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   0x00000001
  25 #define S_MDC         0x00000002
  26 #define S_MDS         0x00000004
  27 #define S_OSC         0x00000008
  28 #define S_OST         0x00000010
  29 #define S_CLASS       0x00000020
  30 #define S_LOG         0x00000040
  31 #define S_LLITE       0x00000080
  32 #define S_RPC         0x00000100
  33 #define S_MGMT        0x00000200
  34 #define S_PORTALS     0x00000400
  35 #define S_SOCKNAL     0x00000800
  36 #define S_QSWNAL      0x00001000
  37 #define S_PINGER      0x00002000
  38 #define S_FILTER      0x00004000
  39 #define S_PTLBD       0x00008000
  40 #define S_ECHO        0x00010000
  41 #define S_LDLM        0x00020000
  42 #define S_LOV         0x00040000
  43 #define S_GMNAL       0x00080000
  44 #define S_PTLROUTER   0x00100000
  45 #define S_COBD        0x00200000
  46 #define S_IBNAL       0x00400000
  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       0x00000001 /* ENTRY/EXIT markers */
  53 #define D_INODE       0x00000002
  54 #define D_SUPER       0x00000004
  55 #define D_EXT2        0x00000008 /* anything from ext2_debug */
  56 #define D_MALLOC      0x00000010 /* print malloc, free information */
  57 #define D_CACHE       0x00000020 /* cache-related items */
  58 #define D_INFO        0x00000040 /* general information */
  59 #define D_IOCTL       0x00000080 /* ioctl related information */
  60 #define D_BLOCKS      0x00000100 /* ext2 block allocation */
  61 #define D_NET         0x00000200 /* network communications */
  62 #define D_WARNING     0x00000400 /* CWARN(...) == CDEBUG (D_WARNING, ...) */
  63 #define D_BUFFS       0x00000800
  64 #define D_OTHER       0x00001000
  65 #define D_DENTRY      0x00002000
  66 #define D_PORTALS     0x00004000 /* ENTRY/EXIT markers */
  67 #define D_PAGE        0x00008000 /* bulk page handling */
  68 #define D_DLMTRACE    0x00010000
  69 #define D_ERROR       0x00020000 /* CERROR(...) == CDEBUG (D_ERROR, ...) */
  70 #define D_EMERG       0x00040000 /* CEMERG(...) == CDEBUG (D_EMERG, ...) */
  71 #define D_HA          0x00080000 /* recovery and failover */
  72 #define D_RPCTRACE    0x00100000 /* for distributed debugging */
  73 #define D_VFSTRACE    0x00200000
  74 #define D_READA       0x00400000 /* 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 CONFIG_SMP
 238 #define LASSERT_SPIN_LOCKED(lock) LASSERT(spin_is_locked(lock))
 239 #else
 240 #define LASSERT_SPIN_LOCKED(lock) do {} while(0)
 241 #endif
 242
 243 #ifdef __arch_um__
 244 #define LBUG_WITH_LOC(file, func, line)                                 \
 245 do {                                                                    \
 246         CEMERG("LBUG - trying to dump log to /tmp/lustre-log\n");       \
 247         portals_debug_dumplog();                                        \
 248         portals_run_lbug_upcall(file, func, line);                      \
 249         panic("LBUG");                                                  \
 250 } while (0)
 251 #else
 252 #define LBUG_WITH_LOC(file, func, line)                                 \
 253 do {                                                                    \
 254         CEMERG("LBUG\n");                                               \
 255         portals_debug_dumplog();                                        \
 256         portals_run_lbug_upcall(file, func, line);                      \
 257         set_task_state(current, TASK_UNINTERRUPTIBLE);                  \
 258         schedule();                                                     \
 259 } while (0)
 260 #endif /* __arch_um__ */
 261
 262 #define LBUG() LBUG_WITH_LOC(__FILE__, __FUNCTION__, __LINE__)
 263
 264 /*
 265  * Memory
 266  */
 267 #ifdef PORTAL_DEBUG
 268 extern atomic_t portal_kmemory;
 269
