hwpmc_mod.c revision 156466
1145256Sjkoshy/*- 2145256Sjkoshy * Copyright (c) 2003-2005 Joseph Koshy 3145256Sjkoshy * All rights reserved. 4145256Sjkoshy * 5145256Sjkoshy * Redistribution and use in source and binary forms, with or without 6145256Sjkoshy * modification, are permitted provided that the following conditions 7145256Sjkoshy * are met: 8145256Sjkoshy * 1. Redistributions of source code must retain the above copyright 9145256Sjkoshy * notice, this list of conditions and the following disclaimer. 10145256Sjkoshy * 2. Redistributions in binary form must reproduce the above copyright 11145256Sjkoshy * notice, this list of conditions and the following disclaimer in the 12145256Sjkoshy * documentation and/or other materials provided with the distribution. 13145256Sjkoshy * 14145256Sjkoshy * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15145256Sjkoshy * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16145256Sjkoshy * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17145256Sjkoshy * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18145256Sjkoshy * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19145256Sjkoshy * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20145256Sjkoshy * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21145256Sjkoshy * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22145256Sjkoshy * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23145256Sjkoshy * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24145256Sjkoshy * SUCH DAMAGE. 25145256Sjkoshy * 26145256Sjkoshy */ 27145256Sjkoshy 28145256Sjkoshy#include <sys/cdefs.h> 29145256Sjkoshy__FBSDID("$FreeBSD: head/sys/dev/hwpmc/hwpmc_mod.c 156466 2006-03-09 02:08:12Z jkoshy $"); 30145256Sjkoshy 31145256Sjkoshy#include <sys/param.h> 32145256Sjkoshy#include <sys/eventhandler.h> 33145256Sjkoshy#include <sys/jail.h> 34145256Sjkoshy#include <sys/kernel.h> 35147191Sjkoshy#include <sys/kthread.h> 36145256Sjkoshy#include <sys/limits.h> 37145256Sjkoshy#include <sys/lock.h> 38145256Sjkoshy#include <sys/malloc.h> 39145256Sjkoshy#include <sys/module.h> 40145256Sjkoshy#include <sys/mutex.h> 41145256Sjkoshy#include <sys/pmc.h> 42145256Sjkoshy#include <sys/pmckern.h> 43147191Sjkoshy#include <sys/pmclog.h> 44145256Sjkoshy#include <sys/proc.h> 45145256Sjkoshy#include <sys/queue.h> 46147191Sjkoshy#include <sys/resourcevar.h> 47145256Sjkoshy#include <sys/sched.h> 48145256Sjkoshy#include <sys/signalvar.h> 49145256Sjkoshy#include <sys/smp.h> 50145256Sjkoshy#include <sys/sx.h> 51145256Sjkoshy#include <sys/sysctl.h> 52145256Sjkoshy#include <sys/sysent.h> 53145256Sjkoshy#include <sys/systm.h> 54147191Sjkoshy#include <sys/vnode.h> 55145256Sjkoshy 56147191Sjkoshy#include <machine/atomic.h> 57145256Sjkoshy#include <machine/md_var.h> 58145256Sjkoshy 59145256Sjkoshy/* 60145256Sjkoshy * Types 61145256Sjkoshy */ 62145256Sjkoshy 63145256Sjkoshyenum pmc_flags { 64145256Sjkoshy PMC_FLAG_NONE = 0x00, /* do nothing */ 65145256Sjkoshy PMC_FLAG_REMOVE = 0x01, /* atomically remove entry from hash */ 66145256Sjkoshy PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */ 67145256Sjkoshy}; 68145256Sjkoshy 69145256Sjkoshy/* 70145256Sjkoshy * The offset in sysent where the syscall is allocated. 71145256Sjkoshy */ 72145256Sjkoshy 73145256Sjkoshystatic int pmc_syscall_num = NO_SYSCALL; 74145256Sjkoshystruct pmc_cpu **pmc_pcpu; /* per-cpu state */ 75145256Sjkoshypmc_value_t *pmc_pcpu_saved; /* saved PMC values: CSW handling */ 76145256Sjkoshy 77145256Sjkoshy#define PMC_PCPU_SAVED(C,R) pmc_pcpu_saved[(R) + md->pmd_npmc*(C)] 78145256Sjkoshy 79145256Sjkoshystruct mtx_pool *pmc_mtxpool; 80145256Sjkoshystatic int *pmc_pmcdisp; /* PMC row dispositions */ 81145256Sjkoshy 82145256Sjkoshy#define PMC_ROW_DISP_IS_FREE(R) (pmc_pmcdisp[(R)] == 0) 83145256Sjkoshy#define PMC_ROW_DISP_IS_THREAD(R) (pmc_pmcdisp[(R)] > 0) 84145256Sjkoshy#define PMC_ROW_DISP_IS_STANDALONE(R) (pmc_pmcdisp[(R)] < 0) 85145256Sjkoshy 86145256Sjkoshy#define PMC_MARK_ROW_FREE(R) do { \ 87145256Sjkoshy pmc_pmcdisp[(R)] = 0; \ 88145256Sjkoshy} while (0) 89145256Sjkoshy 90145256Sjkoshy#define PMC_MARK_ROW_STANDALONE(R) do { \ 91145256Sjkoshy KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 92145256Sjkoshy __LINE__)); \ 93145256Sjkoshy atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 94145256Sjkoshy KASSERT(pmc_pmcdisp[(R)] >= (-mp_ncpus), ("[pmc,%d] row " \ 95145256Sjkoshy "disposition error", __LINE__)); \ 96145256Sjkoshy} while (0) 97145256Sjkoshy 98145256Sjkoshy#define PMC_UNMARK_ROW_STANDALONE(R) do { \ 99145256Sjkoshy atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 100145256Sjkoshy KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 101145256Sjkoshy __LINE__)); \ 102145256Sjkoshy} while (0) 103145256Sjkoshy 104145256Sjkoshy#define PMC_MARK_ROW_THREAD(R) do { \ 105145256Sjkoshy KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 106145256Sjkoshy __LINE__)); \ 107145256Sjkoshy atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 108145256Sjkoshy} while (0) 109145256Sjkoshy 110145256Sjkoshy#define PMC_UNMARK_ROW_THREAD(R) do { \ 111145256Sjkoshy atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 112145256Sjkoshy KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 113145256Sjkoshy __LINE__)); \ 114145256Sjkoshy} while (0) 115145256Sjkoshy 116145256Sjkoshy 117145256Sjkoshy/* various event handlers */ 118145256Sjkoshystatic eventhandler_tag pmc_exit_tag, pmc_fork_tag; 119145256Sjkoshy 120145256Sjkoshy/* Module statistics */ 121145256Sjkoshystruct pmc_op_getdriverstats pmc_stats; 122145256Sjkoshy 123145256Sjkoshy/* Machine/processor dependent operations */ 124145256Sjkoshystruct pmc_mdep *md; 125145256Sjkoshy 126145256Sjkoshy/* 127145256Sjkoshy * Hash tables mapping owner processes and target threads to PMCs. 128145256Sjkoshy */ 129145256Sjkoshy 130145256Sjkoshystruct mtx pmc_processhash_mtx; /* spin mutex */ 131145256Sjkoshystatic u_long pmc_processhashmask; 132145256Sjkoshystatic LIST_HEAD(pmc_processhash, pmc_process) *pmc_processhash; 133145256Sjkoshy 134145256Sjkoshy/* 135145256Sjkoshy * Hash table of PMC owner descriptors. This table is protected by 136145256Sjkoshy * the shared PMC "sx" lock. 137145256Sjkoshy */ 138145256Sjkoshy 139145256Sjkoshystatic u_long pmc_ownerhashmask; 140145256Sjkoshystatic LIST_HEAD(pmc_ownerhash, pmc_owner) *pmc_ownerhash; 141145256Sjkoshy 142145256Sjkoshy/* 143147191Sjkoshy * List of PMC owners with system-wide sampling PMCs. 144147191Sjkoshy */ 145147191Sjkoshy 146147191Sjkoshystatic LIST_HEAD(, pmc_owner) pmc_ss_owners; 147147191Sjkoshy 148147191Sjkoshy 149147191Sjkoshy/* 150145256Sjkoshy * Prototypes 151145256Sjkoshy */ 152145256Sjkoshy 153153110Sru#ifdef DEBUG 154145256Sjkoshystatic int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS); 155145256Sjkoshystatic int pmc_debugflags_parse(char *newstr, char *fence); 156145256Sjkoshy#endif 157145256Sjkoshy 158145256Sjkoshystatic int load(struct module *module, int cmd, void *arg); 159147191Sjkoshystatic int pmc_attach_process(struct proc *p, struct pmc *pm); 160145256Sjkoshystatic struct pmc *pmc_allocate_pmc_descriptor(void); 161147191Sjkoshystatic struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p); 162147191Sjkoshystatic int pmc_attach_one_process(struct proc *p, struct pmc *pm); 163147191Sjkoshystatic int pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, 164147191Sjkoshy int cpu); 165147191Sjkoshystatic int pmc_can_attach(struct pmc *pm, struct proc *p); 166147191Sjkoshystatic void pmc_cleanup(void); 167147191Sjkoshystatic int pmc_detach_process(struct proc *p, struct pmc *pm); 168147191Sjkoshystatic int pmc_detach_one_process(struct proc *p, struct pmc *pm, 169147191Sjkoshy int flags); 170147191Sjkoshystatic void pmc_destroy_owner_descriptor(struct pmc_owner *po); 171147191Sjkoshystatic struct pmc_owner *pmc_find_owner_descriptor(struct proc *p); 172147191Sjkoshystatic int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm); 173145256Sjkoshystatic struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, 174145256Sjkoshy pmc_id_t pmc); 175145256Sjkoshystatic struct pmc_process *pmc_find_process_descriptor(struct proc *p, 176145256Sjkoshy uint32_t mode); 177145774Sjkoshystatic void pmc_force_context_switch(void); 178145256Sjkoshystatic void pmc_link_target_process(struct pmc *pm, 179145256Sjkoshy struct pmc_process *pp); 180147191Sjkoshystatic void pmc_maybe_remove_owner(struct pmc_owner *po); 181147191Sjkoshystatic void pmc_process_csw_in(struct thread *td); 182147191Sjkoshystatic void pmc_process_csw_out(struct thread *td); 183145256Sjkoshystatic void pmc_process_exit(void *arg, struct proc *p); 184145256Sjkoshystatic void pmc_process_fork(void *arg, struct proc *p1, 185145256Sjkoshy struct proc *p2, int n); 186147191Sjkoshystatic void pmc_process_samples(int cpu); 187147191Sjkoshystatic void pmc_release_pmc_descriptor(struct pmc *pmc); 188147191Sjkoshystatic void pmc_remove_owner(struct pmc_owner *po); 189147191Sjkoshystatic void pmc_remove_process_descriptor(struct pmc_process *pp); 190147191Sjkoshystatic void pmc_restore_cpu_binding(struct pmc_binding *pb); 191147191Sjkoshystatic void pmc_save_cpu_binding(struct pmc_binding *pb); 192147191Sjkoshystatic void pmc_select_cpu(int cpu); 193145256Sjkoshystatic int pmc_start(struct pmc *pm); 194145256Sjkoshystatic int pmc_stop(struct pmc *pm); 195147191Sjkoshystatic int pmc_syscall_handler(struct thread *td, void *syscall_args); 196147191Sjkoshystatic void pmc_unlink_target_process(struct pmc *pmc, 197147191Sjkoshy struct pmc_process *pp); 198145256Sjkoshy 199145256Sjkoshy/* 200145256Sjkoshy * Kernel tunables and sysctl(8) interface. 201145256Sjkoshy */ 202145256Sjkoshy 203145256SjkoshySYSCTL_NODE(_kern, OID_AUTO, hwpmc, CTLFLAG_RW, 0, "HWPMC parameters"); 204145256Sjkoshy 205153110Sru#ifdef DEBUG 206147191Sjkoshystruct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS; 207145256Sjkoshychar pmc_debugstr[PMC_DEBUG_STRSIZE]; 208145256SjkoshyTUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr, 209145256Sjkoshy sizeof(pmc_debugstr)); 210145256SjkoshySYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags, 211145256Sjkoshy CTLTYPE_STRING|CTLFLAG_RW|CTLFLAG_TUN, 212145256Sjkoshy 0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags"); 213145256Sjkoshy#endif 214145256Sjkoshy 215145256Sjkoshy/* 216147191Sjkoshy * kern.hwpmc.hashrows -- determines the number of rows in the 217145256Sjkoshy * of the hash table used to look up threads 218145256Sjkoshy */ 219145256Sjkoshy 220145256Sjkoshystatic int pmc_hashsize = PMC_HASH_SIZE; 221145256SjkoshyTUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "hashsize", &pmc_hashsize); 222145256SjkoshySYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_TUN|CTLFLAG_RD, 223145256Sjkoshy &pmc_hashsize, 0, "rows in hash tables"); 224145256Sjkoshy 225145256Sjkoshy/* 226147191Sjkoshy * kern.hwpmc.nsamples --- number of PC samples per CPU 227145256Sjkoshy */ 228145256Sjkoshy 229147191Sjkoshystatic int pmc_nsamples = PMC_NSAMPLES; 230147191SjkoshyTUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nsamples", &pmc_nsamples); 231147191SjkoshySYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_TUN|CTLFLAG_RD, 232147191Sjkoshy &pmc_nsamples, 0, "number of PC samples per CPU"); 233145256Sjkoshy 234145256Sjkoshy/* 235147191Sjkoshy * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool. 236145256Sjkoshy */ 237145256Sjkoshy 238145256Sjkoshystatic int pmc_mtxpool_size = PMC_MTXPOOL_SIZE; 239145256SjkoshyTUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "mtxpoolsize", &pmc_mtxpool_size); 240145256SjkoshySYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_TUN|CTLFLAG_RD, 241145256Sjkoshy &pmc_mtxpool_size, 0, "size of spin mutex pool"); 242145256Sjkoshy 243145256Sjkoshy 244145256Sjkoshy/* 245145256Sjkoshy * security.bsd.unprivileged_syspmcs -- allow non-root processes to 246145256Sjkoshy * allocate system-wide PMCs. 247145256Sjkoshy * 248145256Sjkoshy * Allowing unprivileged processes to allocate system PMCs is convenient 249145256Sjkoshy * if system-wide measurements need to be taken concurrently with other 250145256Sjkoshy * per-process measurements. This feature is turned off by default. 251145256Sjkoshy */ 252145256Sjkoshy 253145256SjkoshySYSCTL_DECL(_security_bsd); 254145256Sjkoshy 255145256Sjkoshystatic int pmc_unprivileged_syspmcs = 0; 256145256SjkoshyTUNABLE_INT("security.bsd.unprivileged_syspmcs", &pmc_unprivileged_syspmcs); 257145256SjkoshySYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RW, 258145256Sjkoshy &pmc_unprivileged_syspmcs, 0, 259145256Sjkoshy "allow unprivileged process to allocate system PMCs"); 260145256Sjkoshy 261147191Sjkoshy/* 262147191Sjkoshy * Hash function. Discard the lower 2 bits of the pointer since 263147191Sjkoshy * these are always zero for our uses. The hash multiplier is 264147191Sjkoshy * round((2^LONG_BIT) * ((sqrt(5)-1)/2)). 265147191Sjkoshy */ 266145256Sjkoshy 267145256Sjkoshy#if LONG_BIT == 64 268145256Sjkoshy#define _PMC_HM 11400714819323198486u 269145256Sjkoshy#elif LONG_BIT == 32 270145256Sjkoshy#define _PMC_HM 2654435769u 271145256Sjkoshy#else 272145256Sjkoshy#error Must know the size of 'long' to compile 273145256Sjkoshy#endif 274145256Sjkoshy 275145256Sjkoshy#define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M)) 276145256Sjkoshy 277145256Sjkoshy/* 278145256Sjkoshy * Syscall structures 279145256Sjkoshy */ 280145256Sjkoshy 281145256Sjkoshy/* The `sysent' for the new syscall */ 282145256Sjkoshystatic struct sysent pmc_sysent = { 283145256Sjkoshy 2, /* sy_narg */ 284145256Sjkoshy pmc_syscall_handler /* sy_call */ 285145256Sjkoshy}; 286145256Sjkoshy 287145256Sjkoshystatic struct syscall_module_data pmc_syscall_mod = { 288145256Sjkoshy load, 289145256Sjkoshy NULL, 290145256Sjkoshy &pmc_syscall_num, 291145256Sjkoshy &pmc_sysent, 292145256Sjkoshy { 0, NULL } 293145256Sjkoshy}; 294145256Sjkoshy 295145256Sjkoshystatic moduledata_t pmc_mod = { 296145256Sjkoshy PMC_MODULE_NAME, 297145256Sjkoshy syscall_module_handler, 298145256Sjkoshy &pmc_syscall_mod 299145256Sjkoshy}; 300145256Sjkoshy 301145256SjkoshyDECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY); 302145256SjkoshyMODULE_VERSION(pmc, PMC_VERSION); 303145256Sjkoshy 304153110Sru#ifdef DEBUG 305147191Sjkoshyenum pmc_dbgparse_state { 306147191Sjkoshy PMCDS_WS, /* in whitespace */ 307147191Sjkoshy PMCDS_MAJOR, /* seen a major keyword */ 308147191Sjkoshy PMCDS_MINOR 309147191Sjkoshy}; 310147191Sjkoshy 311145256Sjkoshystatic int 312145256Sjkoshypmc_debugflags_parse(char *newstr, char *fence) 313145256Sjkoshy{ 314145313Sjkoshy char c, *p, *q; 315147191Sjkoshy struct pmc_debugflags *tmpflags; 316147191Sjkoshy int error, found, *newbits, tmp; 317147191Sjkoshy size_t kwlen; 318145256Sjkoshy 319147191Sjkoshy MALLOC(tmpflags, struct pmc_debugflags *, sizeof(*tmpflags), 320147191Sjkoshy M_PMC, M_WAITOK|M_ZERO); 321145256Sjkoshy 322145256Sjkoshy p = newstr; 323147191Sjkoshy error = 0; 324145256Sjkoshy 325147191Sjkoshy for (; p < fence && (c = *p); p++) { 326145256Sjkoshy 327147191Sjkoshy /* skip white space */ 328147191Sjkoshy if (c == ' ' || c == '\t') 329147191Sjkoshy continue; 330147191Sjkoshy 331147191Sjkoshy /* look for a keyword followed by "=" */ 332147191Sjkoshy for (q = p; p < fence && (c = *p) && c != '='; p++) 333147191Sjkoshy ; 334147191Sjkoshy if (c != '=') { 335147191Sjkoshy error = EINVAL; 336147191Sjkoshy goto done; 337145256Sjkoshy } 338145256Sjkoshy 339147191Sjkoshy kwlen = p - q; 340147191Sjkoshy newbits = NULL; 341145256Sjkoshy 342147191Sjkoshy /* lookup flag group name */ 343147191Sjkoshy#define DBG_SET_FLAG_MAJ(S,F) \ 344147191Sjkoshy if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 345147191Sjkoshy newbits = &tmpflags->pdb_ ## F; 346145256Sjkoshy 347147191Sjkoshy DBG_SET_FLAG_MAJ("cpu", CPU); 348147191Sjkoshy DBG_SET_FLAG_MAJ("csw", CSW); 349147191Sjkoshy DBG_SET_FLAG_MAJ("logging", LOG); 350147191Sjkoshy DBG_SET_FLAG_MAJ("module", MOD); 351147191Sjkoshy DBG_SET_FLAG_MAJ("md", MDP); 352147191Sjkoshy DBG_SET_FLAG_MAJ("owner", OWN); 353147191Sjkoshy DBG_SET_FLAG_MAJ("pmc", PMC); 354147191Sjkoshy DBG_SET_FLAG_MAJ("process", PRC); 355147191Sjkoshy DBG_SET_FLAG_MAJ("sampling", SAM); 356145256Sjkoshy 357147191Sjkoshy if (newbits == NULL) { 358147191Sjkoshy error = EINVAL; 359147191Sjkoshy goto done; 360145256Sjkoshy } 361145256Sjkoshy 362147191Sjkoshy p++; /* skip the '=' */ 363145256Sjkoshy 364147191Sjkoshy /* Now parse the individual flags */ 365147191Sjkoshy tmp = 0; 366147191Sjkoshy newflag: 367147191Sjkoshy for (q = p; p < fence && (c = *p); p++) 368147191Sjkoshy if (c == ' ' || c == '\t' || c == ',') 369147191Sjkoshy break; 370147191Sjkoshy 371147191Sjkoshy /* p == fence or c == ws or c == "," or c == 0 */ 372147191Sjkoshy 373147191Sjkoshy if ((kwlen = p - q) == 0) { 374147191Sjkoshy *newbits = tmp; 375147191Sjkoshy continue; 376147191Sjkoshy } 377147191Sjkoshy 378147191Sjkoshy found = 0; 379147191Sjkoshy#define DBG_SET_FLAG_MIN(S,F) \ 380147191Sjkoshy if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 381147191Sjkoshy tmp |= found = (1 << PMC_DEBUG_MIN_ ## F) 382147191Sjkoshy 383147191Sjkoshy /* a '*' denotes all possible flags in the group */ 384147191Sjkoshy if (kwlen == 1 && *q == '*') 385147191Sjkoshy tmp = found = ~0; 386147191Sjkoshy /* look for individual flag names */ 387147191Sjkoshy DBG_SET_FLAG_MIN("allocaterow", ALR); 388147191Sjkoshy DBG_SET_FLAG_MIN("allocate", ALL); 389147191Sjkoshy DBG_SET_FLAG_MIN("attach", ATT); 390147191Sjkoshy DBG_SET_FLAG_MIN("bind", BND); 391147191Sjkoshy DBG_SET_FLAG_MIN("config", CFG); 392147191Sjkoshy DBG_SET_FLAG_MIN("exec", EXC); 393147191Sjkoshy DBG_SET_FLAG_MIN("exit", EXT); 394147191Sjkoshy DBG_SET_FLAG_MIN("find", FND); 395147191Sjkoshy DBG_SET_FLAG_MIN("flush", FLS); 396147191Sjkoshy DBG_SET_FLAG_MIN("fork", FRK); 397147191Sjkoshy DBG_SET_FLAG_MIN("getbuf", GTB); 398147191Sjkoshy DBG_SET_FLAG_MIN("hook", PMH); 399147191Sjkoshy DBG_SET_FLAG_MIN("init", INI); 400147191Sjkoshy DBG_SET_FLAG_MIN("intr", INT); 401147191Sjkoshy DBG_SET_FLAG_MIN("linktarget", TLK); 402147191Sjkoshy DBG_SET_FLAG_MIN("mayberemove", OMR); 403147191Sjkoshy DBG_SET_FLAG_MIN("ops", OPS); 404147191Sjkoshy DBG_SET_FLAG_MIN("read", REA); 405147191Sjkoshy DBG_SET_FLAG_MIN("register", REG); 406147191Sjkoshy DBG_SET_FLAG_MIN("release", REL); 407147191Sjkoshy DBG_SET_FLAG_MIN("remove", ORM); 408147191Sjkoshy DBG_SET_FLAG_MIN("sample", SAM); 409147191Sjkoshy DBG_SET_FLAG_MIN("scheduleio", SIO); 410147191Sjkoshy DBG_SET_FLAG_MIN("select", SEL); 411147191Sjkoshy DBG_SET_FLAG_MIN("signal", SIG); 412147191Sjkoshy DBG_SET_FLAG_MIN("swi", SWI); 413147191Sjkoshy DBG_SET_FLAG_MIN("swo", SWO); 414147191Sjkoshy DBG_SET_FLAG_MIN("start", STA); 415147191Sjkoshy DBG_SET_FLAG_MIN("stop", STO); 416147191Sjkoshy DBG_SET_FLAG_MIN("syscall", PMS); 417147191Sjkoshy DBG_SET_FLAG_MIN("unlinktarget", TUL); 418147191Sjkoshy DBG_SET_FLAG_MIN("write", WRI); 419147191Sjkoshy if (found == 0) { 420147191Sjkoshy /* unrecognized flag name */ 421147191Sjkoshy error = EINVAL; 422147191Sjkoshy goto done; 423147191Sjkoshy } 424147191Sjkoshy 425147191Sjkoshy if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */ 426147191Sjkoshy *newbits = tmp; 427147191Sjkoshy continue; 428147191Sjkoshy } 429147191Sjkoshy 430147191Sjkoshy p++; 431147191Sjkoshy goto newflag; 432145256Sjkoshy } 433145256Sjkoshy 434147191Sjkoshy /* save the new flag set */ 435147191Sjkoshy bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags)); 436145256Sjkoshy 437147191Sjkoshy done: 438147191Sjkoshy FREE(tmpflags, M_PMC); 439147191Sjkoshy return error; 440145256Sjkoshy} 441145256Sjkoshy 442145256Sjkoshystatic int 443145256Sjkoshypmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS) 444145256Sjkoshy{ 445145256Sjkoshy char *fence, *newstr; 446145256Sjkoshy int error; 447145256Sjkoshy unsigned int n; 448145256Sjkoshy 449145256Sjkoshy (void) arg1; (void) arg2; /* unused parameters */ 450145256Sjkoshy 451145256Sjkoshy n = sizeof(pmc_debugstr); 452145256Sjkoshy MALLOC(newstr, char *, n, M_PMC, M_ZERO|M_WAITOK); 453147191Sjkoshy (void) strlcpy(newstr, pmc_debugstr, n); 454145256Sjkoshy 455145256Sjkoshy error = sysctl_handle_string(oidp, newstr, n, req); 456145256Sjkoshy 457145256Sjkoshy /* if there is a new string, parse and copy it */ 458145256Sjkoshy if (error == 0 && req->newptr != NULL) { 459147191Sjkoshy fence = newstr + (n < req->newlen ? n : req->newlen + 1); 460145256Sjkoshy if ((error = pmc_debugflags_parse(newstr, fence)) == 0) 461145256Sjkoshy (void) strlcpy(pmc_debugstr, newstr, 462145256Sjkoshy sizeof(pmc_debugstr)); 463145256Sjkoshy } 464145256Sjkoshy 465145256Sjkoshy FREE(newstr, M_PMC); 466145256Sjkoshy 467145256Sjkoshy return error; 468145256Sjkoshy} 469145256Sjkoshy#endif 470145256Sjkoshy 471145256Sjkoshy/* 472145256Sjkoshy * Concurrency Control 473145256Sjkoshy * 474145256Sjkoshy * The driver manages the following data structures: 475145256Sjkoshy * 476145256Sjkoshy * - target process descriptors, one per target process 477145256Sjkoshy * - owner process descriptors (and attached lists), one per owner process 478145256Sjkoshy * - lookup hash tables for owner and target processes 479145256Sjkoshy * - PMC descriptors (and attached lists) 480145256Sjkoshy * - per-cpu hardware state 481145256Sjkoshy * - the 'hook' variable through which the kernel calls into 482145256Sjkoshy * this module 483145256Sjkoshy * - the machine hardware state (managed by the MD layer) 484145256Sjkoshy * 485145256Sjkoshy * These data structures are accessed from: 486145256Sjkoshy * 487145256Sjkoshy * - thread context-switch code 488145256Sjkoshy * - interrupt handlers (possibly on multiple cpus) 489145256Sjkoshy * - kernel threads on multiple cpus running on behalf of user 490145256Sjkoshy * processes doing system calls 491145256Sjkoshy * - this driver's private kernel threads 492145256Sjkoshy * 493145256Sjkoshy * = Locks and Locking strategy = 494145256Sjkoshy * 495145256Sjkoshy * The driver uses four locking strategies for its operation: 496145256Sjkoshy * 497145256Sjkoshy * - There is a 'global' SX lock "pmc_sx" that is used to protect 498145256Sjkoshy * the its 'meta-data'. 