lustre/portals/include/linux/kp30.h

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  */
   4 #ifndef _KP30_INCLUDED
   5 #define _KP30_INCLUDED
   6
   7 #define PORTAL_DEBUG
   8
   9 #ifndef offsetof
  10 # define offsetof(typ,memb)     ((int)((char *)&(((typ *)0)->memb)))
  11 #endif
  12
  13 #define LOWEST_BIT_SET(x)       ((x) & ~((x) - 1))
  14
  15 /*
  16  *  Debugging
  17  */
  18 extern unsigned int portal_subsystem_debug;
  19 extern unsigned int portal_stack;
  20 extern unsigned int portal_debug;
  21 extern unsigned int portal_printk;
  22 extern unsigned int portal_cerror;
  23 /* Debugging subsystems (32 bits, non-overlapping) */
  24 #define S_UNDEFINED    (1 << 0)
  25 #define S_MDC          (1 << 1)
  26 #define S_MDS          (1 << 2)
  27 #define S_OSC          (1 << 3)
  28 #define S_OST          (1 << 4)
  29 #define S_CLASS        (1 << 5)
  30 #define S_LOG          (1 << 6)
  31 #define S_LLITE        (1 << 7)
  32 #define S_RPC          (1 << 8)
  33 #define S_MGMT         (1 << 9)
  34 #define S_PORTALS     (1 << 10)
  35 #define S_SOCKNAL     (1 << 11)
  36 #define S_QSWNAL      (1 << 12)
  37 #define S_PINGER      (1 << 13)
  38 #define S_FILTER      (1 << 14)
  39 #define S_PTLBD       (1 << 15)
  40 #define S_ECHO        (1 << 16)
  41 #define S_LDLM        (1 << 17)
  42 #define S_LOV         (1 << 18)
  43 #define S_GMNAL       (1 << 19)
  44 #define S_PTLROUTER   (1 << 20)
  45 #define S_COBD        (1 << 21)
  46 #define S_IBNAL       (1 << 22)
  47
  48 /* If you change these values, please keep portals/utils/debug.c
  49  * up to date! */
  50
  51 /* Debugging masks (32 bits, non-overlapping) */
  52 #define D_TRACE     (1 << 0) /* ENTRY/EXIT markers */
  53 #define D_INODE     (1 << 1)
  54 #define D_SUPER     (1 << 2)
  55 #define D_EXT2      (1 << 3) /* anything from ext2_debug */
  56 #define D_MALLOC    (1 << 4) /* print malloc, free information */
  57 #define D_CACHE     (1 << 5) /* cache-related items */
  58 #define D_INFO      (1 << 6) /* general information */
  59 #define D_IOCTL     (1 << 7) /* ioctl related information */
  60 #define D_BLOCKS    (1 << 8) /* ext2 block allocation */
  61 #define D_NET       (1 << 9) /* network communications */
  62 #define D_WARNING   (1 << 10) /* CWARN(...) == CDEBUG (D_WARNING, ...) */
  63 #define D_BUFFS     (1 << 11)
  64 #define D_OTHER     (1 << 12)
  65 #define D_DENTRY    (1 << 13)
  66 #define D_PORTALS   (1 << 14) /* ENTRY/EXIT markers */
  67 #define D_PAGE      (1 << 15) /* bulk page handling */
  68 #define D_DLMTRACE  (1 << 16)
  69 #define D_ERROR     (1 << 17) /* CERROR(...) == CDEBUG (D_ERROR, ...) */
  70 #define D_EMERG     (1 << 18) /* CEMERG(...) == CDEBUG (D_EMERG, ...) */
  71 #define D_HA        (1 << 19) /* recovery and failover */
  72 #define D_RPCTRACE  (1 << 20) /* for distributed debugging */
  73 #define D_VFSTRACE  (1 << 21)
  74 #define D_READA     (1 << 22) /* read-ahead */
  75
  76 #ifdef __KERNEL__
  77 # include <linux/sched.h> /* THREAD_SIZE */
  78 #else
  79 # ifndef THREAD_SIZE /* x86_64 has THREAD_SIZE in userspace */
  80 #  define THREAD_SIZE 8192
  81 # endif
  82 #endif
  83
  84 #define LUSTRE_TRACE_SIZE (THREAD_SIZE >> 5)
  85
  86 #ifdef __KERNEL__
  87 # ifdef  __ia64__
  88 #  define CDEBUG_STACK (THREAD_SIZE -                                      \
  89                         ((unsigned long)__builtin_dwarf_cfa() &            \
  90                          (THREAD_SIZE - 1)))
  91 # else
  92 #  define CDEBUG_STACK (THREAD_SIZE -                                      \
  93                         ((unsigned long)__builtin_frame_address(0) &       \
  94                          (THREAD_SIZE - 1)))
  95 # endif
  96
  97 #define CHECK_STACK(stack)                                                    \
  98         do {                                                                  \
  99                 if ((stack) > 3*THREAD_SIZE/4 && (stack) > portal_stack) {    \
 100                         portals_debug_msg(DEBUG_SUBSYSTEM, D_WARNING,         \
 101                                           __FILE__, __FUNCTION__, __LINE__,   \
 102                                           (stack),"maximum lustre stack %u\n",\
 103                                           portal_stack = (stack));            \
 104                       /*panic("LBUG");*/                                      \
 105                 }                                                             \
 106         } while (0)
 107 #else /* __KERNEL__ */
 108 #define CHECK_STACK(stack) do { } while(0)
 109 #define CDEBUG_STACK (0L)
 110 #endif /* __KERNEL__ */
 111
 112 #if 1
 113 #define CDEBUG(mask, format, a...)                                            \
 114 do {                                                                          \
 115         if (portal_cerror == 0)                                               \
 116                 break;                                                        \
 117         CHECK_STACK(CDEBUG_STACK);                                            \
 118         if (((mask) & (D_ERROR | D_EMERG | D_WARNING)) ||                     \
 119             (portal_debug & (mask) &&                                         \
 120              portal_subsystem_debug & DEBUG_SUBSYSTEM))                       \
 121                 portals_debug_msg(DEBUG_SUBSYSTEM, mask,                      \
 122                                   __FILE__, __FUNCTION__, __LINE__,           \
 123                                   CDEBUG_STACK, format, ## a);                \
 124 } while (0)
 125
 126 #define CWARN(format, a...) CDEBUG(D_WARNING, format, ## a)
 127 #define CERROR(format, a...) CDEBUG(D_ERROR, format, ## a)
 128 #define CEMERG(format, a...) CDEBUG(D_EMERG, format, ## a)
 129
 130 #define GOTO(label, rc)                                                 \
 131 do {                                                                    \
 132         long GOTO__ret = (long)(rc);                                    \
 133         CDEBUG(D_TRACE,"Process leaving via %s (rc=%lu : %ld : %lx)\n", \
 134                #label, (unsigned long)GOTO__ret, (signed long)GOTO__ret,\