 270 # define portal_kmem_inc(ptr, size)                                           \
 271 do {                                                                          \
 272         atomic_add(size, &portal_kmemory);                                    \
 273 } while (0)
 274
 275 # define portal_kmem_dec(ptr, size) do {                                      \
 276         atomic_sub(size, &portal_kmemory);                                    \
 277 } while (0)
 278
 279 #else
 280 # define portal_kmem_inc(ptr, size) do {} while (0)
 281 # define portal_kmem_dec(ptr, size) do {} while (0)
 282 #endif /* PORTAL_DEBUG */
 283
 284 #define PORTAL_VMALLOC_SIZE        16384
 285
 286 #define PORTAL_ALLOC_GFP(ptr, size, mask)                                 \
 287 do {                                                                      \
 288         LASSERT(!in_interrupt() ||                                        \
 289                (size <= PORTAL_VMALLOC_SIZE && mask == GFP_ATOMIC));      \
 290         if ((size) > PORTAL_VMALLOC_SIZE)                                 \
 291                 (ptr) = vmalloc(size);                                    \
 292         else                                                              \
 293                 (ptr) = kmalloc((size), (mask));                          \
 294         if ((ptr) == NULL) {                                              \
 295                 CERROR("PORTALS: out of memory at %s:%d (tried to alloc '"\
 296                        #ptr "' = %d)\n", __FILE__, __LINE__, (int)(size));\
 297                 CERROR("PORTALS: %d total bytes allocated by portals\n",  \
 298                        atomic_read(&portal_kmemory));                     \
 299         } else {                                                          \
 300                 portal_kmem_inc((ptr), (size));                           \
 301                 memset((ptr), 0, (size));                                 \
 302         }                                                                 \
 303         CDEBUG(D_MALLOC, "kmalloced '" #ptr "': %d at %p (tot %d).\n",    \
 304                (int)(size), (ptr), atomic_read (&portal_kmemory));        \
 305 } while (0)
 306
 307 #define PORTAL_ALLOC(ptr, size) \
 308         PORTAL_ALLOC_GFP(ptr, size, GFP_NOFS)
 309
 310 #define PORTAL_ALLOC_ATOMIC(ptr, size) \
 311         PORTAL_ALLOC_GFP(ptr, size, GFP_ATOMIC)
 312
 313 #define PORTAL_FREE(ptr, size)                                          \
 314 do {                                                                    \
 315         int s = (size);                                                 \
 316         if ((ptr) == NULL) {                                            \
 317                 CERROR("PORTALS: free NULL '" #ptr "' (%d bytes) at "   \
 318                        "%s:%d\n", s, __FILE__, __LINE__);               \
 319                 break;                                                  \
 320         }                                                               \
 321         if (s > PORTAL_VMALLOC_SIZE)                                    \
 322                 vfree(ptr);                                             \
 323         else                                                            \
 324                 kfree(ptr);                                             \
 325         portal_kmem_dec((ptr), s);                                      \
 326         CDEBUG(D_MALLOC, "kfreed '" #ptr "': %d at %p (tot %d).\n",     \
 327                s, (ptr), atomic_read(&portal_kmemory));                 \
 328 } while (0)
 329
 330 /* ------------------------------------------------------------------- */
 331
 332 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 333
 334 #define PORTAL_SYMBOL_REGISTER(x) inter_module_register(#x, THIS_MODULE, &x)
