lnet/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_ERROR,           \
 101                                           __FILE__, __FUNCTION__, __LINE__,   \
 102                                           (stack),                            \
 103                                           "maximum lustre stack %u\n",        \
 104                                           portal_stack = (stack));            \
 105                       /*panic("LBUG");*/                                      \
 106                 }                                                             \
 107         } while (0)
 108 #else /* __KERNEL__ */
 109 #define CHECK_STACK(stack) do { } while(0)
 110 #define CDEBUG_STACK (0L)
 111 #endif /* __KERNEL__ */
 112
 113 #if 1
 114 #define CDEBUG(mask, format, a...)                                            \
 115 do {                                                                          \
 116         if (portal_cerror == 0)                                               \
 117                 break;                                                        \
 118         CHECK_STACK(CDEBUG_STACK);                                            \
 119         if (!(mask) || ((mask) & (D_ERROR | D_EMERG | D_WARNING)) ||          \
 120             (portal_debug & (mask) &&                                         \
 121              portal_subsystem_debug & DEBUG_SUBSYSTEM))                       \
 122                 portals_debug_msg(DEBUG_SUBSYSTEM, mask,                      \
 123                                   __FILE__, __FUNCTION__, __LINE__,           \
 124                                   CDEBUG_STACK, format, ## a);                \
 125 } while (0)
 126
 127 #define CWARN(format, a...) CDEBUG(D_WARNING, format, ## a)
 128 #define CERROR(format, a...) CDEBUG(D_ERROR, format, ## a)
 129 #define CEMERG(format, a...) CDEBUG(D_EMERG, format, ## a)
 130
 131 #define GOTO(label, rc)                                                 \
 132 do {                                                                    \
 133         long GOTO__ret = (long)(rc);                                    \
 134         CDEBUG(D_TRACE,"Process leaving via %s (rc=%lu : %ld : %lx)\n", \
 135                #label, (unsigned long)GOTO__ret, (signed long)GOTO__ret,\
 136                (signed long)GOTO__ret);                                 \
 137         goto label;                                                     \
 138 } while (0)
 139
 140 #define RETURN(rc)                                                      \
 141 do {                                                                    \
 142         typeof(rc) RETURN__ret = (rc);                                  \
 143         CDEBUG(D_TRACE, "Process leaving (rc=%lu : %ld : %lx)\n",       \
 144                (long)RETURN__ret, (long)RETURN__ret, (long)RETURN__ret);\
 145         return RETURN__ret;                                             \
 146 } while (0)
 147
 148 #define ENTRY                                                           \
 149 do {                                                                    \
 150         CDEBUG(D_TRACE, "Process entered\n");                           \
 151 } while (0)
 152
 153 #define EXIT                                                            \
 154 do {                                                                    \
 155         CDEBUG(D_TRACE, "Process leaving\n");                           \
 156 } while(0)
 157 #else
 158 #define CDEBUG(mask, format, a...)      do { } while (0)
 159 #define CWARN(format, a...)             do { } while (0)
 160 #define CERROR(format, a...)            printk("<3>" format, ## a)
 161 #define CEMERG(format, a...)            printk("<0>" format, ## a)
 162 #define GOTO(label, rc)                 do { (void)(rc); goto label; } while (0)
 163 #define RETURN(rc)                      return (rc)
 164 #define ENTRY                           do { } while (0)
 165 #define EXIT                            do { } while (0)
 166 #endif
 167
 168 #ifdef __KERNEL__
 169 # include <linux/vmalloc.h>
 170 # include <linux/time.h>
 171 # include <linux/slab.h>
 172 # include <linux/interrupt.h>
 173 # include <linux/highmem.h>
 174 # include <linux/module.h>
 175 # include <linux/version.h>
 176 # include <portals/lib-nal.h>
 177 # include <linux/smp_lock.h>
 178 # include <asm/atomic.h>
 179
 180 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 181 #define schedule_work schedule_task
 182 #define prepare_work(wq,cb,cbdata)                                            \
 183 do {                                                                          \
 184         INIT_TQUEUE((wq), 0, 0);                                              \
 185         PREPARE_TQUEUE((wq), (cb), (cbdata));                                 \
 186 } while (0)
 187
 188 #define ll_invalidate_inode_pages invalidate_inode_pages
 189 #define PageUptodate Page_Uptodate
 190 #define our_recalc_sigpending(current) recalc_sigpending(current)
 191 #define num_online_cpus() smp_num_cpus
 192 static inline void our_cond_resched(void)
 193 {
 194         if (current->need_resched)
 195                schedule ();
 196 }
 197 #define work_struct_t       struct tq_struct
 198
 199 #else
 200
 201 #define prepare_work(wq,cb,cbdata)                                            \
 202 do {                                                                          \
 203         INIT_WORK((wq), (void *)(cb), (void *)(cbdata));                      \
 204 } while (0)
 205 #define ll_invalidate_inode_pages(inode) invalidate_inode_pages((inode)->i_mapping)
 206 #define wait_on_page wait_on_page_locked
 207 #define our_recalc_sigpending(current) recalc_sigpending()
 208 #define strtok(a,b) strpbrk(a, b)
 209 static inline void our_cond_resched(void)
 210 {
 211         cond_resched();