499145256Sjkoshy * 500145256Sjkoshy * Calls into the module (via syscall() or by the kernel) start with 501145256Sjkoshy * this lock being held in exclusive mode. Depending on the requested 502145256Sjkoshy * operation, the lock may be downgraded to 'shared' mode to allow 503145256Sjkoshy * more concurrent readers into the module. 504145256Sjkoshy * 505145256Sjkoshy * This SX lock is held in exclusive mode for any operations that 506145256Sjkoshy * modify the linkages between the driver's internal data structures. 507145256Sjkoshy * 508145256Sjkoshy * The 'pmc_hook' function pointer is also protected by this lock. 509145256Sjkoshy * It is only examined with the sx lock held in exclusive mode. The 510145256Sjkoshy * kernel module is allowed to be unloaded only with the sx lock 511145256Sjkoshy * held in exclusive mode. In normal syscall handling, after 512145256Sjkoshy * acquiring the pmc_sx lock we first check that 'pmc_hook' is 513145256Sjkoshy * non-null before proceeding. This prevents races between the 514145256Sjkoshy * thread unloading the module and other threads seeking to use the 515145256Sjkoshy * module. 516145256Sjkoshy * 517145256Sjkoshy * - Lookups of target process structures and owner process structures 518145256Sjkoshy * cannot use the global "pmc_sx" SX lock because these lookups need 519145256Sjkoshy * to happen during context switches and in other critical sections 520145256Sjkoshy * where sleeping is not allowed. We protect these lookup tables 521145256Sjkoshy * with their own private spin-mutexes, "pmc_processhash_mtx" and 522145256Sjkoshy * "pmc_ownerhash_mtx". These are 'leaf' mutexes, in that no other 523145256Sjkoshy * lock is acquired with these locks held. 524145256Sjkoshy * 525145256Sjkoshy * - Interrupt handlers work in a lock free manner. At interrupt 526145256Sjkoshy * time, handlers look at the PMC pointer (phw->phw_pmc) configured 527145256Sjkoshy * when the PMC was started. If this pointer is NULL, the interrupt 528145256Sjkoshy * is ignored after updating driver statistics. We ensure that this 529145256Sjkoshy * pointer is set (using an atomic operation if necessary) before the 530145256Sjkoshy * PMC hardware is started. Conversely, this pointer is unset atomically 531145256Sjkoshy * only after the PMC hardware is stopped. 532145256Sjkoshy * 533145256Sjkoshy * We ensure that everything needed for the operation of an 534145256Sjkoshy * interrupt handler is available without it needing to acquire any 535145256Sjkoshy * locks. We also ensure that a PMC's software state is destroyed only 536145256Sjkoshy * after the PMC is taken off hardware (on all CPUs). 537145256Sjkoshy * 538145256Sjkoshy * - Context-switch handling with process-private PMCs needs more 539145256Sjkoshy * care. 540145256Sjkoshy * 541145256Sjkoshy * A given process may be the target of multiple PMCs. For example, 542145256Sjkoshy * PMCATTACH and PMCDETACH may be requested by a process on one CPU 543145256Sjkoshy * while the target process is running on another. A PMC could also 544145256Sjkoshy * be getting released because its owner is exiting. We tackle 545145256Sjkoshy * these situations in the following manner: 546145256Sjkoshy * 547145256Sjkoshy * - each target process structure 'pmc_process' has an array 548145256Sjkoshy * of 'struct pmc *' pointers, one for each hardware PMC. 549145256Sjkoshy * 550145256Sjkoshy * - At context switch IN time, each "target" PMC in RUNNING state 551145256Sjkoshy * gets started on hardware and a pointer to each PMC is copied into 552145256Sjkoshy * the per-cpu phw array. The 'runcount' for the PMC is 553145256Sjkoshy * incremented. 554145256Sjkoshy * 555145256Sjkoshy * - At context switch OUT time, all process-virtual PMCs are stopped 556145256Sjkoshy * on hardware. The saved value is added to the PMCs value field 557145256Sjkoshy * only if the PMC is in a non-deleted state (the PMCs state could 558145256Sjkoshy * have changed during the current time slice). 559145256Sjkoshy * 560145256Sjkoshy * Note that since in-between a switch IN on a processor and a switch 561145256Sjkoshy * OUT, the PMC could have been released on another CPU. Therefore 562145256Sjkoshy * context switch OUT always looks at the hardware state to turn 563145256Sjkoshy * OFF PMCs and will update a PMC's saved value only if reachable 564145256Sjkoshy * from the target process record. 565145256Sjkoshy * 566145256Sjkoshy * - OP PMCRELEASE could be called on a PMC at any time (the PMC could 567145256Sjkoshy * be attached to many processes at the time of the call and could 568145256Sjkoshy * be active on multiple CPUs). 569145256Sjkoshy * 570145256Sjkoshy * We prevent further scheduling of the PMC by marking it as in 571145256Sjkoshy * state 'DELETED'. If the runcount of the PMC is non-zero then 572145256Sjkoshy * this PMC is currently running on a CPU somewhere. The thread 573145256Sjkoshy * doing the PMCRELEASE operation waits by repeatedly doing an 574145256Sjkoshy * tsleep() till the runcount comes to zero. 575145256Sjkoshy * 576145256Sjkoshy */ 577145256Sjkoshy 578145256Sjkoshy/* 579145256Sjkoshy * save the cpu binding of the current kthread 580145256Sjkoshy */ 581145256Sjkoshy 582145256Sjkoshystatic void 583145256Sjkoshypmc_save_cpu_binding(struct pmc_binding *pb) 584145256Sjkoshy{ 585145256Sjkoshy PMCDBG(CPU,BND,2, "%s", "save-cpu"); 586145256Sjkoshy mtx_lock_spin(&sched_lock); 587145256Sjkoshy pb->pb_bound = sched_is_bound(curthread); 588145256Sjkoshy pb->pb_cpu = curthread->td_oncpu; 589145256Sjkoshy mtx_unlock_spin(&sched_lock); 590145256Sjkoshy PMCDBG(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu); 591145256Sjkoshy} 592145256Sjkoshy 593145256Sjkoshy/* 594145256Sjkoshy * restore the cpu binding of the current thread 595145256Sjkoshy */ 596145256Sjkoshy 597145256Sjkoshystatic void 598145256Sjkoshypmc_restore_cpu_binding(struct pmc_binding *pb) 599145256Sjkoshy{ 600145256Sjkoshy PMCDBG(CPU,BND,2, "restore-cpu curcpu=%d restore=%d", 601145256Sjkoshy curthread->td_oncpu, pb->pb_cpu); 602145256Sjkoshy mtx_lock_spin(&sched_lock); 603145256Sjkoshy if (pb->pb_bound) 604145256Sjkoshy sched_bind(curthread, pb->pb_cpu); 605145256Sjkoshy else 606145256Sjkoshy sched_unbind(curthread); 607145256Sjkoshy mtx_unlock_spin(&sched_lock); 608145256Sjkoshy PMCDBG(CPU,BND,2, "%s", "restore-cpu done"); 609145256Sjkoshy} 610145256Sjkoshy 611145256Sjkoshy/* 612145256Sjkoshy * move execution over the specified cpu and bind it there. 613145256Sjkoshy */ 614145256Sjkoshy 615145256Sjkoshystatic void 616145256Sjkoshypmc_select_cpu(int cpu) 617145256Sjkoshy{ 618145256Sjkoshy KASSERT(cpu >= 0 && cpu < mp_ncpus, 619145256Sjkoshy ("[pmc,%d] bad cpu number %d", __LINE__, cpu)); 620145256Sjkoshy 621145256Sjkoshy /* never move to a disabled CPU */ 622145256Sjkoshy KASSERT(pmc_cpu_is_disabled(cpu) == 0, ("[pmc,%d] selecting " 623145256Sjkoshy "disabled CPU %d", __LINE__, cpu)); 624145256Sjkoshy 625145256Sjkoshy PMCDBG(CPU,SEL,2, "select-cpu cpu=%d", cpu); 626145256Sjkoshy mtx_lock_spin(&sched_lock); 627145256Sjkoshy sched_bind(curthread, cpu); 628145256Sjkoshy mtx_unlock_spin(&sched_lock); 629145256Sjkoshy 630145256Sjkoshy KASSERT(curthread->td_oncpu == cpu, 631145256Sjkoshy ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__, 632145256Sjkoshy cpu, curthread->td_oncpu)); 633145256Sjkoshy 634145256Sjkoshy PMCDBG(CPU,SEL,2, "select-cpu cpu=%d ok", cpu); 635145256Sjkoshy} 636145256Sjkoshy 637145256Sjkoshy/* 638145774Sjkoshy * Force a context switch. 639145774Sjkoshy * 640145774Sjkoshy * We do this by tsleep'ing for 1 tick -- invoking mi_switch() is not 641145774Sjkoshy * guaranteed to force a context switch. 642145774Sjkoshy */ 643145774Sjkoshy 644145774Sjkoshystatic void 645145774Sjkoshypmc_force_context_switch(void) 646145774Sjkoshy{ 647145774Sjkoshy u_char curpri; 648145774Sjkoshy 649145774Sjkoshy mtx_lock_spin(&sched_lock); 650145774Sjkoshy curpri = curthread->td_priority; 651145774Sjkoshy mtx_unlock_spin(&sched_lock); 652145774Sjkoshy 653145774Sjkoshy (void) tsleep((void *) pmc_force_context_switch, curpri, 654145774Sjkoshy "pmcctx", 1); 655145774Sjkoshy 656145774Sjkoshy} 657145774Sjkoshy 658145774Sjkoshy/* 659147191Sjkoshy * Get the file name for an executable. This is a simple wrapper 660147191Sjkoshy * around vn_fullpath(9). 661145256Sjkoshy */ 662145256Sjkoshy 663147191Sjkoshystatic void 664147708Sjkoshypmc_getfilename(struct vnode *v, char **fullpath, char **freepath) 665145256Sjkoshy{ 666145256Sjkoshy struct thread *td; 667145256Sjkoshy 668147191Sjkoshy td = curthread; 669147191Sjkoshy *fullpath = "unknown"; 670147191Sjkoshy *freepath = NULL; 671148088Sjkoshy vn_lock(v, LK_CANRECURSE | LK_EXCLUSIVE | LK_RETRY, td); 672147708Sjkoshy vn_fullpath(td, v, fullpath, freepath); 673147708Sjkoshy VOP_UNLOCK(v, 0, td); 674145256Sjkoshy} 675145256Sjkoshy 676145256Sjkoshy/* 677145256Sjkoshy * remove an process owning PMCs 678145256Sjkoshy */ 679145256Sjkoshy 680145256Sjkoshyvoid 681145256Sjkoshypmc_remove_owner(struct pmc_owner *po) 682145256Sjkoshy{ 683147191Sjkoshy struct pmc *pm, *tmp; 684145256Sjkoshy 685145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 686145256Sjkoshy 687145256Sjkoshy PMCDBG(OWN,ORM,1, "remove-owner po=%p", po); 688145256Sjkoshy 689145256Sjkoshy /* Remove descriptor from the owner hash table */ 690145256Sjkoshy LIST_REMOVE(po, po_next); 691145256Sjkoshy 692147191Sjkoshy /* release all owned PMC descriptors */ 693147191Sjkoshy LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) { 694147191Sjkoshy PMCDBG(OWN,ORM,2, "pmc=%p", pm); 695147191Sjkoshy KASSERT(pm->pm_owner == po, 696147191Sjkoshy ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po)); 697145256Sjkoshy 698147191Sjkoshy pmc_release_pmc_descriptor(pm); /* will unlink from the list */ 699145256Sjkoshy } 700145256Sjkoshy 701147191Sjkoshy KASSERT(po->po_sscount == 0, 702147191Sjkoshy ("[pmc,%d] SS count not zero", __LINE__)); 703145256Sjkoshy KASSERT(LIST_EMPTY(&po->po_pmcs), 704147191Sjkoshy ("[pmc,%d] PMC list not empty", __LINE__)); 705145256Sjkoshy 706147191Sjkoshy /* de-configure the log file if present */ 707145774Sjkoshy if (po->po_flags & PMC_PO_OWNS_LOGFILE) 708147191Sjkoshy pmclog_deconfigure_log(po); 709145256Sjkoshy} 710145256Sjkoshy 711145256Sjkoshy/* 712145256Sjkoshy * remove an owner process record if all conditions are met. 713145256Sjkoshy */ 714145256Sjkoshy 715145256Sjkoshystatic void 716145256Sjkoshypmc_maybe_remove_owner(struct pmc_owner *po) 717145256Sjkoshy{ 718145256Sjkoshy 719145256Sjkoshy PMCDBG(OWN,OMR,1, "maybe-remove-owner po=%p", po); 720145256Sjkoshy 721145256Sjkoshy /* 722145256Sjkoshy * Remove owner record if 723145256Sjkoshy * - this process does not own any PMCs 724145256Sjkoshy * - this process has not allocated a system-wide sampling buffer 725145256Sjkoshy */ 726145256Sjkoshy 727145256Sjkoshy if (LIST_EMPTY(&po->po_pmcs) && 728145774Sjkoshy ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) { 729145256Sjkoshy pmc_remove_owner(po); 730147191Sjkoshy pmc_destroy_owner_descriptor(po); 731145256Sjkoshy } 732145256Sjkoshy} 733145256Sjkoshy 734145256Sjkoshy/* 735145256Sjkoshy * Add an association between a target process and a PMC. 736145256Sjkoshy */ 737145256Sjkoshy 738145256Sjkoshystatic void 739145256Sjkoshypmc_link_target_process(struct pmc *pm, struct pmc_process *pp) 740145256Sjkoshy{ 741145256Sjkoshy int ri; 742145256Sjkoshy struct pmc_target *pt; 743145256Sjkoshy 744145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 745145256Sjkoshy 746145256Sjkoshy KASSERT(pm != NULL && pp != NULL, 747145256Sjkoshy ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 748147191Sjkoshy KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 749147191Sjkoshy ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d", 750147191Sjkoshy __LINE__, pm, pp->pp_proc->p_pid)); 751145256Sjkoshy KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt < ((int) md->pmd_npmc - 1), 752145256Sjkoshy ("[pmc,%d] Illegal reference count %d for process record %p", 753145256Sjkoshy __LINE__, pp->pp_refcnt, (void *) pp)); 754145256Sjkoshy 755145774Sjkoshy ri = PMC_TO_ROWINDEX(pm); 756145256Sjkoshy 757145256Sjkoshy PMCDBG(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p", 758145256Sjkoshy pm, ri, pp); 759145256Sjkoshy 760153110Sru#ifdef DEBUG 761145256Sjkoshy LIST_FOREACH(pt, &pm->pm_targets, pt_next) 762145256Sjkoshy if (pt->pt_process == pp) 763145256Sjkoshy KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets", 764145256Sjkoshy __LINE__, pp, pm)); 765145256Sjkoshy#endif 766145256Sjkoshy 767145256Sjkoshy MALLOC(pt, struct pmc_target *, sizeof(struct pmc_target), 768145256Sjkoshy M_PMC, M_ZERO|M_WAITOK); 769145256Sjkoshy 770145256Sjkoshy pt->pt_process = pp; 771145256Sjkoshy 772145256Sjkoshy LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next); 773145256Sjkoshy 774148067Sjhb atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc, 775148067Sjhb (uintptr_t)pm); 776145256Sjkoshy 777145615Sjkoshy if (pm->pm_owner->po_owner == pp->pp_proc) 778145774Sjkoshy pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER; 779145615Sjkoshy 780147191Sjkoshy /* 781147191Sjkoshy * Initialize the per-process values at this row index. 782147191Sjkoshy */ 783147191Sjkoshy pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ? 784147191Sjkoshy pm->pm_sc.pm_reloadcount : 0; 785147191Sjkoshy 786145256Sjkoshy pp->pp_refcnt++; 787145256Sjkoshy 788145256Sjkoshy} 789145256Sjkoshy 790145256Sjkoshy/* 791145256Sjkoshy * Removes the association between a target process and a PMC. 792145256Sjkoshy */ 793145256Sjkoshy 794145256Sjkoshystatic void 795145256Sjkoshypmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp) 796145256Sjkoshy{ 797145256Sjkoshy int ri; 798147191Sjkoshy struct proc *p; 799145256Sjkoshy struct pmc_target *ptgt; 800145256Sjkoshy 801145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 802145256Sjkoshy 803145256Sjkoshy KASSERT(pm != NULL && pp != NULL, 804145256Sjkoshy ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 805145256Sjkoshy 806145256Sjkoshy KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt < (int) md->pmd_npmc, 807145256Sjkoshy ("[pmc,%d] Illegal ref count %d on process record %p", 808145256Sjkoshy __LINE__, pp->pp_refcnt, (void *) pp)); 809145256Sjkoshy 810145774Sjkoshy ri = PMC_TO_ROWINDEX(pm); 811145256Sjkoshy 812145256Sjkoshy PMCDBG(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p", 813145256Sjkoshy pm, ri, pp); 814145256Sjkoshy 815145256Sjkoshy KASSERT(pp->pp_pmcs[ri].pp_pmc == pm, 816145256Sjkoshy ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__, 817145256Sjkoshy ri, pm, pp->pp_pmcs[ri].pp_pmc)); 818145256Sjkoshy 819145256Sjkoshy pp->pp_pmcs[ri].pp_pmc = NULL; 820145256Sjkoshy pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0; 821145256Sjkoshy 822145774Sjkoshy /* Remove owner-specific flags */ 823145774Sjkoshy if (pm->pm_owner->po_owner == pp->pp_proc) { 824145774Sjkoshy pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS; 825145774Sjkoshy pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER; 826145774Sjkoshy } 827145615Sjkoshy 828145256Sjkoshy pp->pp_refcnt--; 829145256Sjkoshy 830145256Sjkoshy /* Remove the target process from the PMC structure */ 831145256Sjkoshy LIST_FOREACH(ptgt, &pm->pm_targets, pt_next) 832145256Sjkoshy if (ptgt->pt_process == pp) 833145256Sjkoshy break; 834145256Sjkoshy 835145256Sjkoshy KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found " 836145256Sjkoshy "in pmc %p", __LINE__, pp->pp_proc, pp, pm)); 837145256Sjkoshy 838145256Sjkoshy LIST_REMOVE(ptgt, pt_next); 839145256Sjkoshy FREE(ptgt, M_PMC); 840145256Sjkoshy 841147191Sjkoshy /* if the PMC now lacks targets, send the owner a SIGIO */ 842147191Sjkoshy if (LIST_EMPTY(&pm->pm_targets)) { 843147191Sjkoshy p = pm->pm_owner->po_owner; 844147191Sjkoshy PROC_LOCK(p); 845147191Sjkoshy psignal(p, SIGIO); 846147191Sjkoshy PROC_UNLOCK(p); 847145256Sjkoshy 848147191Sjkoshy PMCDBG(PRC,SIG,2, "signalling proc=%p signal=%d", p, 849147191Sjkoshy SIGIO); 850145256Sjkoshy } 851145256Sjkoshy} 852145256Sjkoshy 853145256Sjkoshy/* 854145256Sjkoshy * Check if PMC 'pm' may be attached to target process 't'. 855145256Sjkoshy */ 856145256Sjkoshy 857145256Sjkoshystatic int 858145256Sjkoshypmc_can_attach(struct pmc *pm, struct proc *t) 859145256Sjkoshy{ 860145256Sjkoshy struct proc *o; /* pmc owner */ 861145256Sjkoshy struct ucred *oc, *tc; /* owner, target credentials */ 862145256Sjkoshy int decline_attach, i; 863145256Sjkoshy 864145256Sjkoshy /* 865145256Sjkoshy * A PMC's owner can always attach that PMC to itself. 866145256Sjkoshy */ 867145256Sjkoshy 868145256Sjkoshy if ((o = pm->pm_owner->po_owner) == t) 869145256Sjkoshy return 0; 870145256Sjkoshy 871145256Sjkoshy PROC_LOCK(o); 872145256Sjkoshy oc = o->p_ucred; 873145256Sjkoshy crhold(oc); 874145256Sjkoshy PROC_UNLOCK(o); 875145256Sjkoshy 876145256Sjkoshy PROC_LOCK(t); 877145256Sjkoshy tc = t->p_ucred; 878145256Sjkoshy crhold(tc); 879145256Sjkoshy PROC_UNLOCK(t); 880145256Sjkoshy 881145256Sjkoshy /* 882145256Sjkoshy * The effective uid of the PMC owner should match at least one 883145256Sjkoshy * of the {effective,real,saved} uids of the target process. 884145256Sjkoshy */ 885145256Sjkoshy 886145256Sjkoshy decline_attach = oc->cr_uid != tc->cr_uid && 887145256Sjkoshy oc->cr_uid != tc->cr_svuid && 888145256Sjkoshy oc->cr_uid != tc->cr_ruid; 889145256Sjkoshy 890145256Sjkoshy /* 891145256Sjkoshy * Every one of the target's group ids, must be in the owner's 892145256Sjkoshy * group list. 893145256Sjkoshy */ 894145256Sjkoshy for (i = 0; !decline_attach && i < tc->cr_ngroups; i++) 895145256Sjkoshy decline_attach = !groupmember(tc->cr_groups[i], oc); 896145256Sjkoshy 897145256Sjkoshy /* check the read and saved gids too */ 898145256Sjkoshy if (decline_attach == 0) 899145256Sjkoshy decline_attach = !groupmember(tc->cr_rgid, oc) || 900145256Sjkoshy !groupmember(tc->cr_svgid, oc); 901145256Sjkoshy 902145256Sjkoshy crfree(tc); 903145256Sjkoshy crfree(oc); 904145256Sjkoshy 905145256Sjkoshy return !decline_attach; 906145256Sjkoshy} 907145256Sjkoshy 908145256Sjkoshy/* 909145256Sjkoshy * Attach a process to a PMC. 910145256Sjkoshy */ 911145256Sjkoshy 912145256Sjkoshystatic int 913145256Sjkoshypmc_attach_one_process(struct proc *p, struct pmc *pm) 914145256Sjkoshy{ 915145256Sjkoshy int ri; 916147191Sjkoshy char *fullpath, *freepath; 917145256Sjkoshy struct pmc_process *pp; 918145256Sjkoshy 919145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 920145256Sjkoshy 921145256Sjkoshy PMCDBG(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm, 922145774Sjkoshy PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 923145256Sjkoshy 924145256Sjkoshy /* 925145256Sjkoshy * Locate the process descriptor corresponding to process 'p', 926145256Sjkoshy * allocating space as needed. 927145256Sjkoshy * 928145256Sjkoshy * Verify that rowindex 'pm_rowindex' is free in the process 929145256Sjkoshy * descriptor. 930145256Sjkoshy * 931145256Sjkoshy * If not, allocate space for a descriptor and link the 932145256Sjkoshy * process descriptor and PMC. 933145256Sjkoshy */ 934145774Sjkoshy ri = PMC_TO_ROWINDEX(pm); 935145256Sjkoshy 936145256Sjkoshy if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) 937145256Sjkoshy return ENOMEM; 938145256Sjkoshy 939145256Sjkoshy if (pp->pp_pmcs[ri].pp_pmc == pm) /* already present at slot [ri] */ 940145256Sjkoshy return EEXIST; 941145256Sjkoshy 942145256Sjkoshy if (pp->pp_pmcs[ri].pp_pmc != NULL) 943145256Sjkoshy return EBUSY; 944145256Sjkoshy 945145256Sjkoshy pmc_link_target_process(pm, pp); 946145256Sjkoshy 947147191Sjkoshy if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) && 948147191Sjkoshy (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0) 949147191Sjkoshy pm->pm_flags |= PMC_F_NEEDS_LOGFILE; 950147191Sjkoshy 951147191Sjkoshy pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */ 952147191Sjkoshy 953147191Sjkoshy /* issue an attach event to a configured log file */ 954147191Sjkoshy if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) { 955147708Sjkoshy pmc_getfilename(p->p_textvp, &fullpath, &freepath); 956147191Sjkoshy pmclog_process_pmcattach(pm, p->p_pid, fullpath); 957147191Sjkoshy if (freepath) 958147191Sjkoshy FREE(freepath, M_TEMP); 959147191Sjkoshy } 960145256Sjkoshy /* mark process as using HWPMCs */ 961145256Sjkoshy PROC_LOCK(p); 962145256Sjkoshy p->p_flag |= P_HWPMC; 963145256Sjkoshy PROC_UNLOCK(p); 964145256Sjkoshy 965145256Sjkoshy return 0; 966145256Sjkoshy} 967145256Sjkoshy 968145256Sjkoshy/* 969145256Sjkoshy * Attach a process and optionally its children 970145256Sjkoshy */ 971145256Sjkoshy 972145256Sjkoshystatic int 973145256Sjkoshypmc_attach_process(struct proc *p, struct pmc *pm) 974145256Sjkoshy{ 975145256Sjkoshy int error; 976145256Sjkoshy struct proc *top; 977145256Sjkoshy 978145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 979145256Sjkoshy 980145256Sjkoshy PMCDBG(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm, 981145774Sjkoshy PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 982145256Sjkoshy 983145774Sjkoshy 984145774Sjkoshy /* 985145774Sjkoshy * If this PMC successfully allowed a GETMSR operation 986145774Sjkoshy * in the past, disallow further ATTACHes. 