1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License, Version 1.0 only 6 * (the "License"). You may not use this file except in compliance 7 * with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or http://www.opensolaris.org/os/licensing. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright 2005 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 #pragma ident "%Z%%M% %I% %E% SMI" 28 29 #include <sys/systm.h> 30 #include <sys/conf.h> 31 #include <sys/stat.h> 32 #include <sys/ddi.h> 33 #include <sys/sunddi.h> 34 #include <sys/modctl.h> 35 #include <sys/cpu_module.h> 36 #include <vm/hat_sfmmu.h> 37 #include <vm/seg_kmem.h> 38 #include <vm/seg_kpm.h> 39 #include <vm/vm_dep.h> 40 #include <sys/machsystm.h> 41 #include <sys/machasi.h> 42 #include <sys/sysmacros.h> 43 #include <sys/callb.h> 44 #include <sys/archsystm.h> 45 #include <sys/trapstat.h> 46 #ifdef sun4v 47 #include <sys/hypervisor_api.h> 48 #endif 49 50 /* BEGIN CSTYLED */ 51 /* 52 * trapstat: Trap Statistics through Dynamic Trap Table Interposition 53 * ------------------------------------------------------------------- 54 * 55 * Motivation and Overview 56 * 57 * Despite being a fundamental indicator of system behavior, there has 58 * historically been very little insight provided into the frequency and cost 59 * of machine-specific traps. The lack of insight has been especially acute 60 * on UltraSPARC microprocessors: because these microprocessors handle TLB 61 * misses as software traps, the frequency and duration of traps play a 62 * decisive role in the performance of the memory system. As applications have 63 * increasingly outstripped TLB reach, this has become increasingly true. 64 * 65 * Part of the difficulty of observing trap behavior is that the trap handlers 66 * are so frequently called (e.g. millions of times per second) that any 67 * permanently enabled instrumentation would induce an unacceptable performance 68 * degradation. Thus, it is a constraint on any trap observability 69 * infrastructure that it have no probe effect when not explicitly enabled. 70 * 71 * The basic idea, then, is to create an interposing trap table in which each 72 * entry increments a per-trap, in-memory counter and then jumps to the actual, 73 * underlying trap table entry. To enable trapstat, we atomically write to the 74 * trap base address (%tba) register to point to our interposing trap table. 75 * (Note that per-CPU statistics fall out by creating a different trap table 76 * for each CPU.) 77 * 78 * Implementation Details 79 * 80 * While the idea is straight-forward, a nuance of SPARC V9 slightly 81 * complicates the implementation. Unlike its predecessors, SPARC V9 supports 82 * the notion of nested traps. The trap level is kept in the TL register: 83 * during normal operation it is 0; when a trap is taken, the TL register is 84 * incremented by 1. To aid system software, SPARC V9 breaks the trap table 85 * into two halves: the lower half contains the trap handlers for traps taken 86 * when TL is 0; the upper half contains the trap handlers for traps taken 87 * when TL is greater than 0. Each half is further subdivided into two 88 * subsequent halves: the lower half contains the trap handlers for traps 89 * other than those induced by the trap instruction (Tcc variants); the upper 90 * half contains the trap handlers for traps induced by the trap instruction. 91 * This gives a total of four ranges, with each range containing 256 traps: 92 * 93 * +--------------------------------+- 3ff 94 * | | . 95 * | Trap instruction, TL>0 | . 96 * | | . 97 * |- - - - - - - - - - - - - - - - +- 300 98 * |- - - - - - - - - - - - - - - - +- 2ff 99 * | | . 100 * | Non-trap instruction, TL>0 | . 101 * | | . 102 * |- - - - - - - - - - - - - - - - +- 200 103 * |- - - - - - - - - - - - - - - - +- 1ff 104 * | | . 105 * | Trap instruction, TL=0 | . 106 * | | . 107 * |- - - - - - - - - - - - - - - - +- 100 108 * |- - - - - - - - - - - - - - - - +- 0ff 109 * | | . 110 * | Non-trap instruction, TL=0 | . 111 * | | . 112 * +--------------------------------+- 000 113 * 114 * 115 * Solaris, however, doesn't have reason to support trap instructions when 116 * TL>0 (only privileged code may execute at TL>0; not supporting this only 117 * constrains our own implementation). The trap table actually looks like: 118 * 119 * +--------------------------------+- 2ff 120 * | | . 121 * | Non-trap instruction, TL>0 | . 122 * | | . 123 * |- - - - - - - - - - - - - - - - +- 200 124 * |- - - - - - - - - - - - - - - - +- 1ff 125 * | | . 126 * | Trap instruction, TL=0 | . 127 * | | . 128 * |- - - - - - - - - - - - - - - - +- 100 129 * |- - - - - - - - - - - - - - - - +- 0ff 130 * | | . 131 * | Non-trap instruction, TL=0 | . 132 * | | . 133 * +--------------------------------+- 000 134 * 135 * Putatively to aid system software, SPARC V9 has the notion of multiple 136 * sets of global registers. UltraSPARC defines four sets of global 137 * registers: 138 * 139 * Normal Globals 140 * Alternate Globals (AGs) 141 * MMU Globals (MGs) 142 * Interrupt Globals (IGs) 143 * 144 * The set of globals in use is controlled by bits in PSTATE; when TL is 0 145 * (and PSTATE has not been otherwise explicitly modified), the Normal Globals 146 * are in use. When a trap is issued, PSTATE is modified to point to a set of 147 * globals corresponding to the trap type. Most traps correspond to the 148 * Alternate Globals, with a minority corresponding to the MMU Globals, and 149 * only the interrupt-vector trap (vector 0x60) corresponding to the Interrupt 150 * Globals. (The complete mapping can be found in the UltraSPARC I&II User's 151 * Manual.) 152 * 153 * Note that the sets of globals are per trap _type_, not per trap _level_. 154 * Thus, when executing a TL>0 trap handler, one may not have registers 155 * available (for example, both trap-instruction traps and spill traps execute 156 * on the alternate globals; if a trap-instruction trap induces a window spill, 157 * the window spill handler has no available globals). For trapstat, this is 158 * problematic: a register is required to transfer control from one arbitrary 159 * location (in the interposing trap table) to another (in the actual trap 160 * table). 161 * 162 * We solve this problem by exploiting the trap table's location at the bottom 163 * of valid kernel memory (i.e. at KERNELBASE). We locate the interposing trap 164 * tables just below KERNELBASE -- thereby allowing us to use a branch-always 165 * instruction (ba) instead of a jump instruction (jmp) to transfer control 166 * from the TL>0 entries in the interposing trap table to the TL>0 entries in 167 * the actual trap table. (N.B. while this allows trap table interposition to 168 * work, it necessarily limits trapstat to only recording information about 169 * TL=0 traps -- there is no way to increment a counter without using a 170 * register.) Diagrammatically: 171 * 172 * Actual trap table: 173 * 174 * +--------------------------------+- 2ff 175 * | | . 176 * | Non-trap instruction, TL>0 | . <-----------------------+ 177 * | | . <-----------------------|-+ 178 * |- - - - - - - - - - - - - - - - +- 200 <-----------------------|-|-+ 179 * |- - - - - - - - - - - - - - - - +- 1ff | | | 180 * | | . | | | 181 * | Trap instruction, TL=0 | . <-----------------+ | | | 182 * | | . <-----------------|-+ | | | 183 * |- - - - - - - - - - - - - - - - +- 100 <-----------------|-|-+ | | | 184 * |- - - - - - - - - - - - - - - - +- 0ff | | | | | | 185 * | | . | | | | | | 186 * | Non-trap instruction, TL=0 | . <-----------+ | | | | | | 187 * | | . <-----------|-+ | | | | | | 188 * +--------------------------------+- 000 <-----------|-|-+ | | | | | | 189 * KERNELBASE | | | | | | | | | 190 * | | | | | | | | | 191 * | | | | | | | | | 192 * Interposing trap table: | | | | | | | | | 193 * | | | | | | | | | 194 * +--------------------------------+- 2ff | | | | | | | | | 195 * | ... | . | | | | | | | | | 196 * | ... | . | | | | | | | | | 197 * | ... | . | | | | | | | | | 198 * |- - - - - - - - - - - - - - - - +- 203 | | | | | | | | | 199 * | ba,a | -------------|-|-|-|-|-|-+ | | 200 * |- - - - - - - - - - - - - - - - +- 202 | | | | | | | | 201 * | ba,a | -------------|-|-|-|-|-|---+ | 202 * |- - - - - - - - - - - - - - - - +- 201 | | | | | | | 203 * | ba,a | -------------|-|-|-|-|-|-----+ 204 * |- - - - - - - - - - - - - - - - +- 200 | | | | | | 205 * | ... | . | | | | | | 206 * | ... | . | | | | | | 207 * | ... | . | | | | | | 208 * |- - - - - - - - - - - - - - - - +- 103 | | | | | | 209 * | (Increment counter) | | | | | | | 210 * | ba,a | -------------------+ | | 211 * |- - - - - - - - - - - - - - - - +- 102 | | | | | 212 * | (Increment counter) | | | | | | 213 * | ba,a | ---------------------+ | 214 * |- - - - - - - - - - - - - - - - +- 101 | | | | 215 * | (Increment counter) | | | | | 216 * | ba,a | -----------------------+ 217 * |- - - - - - - - - - - - - - - - +- 100 | | | 218 * | ... | . | | | 219 * | ... | . | | | 220 * | ... | . | | | 221 * |- - - - - - - - - - - - - - - - +- 003 | | | 222 * | (Increment counter) | | | | 223 * | ba,a | -------------+ | | 224 * |- - - - - - - - - - - - - - - - +- 002 | | 225 * | (Increment counter) | | | 226 * | ba,a | ---------------+ | 227 * |- - - - - - - - - - - - - - - - +- 001 | 228 * | (Increment counter) | | 229 * | ba,a | -----------------+ 230 * +--------------------------------+- 000 231 * KERNELBASE - tstat_total_size 232 * 233 * tstat_total_size is the number of pages required for each trap table. It 234 * must be true that KERNELBASE - tstat_total_size is less than the maximum 235 * branch displacement; if each CPU were to consume a disjoint virtual range 236 * below KERNELBASE for its trap table, we could support at most 237 * (maximum_branch_displacement / tstat_total_size) CPUs. The maximum branch 238 * displacement for Bicc variants is just under eight megabytes, and (because 239 * the %tba must be 32K aligned), tstat_total_size must be at least 32K; if 240 * each CPU were to consume a disjoint virtual range, we would have an 241 * unacceptably low upper bound of 256 CPUs. 242 * 243 * While there are tricks that one could use to address this constraint (e.g., 244 * creating trampolines every maximum_branch_displacement bytes), we instead 245 * solve this by not permitting each CPU to consume a disjoint virtual range. 