 135                (signed long)GOTO__ret);                                 \
 136         goto label;                                                     \
 137 } while (0)
 138
 139 #define RETURN(rc)                                                      \
 140 do {                                                                    \
 141         typeof(rc) RETURN__ret = (rc);                                  \
 142         CDEBUG(D_TRACE, "Process leaving (rc=%lu : %ld : %lx)\n",       \
 143                (long)RETURN__ret, (long)RETURN__ret, (long)RETURN__ret);\
 144         return RETURN__ret;                                             \
 145 } while (0)
 146
 147 #define ENTRY                                                           \
 148 do {                                                                    \
 149         CDEBUG(D_TRACE, "Process entered\n");                           \
 150 } while (0)
 151
 152 #define EXIT                                                            \
 153 do {                                                                    \
 154         CDEBUG(D_TRACE, "Process leaving\n");                           \
 155 } while(0)
 156 #else
 157 #define CDEBUG(mask, format, a...)      do { } while (0)
 158 #define CWARN(format, a...)             do { } while (0)
 159 #define CERROR(format, a...)            printk("<3>" format, ## a)
 160 #define CEMERG(format, a...)            printk("<0>" format, ## a)
 161 #define GOTO(label, rc)                 do { (void)(rc); goto label; } while (0)
 162 #define RETURN(rc)                      return (rc)
 163 #define ENTRY                           do { } while (0)
 164 #define EXIT                            do { } while (0)
 165 #endif
 166
 167 #ifdef __KERNEL__
 168 # include <linux/vmalloc.h>
 169 # include <linux/time.h>
 170 # include <linux/slab.h>
 171 # include <linux/interrupt.h>
 172 # include <linux/highmem.h>
 173 # include <linux/module.h>
 174 # include <linux/version.h>
 175 # include <portals/lib-nal.h>
 176 # include <linux/smp_lock.h>
 177 # include <asm/atomic.h>
 178
 179 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 180 #define schedule_work schedule_task
 181 #define prepare_work(wq,cb,cbdata)                                            \
 182 do {                                                                          \
 183         INIT_TQUEUE((wq), 0, 0);                                              \
 184         PREPARE_TQUEUE((wq), (cb), (cbdata));                                 \
 185 } while (0)
 186
 187 #define PageUptodate Page_Uptodate
 188 #define our_recalc_sigpending(current) recalc_sigpending(current)
 189 #define num_online_cpus() smp_num_cpus
 190 static inline void our_cond_resched(void)
 191 {
 192         if (current->need_resched)
 193                schedule ();
 194 }
 195 #define work_struct_t       struct tq_struct
 196
 197 #else
 198
 199 #define prepare_work(wq,cb,cbdata)                                            \
 200 do {                                                                          \
 201         INIT_WORK((wq), (void *)(cb), (void *)(cbdata));                      \
 202 } while (0)
 203 #define wait_on_page wait_on_page_locked
 204 #define our_recalc_sigpending(current) recalc_sigpending()
 205 #define strtok(a,b) strpbrk(a, b)
 206 static inline void our_cond_resched(void)
 207 {
 208         cond_resched();
 209 }
 210 #define work_struct_t      struct work_struct
 211
 212 #endif /* LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0) */
 213
 214 #ifdef PORTAL_DEBUG
 215 extern void kportal_assertion_failed(char *expr, char *file, const char *func,
 216                                      const int line);
 217 #define LASSERT(e) ((e) ? 0 : kportal_assertion_failed( #e , __FILE__,  \
 218                                                         __FUNCTION__, __LINE__))
 219 /* it would be great to dump_stack() here, but some kernels
 220  * export it as show_stack() and I can't be bothered to
 221  * proprely engage in that dance right now */
 222 #define LASSERTF(cond, fmt...)                                                \
 223         do {                                                                  \
 224                 if (unlikely(!(cond))) {                                      \
 225                         portals_debug_msg(0, D_EMERG,  __FILE__, __FUNCTION__,\
 226                                           __LINE__,  CDEBUG_STACK,            \
 227                                           "ASSERTION(" #cond ") failed:" fmt);\
 228                         LBUG();                                               \
 229                 }                                                             \
 230         } while (0)
 231
 232 #else
 233 #define LASSERT(e)
 234 #define LASSERTF(cond, fmt...) do { } while (0)
 235 #endif
 236
 237 #ifdef __arch_um__
 238 #define LBUG_WITH_LOC(file, func, line)                                 \
 239 do {                                                                    \
 240         CEMERG("LBUG - trying to dump log to /tmp/lustre-log\n");       \
 241         portals_debug_dumplog();                                        \
 242         portals_run_lbug_upcall(file, func, line);                      \
 243         panic("LBUG");                                                  \
 244 } while (0)
 245 #else
 246 #define LBUG_WITH_LOC(file, func, line)                                 \
 247 do {                                                                    \
 248         CEMERG("LBUG\n");                                               \
 249         portals_debug_dumplog();                                        \
 250         portals_run_lbug_upcall(file, func, line);                      \
 251         set_task_state(current, TASK_UNINTERRUPTIBLE);                  \
 252         schedule();                                                     \
 253 } while (0)
 254 #endif /* __arch_um__ */
 255
 256 #define LBUG() LBUG_WITH_LOC(__FILE__, __FUNCTION__, __LINE__)
 257
 258 /*
 259  * Memory
 260  */
 261 #ifdef PORTAL_DEBUG
 262 extern atomic_t portal_kmemory;
 263
 264 # define portal_kmem_inc(ptr, size)                                           \