 335 #define PORTAL_SYMBOL_UNREGISTER(x) inter_module_unregister(#x)
 336
 337 #define PORTAL_SYMBOL_GET(x) ((typeof(&x))inter_module_get(#x))
 338 #define PORTAL_SYMBOL_PUT(x) inter_module_put(#x)
 339
 340 #define PORTAL_MODULE_USE       MOD_INC_USE_COUNT
 341 #define PORTAL_MODULE_UNUSE     MOD_DEC_USE_COUNT
 342 #else
 343
 344 #define PORTAL_SYMBOL_REGISTER(x)
 345 #define PORTAL_SYMBOL_UNREGISTER(x)
 346
 347 #define PORTAL_SYMBOL_GET(x) symbol_get(x)
 348 #define PORTAL_SYMBOL_PUT(x) symbol_put(x)
 349
 350 #define PORTAL_MODULE_USE       try_module_get(THIS_MODULE)
 351 #define PORTAL_MODULE_UNUSE     module_put(THIS_MODULE)
 352
 353 #endif
 354
 355 /******************************************************************************/
 356 /* Kernel Portals Router interface */
 357
 358 typedef void (*kpr_fwd_callback_t)(void *arg, int error); // completion callback
 359
 360 /* space for routing targets to stash "stuff" in a forwarded packet */
 361 typedef union {
 362         long long        _alignment;
 363         void            *_space[16];            /* scale with CPU arch */
 364 } kprfd_scratch_t;
 365
 366 /* Kernel Portals Routing Forwarded message Descriptor */
 367 typedef struct {
 368         struct list_head     kprfd_list;        /* stash in queues (routing target can use) */
 369         ptl_nid_t            kprfd_target_nid;  /* final destination NID */
 370         ptl_nid_t            kprfd_gateway_nid; /* gateway NID */
 371         ptl_hdr_t           *kprfd_hdr;         /* header in wire byte order */
 372         int                  kprfd_nob;         /* # payload bytes */
 373         int                  kprfd_niov;        /* # payload frags */
 374         ptl_kiov_t          *kprfd_kiov;        /* payload fragments */
 375         void                *kprfd_router_arg;  /* originating NAL's router arg */
 376         kpr_fwd_callback_t   kprfd_callback;    /* completion callback */
 377         void                *kprfd_callback_arg; /* completion callback arg */
 378         kprfd_scratch_t      kprfd_scratch;     /* scratchpad for routing targets */
 379 } kpr_fwd_desc_t;
 380
 381 typedef void  (*kpr_fwd_t)(void *arg, kpr_fwd_desc_t *fwd);
 382 typedef void  (*kpr_notify_t)(void *arg, ptl_nid_t peer, int alive);
 383
 384 /* NAL's routing interface (Kernel Portals Routing Nal Interface) */
 385 typedef const struct {
 386         int             kprni_nalid;    /* NAL's id */
 387         void           *kprni_arg;      /* Arg to pass when calling into NAL */
 388         kpr_fwd_t       kprni_fwd;      /* NAL's forwarding entrypoint */
 389         kpr_notify_t    kprni_notify;   /* NAL's notification entrypoint */
 390 } kpr_nal_interface_t;
 391
 392 /* Router's routing interface (Kernel Portals Routing Router Interface) */
 393 typedef const struct {
 394         /* register the calling NAL with the router and get back the handle for
 395          * subsequent calls */
 396         int     (*kprri_register) (kpr_nal_interface_t *nal_interface,
 397                                    void **router_arg);
 398
 399         /* ask the router to find a gateway that forwards to 'nid' and is a
 400          * peer of the calling NAL; assume caller will send 'nob' bytes of
 401          * payload there */
 402         int     (*kprri_lookup) (void *router_arg, ptl_nid_t nid, int nob,
 403                                  ptl_nid_t *gateway_nid);
 404
 405         /* hand a packet over to the router for forwarding */
 406         kpr_fwd_t kprri_fwd_start;
 407
 408         /* hand a packet back to the router for completion */
 409         void    (*kprri_fwd_done) (void *router_arg, kpr_fwd_desc_t *fwd,
 410                                    int error);
 411
 412         /* notify the router about peer state */