 212 }
 213 #define work_struct_t      struct work_struct
 214
 215 #endif /* LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0) */
 216
 217 #ifdef PORTAL_DEBUG
 218 extern void kportal_assertion_failed(char *expr, char *file, const char *func,
 219                                      const int line);
 220 #define LASSERT(e) ((e) ? 0 : kportal_assertion_failed( #e , __FILE__,  \
 221                                                         __FUNCTION__, __LINE__))
 222 /* it would be great to dump_stack() here, but some kernels
 223  * export it as show_stack() and I can't be bothered to
 224  * proprely engage in that dance right now */
 225 #define LASSERTF(cond, fmt...)                                                \
 226         do {                                                                  \
 227                 if (unlikely(!(cond))) {                                      \
 228                         portals_debug_msg(0, D_EMERG,  __FILE__, __FUNCTION__,\
 229                                           __LINE__,  CDEBUG_STACK,            \
 230                                           "ASSERTION(" #cond ") failed:" fmt);\
 231                         LBUG();                                               \
 232                 }                                                             \
 233         } while (0)
 234
 235 #else
 236 #define LASSERT(e)
 237 #define LASSERTF(cond, fmt...) do { } while (0)
 238 #endif
 239
 240 #ifdef __arch_um__
 241 #define LBUG_WITH_LOC(file, func, line)                                 \
 242 do {                                                                    \
 243         CEMERG("LBUG - trying to dump log to /tmp/lustre-log\n");       \
 244         portals_debug_dumplog();                                        \
 245         portals_run_lbug_upcall(file, func, line);                      \
 246         panic("LBUG");                                                  \
 247 } while (0)
 248 #else
 249 #define LBUG_WITH_LOC(file, func, line)                                 \
 250 do {                                                                    \
 251         CEMERG("LBUG\n");                                               \
 252         portals_debug_dumplog();                                        \
 253         portals_run_lbug_upcall(file, func, line);                      \
 254         set_task_state(current, TASK_UNINTERRUPTIBLE);                  \
 255         schedule();                                                     \
 256 } while (0)
 257 #endif /* __arch_um__ */
 258
 259 #define LBUG() LBUG_WITH_LOC(__FILE__, __FUNCTION__, __LINE__)
 260
 261 /*
 262  * Memory
 263  */
 264 #ifdef PORTAL_DEBUG
 265 extern atomic_t portal_kmemory;
 266
 267 # define portal_kmem_inc(ptr, size)                                           \
 268 do {                                                                          \
 269         atomic_add(size, &portal_kmemory);                                    \
 270 } while (0)
 271
 272 # define portal_kmem_dec(ptr, size) do {                                      \
 273         atomic_sub(size, &portal_kmemory);                                    \
 274 } while (0)
 275
 276 #else
 277 # define portal_kmem_inc(ptr, size) do {} while (0)
 278 # define portal_kmem_dec(ptr, size) do {} while (0)
 279 #endif /* PORTAL_DEBUG */
 280
 281 #define PORTAL_VMALLOC_SIZE        16384
 282
 283 #define PORTAL_ALLOC_GFP(ptr, size, mask)                                 \
 284 do {                                                                      \
 285         LASSERT (!in_interrupt());                                        \
 286         if ((size) > PORTAL_VMALLOC_SIZE)                                 \
 287                 (ptr) = vmalloc(size);                                    \
 288         else                                                              \
 289                 (ptr) = kmalloc((size), (mask));                          \
 290         if ((ptr) == NULL) {                                              \
 291                 CERROR("PORTALS: out of memory at %s:%d (tried to alloc '"\
 292                        #ptr "' = %d)\n", __FILE__, __LINE__, (int)(size));\
 293                 CERROR("PORTALS: %d total bytes allocated by portals\n",  \
 294                        atomic_read(&portal_kmemory));                     \
 295         } else {                                                          \
 296                 portal_kmem_inc((ptr), (size));                           \
 297                 memset((ptr), 0, (size));                                 \
 298         }                                                                 \
 299         CDEBUG(D_MALLOC, "kmalloced '" #ptr "': %d at %p (tot %d).\n",    \
 300                (int)(size), (ptr), atomic_read (&portal_kmemory));        \
 301 } while (0)
 302
 303 #define PORTAL_ALLOC(ptr, size) \
 304         PORTAL_ALLOC_GFP(ptr, size, GFP_NOFS)
 305
 306 #define PORTAL_ALLOC_ATOMIC(ptr, size) \
 307         PORTAL_ALLOC_GFP(ptr, size, GFP_ATOMIC)
 308
 309 #define PORTAL_FREE(ptr, size)                                          \
 310 do {                                                                    \
 311         int s = (size);                                                 \
 312         if ((ptr) == NULL) {                                            \
 313                 CERROR("PORTALS: free NULL '" #ptr "' (%d bytes) at "   \
 314                        "%s:%d\n", s, __FILE__, __LINE__);               \
 315                 break;                                                  \
 316         }                                                               \
 317         if (s > PORTAL_VMALLOC_SIZE)                                    \