246 * Rather, we have each CPU's interposing trap table use the _same_ virtual 247 * range, but we back the trap tables with disjoint physical memory. Normally, 248 * such one-to-many virtual-to-physical mappings are illegal; this is 249 * permissible here only because the pages for the interposing trap table are 250 * necessarily locked in the TLB. (The CPUs thus never have the opportunity to 251 * discover that they have conflicting translations.) 252 * 253 * On CMT architectures in which CPUs can share MMUs, the above trick will not 254 * work: two CPUs that share an MMU cannot have the same virtual address map 255 * to disjoint physical pages. On these architectures, any CPUs sharing the 256 * same MMU must consume a disjoint 32K virtual address range -- limiting the 257 * number of CPUs sharing an MMU on these architectures to 256 due to the 258 * branch displacement limitation described above. On the sun4v architecture, 259 * there is a further limitation: a guest may not have more than eight locked 260 * TLB entries per MMU. To allow operation under this restriction, the 261 * interposing trap table and the trap statistics are each accessed through 262 * a single 4M TLB entry. This limits the footprint to two locked entries 263 * (one for the I-TLB and one for the D-TLB), but further restricts the number 264 * of CPUs to 128 per MMU. However, support for more than 128 CPUs can easily 265 * be added via a hybrid scheme, where the same 4M virtual address is used 266 * on different MMUs. 267 * 268 * 269 * TLB Statistics 270 * 271 * Because TLB misses are an important component of system performance, we wish 272 * to know much more about these traps than simply the number received. 273 * Specifically, we wish to know: 274 * 275 * (a) The amount of time spent executing the TLB miss handler 276 * (b) TLB misses versus TSB misses 277 * (c) Kernel-level misses versus user-level misses 278 * (d) Misses per pagesize 279 * 280 * TLB Statistics: Time Spent Executing 281 * 282 * To accurately determine the amount of time spent executing the TLB miss 283 * handler, one must get a timestamp on trap entry and trap exit, subtract the 284 * latter from the former, and add the result to an accumulating count. 285 * Consider flow of control during normal TLB miss processing (where "ldx 286 * [%g2], %g2" is an arbitrary TLB-missing instruction): 287 * 288 * + - - - - - - - -+ 289 * : : 290 * : ldx [%g2], %g2 :<-------------------------------------------------------+ 291 * : : Return from trap: | 292 * + - - - - - - - -+ TL <- TL - 1 (0) | 293 * | %pc <- TSTATE[TL].TPC (address of load) | 294 * | TLB miss: | 295 * | TL <- TL + 1 (1) | 296 * | %pc <- TLB-miss-trap-handler | 297 * | | 298 * v | 299 * + - - - - - - - - - - - - - - - + | 300 * : : | 301 * : Lookup VA in TSB : | 302 * : If (hit) : | 303 * : Fill TLB : | 304 * : Else : | 305 * : Lookup VA (hme hash table : | 306 * : or segkpm) : | 307 * : Fill TLB : | 308 * : Endif : | 309 * : Issue "retry" ---------------------------------------------------------+ 310 * : : 311 * + - - - - - - - - - - - - - - - + 312 * TLB-miss-trap-handler 313 * 314 * 315 * As the above diagram indicates, interposing on the trap table allows one 316 * only to determine a timestamp on trap _entry_: when the TLB miss handler 317 * has completed filling the TLB, a "retry" will be issued, and control will 318 * transfer immediately back to the missing %pc. 319 * 320 * To obtain a timestamp on trap exit, we must then somehow interpose between 321 * the "retry" and the subsequent control transfer to the TLB-missing 322 * instruction. To do this, we _push_ a trap level. The basic idea is to 323 * spoof a TLB miss by raising TL, setting the %tpc to be within text 324 * controlled by trapstat (the "TLB return entry") and branching to the 325 * underlying TLB miss handler. When the TLB miss handler issues its "retry", 326 * control will transfer not to the TLB-missing instruction, but rather to the 327 * TLB return entry. This code can then obtain a timestamp, and issue its own 328 * "retry" -- thereby correctly returning to the TLB-missing instruction. 329 * Here is the above TLB miss flow control diagram modified to reflect 330 * trapstat's operation: 331 * 332 * + - - - - - - - -+ 333 * : : 334 * : ldx [%g2], %g2 :<-------------------------------------------------------+ 335 * : : Return from trap: | 336 * + - - - - - - - -+ TL <- TL - 1 (0) | 337 * | %pc <- TSTATE[TL].TPC (address of load) | 338 * | TLB miss: | 339 * | TL <- TL + 1 (1) | 340 * | %pc <- TLB-miss-trap-handler (trapstat) | 341 * | | 342 * v TLB-return-entry (trapstat) | 343 * + - - - - - - - - - - - - - - - - - - + + - - - - - - - - - - - - - + | 344 * : : : : | 345 * : Record timestamp : : Record timestamp : | 346 * : TL <- 2 : : Take timestamp difference : | 347 * : TSTATE[1].TPC <- TLB-return-entry : : Add to running total : | 348 * : ba,a TLB-miss-trap-handler -----------+ : Issue "retry" --------------+ 349 * : : | : : 350 * + - - - - - - - - - - - - - - - - - - + | + - - - - - - - - - - - - - + 351 * TLB-miss-trap-handler | ^ 352 * (trapstat) | | 353 * | | 354 * | | 355 * +-----------------------+ | 356 * | | 357 * | | 358 * v | 359 * + - - - - - - - - - - - - - - - + | 360 * : : | 361 * : Lookup VA in TSB : | 362 * : If (hit) : | 363 * : Fill TLB : | 364 * : Else : | 365 * : Lookup VA (hme hash table : | 366 * : or segkpm) : | 367 * : Fill TLB : | 368 * : Endif : | 369 * : Issue "retry" ------------------------------------------+ 370 * : : Return from trap: 371 * + - - - - - - - - - - - - - - - + TL <- TL - 1 (1) 372 * TLB-miss-trap-handler %pc <- TSTATE[TL].TPC (TLB-return-entry) 373 * 374 * 375 * A final subterfuge is required to complete our artifice: if we miss in 376 * the TLB, the TSB _and_ the subsequent hash or segkpm lookup (that is, if 377 * there is no valid translation for the TLB-missing address), common system 378 * software will need to accurately determine the %tpc as part of its page 379 * fault handling. We therefore modify the kernel to check the %tpc in this 380 * case: if the %tpc falls within the VA range controlled by trapstat and 381 * the TL is 2, TL is simply lowered back to 1 (this check is implemented 382 * by the TSTAT_CHECK_TL1 macro). Lowering TL to 1 has the effect of 383 * discarding the state pushed by trapstat. 384 * 385 * TLB Statistics: TLB Misses versus TSB Misses 386 * 387 * Distinguishing TLB misses from TSB misses requires further interposition 388 * on the TLB miss handler: we cannot know a priori or a posteriori if a 389 * given VA will or has hit in the TSB. 390 * 391 * We achieve this distinction by adding a second TLB return entry almost 392 * identical to the first -- differing only in the address to which it 393 * stores its results. We then modify the TLB miss handlers of the kernel 394 * such that they check the %tpc when they determine that a TLB miss has 395 * subsequently missed in the TSB: if the %tpc lies within trapstat's VA 396 * range and TL is 2 (that is, if trapstat is running), the TLB miss handler 397 * _increments_ the %tpc by the size of the TLB return entry. The ensuing 398 * "retry" will thus transfer control to the second TLB return entry, and 399 * the time spent in the handler will be accumulated in a memory location 400 * specific to TSB misses. 401 * 402 * N.B.: To minimize the amount of knowledge the kernel must have of trapstat, 403 * we do not allow the kernel to hard-code the size of the TLB return entry. 404 * Rather, the actual tsbmiss handler executes a known instruction at the 405 * corresponding tsbmiss patch points (see the tstat_tsbmiss_patch_table) with 406 * the %tpc in %g7: when trapstat is not running, these points contain the 407 * harmless TSTAT_TSBMISS_INSTR instruction ("add %g7, 0, %g7"). Before 408 * running, trapstat modifies the instructions at these patch points such 409 * that the simm13 equals the size of the TLB return entry. 