987145774Sjkoshy */ 988145774Sjkoshy 989145774Sjkoshy if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0) 990145774Sjkoshy return EPERM; 991145774Sjkoshy 992145256Sjkoshy if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 993145256Sjkoshy return pmc_attach_one_process(p, pm); 994145256Sjkoshy 995145256Sjkoshy /* 996145256Sjkoshy * Traverse all child processes, attaching them to 997145256Sjkoshy * this PMC. 998145256Sjkoshy */ 999145256Sjkoshy 1000145256Sjkoshy sx_slock(&proctree_lock); 1001145256Sjkoshy 1002145256Sjkoshy top = p; 1003145256Sjkoshy 1004145256Sjkoshy for (;;) { 1005145256Sjkoshy if ((error = pmc_attach_one_process(p, pm)) != 0) 1006145256Sjkoshy break; 1007145256Sjkoshy if (!LIST_EMPTY(&p->p_children)) 1008145256Sjkoshy p = LIST_FIRST(&p->p_children); 1009145256Sjkoshy else for (;;) { 1010145256Sjkoshy if (p == top) 1011145256Sjkoshy goto done; 1012145256Sjkoshy if (LIST_NEXT(p, p_sibling)) { 1013145256Sjkoshy p = LIST_NEXT(p, p_sibling); 1014145256Sjkoshy break; 1015145256Sjkoshy } 1016145256Sjkoshy p = p->p_pptr; 1017145256Sjkoshy } 1018145256Sjkoshy } 1019145256Sjkoshy 1020145256Sjkoshy if (error) 1021145256Sjkoshy (void) pmc_detach_process(top, pm); 1022145256Sjkoshy 1023145256Sjkoshy done: 1024145256Sjkoshy sx_sunlock(&proctree_lock); 1025145256Sjkoshy return error; 1026145256Sjkoshy} 1027145256Sjkoshy 1028145256Sjkoshy/* 1029145256Sjkoshy * Detach a process from a PMC. If there are no other PMCs tracking 1030145256Sjkoshy * this process, remove the process structure from its hash table. If 1031145256Sjkoshy * 'flags' contains PMC_FLAG_REMOVE, then free the process structure. 1032145256Sjkoshy */ 1033145256Sjkoshy 1034145256Sjkoshystatic int 1035145256Sjkoshypmc_detach_one_process(struct proc *p, struct pmc *pm, int flags) 1036145256Sjkoshy{ 1037145256Sjkoshy int ri; 1038145256Sjkoshy struct pmc_process *pp; 1039145256Sjkoshy 1040145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 1041145256Sjkoshy 1042145256Sjkoshy KASSERT(pm != NULL, 1043145256Sjkoshy ("[pmc,%d] null pm pointer", __LINE__)); 1044145256Sjkoshy 1045145774Sjkoshy ri = PMC_TO_ROWINDEX(pm); 1046145774Sjkoshy 1047145256Sjkoshy PMCDBG(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x", 1048145774Sjkoshy pm, ri, p, p->p_pid, p->p_comm, flags); 1049145256Sjkoshy 1050145256Sjkoshy if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) 1051145256Sjkoshy return ESRCH; 1052145256Sjkoshy 1053145256Sjkoshy if (pp->pp_pmcs[ri].pp_pmc != pm) 1054145256Sjkoshy return EINVAL; 1055145256Sjkoshy 1056145256Sjkoshy pmc_unlink_target_process(pm, pp); 1057145256Sjkoshy 1058147191Sjkoshy /* Issue a detach entry if a log file is configured */ 1059147191Sjkoshy if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) 1060147191Sjkoshy pmclog_process_pmcdetach(pm, p->p_pid); 1061147191Sjkoshy 1062145256Sjkoshy /* 1063145256Sjkoshy * If there are no PMCs targetting this process, we remove its 1064145256Sjkoshy * descriptor from the target hash table and unset the P_HWPMC 1065145256Sjkoshy * flag in the struct proc. 1066145256Sjkoshy */ 1067145256Sjkoshy KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt < (int) md->pmd_npmc, 1068145256Sjkoshy ("[pmc,%d] Illegal refcnt %d for process struct %p", 1069145256Sjkoshy __LINE__, pp->pp_refcnt, pp)); 1070145256Sjkoshy 1071145256Sjkoshy if (pp->pp_refcnt != 0) /* still a target of some PMC */ 1072145256Sjkoshy return 0; 1073145256Sjkoshy 1074145256Sjkoshy pmc_remove_process_descriptor(pp); 1075145256Sjkoshy 1076145256Sjkoshy if (flags & PMC_FLAG_REMOVE) 1077145256Sjkoshy FREE(pp, M_PMC); 1078145256Sjkoshy 1079145256Sjkoshy PROC_LOCK(p); 1080145256Sjkoshy p->p_flag &= ~P_HWPMC; 1081145256Sjkoshy PROC_UNLOCK(p); 1082145256Sjkoshy 1083145256Sjkoshy return 0; 1084145256Sjkoshy} 1085145256Sjkoshy 1086145256Sjkoshy/* 1087145256Sjkoshy * Detach a process and optionally its descendants from a PMC. 1088145256Sjkoshy */ 1089145256Sjkoshy 1090145256Sjkoshystatic int 1091145256Sjkoshypmc_detach_process(struct proc *p, struct pmc *pm) 1092145256Sjkoshy{ 1093145256Sjkoshy struct proc *top; 1094145256Sjkoshy 1095145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 1096145256Sjkoshy 1097145256Sjkoshy PMCDBG(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm, 1098145774Sjkoshy PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1099145256Sjkoshy 1100145256Sjkoshy if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1101145256Sjkoshy return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); 1102145256Sjkoshy 1103145256Sjkoshy /* 1104145256Sjkoshy * Traverse all children, detaching them from this PMC. We 1105145256Sjkoshy * ignore errors since we could be detaching a PMC from a 1106145256Sjkoshy * partially attached proc tree. 1107145256Sjkoshy */ 1108145256Sjkoshy 1109145256Sjkoshy sx_slock(&proctree_lock); 1110145256Sjkoshy 1111145256Sjkoshy top = p; 1112145256Sjkoshy 1113145256Sjkoshy for (;;) { 1114145256Sjkoshy (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); 1115145256Sjkoshy 1116145256Sjkoshy if (!LIST_EMPTY(&p->p_children)) 1117145256Sjkoshy p = LIST_FIRST(&p->p_children); 1118145256Sjkoshy else for (;;) { 1119145256Sjkoshy if (p == top) 1120145256Sjkoshy goto done; 1121145256Sjkoshy if (LIST_NEXT(p, p_sibling)) { 1122145256Sjkoshy p = LIST_NEXT(p, p_sibling); 1123145256Sjkoshy break; 1124145256Sjkoshy } 1125145256Sjkoshy p = p->p_pptr; 1126145256Sjkoshy } 1127145256Sjkoshy } 1128145256Sjkoshy 1129145256Sjkoshy done: 1130145256Sjkoshy sx_sunlock(&proctree_lock); 1131147191Sjkoshy 1132147191Sjkoshy if (LIST_EMPTY(&pm->pm_targets)) 1133147191Sjkoshy pm->pm_flags &= ~PMC_F_ATTACH_DONE; 1134147191Sjkoshy 1135145256Sjkoshy return 0; 1136145256Sjkoshy} 1137145256Sjkoshy 1138147191Sjkoshy 1139145256Sjkoshy/* 1140147191Sjkoshy * Thread context switch IN 1141145256Sjkoshy */ 1142145256Sjkoshy 1143147191Sjkoshystatic void 1144147191Sjkoshypmc_process_csw_in(struct thread *td) 1145147191Sjkoshy{ 1146147191Sjkoshy int cpu; 1147147191Sjkoshy unsigned int ri; 1148147191Sjkoshy struct pmc *pm; 1149147191Sjkoshy struct proc *p; 1150147191Sjkoshy struct pmc_cpu *pc; 1151147191Sjkoshy struct pmc_hw *phw; 1152147191Sjkoshy struct pmc_process *pp; 1153147191Sjkoshy pmc_value_t newvalue; 1154145256Sjkoshy 1155147191Sjkoshy p = td->td_proc; 1156145256Sjkoshy 1157147191Sjkoshy if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL) 1158147191Sjkoshy return; 1159145256Sjkoshy 1160147191Sjkoshy KASSERT(pp->pp_proc == td->td_proc, 1161147191Sjkoshy ("[pmc,%d] not my thread state", __LINE__)); 1162145256Sjkoshy 1163147191Sjkoshy critical_enter(); /* no preemption from this point */ 1164145256Sjkoshy 1165147191Sjkoshy cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1166145256Sjkoshy 1167147191Sjkoshy PMCDBG(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1168147191Sjkoshy p->p_pid, p->p_comm, pp); 1169145256Sjkoshy 1170147191Sjkoshy KASSERT(cpu >= 0 && cpu < mp_ncpus, 1171147191Sjkoshy ("[pmc,%d] wierd CPU id %d", __LINE__, cpu)); 1172145256Sjkoshy 1173147191Sjkoshy pc = pmc_pcpu[cpu]; 1174145256Sjkoshy 1175147191Sjkoshy for (ri = 0; ri < md->pmd_npmc; ri++) { 1176145256Sjkoshy 1177147191Sjkoshy if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) 1178147191Sjkoshy continue; 1179147191Sjkoshy 1180147191Sjkoshy KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 1181147191Sjkoshy ("[pmc,%d] Target PMC in non-virtual mode (%d)", 1182147191Sjkoshy __LINE__, PMC_TO_MODE(pm))); 1183147191Sjkoshy 1184147191Sjkoshy KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1185147191Sjkoshy ("[pmc,%d] Row index mismatch pmc %d != ri %d", 1186147191Sjkoshy __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1187147191Sjkoshy 1188145256Sjkoshy /* 1189147191Sjkoshy * Only PMCs that are marked as 'RUNNING' need 1190147191Sjkoshy * be placed on hardware. 1191145256Sjkoshy */ 1192145256Sjkoshy 1193147191Sjkoshy if (pm->pm_state != PMC_STATE_RUNNING) 1194147191Sjkoshy continue; 1195145256Sjkoshy 1196147191Sjkoshy /* increment PMC runcount */ 1197147191Sjkoshy atomic_add_rel_32(&pm->pm_runcount, 1); 1198145256Sjkoshy 1199147191Sjkoshy /* configure the HWPMC we are going to use. */ 1200147191Sjkoshy md->pmd_config_pmc(cpu, ri, pm); 1201145256Sjkoshy 1202147191Sjkoshy phw = pc->pc_hwpmcs[ri]; 1203145256Sjkoshy 1204147191Sjkoshy KASSERT(phw != NULL, 1205147191Sjkoshy ("[pmc,%d] null hw pointer", __LINE__)); 1206145256Sjkoshy 1207147191Sjkoshy KASSERT(phw->phw_pmc == pm, 1208147191Sjkoshy ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__, 1209147191Sjkoshy phw->phw_pmc, pm)); 1210145256Sjkoshy 1211147191Sjkoshy /* 1212147191Sjkoshy * Write out saved value and start the PMC. 1213147191Sjkoshy * 1214147191Sjkoshy * Sampling PMCs use a per-process value, while 1215147191Sjkoshy * counting mode PMCs use a per-pmc value that is 1216147191Sjkoshy * inherited across descendants. 1217147191Sjkoshy */ 1218147191Sjkoshy if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 1219147191Sjkoshy mtx_pool_lock_spin(pmc_mtxpool, pm); 1220147191Sjkoshy newvalue = PMC_PCPU_SAVED(cpu,ri) = 1221147191Sjkoshy pp->pp_pmcs[ri].pp_pmcval; 1222147191Sjkoshy mtx_pool_unlock_spin(pmc_mtxpool, pm); 1223147191Sjkoshy } else { 1224147191Sjkoshy KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC, 1225147191Sjkoshy ("[pmc,%d] illegal mode=%d", __LINE__, 1226147191Sjkoshy PMC_TO_MODE(pm))); 1227147191Sjkoshy mtx_pool_lock_spin(pmc_mtxpool, pm); 1228147191Sjkoshy newvalue = PMC_PCPU_SAVED(cpu, ri) = 1229147191Sjkoshy pm->pm_gv.pm_savedvalue; 1230147191Sjkoshy mtx_pool_unlock_spin(pmc_mtxpool, pm); 1231147191Sjkoshy } 1232145256Sjkoshy 1233147191Sjkoshy PMCDBG(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue); 1234145256Sjkoshy 1235147191Sjkoshy md->pmd_write_pmc(cpu, ri, newvalue); 1236147191Sjkoshy md->pmd_start_pmc(cpu, ri); 1237147191Sjkoshy } 1238145256Sjkoshy 1239147191Sjkoshy /* 1240147191Sjkoshy * perform any other architecture/cpu dependent thread 1241147191Sjkoshy * switch-in actions. 1242147191Sjkoshy */ 1243145256Sjkoshy 1244147191Sjkoshy (void) (*md->pmd_switch_in)(pc, pp); 1245145256Sjkoshy 1246147191Sjkoshy critical_exit(); 1247145256Sjkoshy 1248147191Sjkoshy} 1249145256Sjkoshy 1250147191Sjkoshy/* 1251147191Sjkoshy * Thread context switch OUT. 1252147191Sjkoshy */ 1253145256Sjkoshy 1254147191Sjkoshystatic void 1255147191Sjkoshypmc_process_csw_out(struct thread *td) 1256147191Sjkoshy{ 1257147191Sjkoshy int cpu; 1258147191Sjkoshy enum pmc_mode mode; 1259147191Sjkoshy unsigned int ri; 1260147191Sjkoshy struct pmc *pm; 1261147191Sjkoshy struct proc *p; 1262147191Sjkoshy struct pmc_cpu *pc; 1263147191Sjkoshy struct pmc_process *pp; 1264147191Sjkoshy int64_t tmp; 1265147191Sjkoshy pmc_value_t newvalue; 1266145256Sjkoshy 1267147191Sjkoshy /* 1268147191Sjkoshy * Locate our process descriptor; this may be NULL if 1269147191Sjkoshy * this process is exiting and we have already removed 1270147191Sjkoshy * the process from the target process table. 1271147191Sjkoshy * 1272147191Sjkoshy * Note that due to kernel preemption, multiple 1273147191Sjkoshy * context switches may happen while the process is 1274147191Sjkoshy * exiting. 1275147191Sjkoshy * 1276147191Sjkoshy * Note also that if the target process cannot be 1277147191Sjkoshy * found we still need to deconfigure any PMCs that 1278147191Sjkoshy * are currently running on hardware. 1279147191Sjkoshy */ 1280145256Sjkoshy 1281147191Sjkoshy p = td->td_proc; 1282147191Sjkoshy pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE); 1283145256Sjkoshy 1284147191Sjkoshy /* 1285147191Sjkoshy * save PMCs 1286147191Sjkoshy */ 1287145256Sjkoshy 1288147191Sjkoshy critical_enter(); 1289145774Sjkoshy 1290147191Sjkoshy cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1291145256Sjkoshy 1292147191Sjkoshy PMCDBG(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1293147191Sjkoshy p->p_pid, p->p_comm, pp); 1294145615Sjkoshy 1295147191Sjkoshy KASSERT(cpu >= 0 && cpu < mp_ncpus, 1296147191Sjkoshy ("[pmc,%d wierd CPU id %d", __LINE__, cpu)); 1297145615Sjkoshy 1298147191Sjkoshy pc = pmc_pcpu[cpu]; 1299145615Sjkoshy 1300147191Sjkoshy /* 1301147191Sjkoshy * When a PMC gets unlinked from a target PMC, it will 1302147191Sjkoshy * be removed from the target's pp_pmc[] array. 1303147191Sjkoshy * 1304147191Sjkoshy * However, on a MP system, the target could have been 1305147191Sjkoshy * executing on another CPU at the time of the unlink. 1306147191Sjkoshy * So, at context switch OUT time, we need to look at 1307147191Sjkoshy * the hardware to determine if a PMC is scheduled on 1308147191Sjkoshy * it. 1309147191Sjkoshy */ 1310145256Sjkoshy 1311147191Sjkoshy for (ri = 0; ri < md->pmd_npmc; ri++) { 1312145256Sjkoshy 1313147191Sjkoshy pm = NULL; 1314147191Sjkoshy (void) (*md->pmd_get_config)(cpu, ri, &pm); 1315145256Sjkoshy 1316147191Sjkoshy if (pm == NULL) /* nothing at this row index */ 1317147191Sjkoshy continue; 1318145256Sjkoshy 1319147191Sjkoshy mode = PMC_TO_MODE(pm); 1320147191Sjkoshy if (!PMC_IS_VIRTUAL_MODE(mode)) 1321147191Sjkoshy continue; /* not a process virtual PMC */ 1322145774Sjkoshy 1323147191Sjkoshy KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1324147191Sjkoshy ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", 1325147191Sjkoshy __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1326145256Sjkoshy 1327147191Sjkoshy /* Stop hardware if not already stopped */ 1328147867Sjkoshy if (pm->pm_stalled == 0) 1329147191Sjkoshy md->pmd_stop_pmc(cpu, ri); 1330147191Sjkoshy 1331147191Sjkoshy /* reduce this PMC's runcount */ 1332147191Sjkoshy atomic_subtract_rel_32(&pm->pm_runcount, 1); 1333147191Sjkoshy 1334145256Sjkoshy /* 1335147191Sjkoshy * If this PMC is associated with this process, 1336147191Sjkoshy * save the reading. 1337145256Sjkoshy */ 1338145256Sjkoshy 1339147191Sjkoshy if (pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) { 1340147191Sjkoshy 1341147191Sjkoshy KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 1342147191Sjkoshy ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, 1343147191Sjkoshy pm, ri, pp->pp_pmcs[ri].pp_pmc)); 1344147191Sjkoshy 1345147191Sjkoshy KASSERT(pp->pp_refcnt > 0, 1346147191Sjkoshy ("[pmc,%d] pp refcnt = %d", __LINE__, 1347147191Sjkoshy pp->pp_refcnt)); 1348147191Sjkoshy 1349147191Sjkoshy md->pmd_read_pmc(cpu, ri, &newvalue); 1350147191Sjkoshy 1351147191Sjkoshy tmp = newvalue - PMC_PCPU_SAVED(cpu,ri); 1352147191Sjkoshy 1353147191Sjkoshy PMCDBG(CSW,SWI,1,"cpu=%d ri=%d tmp=%jd", cpu, ri, 1354147191Sjkoshy tmp); 1355147191Sjkoshy 1356147191Sjkoshy if (mode == PMC_MODE_TS) { 1357147191Sjkoshy 1358147191Sjkoshy /* 1359147191Sjkoshy * For sampling process-virtual PMCs, 1360147191Sjkoshy * we expect the count to be 1361147191Sjkoshy * decreasing as the 'value' 1362147191Sjkoshy * programmed into the PMC is the 1363147191Sjkoshy * number of events to be seen till 1364147191Sjkoshy * the next sampling interrupt. 1365147191Sjkoshy */ 1366147191Sjkoshy if (tmp < 0) 1367147191Sjkoshy tmp += pm->pm_sc.pm_reloadcount; 1368147191Sjkoshy mtx_pool_lock_spin(pmc_mtxpool, pm); 1369147191Sjkoshy pp->pp_pmcs[ri].pp_pmcval -= tmp; 1370147191Sjkoshy if ((int64_t) pp->pp_pmcs[ri].pp_pmcval < 0) 1371147191Sjkoshy pp->pp_pmcs[ri].pp_pmcval += 1372147191Sjkoshy pm->pm_sc.pm_reloadcount; 1373147191Sjkoshy mtx_pool_unlock_spin(pmc_mtxpool, pm); 1374147191Sjkoshy 1375147191Sjkoshy } else { 1376147191Sjkoshy 1377147191Sjkoshy /* 1378147191Sjkoshy * For counting process-virtual PMCs, 1379147191Sjkoshy * we expect the count to be 1380147191Sjkoshy * increasing monotonically, modulo a 64 1381147191Sjkoshy * bit wraparound. 1382147191Sjkoshy */ 1383147191Sjkoshy KASSERT((int64_t) tmp >= 0, 1384147191Sjkoshy ("[pmc,%d] negative increment cpu=%d " 1385147191Sjkoshy "ri=%d newvalue=%jx saved=%jx " 1386147191Sjkoshy "incr=%jx", __LINE__, cpu, ri, 1387147191Sjkoshy newvalue, PMC_PCPU_SAVED(cpu,ri), tmp)); 1388147191Sjkoshy 1389147191Sjkoshy mtx_pool_lock_spin(pmc_mtxpool, pm); 1390147191Sjkoshy pm->pm_gv.pm_savedvalue += tmp; 1391147191Sjkoshy pp->pp_pmcs[ri].pp_pmcval += tmp; 1392147191Sjkoshy mtx_pool_unlock_spin(pmc_mtxpool, pm); 1393147191Sjkoshy 1394147191Sjkoshy if (pm->pm_flags & PMC_F_LOG_PROCCSW) 1395147191Sjkoshy pmclog_process_proccsw(pm, pp, tmp); 1396147191Sjkoshy } 1397145256Sjkoshy } 1398145256Sjkoshy 1399147191Sjkoshy /* mark hardware as free */ 1400147191Sjkoshy md->pmd_config_pmc(cpu, ri, NULL); 1401145256Sjkoshy } 1402145256Sjkoshy 1403145256Sjkoshy /* 1404147191Sjkoshy * perform any other architecture/cpu dependent thread 1405147191Sjkoshy * switch out functions. 1406147191Sjkoshy */ 1407147191Sjkoshy 1408147191Sjkoshy (void) (*md->pmd_switch_out)(pc, pp); 1409147191Sjkoshy 1410147191Sjkoshy critical_exit(); 1411147191Sjkoshy} 1412147191Sjkoshy 1413147191Sjkoshy/* 1414147191Sjkoshy * The 'hook' invoked from the kernel proper 1415147191Sjkoshy */ 1416147191Sjkoshy 1417147191Sjkoshy 1418153110Sru#ifdef DEBUG 1419147191Sjkoshyconst char *pmc_hooknames[] = { 1420147191Sjkoshy "", 1421147191Sjkoshy "EXIT", 1422147191Sjkoshy "EXEC", 1423147191Sjkoshy "FORK", 1424147191Sjkoshy "CSW-IN", 1425147191Sjkoshy "CSW-OUT", 1426147191Sjkoshy "SAMPLE" 1427147191Sjkoshy}; 1428147191Sjkoshy#endif 1429147191Sjkoshy 1430147191Sjkoshystatic int 1431147191Sjkoshypmc_hook_handler(struct thread *td, int function, void *arg) 1432147191Sjkoshy{ 1433147191Sjkoshy 1434147191Sjkoshy PMCDBG(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function, 1435147191Sjkoshy pmc_hooknames[function], arg); 1436147191Sjkoshy 1437147191Sjkoshy switch (function) 1438147191Sjkoshy { 1439147191Sjkoshy 1440147191Sjkoshy /* 1441145256Sjkoshy * Process exec() 1442145256Sjkoshy */ 1443145256Sjkoshy 1444145256Sjkoshy case PMC_FN_PROCESS_EXEC: 1445145256Sjkoshy { 1446147191Sjkoshy char *fullpath, *freepath; 1447145256Sjkoshy unsigned int ri; 1448147191Sjkoshy int is_using_hwpmcs; 1449145256Sjkoshy struct pmc *pm; 1450145256Sjkoshy struct proc *p; 1451145256Sjkoshy struct pmc_owner *po; 1452145256Sjkoshy struct pmc_process *pp; 1453147708Sjkoshy struct pmckern_procexec *pk; 1454145256Sjkoshy 1455145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 1456145256Sjkoshy 1457147191Sjkoshy p = td->td_proc; 1458147708Sjkoshy pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1459147191Sjkoshy 1460147708Sjkoshy pk = (struct pmckern_procexec *) arg; 1461147708Sjkoshy 1462147191Sjkoshy /* Inform owners of SS mode PMCs of the exec event. */ 1463147191Sjkoshy LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 1464147191Sjkoshy if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1465147708Sjkoshy pmclog_process_procexec(po, PMC_ID_INVALID, 1466147708Sjkoshy p->p_pid, pk->pm_entryaddr, fullpath); 1467147191Sjkoshy 1468147191Sjkoshy PROC_LOCK(p); 1469147191Sjkoshy is_using_hwpmcs = p->p_flag & P_HWPMC; 1470147191Sjkoshy PROC_UNLOCK(p); 1471147191Sjkoshy 1472147191Sjkoshy if (!is_using_hwpmcs) { 1473147191Sjkoshy if (freepath) 1474147191Sjkoshy FREE(freepath, M_TEMP); 1475147191Sjkoshy break; 1476147191Sjkoshy } 1477147191Sjkoshy 1478145256Sjkoshy /* 1479145256Sjkoshy * PMCs are not inherited across an exec(): remove any 1480145256Sjkoshy * PMCs that this process is the owner of. 1481145256Sjkoshy */ 1482145256Sjkoshy 1483145256Sjkoshy if ((po = pmc_find_owner_descriptor(p)) != NULL) { 1484145256Sjkoshy pmc_remove_owner(po); 1485147191Sjkoshy pmc_destroy_owner_descriptor(po); 1486145256Sjkoshy } 1487145256Sjkoshy 1488145256Sjkoshy /* 1489154483Sjkoshy * If the process being exec'ed is not the target of any 1490154483Sjkoshy * PMC, we are done. 1491145256Sjkoshy */ 1492154483Sjkoshy if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) { 1493154483Sjkoshy if (freepath) 1494154483Sjkoshy FREE(freepath, M_TEMP); 1495145256Sjkoshy break; 1496154483Sjkoshy } 1497145256Sjkoshy 1498147191Sjkoshy /* 1499147191Sjkoshy * Log the exec event to all monitoring owners. Skip 1500147191Sjkoshy * owners who have already recieved the event because 1501154483Sjkoshy * they had system sampling PMCs active. 