 318                 vfree(ptr);                                             \
 319         else                                                            \
 320                 kfree(ptr);                                             \
 321         portal_kmem_dec((ptr), s);                                      \
 322         CDEBUG(D_MALLOC, "kfreed '" #ptr "': %d at %p (tot %d).\n",     \
 323                s, (ptr), atomic_read(&portal_kmemory));                 \
 324 } while (0)
 325
 326 /* ------------------------------------------------------------------- */
 327
 328 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 329
 330 #define PORTAL_SYMBOL_REGISTER(x) inter_module_register(#x, THIS_MODULE, &x)
 331 #define PORTAL_SYMBOL_UNREGISTER(x) inter_module_unregister(#x)
 332
 333 #define PORTAL_SYMBOL_GET(x) ((typeof(&x))inter_module_get(#x))
 334 #define PORTAL_SYMBOL_PUT(x) inter_module_put(#x)
 335
 336 #define PORTAL_MODULE_USE       MOD_INC_USE_COUNT
 337 #define PORTAL_MODULE_UNUSE     MOD_DEC_USE_COUNT
 338 #else
 339
 340 #define PORTAL_SYMBOL_REGISTER(x)
 341 #define PORTAL_SYMBOL_UNREGISTER(x)
 342
 343 #define PORTAL_SYMBOL_GET(x) symbol_get(x)
 344 #define PORTAL_SYMBOL_PUT(x) symbol_put(x)
 345
 346 #define PORTAL_MODULE_USE       try_module_get(THIS_MODULE)
 347 #define PORTAL_MODULE_UNUSE     module_put(THIS_MODULE)
 348
 349 #endif
 350
 351 /******************************************************************************/
 352 /* Kernel Portals Router interface */
 353
 354 typedef void (*kpr_fwd_callback_t)(void *arg, int error); // completion callback
 355
 356 /* space for routing targets to stash "stuff" in a forwarded packet */
 357 typedef union {
 358         long long        _alignment;
 359         void            *_space[16];            /* scale with CPU arch */
 360 } kprfd_scratch_t;
 361
 362 /* Kernel Portals Routing Forwarded message Descriptor */
 363 typedef struct {
 364         struct list_head     kprfd_list;        /* stash in queues (routing target can use) */
 365         ptl_nid_t            kprfd_target_nid;  /* final destination NID */
 366         ptl_nid_t            kprfd_gateway_nid; /* gateway NID */
 367         int                  kprfd_nob;         /* # message bytes (including header) */
 368         int                  kprfd_niov;        /* # message frags (including header) */
 369         struct iovec        *kprfd_iov;         /* message fragments */
 370         void                *kprfd_router_arg;  // originating NAL's router arg
 371         kpr_fwd_callback_t   kprfd_callback;    /* completion callback */
 372         void                *kprfd_callback_arg; /* completion callback arg */
 373         kprfd_scratch_t      kprfd_scratch;    // scratchpad for routing targets
 374 } kpr_fwd_desc_t;
 375
 376 typedef void  (*kpr_fwd_t)(void *arg, kpr_fwd_desc_t *fwd);
 377 typedef void  (*kpr_notify_t)(void *arg, ptl_nid_t peer, int alive);
 378
 379 /* NAL's routing interface (Kernel Portals Routing Nal Interface) */
 380 typedef const struct {
 381         int             kprni_nalid;    /* NAL's id */
 382         void           *kprni_arg;      /* Arg to pass when calling into NAL */
 383         kpr_fwd_t       kprni_fwd;      /* NAL's forwarding entrypoint */
 384         kpr_notify_t    kprni_notify;   /* NAL's notification entrypoint */
 385 } kpr_nal_interface_t;
 386
 387 /* Router's routing interface (Kernel Portals Routing Router Interface) */
 388 typedef const struct {
 389         /* register the calling NAL with the router and get back the handle for
 390          * subsequent calls */
 391         int     (*kprri_register) (kpr_nal_interface_t *nal_interface,
 392                                    void **router_arg);
 393
 394         /* ask the router to find a gateway that forwards to 'nid' and is a
 395          * peer of the calling NAL; assume caller will send 'nob' bytes of
 396          * payload there */
 397         int     (*kprri_lookup) (void *router_arg, ptl_nid_t nid, int nob,
 398                                  ptl_nid_t *gateway_nid);
 399
 400         /* hand a packet over to the router for forwarding */
 401         kpr_fwd_t kprri_fwd_start;
 402
 403         /* hand a packet back to the router for completion */
 404         void    (*kprri_fwd_done) (void *router_arg, kpr_fwd_desc_t *fwd,
 405                                    int error);
 406
 407         /* notify the router about peer state */
 408         void    (*kprri_notify) (void *router_arg, ptl_nid_t peer,
 409                                  int alive, time_t when);
 410
 411         /* the calling NAL is shutting down */
 412         void    (*kprri_shutdown) (void *router_arg);
 413
 414         /* deregister the calling NAL with the router */
 415         void    (*kprri_deregister) (void *router_arg);
 416
 417 } kpr_router_interface_t;
 418
 419 /* Convenient struct for NAL to stash router interface/args */
 420 typedef struct {
 421         kpr_router_interface_t  *kpr_interface;
 422         void                    *kpr_arg;
 423 } kpr_router_t;
 424
 425 /* Router's control interface (Kernel Portals Routing Control Interface) */
 426 typedef const struct {
 427         int     (*kprci_add_route)(int gateway_nal, ptl_nid_t gateway_nid,
 428                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
 429         int     (*kprci_del_route)(int gateway_nal, ptl_nid_t gateway_nid,
 430                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
 431         int     (*kprci_get_route)(int index, int *gateway_nal,
 432                                    ptl_nid_t *gateway,
 433                                    ptl_nid_t *lo_nid, ptl_nid_t *hi_nid,
 434                                    int *alive);
 435         int     (*kprci_notify)(int gateway_nal, ptl_nid_t gateway_nid,
 436                                 int alive, time_t when);