410 * 411 * TLB Statistics: Kernel-level Misses versus User-level Misses 412 * 413 * Differentiating user-level misses from kernel-level misses employs a 414 * similar technique, but is simplified by the ability to distinguish a 415 * user-level miss from a kernel-level miss a priori by reading the context 416 * register: we implement kernel-/user-level differentiation by again doubling 417 * the number of TLB return entries, and setting the %tpc to the appropriate 418 * TLB return entry in trapstat's TLB miss handler. Together with the doubling 419 * of entries required for TLB-miss/TSB-miss differentiation, this yields a 420 * total of four TLB return entries: 421 * 422 * Level TSB hit? Structure member 423 * ------------------------------------------------------------ 424 * Kernel Yes tstat_tlbret_t.ttlbr_ktlb 425 * Kernel No tstat_tlbret_t.ttlbr_ktsb 426 * User Yes tstat_tlbret_t.ttlbr_utlb 427 * User No tstat_tlbret_t.ttlbr_utsb 428 * 429 * TLB Statistics: Misses per Pagesize 430 * 431 * As with the TLB-/TSB-miss differentiation, we have no way of determining 432 * pagesize a priori. This is therefore implemented by mandating a new rule: 433 * whenever the kernel fills the TLB in its TLB miss handler, the TTE 434 * corresponding to the TLB-missing VA must be in %g5 when the handler 435 * executes its "retry". This allows the TLB return entry to determine 436 * pagesize by simply looking at the pagesize field in the TTE stored in 437 * %g5. 438 * 439 * TLB Statistics: Probe Effect 440 * 441 * As one might imagine, gathering TLB statistics by pushing a trap level 442 * induces significant probe effect. To account for this probe effect, 443 * trapstat attempts to observe it by executing a code sequence with a known 444 * number of TLB misses both before and after interposing on the trap table. 445 * This allows trapstat to determine a per-trap probe effect which can then be 446 * factored into the "%tim" fields of the trapstat command. 447 * 448 * Note that on sun4v platforms, TLB misses are normally handled by the 449 * hypervisor or the hardware TSB walker. Thus no fast MMU miss information 450 * is reported for normal operation. However, when trapstat is invoked with 451 * -t or -T option to collect detailed TLB statistics, kernel takes 452 * over TLB miss handling. This results in significantly more overhead 453 * and TLB statistics may not be as accurate as on sun4u platforms. 454 * 455 * Locking 456 * 457 * The implementation uses two locks: tstat_lock (a local lock) and the global 458 * cpu_lock. tstat_lock is used to assure trapstat's consistency in the 459 * presence of multithreaded /dev/trapstat consumers (while as of this writing 460 * the only consumer of /dev/trapstat is single threaded, it is obviously 461 * necessary to correctly support multithreaded access). cpu_lock is held 462 * whenever CPUs are being manipulated directly, to prevent them from 463 * disappearing in the process. Because trapstat's DR callback 464 * (trapstat_cpu_setup()) must grab tstat_lock and is called with cpu_lock 465 * held, the lock ordering is necessarily cpu_lock before tstat_lock. 466 * 467 */ 468 /* END CSTYLED */ 469 470 static dev_info_t *tstat_devi; /* saved in xxattach() for xxinfo() */ 471 static int tstat_open; /* set if driver is open */ 472 static kmutex_t tstat_lock; /* serialize access */ 473 static vmem_t *tstat_arena; /* arena for TLB-locked pages */ 474 static tstat_percpu_t *tstat_percpu; /* per-CPU data */ 475 static int tstat_running; /* set if trapstat is running */ 476 static tstat_data_t *tstat_buffer; /* staging buffer for outgoing data */ 477 static int tstat_options; /* bit-wise indication of options */ 478 static int *tstat_enabled; /* map of enabled trap entries */ 479 static int tstat_tsbmiss_patched; /* tsbmiss patch flag */ 480 static callb_id_t tstat_cprcb; /* CPR callback */ 481 static char *tstat_probe_area; /* VA range used for probe effect */ 482 static caddr_t tstat_probe_phys; /* physical to back above VA */ 483 static hrtime_t tstat_probe_time; /* time spent on probe effect */ 484 static hrtime_t tstat_probe_before[TSTAT_PROBE_NLAPS]; 485 static hrtime_t tstat_probe_after[TSTAT_PROBE_NLAPS]; 486 static uint_t tstat_pgszs; /* # of kernel page sizes */ 487 static uint_t tstat_user_pgszs; /* # of user page sizes */ 488 489 /* 490 * sizeof tstat_data_t + pgsz data for the kernel. For simplicity's sake, when 491 * we collect data, we do it based upon szc, but when we report data back to 492 * userland, we have to do it based upon the userszc which may not match. 493 * So, these two variables are for internal use and exported use respectively. 494 */ 495 static size_t tstat_data_t_size; 496 static size_t tstat_data_t_exported_size; 497 498 static size_t tstat_data_pages; /* number of pages of tstat data */ 499 static size_t tstat_data_size; /* tstat data size in bytes */ 500 static size_t tstat_total_pages; /* #data pages + #instr pages */ 501 static size_t tstat_total_size; /* tstat data size + instr size */ 502 #ifdef sun4v 503 static caddr_t tstat_va; /* VA of memory reserved for TBA */ 504 static pfn_t tstat_pfn; /* PFN of memory reserved for TBA */ 505 #endif 506 507 /* 508 * In the above block comment, see "TLB Statistics: TLB Misses versus 509 * TSB Misses" for an explanation of the tsbmiss patch points. 510 */ 511 extern uint32_t tsbmiss_trapstat_patch_point; 512 extern uint32_t tsbmiss_trapstat_patch_point_kpm; 513 extern uint32_t tsbmiss_trapstat_patch_point_kpm_small; 514 515 /* 516 * Trapstat tsbmiss patch table 517 */ 518 tstat_tsbmiss_patch_entry_t tstat_tsbmiss_patch_table[] = { 519 {(uint32_t *)&tsbmiss_trapstat_patch_point, 0}, 520 {(uint32_t *)&tsbmiss_trapstat_patch_point_kpm, 0}, 521 {(uint32_t *)&tsbmiss_trapstat_patch_point_kpm_small, 0}, 522 {(uint32_t *)NULL, 0} 523 }; 524 525 /* 526 * We define some general SPARC-specific constants to allow more readable 527 * relocations. 528 */ 529 #define NOP 0x01000000 530 #define HI22(v) ((uint32_t)(v) >> 10) 531 #define LO10(v) ((uint32_t)(v) & 0x3ff) 532 #define LO12(v) ((uint32_t)(v) & 0xfff) 533 #define DISP22(from, to) \ 534 ((((uintptr_t)(to) - (uintptr_t)(from)) >> 2) & 0x3fffff) 535 #define ASI(asi) ((asi) << 5) 536 537 /* 538 * The interposing trap table must be locked in the I-TLB, and any data 539 * referred to in the interposing trap handler must be locked in the D-TLB. 540 * This function locks these pages in the appropriate TLBs by creating TTEs 541 * from whole cloth, and manually loading them into the TLB. This function is 542 * called from cross call context. 543 * 544 * On sun4v platforms, we use 4M page size mappings to minimize the number 545 * of locked down entries (i.e. permanent mappings). Each CPU uses a 546 * reserved portion of that 4M page for its TBA and data. 547 */ 548 static void 549 trapstat_load_tlb(void) 550 { 551 #ifndef sun4v 552 int i; 553 #else 554 uint64_t ret; 555 #endif 556 tte_t tte; 557 tstat_percpu_t *tcpu = &tstat_percpu[CPU->cpu_id]; 558 caddr_t va = tcpu->tcpu_vabase; 559 560 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_ALLOCATED); 561 ASSERT(!(tcpu->tcpu_flags & TSTAT_CPU_ENABLED)); 562 563 #ifndef sun4v 564 for (i = 0; i < tstat_total_pages; i++, va += MMU_PAGESIZE) { 565 tte.tte_inthi = TTE_VALID_INT | TTE_SZ_INT(TTE8K) | 566 TTE_PFN_INTHI(tcpu->tcpu_pfn[i]); 567 if (i < TSTAT_INSTR_PAGES) { 568 tte.tte_intlo = TTE_PFN_INTLO(tcpu->tcpu_pfn[i]) | 569 TTE_LCK_INT | TTE_CP_INT | TTE_PRIV_INT; 570 sfmmu_itlb_ld(va, KCONTEXT, &tte); 571 } else { 572 tte.tte_intlo = TTE_PFN_INTLO(tcpu->tcpu_pfn[i]) | 573 TTE_LCK_INT | TTE_CP_INT | TTE_CV_INT | 574 TTE_PRIV_INT | TTE_HWWR_INT; 575 sfmmu_dtlb_ld(va, KCONTEXT, &tte); 576 } 577 } 578 #else /* sun4v */ 579 tte.tte_inthi = TTE_VALID_INT | TTE_PFN_INTHI(tstat_pfn); 580 tte.tte_intlo = TTE_PFN_INTLO(tstat_pfn) | TTE_CP_INT | 581 TTE_CV_INT | TTE_PRIV_INT | TTE_HWWR_INT | 582 TTE_SZ_INTLO(TTE4M); 583 ret = hv_mmu_map_perm_addr(va, KCONTEXT, *(uint64_t *)&tte, 584 MAP_ITLB | MAP_DTLB); 585 586 if (ret != H_EOK) 587 cmn_err(CE_PANIC, "trapstat: cannot map new TBA " 588 "for cpu %d (error: 0x%lx)", CPU->cpu_id, ret); 589 #endif /* sun4v */ 590 } 591 592 /* 593 * As mentioned in the "TLB Statistics: TLB Misses versus TSB Misses" section 594 * of the block comment, TLB misses are differentiated from TSB misses in 595 * part by hot-patching the instructions at the tsbmiss patch points (see 596 * tstat_tsbmiss_patch_table). This routine is used both to initially patch 597 * the instructions, and to patch them back to their original values upon 598 * restoring the original trap table. 599 */ 600 static void 601 trapstat_hotpatch() 602 { 603 uint32_t instr; 604 uint32_t simm13; 605 tstat_tsbmiss_patch_entry_t *ep; 606 607 ASSERT(MUTEX_HELD(&tstat_lock)); 608 609 if (!(tstat_options & TSTAT_OPT_TLBDATA)) 610 return; 611 612 if (!tstat_tsbmiss_patched) { 613 /* 614 * We haven't patched the TSB paths; do so now. 615 */ 616 /*CONSTCOND*/ 617 ASSERT(offsetof(tstat_tlbret_t, ttlbr_ktsb) - 618 offsetof(tstat_tlbret_t, ttlbr_ktlb) == 619 offsetof(tstat_tlbret_t, ttlbr_utsb) - 620 offsetof(tstat_tlbret_t, ttlbr_utlb)); 621 622 simm13 = offsetof(tstat_tlbret_t, ttlbr_ktsb) - 623 offsetof(tstat_tlbret_t, ttlbr_ktlb); 624 625 for (ep = tstat_tsbmiss_patch_table; ep->tpe_addr; ep++) { 626 ASSERT(ep->tpe_instr == 0); 627 instr = ep->tpe_instr = *ep->tpe_addr; 628 629 /* 630 * Assert that the instruction we're about to patch is 631 * "add %g7, 0, %g7" (0x8e01e000). 632 */ 633 ASSERT(instr == TSTAT_TSBMISS_INSTR); 634 635 instr |= simm13; 636 hot_patch_kernel_text((caddr_t)ep->tpe_addr, 637 instr, sizeof (instr)); 638 } 639 640 tstat_tsbmiss_patched = 1; 641 642 } else { 643 /* 644 * Remove patches from the TSB paths. 