 265 do {                                                                          \
 266         atomic_add(size, &portal_kmemory);                                    \
 267 } while (0)
 268
 269 # define portal_kmem_dec(ptr, size) do {                                      \
 270         atomic_sub(size, &portal_kmemory);                                    \
 271 } while (0)
 272
 273 #else
 274 # define portal_kmem_inc(ptr, size) do {} while (0)
 275 # define portal_kmem_dec(ptr, size) do {} while (0)
 276 #endif /* PORTAL_DEBUG */
 277
 278 #define PORTAL_VMALLOC_SIZE        16384
 279
 280 #define PORTAL_ALLOC_GFP(ptr, size, mask)                                 \
 281 do {                                                                      \
 282         LASSERT(!in_interrupt() ||                                        \
 283                (size <= PORTAL_VMALLOC_SIZE && mask == GFP_ATOMIC));      \
 284         if ((size) > PORTAL_VMALLOC_SIZE)                                 \
 285                 (ptr) = vmalloc(size);                                    \
 286         else                                                              \
 287                 (ptr) = kmalloc((size), (mask));                          \
 288         if ((ptr) == NULL) {                                              \
 289                 CERROR("PORTALS: out of memory at %s:%d (tried to alloc '"\
 290                        #ptr "' = %d)\n", __FILE__, __LINE__, (int)(size));\
 291                 CERROR("PORTALS: %d total bytes allocated by portals\n",  \
 292                        atomic_read(&portal_kmemory));                     \
 293         } else {                                                          \
 294                 portal_kmem_inc((ptr), (size));                           \
 295                 memset((ptr), 0, (size));                                 \
 296         }                                                                 \
 297         CDEBUG(D_MALLOC, "kmalloced '" #ptr "': %d at %p (tot %d).\n",    \
 298                (int)(size), (ptr), atomic_read (&portal_kmemory));        \
 299 } while (0)
 300
 301 #define PORTAL_ALLOC(ptr, size) \
 302         PORTAL_ALLOC_GFP(ptr, size, GFP_NOFS)
 303
 304 #define PORTAL_ALLOC_ATOMIC(ptr, size) \
 305         PORTAL_ALLOC_GFP(ptr, size, GFP_ATOMIC)
 306
 307 #define PORTAL_FREE(ptr, size)                                          \
 308 do {                                                                    \
 309         int s = (size);                                                 \
 310         if ((ptr) == NULL) {                                            \
 311                 CERROR("PORTALS: free NULL '" #ptr "' (%d bytes) at "   \
 312                        "%s:%d\n", s, __FILE__, __LINE__);               \
 313                 break;                                                  \
 314         }                                                               \
 315         if (s > PORTAL_VMALLOC_SIZE)                                    \
 316                 vfree(ptr);                                             \
 317         else                                                            \
 318                 kfree(ptr);                                             \
 319         portal_kmem_dec((ptr), s);                                      \
 320         CDEBUG(D_MALLOC, "kfreed '" #ptr "': %d at %p (tot %d).\n",     \
 321                s, (ptr), atomic_read(&portal_kmemory));                 \
 322 } while (0)
 323
 324 /* ------------------------------------------------------------------- */
 325
 326 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 327
 328 #define PORTAL_SYMBOL_REGISTER(x) inter_module_register(#x, THIS_MODULE, &x)
 329 #define PORTAL_SYMBOL_UNREGISTER(x) inter_module_unregister(#x)
 330
 331 #define PORTAL_SYMBOL_GET(x) ((typeof(&x))inter_module_get(#x))
 332 #define PORTAL_SYMBOL_PUT(x) inter_module_put(#x)
 333
 334 #define PORTAL_MODULE_USE       MOD_INC_USE_COUNT
 335 #define PORTAL_MODULE_UNUSE     MOD_DEC_USE_COUNT
 336 #else
 337
 338 #define PORTAL_SYMBOL_REGISTER(x)
 339 #define PORTAL_SYMBOL_UNREGISTER(x)
 340
 341 #define PORTAL_SYMBOL_GET(x) symbol_get(x)
 342 #define PORTAL_SYMBOL_PUT(x) symbol_put(x)
 343
 344 #define PORTAL_MODULE_USE       try_module_get(THIS_MODULE)
 345 #define PORTAL_MODULE_UNUSE     module_put(THIS_MODULE)
 346
 347 #endif
 348
 349 /******************************************************************************/
 350 /* Kernel Portals Router interface */
 351
 352 typedef void (*kpr_fwd_callback_t)(void *arg, int error); // completion callback
 353
 354 /* space for routing targets to stash "stuff" in a forwarded packet */
 355 typedef union {
 356         long long        _alignment;
 357         void            *_space[16];            /* scale with CPU arch */
 358 } kprfd_scratch_t;
 359
 360 /* Kernel Portals Routing Forwarded message Descriptor */
 361 typedef struct {
 362         struct list_head     kprfd_list;        /* stash in queues (routing target can use) */
 363         ptl_nid_t            kprfd_target_nid;  /* final destination NID */
 364         ptl_nid_t            kprfd_gateway_nid; /* gateway NID */
 365         int                  kprfd_nob;         /* # message bytes (including header) */
 366         int                  kprfd_niov;        /* # message frags (including header) */
 367         struct iovec        *kprfd_iov;         /* message fragments */
 368         void                *kprfd_router_arg;  // originating NAL's router arg
 369         kpr_fwd_callback_t   kprfd_callback;    /* completion callback */
 370         void                *kprfd_callback_arg; /* completion callback arg */
 371         kprfd_scratch_t      kprfd_scratch;    // scratchpad for routing targets
 372 } kpr_fwd_desc_t;
 373
 374 typedef void  (*kpr_fwd_t)(void *arg, kpr_fwd_desc_t *fwd);
 375 typedef void  (*kpr_notify_t)(void *arg, ptl_nid_t peer, int alive);
 376
 377 /* NAL's routing interface (Kernel Portals Routing Nal Interface) */
 378 typedef const struct {
 379         int             kprni_nalid;    /* NAL's id */
 380         void           *kprni_arg;      /* Arg to pass when calling into NAL */