 437 } kpr_control_interface_t;
 438
 439 extern kpr_control_interface_t  kpr_control_interface;
 440 extern kpr_router_interface_t   kpr_router_interface;
 441
 442 static inline int
 443 kpr_register (kpr_router_t *router, kpr_nal_interface_t *nalif)
 444 {
 445         int    rc;
 446
 447         router->kpr_interface = PORTAL_SYMBOL_GET (kpr_router_interface);
 448         if (router->kpr_interface == NULL)
 449                 return (-ENOENT);
 450
 451         rc = (router->kpr_interface)->kprri_register (nalif, &router->kpr_arg);
 452         if (rc != 0)
 453                 router->kpr_interface = NULL;
 454
 455         PORTAL_SYMBOL_PUT (kpr_router_interface);
 456         return (rc);
 457 }
 458
 459 static inline int
 460 kpr_routing (kpr_router_t *router)
 461 {
 462         return (router->kpr_interface != NULL);
 463 }
 464
 465 static inline int
 466 kpr_lookup (kpr_router_t *router, ptl_nid_t nid, int nob, ptl_nid_t *gateway_nid)
 467 {
 468         if (!kpr_routing (router))
 469                 return (-ENETUNREACH);
 470
 471         return (router->kpr_interface->kprri_lookup(router->kpr_arg, nid, nob,
 472                                                     gateway_nid));
 473 }
 474
 475 static inline void
 476 kpr_fwd_init (kpr_fwd_desc_t *fwd, ptl_nid_t nid,
 477               int nob, int niov, struct iovec *iov,
 478               kpr_fwd_callback_t callback, void *callback_arg)
 479 {
 480         fwd->kprfd_target_nid   = nid;
 481         fwd->kprfd_gateway_nid  = nid;
 482         fwd->kprfd_nob          = nob;
 483         fwd->kprfd_niov         = niov;
 484         fwd->kprfd_iov          = iov;
 485         fwd->kprfd_callback     = callback;
 486         fwd->kprfd_callback_arg = callback_arg;
 487 }
 488
 489 static inline void
 490 kpr_fwd_start (kpr_router_t *router, kpr_fwd_desc_t *fwd)
 491 {
 492         if (!kpr_routing (router))
 493                 fwd->kprfd_callback (fwd->kprfd_callback_arg, -ENETUNREACH);
 494         else
 495                 router->kpr_interface->kprri_fwd_start (router->kpr_arg, fwd);
 496 }
 497
 498 static inline void
 499 kpr_fwd_done (kpr_router_t *router, kpr_fwd_desc_t *fwd, int error)
 500 {
 501         LASSERT (kpr_routing (router));
 502         router->kpr_interface->kprri_fwd_done (router->kpr_arg, fwd, error);
 503 }
 504
 505 static inline void
 506 kpr_notify (kpr_router_t *router,
 507             ptl_nid_t peer, int alive, time_t when)
 508 {
 509         if (!kpr_routing (router))
 510                 return;
 511
 512         router->kpr_interface->kprri_notify(router->kpr_arg, peer, alive, when);
 513 }
 514
 515 static inline void
 516 kpr_shutdown (kpr_router_t *router)
 517 {
 518         if (kpr_routing (router))
 519                 router->kpr_interface->kprri_shutdown (router->kpr_arg);
 520 }
 521
 522 static inline void
 523 kpr_deregister (kpr_router_t *router)
 524 {
 525         if (!kpr_routing (router))
 526                 return;
 527         router->kpr_interface->kprri_deregister (router->kpr_arg);
 528         router->kpr_interface = NULL;
 529 }
 530
 531 /******************************************************************************/
 532
 533 #ifdef PORTALS_PROFILING
 534 #define prof_enum(FOO) PROF__##FOO
 535 enum {
 536         prof_enum(our_recvmsg),
 537         prof_enum(our_sendmsg),
 538         prof_enum(socknal_recv),
 539         prof_enum(lib_parse),
 540         prof_enum(conn_list_walk),
 541         prof_enum(memcpy),
 542         prof_enum(lib_finalize),
 543         prof_enum(pingcli_time),
 544         prof_enum(gmnal_send),
 545         prof_enum(gmnal_recv),
 546         MAX_PROFS
 547 };
 548
 549 struct prof_ent {
 550         char *str;
 551         /* hrmph.  wrap-tastic. */
 552         u32       starts;
 553         u32       finishes;
 554         cycles_t  total_cycles;
 555         cycles_t  start;
 556         cycles_t  end;
 557 };
 558
 559 extern struct prof_ent prof_ents[MAX_PROFS];
 560
 561 #define PROF_START(FOO)                                         \
 562         do {                                                    \
 563                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
 564                 pe->starts++;                                   \
 565                 pe->start = get_cycles();                       \
 566         } while (0)
 567
 568 #define PROF_FINISH(FOO)                                        \
 569         do {                                                    \
 570                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
 571                 pe->finishes++;                                 \
 572                 pe->end = get_cycles();                         \
 573                 pe->total_cycles += (pe->end - pe->start);      \
 574         } while (0)
 575 #else /* !PORTALS_PROFILING */
 576 #define PROF_START(FOO) do {} while(0)
 577 #define PROF_FINISH(FOO) do {} while(0)
 578 #endif /* PORTALS_PROFILING */
 579
 580 /* debug.c */
 581 extern spinlock_t stack_backtrace_lock;
 582
 583 char *portals_debug_dumpstack(void);
 584 void portals_run_upcall(char **argv);
 585 void portals_run_lbug_upcall(char * file, const char *fn, const int line);
 586 void portals_debug_dumplog(void);
 587 int portals_debug_init(unsigned long bufsize);
 588 int portals_debug_cleanup(void);
 589 int portals_debug_clear_buffer(void);
 590 int portals_debug_mark_buffer(char *text);
 591 int portals_debug_set_daemon(unsigned int cmd, unsigned int length,
 592                              char *file, unsigned int size);
 593 __s32 portals_debug_copy_to_user(char *buf, unsigned long len);
 594 #if (__GNUC__)
 595 /* Use the special GNU C __attribute__ hack to have the compiler check the
 596  * printf style argument string against the actual argument count and
 597  * types.