645 */ 646 for (ep = tstat_tsbmiss_patch_table; ep->tpe_addr; ep++) { 647 ASSERT(ep->tpe_instr == TSTAT_TSBMISS_INSTR); 648 hot_patch_kernel_text((caddr_t)ep->tpe_addr, 649 ep->tpe_instr, sizeof (instr)); 650 ep->tpe_instr = 0; 651 } 652 653 tstat_tsbmiss_patched = 0; 654 } 655 } 656 657 /* 658 * This is the routine executed to clock the performance of the trap table, 659 * executed both before and after interposing on the trap table to attempt to 660 * determine probe effect. The probe effect is used to adjust the "%tim" 661 * fields of trapstat's -t and -T output; we only use TLB misses to clock the 662 * trap table. We execute the inner loop (which is designed to exceed the 663 * TLB's reach) nlaps times, taking the best time as our time (thereby 664 * factoring out the effects of interrupts, cache misses or other perturbing 665 * events. 666 */ 667 static hrtime_t 668 trapstat_probe_laps(int nlaps, hrtime_t *buf) 669 { 670 int i, j = 0; 671 hrtime_t ts, best = INT64_MAX; 672 673 while (nlaps--) { 674 ts = rdtick(); 675 676 for (i = 0; i < TSTAT_PROBE_SIZE; i += MMU_PAGESIZE) 677 *((volatile char *)&tstat_probe_area[i]); 678 679 if ((ts = rdtick() - ts) < best) 680 best = ts; 681 buf[j++] = ts; 682 } 683 684 return (best); 685 } 686 687 /* 688 * This routine determines the probe effect by calling trapstat_probe_laps() 689 * both without and with the interposing trap table. Note that this is 690 * called from a cross call on the desired CPU, and that it is called on 691 * every CPU (this is necessary because the probe effect may differ from 692 * one CPU to another). 693 */ 694 static void 695 trapstat_probe() 696 { 697 tstat_percpu_t *tcpu = &tstat_percpu[CPU->cpu_id]; 698 hrtime_t before, after; 699 700 if (!(tcpu->tcpu_flags & TSTAT_CPU_SELECTED)) 701 return; 702 703 if (tstat_probe_area == NULL || (tstat_options & TSTAT_OPT_NOGO)) 704 return; 705 706 /* 707 * We very much expect the %tba to be KERNELBASE; this is a 708 * precautionary measure to assure that trapstat doesn't melt the 709 * machine should the %tba point unexpectedly elsewhere. 710 */ 711 if (get_tba() != (caddr_t)KERNELBASE) 712 return; 713 714 /* 715 * Preserve this CPU's data before destroying it by enabling the 716 * interposing trap table. We can safely use tstat_buffer because 717 * the caller of the trapstat_probe() cross call is holding tstat_lock. 718 */ 719 bcopy(tcpu->tcpu_data, tstat_buffer, tstat_data_t_size); 720 721 tstat_probe_time = gethrtime(); 722 723 before = trapstat_probe_laps(TSTAT_PROBE_NLAPS, tstat_probe_before); 724 (void) set_tba(tcpu->tcpu_ibase); 725 726 after = trapstat_probe_laps(TSTAT_PROBE_NLAPS, tstat_probe_after); 727 (void) set_tba((caddr_t)KERNELBASE); 728 729 tstat_probe_time = gethrtime() - tstat_probe_time; 730 731 bcopy(tstat_buffer, tcpu->tcpu_data, tstat_data_t_size); 732 tcpu->tcpu_data->tdata_peffect = (after - before) / TSTAT_PROBE_NPAGES; 733 } 734 735 static void 736 trapstat_probe_alloc() 737 { 738 pfn_t pfn; 739 caddr_t va; 740 int i; 741 742 ASSERT(MUTEX_HELD(&tstat_lock)); 743 ASSERT(tstat_probe_area == NULL); 744 ASSERT(tstat_probe_phys == NULL); 745 746 if (!(tstat_options & TSTAT_OPT_TLBDATA)) 747 return; 748 749 /* 750 * Grab some virtual from the heap arena. 751 */ 752 tstat_probe_area = vmem_alloc(heap_arena, TSTAT_PROBE_SIZE, VM_SLEEP); 753 va = tstat_probe_area; 754 755 /* 756 * Grab a single physical page. 757 */ 758 tstat_probe_phys = vmem_alloc(tstat_arena, MMU_PAGESIZE, VM_SLEEP); 759 pfn = hat_getpfnum(kas.a_hat, tstat_probe_phys); 760 761 /* 762 * Now set the translation for every page in our virtual range 763 * to be our allocated physical page. 764 */ 765 for (i = 0; i < TSTAT_PROBE_NPAGES; i++) { 766 hat_devload(kas.a_hat, va, MMU_PAGESIZE, pfn, PROT_READ, 767 HAT_LOAD_NOCONSIST | HAT_LOAD_LOCK); 768 va += MMU_PAGESIZE; 769 } 770 } 771 772 static void 773 trapstat_probe_free() 774 { 775 caddr_t va; 776 int i; 777 778 ASSERT(MUTEX_HELD(&tstat_lock)); 779 780 if ((va = tstat_probe_area) == NULL) 781 return; 782 783 for (i = 0; i < TSTAT_PROBE_NPAGES; i++) { 784 hat_unload(kas.a_hat, va, MMU_PAGESIZE, HAT_UNLOAD_UNLOCK); 785 va += MMU_PAGESIZE; 786 } 787 788 vmem_free(tstat_arena, tstat_probe_phys, MMU_PAGESIZE); 789 vmem_free(heap_arena, tstat_probe_area, TSTAT_PROBE_SIZE); 790 791 tstat_probe_phys = NULL; 792 tstat_probe_area = NULL; 793 } 794 795 /* 796 * This routine actually enables a CPU by setting its %tba to be the 797 * CPU's interposing trap table. It is called out of cross call context. 798 */ 799 static void 800 trapstat_enable() 801 { 802 tstat_percpu_t *tcpu = &tstat_percpu[CPU->cpu_id]; 803 804 if (!(tcpu->tcpu_flags & TSTAT_CPU_SELECTED)) 805 return; 806 807 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_ALLOCATED); 808 ASSERT(!(tcpu->tcpu_flags & TSTAT_CPU_ENABLED)); 809 810 if (get_tba() != (caddr_t)KERNELBASE) 811 return; 812 813 if (!(tstat_options & TSTAT_OPT_NOGO)) 814 (void) set_tba(tcpu->tcpu_ibase); 815 tcpu->tcpu_flags |= TSTAT_CPU_ENABLED; 816 #ifdef sun4v 817 if (tstat_options & (TSTAT_OPT_TLBDATA | TSTAT_OPT_NOGO)) { 818 /* 819 * On sun4v platforms, TLB misses are normally handled by the 820 * hypervisor or the hardware -- provided one or more TSBs 821 * have been setup and communicated via hv_set_ctx0 and 822 * hv_set_nonctx0 API. However, as part of collecting TLB 823 * statistics, we disabled this miss processing by telling the 824 * hypervisor that there was not a TSB; we now need to 825 * communicate the proper kernel/user TSB information to 826 * resume efficient operation. 827 * 828 * While we restore kernel TSB information immediately, to 829 * avoid any locking dependency, we don't restore user TSB 830 * information right away. Rather, we simply clear the 831 * TSTAT_TLB_STATS flag so that the user TSB information is 832 * automatically restored on the next context switch. 833 * 834 * Note that the call to restore kernel TSB information is not 835 * expected to fail. Even in the event of failure, the system 836 * will still continue to function properly, if in a state of 837 * reduced performance due to the guest kernel handling all 838 * TLB misses. 839 */ 840 cpu_t *cp = CPU; 841 842 cp->cpu_m.cpu_tstat_flags |= TSTAT_TLB_STATS; 843 (void) hv_set_ctx0(NULL, NULL); 844 (void) hv_set_ctxnon0(NULL, NULL); 845 } 846 #endif 847 } 848 849 /* 850 * This routine disables a CPU (vis a vis trapstat) by setting its %tba to be 851 * the actual, underlying trap table. It is called out of cross call context. 852 */ 853 static void 854 trapstat_disable() 855 { 856 tstat_percpu_t *tcpu = &tstat_percpu[CPU->cpu_id]; 857 858 if (!(tcpu->tcpu_flags & TSTAT_CPU_ENABLED)) 859 return; 860 861 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_SELECTED); 862 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_ALLOCATED); 863 864 if (!(tstat_options & TSTAT_OPT_NOGO)) 865 (void) set_tba((caddr_t)KERNELBASE); 866 867 tcpu->tcpu_flags &= ~TSTAT_CPU_ENABLED; 868 869 #ifdef sun4v 870 if (tstat_options & (TSTAT_OPT_TLBDATA | TSTAT_OPT_NOGO)) { 871 /* 872 * On sun4v platforms, TlB misses are normally handled by 873 * the hypervisor or the hardware provided one or more TSBs 874 * have been setup and communicated via hv_set_ctx0 and 875 * hv_set_nonctx0 API. However, as part of collecting TLB 876 * statistics, we disabled that by faking NO TSB and we 877 * need to communicate proper kernel/user TSB information 878 * so that TLB misses can be handled by the hypervisor or 879 * the hardware more efficiently. 880 * 881 * We restore kernel TSB information right away. However, 882 * to minimize any locking dependency, we don't restore 883 * user TSB information right away. Instead, we simply 884 * clear the TSTAT_TLB_STATS flag so that the user TSB 885 * information is automatically restored on next context 886 * switch. 887 * 888 * Note that the call to restore kernel TSB information 889 * will normally not fail, unless wrong information is 890 * passed here. In that scenario, system will still 891 * continue to function properly with the exception of 892 * kernel handling all the TLB misses. 893 */ 894 struct hv_tsb_block *hvbp = &ksfmmup->sfmmu_hvblock; 895 cpu_t *cp = CPU; 896 897 cp->cpu_m.cpu_tstat_flags &= ~TSTAT_TLB_STATS; 898 (void) hv_set_ctx0(hvbp->hv_tsb_info_cnt, hvbp->hv_tsb_info_pa); 899 } 900 #endif 901 } 902 903 /* 904 * We use %tick as the time base when recording the time spent executing 905 * the trap handler. %tick, however, is not necessarily kept in sync 906 * across CPUs (indeed, different CPUs may have different %tick frequencies). 907 * We therefore cross call onto a CPU to get a snapshot of its data to 908 * copy out; this is the routine executed out of that cross call. 909 */ 910 static void 911 trapstat_snapshot() 912 { 913 tstat_percpu_t *tcpu = &tstat_percpu[CPU->cpu_id]; 914 tstat_data_t *data = tcpu->tcpu_data; 915 916 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_SELECTED); 917 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_ALLOCATED); 918 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_ENABLED); 919 920 data->tdata_snapts = gethrtime(); 921 data->tdata_snaptick = rdtick(); 922 bcopy(data, tstat_buffer, tstat_data_t_size); 923 } 924 925 /* 926 * The TSTAT_RETENT_* constants define offsets in the TLB return entry. 