 413         void    (*kprri_notify) (void *router_arg, ptl_nid_t peer,
 414                                  int alive, time_t when);
 415
 416         /* the calling NAL is shutting down */
 417         void    (*kprri_shutdown) (void *router_arg);
 418
 419         /* deregister the calling NAL with the router */
 420         void    (*kprri_deregister) (void *router_arg);
 421
 422 } kpr_router_interface_t;
 423
 424 /* Convenient struct for NAL to stash router interface/args */
 425 typedef struct {
 426         kpr_router_interface_t  *kpr_interface;
 427         void                    *kpr_arg;
 428 } kpr_router_t;
 429
 430 /* Router's control interface (Kernel Portals Routing Control Interface) */
 431 typedef const struct {
 432         int     (*kprci_add_route)(int gateway_nal, ptl_nid_t gateway_nid,
 433                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
 434         int     (*kprci_del_route)(int gateway_nal, ptl_nid_t gateway_nid,
 435                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
 436         int     (*kprci_get_route)(int index, int *gateway_nal,
 437                                    ptl_nid_t *gateway,
 438                                    ptl_nid_t *lo_nid, ptl_nid_t *hi_nid,
 439                                    int *alive);
 440         int     (*kprci_notify)(int gateway_nal, ptl_nid_t gateway_nid,
 441                                 int alive, time_t when);
 442 } kpr_control_interface_t;
 443
 444 extern kpr_control_interface_t  kpr_control_interface;
 445 extern kpr_router_interface_t   kpr_router_interface;
 446
 447 static inline int
 448 kpr_register (kpr_router_t *router, kpr_nal_interface_t *nalif)
 449 {
 450         int    rc;
 451
 452         router->kpr_interface = PORTAL_SYMBOL_GET (kpr_router_interface);
 453         if (router->kpr_interface == NULL)
 454                 return (-ENOENT);
 455
 456         rc = (router->kpr_interface)->kprri_register (nalif, &router->kpr_arg);
 457         if (rc != 0)
 458                 router->kpr_interface = NULL;
 459
 460         PORTAL_SYMBOL_PUT (kpr_router_interface);
 461         return (rc);
 462 }
 463
 464 static inline int
 465 kpr_routing (kpr_router_t *router)
 466 {
 467         return (router->kpr_interface != NULL);
 468 }
 469
 470 static inline int
 471 kpr_lookup (kpr_router_t *router, ptl_nid_t nid, int nob, ptl_nid_t *gateway_nid)
 472 {
 473         if (!kpr_routing (router))
 474                 return (-ENETUNREACH);
 475
 476         return (router->kpr_interface->kprri_lookup(router->kpr_arg, nid, nob,
 477                                                     gateway_nid));
 478 }
 479
 480 static inline void
 481 kpr_fwd_init (kpr_fwd_desc_t *fwd, ptl_nid_t nid, ptl_hdr_t *hdr,
 482               int nob, int niov, ptl_kiov_t *kiov,
 483               kpr_fwd_callback_t callback, void *callback_arg)
 484 {
 485         fwd->kprfd_target_nid   = nid;
 486         fwd->kprfd_gateway_nid  = nid;
 487         fwd->kprfd_hdr          = hdr;
 488         fwd->kprfd_nob          = nob;
 489         fwd->kprfd_niov         = niov;
 490         fwd->kprfd_kiov         = kiov;
 491         fwd->kprfd_callback     = callback;
 492         fwd->kprfd_callback_arg = callback_arg;
 493 }
 494
 495 static inline void
 496 kpr_fwd_start (kpr_router_t *router, kpr_fwd_desc_t *fwd)
 497 {
 498         if (!kpr_routing (router))
 499                 fwd->kprfd_callback (fwd->kprfd_callback_arg, -ENETUNREACH);
 500         else
 501                 router->kpr_interface->kprri_fwd_start (router->kpr_arg, fwd);
 502 }
 503
 504 static inline void
 505 kpr_fwd_done (kpr_router_t *router, kpr_fwd_desc_t *fwd, int error)
 506 {
 507         LASSERT (kpr_routing (router));
 508         router->kpr_interface->kprri_fwd_done (router->kpr_arg, fwd, error);
