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 (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  24  */
  25 
  26 #include <sys/types.h>
  27 #include <sys/param.h>
  28 #include <sys/cmn_err.h>
  29 #include <sys/mutex.h>
  30 #include <sys/systm.h>
  31 #include <sys/sysmacros.h>
  32 #include <sys/machsystm.h>
  33 #include <sys/archsystm.h>
  34 #include <sys/x_call.h>
  35 #include <sys/promif.h>
  36 #include <sys/prom_isa.h>
  37 #include <sys/privregs.h>
  38 #include <sys/vmem.h>
  39 #include <sys/atomic.h>
  40 #include <sys/panic.h>
  41 #include <sys/rwlock.h>
  42 #include <sys/reboot.h>
  43 #include <sys/kdi.h>
  44 #include <sys/kdi_machimpl.h>
  45 
  46 /*
  47  * We are called with a pointer to a cell-sized argument array.
  48  * The service name (the first element of the argument array) is
  49  * the name of the callback being invoked.  When called, we are
  50  * running on the firmwares trap table as a trusted subroutine
  51  * of the firmware.
  52  *
  53  * We define entry points to allow callback handlers to be dynamically
  54  * added and removed, to support obpsym, which is a separate module
  55  * and can be dynamically loaded and unloaded and registers its
  56  * callback handlers dynamically.
  57  *
  58  * Note: The actual callback handler we register, is the assembly lang.
  59  * glue, callback_handler, which takes care of switching from a 64
  60  * bit stack and environment to a 32 bit stack and environment, and
  61  * back again, if the callback handler returns. callback_handler calls
  62  * vx_handler to process the callback.
  63  */
  64 
  65 static kmutex_t vx_cmd_lock;    /* protect vx_cmd table */
  66 
  67 #define VX_CMD_MAX      10
  68 #define ENDADDR(a)      &a[sizeof (a) / sizeof (a[0])]
  69 #define vx_cmd_end      ((struct vx_cmd *)(ENDADDR(vx_cmd)))
  70 
  71 static struct vx_cmd {
  72         char    *service;       /* Service name */
  73         int     take_tba;       /* If Non-zero we take over the tba */
  74         void    (*func)(cell_t *argument_array);
  75 } vx_cmd[VX_CMD_MAX+1];
  76 
  77 void
  78 init_vx_handler(void)
  79 {
  80         extern int callback_handler(cell_t *arg_array);
  81 
  82         /*
  83          * initialize the lock protecting additions and deletions from
  84          * the vx_cmd table.  At callback time we don't need to grab
  85          * this lock.  Callback handlers do not need to modify the
  86          * callback handler table.
  87          */
  88         mutex_init(&vx_cmd_lock, NULL, MUTEX_DEFAULT, NULL);
  89 
  90         /*
  91          * Tell OBP about our callback handler.
  92          */
  93         (void) prom_set_callback((void *)callback_handler);
  94 }
  95 
  96 /*
  97  * Add a kernel callback handler to the kernel's list.
  98  * The table is static, so if you add a callback handler, increase
  99  * the value of VX_CMD_MAX. Find the first empty slot and use it.
 100  */
 101 void
 102 add_vx_handler(char *name, int flag, void (*func)(cell_t *))
 103 {
 104         struct vx_cmd *vp;
 105 
 106         mutex_enter(&vx_cmd_lock);
 107         for (vp = vx_cmd; vp < vx_cmd_end; vp++) {
 108                 if (vp->service == NULL) {
 109                         vp->service = name;
 110                         vp->take_tba = flag;
 111                         vp->func = func;
 112                         mutex_exit(&vx_cmd_lock);
 113                         return;
 114                 }
 115         }
 116         mutex_exit(&vx_cmd_lock);
 117 
 118 #ifdef  DEBUG
 119 
 120         /*
 121          * There must be enough entries to handle all callback entries.
 122          * Increase VX_CMD_MAX if this happens. This shouldn't happen.
 123          */
 124         cmn_err(CE_PANIC, "add_vx_handler <%s>", name);
 125         /* NOTREACHED */
 126 
 127 #else   /* DEBUG */
 128 
 129         cmn_err(CE_WARN, "add_vx_handler: Can't add callback hander <%s>",
 130             name);
 131 
 132 #endif  /* DEBUG */
 133 
 134 }
 135 
 136 /*
 137  * Remove a vx_handler function -- find the name string in the table,
 138  * and clear it.