927 * They are used only in trapstat_tlbretent() (below) and #undef'd 928 * immediately afterwards. Any change to "retent" in trapstat_tlbretent() 929 * will likely require changes to these constants. 930 */ 931 932 #ifndef sun4v 933 #define TSTAT_RETENT_STATHI 1 934 #define TSTAT_RETENT_STATLO 2 935 #define TSTAT_RETENT_SHIFT 8 936 #define TSTAT_RETENT_COUNT_LD 10 937 #define TSTAT_RETENT_COUNT_ST 12 938 #define TSTAT_RETENT_TMPTSHI 13 939 #define TSTAT_RETENT_TMPTSLO 14 940 #define TSTAT_RETENT_TIME_LD 16 941 #define TSTAT_RETENT_TIME_ST 18 942 #else /* sun4v */ 943 #define TSTAT_RETENT_STATHI 1 944 #define TSTAT_RETENT_STATLO 2 945 #define TSTAT_RETENT_SHIFT 5 946 #define TSTAT_RETENT_COUNT_LD 7 947 #define TSTAT_RETENT_COUNT_ST 9 948 #define TSTAT_RETENT_TMPTSHI 10 949 #define TSTAT_RETENT_TMPTSLO 11 950 #define TSTAT_RETENT_TIME_LD 13 951 #define TSTAT_RETENT_TIME_ST 15 952 #endif /* sun4v */ 953 954 static void 955 trapstat_tlbretent(tstat_percpu_t *tcpu, tstat_tlbretent_t *ret, 956 tstat_missdata_t *data) 957 { 958 uint32_t *ent = ret->ttlbrent_instr, shift; 959 uintptr_t base, tmptick = TSTAT_DATA_OFFS(tcpu, tdata_tmptick); 960 961 /* 962 * This is the entry executed upon return from the TLB/TSB miss 963 * handler (i.e. the code interpositioned between the "retry" and 964 * the actual return to the TLB-missing instruction). Detail on its 965 * theory of operation can be found in the "TLB Statistics" section 966 * of the block comment. Note that we expect the TTE just loaded 967 * into the TLB to be in %g5; all other globals are available as 968 * scratch. Finally, note that the page size information in sun4v is 969 * located in the lower bits of the TTE -- requiring us to have a 970 * different return entry on sun4v. 971 */ 972 static const uint32_t retent[TSTAT_TLBRET_NINSTR] = { 973 #ifndef sun4v 974 0x87410000, /* rd %tick, %g3 */ 975 0x03000000, /* sethi %hi(stat), %g1 */ 976 0x82106000, /* or %g1, %lo(stat), %g1 */ 977 0x89297001, /* sllx %g5, 1, %g4 */ 978 0x8931303e, /* srlx %g4, 62, %g4 */ 979 0x8531702e, /* srlx %g5, 46, %g2 */ 980 0x8408a004, /* and %g2, 4, %g2 */ 981 0x88110002, /* or %g4, %g2, %g4 */ 982 0x89292000, /* sll %g4, shift, %g4 */ 983 0x82004004, /* add %g1, %g4, %g1 */ 984 0xc4586000, /* ldx [%g1 + tmiss_count], %g2 */ 985 0x8400a001, /* add %g2, 1, %g2 */ 986 0xc4706000, /* stx %g2, [%g1 + tmiss_count] */ 987 0x0d000000, /* sethi %hi(tdata_tmptick), %g6 */ 988 0xc459a000, /* ldx [%g6 + %lo(tdata_tmptick)], %g2 */ 989 0x8620c002, /* sub %g3, %g2, %g3 */ 990 0xc4586000, /* ldx [%g1 + tmiss_time], %g2 */ 991 0x84008003, /* add %g2, %g3, %g2 */ 992 0xc4706000, /* stx %g2, [%g1 + tmiss_time] */ 993 0x83f00000 /* retry */ 994 #else /* sun4v */ 995 0x87410000, /* rd %tick, %g3 */ 996 0x03000000, /* sethi %hi(stat), %g1 */ 997 0x82106000, /* or %g1, %lo(stat), %g1 */ 998 0x8929703d, /* sllx %g5, 61, %g4 */ 999 0x8931303d, /* srlx %g4, 61, %g4 */ 1000 0x89292000, /* sll %g4, shift, %g4 */ 1001 0x82004004, /* add %g1, %g4, %g1 */ 1002 0xc4586000, /* ldx [%g1 + tmiss_count], %g2 */ 1003 0x8400a001, /* add %g2, 1, %g2 */ 1004 0xc4706000, /* stx %g2, [%g1 + tmiss_count] */ 1005 0x0d000000, /* sethi %hi(tdata_tmptick), %g6 */ 1006 0xc459a000, /* ldx [%g6 + %lo(tdata_tmptick)], %g2 */ 1007 0x8620c002, /* sub %g3, %g2, %g3 */ 1008 0xc4586000, /* ldx [%g1 + tmiss_time], %g2 */ 1009 0x84008003, /* add %g2, %g3, %g2 */ 1010 0xc4706000, /* stx %g2, [%g1 + tmiss_time] */ 1011 0x83f00000 /* retry */ 1012 #endif /* sun4v */ 1013 }; 1014 1015 ASSERT(MUTEX_HELD(&tstat_lock)); 1016 /*CONSTCOND*/ 1017 ASSERT(offsetof(tstat_missdata_t, tmiss_count) <= LO10(-1)); 1018 /*CONSTCOND*/ 1019 ASSERT(offsetof(tstat_missdata_t, tmiss_time) <= LO10(-1)); 1020 /*CONSTCOND*/ 1021 ASSERT(!((sizeof (tstat_pgszdata_t) - 1) & sizeof (tstat_pgszdata_t))); 1022 1023 for (shift = 1; (1 << shift) != sizeof (tstat_pgszdata_t); shift++) 1024 continue; 1025 1026 base = (uintptr_t)tcpu->tcpu_dbase + 1027 ((uintptr_t)data - (uintptr_t)tcpu->tcpu_data); 1028 1029 bcopy(retent, ent, sizeof (retent)); 1030 1031 ent[TSTAT_RETENT_STATHI] |= HI22(base); 1032 ent[TSTAT_RETENT_STATLO] |= LO10(base); 1033 ent[TSTAT_RETENT_SHIFT] |= shift; 1034 /* LINTED E_EXPR_NULL_EFFECT */ 1035 ent[TSTAT_RETENT_COUNT_LD] |= offsetof(tstat_missdata_t, tmiss_count); 1036 /* LINTED E_EXPR_NULL_EFFECT */ 1037 ent[TSTAT_RETENT_COUNT_ST] |= offsetof(tstat_missdata_t, tmiss_count); 1038 ent[TSTAT_RETENT_TMPTSHI] |= HI22(tmptick); 1039 ent[TSTAT_RETENT_TMPTSLO] |= LO10(tmptick); 1040 ent[TSTAT_RETENT_TIME_LD] |= offsetof(tstat_missdata_t, tmiss_time); 1041 ent[TSTAT_RETENT_TIME_ST] |= offsetof(tstat_missdata_t, tmiss_time); 1042 } 1043 1044 #undef TSTAT_RETENT_STATHI 1045 #undef TSTAT_RETENT_STATLO 1046 #undef TSTAT_RETENT_SHIFT 1047 #undef TSTAT_RETENT_COUNT_LD 1048 #undef TSTAT_RETENT_COUNT_ST 1049 #undef TSTAT_RETENT_TMPTSHI 1050 #undef TSTAT_RETENT_TMPTSLO 1051 #undef TSTAT_RETENT_TIME_LD 1052 #undef TSTAT_RETENT_TIME_ST 1053 1054 /* 1055 * The TSTAT_TLBENT_* constants define offsets in the TLB entry. They are 1056 * used only in trapstat_tlbent() (below) and #undef'd immediately afterwards. 1057 * Any change to "tlbent" in trapstat_tlbent() will likely require changes 1058 * to these constants. 1059 */ 1060 1061 #ifndef sun4v 1062 #define TSTAT_TLBENT_STATHI 0 1063 #define TSTAT_TLBENT_STATLO_LD 1 1064 #define TSTAT_TLBENT_STATLO_ST 3 1065 #define TSTAT_TLBENT_MMUASI 15 1066 #define TSTAT_TLBENT_TPCHI 18 1067 #define TSTAT_TLBENT_TPCLO_USER 19 1068 #define TSTAT_TLBENT_TPCLO_KERN 21 1069 #define TSTAT_TLBENT_TSHI 25 1070 #define TSTAT_TLBENT_TSLO 27 1071 #define TSTAT_TLBENT_BA 28 1072 #else /* sun4v */ 1073 #define TSTAT_TLBENT_STATHI 0 1074 #define TSTAT_TLBENT_STATLO_LD 1 1075 #define TSTAT_TLBENT_STATLO_ST 3 1076 #define TSTAT_TLBENT_TAGTARGET 19 1077 #define TSTAT_TLBENT_TPCHI 21 1078 #define TSTAT_TLBENT_TPCLO_USER 22 1079 #define TSTAT_TLBENT_TPCLO_KERN 24 1080 #define TSTAT_TLBENT_TSHI 28 1081 #define TSTAT_TLBENT_TSLO 30 1082 #define TSTAT_TLBENT_BA 31 1083 #endif /* sun4v */ 1084 1085 static void 1086 trapstat_tlbent(tstat_percpu_t *tcpu, int entno) 1087 { 1088 uint32_t *ent; 1089 uintptr_t orig, va, baoffs; 1090 int itlb = entno == TSTAT_ENT_ITLBMISS; 1091 int entoffs = entno << TSTAT_ENT_SHIFT; 1092 uintptr_t tmptick, stat, tpc, utpc; 1093 tstat_pgszdata_t *data = &tcpu->tcpu_data->tdata_pgsz[0]; 1094 tstat_tlbdata_t *udata, *kdata; 1095 tstat_tlbret_t *ret; 1096 #ifndef sun4v 1097 uint32_t asi = itlb ? ASI(ASI_IMMU) : ASI(ASI_DMMU); 1098 #else 1099 uint32_t tagtarget_off = itlb ? MMFSA_I_CTX : MMFSA_D_CTX; 1100 #endif 1101 1102 /* 1103 * When trapstat is run with TLB statistics, this is the entry for 1104 * both I- and D-TLB misses; this code performs trap level pushing, 1105 * as described in the "TLB Statistics" section of the block comment. 1106 * This code is executing at TL 1; %tstate[0] contains the saved 1107 * state at the time of the TLB miss. Pushing trap level 1 (and thus 1108 * raising TL to 2) requires us to fill in %tstate[1] with our %pstate, 1109 * %cwp and %asi. We leave %tt unchanged, and we set %tpc and %tnpc to 1110 * the appropriate TLB return entry (based on the context of the miss). 1111 * Finally, we sample %tick, and stash it in the tdata_tmptick member 1112 * the per-CPU tstat_data structure. tdata_tmptick will be used in 1113 * the TLB return entry to determine the amount of time spent in the 1114 * TLB miss handler. 1115 * 1116 * Note that on sun4v platforms, we must obtain the context information 1117 * from the MMU fault status area. (The base address of this MMU fault 1118 * status area is kept in the scratchpad register 0.) 1119 */ 1120 static const uint32_t tlbent[] = { 1121 #ifndef sun4v 1122 0x03000000, /* sethi %hi(stat), %g1 */ 1123 0xc4586000, /* ldx [%g1 + %lo(stat)], %g2 */ 1124 0x8400a001, /* add %g2, 1, %g2 */ 1125 0xc4706000, /* stx %g2, [%g1 + %lo(stat)] */ 1126 0x85524000, /* rdpr %cwp, %g2 */ 1127 0x87518000, /* rdpr %pstate, %g3 */ 1128 0x8728f008, /* sllx %g3, 8, %g3 */ 1129 0x84108003, /* or %g2, %g3, %g2 */ 1130 0x8740c000, /* rd %asi, %g3 */ 1131 0x8728f018, /* sllx %g3, 24, %g3 */ 1132 0x84108003, /* or %g2, %g3, %g2 */ 1133 0x8350c000, /* rdpr %tt, %g1 */ 1134 0x8f902002, /* wrpr %g0, 2, %tl */ 1135 0x85908000, /* wrpr %g2, %g0, %tstate */ 1136 0x87904000, /* wrpr %g1, %g0, %tt */ 1137 0xc2d80000, /* ldxa [%g0]ASI_MMU, %g1 */ 1138 0x83307030, /* srlx %g1, CTXSHIFT, %g1 */ 1139 0x02c04004, /* brz,pn %g1, .+0x10 */ 1140 0x03000000, /* sethi %hi(new_tpc), %g1 */ 1141 0x82106000, /* or %g1, %lo(new_tpc), %g1 */ 1142 0x30800002, /* ba,a .+0x8 */ 1143 0x82106000, /* or %g1, %lo(new_tpc), %g1 */ 1144 0x81904000, /* wrpr %g1, %g0, %tpc */ 1145 0x82006004, /* add %g1, 4, %g1 */ 1146 0x83904000, /* wrpr %g1, %g0, %tnpc */ 1147 0x03000000, /* sethi %hi(tmptick), %g1 */ 1148 0x85410000, /* rd %tick, %g2 */ 1149 0xc4706000, /* stx %g2, [%g1 + %lo(tmptick)] */ 1150 0x30800000, /* ba,a addr */ 1151 NOP, NOP, NOP 1152 #else /* sun4v */ 1153 0x03000000, /* sethi %hi(stat), %g1 */ 1154 0xc4586000, /* ldx [%g1 + %lo(stat)], %g2 */ 1155 0x8400a001, /* add %g2, 1, %g2 */ 1156 0xc4706000, /* stx %g2, [%g1 + %lo(stat)] */ 1157 0x85524000, /* rdpr %cwp, %g2 */ 1158 0x87518000, /* rdpr %pstate, %g3 */ 1159 0x8728f008, /* sllx %g3, 8, %g3 */ 1160 0x84108003, /* or %g2, %g3, %g2 */ 1161 0x8740c000, /* rd %asi, %g3 */ 1162 0x8728f018, /* sllx %g3, 24, %g3 */ 1163 0x83540000, /* rdpr %gl, %g1 */ 1164 0x83287028, /* sllx %g1, 40, %g1 */ 1165 0x86104003, /* or %g1, %g3, %g3 */ 1166 0x84108003, /* or %g2, %g3, %g2 */ 1167 0x8350c000, /* rdpr %tt, %g1 */ 1168 0x8f902002, /* wrpr %g0, 2, %tl */ 1169 0x85908000, /* wrpr %g2, %g0, %tstate */ 1170 0x87904000, /* wrpr %g1, %g0, %tt */ 1171 0xc2d80400, /* ldxa [%g0]ASI_SCRATCHPAD, %g1 */ 1172 0xc2586000, /* ldx [%g1 + MMFSA_?