1502147191Sjkoshy */ 1503147191Sjkoshy for (ri = 0; ri < md->pmd_npmc; ri++) 1504147191Sjkoshy if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 1505147191Sjkoshy po = pm->pm_owner; 1506147191Sjkoshy if (po->po_sscount == 0 && 1507147191Sjkoshy po->po_flags & PMC_PO_OWNS_LOGFILE) 1508147708Sjkoshy pmclog_process_procexec(po, pm->pm_id, 1509147708Sjkoshy p->p_pid, pk->pm_entryaddr, 1510147191Sjkoshy fullpath); 1511147191Sjkoshy } 1512147191Sjkoshy 1513147191Sjkoshy if (freepath) 1514147191Sjkoshy FREE(freepath, M_TEMP); 1515147191Sjkoshy 1516145256Sjkoshy 1517145256Sjkoshy PMCDBG(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d", 1518147708Sjkoshy p, p->p_pid, p->p_comm, pk->pm_credentialschanged); 1519145256Sjkoshy 1520147708Sjkoshy if (pk->pm_credentialschanged == 0) /* no change */ 1521145256Sjkoshy break; 1522145256Sjkoshy 1523145256Sjkoshy /* 1524145256Sjkoshy * If the newly exec()'ed process has a different credential 1525145256Sjkoshy * than before, allow it to be the target of a PMC only if 1526145256Sjkoshy * the PMC's owner has sufficient priviledge. 1527145256Sjkoshy */ 1528145256Sjkoshy 1529145256Sjkoshy for (ri = 0; ri < md->pmd_npmc; ri++) 1530145256Sjkoshy if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) 1531145256Sjkoshy if (pmc_can_attach(pm, td->td_proc) != 0) 1532145256Sjkoshy pmc_detach_one_process(td->td_proc, 1533145256Sjkoshy pm, PMC_FLAG_NONE); 1534145256Sjkoshy 1535145256Sjkoshy KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt < (int) md->pmd_npmc, 1536145256Sjkoshy ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__, 1537145256Sjkoshy pp->pp_refcnt, pp)); 1538145256Sjkoshy 1539145256Sjkoshy /* 1540145256Sjkoshy * If this process is no longer the target of any 1541145256Sjkoshy * PMCs, we can remove the process entry and free 1542145256Sjkoshy * up space. 1543145256Sjkoshy */ 1544145256Sjkoshy 1545145256Sjkoshy if (pp->pp_refcnt == 0) { 1546145256Sjkoshy pmc_remove_process_descriptor(pp); 1547145256Sjkoshy FREE(pp, M_PMC); 1548147191Sjkoshy break; 1549145256Sjkoshy } 1550145256Sjkoshy 1551145256Sjkoshy } 1552145256Sjkoshy break; 1553145256Sjkoshy 1554145256Sjkoshy case PMC_FN_CSW_IN: 1555147191Sjkoshy pmc_process_csw_in(td); 1556147191Sjkoshy break; 1557145256Sjkoshy 1558147191Sjkoshy case PMC_FN_CSW_OUT: 1559147191Sjkoshy pmc_process_csw_out(td); 1560147191Sjkoshy break; 1561145256Sjkoshy 1562145256Sjkoshy /* 1563147191Sjkoshy * Process accumulated PC samples. 1564147191Sjkoshy * 1565147191Sjkoshy * This function is expected to be called by hardclock() for 1566147191Sjkoshy * each CPU that has accumulated PC samples. 1567147191Sjkoshy * 1568147191Sjkoshy * This function is to be executed on the CPU whose samples 1569147191Sjkoshy * are being processed. 1570145256Sjkoshy */ 1571147191Sjkoshy case PMC_FN_DO_SAMPLES: 1572145256Sjkoshy 1573145256Sjkoshy /* 1574147191Sjkoshy * Clear the cpu specific bit in the CPU mask before 1575147191Sjkoshy * do the rest of the processing. If the NMI handler 1576147191Sjkoshy * gets invoked after the "atomic_clear_int()" call 1577147191Sjkoshy * below but before "pmc_process_samples()" gets 1578147191Sjkoshy * around to processing the interrupt, then we will 1579147191Sjkoshy * come back here at the next hardclock() tick (and 1580147191Sjkoshy * may find nothing to do if "pmc_process_samples()" 1581147191Sjkoshy * had already processed the interrupt). We don't 1582147191Sjkoshy * lose the interrupt sample. 1583145256Sjkoshy */ 1584147191Sjkoshy atomic_clear_int(&pmc_cpumask, (1 << PCPU_GET(cpuid))); 1585147191Sjkoshy pmc_process_samples(PCPU_GET(cpuid)); 1586147191Sjkoshy break; 1587145256Sjkoshy 1588145256Sjkoshy default: 1589153110Sru#ifdef DEBUG 1590145256Sjkoshy KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function)); 1591145256Sjkoshy#endif 1592145256Sjkoshy break; 1593145256Sjkoshy 1594145256Sjkoshy } 1595145256Sjkoshy 1596145256Sjkoshy return 0; 1597145256Sjkoshy} 1598145256Sjkoshy 1599145256Sjkoshy/* 1600145256Sjkoshy * allocate a 'struct pmc_owner' descriptor in the owner hash table. 1601145256Sjkoshy */ 1602145256Sjkoshy 1603145256Sjkoshystatic struct pmc_owner * 1604145256Sjkoshypmc_allocate_owner_descriptor(struct proc *p) 1605145256Sjkoshy{ 1606145256Sjkoshy uint32_t hindex; 1607145256Sjkoshy struct pmc_owner *po; 1608145256Sjkoshy struct pmc_ownerhash *poh; 1609145256Sjkoshy 1610145256Sjkoshy hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 1611145256Sjkoshy poh = &pmc_ownerhash[hindex]; 1612145256Sjkoshy 1613145256Sjkoshy /* allocate space for N pointers and one descriptor struct */ 1614145256Sjkoshy MALLOC(po, struct pmc_owner *, sizeof(struct pmc_owner), 1615147191Sjkoshy M_PMC, M_ZERO|M_WAITOK); 1616145256Sjkoshy 1617147191Sjkoshy po->po_sscount = po->po_error = po->po_flags = 0; 1618147191Sjkoshy po->po_file = NULL; 1619145256Sjkoshy po->po_owner = p; 1620147191Sjkoshy po->po_kthread = NULL; 1621145256Sjkoshy LIST_INIT(&po->po_pmcs); 1622145256Sjkoshy LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */ 1623145256Sjkoshy 1624147191Sjkoshy TAILQ_INIT(&po->po_logbuffers); 1625147191Sjkoshy mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc", MTX_SPIN); 1626147191Sjkoshy 1627145256Sjkoshy PMCDBG(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p", 1628145256Sjkoshy p, p->p_pid, p->p_comm, po); 1629145256Sjkoshy 1630145256Sjkoshy return po; 1631145256Sjkoshy} 1632145256Sjkoshy 1633147191Sjkoshystatic void 1634147191Sjkoshypmc_destroy_owner_descriptor(struct pmc_owner *po) 1635147191Sjkoshy{ 1636147191Sjkoshy 1637147191Sjkoshy PMCDBG(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)", 1638147191Sjkoshy po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm); 1639147191Sjkoshy 1640147191Sjkoshy mtx_destroy(&po->po_mtx); 1641147191Sjkoshy FREE(po, M_PMC); 1642147191Sjkoshy} 1643147191Sjkoshy 1644145256Sjkoshy/* 1645145256Sjkoshy * find the descriptor corresponding to process 'p', adding or removing it 1646145256Sjkoshy * as specified by 'mode'. 1647145256Sjkoshy */ 1648145256Sjkoshy 1649145256Sjkoshystatic struct pmc_process * 1650145256Sjkoshypmc_find_process_descriptor(struct proc *p, uint32_t mode) 1651145256Sjkoshy{ 1652145256Sjkoshy uint32_t hindex; 1653145256Sjkoshy struct pmc_process *pp, *ppnew; 1654145256Sjkoshy struct pmc_processhash *pph; 1655145256Sjkoshy 1656145256Sjkoshy hindex = PMC_HASH_PTR(p, pmc_processhashmask); 1657145256Sjkoshy pph = &pmc_processhash[hindex]; 1658145256Sjkoshy 1659145256Sjkoshy ppnew = NULL; 1660145256Sjkoshy 1661145256Sjkoshy /* 1662145256Sjkoshy * Pre-allocate memory in the FIND_ALLOCATE case since we 1663145256Sjkoshy * cannot call malloc(9) once we hold a spin lock. 1664145256Sjkoshy */ 1665145256Sjkoshy 1666145256Sjkoshy if (mode & PMC_FLAG_ALLOCATE) { 1667145256Sjkoshy /* allocate additional space for 'n' pmc pointers */ 1668145256Sjkoshy MALLOC(ppnew, struct pmc_process *, 1669145256Sjkoshy sizeof(struct pmc_process) + md->pmd_npmc * 1670145256Sjkoshy sizeof(struct pmc_targetstate), M_PMC, M_ZERO|M_WAITOK); 1671145256Sjkoshy } 1672145256Sjkoshy 1673145256Sjkoshy mtx_lock_spin(&pmc_processhash_mtx); 1674145256Sjkoshy LIST_FOREACH(pp, pph, pp_next) 1675145256Sjkoshy if (pp->pp_proc == p) 1676145256Sjkoshy break; 1677145256Sjkoshy 1678145256Sjkoshy if ((mode & PMC_FLAG_REMOVE) && pp != NULL) 1679145256Sjkoshy LIST_REMOVE(pp, pp_next); 1680145256Sjkoshy 1681145256Sjkoshy if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL && 1682145256Sjkoshy ppnew != NULL) { 1683145256Sjkoshy ppnew->pp_proc = p; 1684145256Sjkoshy LIST_INSERT_HEAD(pph, ppnew, pp_next); 1685145256Sjkoshy pp = ppnew; 1686145256Sjkoshy ppnew = NULL; 1687145256Sjkoshy } 1688145256Sjkoshy mtx_unlock_spin(&pmc_processhash_mtx); 1689145256Sjkoshy 1690145256Sjkoshy if (pp != NULL && ppnew != NULL) 1691145256Sjkoshy FREE(ppnew, M_PMC); 1692145256Sjkoshy 1693145256Sjkoshy return pp; 1694145256Sjkoshy} 1695145256Sjkoshy 1696145256Sjkoshy/* 1697145256Sjkoshy * remove a process descriptor from the process hash table. 1698145256Sjkoshy */ 1699145256Sjkoshy 1700145256Sjkoshystatic void 1701145256Sjkoshypmc_remove_process_descriptor(struct pmc_process *pp) 1702145256Sjkoshy{ 1703145256Sjkoshy KASSERT(pp->pp_refcnt == 0, 1704145256Sjkoshy ("[pmc,%d] Removing process descriptor %p with count %d", 1705145256Sjkoshy __LINE__, pp, pp->pp_refcnt)); 1706145256Sjkoshy 1707145256Sjkoshy mtx_lock_spin(&pmc_processhash_mtx); 1708145256Sjkoshy LIST_REMOVE(pp, pp_next); 1709145256Sjkoshy mtx_unlock_spin(&pmc_processhash_mtx); 1710145256Sjkoshy} 1711145256Sjkoshy 1712145256Sjkoshy 1713145256Sjkoshy/* 1714145256Sjkoshy * find an owner descriptor corresponding to proc 'p' 1715145256Sjkoshy */ 1716145256Sjkoshy 1717145256Sjkoshystatic struct pmc_owner * 1718145256Sjkoshypmc_find_owner_descriptor(struct proc *p) 1719145256Sjkoshy{ 1720145256Sjkoshy uint32_t hindex; 1721145256Sjkoshy struct pmc_owner *po; 1722145256Sjkoshy struct pmc_ownerhash *poh; 1723145256Sjkoshy 1724145256Sjkoshy hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 1725145256Sjkoshy poh = &pmc_ownerhash[hindex]; 1726145256Sjkoshy 1727145256Sjkoshy po = NULL; 1728145256Sjkoshy LIST_FOREACH(po, poh, po_next) 1729145256Sjkoshy if (po->po_owner == p) 1730145256Sjkoshy break; 1731145256Sjkoshy 1732145256Sjkoshy PMCDBG(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> " 1733145256Sjkoshy "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po); 1734145256Sjkoshy 1735145256Sjkoshy return po; 1736145256Sjkoshy} 1737145256Sjkoshy 1738145256Sjkoshy/* 1739145256Sjkoshy * pmc_allocate_pmc_descriptor 1740145256Sjkoshy * 1741145256Sjkoshy * Allocate a pmc descriptor and initialize its 1742145256Sjkoshy * fields. 1743145256Sjkoshy */ 1744145256Sjkoshy 1745145256Sjkoshystatic struct pmc * 1746145256Sjkoshypmc_allocate_pmc_descriptor(void) 1747145256Sjkoshy{ 1748145256Sjkoshy struct pmc *pmc; 1749145256Sjkoshy 1750145256Sjkoshy MALLOC(pmc, struct pmc *, sizeof(struct pmc), M_PMC, M_ZERO|M_WAITOK); 1751145256Sjkoshy 1752145256Sjkoshy if (pmc != NULL) { 1753145256Sjkoshy pmc->pm_owner = NULL; 1754145256Sjkoshy LIST_INIT(&pmc->pm_targets); 1755145256Sjkoshy } 1756145256Sjkoshy 1757145256Sjkoshy PMCDBG(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc); 1758145256Sjkoshy 1759145256Sjkoshy return pmc; 1760145256Sjkoshy} 1761145256Sjkoshy 1762145256Sjkoshy/* 1763145256Sjkoshy * Destroy a pmc descriptor. 1764145256Sjkoshy */ 1765145256Sjkoshy 1766145256Sjkoshystatic void 1767145256Sjkoshypmc_destroy_pmc_descriptor(struct pmc *pm) 1768145256Sjkoshy{ 1769145256Sjkoshy (void) pm; 1770145256Sjkoshy 1771153110Sru#ifdef DEBUG 1772145256Sjkoshy KASSERT(pm->pm_state == PMC_STATE_DELETED || 1773145256Sjkoshy pm->pm_state == PMC_STATE_FREE, 1774145256Sjkoshy ("[pmc,%d] destroying non-deleted PMC", __LINE__)); 1775145256Sjkoshy KASSERT(LIST_EMPTY(&pm->pm_targets), 1776145256Sjkoshy ("[pmc,%d] destroying pmc with targets", __LINE__)); 1777145256Sjkoshy KASSERT(pm->pm_owner == NULL, 1778145256Sjkoshy ("[pmc,%d] destroying pmc attached to an owner", __LINE__)); 1779145256Sjkoshy KASSERT(pm->pm_runcount == 0, 1780145256Sjkoshy ("[pmc,%d] pmc has non-zero run count %d", __LINE__, 1781145256Sjkoshy pm->pm_runcount)); 1782145256Sjkoshy#endif 1783145256Sjkoshy} 1784145256Sjkoshy 1785147191Sjkoshystatic void 1786147191Sjkoshypmc_wait_for_pmc_idle(struct pmc *pm) 1787147191Sjkoshy{ 1788153110Sru#ifdef DEBUG 1789147191Sjkoshy volatile int maxloop; 1790147191Sjkoshy 1791147191Sjkoshy maxloop = 100 * mp_ncpus; 1792147191Sjkoshy#endif 1793147191Sjkoshy 1794147191Sjkoshy /* 1795147191Sjkoshy * Loop (with a forced context switch) till the PMC's runcount 1796147191Sjkoshy * comes down to zero. 1797147191Sjkoshy */ 1798147191Sjkoshy while (atomic_load_acq_32(&pm->pm_runcount) > 0) { 1799153110Sru#ifdef DEBUG 1800147191Sjkoshy maxloop--; 1801147191Sjkoshy KASSERT(maxloop > 0, 1802147191Sjkoshy ("[pmc,%d] (ri%d, rc%d) waiting too long for " 1803147191Sjkoshy "pmc to be free", __LINE__, 1804147191Sjkoshy PMC_TO_ROWINDEX(pm), pm->pm_runcount)); 1805147191Sjkoshy#endif 1806147191Sjkoshy pmc_force_context_switch(); 1807147191Sjkoshy } 1808147191Sjkoshy} 1809147191Sjkoshy 1810145256Sjkoshy/* 1811145256Sjkoshy * This function does the following things: 1812145256Sjkoshy * 1813145256Sjkoshy * - detaches the PMC from hardware 1814145256Sjkoshy * - unlinks all target threads that were attached to it 1815145256Sjkoshy * - removes the PMC from its owner's list 1816145256Sjkoshy * - destroy's the PMC private mutex 1817145256Sjkoshy * 1818145256Sjkoshy * Once this function completes, the given pmc pointer can be safely 1819145256Sjkoshy * FREE'd by the caller. 1820145256Sjkoshy */ 1821145256Sjkoshy 1822145256Sjkoshystatic void 1823145256Sjkoshypmc_release_pmc_descriptor(struct pmc *pm) 1824145256Sjkoshy{ 1825145256Sjkoshy u_int ri, cpu; 1826145774Sjkoshy enum pmc_mode mode; 1827145256Sjkoshy struct pmc_hw *phw; 1828147191Sjkoshy struct pmc_owner *po; 1829145256Sjkoshy struct pmc_process *pp; 1830145256Sjkoshy struct pmc_target *ptgt, *tmp; 1831145256Sjkoshy struct pmc_binding pb; 1832145256Sjkoshy 1833145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 1834145256Sjkoshy 1835145256Sjkoshy KASSERT(pm, ("[pmc,%d] null pmc", __LINE__)); 1836145256Sjkoshy 1837145774Sjkoshy ri = PMC_TO_ROWINDEX(pm); 1838145774Sjkoshy mode = PMC_TO_MODE(pm); 1839145256Sjkoshy 1840145256Sjkoshy PMCDBG(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri, 1841145774Sjkoshy mode); 1842145256Sjkoshy 1843145256Sjkoshy /* 1844145256Sjkoshy * First, we take the PMC off hardware. 1845145256Sjkoshy */ 1846145301Simp cpu = 0; 1847145774Sjkoshy if (PMC_IS_SYSTEM_MODE(mode)) { 1848145256Sjkoshy 1849145256Sjkoshy /* 1850145256Sjkoshy * A system mode PMC runs on a specific CPU. Switch 1851145256Sjkoshy * to this CPU and turn hardware off. 1852145256Sjkoshy */ 1853145256Sjkoshy pmc_save_cpu_binding(&pb); 1854145256Sjkoshy 1855145774Sjkoshy cpu = PMC_TO_CPU(pm); 1856145256Sjkoshy 1857147191Sjkoshy pmc_select_cpu(cpu); 1858145256Sjkoshy 1859147191Sjkoshy /* switch off non-stalled CPUs */ 1860147191Sjkoshy if (pm->pm_state == PMC_STATE_RUNNING && 1861147867Sjkoshy pm->pm_stalled == 0) { 1862145256Sjkoshy 1863145256Sjkoshy phw = pmc_pcpu[cpu]->pc_hwpmcs[ri]; 1864145256Sjkoshy 1865145256Sjkoshy KASSERT(phw->phw_pmc == pm, 1866145256Sjkoshy ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)", 1867145256Sjkoshy __LINE__, ri, phw->phw_pmc, pm)); 1868145256Sjkoshy PMCDBG(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri); 1869145256Sjkoshy 1870145256Sjkoshy critical_enter(); 1871145256Sjkoshy md->pmd_stop_pmc(cpu, ri); 1872145256Sjkoshy critical_exit(); 1873145256Sjkoshy } 1874145256Sjkoshy 1875145256Sjkoshy PMCDBG(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri); 1876145256Sjkoshy 1877145256Sjkoshy critical_enter(); 1878145256Sjkoshy md->pmd_config_pmc(cpu, ri, NULL); 1879145256Sjkoshy critical_exit(); 1880145256Sjkoshy 1881147191Sjkoshy /* adjust the global and process count of SS mode PMCs */ 1882147191Sjkoshy if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) { 1883147191Sjkoshy po = pm->pm_owner; 1884147191Sjkoshy po->po_sscount--; 1885147191Sjkoshy if (po->po_sscount == 0) { 1886147191Sjkoshy atomic_subtract_rel_int(&pmc_ss_count, 1); 1887147191Sjkoshy LIST_REMOVE(po, po_ssnext); 1888147191Sjkoshy } 1889147191Sjkoshy } 1890147191Sjkoshy 1891145256Sjkoshy pm->pm_state = PMC_STATE_DELETED; 1892145256Sjkoshy 1893145256Sjkoshy pmc_restore_cpu_binding(&pb); 1894145256Sjkoshy 1895147191Sjkoshy /* 1896147191Sjkoshy * We could have references to this PMC structure in 1897147191Sjkoshy * the per-cpu sample queues. Wait for the queue to 1898147191Sjkoshy * drain. 1899147191Sjkoshy */ 1900147191Sjkoshy pmc_wait_for_pmc_idle(pm); 1901147191Sjkoshy 1902145774Sjkoshy } else if (PMC_IS_VIRTUAL_MODE(mode)) { 1903145256Sjkoshy 1904145256Sjkoshy /* 1905145256Sjkoshy * A virtual PMC could be running on multiple CPUs at 1906145256Sjkoshy * a given instant. 1907145256Sjkoshy * 1908145256Sjkoshy * By marking its state as DELETED, we ensure that 1909145256Sjkoshy * this PMC is never further scheduled on hardware. 1910145256Sjkoshy * 1911145256Sjkoshy * Then we wait till all CPUs are done with this PMC. 1912145256Sjkoshy */ 1913145256Sjkoshy pm->pm_state = PMC_STATE_DELETED; 1914145256Sjkoshy 1915145256Sjkoshy 1916147191Sjkoshy /* Wait for the PMCs runcount to come to zero. */ 1917147191Sjkoshy pmc_wait_for_pmc_idle(pm); 1918145256Sjkoshy 1919145256Sjkoshy /* 1920145256Sjkoshy * At this point the PMC is off all CPUs and cannot be 1921145256Sjkoshy * freshly scheduled onto a CPU. It is now safe to 1922145256Sjkoshy * unlink all targets from this PMC. If a 1923145256Sjkoshy * process-record's refcount falls to zero, we remove 1924145256Sjkoshy * it from the hash table. The module-wide SX lock 1925145256Sjkoshy * protects us from races. 1926145256Sjkoshy */ 1927145256Sjkoshy LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) { 1928145256Sjkoshy pp = ptgt->pt_process; 1929145256Sjkoshy pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */ 1930145256Sjkoshy 1931145256Sjkoshy PMCDBG(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt); 1932145256Sjkoshy 1933145256Sjkoshy /* 1934145256Sjkoshy * If the target process record shows that no 1935145256Sjkoshy * PMCs are attached to it, reclaim its space. 1936145256Sjkoshy */ 1937145256Sjkoshy 1938145256Sjkoshy if (pp->pp_refcnt == 0) { 1939145256Sjkoshy pmc_remove_process_descriptor(pp); 1940145256Sjkoshy FREE(pp, M_PMC); 1941145256Sjkoshy } 1942145256Sjkoshy } 1943145256Sjkoshy 1944145256Sjkoshy cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */ 1945145256Sjkoshy 1946145256Sjkoshy } 1947145256Sjkoshy 1948145256Sjkoshy /* 1949145256Sjkoshy * Release any MD resources 1950145256Sjkoshy */ 1951145256Sjkoshy 1952145256Sjkoshy (void) md->pmd_release_pmc(cpu, ri, pm); 1953145256Sjkoshy 1954145256Sjkoshy /* 1955145256Sjkoshy * Update row disposition 1956145256Sjkoshy */ 1957145256Sjkoshy 1958145774Sjkoshy if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) 1959145256Sjkoshy PMC_UNMARK_ROW_STANDALONE(ri); 1960145256Sjkoshy else 1961145256Sjkoshy PMC_UNMARK_ROW_THREAD(ri); 1962145256Sjkoshy 1963145256Sjkoshy /* unlink from the owner's list */ 1964147191Sjkoshy if (pm->pm_owner) { 1965147191Sjkoshy LIST_REMOVE(pm, pm_next); 1966147191Sjkoshy pm->pm_owner = NULL; 1967147191Sjkoshy } 1968145256Sjkoshy 1969145256Sjkoshy pmc_destroy_pmc_descriptor(pm); 1970145256Sjkoshy} 1971145256Sjkoshy 1972145256Sjkoshy/* 1973145256Sjkoshy * Register an owner and a pmc. 1974145256Sjkoshy */ 1975145256Sjkoshy 1976145256Sjkoshystatic int 1977145256Sjkoshypmc_register_owner(struct proc *p, struct pmc *pmc) 1978145256Sjkoshy{ 1979145256Sjkoshy struct pmc_owner *po; 1980145256Sjkoshy 1981145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 1982145256Sjkoshy 1983145774Sjkoshy if ((po = pmc_find_owner_descriptor(p)) == NULL) 1984147191Sjkoshy if ((po = pmc_allocate_owner_descriptor(p)) == NULL) 1985145256Sjkoshy return ENOMEM; 1986145256Sjkoshy 1987145256Sjkoshy KASSERT(pmc->pm_owner == NULL, 1988145256Sjkoshy ("[pmc,%d] attempting to own an initialized PMC", __LINE__)); 1989145256Sjkoshy pmc->pm_owner = po; 1990145256Sjkoshy 1991147191Sjkoshy LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next); 1992145256Sjkoshy 1993145256Sjkoshy PROC_LOCK(p); 1994145256Sjkoshy p->p_flag |= P_HWPMC; 1995145256Sjkoshy PROC_UNLOCK(p); 1996145256Sjkoshy 1997147191Sjkoshy if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1998147191Sjkoshy pmclog_process_pmcallocate(pmc); 1999145256Sjkoshy 2000147191Sjkoshy PMCDBG(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p", 2001147191Sjkoshy po, pmc); 2002147191Sjkoshy 2003145256Sjkoshy return 0; 2004145256Sjkoshy} 2005145256Sjkoshy 2006145256Sjkoshy/* 2007145256Sjkoshy * Return the current row disposition: 2008145256Sjkoshy * == 0 => FREE 2009145256Sjkoshy * > 0 => PROCESS MODE 2010145256Sjkoshy * < 0 => SYSTEM MODE 2011145256Sjkoshy */ 2012145256Sjkoshy 2013145256Sjkoshyint 2014145256Sjkoshypmc_getrowdisp(int ri) 2015145256Sjkoshy{ 2016145256Sjkoshy return pmc_pmcdisp[ri]; 2017145256Sjkoshy} 2018145256Sjkoshy 2019145256Sjkoshy/* 2020145256Sjkoshy * Check if a PMC at row index 'ri' can be allocated to the current 2021145256Sjkoshy * process. 2022145256Sjkoshy * 2023145256Sjkoshy * Allocation can fail if: 2024145256Sjkoshy * - the current process is already being profiled by a PMC at index 'ri', 2025145256Sjkoshy * attached to it via OP_PMCATTACH. 2026145256Sjkoshy * - the current process has already allocated a PMC at index 'ri' 2027145256Sjkoshy * via OP_ALLOCATE. 2028145256Sjkoshy */ 2029145256Sjkoshy 2030145256Sjkoshystatic int 2031145774Sjkoshypmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu) 2032145256Sjkoshy{ 2033145774Sjkoshy enum pmc_mode mode; 2034145774Sjkoshy struct pmc *pm; 2035145256Sjkoshy struct pmc_owner *po; 2036145256Sjkoshy struct pmc_process *pp; 2037145256Sjkoshy 2038145774Sjkoshy PMCDBG(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d " 2039145774Sjkoshy "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu); 2040145256Sjkoshy 2041145774Sjkoshy /* 2042145774Sjkoshy * We shouldn't have already allocated a process-mode PMC at 2043145774Sjkoshy * row index 'ri'. 