /* */ /* Return a string to the calling routine that contains the */ /* identifying information for the host we are running on. This */ /* information will then either be displayed locally, or returned to */ /* a remote caller for display there. */ /* */ /***********************************************************************/ char * get_id() { static char id_string[80]; #ifdef WIN32 char system_name[MAX_COMPUTERNAME_LENGTH+1] ; DWORD name_len = MAX_COMPUTERNAME_LENGTH + 1 ; #else struct utsname system_name; #endif /* WIN32 */ #ifdef WIN32 SYSTEM_INFO SystemInfo; GetSystemInfo( &SystemInfo ) ; if ( !GetComputerName(system_name , &name_len) ) strcpy(system_name , "no_name") ; #else if (uname(&system_name) <0) { perror("identify_local: uname"); exit(1); } #endif /* WIN32 */ snprintf(id_string, sizeof(id_string), #ifdef WIN32 "%-15s%-15s%d.%d%d", "Windows NT", system_name , GetVersion() & 0xFF , GetVersion() & 0xFF00 , SystemInfo.dwProcessorType #else "%-15s%-15s%-15s%-15s%-15s", system_name.sysname, system_name.nodename, system_name.release, system_name.version, system_name.machine #endif /* WIN32 */ ); return (id_string); } /***********************************************************************/ /* */ /* identify_local() */ /* */ /* Display identifying information about the local host to the user. */ /* At first release, this information will be the same as that which */ /* is returned by the uname -a command, with the exception of the */ /* idnumber field, which seems to be a non-POSIX item, and hence */ /* non-portable. */ /* */ /***********************************************************************/ void identify_local() { char *local_id; local_id = get_id(); fprintf(where,"Local Information \n\ Sysname Nodename Release Version Machine\n"); fprintf(where,"%s\n", local_id); } /***********************************************************************/ /* */ /* identify_remote() */ /* */ /* Display identifying information about the remote host to the user. */ /* At first release, this information will be the same as that which */ /* is returned by the uname -a command, with the exception of the */ /* idnumber field, which seems to be a non-POSIX item, and hence */ /* non-portable. A request is sent to the remote side, which will */ /* return a string containing the utsname information in a */ /* pre-formatted form, which is then displayed after the header. */ /* */ /***********************************************************************/ void identify_remote() { char *remote_id=""; /* send a request for node info to the remote */ netperf_request.content.request_type = NODE_IDENTIFY; send_request(); /* and now wait for the reply to come back */ recv_response(); if (netperf_response.content.serv_errno) { Set_errno(netperf_response.content.serv_errno); perror("identify_remote: on remote"); exit(1); } fprintf(where,"Remote Information \n\ Sysname Nodename Release Version Machine\n"); fprintf(where,"%s", remote_id); } void cpu_start(int measure_cpu) { gettimeofday(&time1, &tz); if (measure_cpu) { cpu_util_init(); measuring_cpu = 1; cpu_method = get_cpu_method(); cpu_start_internal(); } } void cpu_stop(int measure_cpu, float *elapsed) { int sec, usec; if (measure_cpu) { cpu_stop_internal(); cpu_util_terminate(); } gettimeofday(&time2, &tz); if (time2.tv_usec < time1.tv_usec) { time2.tv_usec += 1000000; time2.tv_sec -= 1; } sec = time2.tv_sec - time1.tv_sec; usec = time2.tv_usec - time1.tv_usec; lib_elapsed = (float)sec + ((float)usec/(float)1000000.0); #ifdef WIN32 if (timed_out) lib_elapsed-=PAD_TIME/2; #endif *elapsed = lib_elapsed; } double calc_thruput_interval(double units_received,double elapsed) { double divisor; /* We will calculate the thruput in libfmt units/second */ switch (libfmt) { case 'K': divisor = 1024.0; break; case 'M': divisor = 1024.0 * 1024.0; break; case 'G': divisor = 1024.0 * 1024.0 * 1024.0; break; case 'k': divisor = 1000.0 / 8.0; break; case 'm': divisor = 1000.0 * 1000.0 / 8.0; break; case 'g': divisor = 1000.0 * 1000.0 * 1000.0 / 8.0; break; case 'x': case 'b': case 'B': divisor = 1.0; break; case 'u': /* latency in microseconds a bit squirrely but we don't want to really muck with things for the default return statement. invert transactions per second and multiply to get microseconds per transaction */ return (1 / (units_received / elapsed)) * 1000000.0; default: divisor = 1024.0; } return (units_received / divisor / elapsed); } double calc_thruput(double units_received) { return(calc_thruput_interval(units_received,lib_elapsed)); } /* these "_omni" versions are ones which understand 'x' as a unit, meaning transactions/s. we have a separate routine rather than convert the existing routine so we don't have to go and change _all_ the nettest_foo.c files at one time. raj 2007-06-08 */ double calc_thruput_interval_omni(double units_received,double elapsed) { double divisor; /* We will calculate the thruput in libfmt units/second */ switch (libfmt) { case 'K': divisor = 1024.0; break; case 'M': divisor = 1024.0 * 1024.0; break; case 'G': divisor = 1024.0 * 1024.0 * 1024.0; break; case 'k': divisor = 1000.0 / 8.0; break; case 'm': divisor = 1000.0 * 1000.0 / 8.0; break; case 'g': divisor = 1000.0 * 1000.0 * 1000.0 / 8.0; break; case 'x': case 'b': case 'B': divisor = 1.0; break; case 'u': /* latency in microseconds a bit squirrely but we don't want to really muck with things for the default return statement. invert transactions per second and multiply to get microseconds per transaction */ return (1 / (units_received / elapsed)) * 1000000.0; default: fprintf(where, "WARNING calc_throughput_internal_omni: unknown units %c\n", libfmt); fflush(where); divisor = 1024.0; } return (units_received / divisor / elapsed); } double calc_thruput_omni(double units_received) { return(calc_thruput_interval_omni(units_received,lib_elapsed)); } float calc_cpu_util(float elapsed_time) { float temp_util; int i; temp_util = calc_cpu_util_internal(elapsed_time); /* now, what was the most utilized CPU and its util? */ for (i = 0; i < MAXCPUS; i++) { if (lib_local_per_cpu_util[i] > lib_local_cpu_stats.peak_cpu_util) { lib_local_cpu_stats.peak_cpu_util = lib_local_per_cpu_util[i]; lib_local_cpu_stats.peak_cpu_id = lib_cpu_map[i]; } } return temp_util; } float calc_service_demand_internal(double unit_divisor, double units_sent, float elapsed_time, float cpu_utilization, int num_cpus) { double service_demand; double thruput; if (debug) { fprintf(where, "calc_service_demand called: units_sent = %f\n" " elapsed_time = %f\n" " cpu_util = %f\n" " num cpu = %d\n", units_sent, elapsed_time, cpu_utilization, num_cpus); fflush(where); } if (num_cpus == 0) num_cpus = lib_num_loc_cpus; if (elapsed_time == 0.0) { elapsed_time = lib_elapsed; } if (cpu_utilization == 0.0) { cpu_utilization = lib_local_cpu_stats.cpu_util; } thruput = (units_sent / (double) unit_divisor / (double) elapsed_time); /* on MP systems, it is necessary to multiply the service demand by the number of CPU's. at least, I believe that to be the case:) raj 10/95 */ /* thruput has a "per second" component. if we were using 100% ( 100.0) of the CPU in a second, that would be 1 second, or 1 millisecond, so we multiply cpu_utilization by 10 to go to milliseconds, or 10,000 to go to micro seconds. With revision 2.1, the service demand measure goes to microseconds per unit. raj 12/95 */ service_demand = (cpu_utilization*10000.0/thruput) * (float) num_cpus; if (debug) { fprintf(where, "calc_service_demand using: units_sent = %f\n" " elapsed_time = %f\n" " cpu_util = %f\n" " num cpu = %d\n" "calc_service_demand got: thruput = %f\n" " servdem = %f\n", units_sent, elapsed_time, cpu_utilization, num_cpus, thruput, service_demand); fflush(where); } return (float)service_demand; } float calc_service_demand(double units_sent, float elapsed_time, float cpu_utilization, int num_cpus) { double unit_divisor = (double)1024.0; return(calc_service_demand_internal(unit_divisor, units_sent, elapsed_time, cpu_utilization, num_cpus)); } /* use the value of libfmt to determine the unit_divisor */ float calc_service_demand_fmt(double units_sent, float elapsed_time, float cpu_utilization, int num_cpus) { double unit_divisor; if ('x' == libfmt) unit_divisor = 1.0; else unit_divisor = 1024.0; return(calc_service_demand_internal(unit_divisor, units_sent, elapsed_time, cpu_utilization, num_cpus)); } float calibrate_local_cpu(float local_cpu_rate) { lib_num_loc_cpus = get_num_cpus(); lib_use_idle = 0; #ifdef USE_LOOPER cpu_util_init(); lib_use_idle = 1; #endif /* USE_LOOPER */ if (local_cpu_rate > 0) { /* The user think that he knows what the cpu rate is. We assume that all