 509 }
 510
 511 static inline void
 512 kpr_notify (kpr_router_t *router,
 513             ptl_nid_t peer, int alive, time_t when)
 514 {
 515         if (!kpr_routing (router))
 516                 return;
 517
 518         router->kpr_interface->kprri_notify(router->kpr_arg, peer, alive, when);
 519 }
 520
 521 static inline void
 522 kpr_shutdown (kpr_router_t *router)
 523 {
 524         if (kpr_routing (router))
 525                 router->kpr_interface->kprri_shutdown (router->kpr_arg);
 526 }
 527
 528 static inline void
 529 kpr_deregister (kpr_router_t *router)
 530 {
 531         if (!kpr_routing (router))
 532                 return;
 533         router->kpr_interface->kprri_deregister (router->kpr_arg);
 534         router->kpr_interface = NULL;
 535 }
 536
 537 /******************************************************************************/
 538
 539 #ifdef PORTALS_PROFILING
 540 #define prof_enum(FOO) PROF__##FOO
 541 enum {
 542         prof_enum(our_recvmsg),
 543         prof_enum(our_sendmsg),
 544         prof_enum(socknal_recv),
 545         prof_enum(lib_parse),
 546         prof_enum(conn_list_walk),
 547         prof_enum(memcpy),
 548         prof_enum(lib_finalize),
 549         prof_enum(pingcli_time),
 550         prof_enum(gmnal_send),
 551         prof_enum(gmnal_recv),
 552         MAX_PROFS
 553 };
 554
 555 struct prof_ent {
 556         char *str;
 557         /* hrmph.  wrap-tastic. */
 558         u32       starts;
 559         u32       finishes;
 560         cycles_t  total_cycles;
 561         cycles_t  start;
 562         cycles_t  end;
 563 };
 564
 565 extern struct prof_ent prof_ents[MAX_PROFS];
 566
 567 #define PROF_START(FOO)                                         \
 568         do {                                                    \
 569                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
 570                 pe->starts++;                                   \
 571                 pe->start = get_cycles();                       \
 572         } while (0)
 573
 574 #define PROF_FINISH(FOO)                                        \
 575         do {                                                    \
 576                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
 577                 pe->finishes++;                                 \
 578                 pe->end = get_cycles();                         \
 579                 pe->total_cycles += (pe->end - pe->start);      \
 580         } while (0)
 581 #else /* !PORTALS_PROFILING */
 582 #define PROF_START(FOO) do {} while(0)
 583 #define PROF_FINISH(FOO) do {} while(0)
 584 #endif /* PORTALS_PROFILING */
 585
 586 /* debug.c */
 587 extern spinlock_t stack_backtrace_lock;
 588
 589 char *portals_debug_dumpstack(void);
 590 void portals_run_upcall(char **argv);
 591 void portals_run_lbug_upcall(char * file, const char *fn, const int line);
 592 void portals_debug_dumplog(void);
 593 int portals_debug_init(unsigned long bufsize);
 594 int portals_debug_cleanup(void);
 595 int portals_debug_clear_buffer(void);
 596 int portals_debug_mark_buffer(char *text);
 597 int portals_debug_set_daemon(unsigned int cmd, unsigned int length,
 598                              char *file, unsigned int size);
 599 __s32 portals_debug_copy_to_user(char *buf, unsigned long len);
 600 #if (__GNUC__)
 601 /* Use the special GNU C __attribute__ hack to have the compiler check the
 602  * printf style argument string against the actual argument count and
 603  * types.
 604  */
 605 #ifdef printf
 606 # warning printf has been defined as a macro...
 607 # undef printf
 608 #endif
 609 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
 610                        const int line, unsigned long stack,
 611                        char *format, ...)