 139  */
 140 void
 141 remove_vx_handler(char *name)
 142 {
 143         struct vx_cmd *vp;
 144 
 145         mutex_enter(&vx_cmd_lock);
 146         for (vp = vx_cmd; vp < vx_cmd_end; vp++) {
 147                 if (vp->service == NULL)
 148                         continue;
 149                 if (strcmp(vp->service, name) != 0)
 150                         continue;
 151                 vp->service = 0;
 152                 vp->take_tba = 0;
 153                 vp->func = 0;
 154                 mutex_exit(&vx_cmd_lock);
 155                 return;
 156         }
 157         mutex_exit(&vx_cmd_lock);
 158         cmn_err(CE_WARN, "remove_vx_handler: <%s> not found", name);
 159 }
 160 
 161 int
 162 vx_handler(cell_t *argument_array)
 163 {
 164         char *name;
 165         struct vx_cmd *vp;
 166         void *old_tba;
 167 
 168         name = p1275_cell2ptr(*argument_array);
 169 
 170         for (vp = vx_cmd; vp < vx_cmd_end; vp++) {
 171                 if (vp->service == (char *)0)
 172                         continue;
 173                 if (strcmp(vp->service, name) != 0)
 174                         continue;
 175                 if (vp->take_tba != 0)  {
 176                         reestablish_curthread();
 177                         if (tba_taken_over != 0)
 178                                 old_tba = set_tba((void *)&trap_table);
 179                 }
 180                 vp->func(argument_array);
 181                 if ((vp->take_tba != 0) && (tba_taken_over != 0))
 182                         (void) set_tba(old_tba);
 183                 return (0);     /* Service name was known */
 184         }
 185 
 186         return (-1);            /* Service name unknown */
 187 }
 188 
 189 /*
 190  * PROM Locking Primitives
 191  *
 192  * These routines are called immediately before and immediately after calling
 193  * into the firmware.  The firmware is single-threaded and assumes that the
 194  * kernel will implement locking to prevent simultaneous service calls.  In
 195  * addition, some service calls (particularly character rendering) can be
 196  * slow, so we would like to sleep if we cannot acquire the lock to allow the
 197  * caller's CPU to continue to perform useful work in the interim.  Service
 198  * routines may also be called early in boot as part of slave CPU startup
 199  * when mutexes and cvs are not yet available (i.e. they are still running on
 200  * the prom's TLB handlers and cannot touch curthread).  Therefore, these
 201  * routines must reduce to a simple compare-and-swap spin lock when necessary.
 202  * Finally, kernel code may wish to acquire the firmware lock before executing
 203  * a block of code that includes service calls, so we also allow the firmware
 204  * lock to be acquired recursively by the owning CPU after disabling preemption.
 205  *
 206  * To meet these constraints, the lock itself is implemented as a compare-and-
 207  * swap spin lock on the global prom_cpu pointer.  We implement recursion by
 208  * atomically incrementing the integer prom_holdcnt after acquiring the lock.
 209  * If the current CPU is an "adult" (determined by testing cpu_m.mutex_ready),
 210  * we disable preemption before acquiring the lock and leave it disabled once
 211  * the lock is held.  The kern_postprom() routine then enables preemption if
 212  * we drop the lock and prom_holdcnt returns to zero.  If the current CPU is
 213  * an adult and the lock is held by another adult CPU, we can safely sleep
 214  * until the lock is released.  To do so, we acquire the adaptive prom_mutex
 215  * and then sleep on prom_cv.  Therefore, service routines must not be called
 216  * from above LOCK_LEVEL on any adult CPU.  Finally, if recursive entry is
 217  * attempted on an adult CPU, we must also verify that curthread matches the
 218  * saved prom_thread (the original owner) to ensure that low-level interrupt
 219  * threads do not step on other threads running on the same CPU.
 220  */
 221 
 222 static cpu_t *volatile prom_cpu;
 223 static kthread_t *volatile prom_thread;
 224 static uint32_t prom_holdcnt;
 225 static kmutex_t prom_mutex;
 226 static kcondvar_t prom_cv;
 227 
 228 /*
 229  * The debugger uses PROM services, and is thus unable to run if any of the
 230  * CPUs on the system are executing in the PROM at the time of debugger entry.
 231  * If a CPU is determined to be in the PROM when the debugger is entered,
 232  * prom_return_enter_debugger will be set, thus triggering a programmed debugger
 233  * entry when the given CPU returns from the PROM.  That CPU is then released by
 234  * the debugger, and is allowed to complete PROM-related work.