_CTX], %g1 */ 1173 0x02c04004, /* brz,pn %g1, .+0x10 */ 1174 0x03000000, /* sethi %hi(new_tpc), %g1 */ 1175 0x82106000, /* or %g1, %lo(new_tpc), %g1 */ 1176 0x30800002, /* ba,a .+0x8 */ 1177 0x82106000, /* or %g1, %lo(new_tpc), %g1 */ 1178 0x81904000, /* wrpr %g1, %g0, %tpc */ 1179 0x82006004, /* add %g1, 4, %g1 */ 1180 0x83904000, /* wrpr %g1, %g0, %tnpc */ 1181 0x03000000, /* sethi %hi(tmptick), %g1 */ 1182 0x85410000, /* rd %tick, %g2 */ 1183 0xc4706000, /* stx %g2, [%g1 + %lo(tmptick)] */ 1184 0x30800000 /* ba,a addr */ 1185 #endif /* sun4v */ 1186 }; 1187 1188 ASSERT(MUTEX_HELD(&tstat_lock)); 1189 ASSERT(entno == TSTAT_ENT_ITLBMISS || entno == TSTAT_ENT_DTLBMISS); 1190 1191 stat = TSTAT_DATA_OFFS(tcpu, tdata_traps) + entoffs; 1192 tmptick = TSTAT_DATA_OFFS(tcpu, tdata_tmptick); 1193 1194 if (itlb) { 1195 ret = &tcpu->tcpu_instr->tinst_itlbret; 1196 udata = &data->tpgsz_user.tmode_itlb; 1197 kdata = &data->tpgsz_kernel.tmode_itlb; 1198 tpc = TSTAT_INSTR_OFFS(tcpu, tinst_itlbret.ttlbr_ktlb); 1199 } else { 1200 ret = &tcpu->tcpu_instr->tinst_dtlbret; 1201 udata = &data->tpgsz_user.tmode_dtlb; 1202 kdata = &data->tpgsz_kernel.tmode_dtlb; 1203 tpc = TSTAT_INSTR_OFFS(tcpu, tinst_dtlbret.ttlbr_ktlb); 1204 } 1205 1206 utpc = tpc + offsetof(tstat_tlbret_t, ttlbr_utlb) - 1207 offsetof(tstat_tlbret_t, ttlbr_ktlb); 1208 1209 ASSERT(HI22(tpc) == HI22(utpc)); 1210 1211 ent = (uint32_t *)((uintptr_t)tcpu->tcpu_instr + entoffs); 1212 orig = KERNELBASE + entoffs; 1213 va = (uintptr_t)tcpu->tcpu_ibase + entoffs; 1214 baoffs = TSTAT_TLBENT_BA * sizeof (uint32_t); 1215 1216 bcopy(tlbent, ent, sizeof (tlbent)); 1217 1218 ent[TSTAT_TLBENT_STATHI] |= HI22(stat); 1219 ent[TSTAT_TLBENT_STATLO_LD] |= LO10(stat); 1220 ent[TSTAT_TLBENT_STATLO_ST] |= LO10(stat); 1221 #ifndef sun4v 1222 ent[TSTAT_TLBENT_MMUASI] |= asi; 1223 #else 1224 ent[TSTAT_TLBENT_TAGTARGET] |= tagtarget_off; 1225 #endif 1226 ent[TSTAT_TLBENT_TPCHI] |= HI22(tpc); 1227 ent[TSTAT_TLBENT_TPCLO_USER] |= LO10(utpc); 1228 ent[TSTAT_TLBENT_TPCLO_KERN] |= LO10(tpc); 1229 ent[TSTAT_TLBENT_TSHI] |= HI22(tmptick); 1230 ent[TSTAT_TLBENT_TSLO] |= LO10(tmptick); 1231 ent[TSTAT_TLBENT_BA] |= DISP22(va + baoffs, orig); 1232 1233 /* 1234 * And now set up the TLB return entries. 1235 */ 1236 trapstat_tlbretent(tcpu, &ret->ttlbr_ktlb, &kdata->ttlb_tlb); 1237 trapstat_tlbretent(tcpu, &ret->ttlbr_ktsb, &kdata->ttlb_tsb); 1238 trapstat_tlbretent(tcpu, &ret->ttlbr_utlb, &udata->ttlb_tlb); 1239 trapstat_tlbretent(tcpu, &ret->ttlbr_utsb, &udata->ttlb_tsb); 1240 } 1241 1242 #undef TSTAT_TLBENT_STATHI 1243 #undef TSTAT_TLBENT_STATLO_LD 1244 #undef TSTAT_TLBENT_STATLO_ST 1245 #ifndef sun4v 1246 #undef TSTAT_TLBENT_MMUASI 1247 #else 1248 #undef TSTAT_TLBENT_TAGTARGET 1249 #endif 1250 #undef TSTAT_TLBENT_TPCHI 1251 #undef TSTAT_TLBENT_TPCLO_USER 1252 #undef TSTAT_TLBENT_TPCLO_KERN 1253 #undef TSTAT_TLBENT_TSHI 1254 #undef TSTAT_TLBENT_TSLO 1255 #undef TSTAT_TLBENT_BA 1256 1257 /* 1258 * The TSTAT_ENABLED_* constants define offsets in the enabled entry; the 1259 * TSTAT_DISABLED_BA constant defines an offset in the disabled entry. Both 1260 * sets of constants are used only in trapstat_make_traptab() (below) and 1261 * #undef'd immediately afterwards. Any change to "enabled" or "disabled" 1262 * in trapstat_make_traptab() will likely require changes to these constants. 1263 */ 1264 #define TSTAT_ENABLED_STATHI 0 1265 #define TSTAT_ENABLED_STATLO_LD 1 1266 #define TSTAT_ENABLED_STATLO_ST 3 1267 #define TSTAT_ENABLED_BA 4 1268 #define TSTAT_DISABLED_BA 0 1269 1270 static void 1271 trapstat_make_traptab(tstat_percpu_t *tcpu) 1272 { 1273 uint32_t *ent; 1274 uint64_t *stat; 1275 uintptr_t orig, va, en_baoffs, dis_baoffs; 1276 int nent; 1277 1278 /* 1279 * This is the entry in the interposing trap table for enabled trap 1280 * table entries. It loads a counter, increments it and stores it 1281 * back before branching to the actual trap table entry. 1282 */ 1283 static const uint32_t enabled[TSTAT_ENT_NINSTR] = { 1284 0x03000000, /* sethi %hi(stat), %g1 */ 1285 0xc4586000, /* ldx [%g1 + %lo(stat)], %g2 */ 1286 0x8400a001, /* add %g2, 1, %g2 */ 1287 0xc4706000, /* stx %g2, [%g1 + %lo(stat)] */ 1288 0x30800000, /* ba,a addr */ 1289 NOP, NOP, NOP 1290 }; 1291 1292 /* 1293 * This is the entry in the interposing trap table for disabled trap 1294 * table entries. It simply branches to the actual, underlying trap 1295 * table entry. As explained in the "Implementation Details" section 1296 * of the block comment, all TL>0 traps _must_ use the disabled entry; 1297 * additional entries may be explicitly disabled through the use 1298 * of TSTATIOC_ENTRY/TSTATIOC_NOENTRY. 1299 */ 1300 static const uint32_t disabled[TSTAT_ENT_NINSTR] = { 1301 0x30800000, /* ba,a addr */ 1302 NOP, NOP, NOP, NOP, NOP, NOP, NOP, 1303 }; 1304 1305 ASSERT(MUTEX_HELD(&tstat_lock)); 1306 1307 ent = tcpu->tcpu_instr->tinst_traptab; 1308 stat = (uint64_t *)TSTAT_DATA_OFFS(tcpu, tdata_traps); 1309 orig = KERNELBASE; 1310 va = (uintptr_t)tcpu->tcpu_ibase; 1311 en_baoffs = TSTAT_ENABLED_BA * sizeof (uint32_t); 1312 dis_baoffs = TSTAT_DISABLED_BA * sizeof (uint32_t); 1313 1314 for (nent = 0; nent < TSTAT_TOTAL_NENT; nent++) { 1315 if (tstat_enabled[nent]) { 1316 bcopy(enabled, ent, sizeof (enabled)); 1317 ent[TSTAT_ENABLED_STATHI] |= HI22(stat); 1318 ent[TSTAT_ENABLED_STATLO_LD] |= LO10(stat); 1319 ent[TSTAT_ENABLED_STATLO_ST] |= LO10(stat); 1320 ent[TSTAT_ENABLED_BA] |= DISP22(va + en_baoffs, orig); 1321 } else { 1322 bcopy(disabled, ent, sizeof (disabled)); 1323 ent[TSTAT_DISABLED_BA] |= DISP22(va + dis_baoffs, orig); 1324 } 1325 1326 stat++; 1327 orig += sizeof (enabled); 1328 ent += sizeof (enabled) / sizeof (*ent); 1329 va += sizeof (enabled); 1330 } 1331 } 1332 1333 #undef TSTAT_ENABLED_STATHI 1334 #undef TSTAT_ENABLED_STATLO_LD 1335 #undef TSTAT_ENABLED_STATLO_ST 1336 #undef TSTAT_ENABLED_BA 1337 #undef TSTAT_DISABLED_BA 1338 1339 static void 1340 trapstat_setup(processorid_t cpu) 1341 { 1342 tstat_percpu_t *tcpu = &tstat_percpu[cpu]; 1343 #ifndef sun4v 1344 int i; 1345 caddr_t va; 1346 pfn_t *pfn; 1347 #endif 1348 1349 ASSERT(tcpu->tcpu_pfn == NULL); 1350 ASSERT(tcpu->tcpu_instr == NULL); 1351 ASSERT(tcpu->tcpu_data == NULL); 1352 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_SELECTED); 1353 ASSERT(!(tcpu->tcpu_flags & TSTAT_CPU_ALLOCATED)); 1354 ASSERT(MUTEX_HELD(&cpu_lock)); 1355 ASSERT(MUTEX_HELD(&tstat_lock)); 1356 1357 /* 1358 * The lower fifteen bits of the %tba are always read as zero; we must 1359 * align our instruction base address appropriately. 1360 */ 1361 #ifndef sun4v 1362 tcpu->tcpu_ibase = (caddr_t)((KERNELBASE - tstat_total_size) 1363 & TSTAT_TBA_MASK); 1364 tcpu->tcpu_dbase = tcpu->tcpu_ibase + TSTAT_INSTR_SIZE; 1365 tcpu->tcpu_vabase = tcpu->tcpu_ibase; 1366 1367 tcpu->tcpu_pfn = vmem_alloc(tstat_arena, tstat_total_pages, VM_SLEEP); 1368 bzero(tcpu->tcpu_pfn, tstat_total_pages); 1369 pfn = tcpu->tcpu_pfn; 1370 1371 tcpu->tcpu_instr = vmem_alloc(tstat_arena, TSTAT_INSTR_SIZE, VM_SLEEP); 1372 1373 va = (caddr_t)tcpu->tcpu_instr; 1374 for (i = 0; i < TSTAT_INSTR_PAGES; i++, va += MMU_PAGESIZE) 1375 *pfn++ = hat_getpfnum(kas.a_hat, va); 1376 1377 /* 1378 * We must be sure that the pages that we will use to examine the data 1379 * have the same virtual color as the pages to which the data is being 1380 * recorded, hence the alignment and phase constraints on the 1381 * allocation. 1382 */ 1383 tcpu->tcpu_data = vmem_xalloc(tstat_arena, tstat_data_size, 1384 shm_alignment, (uintptr_t)tcpu->tcpu_dbase & (shm_alignment - 1), 1385 0, 0, NULL, VM_SLEEP); 1386 bzero(tcpu->tcpu_data, tstat_data_size); 1387 tcpu->tcpu_data->tdata_cpuid = cpu; 1388 1389 va = (caddr_t)tcpu->tcpu_data; 1390 for (i = 0; i < tstat_data_pages; i++, va += MMU_PAGESIZE) 1391 *pfn++ = hat_getpfnum(kas.a_hat, va); 1392 #else /* sun4v */ 1393 ASSERT(!(tstat_total_size > (1 + ~TSTAT_TBA_MASK))); 1394 tcpu->tcpu_vabase = (caddr_t)(KERNELBASE - MMU_PAGESIZE4M); 1395 tcpu->tcpu_ibase = tcpu->tcpu_vabase + (cpu * (1 + ~TSTAT_TBA_MASK)); 1396 tcpu->tcpu_dbase = tcpu->tcpu_ibase + TSTAT_INSTR_SIZE; 1397 1398 tcpu->tcpu_pfn = &tstat_pfn; 1399 tcpu->tcpu_instr = (tstat_instr_t *)(tstat_va + (cpu * 1400 (1 + ~TSTAT_TBA_MASK))); 1401 tcpu->tcpu_data = (tstat_data_t *)(tstat_va + (cpu * 1402 (1 + ~TSTAT_TBA_MASK)) + TSTAT_INSTR_SIZE); 1403 bzero(tcpu->tcpu_data, tstat_data_size); 1404 tcpu->tcpu_data->tdata_cpuid = cpu; 1405 #endif /* sun4v */ 1406 1407 /* 1408 * Now that we have all of the instruction and data pages allocated, 1409 * make the trap table from scratch. 1410 */ 1411 trapstat_make_traptab(tcpu); 1412 1413 if (tstat_options & TSTAT_OPT_TLBDATA) { 1414 /* 1415 * TLB Statistics have been specified; set up the I- and D-TLB 1416 * entries and corresponding TLB return entries. 