the processors of an MP system are essentially the same - for this reason we do not have a per processor maxrate. if the machine has processors which are different in performance, the CPU utilization will be skewed. raj 4/95 */ lib_local_maxrate = local_cpu_rate; } else { /* if neither USE_LOOPER nor USE_PSTAT are defined, we return a 0.0 to indicate that times or getrusage should be used. raj 4/95 */ lib_local_maxrate = (float)0.0; #if defined(USE_PROC_STAT) || defined(USE_LOOPER) || defined(USE_PSTAT) || defined(USE_KSTAT) || defined(USE_PERFSTAT) || defined(USE_SYSCTL) lib_local_maxrate = calibrate_idle_rate(4,10); #endif } return lib_local_maxrate; } float calibrate_remote_cpu() { float remrate; netperf_request.content.request_type = CPU_CALIBRATE; send_request(); /* we know that calibration will last at least 40 seconds, so go to sleep for that long so the 60 second select in recv_response will not pop. raj 7/95 */ /* we know that CPU calibration may last as long as 40 seconds, so make sure we "select" for at least that long while looking for the response. raj 2005-05-16 */ recv_response_timed(40); if (netperf_response.content.serv_errno) { /* initially, silently ignore remote errors and pass back a zero to the caller this should allow us to mix rev 1.0 and rev 1.1 netperfs... */ return((float)0.0); } else { /* the rate is the first word of the test_specific data */ bcopy((char *)netperf_response.content.test_specific_data, (char *)&remrate, sizeof(remrate)); bcopy((char *)netperf_response.content.test_specific_data + sizeof(remrate), (char *)&lib_num_rem_cpus, sizeof(lib_num_rem_cpus)); /* remrate = (float) netperf_response.content.test_specific_data[0]; */ return(remrate); } } #ifndef WIN32 /* WIN32 requires that at least one of the file sets to select be non-null. Since msec_sleep routine is only called by nettest_dlpi & nettest_unix, let's duck this issue. */ int msec_sleep( int msecs ) { int rval ; struct timeval timeout; timeout.tv_sec = msecs / 1000; timeout.tv_usec = (msecs - (msecs/1000) *1000) * 1000; if ((rval = select(0, 0, 0, 0, &timeout))) { if ( SOCKET_EINTR(rval) ) { return(1); } perror("msec_sleep: select"); exit(1); } return(0); } #endif /* WIN32 */ #if defined(WANT_INTERVALS) || defined(WANT_DEMO) int demo_mode; /* are we actually in demo mode? = 0 == not in demo mode; 1 == classic unit based demo mode; 2 == always timestamp demo mode */ double demo_interval = 1000000.0; /* what is the desired interval to display interval results. default is one second in units of microseconds */ double demo_units = 0.0; /* what is our current best guess as to how many work units must be done to be near the desired reporting interval? */ double units_this_tick; #endif #ifdef WANT_DEMO #ifdef HAVE_GETHRTIME static hrtime_t demo_one; static hrtime_t demo_two; static hrtime_t *demo_one_ptr = &demo_one; static hrtime_t *demo_two_ptr = &demo_two; static hrtime_t *temp_demo_ptr = &demo_one; #elif defined(WIN32) static LARGE_INTEGER demo_one; static LARGE_INTEGER demo_two; static LARGE_INTEGER *demo_one_ptr = &demo_one; static LARGE_INTEGER *demo_two_ptr = &demo_two; static LARGE_INTEGER *temp_demo_ptr = &demo_one; #else static struct timeval demo_one; static struct timeval demo_two; static struct timeval *demo_one_ptr = &demo_one; static struct timeval *demo_two_ptr = &demo_two; static struct timeval *temp_demo_ptr = &demo_one; #endif void demo_first_timestamp() { HIST_timestamp(demo_one_ptr); } void demo_reset() { if (debug) { fprintf(where, "Resetting interim results\n"); fflush(where); } units_this_tick = 0; demo_first_timestamp(); } /* for a _STREAM test, "a" should be lss_size and "b" should be rsr_size. for a _MAERTS test, "a" should be lsr_size and "b" should be rss_size. raj 2005-04-06 */ void demo_stream_setup(uint32_t a, uint32_t b) { if ((demo_mode) && (demo_units == 0)) { /* take our default value of demo_units to be the larger of twice the remote's SO_RCVBUF or twice our SO_SNDBUF */ if (a > b) { demo_units = 2*a; } else { demo_units = 2*b; } } } #ifdef WIN32 __forceinline void demo_interval_display(double actual_interval) #else inline void demo_interval_display(double actual_interval) #endif { static int count = 0; struct timeval now; gettimeofday(&now,NULL); switch (netperf_output_mode) { case HUMAN: fprintf(where, "Interim result: %7.2f %s/s over %.3f seconds ending at %ld.