 598  */
 599 #ifdef printf
 600 # warning printf has been defined as a macro...
 601 # undef printf
 602 #endif
 603 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
 604                        const int line, unsigned long stack,
 605                        char *format, ...)
 606         __attribute__ ((format (printf, 7, 8)));
 607 #else
 608 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
 609                        const int line, unsigned long stack,
 610                        const char *format, ...);
 611 #endif /* __GNUC__ */
 612 void portals_debug_set_level(unsigned int debug_level);
 613
 614 # define fprintf(a, format, b...) CDEBUG(D_OTHER, format , ## b)
 615 # define printf(format, b...) CDEBUG(D_OTHER, format , ## b)
 616 # define time(a) CURRENT_TIME
 617
 618 extern void kportal_daemonize (char *name);
 619 extern void kportal_blockallsigs (void);
 620
 621 #else  /* !__KERNEL__ */
 622 # include <stdio.h>
 623 # include <stdlib.h>
 624 #ifndef __CYGWIN__
 625 # include <stdint.h>
 626 #else
 627 # include <cygwin-ioctl.h>
 628 #endif
 629 # include <unistd.h>
 630 # include <time.h>
 631 # include <asm/types.h>
 632 # ifndef DEBUG_SUBSYSTEM
 633 #  define DEBUG_SUBSYSTEM S_UNDEFINED
 634 # endif
 635 # ifdef PORTAL_DEBUG
 636 #  undef NDEBUG
 637 #  include <assert.h>
 638 #  define LASSERT(e)     assert(e)
 639 #  define LASSERTF(cond, args...)     assert(cond)
 640 # else
 641 #  define LASSERT(e)
 642 #  define LASSERTF(cond, args...) do { } while (0)
 643 # endif
 644 # define printk(format, args...) printf (format, ## args)
 645 # define PORTAL_ALLOC(ptr, size) do { (ptr) = malloc(size); } while (0);
 646 # define PORTAL_FREE(a, b) do { free(a); } while (0);
 647 # define portals_debug_msg(subsys, mask, file, fn, line, stack, format, a...) \
 648     printf("%02x:%06x (@%lu %s:%s,l. %d %d %lu): " format,                    \
 649            (subsys), (mask), (long)time(0), file, fn, line,                   \
 650            getpid() , stack, ## a);
 651 #endif
 652
 653 /* support decl needed both by kernel and liblustre */
 654 char *portals_nid2str(int nal, ptl_nid_t nid, char *str);
 655
 656 #ifndef CURRENT_TIME
 657 # define CURRENT_TIME time(0)
 658 #endif
 659
 660 /******************************************************************************/
 661 /* Light-weight trace
 662  * Support for temporary event tracing with minimal Heisenberg effect. */
 663 #define LWT_SUPPORT  0
 664
 665 #define LWT_MEMORY   (64<<20)
 666 #define LWT_MAX_CPUS 4
 667
 668 typedef struct {
 669         cycles_t    lwte_when;
 670         char       *lwte_where;
 671         void       *lwte_task;
 672         long        lwte_p1;
 673         long        lwte_p2;
 674         long        lwte_p3;
 675         long        lwte_p4;
 676 #if BITS_PER_LONG > 32
 677         long        lwte_pad;
 678 #endif
 679 } lwt_event_t;
 680
 681 #if LWT_SUPPORT
 682 #ifdef __KERNEL__
 683 #define LWT_EVENTS_PER_PAGE (PAGE_SIZE / sizeof (lwt_event_t))
 684
 685 typedef struct _lwt_page {
 686         struct list_head     lwtp_list;
 687         struct page         *lwtp_page;
 688         lwt_event_t         *lwtp_events;
 689 } lwt_page_t;
 690
 691 typedef struct {
 692         int                lwtc_current_index;
 693         lwt_page_t        *lwtc_current_page;
 694 } lwt_cpu_t;
 695
 696 extern int       lwt_enabled;
 697 extern lwt_cpu_t lwt_cpus[];
 698
 699 extern int  lwt_init (void);
 700 extern void lwt_fini (void);
 701 extern int  lwt_lookup_string (int *size, char *knlptr,
 702                                char *usrptr, int usrsize);
 703 extern int  lwt_control (int enable, int clear);
 704 extern int  lwt_snapshot (cycles_t *now, int *ncpu, int *total_size,
 705                           void *user_ptr, int user_size);
 706
 707 /* Note that we _don't_ define LWT_EVENT at all if LWT_SUPPORT isn't set.