 381         kpr_fwd_t       kprni_fwd;      /* NAL's forwarding entrypoint */
 382         kpr_notify_t    kprni_notify;   /* NAL's notification entrypoint */
 383 } kpr_nal_interface_t;
 384
 385 /* Router's routing interface (Kernel Portals Routing Router Interface) */
 386 typedef const struct {
 387         /* register the calling NAL with the router and get back the handle for
 388          * subsequent calls */
 389         int     (*kprri_register) (kpr_nal_interface_t *nal_interface,
 390                                    void **router_arg);
 391
 392         /* ask the router to find a gateway that forwards to 'nid' and is a
 393          * peer of the calling NAL; assume caller will send 'nob' bytes of
 394          * payload there */
 395         int     (*kprri_lookup) (void *router_arg, ptl_nid_t nid, int nob,
 396                                  ptl_nid_t *gateway_nid);
 397
 398         /* hand a packet over to the router for forwarding */
 399         kpr_fwd_t kprri_fwd_start;
 400
 401         /* hand a packet back to the router for completion */
 402         void    (*kprri_fwd_done) (void *router_arg, kpr_fwd_desc_t *fwd,
 403                                    int error);
 404
 405         /* notify the router about peer state */
 406         void    (*kprri_notify) (void *router_arg, ptl_nid_t peer,
 407                                  int alive, time_t when);
 408
 409         /* the calling NAL is shutting down */
 410         void    (*kprri_shutdown) (void *router_arg);
 411
 412         /* deregister the calling NAL with the router */
 413         void    (*kprri_deregister) (void *router_arg);
 414
 415 } kpr_router_interface_t;
 416
 417 /* Convenient struct for NAL to stash router interface/args */
 418 typedef struct {
 419         kpr_router_interface_t  *kpr_interface;
 420         void                    *kpr_arg;
 421 } kpr_router_t;
 422
 423 /* Router's control interface (Kernel Portals Routing Control Interface) */
 424 typedef const struct {
 425         int     (*kprci_add_route)(int gateway_nal, ptl_nid_t gateway_nid,
 426                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
 427         int     (*kprci_del_route)(int gateway_nal, ptl_nid_t gateway_nid,
 428                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
 429         int     (*kprci_get_route)(int index, int *gateway_nal,
 430                                    ptl_nid_t *gateway,
 431                                    ptl_nid_t *lo_nid, ptl_nid_t *hi_nid,
 432                                    int *alive);
 433         int     (*kprci_notify)(int gateway_nal, ptl_nid_t gateway_nid,
 434                                 int alive, time_t when);
 435 } kpr_control_interface_t;
 436
 437 extern kpr_control_interface_t  kpr_control_interface;
 438 extern kpr_router_interface_t   kpr_router_interface;
 439
 440 static inline int
 441 kpr_register (kpr_router_t *router, kpr_nal_interface_t *nalif)
 442 {
 443         int    rc;
 444
 445         router->kpr_interface = PORTAL_SYMBOL_GET (kpr_router_interface);
 446         if (router->kpr_interface == NULL)
 447                 return (-ENOENT);
 448
 449         rc = (router->kpr_interface)->kprri_register (nalif, &router->kpr_arg);
 450         if (rc != 0)
 451                 router->kpr_interface = NULL;
 452
 453         PORTAL_SYMBOL_PUT (kpr_router_interface);
 454         return (rc);
 455 }
 456
 457 static inline int
 458 kpr_routing (kpr_router_t *router)
 459 {
 460         return (router->kpr_interface != NULL);
 461 }
 462
 463 static inline int
 464 kpr_lookup (kpr_router_t *router, ptl_nid_t nid, int nob, ptl_nid_t *gateway_nid)
 465 {
 466         if (!kpr_routing (router))
 467                 return (-ENETUNREACH);
 468
 469         return (router->kpr_interface->kprri_lookup(router->kpr_arg, nid, nob,
 470                                                     gateway_nid));
 471 }
 472
 473 static inline void
 474 kpr_fwd_init (kpr_fwd_desc_t *fwd, ptl_nid_t nid,
 475               int nob, int niov, struct iovec *iov,
 476               kpr_fwd_callback_t callback, void *callback_arg)
 477 {
 478         fwd->kprfd_target_nid   = nid;
 479         fwd->kprfd_gateway_nid  = nid;
 480         fwd->kprfd_nob          = nob;
 481         fwd->kprfd_niov         = niov;
 482         fwd->kprfd_iov          = iov;
 483         fwd->kprfd_callback     = callback;
 484         fwd->kprfd_callback_arg = callback_arg;
 485 }
 486
 487 static inline void
 488 kpr_fwd_start (kpr_router_t *router, kpr_fwd_desc_t *fwd)
 489 {
 490         if (!kpr_routing (router))
 491                 fwd->kprfd_callback (fwd->kprfd_callback_arg, -ENETUNREACH);
 492         else
 493                 router->kpr_interface->kprri_fwd_start (router->kpr_arg, fwd);
 494 }
 495
 496 static inline void
 497 kpr_fwd_done (kpr_router_t *router, kpr_fwd_desc_t *fwd, int error)
 498 {
 499         LASSERT (kpr_routing (router));
 500         router->kpr_interface->kprri_fwd_done (router->kpr_arg, fwd, error);
 501 }
 502
 503 static inline void
 504 kpr_notify (kpr_router_t *router,
 505             ptl_nid_t peer, int alive, time_t when)
 506 {
 507         if (!kpr_routing (router))
 508                 return;
 509
 510         router->kpr_interface->kprri_notify(router->kpr_arg, peer, alive, when);
 511 }
 512
 513 static inline void
 514 kpr_shutdown (kpr_router_t *router)
 515 {
 516         if (kpr_routing (router))
 517                 router->kpr_interface->kprri_shutdown (router->kpr_arg);
 518 }
 519
 520 static inline void
 521 kpr_deregister (kpr_router_t *router)
 522 {
 523         if (!kpr_routing (router))
 524                 return;
 525         router->kpr_interface->kprri_deregister (router->kpr_arg);
 526         router->kpr_interface = NULL;
 527 }
 528
 529 /******************************************************************************/
 530
 531 #ifdef PORTALS_PROFILING
 532 #define prof_enum(FOO) PROF__##FOO
 533 enum {
 534         prof_enum(our_recvmsg),
 535         prof_enum(our_sendmsg),
 536         prof_enum(socknal_recv),
 537         prof_enum(lib_parse),
 538         prof_enum(conn_list_walk),
 539         prof_enum(memcpy),