 235  */
 236 int prom_exit_enter_debugger;
 237 
 238 void
 239 kern_preprom(void)
 240 {
 241         for (;;) {
 242                 /*
 243                  * Load the current CPU pointer and examine the mutex_ready bit.
 244                  * It doesn't matter if we are preempted here because we are
 245                  * only trying to determine if we are in the *set* of mutex
 246                  * ready CPUs.  We cannot disable preemption until we confirm
 247                  * that we are running on a CPU in this set, since a call to
 248                  * kpreempt_disable() requires access to curthread.
 249                  */
 250                 processorid_t cpuid = getprocessorid();
 251                 cpu_t *cp = cpu[cpuid];
 252                 cpu_t *prcp;
 253 
 254                 if (panicstr)
 255                         return; /* just return if we are currently panicking */
 256 
 257                 if (CPU_IN_SET(cpu_ready_set, cpuid) && cp->cpu_m.mutex_ready) {
 258                         /*
 259                          * Disable premption, and reload the current CPU.  We
 260                          * can't move from a mutex_ready cpu to a non-ready cpu
 261                          * so we don't need to re-check cp->cpu_m.mutex_ready.
 262                          */
 263                         kpreempt_disable();
 264                         cp = CPU;
 265                         ASSERT(cp->cpu_m.mutex_ready);
 266 
 267                         /*
 268                          * Try the lock.  If we don't get the lock, re-enable
 269                          * preemption and see if we should sleep.  If we are
 270                          * already the lock holder, remove the effect of the
 271                          * previous kpreempt_disable() before returning since
 272                          * preemption was disabled by an earlier kern_preprom.
 273                          */
 274                         prcp = atomic_cas_ptr((void *)&prom_cpu, NULL, cp);
 275                         if (prcp == NULL ||
 276                             (prcp == cp && prom_thread == curthread)) {
 277                                 if (prcp == cp)
 278                                         kpreempt_enable();
 279                                 break;
 280                         }
 281 
 282                         kpreempt_enable();
 283 
 284                         /*
 285                          * We have to be very careful here since both prom_cpu
 286                          * and prcp->cpu_m.mutex_ready can be changed at any
 287                          * time by a non mutex_ready cpu holding the lock.
 288                          * If the owner is mutex_ready, holding prom_mutex
 289                          * prevents kern_postprom() from completing.  If the
 290                          * owner isn't mutex_ready, we only know it will clear
 291                          * prom_cpu before changing cpu_m.mutex_ready, so we
 292                          * issue a membar after checking mutex_ready and then
 293                          * re-verify that prom_cpu is still held by the same
 294                          * cpu before actually proceeding to cv_wait().
 295                          */
 296                         mutex_enter(&prom_mutex);
 297                         prcp = prom_cpu;
 298                         if (prcp != NULL && prcp->cpu_m.mutex_ready != 0) {
 299                                 membar_consumer();
 300                                 if (prcp == prom_cpu)
 301                                         cv_wait(&prom_cv, &prom_mutex);
 302                         }
 303                         mutex_exit(&prom_mutex);
 304 
 305                 } else {
 306                         /*
 307                          * If we are not yet mutex_ready, just attempt to grab
 308                          * the lock.  If we get it or already hold it, break.
 309                          */
 310                         ASSERT(getpil() == PIL_MAX);
 311                         prcp = atomic_cas_ptr((void *)&prom_cpu, NULL, cp);
 312                         if (prcp == NULL || prcp == cp)
 313                                 break;
 314                 }
 315         }
 316 
 317         /*
 318          * We now hold the prom_cpu lock.  Increment the hold count by one
 319          * and assert our current state before returning to the caller.
 320          */
 321         atomic_inc_32(&prom_holdcnt);
 322         ASSERT(prom_holdcnt >= 1);
 323         prom_thread = curthread;
 324 }
 325 
 326 /*
 327  * Drop the prom lock if it is held by the current CPU.  If the lock is held
 328  * recursively, return without clearing prom_cpu.  If the hold count is now
 329  * zero, clear prom_cpu and cv_signal any waiting CPU.