 708  * This stuff is meant for finding specific problems; it never stays in
 709  * production code... */
 710
 711 #define LWTSTR(n)       #n
 712 #define LWTWHERE(f,l)   f ":" LWTSTR(l)
 713
 714 #define LWT_EVENT(p1, p2, p3, p4)                                       \
 715 do {                                                                    \
 716         unsigned long    flags;                                         \
 717         lwt_cpu_t       *cpu;                                           \
 718         lwt_page_t      *p;                                             \
 719         lwt_event_t     *e;                                             \
 720                                                                         \
 721         local_irq_save (flags);                                         \
 722                                                                         \
 723         if (lwt_enabled) {                                              \
 724                 cpu = &lwt_cpus[smp_processor_id()];                    \
 725                 p = cpu->lwtc_current_page;                             \
 726                 e = &p->lwtp_events[cpu->lwtc_current_index++];         \
 727                                                                         \
 728                 if (cpu->lwtc_current_index >= LWT_EVENTS_PER_PAGE) {   \
 729                         cpu->lwtc_current_page =                        \
 730                                 list_entry (p->lwtp_list.next,          \
 731                                             lwt_page_t, lwtp_list);     \
 732                         cpu->lwtc_current_index = 0;                    \
 733                 }                                                       \
 734                                                                         \
 735                 e->lwte_when  = get_cycles();                           \
 736                 e->lwte_where = LWTWHERE(__FILE__,__LINE__);            \
 737                 e->lwte_task  = current;                                \
 738                 e->lwte_p1    = (long)(p1);                             \
 739                 e->lwte_p2    = (long)(p2);                             \
 740                 e->lwte_p3    = (long)(p3);                             \
 741                 e->lwte_p4    = (long)(p4);                             \
 742         }                                                               \
 743                                                                         \
 744         local_irq_restore (flags);                                      \
 745 } while (0)
 746 #else  /* __KERNEL__ */
 747 #define LWT_EVENT(p1,p2,p3,p4)     /* no userland implementation yet */
 748 #endif /* __KERNEL__ */
 749 #endif /* LWT_SUPPORT */
 750
 751 struct portals_device_userstate
 752 {
 753         int          pdu_memhog_pages;
 754         struct page *pdu_memhog_root_page;
 755 };
 756
 757 #include <linux/portals_lib.h>
 758
 759 /*
 760  * USER LEVEL STUFF BELOW
 761  */
 762
 763 #define PORTALS_CFG_VERSION 0x00010001;
 764
 765 struct portals_cfg {
 766         __u32 pcfg_version;
 767         __u32 pcfg_command;
 768
 769         __u32 pcfg_nal;
 770         __u32 pcfg_flags;
 771
 772         __u32 pcfg_gw_nal;
 773         __u64 pcfg_nid;
 774         __u64 pcfg_nid2;
 775         __u64 pcfg_nid3;
 776         __u32 pcfg_id;
 777         __u32 pcfg_misc;
 778         __u32 pcfg_fd;
 779         __u32 pcfg_count;
 780         __u32 pcfg_size;
 781         __u32 pcfg_wait;
 782
 783         __u32 pcfg_plen1; /* buffers in userspace */
 784         char *pcfg_pbuf1;
 785         __u32 pcfg_plen2; /* buffers in userspace */
 786         char *pcfg_pbuf2;
 787 };
 788
 789 #define PCFG_INIT(pcfg, cmd)                            \
 790 do {                                                    \
 791         memset(&pcfg, 0, sizeof(pcfg));                 \
 792         pcfg.pcfg_version = PORTALS_CFG_VERSION;        \
 793         pcfg.pcfg_command = (cmd);                      \
 794                                                         \
 795 } while (0)
 796
 797 #define PORTAL_IOCTL_VERSION 0x00010007
 798 #define PING_SYNC       0
 799 #define PING_ASYNC      1
 800
 801 struct portal_ioctl_data {
 802         __u32 ioc_len;
 803         __u32 ioc_version;
 804         __u64 ioc_nid;
 805         __u64 ioc_nid2;
 806         __u64 ioc_nid3;
 807         __u32 ioc_count;
 808         __u32 ioc_nal;
 809         __u32 ioc_nal_cmd;
 810         __u32 ioc_fd;
 811         __u32 ioc_id;
 812
 813         __u32 ioc_flags;
 814         __u32 ioc_size;
 815
 816         __u32 ioc_wait;
 817         __u32 ioc_timeout;
 818         __u32 ioc_misc;
 819
 820         __u32 ioc_inllen1;
 821         char *ioc_inlbuf1;
 822         __u32 ioc_inllen2;
 823         char *ioc_inlbuf2;
 824
 825         __u32 ioc_plen1; /* buffers in userspace */
 826         char *ioc_pbuf1;
 827         __u32 ioc_plen2; /* buffers in userspace */
 828         char *ioc_pbuf2;
 829
 830         char ioc_bulk[0];
 831 };
 832
 833 struct portal_ioctl_hdr {
 834         __u32 ioc_len;
 835         __u32 ioc_version;