%.3ld\n", calc_thruput_interval(units_this_tick, actual_interval/1000000.0), format_units(), actual_interval/1000000.0, now.tv_sec, (long) now.tv_usec/1000); break; case CSV: fprintf(where, "%7.2f,%s/s,%.3f,%ld.%.3ld\n", calc_thruput_interval(units_this_tick, actual_interval/1000000.0), format_units(), actual_interval/1000000.0, now.tv_sec, (long) now.tv_usec/1000); break; case KEYVAL: fprintf(where, "NETPERF_INTERIM_RESULT[%d]=%.2f\n" "NETPERF_UNITS[%d]=%s/s\n" "NETPERF_INTERVAL[%d]=%.3f\n" "NETPERF_ENDING[%d]=%ld.%.3ld\n", count, calc_thruput_interval(units_this_tick, actual_interval/1000000.0), count, format_units(), count, actual_interval/1000000.0, count, now.tv_sec, (long) now.tv_usec/1000); count += 1; break; default: fprintf(where, "Hey Ricky you not fine, theres a bug at demo time. Hey Ricky!"); fflush(where); exit(-1); } fflush(where); } /* this has gotten long enough to warrant being an inline function rather than a macro, and it has been enough years since all the important compilers have supported such a construct so it should not be a big deal. raj 2012-01-23 */ #ifdef WIN32 /* It would seem that the Microsoft compiler will not inline across source files. So there is little point in having an inline directive in that situation. Of course that makes me wonder if an inline directive has to appear in netlib.h... */ void demo_interval_tick(uint32_t units) #else inline void demo_interval_tick(uint32_t units) #endif { double actual_interval = 0.0; switch (demo_mode) { case 0: return; case 1: /* use the unit accumulation first */ units_this_tick += units; if (units_this_tick >= demo_units) { /* time to possibly update demo_units and maybe output an interim result */ HIST_timestamp(demo_two_ptr); actual_interval = delta_micro(demo_one_ptr,demo_two_ptr); /* we always want to fine-tune demo_units here whether we emit an interim result or not. if we are short, this will lengthen demo_units. if we are long, this will shorten it */ demo_units = demo_units * (demo_interval / actual_interval); } else return; break; case 2: /* Always timestamp */ units_this_tick += units; HIST_timestamp(demo_two_ptr); actual_interval = delta_micro(demo_one_ptr,demo_two_ptr); break; default: fprintf(where, "Unexpected value of demo_mode of %d. Please report this as a bug.\n", demo_mode); fflush(where); exit(-1); } /* units == 0 will be when we have completed a test. we want to emit a final interim results if there is anything to report */ if (actual_interval >= demo_interval) { /* time to emit an interim result, giving the current time to the millisecond for compatability with RRD */ demo_interval_display(actual_interval); units_this_tick = 0.0; /* now get a new starting timestamp. we could be clever and swap pointers - the math we do probably does not take all that long, but for now this will suffice */ temp_demo_ptr = demo_one_ptr; demo_one_ptr = demo_two_ptr; demo_two_ptr = temp_demo_ptr; } } void demo_interval_final() { double actual_interval; switch (demo_mode) { case 0: return; case 1: case 2: if (units_this_tick > 0.0) { HIST_timestamp(demo_two_ptr); actual_interval = delta_micro(demo_one_ptr,demo_two_ptr); demo_interval_display(actual_interval); units_this_tick = 0.0; } } } void demo_stream_interval(uint32_t units) { demo_interval_tick(units); } void demo_rr_setup(uint32_t a) { if ((demo_mode) && (demo_units == 0)) { /* take whatever we are given */ demo_units = a; } } void demo_rr_interval(uint32_t units) { demo_interval_tick(units); } #endif /* hist.c Given a time difference in microseconds, increment one of 61 different buckets: 0 - 9 in increments of 1 usec 0 - 9 in increments of 10 usecs 0 - 9 in increments of 100 usecs 1 - 9 in increments of 1 msec 1 - 9 in increments of 10 msecs 1 - 9 in increments of 100 msecs 1 - 9 in increments of 1 sec 1 - 9 in increments of 10 sec > 100 secs This will allow any time to be recorded to within an accuracy of 10%, and provides a compact representation for capturing the distribution of a large number of time differences (e.g. request-response latencies). Colin Low 10/6/93 Rick Jones 2004-06-15 extend to unit and ten usecs */ /* #include "sys.h" */ /*#define HIST_TEST*/ HIST HIST_new_n(int max_outstanding) { HIST h; if((h = (HIST) malloc(sizeof(struct histogram_struct))) == NULL) { perror("HIST_new_n - histogram_struct malloc failed"); exit(1); } HIST_clear(h); /* we never want to have a full queue, so will trade a little space for that. one day we may still have to check for a full queue */ h->limit = max_outstanding + 1; /* now allocate the time_ones based on h->limit */ #ifdef HAVE_GETHRTIME h->time_ones = (hrtime_t *) malloc(h->limit * sizeof(hrtime_t)); #elif HAVE_GET_HRT h->time_ones = (hrt_t *) malloc(h->limit * sizeof(hrt_t)); #elif defined(WIN32) h->time_ones = (LARGE_INTEGER *) malloc(h->limit * sizeof(LARGE_INTEGER)); #else h->time_ones = (struct timeval *) malloc(h->limit * sizeof(struct timeval)); #endif /* HAVE_GETHRTIME */ if (h->time_ones == NULL) { perror("HIST_new_n - time_ones malloc failed"); exit(1); } return h; } HIST HIST_new(void){ return HIST_new_n(0); } void HIST_clear(HIST h){ int i; for(i = 0; i < HIST_NUM_OF_BUCKET; i++){ h->unit_usec[i] = 0; h->ten_usec[i] = 0; h->hundred_usec[i] = 0; h->unit_msec[i] = 0; h->ten_msec[i] = 0; h->hundred_msec[i] = 0; h->unit_sec[i] = 0; h->ten_sec[i] = 0; } h->ridiculous = 0; h->total = 0; h->sum = 0; h->sumsquare = 0; h->hmin = 0; h->hmax = 0; h->limit = 0; h->count = 0; h->producer = 0; h->consumer = 0; h->time_ones = NULL; } void HIST_purge(HIST h) { h->count = 0; h->producer = 0; h->consumer = 0; } void HIST_add(register HIST h, int time_delta){ register float val; register int base = HIST_NUM_OF_BUCKET / 10; /* check for < 0 added via VMware ESX patches. */ /* hoisted up to the top because we do not want to count any ridiculous values in the actual statistics. right? raj 2011-07-28 */ if (time_delta < 0) { h->ridiculous++; return; } if (!h->total) h->hmin = h->hmax = time_delta; h->total++; h->sum += time_delta; /* am I just being paranoid about the overhead of pow() when we aren't all that interested in the statistics derived from it? raj 20100914 */ if (keep_statistics) { h->sumsquare += pow(time_delta, 2); } h->hmin = ((h->hmin < time_delta) ? h->hmin : time_delta); h->hmax = ((h->hmax > time_delta) ? h->hmax : time_delta); val = (float) time_delta; if(val < 10) h->unit_usec[(int)(val * base)]++; else { val /= 10; if(val < 10) h->ten_usec[(int)(val * base)]++; else { val /= 10; if(val < 10) h->hundred_usec[(int)(val * base)]++; else { val /= 10; if(val < 10) h->unit_msec[(int)(val * base)]++; else { val /= 10; if(val < 10) h->ten_msec[(int)(val * base)]++; else { val /= 10; if(val < 10) h->hundred_msec[(int)(val * base)]++; else { val /= 10; if(val < 10) h->unit_sec[(int)(val * base)]++; else { val /= 10; if(val < 10) h->ten_sec[(int)(val * base)]++; else h->ridiculous++; } } } } } } } } void output_row(FILE *fd, char *title, int *row){ register int i; register int j; register int base = HIST_NUM_OF_BUCKET / 10; register int sum; fprintf(where,"%s", title); for(i = 0; i < 10; i++){ sum = 0; for (j = i * base; j < (i + 1) * base; j++) { sum += row[j]; } fprintf(where,": %4d", sum); } fprintf(where,"\n"); } int sum_row(int *row) { int sum = 0; int i; for (i = 0; i < HIST_NUM_OF_BUCKET; i++) sum += row[i]; return(sum); } void HIST_report(HIST h){ #ifndef OLD_HISTOGRAM output_row(stdout, "UNIT_USEC ", h->unit_usec); output_row(stdout, "TEN_USEC ", h->ten_usec); output_row(stdout, "HUNDRED_USEC ", h->hundred_usec); #else h->hundred_usec[0] += sum_row(h->unit_usec); h->hundred_usec[0] += sum_row(h->ten_usec); output_row(stdout, "TENTH_MSEC ", h->hundred_usec); #endif output_row(stdout, "UNIT_MSEC ", h->unit_msec); output_row(stdout, "TEN_MSEC ", h->ten_msec); output_row(stdout, "HUNDRED_MSEC ", h->hundred_msec); output_row(stdout, "UNIT_SEC ", h->unit_sec); output_row(stdout, "TEN_SEC ", h->ten_sec); fprintf(where,">100_SECS: %d\n", h->ridiculous); fprintf(where,"HIST_TOTAL: %d\n", h->total); if (debug) { fprintf(where, "sum %"PRIi64", sumsquare %f, limit %d count %d\n", h->sum, h->sumsquare, h->limit, h->count); } } /* search buckets for each unit */ int HIST_search_bucket(int *unit, int num, int *last, int *current, double scale){ int base = HIST_NUM_OF_BUCKET / 10; int i; for (i = 0; i < HIST_NUM_OF_BUCKET; i++){ *last = *current; *current += unit[i]; if (*current >= num) return (int)((i + (double)(num - *last)/(*current - *last)) * scale/base); } return 0; } /* get percentile from histogram */ int HIST_get_percentile(HIST h, const double percentile){ double