 330  */
 331 void
 332 kern_postprom(void)
 333 {
 334         processorid_t cpuid = getprocessorid();
 335         cpu_t *cp = cpu[cpuid];
 336 
 337         if (panicstr)
 338                 return; /* do not modify lock further if we have panicked */
 339 
 340         if (prom_cpu != cp)
 341                 panic("kern_postprom: not owner, cp=%p owner=%p",
 342                     (void *)cp, (void *)prom_cpu);
 343 
 344         if (prom_holdcnt == 0)
 345                 panic("kern_postprom: prom_holdcnt == 0, owner=%p",
 346                     (void *)prom_cpu);
 347 
 348         if (atomic_dec_32_nv(&prom_holdcnt) != 0)
 349                 return; /* prom lock is held recursively by this CPU */
 350 
 351         if ((boothowto & RB_DEBUG) && prom_exit_enter_debugger)
 352                 kmdb_enter();
 353 
 354         prom_thread = NULL;
 355         membar_producer();
 356 
 357         prom_cpu = NULL;
 358         membar_producer();
 359 
 360         if (CPU_IN_SET(cpu_ready_set, cpuid) && cp->cpu_m.mutex_ready) {
 361                 mutex_enter(&prom_mutex);
 362                 cv_signal(&prom_cv);
 363                 mutex_exit(&prom_mutex);
 364                 kpreempt_enable();
 365         }
 366 }
 367 
 368 /*
 369  * If the frame buffer device is busy, briefly capture the other CPUs so that
 370  * another CPU executing code to manipulate the device does not execute at the
 371  * same time we are rendering characters.  Refer to the comments and code in
 372  * common/os/console.c for more information on these callbacks.
 373  *
 374  * Notice that we explicitly acquire the PROM lock using kern_preprom() prior
 375  * to idling other CPUs.  The idling mechanism will cross-trap the other CPUs
 376  * and have them spin at MAX(%pil, XCALL_PIL), so we must be sure that none of
 377  * them are holding the PROM lock before we idle them and then call into the
 378  * PROM routines that render characters to the frame buffer.
 379  */
 380 int
 381 console_enter(int busy)
 382 {
 383         int s = 0;
 384 
 385         if (busy && panicstr == NULL) {
 386                 kern_preprom();
 387                 s = splhi();
 388                 idle_other_cpus();
 389         }
 390 
 391         return (s);
 392 }
 393 
 394 void
 395 console_exit(int busy, int spl)
 396 {
 397         if (busy && panicstr == NULL) {
 398                 resume_other_cpus();
 399                 splx(spl);
 400                 kern_postprom();
 401         }
 402 }
 403 
 404 /*
 405  * This routine is a special form of pause_cpus().  It ensures that
 406  * prom functions are callable while the cpus are paused.
 407  */
 408 void
 409 promsafe_pause_cpus(void)
 410 {
 411         pause_cpus(NULL);
 412 
 413         /* If some other cpu is entering or is in the prom, spin */
 414         while (prom_cpu || mutex_owner(&prom_mutex)) {
 415 
 416                 start_cpus();
 417                 mutex_enter(&prom_mutex);
 418 
 419                 /* Wait for other cpu to exit prom */
 420                 while (prom_cpu)
 421                         cv_wait(&prom_cv, &prom_mutex);
 422 
 423                 mutex_exit(&prom_mutex);
 424                 pause_cpus(NULL);
 425         }
 426 
 427         /* At this point all cpus are paused and none are in the prom */
 428 }
 429 
 430 /*
 431  * This routine is a special form of xc_attention().  It ensures that
 432  * prom functions are callable while the cpus are at attention.
 433  */
 434 void
 435 promsafe_xc_attention(cpuset_t cpuset)
 436 {
 437         xc_attention(cpuset);
 438 
 439         /* If some other cpu is entering or is in the prom, spin */
 440         while (prom_cpu || mutex_owner(&prom_mutex)) {
 441 
 442                 xc_dismissed(cpuset);
 443                 mutex_enter(&prom_mutex);
 444 
 445                 /* Wait for other cpu to exit prom */
 446                 while (prom_cpu)
 447                         cv_wait(&prom_cv, &prom_mutex);
 448 
 449                 mutex_exit(&prom_mutex);
 450                 xc_attention(cpuset);
 451         }
 452 
 453         /* At this point all cpus are paused and none are in the prom */
 454 }
 455 
 456 
 457 #if defined(PROM_32BIT_ADDRS)
 458 
 459 #include <sys/promimpl.h>
 460 #include <vm/seg_kmem.h>
 461 #include <sys/kmem.h>
 462 #include <sys/bootconf.h>
 463 
 464 /*
 465  * These routines are only used to workaround "poor feature interaction"
 466  * in OBP.  See bug 4115680 for details.