 612         __attribute__ ((format (printf, 7, 8)));
 613 #else
 614 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
 615                        const int line, unsigned long stack,
 616                        const char *format, ...);
 617 #endif /* __GNUC__ */
 618 void portals_debug_set_level(unsigned int debug_level);
 619
 620 # define fprintf(a, format, b...) CDEBUG(D_OTHER, format , ## b)
 621 # define printf(format, b...) CDEBUG(D_OTHER, format , ## b)
 622 # define time(a) CURRENT_TIME
 623
 624 extern void kportal_daemonize (char *name);
 625 extern void kportal_blockallsigs (void);
 626
 627 #else  /* !__KERNEL__ */
 628 # include <stdio.h>
 629 # include <stdlib.h>
 630 #ifndef __CYGWIN__
 631 # include <stdint.h>
 632 #else
 633 # include <cygwin-ioctl.h>
 634 #endif
 635 # include <unistd.h>
 636 # include <time.h>
 637 # include <asm/types.h>
 638 # ifndef DEBUG_SUBSYSTEM
 639 #  define DEBUG_SUBSYSTEM S_UNDEFINED
 640 # endif
 641 # ifdef PORTAL_DEBUG
 642 #  undef NDEBUG
 643 #  include <assert.h>
 644 #  define LASSERT(e)     assert(e)
 645 #  define LASSERTF(cond, args...)     assert(cond)
 646 # else
 647 #  define LASSERT(e)
 648 #  define LASSERTF(cond, args...) do { } while (0)
 649 # endif
 650 # define printk(format, args...) printf (format, ## args)
 651 # define PORTAL_ALLOC(ptr, size) do { (ptr) = malloc(size); } while (0);
 652 # define PORTAL_FREE(a, b) do { free(a); } while (0);
 653 # define portals_debug_msg(subsys, mask, file, fn, line, stack, format, a...) \
 654     printf("%02x:%06x (@%lu %s:%s,l. %d %d %lu): " format,                    \
 655            (subsys), (mask), (long)time(0), file, fn, line,                   \
 656            getpid() , stack, ## a);
 657 #endif
 658
 659 /* support decl needed both by kernel and liblustre */
 660 char *portals_nid2str(int nal, ptl_nid_t nid, char *str);
 661
 662 #ifndef CURRENT_TIME
 663 # define CURRENT_TIME time(0)
 664 #endif
 665
 666 /******************************************************************************/
 667 /* Light-weight trace
 668  * Support for temporary event tracing with minimal Heisenberg effect. */
 669 #define LWT_SUPPORT  0
 670
 671 #define LWT_MEMORY   (64<<20)
 672 #define LWT_MAX_CPUS 4
 673
 674 typedef struct {
 675         cycles_t    lwte_when;
 676         char       *lwte_where;
 677         void       *lwte_task;
 678         long        lwte_p1;
 679         long        lwte_p2;
 680         long        lwte_p3;
 681         long        lwte_p4;
 682 #if BITS_PER_LONG > 32
 683         long        lwte_pad;
 684 #endif
 685 } lwt_event_t;
 686
 687 #if LWT_SUPPORT
 688 #ifdef __KERNEL__
 689 #define LWT_EVENTS_PER_PAGE (PAGE_SIZE / sizeof (lwt_event_t))
 690
 691 typedef struct _lwt_page {
 692         struct list_head     lwtp_list;
 693         struct page         *lwtp_page;
 694         lwt_event_t         *lwtp_events;
 695 } lwt_page_t;
 696
 697 typedef struct {
 698         int                lwtc_current_index;
 699         lwt_page_t        *lwtc_current_page;
 700 } lwt_cpu_t;
 701
 702 extern int       lwt_enabled;
 703 extern lwt_cpu_t lwt_cpus[];
 704
 705 extern int  lwt_init (void);
 706 extern void lwt_fini (void);
 707 extern int  lwt_lookup_string (int *size, char *knlptr,
 708                                char *usrptr, int usrsize);
 709 extern int  lwt_control (int enable, int clear);
 710 extern int  lwt_snapshot (cycles_t *now, int *ncpu, int *total_size,
 711                           void *user_ptr, int user_size);
 712
 713 /* Note that we _don't_ define LWT_EVENT at all if LWT_SUPPORT isn't set.