 836 };
 837
 838 struct portals_debug_ioctl_data
 839 {
 840         struct portal_ioctl_hdr hdr;
 841         unsigned int subs;
 842         unsigned int debug;
 843 };
 844
 845 #define PORTAL_IOC_INIT(data)                           \
 846 do {                                                    \
 847         memset(&data, 0, sizeof(data));                 \
 848         data.ioc_version = PORTAL_IOCTL_VERSION;        \
 849         data.ioc_len = sizeof(data);                    \
 850 } while (0)
 851
 852 /* FIXME check conflict with lustre_lib.h */
 853 #define PTL_IOC_DEBUG_MASK             _IOWR('f', 250, long)
 854
 855 static inline int portal_ioctl_packlen(struct portal_ioctl_data *data)
 856 {
 857         int len = sizeof(*data);
 858         len += size_round(data->ioc_inllen1);
 859         len += size_round(data->ioc_inllen2);
 860         return len;
 861 }
 862
 863 static inline int portal_ioctl_is_invalid(struct portal_ioctl_data *data)
 864 {
 865         if (data->ioc_len > (1<<30)) {
 866                 CERROR ("PORTALS ioctl: ioc_len larger than 1<<30\n");
 867                 return 1;
 868         }
 869         if (data->ioc_inllen1 > (1<<30)) {
 870                 CERROR ("PORTALS ioctl: ioc_inllen1 larger than 1<<30\n");
 871                 return 1;
 872         }
 873         if (data->ioc_inllen2 > (1<<30)) {
 874                 CERROR ("PORTALS ioctl: ioc_inllen2 larger than 1<<30\n");
 875                 return 1;
 876         }
 877         if (data->ioc_inlbuf1 && !data->ioc_inllen1) {
 878                 CERROR ("PORTALS ioctl: inlbuf1 pointer but 0 length\n");
 879                 return 1;
 880         }
 881         if (data->ioc_inlbuf2 && !data->ioc_inllen2) {
 882                 CERROR ("PORTALS ioctl: inlbuf2 pointer but 0 length\n");
 883                 return 1;
 884         }
 885         if (data->ioc_pbuf1 && !data->ioc_plen1) {
 886                 CERROR ("PORTALS ioctl: pbuf1 pointer but 0 length\n");
 887                 return 1;
 888         }
 889         if (data->ioc_pbuf2 && !data->ioc_plen2) {
 890                 CERROR ("PORTALS ioctl: pbuf2 pointer but 0 length\n");
 891                 return 1;
 892         }
 893         if (data->ioc_plen1 && !data->ioc_pbuf1) {
 894                 CERROR ("PORTALS ioctl: plen1 nonzero but no pbuf1 pointer\n");
 895                 return 1;
 896         }
 897         if (data->ioc_plen2 && !data->ioc_pbuf2) {
 898                 CERROR ("PORTALS ioctl: plen2 nonzero but no pbuf2 pointer\n");
 899                 return 1;
 900         }
 901         if (portal_ioctl_packlen(data) != data->ioc_len ) {
 902                 CERROR ("PORTALS ioctl: packlen != ioc_len\n");
 903                 return 1;
 904         }
 905         if (data->ioc_inllen1 &&
 906             data->ioc_bulk[data->ioc_inllen1 - 1] != '\0') {
 907                 CERROR ("PORTALS ioctl: inlbuf1 not 0 terminated\n");
 908                 return 1;
 909         }
 910         if (data->ioc_inllen2 &&
 911             data->ioc_bulk[size_round(data->ioc_inllen1) +
 912                            data->ioc_inllen2 - 1] != '\0') {
 913                 CERROR ("PORTALS ioctl: inlbuf2 not 0 terminated\n");
 914                 return 1;
 915         }
 916         return 0;
 917 }
 918
 919 #ifndef __KERNEL__
 920 static inline int portal_ioctl_pack(struct portal_ioctl_data *data, char **pbuf,
 921                                     int max)
 922 {
 923         char *ptr;
 924         struct portal_ioctl_data *overlay;
 925         data->ioc_len = portal_ioctl_packlen(data);
 926         data->ioc_version = PORTAL_IOCTL_VERSION;
 927
 928         if (*pbuf && portal_ioctl_packlen(data) > max)
 929                 return 1;
 930         if (*pbuf == NULL) {
 931                 *pbuf = malloc(data->ioc_len);
 932         }
 933         if (!*pbuf)
 934                 return 1;
 935         overlay = (struct portal_ioctl_data *)*pbuf;
 936         memcpy(*pbuf, data, sizeof(*data));
 937
 938         ptr = overlay->ioc_bulk;
 939         if (data->ioc_inlbuf1)
 940                 LOGL(data->ioc_inlbuf1, data->ioc_inllen1, ptr);
 941         if (data->ioc_inlbuf2)
 942                 LOGL(data->ioc_inlbuf2, data->ioc_inllen2, ptr);
 943         if (portal_ioctl_is_invalid(overlay))
 944                 return 1;
 945
 946         return 0;
 947 }
 948 #else
 949 #include <asm/uaccess.h>
 950
 951 /* buffer MUST be at least the size of portal_ioctl_hdr */
 952 static inline int portal_ioctl_getdata(char *buf, char *end, void *arg)
 953 {
 954         struct portal_ioctl_hdr *hdr;
 955         struct portal_ioctl_data *data;
 956         int err;
 957         ENTRY;
 958
 959         hdr = (struct portal_ioctl_hdr *)buf;
 960         data = (struct portal_ioctl_data *)buf;
 961
 962         err = copy_from_user(buf, (void *)arg, sizeof(*hdr));
 963         if ( err ) {
 964                 EXIT;
 965                 return err;
 966         }
 967
 968         if (hdr->ioc_version != PORTAL_IOCTL_VERSION) {
 969                 CERROR ("PORTALS: version mismatch kernel vs application\n");
 970                 return -EINVAL;
 971         }