win_kludge = percentile * (double) h->total; int num = (int) win_kludge; int last = 0; int current = 0; int result; if (!num) return 0; /* search in unit usec range */ result = HIST_search_bucket(h->unit_usec, num, &last, ¤t, 1e0); if (result) return result; /* search in ten usec range */ result = HIST_search_bucket(h->ten_usec, num, &last, ¤t, 1e1); if (result) return result; /* search in ten hundred usec range */ result = HIST_search_bucket(h->hundred_usec, num, &last, ¤t, 1e2); if (result) return result; /* search in unic msec range */ result = HIST_search_bucket(h->unit_msec, num, &last, ¤t, 1e3); if (result) return result; /* search in ten msec range */ result = HIST_search_bucket(h->ten_msec, num, &last, ¤t, 1e4); if (result) return result; /* search in hundred msec range */ result = HIST_search_bucket(h->hundred_msec, num, &last, ¤t, 1e5); if (result) return result; /* search in unit sec range */ result = HIST_search_bucket(h->unit_sec, num, &last, ¤t, 1e6); if (result) return result; /* search in ten sec range */ result = HIST_search_bucket(h->ten_sec, num, &last, ¤t, 1e7); if (result) return result; return (int)(1e8); } /* get basic stats */ void HIST_get_stats(HIST h, int *min, int *max, double *mean, double *stddev){ *min = h->hmin; *max = h->hmax; if (h->total){ *mean = (double)h->sum / (double)h->total; *stddev = (h->sumsquare * h->total - pow((double)h->sum, 2)) / pow(h->total, 2); *stddev = sqrt(*stddev); } else{ *mean = 0; *stddev = 0; } } /* with the advent of sit-and-spin intervals support, we might as well make these things available all the time, not just for demo or histogram modes. raj 2006-02-06 */ #ifdef HAVE_GETHRTIME void HIST_timestamp(hrtime_t *timestamp) { *timestamp = gethrtime(); } int delta_micro(hrtime_t *begin, hrtime_t *end) { long nsecs; nsecs = (*end) - (*begin); return(nsecs/1000); } #elif defined(HAVE_GET_HRT) #include "hrt.h" void HIST_timestamp(hrt_t *timestamp) { *timestamp = get_hrt(); } int delta_micro(hrt_t *begin, hrt_t *end) { return((int)get_hrt_delta(*end,*begin)); } #elif defined(WIN32) void HIST_timestamp(LARGE_INTEGER *timestamp) { QueryPerformanceCounter(timestamp); } int delta_micro(LARGE_INTEGER *begin, LARGE_INTEGER *end) { LARGE_INTEGER DeltaTimestamp; static LARGE_INTEGER TickHz = {{0,0}}; if (TickHz.QuadPart == 0) { QueryPerformanceFrequency(&TickHz); } /*+*+ Rick; this will overflow after ~2000 seconds, is that good enough? Spencer: Yes, that should be more than good enough for histogram support */ DeltaTimestamp.QuadPart = (end->QuadPart - begin->QuadPart) * 1000000/TickHz.QuadPart; assert((DeltaTimestamp.HighPart == 0) && ((int)DeltaTimestamp.LowPart >= 0)); return (int)DeltaTimestamp.LowPart; } #else void HIST_timestamp(struct timeval *timestamp) { gettimeofday(timestamp,NULL); } /* return the difference (in micro seconds) between two timeval */ /* timestamps */ int delta_micro(struct timeval *begin,struct timeval *end) { int usecs, secs; if (end->tv_usec < begin->tv_usec) { /* borrow a second from the tv_sec */ end->tv_usec += 1000000; end->tv_sec--; } usecs = end->tv_usec - begin->tv_usec; secs = end->tv_sec - begin->tv_sec; usecs += (secs * 1000000); return(usecs); } #endif /* HAVE_GETHRTIME */ void HIST_timestamp_start(HIST h) { if (NULL == h) { fprintf(where,"HIST_timestamp_start called with NULL histogram\n"); fflush(where); exit(-1); } if (h->count == h->limit) { fprintf(where,"HIST_timestamp_start called with full time_ones\n"); } HIST_timestamp(&(h->time_ones[h->producer])); h->producer += 1; h->producer %= h->limit; h->count += 1; } /* snap an ending timestamp and add the delta to the histogram */ void HIST_timestamp_stop_add(HIST h) { if (NULL == h) { fprintf(where,"HIST_timestamp_stop called with NULL histogram\n"); fflush(where); exit(-1); } if (h->consumer == h->producer) { fprintf(where, "HIST_timestamp_stop called with empty time_ones consumer %d producer %d\n", h->consumer, h->producer); fflush(where); exit(-1); } /* take our stopping timestamp */ HIST_timestamp(&(h->time_two)); /* now add it */ HIST_add(h,delta_micro(&(h->time_ones[h->consumer]),&(h->time_two))); h->consumer += 1; h->consumer %= h->limit; h->count -= 1; } /* these routines for confidence intervals are courtesy of IBM. They have been modified slightly for more general usage beyond TCP/UDP tests. raj 11/94 