 540         prof_enum(lib_finalize),
 541         prof_enum(pingcli_time),
 542         prof_enum(gmnal_send),
 543         prof_enum(gmnal_recv),
 544         MAX_PROFS
 545 };
 546
 547 struct prof_ent {
 548         char *str;
 549         /* hrmph.  wrap-tastic. */
 550         u32       starts;
 551         u32       finishes;
 552         cycles_t  total_cycles;
 553         cycles_t  start;
 554         cycles_t  end;
 555 };
 556
 557 extern struct prof_ent prof_ents[MAX_PROFS];
 558
 559 #define PROF_START(FOO)                                         \
 560         do {                                                    \
 561                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
 562                 pe->starts++;                                   \
 563                 pe->start = get_cycles();                       \
 564         } while (0)
 565
 566 #define PROF_FINISH(FOO)                                        \
 567         do {                                                    \
 568                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
 569                 pe->finishes++;                                 \
 570                 pe->end = get_cycles();                         \
 571                 pe->total_cycles += (pe->end - pe->start);      \
 572         } while (0)
 573 #else /* !PORTALS_PROFILING */
 574 #define PROF_START(FOO) do {} while(0)
 575 #define PROF_FINISH(FOO) do {} while(0)
 576 #endif /* PORTALS_PROFILING */
 577
 578 /* debug.c */
 579 extern spinlock_t stack_backtrace_lock;
 580
 581 char *portals_debug_dumpstack(void);
 582 void portals_run_upcall(char **argv);
 583 void portals_run_lbug_upcall(char * file, const char *fn, const int line);
 584 void portals_debug_dumplog(void);
 585 int portals_debug_init(unsigned long bufsize);
 586 int portals_debug_cleanup(void);
 587 int portals_debug_clear_buffer(void);
 588 int portals_debug_mark_buffer(char *text);
 589 int portals_debug_set_daemon(unsigned int cmd, unsigned int length,
 590                              char *file, unsigned int size);
 591 __s32 portals_debug_copy_to_user(char *buf, unsigned long len);
 592 #if (__GNUC__)
 593 /* Use the special GNU C __attribute__ hack to have the compiler check the
 594  * printf style argument string against the actual argument count and
 595  * types.
 596  */
 597 #ifdef printf
 598 # warning printf has been defined as a macro...
 599 # undef printf
 600 #endif
 601 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
 602                        const int line, unsigned long stack,
 603                        char *format, ...)
 604         __attribute__ ((format (printf, 7, 8)));
 605 #else
 606 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
 607                        const int line, unsigned long stack,
 608                        const char *format, ...);
 609 #endif /* __GNUC__ */
 610 void portals_debug_set_level(unsigned int debug_level);
 611
 612 # define fprintf(a, format, b...) CDEBUG(D_OTHER, format , ## b)
 613 # define printf(format, b...) CDEBUG(D_OTHER, format , ## b)
 614 # define time(a) CURRENT_TIME
 615
 616 extern void kportal_daemonize (char *name);
 617 extern void kportal_blockallsigs (void);
 618
 619 #else  /* !__KERNEL__ */
 620 # include <stdio.h>
 621 # include <stdlib.h>
 622 #ifndef __CYGWIN__
 623 # include <stdint.h>
 624 #else
 625 # include <cygwin-ioctl.h>
 626 #endif
 627 # include <unistd.h>
 628 # include <time.h>
 629 # include <asm/types.h>
 630 # ifndef DEBUG_SUBSYSTEM
 631 #  define DEBUG_SUBSYSTEM S_UNDEFINED
 632 # endif
 633 # ifdef PORTAL_DEBUG
 634 #  undef NDEBUG
 635 #  include <assert.h>
 636 #  define LASSERT(e)     assert(e)
 637 #  define LASSERTF(cond, args...)     assert(cond)
 638 # else
 639 #  define LASSERT(e)
 640 #  define LASSERTF(cond, args...) do { } while (0)
 641 # endif
 642 # define printk(format, args...) printf (format, ## args)
 643 # define PORTAL_ALLOC(ptr, size) do { (ptr) = malloc(size); } while (0);
 644 # define PORTAL_FREE(a, b) do { free(a); } while (0);
 645 # define portals_debug_msg(subsys, mask, file, fn, line, stack, format, a...) \
 646     printf("%02x:%06x (@%lu %s:%s,l. %d %d %lu): " format,                    \
 647            (subsys), (mask), (long)time(0), file, fn, line,                   \
 648            getpid() , stack, ## a);
 649 #endif
 650
 651 /* support decl needed both by kernel and liblustre */
 652 char *portals_nid2str(int nal, ptl_nid_t nid, char *str);
 653
 654 #ifndef CURRENT_TIME
 655 # define CURRENT_TIME time(0)
 656 #endif
 657
 658 /******************************************************************************/
 659 /* Light-weight trace
 660  * Support for temporary event tracing with minimal Heisenberg effect. */
 661 #define LWT_SUPPORT  0
 662
 663 #define LWT_MEMORY   (64<<20)
 664 #define LWT_MAX_CPUS 4
 665
 666 typedef struct {
 667         cycles_t    lwte_when;
 668         char       *lwte_where;
 669         void       *lwte_task;
 670         long        lwte_p1;
 671         long        lwte_p2;
 672         long        lwte_p3;
 673         long        lwte_p4;
 674 #if BITS_PER_LONG > 32
 675         long        lwte_pad;
 676 #endif
 677 } lwt_event_t;
 678
 679 #if LWT_SUPPORT
 680 #ifdef __KERNEL__
 681 #define LWT_EVENTS_PER_PAGE (PAGE_SIZE / sizeof (lwt_event_t))
 682
 683 typedef struct _lwt_page {
 684         struct list_head     lwtp_list;
 685         struct page         *lwtp_page;
 686         lwt_event_t         *lwtp_events;
 687 } lwt_page_t;
 688
 689 typedef struct {
 690         int                lwtc_current_index;
 691         lwt_page_t        *lwtc_current_page;
 692 } lwt_cpu_t;
 693
 694 extern int       lwt_enabled;
 695 extern lwt_cpu_t lwt_cpus[];
 696
 697 extern int  lwt_init (void);
 698 extern void lwt_fini (void);
 699 extern int  lwt_lookup_string (int *size, char *knlptr,
 700                                char *usrptr, int usrsize);
 701 extern int  lwt_control (int enable, int clear);
 702 extern int  lwt_snapshot (cycles_t *now, int *ncpu, int *total_size,
 703                           void *user_ptr, int user_size);
 704
 705 /* Note that we _don't_ define LWT_EVENT at all if LWT_SUPPORT isn't set.