2044145774Sjkoshy * 2045145774Sjkoshy * We shouldn't have allocated a system-wide PMC on the same 2046145774Sjkoshy * CPU and same RI. 2047145774Sjkoshy */ 2048145256Sjkoshy if ((po = pmc_find_owner_descriptor(p)) != NULL) 2049147191Sjkoshy LIST_FOREACH(pm, &po->po_pmcs, pm_next) { 2050145774Sjkoshy if (PMC_TO_ROWINDEX(pm) == ri) { 2051145774Sjkoshy mode = PMC_TO_MODE(pm); 2052145774Sjkoshy if (PMC_IS_VIRTUAL_MODE(mode)) 2053145774Sjkoshy return EEXIST; 2054145774Sjkoshy if (PMC_IS_SYSTEM_MODE(mode) && 2055145774Sjkoshy (int) PMC_TO_CPU(pm) == cpu) 2056145774Sjkoshy return EEXIST; 2057145774Sjkoshy } 2058145774Sjkoshy } 2059145256Sjkoshy 2060145774Sjkoshy /* 2061145774Sjkoshy * We also shouldn't be the target of any PMC at this index 2062145774Sjkoshy * since otherwise a PMC_ATTACH to ourselves will fail. 2063145774Sjkoshy */ 2064145256Sjkoshy if ((pp = pmc_find_process_descriptor(p, 0)) != NULL) 2065145256Sjkoshy if (pp->pp_pmcs[ri].pp_pmc) 2066145256Sjkoshy return EEXIST; 2067145256Sjkoshy 2068145256Sjkoshy PMCDBG(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok", 2069145256Sjkoshy p, p->p_pid, p->p_comm, ri); 2070145256Sjkoshy 2071145256Sjkoshy return 0; 2072145256Sjkoshy} 2073145256Sjkoshy 2074145256Sjkoshy/* 2075145256Sjkoshy * Check if a given PMC at row index 'ri' can be currently used in 2076145256Sjkoshy * mode 'mode'. 2077145256Sjkoshy */ 2078145256Sjkoshy 2079145256Sjkoshystatic int 2080145256Sjkoshypmc_can_allocate_row(int ri, enum pmc_mode mode) 2081145256Sjkoshy{ 2082145256Sjkoshy enum pmc_disp disp; 2083145256Sjkoshy 2084145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 2085145256Sjkoshy 2086145256Sjkoshy PMCDBG(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode); 2087145256Sjkoshy 2088145256Sjkoshy if (PMC_IS_SYSTEM_MODE(mode)) 2089145256Sjkoshy disp = PMC_DISP_STANDALONE; 2090145256Sjkoshy else 2091145256Sjkoshy disp = PMC_DISP_THREAD; 2092145256Sjkoshy 2093145256Sjkoshy /* 2094145256Sjkoshy * check disposition for PMC row 'ri': 2095145256Sjkoshy * 2096145256Sjkoshy * Expected disposition Row-disposition Result 2097145256Sjkoshy * 2098145256Sjkoshy * STANDALONE STANDALONE or FREE proceed 2099145256Sjkoshy * STANDALONE THREAD fail 2100145256Sjkoshy * THREAD THREAD or FREE proceed 2101145256Sjkoshy * THREAD STANDALONE fail 2102145256Sjkoshy */ 2103145256Sjkoshy 2104145256Sjkoshy if (!PMC_ROW_DISP_IS_FREE(ri) && 2105145256Sjkoshy !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) && 2106145256Sjkoshy !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri))) 2107145256Sjkoshy return EBUSY; 2108145256Sjkoshy 2109145256Sjkoshy /* 2110145256Sjkoshy * All OK 2111145256Sjkoshy */ 2112145256Sjkoshy 2113145256Sjkoshy PMCDBG(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode); 2114145256Sjkoshy 2115145256Sjkoshy return 0; 2116145256Sjkoshy 2117145256Sjkoshy} 2118145256Sjkoshy 2119145256Sjkoshy/* 2120145774Sjkoshy * Find a PMC descriptor with user handle 'pmcid' for thread 'td'. 2121145256Sjkoshy */ 2122145256Sjkoshy 2123145256Sjkoshystatic struct pmc * 2124145256Sjkoshypmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid) 2125145256Sjkoshy{ 2126147191Sjkoshy struct pmc *pm; 2127145256Sjkoshy 2128145774Sjkoshy KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc, 2129145774Sjkoshy ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__, 2130145774Sjkoshy PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc)); 2131145256Sjkoshy 2132147191Sjkoshy LIST_FOREACH(pm, &po->po_pmcs, pm_next) 2133147191Sjkoshy if (pm->pm_id == pmcid) 2134147191Sjkoshy return pm; 2135145256Sjkoshy 2136145256Sjkoshy return NULL; 2137145256Sjkoshy} 2138145256Sjkoshy 2139145256Sjkoshystatic int 2140145256Sjkoshypmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc) 2141145256Sjkoshy{ 2142145256Sjkoshy 2143145256Sjkoshy struct pmc *pm; 2144145256Sjkoshy struct pmc_owner *po; 2145145256Sjkoshy 2146145256Sjkoshy PMCDBG(PMC,FND,1, "find-pmc id=%d", pmcid); 2147145256Sjkoshy 2148145256Sjkoshy if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) 2149145256Sjkoshy return ESRCH; 2150145256Sjkoshy 2151145256Sjkoshy if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL) 2152145256Sjkoshy return EINVAL; 2153145256Sjkoshy 2154145256Sjkoshy PMCDBG(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm); 2155145256Sjkoshy 2156145256Sjkoshy *pmc = pm; 2157145256Sjkoshy return 0; 2158145256Sjkoshy} 2159145256Sjkoshy 2160145256Sjkoshy/* 2161145256Sjkoshy * Start a PMC. 2162145256Sjkoshy */ 2163145256Sjkoshy 2164145256Sjkoshystatic int 2165145256Sjkoshypmc_start(struct pmc *pm) 2166145256Sjkoshy{ 2167145256Sjkoshy int error, cpu, ri; 2168145774Sjkoshy enum pmc_mode mode; 2169147191Sjkoshy struct pmc_owner *po; 2170145256Sjkoshy struct pmc_binding pb; 2171145256Sjkoshy 2172145256Sjkoshy KASSERT(pm != NULL, 2173145256Sjkoshy ("[pmc,%d] null pm", __LINE__)); 2174145256Sjkoshy 2175145774Sjkoshy mode = PMC_TO_MODE(pm); 2176145774Sjkoshy ri = PMC_TO_ROWINDEX(pm); 2177145774Sjkoshy error = 0; 2178145256Sjkoshy 2179145774Sjkoshy PMCDBG(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri); 2180145774Sjkoshy 2181147191Sjkoshy po = pm->pm_owner; 2182145256Sjkoshy 2183145774Sjkoshy if (PMC_IS_VIRTUAL_MODE(mode)) { 2184145256Sjkoshy 2185145256Sjkoshy /* 2186147191Sjkoshy * If a PMCATTACH has never been done on this PMC, 2187147191Sjkoshy * attach it to its owner process. 2188145256Sjkoshy */ 2189145256Sjkoshy 2190145256Sjkoshy if (LIST_EMPTY(&pm->pm_targets)) 2191147191Sjkoshy error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH : 2192147191Sjkoshy pmc_attach_process(po->po_owner, pm); 2193145256Sjkoshy 2194145774Sjkoshy /* 2195147191Sjkoshy * Disallow PMCSTART if a logfile is required but has not 2196147191Sjkoshy * been configured yet. 2197145774Sjkoshy */ 2198145256Sjkoshy 2199147191Sjkoshy if (error == 0 && (pm->pm_flags & PMC_F_NEEDS_LOGFILE) && 2200147191Sjkoshy (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 2201147191Sjkoshy error = EDOOFUS; 2202147191Sjkoshy 2203145256Sjkoshy /* 2204147191Sjkoshy * If the PMC is attached to its owner, then force a context 2205147191Sjkoshy * switch to ensure that the MD state gets set correctly. 2206145256Sjkoshy */ 2207145256Sjkoshy 2208147191Sjkoshy if (error == 0) { 2209147191Sjkoshy pm->pm_state = PMC_STATE_RUNNING; 2210147191Sjkoshy if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) 2211147191Sjkoshy pmc_force_context_switch(); 2212147191Sjkoshy } 2213147191Sjkoshy 2214145774Sjkoshy return error; 2215147191Sjkoshy } 2216145256Sjkoshy 2217147191Sjkoshy 2218147191Sjkoshy /* 2219147191Sjkoshy * A system-wide PMC. 2220147191Sjkoshy */ 2221147191Sjkoshy 2222147191Sjkoshy if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) && 2223147191Sjkoshy (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 2224147191Sjkoshy return EDOOFUS; /* programming error */ 2225147191Sjkoshy 2226147191Sjkoshy /* 2227147191Sjkoshy * Add the owner to the global list if this is a system-wide 2228147191Sjkoshy * sampling PMC. 2229147191Sjkoshy */ 2230147191Sjkoshy 2231147191Sjkoshy if (mode == PMC_MODE_SS) { 2232147191Sjkoshy if (po->po_sscount == 0) { 2233147191Sjkoshy LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext); 2234147191Sjkoshy atomic_add_rel_int(&pmc_ss_count, 1); 2235147191Sjkoshy PMCDBG(PMC,OPS,1, "po=%p in global list", po); 2236147191Sjkoshy } 2237147191Sjkoshy po->po_sscount++; 2238145256Sjkoshy } 2239145256Sjkoshy 2240145256Sjkoshy /* 2241147191Sjkoshy * Move to the CPU associated with this 2242145256Sjkoshy * PMC, and start the hardware. 2243145256Sjkoshy */ 2244145256Sjkoshy 2245145256Sjkoshy pmc_save_cpu_binding(&pb); 2246145256Sjkoshy 2247145774Sjkoshy cpu = PMC_TO_CPU(pm); 2248145256Sjkoshy 2249145256Sjkoshy if (pmc_cpu_is_disabled(cpu)) 2250145256Sjkoshy return ENXIO; 2251145256Sjkoshy 2252145256Sjkoshy pmc_select_cpu(cpu); 2253145256Sjkoshy 2254145256Sjkoshy /* 2255145256Sjkoshy * global PMCs are configured at allocation time 2256145256Sjkoshy * so write out the initial value and start the PMC. 2257145256Sjkoshy */ 2258145256Sjkoshy 2259147191Sjkoshy pm->pm_state = PMC_STATE_RUNNING; 2260147191Sjkoshy 2261145774Sjkoshy critical_enter(); 2262145256Sjkoshy if ((error = md->pmd_write_pmc(cpu, ri, 2263145774Sjkoshy PMC_IS_SAMPLING_MODE(mode) ? 2264145256Sjkoshy pm->pm_sc.pm_reloadcount : 2265145256Sjkoshy pm->pm_sc.pm_initial)) == 0) 2266145256Sjkoshy error = md->pmd_start_pmc(cpu, ri); 2267145774Sjkoshy critical_exit(); 2268145256Sjkoshy 2269145256Sjkoshy pmc_restore_cpu_binding(&pb); 2270145256Sjkoshy 2271145256Sjkoshy return error; 2272145256Sjkoshy} 2273145256Sjkoshy 2274145256Sjkoshy/* 2275145256Sjkoshy * Stop a PMC. 2276145256Sjkoshy */ 2277145256Sjkoshy 2278145256Sjkoshystatic int 2279145256Sjkoshypmc_stop(struct pmc *pm) 2280145256Sjkoshy{ 2281145774Sjkoshy int cpu, error, ri; 2282147191Sjkoshy struct pmc_owner *po; 2283145256Sjkoshy struct pmc_binding pb; 2284145256Sjkoshy 2285145256Sjkoshy KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__)); 2286145256Sjkoshy 2287145774Sjkoshy PMCDBG(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm, 2288145774Sjkoshy PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm)); 2289145256Sjkoshy 2290145256Sjkoshy pm->pm_state = PMC_STATE_STOPPED; 2291145256Sjkoshy 2292145256Sjkoshy /* 2293145256Sjkoshy * If the PMC is a virtual mode one, changing the state to 2294145256Sjkoshy * non-RUNNING is enough to ensure that the PMC never gets 2295145256Sjkoshy * scheduled. 2296145256Sjkoshy * 2297145256Sjkoshy * If this PMC is current running on a CPU, then it will 2298145256Sjkoshy * handled correctly at the time its target process is context 2299145256Sjkoshy * switched out. 2300145256Sjkoshy */ 2301145256Sjkoshy 2302145774Sjkoshy if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 2303145256Sjkoshy return 0; 2304145256Sjkoshy 2305145256Sjkoshy /* 2306145256Sjkoshy * A system-mode PMC. Move to the CPU associated with 2307145256Sjkoshy * this PMC, and stop the hardware. We update the 2308145256Sjkoshy * 'initial count' so that a subsequent PMCSTART will 2309145256Sjkoshy * resume counting from the current hardware count. 2310145256Sjkoshy */ 2311145256Sjkoshy 2312145256Sjkoshy pmc_save_cpu_binding(&pb); 2313145256Sjkoshy 2314145774Sjkoshy cpu = PMC_TO_CPU(pm); 2315145256Sjkoshy 2316145774Sjkoshy KASSERT(cpu >= 0 && cpu < mp_ncpus, 2317145774Sjkoshy ("[pmc,%d] illegal cpu=%d", __LINE__, cpu)); 2318145774Sjkoshy 2319145256Sjkoshy if (pmc_cpu_is_disabled(cpu)) 2320145256Sjkoshy return ENXIO; 2321145256Sjkoshy 2322145256Sjkoshy pmc_select_cpu(cpu); 2323145256Sjkoshy 2324145774Sjkoshy ri = PMC_TO_ROWINDEX(pm); 2325145256Sjkoshy 2326145774Sjkoshy critical_enter(); 2327145774Sjkoshy if ((error = md->pmd_stop_pmc(cpu, ri)) == 0) 2328145774Sjkoshy error = md->pmd_read_pmc(cpu, ri, &pm->pm_sc.pm_initial); 2329145774Sjkoshy critical_exit(); 2330145774Sjkoshy 2331145256Sjkoshy pmc_restore_cpu_binding(&pb); 2332145256Sjkoshy 2333147191Sjkoshy po = pm->pm_owner; 2334147191Sjkoshy 2335147191Sjkoshy /* remove this owner from the global list of SS PMC owners */ 2336147191Sjkoshy if (PMC_TO_MODE(pm) == PMC_MODE_SS) { 2337147191Sjkoshy po->po_sscount--; 2338147191Sjkoshy if (po->po_sscount == 0) { 2339147191Sjkoshy atomic_subtract_rel_int(&pmc_ss_count, 1); 2340147191Sjkoshy LIST_REMOVE(po, po_ssnext); 2341147191Sjkoshy PMCDBG(PMC,OPS,2,"po=%p removed from global list", po); 2342147191Sjkoshy } 2343147191Sjkoshy } 2344147191Sjkoshy 2345145256Sjkoshy return error; 2346145256Sjkoshy} 2347145256Sjkoshy 2348145256Sjkoshy 2349153110Sru#ifdef DEBUG 2350145256Sjkoshystatic const char *pmc_op_to_name[] = { 2351145256Sjkoshy#undef __PMC_OP 2352145256Sjkoshy#define __PMC_OP(N, D) #N , 2353145256Sjkoshy __PMC_OPS() 2354145256Sjkoshy NULL 2355145256Sjkoshy}; 2356145256Sjkoshy#endif 2357145256Sjkoshy 2358145256Sjkoshy/* 2359145256Sjkoshy * The syscall interface 2360145256Sjkoshy */ 2361145256Sjkoshy 2362145256Sjkoshy#define PMC_GET_SX_XLOCK(...) do { \ 2363145256Sjkoshy sx_xlock(&pmc_sx); \ 2364145256Sjkoshy if (pmc_hook == NULL) { \ 2365145256Sjkoshy sx_xunlock(&pmc_sx); \ 2366145256Sjkoshy return __VA_ARGS__; \ 2367145256Sjkoshy } \ 2368145256Sjkoshy} while (0) 2369145256Sjkoshy 2370145256Sjkoshy#define PMC_DOWNGRADE_SX() do { \ 2371145256Sjkoshy sx_downgrade(&pmc_sx); \ 2372145256Sjkoshy is_sx_downgraded = 1; \ 2373145256Sjkoshy} while (0) 2374145256Sjkoshy 2375145256Sjkoshystatic int 2376145256Sjkoshypmc_syscall_handler(struct thread *td, void *syscall_args) 2377145256Sjkoshy{ 2378145256Sjkoshy int error, is_sx_downgraded, op; 2379145256Sjkoshy struct pmc_syscall_args *c; 2380145256Sjkoshy void *arg; 2381145256Sjkoshy 2382145256Sjkoshy PMC_GET_SX_XLOCK(ENOSYS); 2383145256Sjkoshy 2384147191Sjkoshy DROP_GIANT(); 2385147191Sjkoshy 2386145256Sjkoshy is_sx_downgraded = 0; 2387145256Sjkoshy 2388145256Sjkoshy c = (struct pmc_syscall_args *) syscall_args; 2389145256Sjkoshy 2390145256Sjkoshy op = c->pmop_code; 2391145256Sjkoshy arg = c->pmop_data; 2392145256Sjkoshy 2393145256Sjkoshy PMCDBG(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op, 2394145256Sjkoshy pmc_op_to_name[op], arg); 2395145256Sjkoshy 2396145256Sjkoshy error = 0; 2397145256Sjkoshy atomic_add_int(&pmc_stats.pm_syscalls, 1); 2398145256Sjkoshy 2399145256Sjkoshy switch(op) 2400145256Sjkoshy { 2401145256Sjkoshy 2402145256Sjkoshy 2403145256Sjkoshy /* 2404145256Sjkoshy * Configure a log file. 2405145256Sjkoshy * 2406145256Sjkoshy * XXX This OP will be reworked. 2407145256Sjkoshy */ 2408145256Sjkoshy 2409145256Sjkoshy case PMC_OP_CONFIGURELOG: 2410145256Sjkoshy { 2411156466Sjkoshy struct pmc *pm; 2412145256Sjkoshy struct pmc_owner *po; 2413145256Sjkoshy struct pmc_op_configurelog cl; 2414145256Sjkoshy struct proc *p; 2415145256Sjkoshy 2416145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 2417145256Sjkoshy 2418145256Sjkoshy if ((error = copyin(arg, &cl, sizeof(cl))) != 0) 2419145256Sjkoshy break; 2420145256Sjkoshy 2421145256Sjkoshy /* mark this process as owning a log file */ 2422145256Sjkoshy p = td->td_proc; 2423145256Sjkoshy if ((po = pmc_find_owner_descriptor(p)) == NULL) 2424147191Sjkoshy if ((po = pmc_allocate_owner_descriptor(p)) == NULL) { 2425147191Sjkoshy error = ENOMEM; 2426147191Sjkoshy break; 2427147191Sjkoshy } 2428145256Sjkoshy 2429147191Sjkoshy /* 2430147191Sjkoshy * If a valid fd was passed in, try to configure that, 2431147191Sjkoshy * otherwise if 'fd' was less than zero and there was 2432147191Sjkoshy * a log file configured, flush its buffers and 2433147191Sjkoshy * de-configure it. 2434147191Sjkoshy */ 2435147191Sjkoshy if (cl.pm_logfd >= 0) 2436147191Sjkoshy error = pmclog_configure_log(po, cl.pm_logfd); 2437147191Sjkoshy else if (po->po_flags & PMC_PO_OWNS_LOGFILE) { 2438147191Sjkoshy pmclog_process_closelog(po); 2439147191Sjkoshy error = pmclog_flush(po); 2440156466Sjkoshy if (error == 0) { 2441156466Sjkoshy LIST_FOREACH(pm, &po->po_pmcs, pm_next) 2442156466Sjkoshy if (pm->pm_flags & PMC_F_NEEDS_LOGFILE) 2443156466Sjkoshy pmc_stop(pm); 2444147191Sjkoshy error = pmclog_deconfigure_log(po); 2445156466Sjkoshy } 2446147191Sjkoshy } else 2447147191Sjkoshy error = EINVAL; 2448147191Sjkoshy } 2449147191Sjkoshy break; 2450147191Sjkoshy 2451147191Sjkoshy 2452147191Sjkoshy /* 2453147191Sjkoshy * Flush a log file. 2454147191Sjkoshy */ 2455147191Sjkoshy 2456147191Sjkoshy case PMC_OP_FLUSHLOG: 2457147191Sjkoshy { 2458147191Sjkoshy struct pmc_owner *po; 2459147191Sjkoshy 2460147191Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 2461147191Sjkoshy 2462147191Sjkoshy if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 2463147191Sjkoshy error = EINVAL; 2464145256Sjkoshy break; 2465147191Sjkoshy } 2466145256Sjkoshy 2467147191Sjkoshy error = pmclog_flush(po); 2468145256Sjkoshy } 2469145256Sjkoshy break; 2470145256Sjkoshy 2471145256Sjkoshy /* 2472145256Sjkoshy * Retrieve hardware configuration. 2473145256Sjkoshy */ 2474145256Sjkoshy 2475145256Sjkoshy case PMC_OP_GETCPUINFO: /* CPU information */ 2476145256Sjkoshy { 2477145256Sjkoshy struct pmc_op_getcpuinfo gci; 2478145256Sjkoshy 2479145256Sjkoshy gci.pm_cputype = md->pmd_cputype; 2480145774Sjkoshy gci.pm_ncpu = mp_ncpus; 2481145256Sjkoshy gci.pm_npmc = md->pmd_npmc; 2482145256Sjkoshy gci.pm_nclass = md->pmd_nclass; 2483145256Sjkoshy bcopy(md->pmd_classes, &gci.pm_classes, 2484145256Sjkoshy sizeof(gci.pm_classes)); 2485145256Sjkoshy error = copyout(&gci, arg, sizeof(gci)); 2486145256Sjkoshy } 2487145256Sjkoshy break; 2488145256Sjkoshy 2489145256Sjkoshy 2490145256Sjkoshy /* 2491145256Sjkoshy * Get module statistics 2492145256Sjkoshy */ 2493145256Sjkoshy 2494145256Sjkoshy case PMC_OP_GETDRIVERSTATS: 2495145256Sjkoshy { 2496145256Sjkoshy struct pmc_op_getdriverstats gms; 2497145256Sjkoshy 2498145256Sjkoshy bcopy(&pmc_stats, &gms, sizeof(gms)); 2499145256Sjkoshy error = copyout(&gms, arg, sizeof(gms)); 2500145256Sjkoshy } 2501145256Sjkoshy break; 2502145256Sjkoshy 2503145256Sjkoshy 2504145256Sjkoshy /* 2505145256Sjkoshy * Retrieve module version number 2506145256Sjkoshy */ 2507145256Sjkoshy 2508145256Sjkoshy case PMC_OP_GETMODULEVERSION: 2509145256Sjkoshy { 2510147191Sjkoshy uint32_t cv, modv; 2511147191Sjkoshy 2512147191Sjkoshy /* retrieve the client's idea of the ABI version */ 2513147191Sjkoshy if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0) 2514147191Sjkoshy break; 2515147191Sjkoshy /* don't service clients newer than our driver */ 2516147191Sjkoshy modv = PMC_VERSION; 2517147191Sjkoshy if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) { 2518147191Sjkoshy error = EPROGMISMATCH; 2519147191Sjkoshy break; 2520147191Sjkoshy } 2521147191Sjkoshy error = copyout(&modv, arg, sizeof(int)); 2522145256Sjkoshy } 2523145256Sjkoshy break; 2524145256Sjkoshy 2525145256Sjkoshy 2526145256Sjkoshy /* 2527145256Sjkoshy * Retrieve the state of all the PMCs on a given 2528145256Sjkoshy * CPU. 2529145256Sjkoshy */ 2530145256Sjkoshy 2531145256Sjkoshy case PMC_OP_GETPMCINFO: 2532145256Sjkoshy { 2533145256Sjkoshy uint32_t cpu, n, npmc; 2534145256Sjkoshy size_t pmcinfo_size; 2535145256Sjkoshy struct pmc *pm; 2536145256Sjkoshy struct pmc_info *p, *pmcinfo; 2537145256Sjkoshy struct pmc_op_getpmcinfo *gpi; 2538145256Sjkoshy struct pmc_owner *po; 2539145256Sjkoshy struct pmc_binding pb; 2540145256Sjkoshy 2541145256Sjkoshy PMC_DOWNGRADE_SX(); 2542145256Sjkoshy 2543145256Sjkoshy gpi = (struct pmc_op_getpmcinfo *) arg; 2544145256Sjkoshy 2545145256Sjkoshy if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0) 2546145256Sjkoshy break; 2547145256Sjkoshy 2548145256Sjkoshy if (cpu >= (unsigned int) mp_ncpus) { 2549145256Sjkoshy error = EINVAL; 2550145256Sjkoshy break; 2551145256Sjkoshy } 2552145256Sjkoshy 2553145256Sjkoshy if (pmc_cpu_is_disabled(cpu)) { 2554145256Sjkoshy error = ENXIO; 2555145256Sjkoshy break; 2556145256Sjkoshy } 2557145256Sjkoshy 2558145256Sjkoshy /* switch to CPU 'cpu' */ 2559145256Sjkoshy pmc_save_cpu_binding(&pb); 2560145256Sjkoshy pmc_select_cpu(cpu); 2561145256Sjkoshy 2562145256Sjkoshy npmc = md->pmd_npmc; 2563145256Sjkoshy 2564145256Sjkoshy pmcinfo_size = npmc * sizeof(struct pmc_info); 2565145256Sjkoshy MALLOC(pmcinfo, struct pmc_info *, pmcinfo_size, M_PMC, 2566145256Sjkoshy M_WAITOK); 2567145256Sjkoshy 2568145256Sjkoshy p = pmcinfo; 2569145256Sjkoshy 2570145256Sjkoshy for (n = 0; n < md->pmd_npmc; n++, p++) { 2571145256Sjkoshy 2572145256Sjkoshy if ((error = md->pmd_describe(cpu, n, p, &pm)) != 0) 2573145256Sjkoshy break; 2574145256Sjkoshy 2575145256Sjkoshy if (PMC_ROW_DISP_IS_STANDALONE(n)) 2576145256Sjkoshy p->pm_rowdisp = PMC_DISP_STANDALONE; 2577145256Sjkoshy else if (PMC_ROW_DISP_IS_THREAD(n)) 2578145256Sjkoshy p->pm_rowdisp = PMC_DISP_THREAD; 2579145256Sjkoshy else 2580145256Sjkoshy p->pm_rowdisp = PMC_DISP_FREE; 2581145256Sjkoshy 2582145256Sjkoshy p->pm_ownerpid = -1; 2583145256Sjkoshy 2584145256Sjkoshy if (pm == NULL) /* no PMC associated */ 2585145256Sjkoshy continue; 2586145256Sjkoshy 2587145256Sjkoshy po = pm->pm_owner; 2588145256Sjkoshy 2589145256Sjkoshy KASSERT(po->po_owner != NULL, 2590145256Sjkoshy ("[pmc,%d] pmc_owner had a null proc pointer", 2591145256Sjkoshy __LINE__)); 2592145256Sjkoshy 2593145256Sjkoshy p->pm_ownerpid = po->po_owner->p_pid; 2594145774Sjkoshy p->pm_mode = PMC_TO_MODE(pm); 2595145256Sjkoshy p->pm_event = pm->pm_event; 2596145256Sjkoshy p->pm_flags = pm->pm_flags; 2597145256Sjkoshy 2598145774Sjkoshy if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 2599145256Sjkoshy p->pm_reloadcount = 2600145256Sjkoshy pm->pm_sc.pm_reloadcount; 2601145256Sjkoshy } 2602145256Sjkoshy 2603145256Sjkoshy pmc_restore_cpu_binding(&pb); 2604145256Sjkoshy 2605145256Sjkoshy /* now copy out the PMC info collected */ 2606145256Sjkoshy if (error == 0) 2607145256Sjkoshy error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size); 2608145256Sjkoshy 2609145256Sjkoshy FREE(pmcinfo, M_PMC); 2610145256Sjkoshy } 2611145256Sjkoshy break; 2612145256Sjkoshy 2613145256Sjkoshy 2614145256Sjkoshy /* 2615145256Sjkoshy * Set the administrative state of a PMC. I.e. whether 2616145256Sjkoshy * the PMC is to be used or not. 2617145256Sjkoshy */ 2618145256Sjkoshy 2619145256Sjkoshy case PMC_OP_PMCADMIN: 2620145256Sjkoshy { 2621145256Sjkoshy int cpu, ri; 2622145256Sjkoshy enum pmc_state request; 2623145256Sjkoshy struct pmc_cpu *pc; 2624145256Sjkoshy struct pmc_hw *phw; 2625145256Sjkoshy struct pmc_op_pmcadmin pma; 2626145256Sjkoshy struct pmc_binding pb; 2627145256Sjkoshy 2628145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 2629145256Sjkoshy 2630145256Sjkoshy KASSERT(td == curthread, 2631145256Sjkoshy ("[pmc,%d] td != curthread", __LINE__)); 2632145256Sjkoshy 2633145256Sjkoshy if (suser(td) || jailed(td->td_ucred)) { 2634145256Sjkoshy error = EPERM; 2635145256Sjkoshy break; 2636145256Sjkoshy } 2637145256Sjkoshy 2638145256Sjkoshy if ((error = copyin(arg, &pma, sizeof(pma))) != 0) 2639145256Sjkoshy break; 2640145256Sjkoshy 2641145256Sjkoshy cpu = pma.pm_cpu; 2642145256Sjkoshy 2643145256Sjkoshy if (cpu < 0 || cpu >= mp_ncpus) { 2644145256Sjkoshy error = EINVAL; 2645145256Sjkoshy break; 2646145256Sjkoshy } 2647145256Sjkoshy 2648145256Sjkoshy if (pmc_cpu_is_disabled(cpu)) { 2649145256Sjkoshy error = ENXIO; 2650145256Sjkoshy break; 2651145256Sjkoshy } 2652145256Sjkoshy 2653145256Sjkoshy request = pma.pm_state; 2654145256Sjkoshy 2655145256Sjkoshy if (request != PMC_STATE_DISABLED && 2656145256Sjkoshy request != PMC_STATE_FREE) { 2657145256Sjkoshy error = EINVAL; 2658145256Sjkoshy break; 2659145256Sjkoshy } 2660145256Sjkoshy 2661145256Sjkoshy ri = pma.pm_pmc; /* pmc id == row index */ 2662145256Sjkoshy if (ri < 0 || ri >= (int) md->pmd_npmc) { 2663145256Sjkoshy error = EINVAL; 2664145256Sjkoshy break; 2665145256Sjkoshy } 2666145256Sjkoshy 2667145256Sjkoshy /* 2668145256Sjkoshy * We can't disable a PMC with a row-index allocated 2669145256Sjkoshy * for process virtual PMCs. 2670145256Sjkoshy */ 2671145256Sjkoshy 2672145256Sjkoshy if (PMC_ROW_DISP_IS_THREAD(ri) && 2673145256Sjkoshy request == PMC_STATE_DISABLED) { 2674145256Sjkoshy error = EBUSY; 2675145256Sjkoshy break; 2676145256Sjkoshy } 2677145256Sjkoshy 2678145256Sjkoshy /* 2679145256Sjkoshy * otherwise, this PMC on this CPU is either free or 2680145256Sjkoshy * in system-wide mode. 2681145256Sjkoshy */ 2682145256Sjkoshy 2683145256Sjkoshy pmc_save_cpu_binding(&pb); 2684145256Sjkoshy pmc_select_cpu(cpu); 2685145256Sjkoshy 2686145256Sjkoshy pc = pmc_pcpu[cpu]; 2687145256Sjkoshy phw = pc->pc_hwpmcs[ri]; 2688145256Sjkoshy 2689145256Sjkoshy /* 2690145256Sjkoshy * XXX do we need some kind of 'forced' disable? 