 467  *
 468  * Many of the promif routines need to allocate temporary buffers
 469  * with 32-bit addresses to pass in/out of the CIF.  The lifetime
 470  * of the buffers is extremely short, they are allocated and freed
 471  * around the CIF call.  We use vmem_alloc() to cache 32-bit memory.
 472  *
 473  * Note the code in promplat_free() to prevent exhausting the 32 bit
 474  * heap during boot.
 475  */
 476 static void *promplat_last_free = NULL;
 477 static size_t promplat_last_size;
 478 static vmem_t *promplat_arena;
 479 static kmutex_t promplat_lock;  /* protect arena, last_free, and last_size */
 480 
 481 void *
 482 promplat_alloc(size_t size)
 483 {
 484 
 485         mutex_enter(&promplat_lock);
 486         if (promplat_arena == NULL) {
 487                 promplat_arena = vmem_create("promplat", NULL, 0, 8,
 488                     segkmem_alloc, segkmem_free, heap32_arena, 0, VM_SLEEP);
 489         }
 490         mutex_exit(&promplat_lock);
 491 
 492         return (vmem_alloc(promplat_arena, size, VM_NOSLEEP));
 493 }
 494 
 495 /*
 496  * Delaying the free() of small allocations gets more mileage
 497  * from pages during boot, otherwise a cycle of allocate/free
 498  * calls could burn through available heap32 space too quickly.
 499  */
 500 void
 501 promplat_free(void *p, size_t size)
 502 {
 503         void *p2 = NULL;
 504         size_t s2;
 505 
 506         /*
 507          * If VM is initialized, clean up any delayed free().
 508          */
 509         if (kvseg.s_base != 0 && promplat_last_free != NULL) {
 510                 mutex_enter(&promplat_lock);
 511                 p2 = promplat_last_free;
 512                 s2 = promplat_last_size;
 513                 promplat_last_free = NULL;
 514                 promplat_last_size = 0;
 515                 mutex_exit(&promplat_lock);
 516                 if (p2 != NULL) {
 517                         vmem_free(promplat_arena, p2, s2);
 518                         p2 = NULL;
 519                 }
 520         }
 521 
 522         /*
 523          * Do the free if VM is initialized or it's a large allocation.
 524          */
 525         if (kvseg.s_base != 0 || size >= PAGESIZE) {
 526                 vmem_free(promplat_arena, p, size);
 527                 return;
 528         }
 529 
 530         /*
 531          * Otherwise, do the last free request and delay this one.
 532          */
 533         mutex_enter(&promplat_lock);
 534         if (promplat_last_free != NULL) {
 535                 p2 = promplat_last_free;
 536                 s2 = promplat_last_size;
 537         }
 538         promplat_last_free = p;
 539         promplat_last_size = size;
 540         mutex_exit(&promplat_lock);
 541 
 542         if (p2 != NULL)
 543                 vmem_free(promplat_arena, p2, s2);
 544 }
 545 
 546 void
 547 promplat_bcopy(const void *src, void *dst, size_t count)
 548 {
 549         bcopy(src, dst, count);
 550 }
 551 
 552 #endif /* PROM_32BIT_ADDRS */
 553 
 554 static prom_generation_cookie_t prom_tree_gen;
 555 static krwlock_t prom_tree_lock;
 556 
 557 int
 558 prom_tree_access(int (*callback)(void *arg, int has_changed), void *arg,
 559     prom_generation_cookie_t *ckp)
 560 {
 561         int chg, rv;
 562 
 563         rw_enter(&prom_tree_lock, RW_READER);
 564         /*
 565          * If the tree has changed since the caller last accessed it
 566          * pass 1 as the second argument to the callback function,
 567          * otherwise 0.
 568          */
 569         if (ckp != NULL && *ckp != prom_tree_gen) {
 570                 *ckp = prom_tree_gen;
 571                 chg = 1;
 572         } else
 573                 chg = 0;
 574         rv = callback(arg, chg);
 575         rw_exit(&prom_tree_lock);
 576         return (rv);
 577 }
 578 
 579 int
 580 prom_tree_update(int (*callback)(void *arg), void *arg)
 581 {
 582         int rv;
 583 
 584         rw_enter(&prom_tree_lock, RW_WRITER);
 585         prom_tree_gen++;
 586         rv = callback(arg);
 587         rw_exit(&prom_tree_lock);
 588         return (rv);
 589 }