1417 */ 1418 trapstat_tlbent(tcpu, TSTAT_ENT_ITLBMISS); 1419 trapstat_tlbent(tcpu, TSTAT_ENT_DTLBMISS); 1420 } 1421 1422 tcpu->tcpu_flags |= TSTAT_CPU_ALLOCATED; 1423 1424 /* 1425 * Finally, get the target CPU to load the locked pages into its TLBs. 1426 */ 1427 xc_one(cpu, (xcfunc_t *)trapstat_load_tlb, 0, 0); 1428 } 1429 1430 static void 1431 trapstat_teardown(processorid_t cpu) 1432 { 1433 tstat_percpu_t *tcpu = &tstat_percpu[cpu]; 1434 #ifndef sun4v 1435 int i; 1436 #endif 1437 caddr_t va = tcpu->tcpu_vabase; 1438 1439 ASSERT(tcpu->tcpu_pfn != NULL); 1440 ASSERT(tcpu->tcpu_instr != NULL); 1441 ASSERT(tcpu->tcpu_data != NULL); 1442 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_SELECTED); 1443 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_ALLOCATED); 1444 ASSERT(!(tcpu->tcpu_flags & TSTAT_CPU_ENABLED)); 1445 ASSERT(MUTEX_HELD(&cpu_lock)); 1446 ASSERT(MUTEX_HELD(&tstat_lock)); 1447 1448 #ifndef sun4v 1449 vmem_free(tstat_arena, tcpu->tcpu_pfn, tstat_total_pages); 1450 vmem_free(tstat_arena, tcpu->tcpu_instr, TSTAT_INSTR_SIZE); 1451 vmem_free(tstat_arena, tcpu->tcpu_data, tstat_data_size); 1452 1453 for (i = 0; i < tstat_total_pages; i++, va += MMU_PAGESIZE) { 1454 xt_one(cpu, vtag_flushpage_tl1, (uint64_t)va, KCONTEXT); 1455 } 1456 #else 1457 xt_one(cpu, vtag_unmap_perm_tl1, (uint64_t)va, KCONTEXT); 1458 #endif 1459 1460 tcpu->tcpu_pfn = NULL; 1461 tcpu->tcpu_instr = NULL; 1462 tcpu->tcpu_data = NULL; 1463 tcpu->tcpu_flags &= ~TSTAT_CPU_ALLOCATED; 1464 } 1465 1466 static int 1467 trapstat_go() 1468 { 1469 cpu_t *cp; 1470 1471 mutex_enter(&cpu_lock); 1472 mutex_enter(&tstat_lock); 1473 1474 if (tstat_running) { 1475 mutex_exit(&tstat_lock); 1476 mutex_exit(&cpu_lock); 1477 return (EBUSY); 1478 } 1479 1480 #ifdef sun4v 1481 /* 1482 * Allocate large page to hold interposing tables 1483 */ 1484 tstat_va = contig_mem_alloc(MMU_PAGESIZE4M); 1485 tstat_pfn = va_to_pfn(tstat_va); 1486 if (tstat_pfn == PFN_INVALID) { 1487 contig_mem_free(tstat_va, MMU_PAGESIZE4M); 1488 return (EAGAIN); 1489 } 1490 #endif 1491 1492 /* 1493 * First, perform any necessary hot patching. 1494 */ 1495 trapstat_hotpatch(); 1496 1497 /* 1498 * Allocate the resources we'll need to measure probe effect. 1499 */ 1500 trapstat_probe_alloc(); 1501 1502 1503 cp = cpu_list; 1504 do { 1505 if (!(tstat_percpu[cp->cpu_id].tcpu_flags & TSTAT_CPU_SELECTED)) 1506 continue; 1507 1508 trapstat_setup(cp->cpu_id); 1509 1510 /* 1511 * Note that due to trapstat_probe()'s use of global data, 1512 * we determine the probe effect on each CPU serially instead 1513 * of in parallel with an xc_all(). 1514 */ 1515 xc_one(cp->cpu_id, (xcfunc_t *)trapstat_probe, 0, 0); 1516 } while ((cp = cp->cpu_next) != cpu_list); 1517 1518 xc_all((xcfunc_t *)trapstat_enable, 0, 0); 1519 1520 trapstat_probe_free(); 1521 tstat_running = 1; 1522 mutex_exit(&tstat_lock); 1523 mutex_exit(&cpu_lock); 1524 1525 return (0); 1526 } 1527 1528 static int 1529 trapstat_stop() 1530 { 1531 int i; 1532 1533 mutex_enter(&cpu_lock); 1534 mutex_enter(&tstat_lock); 1535 if (!tstat_running) { 1536 mutex_exit(&tstat_lock); 1537 mutex_exit(&cpu_lock); 1538 return (ENXIO); 1539 } 1540 1541 xc_all((xcfunc_t *)trapstat_disable, 0, 0); 1542 1543 for (i = 0; i <= max_cpuid; i++) { 1544 if (tstat_percpu[i].tcpu_flags & TSTAT_CPU_ALLOCATED) 1545 trapstat_teardown(i); 1546 } 1547 1548 #ifdef sun4v 1549 contig_mem_free(tstat_va, MMU_PAGESIZE4M); 1550 #endif 1551 trapstat_hotpatch(); 1552 tstat_running = 0; 1553 mutex_exit(&tstat_lock); 1554 mutex_exit(&cpu_lock); 1555 1556 return (0); 1557 } 1558 1559 /* 1560 * This is trapstat's DR CPU configuration callback. It's called (with 1561 * cpu_lock held) to unconfigure a newly powered-off CPU, or to configure a 1562 * powered-off CPU that is to be brought into the system. We need only take 1563 * action in the unconfigure case: because a powered-off CPU will have its 1564 * trap table restored to KERNELBASE if it is ever powered back on, we must 1565 * update the flags to reflect that trapstat is no longer enabled on the 1566 * powered-off CPU. Note that this means that a TSTAT_CPU_ENABLED CPU that 1567 * is unconfigured/powered off and later powered back on/reconfigured will 1568 * _not_ be re-TSTAT_CPU_ENABLED. 1569 */ 1570 static int 1571 trapstat_cpu_setup(cpu_setup_t what, processorid_t cpu) 1572 { 1573 tstat_percpu_t *tcpu = &tstat_percpu[cpu]; 1574 1575 ASSERT(MUTEX_HELD(&cpu_lock)); 1576 mutex_enter(&tstat_lock); 1577 1578 if (!tstat_running) { 1579 mutex_exit(&tstat_lock); 1580 return (0); 1581 } 1582 1583 switch (what) { 1584 case CPU_CONFIG: 1585 ASSERT(!(tcpu->tcpu_flags & TSTAT_CPU_ENABLED)); 1586 break; 1587 1588 case CPU_UNCONFIG: 1589 if (tcpu->tcpu_flags & TSTAT_CPU_ENABLED) 1590 tcpu->tcpu_flags &= ~TSTAT_CPU_ENABLED; 1591 break; 1592 1593 default: 1594 break; 1595 } 1596 1597 mutex_exit(&tstat_lock); 1598 return (0); 1599 } 1600 1601 /* 1602 * This is called before a CPR suspend and after a CPR resume. We don't have 1603 * anything to do before a suspend, but after a restart we must restore the 1604 * trap table to be our interposing trap table. However, we don't actually 1605 * know whether or not the CPUs have been powered off -- this routine may be 1606 * called while restoring from a failed CPR suspend. We thus run through each 1607 * TSTAT_CPU_ENABLED CPU, and explicitly destroy and reestablish its 1608 * interposing trap table. This assures that our state is correct regardless 1609 * of whether or not the CPU has been newly powered on. 1610 */ 1611 /*ARGSUSED*/ 1612 static boolean_t 1613 trapstat_cpr(void *arg, int code) 1614 { 1615 cpu_t *cp; 1616 1617 if (code == CB_CODE_CPR_CHKPT) 1618 return (B_TRUE); 1619 1620 ASSERT(code == CB_CODE_CPR_RESUME); 1621 1622 mutex_enter(&cpu_lock); 1623 mutex_enter(&tstat_lock); 1624 1625 if (!tstat_running) { 1626 mutex_exit(&tstat_lock); 1627 mutex_exit(&cpu_lock); 1628 return (B_TRUE); 1629 } 1630 1631 cp = cpu_list; 1632 do { 1633 tstat_percpu_t *tcpu = &tstat_percpu[cp->cpu_id]; 1634 1635 if (!(tcpu->tcpu_flags & TSTAT_CPU_ENABLED)) 1636 continue; 1637 1638 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_SELECTED); 1639 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_ALLOCATED); 1640 1641 xc_one(cp->cpu_id, (xcfunc_t *)trapstat_disable, 0, 0); 1642 ASSERT(!(tcpu->tcpu_flags & TSTAT_CPU_ENABLED)); 1643 1644 /* 1645 * Preserve this CPU's data in tstat_buffer and rip down its 1646 * interposing trap table. 1647 */ 1648 bcopy(tcpu->tcpu_data, tstat_buffer, tstat_data_t_size); 1649 trapstat_teardown(cp->cpu_id); 1650 ASSERT(!(tcpu->tcpu_flags & TSTAT_CPU_ALLOCATED)); 1651 1652 /* 1653 * Reestablish the interposing trap table and restore the old 1654 * data. 1655 */ 1656 trapstat_setup(cp->cpu_id); 1657 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_ALLOCATED); 1658 bcopy(tstat_buffer, tcpu->tcpu_data, tstat_data_t_size); 1659 1660 xc_one(cp->cpu_id, (xcfunc_t *)trapstat_enable, 0, 0); 1661 } while ((cp = cp->cpu_next) != cpu_list); 1662 1663 mutex_exit(&tstat_lock); 1664 mutex_exit(&cpu_lock); 1665 1666 return (B_TRUE); 1667 } 1668 1669 /*ARGSUSED*/ 1670 static int 1671 trapstat_open(dev_t *devp, int flag, int otyp, cred_t *cred_p) 1672 { 1673 int i; 1674 1675 mutex_enter(&cpu_lock); 1676 mutex_enter(&tstat_lock); 1677 if (tstat_open != 0) { 1678 mutex_exit(&tstat_lock); 1679 mutex_exit(&cpu_lock); 1680 return (EBUSY); 1681 } 1682 1683 /* 1684 * Register this in open() rather than in attach() to prevent deadlock 1685 * with DR code. During attach, I/O device tree locks are grabbed 1686 * before trapstat_attach() is invoked - registering in attach 1687 * will result in the lock order: device tree lock, cpu_lock. 1688 * DR code however requires that cpu_lock be acquired before 1689 * device tree locks. 1690 */ 1691 ASSERT(!tstat_running); 1692 register_cpu_setup_func((cpu_setup_func_t *)trapstat_cpu_setup, NULL); 1693 1694 /* 1695 * Clear all options. And until specific CPUs are specified, we'll 1696 * mark all CPUs as selected. 1697 */ 1698 tstat_options = 0; 1699 1700 for (i = 0; i <= max_cpuid; i++) 1701 tstat_percpu[i].tcpu_flags |= TSTAT_CPU_SELECTED; 1702 1703 /* 1704 * By default, all traps at TL=0 are enabled. Traps at TL>0 must 1705 * be disabled. 1706 */ 1707 for (i = 0; i < TSTAT_TOTAL_NENT; i++) 1708 tstat_enabled[i] = i < TSTAT_NENT ? 1 : 0; 1709 1710 tstat_open = 1; 1711 mutex_exit(&tstat_lock); 1712 mutex_exit(&cpu_lock); 1713 1714 return (0); 1715 } 1716 1717 /*ARGSUSED*/ 1718 static int 1719 trapstat_close(dev_t dev, int flag, int otyp, cred_t *cred_p) 1720 { 1721 (void) trapstat_stop(); 1722 1723 ASSERT(!tstat_running); 1724 1725 mutex_enter(&cpu_lock); 1726 unregister_cpu_setup_func((cpu_setup_func_t *)trapstat_cpu_setup, NULL); 1727 mutex_exit(&cpu_lock); 1728 1729 tstat_open = 0; 1730 return (DDI_SUCCESS); 1731 } 1732 1733 static int 1734 trapstat_option(int option) 1735 { 1736 mutex_enter(&tstat_lock); 1737 1738 if (tstat_running) { 1739 mutex_exit(&tstat_lock); 1740 return (EBUSY); 1741 } 1742 1743 tstat_options |= option; 1744 mutex_exit(&tstat_lock); 1745 1746 return (0); 1747 } 1748 1749 /*ARGSUSED*/ 1750 static int 1751 trapstat_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *crd, int *rval) 1752 { 1753 int i, j, out; 1754 size_t dsize; 1755 1756 switch (cmd) { 1757 case TSTATIOC_GO: 1758 return (trapstat_go()); 1759 1760 case TSTATIOC_NOGO: 1761 return (trapstat_option(TSTAT_OPT_NOGO)); 1762 1763 case TSTATIOC_STOP: 1764 return (trapstat_stop()); 1765 1766 case TSTATIOC_CPU: 1767 if (arg < 0 || arg > max_cpuid) 1768 return (EINVAL); 1769 /*FALLTHROUGH*/ 1770 1771 case TSTATIOC_NOCPU: 1772 mutex_enter(&tstat_lock); 1773 1774 if (tstat_running) { 1775 mutex_exit(&tstat_lock); 1776 return (EBUSY); 1777 } 1778 1779 /* 1780 * If this is the first CPU to be specified (or if we are 1781 * being asked to explicitly de-select CPUs), disable all CPUs. 1782 */ 1783 if (!