2691145256Sjkoshy */ 2692145256Sjkoshy 2693145256Sjkoshy if (phw->phw_pmc == NULL) { 2694145256Sjkoshy if (request == PMC_STATE_DISABLED && 2695145256Sjkoshy (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) { 2696145256Sjkoshy phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED; 2697145256Sjkoshy PMC_MARK_ROW_STANDALONE(ri); 2698145256Sjkoshy } else if (request == PMC_STATE_FREE && 2699145256Sjkoshy (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) { 2700145256Sjkoshy phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED; 2701145256Sjkoshy PMC_UNMARK_ROW_STANDALONE(ri); 2702145256Sjkoshy } 2703145256Sjkoshy /* other cases are a no-op */ 2704145256Sjkoshy } else 2705145256Sjkoshy error = EBUSY; 2706145256Sjkoshy 2707145256Sjkoshy pmc_restore_cpu_binding(&pb); 2708145256Sjkoshy } 2709145256Sjkoshy break; 2710145256Sjkoshy 2711145256Sjkoshy 2712145256Sjkoshy /* 2713145256Sjkoshy * Allocate a PMC. 2714145256Sjkoshy */ 2715145256Sjkoshy 2716145256Sjkoshy case PMC_OP_PMCALLOCATE: 2717145256Sjkoshy { 2718145256Sjkoshy uint32_t caps; 2719145256Sjkoshy u_int cpu; 2720145256Sjkoshy int n; 2721145256Sjkoshy enum pmc_mode mode; 2722145256Sjkoshy struct pmc *pmc; 2723145774Sjkoshy struct pmc_hw *phw; 2724145256Sjkoshy struct pmc_op_pmcallocate pa; 2725145256Sjkoshy struct pmc_binding pb; 2726145256Sjkoshy 2727145256Sjkoshy if ((error = copyin(arg, &pa, sizeof(pa))) != 0) 2728145256Sjkoshy break; 2729145256Sjkoshy 2730145256Sjkoshy caps = pa.pm_caps; 2731145256Sjkoshy mode = pa.pm_mode; 2732145256Sjkoshy cpu = pa.pm_cpu; 2733145256Sjkoshy 2734145256Sjkoshy if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC && 2735145256Sjkoshy mode != PMC_MODE_TS && mode != PMC_MODE_TC) || 2736145256Sjkoshy (cpu != (u_int) PMC_CPU_ANY && cpu >= (u_int) mp_ncpus)) { 2737145256Sjkoshy error = EINVAL; 2738145256Sjkoshy break; 2739145256Sjkoshy } 2740145256Sjkoshy 2741145256Sjkoshy /* 2742145256Sjkoshy * Virtual PMCs should only ask for a default CPU. 2743145256Sjkoshy * System mode PMCs need to specify a non-default CPU. 2744145256Sjkoshy */ 2745145256Sjkoshy 2746145256Sjkoshy if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) || 2747145256Sjkoshy (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) { 2748145256Sjkoshy error = EINVAL; 2749145256Sjkoshy break; 2750145256Sjkoshy } 2751145256Sjkoshy 2752145256Sjkoshy /* 2753145256Sjkoshy * Check that a disabled CPU is not being asked for. 2754145256Sjkoshy */ 2755145256Sjkoshy 2756145256Sjkoshy if (PMC_IS_SYSTEM_MODE(mode) && pmc_cpu_is_disabled(cpu)) { 2757145256Sjkoshy error = ENXIO; 2758145256Sjkoshy break; 2759145256Sjkoshy } 2760145256Sjkoshy 2761145256Sjkoshy /* 2762145256Sjkoshy * Refuse an allocation for a system-wide PMC if this 2763145256Sjkoshy * process has been jailed, or if this process lacks 2764145256Sjkoshy * super-user credentials and the sysctl tunable 2765145256Sjkoshy * 'security.bsd.unprivileged_syspmcs' is zero. 2766145256Sjkoshy */ 2767145256Sjkoshy 2768145256Sjkoshy if (PMC_IS_SYSTEM_MODE(mode)) { 2769145256Sjkoshy if (jailed(curthread->td_ucred)) 2770145256Sjkoshy error = EPERM; 2771145256Sjkoshy else if (suser(curthread) && 2772145256Sjkoshy (pmc_unprivileged_syspmcs == 0)) 2773145256Sjkoshy error = EPERM; 2774145256Sjkoshy } 2775145256Sjkoshy 2776145256Sjkoshy if (error) 2777145256Sjkoshy break; 2778145256Sjkoshy 2779145256Sjkoshy /* 2780145256Sjkoshy * Look for valid values for 'pm_flags' 2781145256Sjkoshy */ 2782145256Sjkoshy 2783147191Sjkoshy if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW | 2784147191Sjkoshy PMC_F_LOG_PROCEXIT)) != 0) { 2785145256Sjkoshy error = EINVAL; 2786145256Sjkoshy break; 2787145256Sjkoshy } 2788145256Sjkoshy 2789147191Sjkoshy /* process logging options are not allowed for system PMCs */ 2790147191Sjkoshy if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags & 2791147191Sjkoshy (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) { 2792147191Sjkoshy error = EINVAL; 2793147191Sjkoshy break; 2794147191Sjkoshy } 2795147191Sjkoshy 2796145256Sjkoshy /* 2797145256Sjkoshy * All sampling mode PMCs need to be able to interrupt the 2798145256Sjkoshy * CPU. 2799145256Sjkoshy */ 2800147191Sjkoshy if (PMC_IS_SAMPLING_MODE(mode)) 2801145256Sjkoshy caps |= PMC_CAP_INTERRUPT; 2802145256Sjkoshy 2803149374Sjkoshy /* A valid class specifier should have been passed in. */ 2804149374Sjkoshy for (n = 0; n < md->pmd_nclass; n++) 2805149374Sjkoshy if (md->pmd_classes[n].pm_class == pa.pm_class) 2806149374Sjkoshy break; 2807149374Sjkoshy if (n == md->pmd_nclass) { 2808149374Sjkoshy error = EINVAL; 2809149374Sjkoshy break; 2810149374Sjkoshy } 2811149374Sjkoshy 2812149374Sjkoshy /* The requested PMC capabilities should be feasible. */ 2813149374Sjkoshy if ((md->pmd_classes[n].pm_caps & caps) != caps) { 2814149374Sjkoshy error = EOPNOTSUPP; 2815149374Sjkoshy break; 2816149374Sjkoshy } 2817149374Sjkoshy 2818145256Sjkoshy PMCDBG(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d", 2819145256Sjkoshy pa.pm_ev, caps, mode, cpu); 2820145256Sjkoshy 2821145256Sjkoshy pmc = pmc_allocate_pmc_descriptor(); 2822145774Sjkoshy pmc->pm_id = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class, 2823145774Sjkoshy PMC_ID_INVALID); 2824145256Sjkoshy pmc->pm_event = pa.pm_ev; 2825145256Sjkoshy pmc->pm_state = PMC_STATE_FREE; 2826145256Sjkoshy pmc->pm_caps = caps; 2827145256Sjkoshy pmc->pm_flags = pa.pm_flags; 2828145256Sjkoshy 2829145256Sjkoshy /* switch thread to CPU 'cpu' */ 2830145256Sjkoshy pmc_save_cpu_binding(&pb); 2831145256Sjkoshy 2832145256Sjkoshy#define PMC_IS_SHAREABLE_PMC(cpu, n) \ 2833145256Sjkoshy (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \ 2834145256Sjkoshy PMC_PHW_FLAG_IS_SHAREABLE) 2835145256Sjkoshy#define PMC_IS_UNALLOCATED(cpu, n) \ 2836145256Sjkoshy (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL) 2837145256Sjkoshy 2838145256Sjkoshy if (PMC_IS_SYSTEM_MODE(mode)) { 2839145256Sjkoshy pmc_select_cpu(cpu); 2840145256Sjkoshy for (n = 0; n < (int) md->pmd_npmc; n++) 2841145256Sjkoshy if (pmc_can_allocate_row(n, mode) == 0 && 2842145256Sjkoshy pmc_can_allocate_rowindex( 2843145774Sjkoshy curthread->td_proc, n, cpu) == 0 && 2844145256Sjkoshy (PMC_IS_UNALLOCATED(cpu, n) || 2845145256Sjkoshy PMC_IS_SHAREABLE_PMC(cpu, n)) && 2846145256Sjkoshy md->pmd_allocate_pmc(cpu, n, pmc, 2847145256Sjkoshy &pa) == 0) 2848145256Sjkoshy break; 2849145256Sjkoshy } else { 2850145256Sjkoshy /* Process virtual mode */ 2851145256Sjkoshy for (n = 0; n < (int) md->pmd_npmc; n++) { 2852145256Sjkoshy if (pmc_can_allocate_row(n, mode) == 0 && 2853145256Sjkoshy pmc_can_allocate_rowindex( 2854145774Sjkoshy curthread->td_proc, n, 2855145774Sjkoshy PMC_CPU_ANY) == 0 && 2856145256Sjkoshy md->pmd_allocate_pmc(curthread->td_oncpu, 2857145256Sjkoshy n, pmc, &pa) == 0) 2858145256Sjkoshy break; 2859145256Sjkoshy } 2860145256Sjkoshy } 2861145256Sjkoshy 2862145256Sjkoshy#undef PMC_IS_UNALLOCATED 2863145256Sjkoshy#undef PMC_IS_SHAREABLE_PMC 2864145256Sjkoshy 2865145256Sjkoshy pmc_restore_cpu_binding(&pb); 2866145256Sjkoshy 2867145256Sjkoshy if (n == (int) md->pmd_npmc) { 2868145256Sjkoshy pmc_destroy_pmc_descriptor(pmc); 2869145256Sjkoshy FREE(pmc, M_PMC); 2870145256Sjkoshy pmc = NULL; 2871145256Sjkoshy error = EINVAL; 2872145256Sjkoshy break; 2873145256Sjkoshy } 2874145256Sjkoshy 2875145774Sjkoshy /* Fill in the correct value in the ID field */ 2876145774Sjkoshy pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n); 2877145256Sjkoshy 2878145774Sjkoshy PMCDBG(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x", 2879145774Sjkoshy pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id); 2880145774Sjkoshy 2881147191Sjkoshy /* Process mode PMCs with logging enabled need log files */ 2882147191Sjkoshy if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW)) 2883147191Sjkoshy pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 2884147191Sjkoshy 2885147191Sjkoshy /* All system mode sampling PMCs require a log file */ 2886147191Sjkoshy if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode)) 2887147191Sjkoshy pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 2888147191Sjkoshy 2889145256Sjkoshy /* 2890145256Sjkoshy * Configure global pmc's immediately 2891145256Sjkoshy */ 2892145256Sjkoshy 2893145774Sjkoshy if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) { 2894145774Sjkoshy 2895145774Sjkoshy pmc_save_cpu_binding(&pb); 2896145774Sjkoshy pmc_select_cpu(cpu); 2897145774Sjkoshy 2898145774Sjkoshy phw = pmc_pcpu[cpu]->pc_hwpmcs[n]; 2899145774Sjkoshy 2900145774Sjkoshy if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 || 2901145774Sjkoshy (error = md->pmd_config_pmc(cpu, n, pmc)) != 0) { 2902145256Sjkoshy (void) md->pmd_release_pmc(cpu, n, pmc); 2903145256Sjkoshy pmc_destroy_pmc_descriptor(pmc); 2904145256Sjkoshy FREE(pmc, M_PMC); 2905145256Sjkoshy pmc = NULL; 2906145774Sjkoshy pmc_restore_cpu_binding(&pb); 2907145774Sjkoshy error = EPERM; 2908145256Sjkoshy break; 2909145256Sjkoshy } 2910145256Sjkoshy 2911145774Sjkoshy pmc_restore_cpu_binding(&pb); 2912145774Sjkoshy } 2913145256Sjkoshy 2914145256Sjkoshy pmc->pm_state = PMC_STATE_ALLOCATED; 2915145256Sjkoshy 2916145256Sjkoshy /* 2917145256Sjkoshy * mark row disposition 2918145256Sjkoshy */ 2919145256Sjkoshy 2920145256Sjkoshy if (PMC_IS_SYSTEM_MODE(mode)) 2921145256Sjkoshy PMC_MARK_ROW_STANDALONE(n); 2922145256Sjkoshy else 2923145256Sjkoshy PMC_MARK_ROW_THREAD(n); 2924145256Sjkoshy 2925145256Sjkoshy /* 2926145256Sjkoshy * Register this PMC with the current thread as its owner. 2927145256Sjkoshy */ 2928145256Sjkoshy 2929145256Sjkoshy if ((error = 2930145256Sjkoshy pmc_register_owner(curthread->td_proc, pmc)) != 0) { 2931145256Sjkoshy pmc_release_pmc_descriptor(pmc); 2932145256Sjkoshy FREE(pmc, M_PMC); 2933145256Sjkoshy pmc = NULL; 2934145256Sjkoshy break; 2935145256Sjkoshy } 2936145256Sjkoshy 2937145256Sjkoshy /* 2938145256Sjkoshy * Return the allocated index. 2939145256Sjkoshy */ 2940145256Sjkoshy 2941145774Sjkoshy pa.pm_pmcid = pmc->pm_id; 2942145256Sjkoshy 2943145256Sjkoshy error = copyout(&pa, arg, sizeof(pa)); 2944145256Sjkoshy } 2945145256Sjkoshy break; 2946145256Sjkoshy 2947145256Sjkoshy 2948145256Sjkoshy /* 2949145256Sjkoshy * Attach a PMC to a process. 2950145256Sjkoshy */ 2951145256Sjkoshy 2952145256Sjkoshy case PMC_OP_PMCATTACH: 2953145256Sjkoshy { 2954145256Sjkoshy struct pmc *pm; 2955145256Sjkoshy struct proc *p; 2956145256Sjkoshy struct pmc_op_pmcattach a; 2957145256Sjkoshy 2958145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 2959145256Sjkoshy 2960145256Sjkoshy if ((error = copyin(arg, &a, sizeof(a))) != 0) 2961145256Sjkoshy break; 2962145256Sjkoshy 2963145256Sjkoshy if (a.pm_pid < 0) { 2964145256Sjkoshy error = EINVAL; 2965145256Sjkoshy break; 2966145256Sjkoshy } else if (a.pm_pid == 0) 2967145256Sjkoshy a.pm_pid = td->td_proc->p_pid; 2968145256Sjkoshy 2969145256Sjkoshy if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0) 2970145256Sjkoshy break; 2971145256Sjkoshy 2972145774Sjkoshy if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) { 2973145256Sjkoshy error = EINVAL; 2974145256Sjkoshy break; 2975145256Sjkoshy } 2976145256Sjkoshy 2977145256Sjkoshy /* PMCs may be (re)attached only when allocated or stopped */ 2978145256Sjkoshy if (pm->pm_state == PMC_STATE_RUNNING) { 2979145256Sjkoshy error = EBUSY; 2980145256Sjkoshy break; 2981145256Sjkoshy } else if (pm->pm_state != PMC_STATE_ALLOCATED && 2982145256Sjkoshy pm->pm_state != PMC_STATE_STOPPED) { 2983145256Sjkoshy error = EINVAL; 2984145256Sjkoshy break; 2985145256Sjkoshy } 2986145256Sjkoshy 2987145256Sjkoshy /* lookup pid */ 2988145256Sjkoshy if ((p = pfind(a.pm_pid)) == NULL) { 2989145256Sjkoshy error = ESRCH; 2990145256Sjkoshy break; 2991145256Sjkoshy } 2992145256Sjkoshy 2993145256Sjkoshy /* 2994145256Sjkoshy * Ignore processes that are working on exiting. 2995145256Sjkoshy */ 2996145256Sjkoshy if (p->p_flag & P_WEXIT) { 2997145256Sjkoshy error = ESRCH; 2998145256Sjkoshy PROC_UNLOCK(p); /* pfind() returns a locked process */ 2999145256Sjkoshy break; 3000145256Sjkoshy } 3001145256Sjkoshy 3002145256Sjkoshy /* 3003145256Sjkoshy * we are allowed to attach a PMC to a process if 3004145256Sjkoshy * we can debug it. 3005145256Sjkoshy */ 3006145256Sjkoshy error = p_candebug(curthread, p); 3007145256Sjkoshy 3008145256Sjkoshy PROC_UNLOCK(p); 3009145256Sjkoshy 3010145256Sjkoshy if (error == 0) 3011145256Sjkoshy error = pmc_attach_process(p, pm); 3012145256Sjkoshy } 3013145256Sjkoshy break; 3014145256Sjkoshy 3015145256Sjkoshy 3016145256Sjkoshy /* 3017145256Sjkoshy * Detach an attached PMC from a process. 3018145256Sjkoshy */ 3019145256Sjkoshy 3020145256Sjkoshy case PMC_OP_PMCDETACH: 3021145256Sjkoshy { 3022145256Sjkoshy struct pmc *pm; 3023145256Sjkoshy struct proc *p; 3024145256Sjkoshy struct pmc_op_pmcattach a; 3025145256Sjkoshy 3026145256Sjkoshy if ((error = copyin(arg, &a, sizeof(a))) != 0) 3027145256Sjkoshy break; 3028145256Sjkoshy 3029145256Sjkoshy if (a.pm_pid < 0) { 3030145256Sjkoshy error = EINVAL; 3031145256Sjkoshy break; 3032145256Sjkoshy } else if (a.pm_pid == 0) 3033145256Sjkoshy a.pm_pid = td->td_proc->p_pid; 3034145256Sjkoshy 3035145256Sjkoshy if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0) 3036145256Sjkoshy break; 3037145256Sjkoshy 3038145256Sjkoshy if ((p = pfind(a.pm_pid)) == NULL) { 3039145256Sjkoshy error = ESRCH; 3040145256Sjkoshy break; 3041145256Sjkoshy } 3042145256Sjkoshy 3043145256Sjkoshy /* 3044145256Sjkoshy * Treat processes that are in the process of exiting 3045145256Sjkoshy * as if they were not present. 3046145256Sjkoshy */ 3047145256Sjkoshy 3048145256Sjkoshy if (p->p_flag & P_WEXIT) 3049145256Sjkoshy error = ESRCH; 3050145256Sjkoshy 3051145256Sjkoshy PROC_UNLOCK(p); /* pfind() returns a locked process */ 3052145256Sjkoshy 3053145256Sjkoshy if (error == 0) 3054145256Sjkoshy error = pmc_detach_process(p, pm); 3055145256Sjkoshy } 3056145256Sjkoshy break; 3057145256Sjkoshy 3058145256Sjkoshy 3059145256Sjkoshy /* 3060147191Sjkoshy * Retrieve the MSR number associated with the counter 3061147191Sjkoshy * 'pmc_id'. This allows processes to directly use RDPMC 3062147191Sjkoshy * instructions to read their PMCs, without the overhead of a 3063147191Sjkoshy * system call. 3064147191Sjkoshy */ 3065147191Sjkoshy 3066147191Sjkoshy case PMC_OP_PMCGETMSR: 3067147191Sjkoshy { 3068147191Sjkoshy int ri; 3069147191Sjkoshy struct pmc *pm; 3070147191Sjkoshy struct pmc_target *pt; 3071147191Sjkoshy struct pmc_op_getmsr gm; 3072147191Sjkoshy 3073147191Sjkoshy PMC_DOWNGRADE_SX(); 3074147191Sjkoshy 3075147191Sjkoshy /* CPU has no 'GETMSR' support */ 3076147191Sjkoshy if (md->pmd_get_msr == NULL) { 3077147191Sjkoshy error = ENOSYS; 3078147191Sjkoshy break; 3079147191Sjkoshy } 3080147191Sjkoshy 3081147191Sjkoshy if ((error = copyin(arg, &gm, sizeof(gm))) != 0) 3082147191Sjkoshy break; 3083147191Sjkoshy 3084147191Sjkoshy if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0) 3085147191Sjkoshy break; 3086147191Sjkoshy 3087147191Sjkoshy /* 3088147191Sjkoshy * The allocated PMC has to be a process virtual PMC, 3089147191Sjkoshy * i.e., of type MODE_T[CS]. Global PMCs can only be 3090147191Sjkoshy * read using the PMCREAD operation since they may be 3091147191Sjkoshy * allocated on a different CPU than the one we could 3092147191Sjkoshy * be running on at the time of the RDPMC instruction. 3093147191Sjkoshy * 3094147191Sjkoshy * The GETMSR operation is not allowed for PMCs that 3095147191Sjkoshy * are inherited across processes. 3096147191Sjkoshy */ 3097147191Sjkoshy 3098147191Sjkoshy if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) || 3099147191Sjkoshy (pm->pm_flags & PMC_F_DESCENDANTS)) { 3100147191Sjkoshy error = EINVAL; 3101147191Sjkoshy break; 3102147191Sjkoshy } 3103147191Sjkoshy 3104147191Sjkoshy /* 3105147191Sjkoshy * It only makes sense to use a RDPMC (or its 3106147191Sjkoshy * equivalent instruction on non-x86 architectures) on 3107147191Sjkoshy * a process that has allocated and attached a PMC to 3108147191Sjkoshy * itself. Conversely the PMC is only allowed to have 3109147191Sjkoshy * one process attached to it -- its owner. 3110147191Sjkoshy */ 3111147191Sjkoshy 3112147191Sjkoshy if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL || 3113147191Sjkoshy LIST_NEXT(pt, pt_next) != NULL || 3114147191Sjkoshy pt->pt_process->pp_proc != pm->pm_owner->po_owner) { 3115147191Sjkoshy error = EINVAL; 3116147191Sjkoshy break; 3117147191Sjkoshy } 3118147191Sjkoshy 3119147191Sjkoshy ri = PMC_TO_ROWINDEX(pm); 3120147191Sjkoshy 3121147191Sjkoshy if ((error = (*md->pmd_get_msr)(ri, &gm.pm_msr)) < 0) 3122147191Sjkoshy break; 3123147191Sjkoshy 3124147191Sjkoshy if ((error = copyout(&gm, arg, sizeof(gm))) < 0) 3125147191Sjkoshy break; 3126147191Sjkoshy 3127147191Sjkoshy /* 3128147191Sjkoshy * Mark our process as using MSRs. Update machine 3129147191Sjkoshy * state using a forced context switch. 3130147191Sjkoshy */ 3131147191Sjkoshy 3132147191Sjkoshy pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS; 3133147191Sjkoshy pmc_force_context_switch(); 3134147191Sjkoshy 3135147191Sjkoshy } 3136147191Sjkoshy break; 3137147191Sjkoshy 3138147191Sjkoshy /* 3139145256Sjkoshy * Release an allocated PMC 3140145256Sjkoshy */ 3141145256Sjkoshy 3142145256Sjkoshy case PMC_OP_PMCRELEASE: 3143145256Sjkoshy { 3144145256Sjkoshy pmc_id_t pmcid; 3145145256Sjkoshy struct pmc *pm; 3146145256Sjkoshy struct pmc_owner *po; 3147145256Sjkoshy struct pmc_op_simple sp; 3148145256Sjkoshy 3149145256Sjkoshy /* 3150145256Sjkoshy * Find PMC pointer for the named PMC. 3151145256Sjkoshy * 3152145256Sjkoshy * Use pmc_release_pmc_descriptor() to switch off the 3153145256Sjkoshy * PMC, remove all its target threads, and remove the 3154145256Sjkoshy * PMC from its owner's list. 3155145256Sjkoshy * 3156145256Sjkoshy * Remove the owner record if this is the last PMC 3157145256Sjkoshy * owned. 3158145256Sjkoshy * 3159145256Sjkoshy * Free up space. 