I would suspect that this code carries an IBM copyright that is much the same as that for the original HP netperf code */ int confidence_iterations; /* for iterations */ double result_confid=-10.0, loc_cpu_confid=-10.0, rem_cpu_confid=-10.0, measured_sum_result=0.0, measured_square_sum_result=0.0, measured_mean_result=0.0, measured_var_result=0.0, measured_sum_local_cpu=0.0, measured_square_sum_local_cpu=0.0, measured_mean_local_cpu=0.0, measured_var_local_cpu=0.0, measured_sum_remote_cpu=0.0, measured_square_sum_remote_cpu=0.0, measured_mean_remote_cpu=0.0, measured_var_remote_cpu=0.0, measured_sum_local_service_demand=0.0, measured_square_sum_local_service_demand=0.0, measured_mean_local_service_demand=0.0, measured_var_local_service_demand=0.0, measured_sum_remote_service_demand=0.0, measured_square_sum_remote_service_demand=0.0, measured_mean_remote_service_demand=0.0, measured_var_remote_service_demand=0.0, measured_sum_local_time=0.0, measured_square_sum_local_time=0.0, measured_mean_local_time=0.0, measured_var_local_time=0.0, measured_mean_remote_time=0.0, measured_fails, measured_local_results, confidence=-10.0; /* interval=0.1; */ /************************************************************************/ /* */ /* Constants for Confidence Intervals */ /* */ /************************************************************************/ void init_stat() { measured_sum_result=0.0; measured_square_sum_result=0.0; measured_mean_result=0.0; measured_var_result=0.0; measured_sum_local_cpu=0.0; measured_square_sum_local_cpu=0.0; measured_mean_local_cpu=0.0; measured_var_local_cpu=0.0; measured_sum_remote_cpu=0.0; measured_square_sum_remote_cpu=0.0; measured_mean_remote_cpu=0.0; measured_var_remote_cpu=0.0; measured_sum_local_service_demand=0.0; measured_square_sum_local_service_demand=0.0; measured_mean_local_service_demand=0.0; measured_var_local_service_demand=0.0; measured_sum_remote_service_demand=0.0; measured_square_sum_remote_service_demand=0.0; measured_mean_remote_service_demand=0.0; measured_var_remote_service_demand=0.0; measured_sum_local_time=0.0; measured_square_sum_local_time=0.0; measured_mean_local_time=0.0; measured_var_local_time=0.0; measured_mean_remote_time=0.0; measured_fails = 0.0; measured_local_results=0.0, confidence=-10.0; } /* this routine does a simple table lookup for some statistical function that I would remember if I stayed awake in my probstats class... raj 11/94 */ double confid(int level, int freedom) { double t99[35],t95[35]; t95[1]=12.706; t95[2]= 4.303; t95[3]= 3.182; t95[4]= 2.776; t95[5]= 2.571; t95[6]= 2.447; t95[7]= 2.365; t95[8]= 2.306; t95[9]= 2.262; t95[10]= 2.228; t95[11]= 2.201; t95[12]= 2.179; t95[13]= 2.160; t95[14]= 2.145; t95[15]= 2.131; t95[16]= 2.120; t95[17]= 2.110; t95[18]= 2.101; t95[19]= 2.093; t95[20]= 2.086; t95[21]= 2.080; t95[22]= 2.074; t95[23]= 2.069; t95[24]= 2.064; t95[25]= 2.060; t95[26]= 2.056; t95[27]= 2.052; t95[28]= 2.048; t95[29]= 2.045; t95[30]= 2.042; t99[1]=63.657; t99[2]= 9.925; t99[3]= 5.841; t99[4]= 4.604; t99[5]= 4.032; t99[6]= 3.707; t99[7]= 3.499; t99[8]= 3.355; t99[9]= 3.250; t99[10]= 3.169; t99[11]= 3.106; t99[12]= 3.055; t99[13]= 3.012; t99[14]= 2.977; t99[15]= 2.947; t99[16]= 2.921; t99[17]= 2.898; t99[18]= 2.878; t99[19]= 2.861; t99[20]= 2.845; t99[21]= 2.831; t99[22]= 2.819; t99[23]= 2.807; t99[24]= 2.797; t99[25]= 2.787; t99[26]= 2.779; t99[27]= 2.771; t99[28]= 2.763; t99[29]= 2.756; t99[30]= 2.750; if(level==95){ return(t95[freedom]); } else if(level==99){ return(t99[freedom]); } else{ return(0); } } void calculate_confidence(int confidence_iterations, float time, double result, float loc_cpu, float rem_cpu, float loc_sd, float rem_sd) { if (debug) { fprintf(where, "calculate_confidence: itr %d; time %f; res %f\n" " lcpu %f; rcpu %f\n" " lsdm %f; rsdm %f\n", confidence_iterations, time, result, loc_cpu, rem_cpu, loc_sd, rem_sd); fflush(where); } /* the test time */ measured_sum_local_time += (double) time; measured_square_sum_local_time += (double) time*time; measured_mean_local_time = (double) measured_sum_local_time/confidence_iterations; measured_var_local_time = (double) measured_square_sum_local_time/confidence_iterations -measured_mean_local_time*measured_mean_local_time; /* the test result */ measured_sum_result += (double) result; measured_square_sum_result += (double) result*result; measured_mean_result = (double) measured_sum_result/confidence_iterations; measured_var_result = (double) measured_square_sum_result/confidence_iterations -measured_mean_result*measured_mean_result; /* local cpu utilization */ measured_sum_local_cpu += (double) loc_cpu; measured_square_sum_local_cpu += (double) loc_cpu*loc_cpu; measured_mean_local_cpu = (double) measured_sum_local_cpu/confidence_iterations; measured_var_local_cpu = (double) measured_square_sum_local_cpu/confidence_iterations -measured_mean_local_cpu*measured_mean_local_cpu; /* remote cpu util */ measured_sum_remote_cpu += (double) rem_cpu; measured_square_sum_remote_cpu+= (double) rem_cpu*rem_cpu; measured_mean_remote_cpu = (double) measured_sum_remote_cpu/confidence_iterations; measured_var_remote_cpu = (double) measured_square_sum_remote_cpu/confidence_iterations -measured_mean_remote_cpu*measured_mean_remote_cpu; /* local service demand */ measured_sum_local_service_demand += (double) loc_sd; measured_square_sum_local_service_demand+= (double) loc_sd*loc_sd; measured_mean_local_service_demand = (double) measured_sum_local_service_demand/confidence_iterations; measured_var_local_service_demand = (double) measured_square_sum_local_service_demand/confidence_iterations -measured_mean_local_service_demand*measured_mean_local_service_demand; /* remote service demand */ measured_sum_remote_service_demand += (double) rem_sd; measured_square_sum_remote_service_demand+= (double) rem_sd*rem_sd; measured_mean_remote_service_demand = (double) measured_sum_remote_service_demand/confidence_iterations; measured_var_remote_service_demand = (double) measured_square_sum_remote_service_demand/confidence_iterations -measured_mean_remote_service_demand*measured_mean_remote_service_demand; if(confidence_iterations>1){ result_confid= (double) interval - 2.0 * confid(confidence_level,confidence_iterations-1)* sqrt(measured_var_result/(confidence_iterations-1.0)) / measured_mean_result; loc_cpu_confid= (double) interval - 2.0 * confid(confidence_level,confidence_iterations-1)* sqrt(measured_var_local_cpu/(confidence_iterations-1.0)) / measured_mean_local_cpu; rem_cpu_confid= (double) interval - 2.0 * confid(confidence_level,confidence_iterations-1)* sqrt(measured_var_remote_cpu/(confidence_iterations-1.0)) / measured_mean_remote_cpu; if(debug){ printf("Conf_itvl %2d: results:%4.1f%% loc_cpu:%4.1f%% rem_cpu:%4.1f%%\n", confidence_iterations, (interval-result_confid)*100.0, (interval-loc_cpu_confid)*100.0, (interval-rem_cpu_confid)*100.0); } /* if the user has requested that we only wait for the result to be confident rather than the result and CPU util(s) then do so. raj 2007-08-08 */ if (!result_confidence_only) { confidence = min(min(result_confid,loc_cpu_confid),rem_cpu_confid); } else { confidence = result_confid; } } } /* here ends the IBM code */ void retrieve_confident_values(float *elapsed_time, double *thruput, float *local_cpu_utilization, float *remote_cpu_utilization, float *local_service_demand, float *remote_service_demand) { *elapsed_time = (float)measured_mean_local_time; *thruput = measured_mean_result; *local_cpu_utilization = (float)measured_mean_local_cpu; *remote_cpu_utilization = (float)measured_mean_remote_cpu; *local_service_demand = (float)measured_mean_local_service_demand; *remote_service_demand = (float)measured_mean_remote_service_demand; } double get_result_confid() { return (double) (100.0 * (interval - result_confid)); } double get_loc_cpu_confid() { return (double) (100.0 * (interval - loc_cpu_confid)); } double get_rem_cpu_confid() { return (double) (100.0 * (interval - rem_cpu_confid)); } /* display_confidence() is called when we could not achieve the desired confidence in the results. it will print the achieved confidence to "where" raj 11/94 */ void display_confidence() { fprintf(where, "!!! WARNING\n" "!!! Desired confidence was not achieved within " "the specified iterations.\n" "!!! This implies that there was variability in " "the test environment that\n" "!!! must be investigated before going further.\n" "!!! Confidence intervals: Throughput : %4.3f%%\n" "!!! Local CPU util : %4.3f%%\n" "!!! Remote CPU util : %4.3f%%\n\n", 100.0 * (interval - result_confid), 100.0 * (interval - loc_cpu_confid), 100.0 * (interval - rem_cpu_confid)); }