 714  * This stuff is meant for finding specific problems; it never stays in
 715  * production code... */
 716
 717 #define LWTSTR(n)       #n
 718 #define LWTWHERE(f,l)   f ":" LWTSTR(l)
 719
 720 #define LWT_EVENT(p1, p2, p3, p4)                                       \
 721 do {                                                                    \
 722         unsigned long    flags;                                         \
 723         lwt_cpu_t       *cpu;                                           \
 724         lwt_page_t      *p;                                             \
 725         lwt_event_t     *e;                                             \
 726                                                                         \
 727         local_irq_save (flags);                                         \
 728                                                                         \
 729         if (lwt_enabled) {                                              \
 730                 cpu = &lwt_cpus[smp_processor_id()];                    \
 731                 p = cpu->lwtc_current_page;                             \
 732                 e = &p->lwtp_events[cpu->lwtc_current_index++];         \
 733                                                                         \
 734                 if (cpu->lwtc_current_index >= LWT_EVENTS_PER_PAGE) {   \
 735                         cpu->lwtc_current_page =                        \
 736                                 list_entry (p->lwtp_list.next,          \
 737                                             lwt_page_t, lwtp_list);     \
 738                         cpu->lwtc_current_index = 0;                    \
 739                 }                                                       \
 740                                                                         \
 741                 e->lwte_when  = get_cycles();                           \
 742                 e->lwte_where = LWTWHERE(__FILE__,__LINE__);            \
 743                 e->lwte_task  = current;                                \
 744                 e->lwte_p1    = (long)(p1);                             \
 745                 e->lwte_p2    = (long)(p2);                             \
 746                 e->lwte_p3    = (long)(p3);                             \
 747                 e->lwte_p4    = (long)(p4);                             \
 748         }                                                               \
 749                                                                         \
 750         local_irq_restore (flags);                                      \
 751 } while (0)
 752 #else  /* __KERNEL__ */
 753 #define LWT_EVENT(p1,p2,p3,p4)     /* no userland implementation yet */
 754 #endif /* __KERNEL__ */
 755 #endif /* LWT_SUPPORT */
 756
 757 struct portals_device_userstate
 758 {
 759         int          pdu_memhog_pages;
 760         struct page *pdu_memhog_root_page;
 761 };
 762
 763 #include <linux/portals_lib.h>
 764
 765 /*
 766  * USER LEVEL STUFF BELOW
 767  */
 768
 769 #define PORTALS_CFG_VERSION 0x00010001;
 770
 771 struct portals_cfg {
 772         __u32 pcfg_version;
 773         __u32 pcfg_command;
 774
 775         __u32 pcfg_nal;
 776         __u32 pcfg_flags;
 777
 778         __u32 pcfg_gw_nal;
 779         __u64 pcfg_nid;
 780         __u64 pcfg_nid2;
 781         __u64 pcfg_nid3;
 782         __u32 pcfg_id;
 783         __u32 pcfg_misc;
 784         __u32 pcfg_fd;
 785         __u32 pcfg_count;
 786         __u32 pcfg_size;
 787         __u32 pcfg_wait;
 788
 789         __u32 pcfg_plen1; /* buffers in userspace */
 790         char *pcfg_pbuf1;
 791         __u32 pcfg_plen2; /* buffers in userspace */