 706  * This stuff is meant for finding specific problems; it never stays in
 707  * production code... */
 708
 709 #define LWTSTR(n)       #n
 710 #define LWTWHERE(f,l)   f ":" LWTSTR(l)
 711
 712 #define LWT_EVENT(p1, p2, p3, p4)                                       \
 713 do {                                                                    \
 714         unsigned long    flags;                                         \
 715         lwt_cpu_t       *cpu;                                           \
 716         lwt_page_t      *p;                                             \
 717         lwt_event_t     *e;                                             \
 718                                                                         \
 719         local_irq_save (flags);                                         \
 720                                                                         \
 721         if (lwt_enabled) {                                              \
 722                 cpu = &lwt_cpus[smp_processor_id()];                    \
 723                 p = cpu->lwtc_current_page;                             \
 724                 e = &p->lwtp_events[cpu->lwtc_current_index++];         \
 725                                                                         \
 726                 if (cpu->lwtc_current_index >= LWT_EVENTS_PER_PAGE) {   \
 727                         cpu->lwtc_current_page =                        \
 728                                 list_entry (p->lwtp_list.next,          \
 729                                             lwt_page_t, lwtp_list);     \
 730                         cpu->lwtc_current_index = 0;                    \
 731                 }                                                       \
 732                                                                         \
 733                 e->lwte_when  = get_cycles();                           \
 734                 e->lwte_where = LWTWHERE(__FILE__,__LINE__);            \
 735                 e->lwte_task  = current;                                \
 736                 e->lwte_p1    = (long)(p1);                             \
 737                 e->lwte_p2    = (long)(p2);                             \
 738                 e->lwte_p3    = (long)(p3);                             \
 739                 e->lwte_p4    = (long)(p4);                             \
 740         }                                                               \
 741                                                                         \
 742         local_irq_restore (flags);                                      \
 743 } while (0)
 744 #else  /* __KERNEL__ */
 745 #define LWT_EVENT(p1,p2,p3,p4)     /* no userland implementation yet */
 746 #endif /* __KERNEL__ */
 747 #endif /* LWT_SUPPORT */
 748
 749 struct portals_device_userstate
 750 {
 751         int          pdu_memhog_pages;
 752         struct page *pdu_memhog_root_page;
 753 };
 754
 755 #include <linux/portals_lib.h>
 756
 757 /*
 758  * USER LEVEL STUFF BELOW
 759  */
 760
 761 #define PORTALS_CFG_VERSION 0x00010001;
 762
 763 struct portals_cfg {
 764         __u32 pcfg_version;
 765         __u32 pcfg_command;
 766
 767         __u32 pcfg_nal;
 768         __u32 pcfg_flags;
 769
 770         __u32 pcfg_gw_nal;
 771         __u64 pcfg_nid;
 772         __u64 pcfg_nid2;
 773         __u64 pcfg_nid3;
 774         __u32 pcfg_id;
 775         __u32 pcfg_misc;
 776         __u32 pcfg_fd;
 777         __u32 pcfg_count;
 778         __u32 pcfg_size;
 779         __u32 pcfg_wait;
 780
 781         __u32 pcfg_plen1; /* buffers in userspace */
 782         char *pcfg_pbuf1;
 783         __u32 pcfg_plen2; /* buffers in userspace */
 784         char *pcfg_pbuf2;
 785 };
 786
 787 #define PCFG_INIT(pcfg, cmd)                            \
 788 do {                                                    \
 789         memset(&pcfg, 0, sizeof(pcfg));                 \
 790         pcfg.pcfg_version = PORTALS_CFG_VERSION;        \
 791         pcfg.pcfg_command = (cmd);                      \
 792                                                         \
 793 } while (0)
 794
 795 #define PORTAL_IOCTL_VERSION 0x00010007
 796 #define PING_SYNC       0
 797 #define PING_ASYNC      1
 798
 799 struct portal_ioctl_data {
 800         __u32 ioc_len;
 801         __u32 ioc_version;
 802         __u64 ioc_nid;
 803         __u64 ioc_nid2;
 804         __u64 ioc_nid3;
 805         __u32 ioc_count;
 806         __u32 ioc_nal;
 807         __u32 ioc_nal_cmd;
 808         __u32 ioc_fd;
 809         __u32 ioc_id;
 810
 811         __u32 ioc_flags;
 812         __u32 ioc_size;
 813
 814         __u32 ioc_wait;
 815         __u32 ioc_timeout;
 816         __u32 ioc_misc;
 817
 818         __u32 ioc_inllen1;
 819         char *ioc_inlbuf1;
 820         __u32 ioc_inllen2;
 821         char *ioc_inlbuf2;
 822
 823         __u32 ioc_plen1; /* buffers in userspace */
 824         char *ioc_pbuf1;
 825         __u32 ioc_plen2; /* buffers in userspace */
 826         char *ioc_pbuf2;
 827
 828         char ioc_bulk[0];
 829 };
 830
 831 struct portal_ioctl_hdr {
 832         __u32 ioc_len;
 833         __u32 ioc_version;