3160145256Sjkoshy */ 3161145256Sjkoshy 3162145256Sjkoshy if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 3163145256Sjkoshy break; 3164145256Sjkoshy 3165145256Sjkoshy pmcid = sp.pm_pmcid; 3166145256Sjkoshy 3167145256Sjkoshy if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 3168145256Sjkoshy break; 3169145256Sjkoshy 3170145256Sjkoshy po = pm->pm_owner; 3171145256Sjkoshy pmc_release_pmc_descriptor(pm); 3172145256Sjkoshy pmc_maybe_remove_owner(po); 3173145256Sjkoshy 3174145256Sjkoshy FREE(pm, M_PMC); 3175145256Sjkoshy } 3176145256Sjkoshy break; 3177145256Sjkoshy 3178145256Sjkoshy 3179145256Sjkoshy /* 3180145256Sjkoshy * Read and/or write a PMC. 3181145256Sjkoshy */ 3182145256Sjkoshy 3183145256Sjkoshy case PMC_OP_PMCRW: 3184145256Sjkoshy { 3185145256Sjkoshy uint32_t cpu, ri; 3186145256Sjkoshy struct pmc *pm; 3187145256Sjkoshy struct pmc_op_pmcrw *pprw; 3188145256Sjkoshy struct pmc_op_pmcrw prw; 3189145256Sjkoshy struct pmc_binding pb; 3190145256Sjkoshy pmc_value_t oldvalue; 3191145256Sjkoshy 3192145256Sjkoshy PMC_DOWNGRADE_SX(); 3193145256Sjkoshy 3194145256Sjkoshy if ((error = copyin(arg, &prw, sizeof(prw))) != 0) 3195145256Sjkoshy break; 3196145256Sjkoshy 3197145301Simp ri = 0; 3198145256Sjkoshy PMCDBG(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid, 3199145256Sjkoshy prw.pm_flags); 3200145256Sjkoshy 3201145256Sjkoshy /* must have at least one flag set */ 3202145256Sjkoshy if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) { 3203145256Sjkoshy error = EINVAL; 3204145256Sjkoshy break; 3205145256Sjkoshy } 3206145256Sjkoshy 3207145256Sjkoshy /* locate pmc descriptor */ 3208145256Sjkoshy if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0) 3209145256Sjkoshy break; 3210145256Sjkoshy 3211145256Sjkoshy /* Can't read a PMC that hasn't been started. */ 3212145256Sjkoshy if (pm->pm_state != PMC_STATE_ALLOCATED && 3213145256Sjkoshy pm->pm_state != PMC_STATE_STOPPED && 3214145256Sjkoshy pm->pm_state != PMC_STATE_RUNNING) { 3215145256Sjkoshy error = EINVAL; 3216145256Sjkoshy break; 3217145256Sjkoshy } 3218145256Sjkoshy 3219145256Sjkoshy /* writing a new value is allowed only for 'STOPPED' pmcs */ 3220145256Sjkoshy if (pm->pm_state == PMC_STATE_RUNNING && 3221145256Sjkoshy (prw.pm_flags & PMC_F_NEWVALUE)) { 3222145256Sjkoshy error = EBUSY; 3223145256Sjkoshy break; 3224145256Sjkoshy } 3225145256Sjkoshy 3226145774Sjkoshy if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { 3227145256Sjkoshy 3228145774Sjkoshy /* 3229145774Sjkoshy * If this PMC is attached to its owner (i.e., 3230145774Sjkoshy * the process requesting this operation) and 3231145774Sjkoshy * is running, then attempt to get an 3232145774Sjkoshy * upto-date reading from hardware for a READ. 3233145774Sjkoshy * Writes are only allowed when the PMC is 3234145774Sjkoshy * stopped, so only update the saved value 3235145774Sjkoshy * field. 3236145774Sjkoshy * 3237145774Sjkoshy * If the PMC is not running, or is not 3238145774Sjkoshy * attached to its owner, read/write to the 3239145774Sjkoshy * savedvalue field. 3240145774Sjkoshy */ 3241145774Sjkoshy 3242145774Sjkoshy ri = PMC_TO_ROWINDEX(pm); 3243145774Sjkoshy 3244145256Sjkoshy mtx_pool_lock_spin(pmc_mtxpool, pm); 3245145774Sjkoshy cpu = curthread->td_oncpu; 3246145774Sjkoshy 3247145774Sjkoshy if (prw.pm_flags & PMC_F_OLDVALUE) { 3248145774Sjkoshy if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) && 3249145774Sjkoshy (pm->pm_state == PMC_STATE_RUNNING)) 3250145774Sjkoshy error = (*md->pmd_read_pmc)(cpu, ri, 3251145774Sjkoshy &oldvalue); 3252145774Sjkoshy else 3253145774Sjkoshy oldvalue = pm->pm_gv.pm_savedvalue; 3254145774Sjkoshy } 3255145256Sjkoshy if (prw.pm_flags & PMC_F_NEWVALUE) 3256145256Sjkoshy pm->pm_gv.pm_savedvalue = prw.pm_value; 3257145774Sjkoshy 3258145256Sjkoshy mtx_pool_unlock_spin(pmc_mtxpool, pm); 3259145256Sjkoshy 3260145256Sjkoshy } else { /* System mode PMCs */ 3261145774Sjkoshy cpu = PMC_TO_CPU(pm); 3262145774Sjkoshy ri = PMC_TO_ROWINDEX(pm); 3263145256Sjkoshy 3264145256Sjkoshy if (pmc_cpu_is_disabled(cpu)) { 3265145256Sjkoshy error = ENXIO; 3266145256Sjkoshy break; 3267145256Sjkoshy } 3268145256Sjkoshy 3269145256Sjkoshy /* move this thread to CPU 'cpu' */ 3270145256Sjkoshy pmc_save_cpu_binding(&pb); 3271145256Sjkoshy pmc_select_cpu(cpu); 3272145256Sjkoshy 3273145774Sjkoshy critical_enter(); 3274145256Sjkoshy /* save old value */ 3275145256Sjkoshy if (prw.pm_flags & PMC_F_OLDVALUE) 3276145256Sjkoshy if ((error = (*md->pmd_read_pmc)(cpu, ri, 3277145256Sjkoshy &oldvalue))) 3278145256Sjkoshy goto error; 3279145256Sjkoshy /* write out new value */ 3280145256Sjkoshy if (prw.pm_flags & PMC_F_NEWVALUE) 3281145256Sjkoshy error = (*md->pmd_write_pmc)(cpu, ri, 3282145256Sjkoshy prw.pm_value); 3283145256Sjkoshy error: 3284145774Sjkoshy critical_exit(); 3285145256Sjkoshy pmc_restore_cpu_binding(&pb); 3286145256Sjkoshy if (error) 3287145256Sjkoshy break; 3288145256Sjkoshy } 3289145256Sjkoshy 3290145256Sjkoshy pprw = (struct pmc_op_pmcrw *) arg; 3291145256Sjkoshy 3292153110Sru#ifdef DEBUG 3293145256Sjkoshy if (prw.pm_flags & PMC_F_NEWVALUE) 3294145256Sjkoshy PMCDBG(PMC,OPS,2, "rw id=%d new %jx -> old %jx", 3295145256Sjkoshy ri, prw.pm_value, oldvalue); 3296145256Sjkoshy else 3297145256Sjkoshy PMCDBG(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue); 3298145256Sjkoshy#endif 3299145256Sjkoshy 3300145256Sjkoshy /* return old value if requested */ 3301145256Sjkoshy if (prw.pm_flags & PMC_F_OLDVALUE) 3302145256Sjkoshy if ((error = copyout(&oldvalue, &pprw->pm_value, 3303145256Sjkoshy sizeof(prw.pm_value)))) 3304145256Sjkoshy break; 3305145256Sjkoshy 3306145256Sjkoshy } 3307145256Sjkoshy break; 3308145256Sjkoshy 3309145256Sjkoshy 3310145256Sjkoshy /* 3311145256Sjkoshy * Set the sampling rate for a sampling mode PMC and the 3312145256Sjkoshy * initial count for a counting mode PMC. 3313145256Sjkoshy */ 3314145256Sjkoshy 3315145256Sjkoshy case PMC_OP_PMCSETCOUNT: 3316145256Sjkoshy { 3317145256Sjkoshy struct pmc *pm; 3318145256Sjkoshy struct pmc_op_pmcsetcount sc; 3319145256Sjkoshy 3320145256Sjkoshy PMC_DOWNGRADE_SX(); 3321145256Sjkoshy 3322145256Sjkoshy if ((error = copyin(arg, &sc, sizeof(sc))) != 0) 3323145256Sjkoshy break; 3324145256Sjkoshy 3325145256Sjkoshy if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0) 3326145256Sjkoshy break; 3327145256Sjkoshy 3328145256Sjkoshy if (pm->pm_state == PMC_STATE_RUNNING) { 3329145256Sjkoshy error = EBUSY; 3330145256Sjkoshy break; 3331145256Sjkoshy } 3332145256Sjkoshy 3333145774Sjkoshy if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 3334145256Sjkoshy pm->pm_sc.pm_reloadcount = sc.pm_count; 3335145256Sjkoshy else 3336145256Sjkoshy pm->pm_sc.pm_initial = sc.pm_count; 3337145256Sjkoshy } 3338145256Sjkoshy break; 3339145256Sjkoshy 3340145256Sjkoshy 3341145256Sjkoshy /* 3342145256Sjkoshy * Start a PMC. 3343145256Sjkoshy */ 3344145256Sjkoshy 3345145256Sjkoshy case PMC_OP_PMCSTART: 3346145256Sjkoshy { 3347145256Sjkoshy pmc_id_t pmcid; 3348145256Sjkoshy struct pmc *pm; 3349145256Sjkoshy struct pmc_op_simple sp; 3350145256Sjkoshy 3351145256Sjkoshy sx_assert(&pmc_sx, SX_XLOCKED); 3352145256Sjkoshy 3353145256Sjkoshy if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 3354145256Sjkoshy break; 3355145256Sjkoshy 3356145256Sjkoshy pmcid = sp.pm_pmcid; 3357145256Sjkoshy 3358145256Sjkoshy if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 3359145256Sjkoshy break; 3360145256Sjkoshy 3361145774Sjkoshy KASSERT(pmcid == pm->pm_id, 3362145774Sjkoshy ("[pmc,%d] pmcid %x != id %x", __LINE__, 3363145774Sjkoshy pm->pm_id, pmcid)); 3364145256Sjkoshy 3365145256Sjkoshy if (pm->pm_state == PMC_STATE_RUNNING) /* already running */ 3366145256Sjkoshy break; 3367145256Sjkoshy else if (pm->pm_state != PMC_STATE_STOPPED && 3368145256Sjkoshy pm->pm_state != PMC_STATE_ALLOCATED) { 3369145256Sjkoshy error = EINVAL; 3370145256Sjkoshy break; 3371145256Sjkoshy } 3372145256Sjkoshy 3373145256Sjkoshy error = pmc_start(pm); 3374145256Sjkoshy } 3375145256Sjkoshy break; 3376145256Sjkoshy 3377145256Sjkoshy 3378145256Sjkoshy /* 3379145256Sjkoshy * Stop a PMC. 3380145256Sjkoshy */ 3381145256Sjkoshy 3382145256Sjkoshy case PMC_OP_PMCSTOP: 3383145256Sjkoshy { 3384145256Sjkoshy pmc_id_t pmcid; 3385145256Sjkoshy struct pmc *pm; 3386145256Sjkoshy struct pmc_op_simple sp; 3387145256Sjkoshy 3388145256Sjkoshy PMC_DOWNGRADE_SX(); 3389145256Sjkoshy 3390145256Sjkoshy if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 3391145256Sjkoshy break; 3392145256Sjkoshy 3393145256Sjkoshy pmcid = sp.pm_pmcid; 3394145256Sjkoshy 3395145256Sjkoshy /* 3396145256Sjkoshy * Mark the PMC as inactive and invoke the MD stop 3397145256Sjkoshy * routines if needed. 3398145256Sjkoshy */ 3399145256Sjkoshy 3400145256Sjkoshy if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 3401145256Sjkoshy break; 3402145256Sjkoshy 3403145774Sjkoshy KASSERT(pmcid == pm->pm_id, 3404145774Sjkoshy ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, 3405145774Sjkoshy pm->pm_id, pmcid)); 3406145256Sjkoshy 3407145256Sjkoshy if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */ 3408145256Sjkoshy break; 3409145256Sjkoshy else if (pm->pm_state != PMC_STATE_RUNNING) { 3410145256Sjkoshy error = EINVAL; 3411145256Sjkoshy break; 3412145256Sjkoshy } 3413145256Sjkoshy 3414145256Sjkoshy error = pmc_stop(pm); 3415145256Sjkoshy } 3416145256Sjkoshy break; 3417145256Sjkoshy 3418145256Sjkoshy 3419145256Sjkoshy /* 3420147867Sjkoshy * Write a user supplied value to the log file. 3421145256Sjkoshy */ 3422145256Sjkoshy 3423145256Sjkoshy case PMC_OP_WRITELOG: 3424145256Sjkoshy { 3425147191Sjkoshy struct pmc_op_writelog wl; 3426147191Sjkoshy struct pmc_owner *po; 3427145256Sjkoshy 3428145256Sjkoshy PMC_DOWNGRADE_SX(); 3429145256Sjkoshy 3430147191Sjkoshy if ((error = copyin(arg, &wl, sizeof(wl))) != 0) 3431145256Sjkoshy break; 3432145256Sjkoshy 3433147191Sjkoshy if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 3434145256Sjkoshy error = EINVAL; 3435145256Sjkoshy break; 3436145256Sjkoshy } 3437145256Sjkoshy 3438147191Sjkoshy if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 3439145774Sjkoshy error = EINVAL; 3440145774Sjkoshy break; 3441145774Sjkoshy } 3442145774Sjkoshy 3443147191Sjkoshy error = pmclog_process_userlog(po, &wl); 3444145256Sjkoshy } 3445145256Sjkoshy break; 3446145256Sjkoshy 3447147191Sjkoshy 3448145256Sjkoshy default: 3449145256Sjkoshy error = EINVAL; 3450145256Sjkoshy break; 3451145256Sjkoshy } 3452145256Sjkoshy 3453145256Sjkoshy if (is_sx_downgraded) 3454145256Sjkoshy sx_sunlock(&pmc_sx); 3455145256Sjkoshy else 3456145256Sjkoshy sx_xunlock(&pmc_sx); 3457145256Sjkoshy 3458145256Sjkoshy if (error) 3459145256Sjkoshy atomic_add_int(&pmc_stats.pm_syscall_errors, 1); 3460145256Sjkoshy 3461147191Sjkoshy PICKUP_GIANT(); 3462147191Sjkoshy 3463145256Sjkoshy return error; 3464145256Sjkoshy} 3465145256Sjkoshy 3466145256Sjkoshy/* 3467145256Sjkoshy * Helper functions 3468145256Sjkoshy */ 3469145256Sjkoshy 3470147191Sjkoshy 3471145256Sjkoshy/* 3472147191Sjkoshy * Interrupt processing. 3473147191Sjkoshy * 3474147191Sjkoshy * Find a free slot in the per-cpu array of PC samples and write the 3475147191Sjkoshy * current (PMC,PID,PC) triple to it. If an event was successfully 3476147191Sjkoshy * added, a bit is set in mask 'pmc_cpumask' denoting that the 3477147191Sjkoshy * DO_SAMPLES hook needs to be invoked from the clock handler. 3478147191Sjkoshy * 3479147191Sjkoshy * This function is meant to be called from an NMI handler. It cannot 3480147191Sjkoshy * use any of the locking primitives supplied by the OS. 3481145256Sjkoshy */ 3482145256Sjkoshy 3483147191Sjkoshyint 3484147708Sjkoshypmc_process_interrupt(int cpu, struct pmc *pm, uintfptr_t pc, int usermode) 3485145256Sjkoshy{ 3486147191Sjkoshy int error, ri; 3487147191Sjkoshy struct thread *td; 3488147191Sjkoshy struct pmc_sample *ps; 3489147191Sjkoshy struct pmc_samplebuffer *psb; 3490145256Sjkoshy 3491147191Sjkoshy error = 0; 3492147191Sjkoshy ri = PMC_TO_ROWINDEX(pm); 3493145256Sjkoshy 3494147191Sjkoshy psb = pmc_pcpu[cpu]->pc_sb; 3495145256Sjkoshy 3496147191Sjkoshy ps = psb->ps_write; 3497147191Sjkoshy if (ps->ps_pc) { /* in use, reader hasn't caught up */ 3498147867Sjkoshy pm->pm_stalled = 1; 3499147191Sjkoshy atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1); 3500147191Sjkoshy PMCDBG(SAM,INT,1,"(spc) cpu=%d pm=%p pc=%jx um=%d wr=%d rd=%d", 3501147708Sjkoshy cpu, pm, (uint64_t) pc, usermode, 3502147191Sjkoshy (int) (psb->ps_write - psb->ps_samples), 3503147191Sjkoshy (int) (psb->ps_read - psb->ps_samples)); 3504147191Sjkoshy error = ENOMEM; 3505147191Sjkoshy goto done; 3506147191Sjkoshy } 3507145256Sjkoshy 3508147191Sjkoshy /* fill in entry */ 3509147191Sjkoshy PMCDBG(SAM,INT,1,"cpu=%d pm=%p pc=%jx um=%d wr=%d rd=%d", cpu, pm, 3510147708Sjkoshy (uint64_t) pc, usermode, 3511147191Sjkoshy (int) (psb->ps_write - psb->ps_samples), 3512147191Sjkoshy (int) (psb->ps_read - psb->ps_samples)); 3513145256Sjkoshy 3514147191Sjkoshy atomic_add_rel_32(&pm->pm_runcount, 1); /* hold onto PMC */ 3515147191Sjkoshy ps->ps_pmc = pm; 3516147191Sjkoshy if ((td = curthread) && td->td_proc) 3517147191Sjkoshy ps->ps_pid = td->td_proc->p_pid; 3518147191Sjkoshy else 3519147191Sjkoshy ps->ps_pid = -1; 3520147191Sjkoshy ps->ps_usermode = usermode; 3521147191Sjkoshy ps->ps_pc = pc; /* mark entry as in use */ 3522145256Sjkoshy 3523147191Sjkoshy /* increment write pointer, modulo ring buffer size */ 3524147191Sjkoshy ps++; 3525147191Sjkoshy if (ps == psb->ps_fence) 3526147191Sjkoshy psb->ps_write = psb->ps_samples; 3527147191Sjkoshy else 3528147191Sjkoshy psb->ps_write = ps; 3529145256Sjkoshy 3530147191Sjkoshy done: 3531147191Sjkoshy /* mark CPU as needing processing */ 3532147191Sjkoshy atomic_set_rel_int(&pmc_cpumask, (1 << cpu)); 3533147191Sjkoshy 3534147191Sjkoshy return error; 3535145256Sjkoshy} 3536145256Sjkoshy 3537147191Sjkoshy 3538145256Sjkoshy/* 3539147191Sjkoshy * Process saved PC samples. 3540145256Sjkoshy */ 3541145256Sjkoshy 3542145256Sjkoshystatic void 3543147191Sjkoshypmc_process_samples(int cpu) 3544145256Sjkoshy{ 3545147191Sjkoshy int n, ri; 3546147191Sjkoshy struct pmc *pm; 3547147191Sjkoshy struct thread *td; 3548147191Sjkoshy struct pmc_owner *po; 3549147191Sjkoshy struct pmc_sample *ps; 3550147191Sjkoshy struct pmc_samplebuffer *psb; 3551145256Sjkoshy 3552147191Sjkoshy KASSERT(PCPU_GET(cpuid) == cpu, 3553147191Sjkoshy ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__, 3554147191Sjkoshy PCPU_GET(cpuid), cpu)); 3555145256Sjkoshy 3556147191Sjkoshy psb = pmc_pcpu[cpu]->pc_sb; 3557147191Sjkoshy 3558147191Sjkoshy for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */ 3559147191Sjkoshy 3560147191Sjkoshy ps = psb->ps_read; 3561147191Sjkoshy if (ps->ps_pc == (uintfptr_t) 0) /* no data */ 3562147191Sjkoshy break; 3563147191Sjkoshy 3564147191Sjkoshy pm = ps->ps_pmc; 3565147191Sjkoshy po = pm->pm_owner; 3566147191Sjkoshy 3567147191Sjkoshy KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 3568147191Sjkoshy ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__, 3569147191Sjkoshy pm, PMC_TO_MODE(pm))); 3570147191Sjkoshy 3571147191Sjkoshy /* Ignore PMCs that have been switched off */ 3572147191Sjkoshy if (pm->pm_state != PMC_STATE_RUNNING) 3573147191Sjkoshy goto entrydone; 3574147191Sjkoshy 3575147191Sjkoshy PMCDBG(SAM,OPS,1,"cpu=%d pm=%p pc=%jx um=%d wr=%d rd=%d", cpu, 3576147708Sjkoshy pm, (uint64_t) ps->ps_pc, ps->ps_usermode, 3577147191Sjkoshy (int) (psb->ps_write - psb->ps_samples), 3578147191Sjkoshy (int) (psb->ps_read - psb->ps_samples)); 3579147191Sjkoshy 3580147191Sjkoshy /* 3581147191Sjkoshy * If this is a process-mode PMC that is attached to 3582147191Sjkoshy * its owner, and if the PC is in user mode, update 3583147191Sjkoshy * profiling statistics like timer-based profiling 3584147191Sjkoshy * would have done. 3585147191Sjkoshy */ 3586147191Sjkoshy if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) { 3587147191Sjkoshy if (ps->ps_usermode) { 3588147191Sjkoshy td = FIRST_THREAD_IN_PROC(po->po_owner); 3589147191Sjkoshy addupc_intr(td, ps->ps_pc, 1); 3590147191Sjkoshy } 3591147191Sjkoshy goto entrydone; 3592147191Sjkoshy } 3593147191Sjkoshy 3594147191Sjkoshy /* 3595147191Sjkoshy * Otherwise, this is either a sampling mode PMC that 3596147191Sjkoshy * is attached to a different process than its owner, 3597147191Sjkoshy * or a system-wide sampling PMC. Dispatch a log 3598147191Sjkoshy * entry to the PMC's owner process. 3599147191Sjkoshy */ 3600147191Sjkoshy 3601147191Sjkoshy pmclog_process_pcsample(pm, ps); 3602147191Sjkoshy 3603147191Sjkoshy entrydone: 3604147191Sjkoshy ps->ps_pc = (uintfptr_t) 0; /* mark entry as free */ 3605147191Sjkoshy atomic_subtract_rel_32(&pm->pm_runcount, 1); 3606147191Sjkoshy 3607147191Sjkoshy /* increment read pointer, modulo sample size */ 3608147191Sjkoshy if (++ps == psb->ps_fence) 3609147191Sjkoshy psb->ps_read = psb->ps_samples; 3610147191Sjkoshy else 3611147191Sjkoshy psb->ps_read = ps; 3612147191Sjkoshy } 3613147191Sjkoshy 3614147191Sjkoshy atomic_add_int(&pmc_stats.pm_log_sweeps, 1); 3615147191Sjkoshy 3616147191Sjkoshy /* Do not re-enable stalled PMCs if we failed to process any samples */ 3617147191Sjkoshy if (n == 0) 3618147191Sjkoshy return; 3619147191Sjkoshy 3620147191Sjkoshy /* 3621147191Sjkoshy * Restart any stalled sampling PMCs on this CPU. 3622147191Sjkoshy * 3623147867Sjkoshy * If the NMI handler sets the pm_stalled field of a PMC after 3624147867Sjkoshy * the check below, we'll end up processing the stalled PMC at 3625147867Sjkoshy * the next hardclock tick. 3626147191Sjkoshy */ 3627147191Sjkoshy for (n = 0; n < md->pmd_npmc; n++) { 3628147191Sjkoshy (void) (*md->pmd_get_config)(cpu,n,&pm); 3629147191Sjkoshy if (pm == NULL || /* !cfg'ed */ 3630147191Sjkoshy pm->pm_state != PMC_STATE_RUNNING || /* !active */ 3631147191Sjkoshy !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */ 3632147867Sjkoshy pm->pm_stalled == 0) /* !stalled */ 3633147191Sjkoshy continue; 3634147191Sjkoshy 3635147867Sjkoshy pm->pm_stalled = 0; 3636147191Sjkoshy ri = PMC_TO_ROWINDEX(pm); 3637147191Sjkoshy (*md->pmd_start_pmc)(cpu, ri); 3638147191Sjkoshy } 3639145256Sjkoshy} 3640145256Sjkoshy 3641145256Sjkoshy/* 3642145256Sjkoshy * Event handlers. 3643145256Sjkoshy */ 3644145256Sjkoshy 3645145256Sjkoshy/* 3646145256Sjkoshy * Handle a process exit. 3647145256Sjkoshy * 3648147191Sjkoshy * Remove this process from all hash tables. If this process 3649147191Sjkoshy * owned any PMCs, turn off those PMCs and deallocate them, 3650147191Sjkoshy * removing any associations with target processes. 3651147191Sjkoshy * 3652147191Sjkoshy * This function will be called by the last 'thread' of a 3653147191Sjkoshy * process. 3654147191Sjkoshy * 3655145256Sjkoshy * XXX This eventhandler gets called early in the exit process. 3656145256Sjkoshy * Consider using a 'hook' invocation from thread_exit() or equivalent 3657145256Sjkoshy * spot. Another negative is that kse_exit doesn't seem to call 3658145256Sjkoshy * exit1() [??]. 3659147191Sjkoshy * 3660145256Sjkoshy */ 3661145256Sjkoshy 3662145256Sjkoshystatic void 3663145256Sjkoshypmc_process_exit(void *arg __unused, struct proc *p) 3664145256Sjkoshy{ 3665145256Sjkoshy int is_using_hwpmcs; 3666147191Sjkoshy int cpu; 3667147191Sjkoshy unsigned int ri; 3668147191Sjkoshy struct pmc *pm; 3669147191Sjkoshy struct pmc_process *pp; 3670147191Sjkoshy struct pmc_owner *po; 3671147191Sjkoshy pmc_value_t newvalue, tmp; 3672145256Sjkoshy 3673145256Sjkoshy PROC_LOCK(p); 3674145256Sjkoshy is_using_hwpmcs = p->p_flag & P_HWPMC; 3675145256Sjkoshy PROC_UNLOCK(p); 3676145256Sjkoshy 3677147191Sjkoshy /* 3678147191Sjkoshy * Log a sysexit event to all SS PMC owners. 3679147191Sjkoshy */ 3680147191Sjkoshy LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 3681147191Sjkoshy if (po->po_flags & PMC_PO_OWNS_LOGFILE) 3682147191Sjkoshy pmclog_process_sysexit(po, p->p_pid); 3683145256Sjkoshy 3684147191Sjkoshy if (!is_using_hwpmcs) 3685147191Sjkoshy return; 3686147191Sjkoshy 3687147191Sjkoshy PMC_GET_SX_XLOCK(); 3688147191Sjkoshy PMCDBG(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid, 3689147191Sjkoshy p->p_comm); 3690147191Sjkoshy 3691147191Sjkoshy /* 3692147191Sjkoshy * Since this code is invoked by the last thread in an exiting 3693147191Sjkoshy * process, we would have context switched IN at some prior 3694147191Sjkoshy * point. However, with PREEMPTION, kernel mode context 3695147191Sjkoshy * switches may happen any time, so we want to disable a 3696147191Sjkoshy * context switch OUT till we get any PMCs targetting this 3697147191Sjkoshy * process off the hardware. 3698147191Sjkoshy * 3699147191Sjkoshy * We also need to atomically remove this process' 3700147191Sjkoshy * entry from our target process hash table, using 3701147191Sjkoshy * PMC_FLAG_REMOVE. 3702147191Sjkoshy */ 3703147191Sjkoshy PMCDBG(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid, 3704147191Sjkoshy p->p_comm); 3705147191Sjkoshy 3706147191Sjkoshy critical_enter(); /* no preemption */ 3707147191Sjkoshy 3708147191Sjkoshy cpu = curthread->td_oncpu; 3709147191Sjkoshy 3710147191Sjkoshy if ((pp = pmc_find_process_descriptor(p, 3711147191Sjkoshy PMC_FLAG_REMOVE)) != NULL) { 3712147191Sjkoshy 3713147191Sjkoshy PMCDBG(PRC,EXT,2, 3714147191Sjkoshy "process-exit proc=%p pmc-process=%p", p, pp); 3715147191Sjkoshy 3716147191Sjkoshy /* 3717147191Sjkoshy * The exiting process could the target of 3718147191Sjkoshy * some PMCs which will be running on 3719147191Sjkoshy * currently executing CPU. 3720147191Sjkoshy * 3721147191Sjkoshy * We need to turn these PMCs off like we 3722147191Sjkoshy * would do at context switch OUT time. 3723147191Sjkoshy */ 3724147191Sjkoshy for (ri = 0; ri < md->pmd_npmc; ri++) { 3725147191Sjkoshy 3726147191Sjkoshy /* 3727147191Sjkoshy * Pick up the pmc pointer from hardware 3728147191Sjkoshy * state similar to the CSW_OUT code. 3729147191Sjkoshy */ 3730147191Sjkoshy pm = NULL; 3731147191Sjkoshy (void) (*md->pmd_get_config)(cpu, ri, &pm); 3732147191Sjkoshy 3733147191Sjkoshy PMCDBG(PRC,EXT,2, "ri=%d pm=%p", ri, pm); 3734147191Sjkoshy 3735147191Sjkoshy if (pm == NULL || 3736147191Sjkoshy !