(tstat_options & TSTAT_OPT_CPU) || cmd == TSTATIOC_NOCPU) { 1784 tstat_options |= TSTAT_OPT_CPU; 1785 1786 for (i = 0; i <= max_cpuid; i++) { 1787 tstat_percpu_t *tcpu = &tstat_percpu[i]; 1788 1789 ASSERT(cmd == TSTATIOC_NOCPU || 1790 (tcpu->tcpu_flags & TSTAT_CPU_SELECTED)); 1791 tcpu->tcpu_flags &= ~TSTAT_CPU_SELECTED; 1792 } 1793 } 1794 1795 if (cmd == TSTATIOC_CPU) 1796 tstat_percpu[arg].tcpu_flags |= TSTAT_CPU_SELECTED; 1797 1798 mutex_exit(&tstat_lock); 1799 1800 return (0); 1801 1802 case TSTATIOC_ENTRY: 1803 mutex_enter(&tstat_lock); 1804 1805 if (tstat_running) { 1806 mutex_exit(&tstat_lock); 1807 return (EBUSY); 1808 } 1809 1810 if (arg >= TSTAT_NENT || arg < 0) { 1811 mutex_exit(&tstat_lock); 1812 return (EINVAL); 1813 } 1814 1815 if (!(tstat_options & TSTAT_OPT_ENTRY)) { 1816 /* 1817 * If this is the first entry that we are explicitly 1818 * enabling, explicitly disable every TL=0 entry. 1819 */ 1820 for (i = 0; i < TSTAT_NENT; i++) 1821 tstat_enabled[i] = 0; 1822 1823 tstat_options |= TSTAT_OPT_ENTRY; 1824 } 1825 1826 tstat_enabled[arg] = 1; 1827 mutex_exit(&tstat_lock); 1828 return (0); 1829 1830 case TSTATIOC_NOENTRY: 1831 mutex_enter(&tstat_lock); 1832 1833 if (tstat_running) { 1834 mutex_exit(&tstat_lock); 1835 return (EBUSY); 1836 } 1837 1838 for (i = 0; i < TSTAT_NENT; i++) 1839 tstat_enabled[i] = 0; 1840 1841 mutex_exit(&tstat_lock); 1842 return (0); 1843 1844 case TSTATIOC_READ: 1845 mutex_enter(&tstat_lock); 1846 1847 if (tstat_options & TSTAT_OPT_TLBDATA) { 1848 dsize = tstat_data_t_exported_size; 1849 } else { 1850 dsize = sizeof (tstat_data_t); 1851 } 1852 1853 for (i = 0, out = 0; i <= max_cpuid; i++) { 1854 tstat_percpu_t *tcpu = &tstat_percpu[i]; 1855 1856 if (!(tcpu->tcpu_flags & TSTAT_CPU_ENABLED)) 1857 continue; 1858 1859 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_SELECTED); 1860 ASSERT(tcpu->tcpu_flags & TSTAT_CPU_ALLOCATED); 1861 1862 tstat_buffer->tdata_cpuid = -1; 1863 xc_one(i, (xcfunc_t *)trapstat_snapshot, 0, 0); 1864 1865 if (tstat_buffer->tdata_cpuid == -1) { 1866 /* 1867 * This CPU is not currently responding to 1868 * cross calls; we have caught it while it is 1869 * being unconfigured. We'll drop tstat_lock 1870 * and pick up and drop cpu_lock. By the 1871 * time we acquire cpu_lock, the DR operation 1872 * will appear consistent and we can assert 1873 * that trapstat_cpu_setup() has cleared 1874 * TSTAT_CPU_ENABLED. 1875 */ 1876 mutex_exit(&tstat_lock); 1877 mutex_enter(&cpu_lock); 1878 mutex_exit(&cpu_lock); 1879 mutex_enter(&tstat_lock); 1880 ASSERT(!(tcpu->tcpu_flags & TSTAT_CPU_ENABLED)); 1881 continue; 1882 } 1883 1884 /* 1885 * Need to compensate for the difference between page 1886 * sizes exported to users and page sizes available 1887 * within the kernel. 1888 */ 1889 if ((tstat_options & TSTAT_OPT_TLBDATA) && 1890 (tstat_pgszs != tstat_user_pgszs)) { 1891 tstat_pgszdata_t *tp; 1892 uint_t szc; 1893 1894 tp = &tstat_buffer->tdata_pgsz[0]; 1895 for (j = 0; j < tstat_user_pgszs; j++) { 1896 if ((szc = USERSZC_2_SZC(j)) != j) { 1897 bcopy(&tp[szc], &tp[j], 1898 sizeof (tstat_pgszdata_t)); 1899 } 1900 } 1901 } 1902 1903 if (copyout(tstat_buffer, (void *)arg, dsize) != 0) { 1904 mutex_exit(&tstat_lock); 1905 return (EFAULT); 1906 } 1907 1908 out++; 1909 arg += dsize; 1910 } 1911 1912 if (out != max_cpuid + 1) { 1913 processorid_t cpuid = -1; 1914 arg += offsetof(tstat_data_t, tdata_cpuid); 1915 1916 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0) { 1917 mutex_exit(&tstat_lock); 1918 return (EFAULT); 1919 } 1920 } 1921 1922 mutex_exit(&tstat_lock); 1923 1924 return (0); 1925 1926 case TSTATIOC_TLBDATA: 1927 return (trapstat_option(TSTAT_OPT_TLBDATA)); 1928 1929 default: 1930 break; 1931 } 1932 1933 return (ENOTTY); 1934 } 1935 1936 /*ARGSUSED*/ 1937 static int 1938 trapstat_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result) 1939 { 1940 int error; 1941 1942 switch (infocmd) { 1943 case DDI_INFO_DEVT2DEVINFO: 1944 *result = (void *)tstat_devi; 1945 error = DDI_SUCCESS; 1946 break; 1947 case DDI_INFO_DEVT2INSTANCE: 1948 *result = (void *)0; 1949 error = DDI_SUCCESS; 1950 break; 1951 default: 1952 error = DDI_FAILURE; 1953 } 1954 return (error); 1955 } 1956 1957 static int 1958 trapstat_attach(dev_info_t *devi, ddi_attach_cmd_t cmd) 1959 { 1960 switch (cmd) { 1961 case DDI_ATTACH: 1962 break; 1963 1964 case DDI_RESUME: 1965 return (DDI_SUCCESS); 1966 1967 default: 1968 return (DDI_FAILURE); 1969 } 1970 1971 if (ddi_create_minor_node(devi, "trapstat", S_IFCHR, 1972 0, DDI_PSEUDO, 0) == DDI_FAILURE) { 1973 ddi_remove_minor_node(devi, NULL); 1974 return (DDI_FAILURE); 1975 } 1976 1977 ddi_report_dev(devi); 1978 tstat_devi = devi; 1979 1980 tstat_pgszs = page_num_pagesizes(); 1981 tstat_user_pgszs = page_num_user_pagesizes(); 1982 tstat_data_t_size = sizeof (tstat_data_t) + 1983 (tstat_pgszs - 1) * sizeof (tstat_pgszdata_t); 1984 tstat_data_t_exported_size = sizeof (tstat_data_t) + 1985 (tstat_user_pgszs - 1) * sizeof (tstat_pgszdata_t); 1986 #ifndef sun4v 1987 tstat_data_pages = (tstat_data_t_size >> MMU_PAGESHIFT) + 1; 1988 tstat_total_pages = TSTAT_INSTR_PAGES + tstat_data_pages; 1989 tstat_data_size = tstat_data_pages * MMU_PAGESIZE; 1990 tstat_total_size = TSTAT_INSTR_SIZE + tstat_data_size; 1991 #else 1992 tstat_data_pages = 0; 1993 tstat_data_size = tstat_data_t_size; 1994 tstat_total_pages = ((TSTAT_INSTR_SIZE + tstat_data_size) >> 1995 MMU_PAGESHIFT) + 1; 1996 tstat_total_size = tstat_total_pages * MMU_PAGESIZE; 1997 #endif 1998 1999 tstat_percpu = kmem_zalloc((max_cpuid + 1) * 2000 sizeof (tstat_percpu_t), KM_SLEEP); 2001 2002 /* 2003 * Create our own arena backed by segkmem to assure a source of 2004 * MMU_PAGESIZE-aligned allocations. We allocate out of the 2005 * heap32_arena to assure that we can address the allocated memory with 2006 * a single sethi/simm13 pair in the interposing trap table entries. 2007 */ 2008 tstat_arena = vmem_create("trapstat", NULL, 0, MMU_PAGESIZE, 2009 segkmem_alloc_permanent, segkmem_free, heap32_arena, 0, VM_SLEEP); 2010 2011 tstat_enabled = kmem_alloc(TSTAT_TOTAL_NENT * sizeof (int), KM_SLEEP); 2012 tstat_buffer = kmem_alloc(tstat_data_t_size, KM_SLEEP); 2013 2014 /* 2015 * CB_CL_CPR_POST_USER is the class that executes from cpr_resume() 2016 * after user threads can be restarted. By executing in this class, 2017 * we are assured of the availability of system services needed to 2018 * resume trapstat (specifically, we are assured that all CPUs are 2019 * restarted and responding to cross calls). 2020 */ 2021 tstat_cprcb = 2022 callb_add(trapstat_cpr, NULL, CB_CL_CPR_POST_USER, "trapstat"); 2023 2024 return (DDI_SUCCESS); 2025 } 2026 2027 static int 2028 trapstat_detach(dev_info_t *devi, ddi_detach_cmd_t cmd) 2029 { 2030 int rval; 2031 2032 ASSERT(devi == tstat_devi); 2033 2034 switch (cmd) { 2035 case DDI_DETACH: 2036 break; 2037 2038 case DDI_SUSPEND: 2039 return (DDI_SUCCESS); 2040 2041 default: 2042 return (DDI_FAILURE); 2043 } 2044 2045 ASSERT(!tstat_running); 2046 2047 rval = callb_delete(tstat_cprcb); 2048 ASSERT(rval == 0); 2049 2050 kmem_free(tstat_buffer, tstat_data_t_size); 2051 kmem_free(tstat_enabled, TSTAT_TOTAL_NENT * sizeof (int)); 2052 vmem_destroy(tstat_arena); 2053 kmem_free(tstat_percpu, (max_cpuid + 1) * sizeof (tstat_percpu_t)); 2054 ddi_remove_minor_node(devi, NULL); 2055 2056 return (DDI_SUCCESS); 2057 } 2058 2059 /* 2060 * Configuration data structures 2061 */ 2062 static struct cb_ops trapstat_cb_ops = { 2063 trapstat_open, /* open */ 2064 trapstat_close, /* close */ 2065 nulldev, /* strategy */ 2066 nulldev, /* print */ 2067 nodev, /* dump */ 2068 nodev, /* read */ 2069 nodev, /* write */ 2070 trapstat_ioctl, /* ioctl */ 2071 nodev, /* devmap */ 2072 nodev, /* mmap */ 2073 nodev, /* segmap */ 2074 nochpoll, /* poll */ 2075 ddi_prop_op, /* cb_prop_op */ 2076 0, /* streamtab */ 2077 D_MP | D_NEW /* Driver compatibility flag */ 2078 }; 2079 2080 static struct dev_ops trapstat_ops = { 2081 DEVO_REV, /* devo_rev, */ 2082 0, /* refcnt */ 2083 trapstat_info, /* getinfo */ 2084 nulldev, /* identify */ 2085 nulldev, /* probe */ 2086 trapstat_attach, /* attach */ 2087 trapstat_detach, /* detach */ 2088 nulldev, /* reset */ 2089 &trapstat_cb_ops, /* cb_ops */ 2090 (struct bus_ops *)0, /* bus_ops */ 2091 }; 2092 2093 static struct modldrv modldrv = { 2094 &mod_driverops, /* Type of module. This one is a driver */ 2095 "Trap Statistics", /* name of module */ 2096 &trapstat_ops, /* driver ops */ 2097 }; 2098 2099 static struct modlinkage modlinkage = { 2100 MODREV_1, (void *)&modldrv, NULL 2101 }; 2102 2103 int 2104 _init(void) 2105 { 2106 return (mod_install(&modlinkage)); 2107 } 2108 2109 int 2110 _fini(void) 2111 { 2112 return (mod_remove(&modlinkage)); 2113 } 2114 2115 int 2116 _info(struct modinfo *modinfop) 2117 { 2118 return (mod_info(&modlinkage, modinfop)); 2119 } 2120