 972
 973         if (hdr->ioc_len + buf >= end) {
 974                 CERROR ("PORTALS: user buffer exceeds kernel buffer\n");
 975                 return -EINVAL;
 976         }
 977
 978
 979         if (hdr->ioc_len < sizeof(struct portal_ioctl_data)) {
 980                 CERROR ("PORTALS: user buffer too small for ioctl\n");
 981                 return -EINVAL;
 982         }
 983
 984         err = copy_from_user(buf, (void *)arg, hdr->ioc_len);
 985         if ( err ) {
 986                 EXIT;
 987                 return err;
 988         }
 989
 990         if (portal_ioctl_is_invalid(data)) {
 991                 CERROR ("PORTALS: ioctl not correctly formatted\n");
 992                 return -EINVAL;
 993         }
 994
 995         if (data->ioc_inllen1) {
 996                 data->ioc_inlbuf1 = &data->ioc_bulk[0];
 997         }
 998
 999         if (data->ioc_inllen2) {
1000                 data->ioc_inlbuf2 = &data->ioc_bulk[0] +
1001                         size_round(data->ioc_inllen1);
1002         }
1003
1004         EXIT;
1005         return 0;
1006 }
1007 #endif
1008
1009 /* ioctls for manipulating snapshots 30- */
1010 #define IOC_PORTAL_TYPE                   'e'
1011 #define IOC_PORTAL_MIN_NR                 30
1012
1013 #define IOC_PORTAL_PING                    _IOWR('e', 30, long)
1014 #define IOC_PORTAL_GET_DEBUG               _IOWR('e', 31, long)
1015 #define IOC_PORTAL_CLEAR_DEBUG             _IOWR('e', 32, long)
1016 #define IOC_PORTAL_MARK_DEBUG              _IOWR('e', 33, long)
1017 #define IOC_PORTAL_PANIC                   _IOWR('e', 34, long)
1018 #define IOC_PORTAL_NAL_CMD                 _IOWR('e', 35, long)
1019 #define IOC_PORTAL_GET_NID                 _IOWR('e', 36, long)
1020 #define IOC_PORTAL_FAIL_NID                _IOWR('e', 37, long)
1021 #define IOC_PORTAL_SET_DAEMON              _IOWR('e', 38, long)
1022 #define IOC_PORTAL_LWT_CONTROL             _IOWR('e', 39, long)
1023 #define IOC_PORTAL_LWT_SNAPSHOT            _IOWR('e', 40, long)
1024 #define IOC_PORTAL_LWT_LOOKUP_STRING       _IOWR('e', 41, long)
1025 #define IOC_PORTAL_MEMHOG                  _IOWR('e', 42, long)
1026 #define IOC_PORTAL_MAX_NR                             42
1027
1028 enum {
1029         QSWNAL    = 1,
1030         SOCKNAL   = 2,
1031         GMNAL     = 3,
1032         /*          4 unused */
1033         TCPNAL    = 5,
1034         SCIMACNAL = 6,
1035         ROUTER    = 7,
1036         IBNAL     = 8,
1037         NAL_ENUM_END_MARKER
1038 };
1039
1040 #ifdef __KERNEL__
1041 extern ptl_handle_ni_t  kqswnal_ni;
1042 extern ptl_handle_ni_t  ksocknal_ni;
1043 extern ptl_handle_ni_t  kgmnal_ni;
1044 extern ptl_handle_ni_t  kibnal_ni;
1045 extern ptl_handle_ni_t  kscimacnal_ni;
1046 #endif
1047
1048 #define PTL_NALFMT_SIZE         16
1049
1050 #define NAL_MAX_NR (NAL_ENUM_END_MARKER - 1)
1051
1052 #define NAL_CMD_REGISTER_PEER_FD     100
1053 #define NAL_CMD_CLOSE_CONNECTION     101
1054 #define NAL_CMD_REGISTER_MYNID       102
1055 #define NAL_CMD_PUSH_CONNECTION      103
1056 #define NAL_CMD_GET_CONN             104
1057 #define NAL_CMD_DEL_AUTOCONN         105
1058 #define NAL_CMD_ADD_AUTOCONN         106
1059 #define NAL_CMD_GET_AUTOCONN         107
1060 #define NAL_CMD_GET_TXDESC           108
1061 #define NAL_CMD_ADD_ROUTE            109
1062 #define NAL_CMD_DEL_ROUTE            110
1063 #define NAL_CMD_GET_ROUTE            111
1064 #define NAL_CMD_NOTIFY_ROUTER        112
1065
1066 enum {
1067         DEBUG_DAEMON_START       =  1,
1068         DEBUG_DAEMON_STOP        =  2,
1069         DEBUG_DAEMON_PAUSE       =  3,
1070         DEBUG_DAEMON_CONTINUE    =  4,
1071 };
1072
1073 /* XXX remove to lustre ASAP */
1074 struct lustre_peer {
1075         ptl_nid_t       peer_nid;
1076         ptl_handle_ni_t peer_ni;
1077 };
1078
1079
1080 /* module.c */
1081 typedef int (*nal_cmd_handler_t)(struct portals_cfg *, void * private);
1082 int kportal_nal_register(int nal, nal_cmd_handler_t handler, void * private);
1083 int kportal_nal_unregister(int nal);
1084
1085 enum cfg_record_type {
1086         PORTALS_CFG_TYPE = 1,
1087         LUSTRE_CFG_TYPE = 123,
1088 };
1089
1090 typedef int (*cfg_record_cb_t)(enum cfg_record_type, int len, void *data);
1091 int kportal_nal_cmd(struct portals_cfg *);
1092
1093 ptl_handle_ni_t *kportal_get_ni (int nal);
1094 void kportal_put_ni (int nal);
1095
1096 #ifdef __CYGWIN__
1097 # ifndef BITS_PER_LONG
1098 #  if (~0UL) == 0xffffffffUL
1099 #   define BITS_PER_LONG 32
1100 #  else
1101 #   define BITS_PER_LONG 64
1102 #  endif
1103 # endif
1104 #endif
1105
1106 #if defined(__x86_64__)
1107 # define LPU64 "%Lu"
1108 # define LPD64 "%Ld"
1109 # define LPX64 "%#Lx"
1110 # define LPSZ  "%lu"
1111 # define LPSSZ "%ld"
1112 #elif (BITS_PER_LONG == 32 || __WORDSIZE == 32)
1113 # define LPU64 "%Lu"
1114 # define LPD64 "%Ld"
1115 # define LPX64 "%#Lx"
1116 # define LPSZ  "%u"
1117 # define LPSSZ "%d"
1118 #elif (BITS_PER_LONG == 64 || __WORDSIZE == 64)
1119 # define LPU64 "%lu"
1120 # define LPD64 "%ld"
1121 # define LPX64 "%#lx"
1122 # define LPSZ  "%lu"
1123 # define LPSSZ "%ld"
1124 #endif
1125 #ifndef LPU64
1126 # error "No word size defined"
1127 #endif
1128
1129 #endif