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 3737147191Sjkoshy continue; 3738147191Sjkoshy 3739147191Sjkoshy PMCDBG(PRC,EXT,2, "ppmcs[%d]=%p pm=%p " 3740147191Sjkoshy "state=%d", ri, pp->pp_pmcs[ri].pp_pmc, 3741147191Sjkoshy pm, pm->pm_state); 3742147191Sjkoshy 3743147191Sjkoshy KASSERT(PMC_TO_ROWINDEX(pm) == ri, 3744147191Sjkoshy ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", 3745147191Sjkoshy __LINE__, PMC_TO_ROWINDEX(pm), ri)); 3746147191Sjkoshy 3747147191Sjkoshy KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 3748147191Sjkoshy ("[pmc,%d] pm %p != pp_pmcs[%d] %p", 3749147191Sjkoshy __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc)); 3750147191Sjkoshy 3751147191Sjkoshy (void) md->pmd_stop_pmc(cpu, ri); 3752147191Sjkoshy 3753147191Sjkoshy KASSERT(pm->pm_runcount > 0, 3754147191Sjkoshy ("[pmc,%d] bad runcount ri %d rc %d", 3755147191Sjkoshy __LINE__, ri, pm->pm_runcount)); 3756147191Sjkoshy 3757147867Sjkoshy /* Stop hardware only if it is actually running */ 3758147191Sjkoshy if (pm->pm_state == PMC_STATE_RUNNING && 3759147867Sjkoshy pm->pm_stalled == 0) { 3760147191Sjkoshy md->pmd_read_pmc(cpu, ri, &newvalue); 3761147191Sjkoshy tmp = newvalue - 3762147191Sjkoshy PMC_PCPU_SAVED(cpu,ri); 3763147191Sjkoshy 3764147191Sjkoshy mtx_pool_lock_spin(pmc_mtxpool, pm); 3765147191Sjkoshy pm->pm_gv.pm_savedvalue += tmp; 3766147191Sjkoshy pp->pp_pmcs[ri].pp_pmcval += tmp; 3767147191Sjkoshy mtx_pool_unlock_spin(pmc_mtxpool, pm); 3768147191Sjkoshy } 3769147191Sjkoshy 3770147191Sjkoshy atomic_subtract_rel_32(&pm->pm_runcount,1); 3771147191Sjkoshy 3772147191Sjkoshy KASSERT((int) pm->pm_runcount >= 0, 3773147191Sjkoshy ("[pmc,%d] runcount is %d", __LINE__, ri)); 3774147191Sjkoshy 3775147191Sjkoshy (void) md->pmd_config_pmc(cpu, ri, NULL); 3776147191Sjkoshy } 3777147191Sjkoshy 3778147191Sjkoshy /* 3779147191Sjkoshy * Inform the MD layer of this pseudo "context switch 3780147191Sjkoshy * out" 3781147191Sjkoshy */ 3782147191Sjkoshy (void) md->pmd_switch_out(pmc_pcpu[cpu], pp); 3783147191Sjkoshy 3784147191Sjkoshy critical_exit(); /* ok to be pre-empted now */ 3785147191Sjkoshy 3786147191Sjkoshy /* 3787147191Sjkoshy * Unlink this process from the PMCs that are 3788147191Sjkoshy * targetting it. This will send a signal to 3789147191Sjkoshy * all PMC owner's whose PMCs are orphaned. 3790147191Sjkoshy * 3791147191Sjkoshy * Log PMC value at exit time if requested. 3792147191Sjkoshy */ 3793147191Sjkoshy for (ri = 0; ri < md->pmd_npmc; ri++) 3794147191Sjkoshy if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 3795147867Sjkoshy if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && 3796147867Sjkoshy PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm))) 3797147191Sjkoshy pmclog_process_procexit(pm, pp); 3798147191Sjkoshy pmc_unlink_target_process(pm, pp); 3799147191Sjkoshy } 3800147191Sjkoshy FREE(pp, M_PMC); 3801147191Sjkoshy 3802147191Sjkoshy } else 3803147191Sjkoshy critical_exit(); /* pp == NULL */ 3804147191Sjkoshy 3805147191Sjkoshy 3806147191Sjkoshy /* 3807147191Sjkoshy * If the process owned PMCs, free them up and free up 3808147191Sjkoshy * memory. 3809147191Sjkoshy */ 3810147191Sjkoshy if ((po = pmc_find_owner_descriptor(p)) != NULL) { 3811147191Sjkoshy pmc_remove_owner(po); 3812147191Sjkoshy pmc_destroy_owner_descriptor(po); 3813145256Sjkoshy } 3814147191Sjkoshy 3815147191Sjkoshy sx_xunlock(&pmc_sx); 3816145256Sjkoshy} 3817145256Sjkoshy 3818145256Sjkoshy/* 3819145256Sjkoshy * Handle a process fork. 3820145256Sjkoshy * 3821145256Sjkoshy * If the parent process 'p1' is under HWPMC monitoring, then copy 3822145256Sjkoshy * over any attached PMCs that have 'do_descendants' semantics. 3823145256Sjkoshy */ 3824145256Sjkoshy 3825145256Sjkoshystatic void 3826147191Sjkoshypmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc, 3827145256Sjkoshy int flags) 3828145256Sjkoshy{ 3829145256Sjkoshy int is_using_hwpmcs; 3830147191Sjkoshy unsigned int ri; 3831147191Sjkoshy uint32_t do_descendants; 3832147191Sjkoshy struct pmc *pm; 3833147191Sjkoshy struct pmc_owner *po; 3834147191Sjkoshy struct pmc_process *ppnew, *ppold; 3835145256Sjkoshy 3836145256Sjkoshy (void) flags; /* unused parameter */ 3837145256Sjkoshy 3838145256Sjkoshy PROC_LOCK(p1); 3839145256Sjkoshy is_using_hwpmcs = p1->p_flag & P_HWPMC; 3840145256Sjkoshy PROC_UNLOCK(p1); 3841145256Sjkoshy 3842147191Sjkoshy /* 3843147191Sjkoshy * If there are system-wide sampling PMCs active, we need to 3844147191Sjkoshy * log all fork events to their owner's logs. 3845147191Sjkoshy */ 3846147191Sjkoshy 3847147191Sjkoshy LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 3848147191Sjkoshy if (po->po_flags & PMC_PO_OWNS_LOGFILE) 3849147191Sjkoshy pmclog_process_procfork(po, p1->p_pid, newproc->p_pid); 3850147191Sjkoshy 3851147191Sjkoshy if (!is_using_hwpmcs) 3852147191Sjkoshy return; 3853147191Sjkoshy 3854147191Sjkoshy PMC_GET_SX_XLOCK(); 3855147191Sjkoshy PMCDBG(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1, 3856147191Sjkoshy p1->p_pid, p1->p_comm, newproc); 3857147191Sjkoshy 3858147191Sjkoshy /* 3859147191Sjkoshy * If the parent process (curthread->td_proc) is a 3860147191Sjkoshy * target of any PMCs, look for PMCs that are to be 3861147191Sjkoshy * inherited, and link these into the new process 3862147191Sjkoshy * descriptor. 3863147191Sjkoshy */ 3864147191Sjkoshy if ((ppold = pmc_find_process_descriptor(curthread->td_proc, 3865147191Sjkoshy PMC_FLAG_NONE)) == NULL) 3866147191Sjkoshy goto done; /* nothing to do */ 3867147191Sjkoshy 3868147191Sjkoshy do_descendants = 0; 3869147191Sjkoshy for (ri = 0; ri < md->pmd_npmc; ri++) 3870147191Sjkoshy if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL) 3871147191Sjkoshy do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS; 3872147191Sjkoshy if (do_descendants == 0) /* nothing to do */ 3873147191Sjkoshy goto done; 3874147191Sjkoshy 3875147191Sjkoshy /* allocate a descriptor for the new process */ 3876147191Sjkoshy if ((ppnew = pmc_find_process_descriptor(newproc, 3877147191Sjkoshy PMC_FLAG_ALLOCATE)) == NULL) 3878147191Sjkoshy goto done; 3879147191Sjkoshy 3880147191Sjkoshy /* 3881147191Sjkoshy * Run through all PMCs that were targeting the old process 3882147191Sjkoshy * and which specified F_DESCENDANTS and attach them to the 3883147191Sjkoshy * new process. 3884147191Sjkoshy * 3885147191Sjkoshy * Log the fork event to all owners of PMCs attached to this 3886147191Sjkoshy * process, if not already logged. 3887147191Sjkoshy */ 3888147191Sjkoshy for (ri = 0; ri < md->pmd_npmc; ri++) 3889147191Sjkoshy if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && 3890147191Sjkoshy (pm->pm_flags & PMC_F_DESCENDANTS)) { 3891147191Sjkoshy pmc_link_target_process(pm, ppnew); 3892147191Sjkoshy po = pm->pm_owner; 3893147191Sjkoshy if (po->po_sscount == 0 && 3894147191Sjkoshy po->po_flags & PMC_PO_OWNS_LOGFILE) 3895147191Sjkoshy pmclog_process_procfork(po, p1->p_pid, 3896147191Sjkoshy newproc->p_pid); 3897147191Sjkoshy } 3898147191Sjkoshy 3899147191Sjkoshy /* 3900147191Sjkoshy * Now mark the new process as being tracked by this driver. 3901147191Sjkoshy */ 3902147191Sjkoshy PROC_LOCK(newproc); 3903147191Sjkoshy newproc->p_flag |= P_HWPMC; 3904147191Sjkoshy PROC_UNLOCK(newproc); 3905147191Sjkoshy 3906147191Sjkoshy done: 3907147191Sjkoshy sx_xunlock(&pmc_sx); 3908145256Sjkoshy} 3909145256Sjkoshy 3910145256Sjkoshy 3911145256Sjkoshy/* 3912145256Sjkoshy * initialization 3913145256Sjkoshy */ 3914145256Sjkoshy 3915145256Sjkoshystatic const char *pmc_name_of_pmcclass[] = { 3916145256Sjkoshy#undef __PMC_CLASS 3917145256Sjkoshy#define __PMC_CLASS(N) #N , 3918145256Sjkoshy __PMC_CLASSES() 3919145256Sjkoshy}; 3920145256Sjkoshy 3921145256Sjkoshystatic int 3922145256Sjkoshypmc_initialize(void) 3923145256Sjkoshy{ 3924147191Sjkoshy int cpu, error, n; 3925145256Sjkoshy struct pmc_binding pb; 3926147191Sjkoshy struct pmc_samplebuffer *sb; 3927145256Sjkoshy 3928145256Sjkoshy md = NULL; 3929145256Sjkoshy error = 0; 3930145256Sjkoshy 3931153110Sru#ifdef DEBUG 3932145256Sjkoshy /* parse debug flags first */ 3933145256Sjkoshy if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags", 3934145256Sjkoshy pmc_debugstr, sizeof(pmc_debugstr))) 3935145256Sjkoshy pmc_debugflags_parse(pmc_debugstr, 3936145256Sjkoshy pmc_debugstr+strlen(pmc_debugstr)); 3937145256Sjkoshy#endif 3938145256Sjkoshy 3939145256Sjkoshy PMCDBG(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION); 3940145256Sjkoshy 3941148562Sjkoshy /* check kernel version */ 3942148562Sjkoshy if (pmc_kernel_version != PMC_VERSION) { 3943148562Sjkoshy if (pmc_kernel_version == 0) 3944148562Sjkoshy printf("hwpmc: this kernel has not been compiled with " 3945148562Sjkoshy "'options HWPMC_HOOKS'.\n"); 3946148562Sjkoshy else 3947148562Sjkoshy printf("hwpmc: kernel version (0x%x) does not match " 3948148562Sjkoshy "module version (0x%x).\n", pmc_kernel_version, 3949148562Sjkoshy PMC_VERSION); 3950148562Sjkoshy return EPROGMISMATCH; 3951148562Sjkoshy } 3952148562Sjkoshy 3953145256Sjkoshy /* 3954145256Sjkoshy * check sysctl parameters 3955145256Sjkoshy */ 3956145256Sjkoshy 3957145256Sjkoshy if (pmc_hashsize <= 0) { 3958147191Sjkoshy (void) printf("hwpmc: tunable hashsize=%d must be greater " 3959147191Sjkoshy "than zero.\n", pmc_hashsize); 3960145256Sjkoshy pmc_hashsize = PMC_HASH_SIZE; 3961145256Sjkoshy } 3962145256Sjkoshy 3963147191Sjkoshy if (pmc_nsamples <= 0 || pmc_nsamples > 65535) { 3964153735Sjkoshy (void) printf("hwpmc: tunable nsamples=%d out of range.\n", 3965153735Sjkoshy pmc_nsamples); 3966147191Sjkoshy pmc_nsamples = PMC_NSAMPLES; 3967147191Sjkoshy } 3968145256Sjkoshy 3969147191Sjkoshy md = pmc_md_initialize(); 3970147191Sjkoshy 3971145256Sjkoshy if (md == NULL || md->pmd_init == NULL) 3972145256Sjkoshy return ENOSYS; 3973145256Sjkoshy 3974145256Sjkoshy /* allocate space for the per-cpu array */ 3975145256Sjkoshy MALLOC(pmc_pcpu, struct pmc_cpu **, mp_ncpus * sizeof(struct pmc_cpu *), 3976145256Sjkoshy M_PMC, M_WAITOK|M_ZERO); 3977145256Sjkoshy 3978145256Sjkoshy /* per-cpu 'saved values' for managing process-mode PMCs */ 3979145256Sjkoshy MALLOC(pmc_pcpu_saved, pmc_value_t *, 3980145256Sjkoshy sizeof(pmc_value_t) * mp_ncpus * md->pmd_npmc, M_PMC, M_WAITOK); 3981145256Sjkoshy 3982145256Sjkoshy /* perform cpu dependent initialization */ 3983145256Sjkoshy pmc_save_cpu_binding(&pb); 3984145256Sjkoshy for (cpu = 0; cpu < mp_ncpus; cpu++) { 3985145256Sjkoshy if (pmc_cpu_is_disabled(cpu)) 3986145256Sjkoshy continue; 3987145256Sjkoshy pmc_select_cpu(cpu); 3988145256Sjkoshy if ((error = md->pmd_init(cpu)) != 0) 3989145256Sjkoshy break; 3990145256Sjkoshy } 3991145256Sjkoshy pmc_restore_cpu_binding(&pb); 3992145256Sjkoshy 3993145256Sjkoshy if (error != 0) 3994145256Sjkoshy return error; 3995145256Sjkoshy 3996147191Sjkoshy /* allocate space for the sample array */ 3997147191Sjkoshy for (cpu = 0; cpu < mp_ncpus; cpu++) { 3998147191Sjkoshy if (pmc_cpu_is_disabled(cpu)) 3999147191Sjkoshy continue; 4000147191Sjkoshy MALLOC(sb, struct pmc_samplebuffer *, 4001147191Sjkoshy sizeof(struct pmc_samplebuffer) + 4002147191Sjkoshy pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 4003147191Sjkoshy M_WAITOK|M_ZERO); 4004147191Sjkoshy 4005147191Sjkoshy sb->ps_read = sb->ps_write = sb->ps_samples; 4006153735Sjkoshy sb->ps_fence = sb->ps_samples + pmc_nsamples; 4007147191Sjkoshy KASSERT(pmc_pcpu[cpu] != NULL, 4008147191Sjkoshy ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); 4009147191Sjkoshy 4010147191Sjkoshy pmc_pcpu[cpu]->pc_sb = sb; 4011147191Sjkoshy } 4012147191Sjkoshy 4013145256Sjkoshy /* allocate space for the row disposition array */ 4014145256Sjkoshy pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc, 4015145256Sjkoshy M_PMC, M_WAITOK|M_ZERO); 4016145256Sjkoshy 4017145256Sjkoshy KASSERT(pmc_pmcdisp != NULL, 4018145256Sjkoshy ("[pmc,%d] pmcdisp allocation returned NULL", __LINE__)); 4019145256Sjkoshy 4020145256Sjkoshy /* mark all PMCs as available */ 4021145256Sjkoshy for (n = 0; n < (int) md->pmd_npmc; n++) 4022145256Sjkoshy PMC_MARK_ROW_FREE(n); 4023145256Sjkoshy 4024145256Sjkoshy /* allocate thread hash tables */ 4025145256Sjkoshy pmc_ownerhash = hashinit(pmc_hashsize, M_PMC, 4026145256Sjkoshy &pmc_ownerhashmask); 4027145256Sjkoshy 4028145256Sjkoshy pmc_processhash = hashinit(pmc_hashsize, M_PMC, 4029145256Sjkoshy &pmc_processhashmask); 4030145256Sjkoshy mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc", MTX_SPIN); 4031145256Sjkoshy 4032147191Sjkoshy LIST_INIT(&pmc_ss_owners); 4033147191Sjkoshy pmc_ss_count = 0; 4034147191Sjkoshy 4035145256Sjkoshy /* allocate a pool of spin mutexes */ 4036145256Sjkoshy pmc_mtxpool = mtx_pool_create("pmc", pmc_mtxpool_size, MTX_SPIN); 4037145256Sjkoshy 4038145256Sjkoshy PMCDBG(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx " 4039145256Sjkoshy "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask, 4040145256Sjkoshy pmc_processhash, pmc_processhashmask); 4041145256Sjkoshy 4042145256Sjkoshy /* register process {exit,fork,exec} handlers */ 4043145256Sjkoshy pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit, 4044145256Sjkoshy pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY); 4045145256Sjkoshy pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork, 4046145256Sjkoshy pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY); 4047145256Sjkoshy 4048147191Sjkoshy /* initialize logging */ 4049147191Sjkoshy pmclog_initialize(); 4050147191Sjkoshy 4051145256Sjkoshy /* set hook functions */ 4052145256Sjkoshy pmc_intr = md->pmd_intr; 4053145256Sjkoshy pmc_hook = pmc_hook_handler; 4054145256Sjkoshy 4055145256Sjkoshy if (error == 0) { 4056145256Sjkoshy printf(PMC_MODULE_NAME ":"); 4057149373Sjkoshy for (n = 0; n < (int) md->pmd_nclass; n++) { 4058149373Sjkoshy printf(" %s/%d/0x%b", 4059145774Sjkoshy pmc_name_of_pmcclass[md->pmd_classes[n].pm_class], 4060149373Sjkoshy md->pmd_nclasspmcs[n], 4061149373Sjkoshy md->pmd_classes[n].pm_caps, 4062149373Sjkoshy "\20" 4063149373Sjkoshy "\1INT\2USR\3SYS\4EDG\5THR" 4064149373Sjkoshy "\6REA\7WRI\10INV\11QUA\12PRC" 4065149373Sjkoshy "\13TAG\14CSC"); 4066149373Sjkoshy } 4067145256Sjkoshy printf("\n"); 4068145256Sjkoshy } 4069145256Sjkoshy 4070145256Sjkoshy return error; 4071145256Sjkoshy} 4072145256Sjkoshy 4073145256Sjkoshy/* prepare to be unloaded */ 4074145256Sjkoshystatic void 4075145256Sjkoshypmc_cleanup(void) 4076145256Sjkoshy{ 4077145256Sjkoshy int cpu; 4078145256Sjkoshy struct pmc_ownerhash *ph; 4079145256Sjkoshy struct pmc_owner *po, *tmp; 4080145256Sjkoshy struct pmc_binding pb; 4081153110Sru#ifdef DEBUG 4082145256Sjkoshy struct pmc_processhash *prh; 4083145256Sjkoshy#endif 4084145256Sjkoshy 4085145256Sjkoshy PMCDBG(MOD,INI,0, "%s", "cleanup"); 4086145256Sjkoshy 4087147191Sjkoshy /* switch off sampling */ 4088147191Sjkoshy atomic_store_rel_int(&pmc_cpumask, 0); 4089147191Sjkoshy pmc_intr = NULL; 4090145256Sjkoshy 4091145256Sjkoshy sx_xlock(&pmc_sx); 4092145256Sjkoshy if (pmc_hook == NULL) { /* being unloaded already */ 4093145256Sjkoshy sx_xunlock(&pmc_sx); 4094145256Sjkoshy return; 4095145256Sjkoshy } 4096145256Sjkoshy 4097145256Sjkoshy pmc_hook = NULL; /* prevent new threads from entering module */ 4098145256Sjkoshy 4099145256Sjkoshy /* deregister event handlers */ 4100145256Sjkoshy EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag); 4101145256Sjkoshy EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag); 4102145256Sjkoshy 4103145256Sjkoshy /* send SIGBUS to all owner threads, free up allocations */ 4104145256Sjkoshy if (pmc_ownerhash) 4105145256Sjkoshy for (ph = pmc_ownerhash; 4106145256Sjkoshy ph <= &pmc_ownerhash[pmc_ownerhashmask]; 4107145256Sjkoshy ph++) { 4108145256Sjkoshy LIST_FOREACH_SAFE(po, ph, po_next, tmp) { 4109145256Sjkoshy pmc_remove_owner(po); 4110145256Sjkoshy 4111145256Sjkoshy /* send SIGBUS to owner processes */ 4112145256Sjkoshy PMCDBG(MOD,INI,2, "cleanup signal proc=%p " 4113145256Sjkoshy "(%d, %s)", po->po_owner, 4114145256Sjkoshy po->po_owner->p_pid, 4115145256Sjkoshy po->po_owner->p_comm); 4116145256Sjkoshy 4117145256Sjkoshy PROC_LOCK(po->po_owner); 4118145256Sjkoshy psignal(po->po_owner, SIGBUS); 4119145256Sjkoshy PROC_UNLOCK(po->po_owner); 4120147191Sjkoshy 4121147191Sjkoshy pmc_destroy_owner_descriptor(po); 4122145256Sjkoshy } 4123145256Sjkoshy } 4124145256Sjkoshy 4125145256Sjkoshy /* reclaim allocated data structures */ 4126145256Sjkoshy if (pmc_mtxpool) 4127145256Sjkoshy mtx_pool_destroy(&pmc_mtxpool); 4128145256Sjkoshy 4129145256Sjkoshy mtx_destroy(&pmc_processhash_mtx); 4130145256Sjkoshy if (pmc_processhash) { 4131153110Sru#ifdef DEBUG 4132145256Sjkoshy struct pmc_process *pp; 4133145256Sjkoshy 4134145256Sjkoshy PMCDBG(MOD,INI,3, "%s", "destroy process hash"); 4135145256Sjkoshy for (prh = pmc_processhash; 4136145256Sjkoshy prh <= &pmc_processhash[pmc_processhashmask]; 4137145256Sjkoshy prh++) 4138145256Sjkoshy LIST_FOREACH(pp, prh, pp_next) 4139145256Sjkoshy PMCDBG(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid); 4140145256Sjkoshy#endif 4141145256Sjkoshy 4142145256Sjkoshy hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask); 4143145256Sjkoshy pmc_processhash = NULL; 4144145256Sjkoshy } 4145145256Sjkoshy 4146145256Sjkoshy if (pmc_ownerhash) { 4147145256Sjkoshy PMCDBG(MOD,INI,3, "%s", "destroy owner hash"); 4148145256Sjkoshy hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask); 4149145256Sjkoshy pmc_ownerhash = NULL; 4150145256Sjkoshy } 4151145256Sjkoshy 4152147191Sjkoshy KASSERT(LIST_EMPTY(&pmc_ss_owners), 4153147191Sjkoshy ("[pmc,%d] Global SS owner list not empty", __LINE__)); 4154147191Sjkoshy KASSERT(pmc_ss_count == 0, 4155147191Sjkoshy ("[pmc,%d] Global SS count not empty", __LINE__)); 4156147191Sjkoshy 4157153735Sjkoshy /* free the per-cpu sample buffers */ 4158153735Sjkoshy for (cpu = 0; cpu < mp_ncpus; cpu++) { 4159153735Sjkoshy if (pmc_cpu_is_disabled(cpu)) 4160153735Sjkoshy continue; 4161153735Sjkoshy KASSERT(pmc_pcpu[cpu]->pc_sb != NULL, 4162153735Sjkoshy ("[pmc,%d] Null cpu sample buffer cpu=%d", __LINE__, 4163153735Sjkoshy cpu)); 4164153735Sjkoshy FREE(pmc_pcpu[cpu]->pc_sb, M_PMC); 4165153735Sjkoshy pmc_pcpu[cpu]->pc_sb = NULL; 4166153735Sjkoshy } 4167153735Sjkoshy 4168145256Sjkoshy /* do processor dependent cleanup */ 4169145256Sjkoshy PMCDBG(MOD,INI,3, "%s", "md cleanup"); 4170145256Sjkoshy if (md) { 4171145256Sjkoshy pmc_save_cpu_binding(&pb); 4172145256Sjkoshy for (cpu = 0; cpu < mp_ncpus; cpu++) { 4173145256Sjkoshy PMCDBG(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p", 4174145256Sjkoshy cpu, pmc_pcpu[cpu]); 4175145256Sjkoshy if (pmc_cpu_is_disabled(cpu)) 4176145256Sjkoshy continue; 4177145256Sjkoshy pmc_select_cpu(cpu); 4178145256Sjkoshy if (pmc_pcpu[cpu]) 4179145256Sjkoshy (void) md->pmd_cleanup(cpu); 4180145256Sjkoshy } 4181145256Sjkoshy FREE(md, M_PMC); 4182145256Sjkoshy md = NULL; 4183145256Sjkoshy pmc_restore_cpu_binding(&pb); 4184145256Sjkoshy } 4185145256Sjkoshy 4186145256Sjkoshy /* deallocate per-cpu structures */ 4187145256Sjkoshy FREE(pmc_pcpu, M_PMC); 4188145256Sjkoshy pmc_pcpu = NULL; 4189145256Sjkoshy 4190145256Sjkoshy FREE(pmc_pcpu_saved, M_PMC); 4191145256Sjkoshy pmc_pcpu_saved = NULL; 4192145256Sjkoshy 4193145256Sjkoshy if (pmc_pmcdisp) { 4194145256Sjkoshy FREE(pmc_pmcdisp, M_PMC); 4195145256Sjkoshy pmc_pmcdisp = NULL; 4196145256Sjkoshy } 4197145256Sjkoshy 4198147191Sjkoshy pmclog_shutdown(); 4199147191Sjkoshy 4200145256Sjkoshy sx_xunlock(&pmc_sx); /* we are done */ 4201145256Sjkoshy} 4202145256Sjkoshy 4203145256Sjkoshy/* 4204145256Sjkoshy * The function called at load/unload. 4205145256Sjkoshy */ 4206145256Sjkoshy 4207145256Sjkoshystatic int 4208145256Sjkoshyload (struct module *module __unused, int cmd, void *arg __unused) 4209145256Sjkoshy{ 4210145256Sjkoshy int error; 4211145256Sjkoshy 4212145256Sjkoshy error = 0; 4213145256Sjkoshy 4214145256Sjkoshy switch (cmd) { 4215145256Sjkoshy case MOD_LOAD : 4216145256Sjkoshy /* initialize the subsystem */ 4217145256Sjkoshy error = pmc_initialize(); 4218145256Sjkoshy if (error != 0) 4219145256Sjkoshy break; 4220145256Sjkoshy PMCDBG(MOD,INI,1, "syscall=%d ncpus=%d", 4221145256Sjkoshy pmc_syscall_num, mp_ncpus); 4222145256Sjkoshy break; 4223145256Sjkoshy 4224145256Sjkoshy 4225145256Sjkoshy case MOD_UNLOAD : 4226145256Sjkoshy case MOD_SHUTDOWN: 4227145256Sjkoshy pmc_cleanup(); 4228145256Sjkoshy PMCDBG(MOD,INI,1, "%s", "unloaded"); 4229145256Sjkoshy break; 4230145256Sjkoshy 4231145256Sjkoshy default : 4232145256Sjkoshy error = EINVAL; /* XXX should panic(9) */ 4233145256Sjkoshy break; 4234145256Sjkoshy } 4235145256Sjkoshy 4236145256Sjkoshy return error; 4237145256Sjkoshy} 4238145256Sjkoshy 4239145256Sjkoshy/* memory pool */ 4240145256SjkoshyMALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module"); 4241