 834 };
 835
 836 struct portals_debug_ioctl_data
 837 {
 838         struct portal_ioctl_hdr hdr;
 839         unsigned int subs;
 840         unsigned int debug;
 841 };
 842
 843 #define PORTAL_IOC_INIT(data)                           \
 844 do {                                                    \
 845         memset(&data, 0, sizeof(data));                 \
 846         data.ioc_version = PORTAL_IOCTL_VERSION;        \
 847         data.ioc_len = sizeof(data);                    \
 848 } while (0)
 849
 850 /* FIXME check conflict with lustre_lib.h */
 851 #define PTL_IOC_DEBUG_MASK             _IOWR('f', 250, long)
 852
 853 static inline int portal_ioctl_packlen(struct portal_ioctl_data *data)
 854 {
 855         int len = sizeof(*data);
 856         len += size_round(data->ioc_inllen1);
 857         len += size_round(data->ioc_inllen2);
 858         return len;
 859 }
 860
 861 static inline int portal_ioctl_is_invalid(struct portal_ioctl_data *data)
 862 {
 863         if (data->ioc_len > (1<<30)) {
 864                 CERROR ("PORTALS ioctl: ioc_len larger than 1<<30\n");
 865                 return 1;
 866         }
 867         if (data->ioc_inllen1 > (1<<30)) {
 868                 CERROR ("PORTALS ioctl: ioc_inllen1 larger than 1<<30\n");
 869                 return 1;
 870         }
 871         if (data->ioc_inllen2 > (1<<30)) {
 872                 CERROR ("PORTALS ioctl: ioc_inllen2 larger than 1<<30\n");
 873                 return 1;
 874         }
 875         if (data->ioc_inlbuf1 && !data->ioc_inllen1) {
 876                 CERROR ("PORTALS ioctl: inlbuf1 pointer but 0 length\n");
 877                 return 1;
 878         }
 879         if (data->ioc_inlbuf2 && !data->ioc_inllen2) {
 880                 CERROR ("PORTALS ioctl: inlbuf2 pointer but 0 length\n");
 881                 return 1;
 882         }
 883         if (data->ioc_pbuf1 && !data->ioc_plen1) {
 884                 CERROR ("PORTALS ioctl: pbuf1 pointer but 0 length\n");
 885                 return 1;
 886         }
 887         if (data->ioc_pbuf2 && !data->ioc_plen2) {
 888                 CERROR ("PORTALS ioctl: pbuf2 pointer but 0 length\n");
 889                 return 1;
 890         }
 891         if (data->ioc_plen1 && !data->ioc_pbuf1) {
 892                 CERROR ("PORTALS ioctl: plen1 nonzero but no pbuf1 pointer\n");
 893                 return 1;
 894         }
 895         if (data->ioc_plen2 && !data->ioc_pbuf2) {
 896                 CERROR ("PORTALS ioctl: plen2 nonzero but no pbuf2 pointer\n");
 897                 return 1;
 898         }
 899         if (portal_ioctl_packlen(data) != data->ioc_len ) {
 900                 CERROR ("PORTALS ioctl: packlen != ioc_len\n");
 901                 return 1;
 902         }
 903         if (data->ioc_inllen1 &&
 904             data->ioc_bulk[data->ioc_inllen1 - 1] != '\0') {
 905                 CERROR ("PORTALS ioctl: inlbuf1 not 0 terminated\n");
 906                 return 1;
 907         }
 908         if (data->ioc_inllen2 &&
 909             data->ioc_bulk[size_round(data->ioc_inllen1) +
 910                            data->ioc_inllen2 - 1] != '\0') {
 911                 CERROR ("PORTALS ioctl: inlbuf2 not 0 terminated\n");
 912                 return 1;
 913         }
 914         return 0;
 915 }
 916
 917 #ifndef __KERNEL__
 918 static inline int portal_ioctl_pack(struct portal_ioctl_data *data, char **pbuf,
 919                                     int max)
 920 {
 921         char *ptr;
 922         struct portal_ioctl_data *overlay;
 923         data->ioc_len = portal_ioctl_packlen(data);
 924         data->ioc_version = PORTAL_IOCTL_VERSION;
 925
 926         if (*pbuf && portal_ioctl_packlen(data) > max)
 927                 return 1;
 928         if (*pbuf == NULL) {
 929                 *pbuf = malloc(data->ioc_len);
 930         }
 931         if (!*pbuf)
 932                 return 1;
 933         overlay = (struct portal_ioctl_data *)*pbuf;
 934         memcpy(*pbuf, data, sizeof(*data));
 935
 936         ptr = overlay->ioc_bulk;
 937         if (data->ioc_inlbuf1)
 938                 LOGL(data->ioc_inlbuf1, data->ioc_inllen1, ptr);
 939         if (data->ioc_inlbuf2)
 940                 LOGL(data->ioc_inlbuf2, data->ioc_inllen2, ptr);
 941         if (portal_ioctl_is_invalid(overlay))
 942                 return 1;
 943
 944         return 0;
 945 }
 946 #else
 947 #include <asm/uaccess.h>
 948
 949 /* buffer MUST be at least the size of portal_ioctl_hdr */
 950 static inline int portal_ioctl_getdata(char *buf, char *end, void *arg)
 951 {
 952         struct portal_ioctl_hdr *hdr;
 953         struct portal_ioctl_data *data;
 954         int err;
 955         ENTRY;
 956
 957         hdr = (struct portal_ioctl_hdr *)buf;
 958         data = (struct portal_ioctl_data *)buf;
 959
 960         err = copy_from_user(buf, (void *)arg, sizeof(*hdr));
 961         if ( err ) {
 962                 EXIT;
 963                 return err;
 964         }
 965
 966         if (hdr->ioc_version != PORTAL_IOCTL_VERSION) {
 967                 CERROR ("PORTALS: version mismatch kernel vs application\n");
 968                 return -EINVAL;
 969         }
 970
 971         if (hdr->ioc_len + buf >= end) {
 972                 CERROR ("PORTALS: user buffer exceeds kernel buffer\n");
 973                 return -EINVAL;
 974         }
 975
 976
 977         if (hdr->ioc_len < sizeof(struct portal_ioctl_data)) {
 978                 CERROR ("PORTALS: user buffer too small for ioctl\n");
 979                 return -EINVAL;
 980         }
 981
 982         err = copy_from_user(buf, (void *)arg, hdr->ioc_len);
 983         if ( err ) {
 984                 EXIT;
 985                 return err;
 986         }
 987
 988         if (portal_ioctl_is_invalid(data)) {
 989                 CERROR ("PORTALS: ioctl not correctly formatted\n");
 990                 return -EINVAL;
 991         }
 992
 993         if (data->ioc_inllen1) {
 994                 data->ioc_inlbuf1 = &data->ioc_bulk[0];
 995         }
 996
 997         if (data->ioc_inllen2) {
 998                 data->ioc_inlbuf2 = &data->ioc_bulk[0] +
 999                         size_round(data->ioc_inllen1);
1000         }
1001
1002         EXIT;
1003         return 0;
1004 }
1005 #endif
1006
1007 /* ioctls for manipulating snapshots 30- */
1008 #define IOC_PORTAL_TYPE                   'e'
1009 #define IOC_PORTAL_MIN_NR                 30
1010
1011 #define IOC_PORTAL_PING                    _IOWR('e', 30, long)
1012 #define IOC_PORTAL_GET_DEBUG               _IOWR('e', 31, long)
1013 #define IOC_PORTAL_CLEAR_DEBUG             _IOWR('e', 32, long)
1014 #define IOC_PORTAL_MARK_DEBUG              _IOWR('e', 33, long)
1015 #define IOC_PORTAL_PANIC                   _IOWR('e', 34, long)
1016 #define IOC_PORTAL_NAL_CMD                 _IOWR('e', 35, long)
1017 #define IOC_PORTAL_GET_NID                 _IOWR('e', 36, long)
1018 #define IOC_PORTAL_FAIL_NID                _IOWR('e', 37, long)
1019 #define IOC_PORTAL_SET_DAEMON              _IOWR('e', 38, long)
1020 #define IOC_PORTAL_LWT_CONTROL             _IOWR('e', 39, long)
1021 #define IOC_PORTAL_LWT_SNAPSHOT            _IOWR('e', 40, long)
1022 #define IOC_PORTAL_LWT_LOOKUP_STRING       _IOWR('e', 41, long)
1023 #define IOC_PORTAL_MEMHOG                  _IOWR('e', 42, long)
1024 #define IOC_PORTAL_MAX_NR                             42
1025
1026 enum {
1027         QSWNAL    = 1,
1028         SOCKNAL   = 2,
1029         GMNAL     = 3,
1030         /*          4 unused */
1031         TCPNAL    = 5,
1032         SCIMACNAL = 6,
1033         ROUTER    = 7,
1034         IBNAL     = 8,
1035         NAL_ENUM_END_MARKER
1036 };
1037
1038 #ifdef __KERNEL__
1039 extern ptl_handle_ni_t  kqswnal_ni;
1040 extern ptl_handle_ni_t  ksocknal_ni;
1041 extern ptl_handle_ni_t  kgmnal_ni;
1042 extern ptl_handle_ni_t  kibnal_ni;
1043 extern ptl_handle_ni_t  kscimacnal_ni;
1044 #endif
1045
1046 #define PTL_NALFMT_SIZE         16
1047
1048 #define NAL_MAX_NR (NAL_ENUM_END_MARKER - 1)
1049
1050 #define NAL_CMD_REGISTER_PEER_FD     100
1051 #define NAL_CMD_CLOSE_CONNECTION     101
1052 #define NAL_CMD_REGISTER_MYNID       102
1053 #define NAL_CMD_PUSH_CONNECTION      103
1054 #define NAL_CMD_GET_CONN             104
1055 #define NAL_CMD_DEL_AUTOCONN         105
1056 #define NAL_CMD_ADD_AUTOCONN         106
1057 #define NAL_CMD_GET_AUTOCONN         107
1058 #define NAL_CMD_GET_TXDESC           108
1059 #define NAL_CMD_ADD_ROUTE            109
1060 #define NAL_CMD_DEL_ROUTE            110
1061 #define NAL_CMD_GET_ROUTE            111
1062 #define NAL_CMD_NOTIFY_ROUTER        112
1063
1064 enum {
1065         DEBUG_DAEMON_START       =  1,
1066         DEBUG_DAEMON_STOP        =  2,
1067         DEBUG_DAEMON_PAUSE       =  3,
1068         DEBUG_DAEMON_CONTINUE    =  4,
1069 };
1070
1071 /* XXX remove to lustre ASAP */
1072 struct lustre_peer {
1073         ptl_nid_t       peer_nid;
1074         ptl_handle_ni_t peer_ni;
1075 };
1076
1077
1078 /* module.c */
1079 typedef int (*nal_cmd_handler_t)(struct portals_cfg *, void * private);
1080 int kportal_nal_register(int nal, nal_cmd_handler_t handler, void * private);
1081 int kportal_nal_unregister(int nal);
1082
1083 enum cfg_record_type {
1084         PORTALS_CFG_TYPE = 1,
1085         LUSTRE_CFG_TYPE = 123,
1086 };
1087
1088 typedef int (*cfg_record_cb_t)(enum cfg_record_type, int len, void *data);
1089 int kportal_nal_cmd(struct portals_cfg *);
1090
1091 ptl_handle_ni_t *kportal_get_ni (int nal);
1092 void kportal_put_ni (int nal);
1093
1094 #ifdef __CYGWIN__
1095 # ifndef BITS_PER_LONG
1096 #  if (~0UL) == 0xffffffffUL
1097 #   define BITS_PER_LONG 32
1098 #  else
1099 #   define BITS_PER_LONG 64
1100 #  endif
1101 # endif
1102 #endif
1103
1104 #if defined(__x86_64__)
1105 # define LPU64 "%Lu"
1106 # define LPD64 "%Ld"
1107 # define LPX64 "%#Lx"
1108 # define LPSZ  "%lu"
1109 # define LPSSZ "%ld"
1110 #elif (BITS_PER_LONG == 32 || __WORDSIZE == 32)
1111 # define LPU64 "%Lu"
1112 # define LPD64 "%Ld"
1113 # define LPX64 "%#Lx"
1114 # define LPSZ  "%u"
1115 # define LPSSZ "%d"
1116 #elif (BITS_PER_LONG == 64 || __WORDSIZE == 64)
1117 # define LPU64 "%lu"
1118 # define LPD64 "%ld"
1119 # define LPX64 "%#lx"
1120 # define LPSZ  "%lu"
1121 # define LPSSZ "%ld"
1122 #endif
1123 #ifndef LPU64
1124 # error "No word size defined"
1125 #endif
1126
1127 #endif