 792         char *pcfg_pbuf2;
 793 };
 794
 795 #define PCFG_INIT(pcfg, cmd)                            \
 796 do {                                                    \
 797         memset(&pcfg, 0, sizeof(pcfg));                 \
 798         pcfg.pcfg_version = PORTALS_CFG_VERSION;        \
 799         pcfg.pcfg_command = (cmd);                      \
 800                                                         \
 801 } while (0)
 802
 803 #define PORTAL_IOCTL_VERSION 0x00010007
 804 #define PING_SYNC       0
 805 #define PING_ASYNC      1
 806
 807 struct portal_ioctl_data {
 808         __u32 ioc_len;
 809         __u32 ioc_version;
 810         __u64 ioc_nid;
 811         __u64 ioc_nid2;
 812         __u64 ioc_nid3;
 813         __u32 ioc_count;
 814         __u32 ioc_nal;
 815         __u32 ioc_nal_cmd;
 816         __u32 ioc_fd;
 817         __u32 ioc_id;
 818
 819         __u32 ioc_flags;
 820         __u32 ioc_size;
 821
 822         __u32 ioc_wait;
 823         __u32 ioc_timeout;
 824         __u32 ioc_misc;
 825
 826         __u32 ioc_inllen1;
 827         char *ioc_inlbuf1;
 828         __u32 ioc_inllen2;
 829         char *ioc_inlbuf2;
 830
 831         __u32 ioc_plen1; /* buffers in userspace */
 832         char *ioc_pbuf1;
 833         __u32 ioc_plen2; /* buffers in userspace */
 834         char *ioc_pbuf2;
 835
 836         char ioc_bulk[0];
 837 };
 838
 839 struct portal_ioctl_hdr {
 840         __u32 ioc_len;
 841         __u32 ioc_version;
 842 };
 843
 844 struct portals_debug_ioctl_data
 845 {
 846         struct portal_ioctl_hdr hdr;
 847         unsigned int subs;
 848         unsigned int debug;
 849 };
 850
 851 #define PORTAL_IOC_INIT(data)                           \
 852 do {                                                    \
 853         memset(&data, 0, sizeof(data));                 \
 854         data.ioc_version = PORTAL_IOCTL_VERSION;        \
 855         data.ioc_len = sizeof(data);                    \
 856 } while (0)
 857
 858 /* FIXME check conflict with lustre_lib.h */
 859 #define PTL_IOC_DEBUG_MASK             _IOWR('f', 250, long)
 860
 861 static inline int portal_ioctl_packlen(struct portal_ioctl_data *data)
 862 {
 863         int len = sizeof(*data);
 864         len += size_round(data->ioc_inllen1);
 865         len += size_round(data->ioc_inllen2);
 866         return len;
 867 }
 868
 869 static inline int portal_ioctl_is_invalid(struct portal_ioctl_data *data)
 870 {
 871         if (data->ioc_len > (1<<30)) {
 872                 CERROR ("PORTALS ioctl: ioc_len larger than 1<<30\n");
 873                 return 1;
 874         }
 875         if (data->ioc_inllen1 > (1<<30)) {
 876                 CERROR ("PORTALS ioctl: ioc_inllen1 larger than 1<<30\n");
 877                 return 1;
 878         }
 879         if (data->ioc_inllen2 > (1<<30)) {
 880                 CERROR ("PORTALS ioctl: ioc_inllen2 larger than 1<<30\n");
 881                 return 1;
 882         }
 883         if (data->ioc_inlbuf1 && !data->ioc_inllen1) {
 884                 CERROR ("PORTALS ioctl: inlbuf1 pointer but 0 length\n");
 885                 return 1;
 886         }
 887         if (data->ioc_inlbuf2 && !data->ioc_inllen2) {
 888                 CERROR ("PORTALS ioctl: inlbuf2 pointer but 0 length\n");
 889                 return 1;
 890         }
 891         if (data->ioc_pbuf1 && !data->ioc_plen1) {
 892                 CERROR ("PORTALS ioctl: pbuf1 pointer but 0 length\n");
 893                 return 1;
 894         }
 895         if (data->ioc_pbuf2 && !data->ioc_plen2) {
 896                 CERROR ("PORTALS ioctl: pbuf2 pointer but 0 length\n");
 897                 return 1;
 898         }
 899         if (data->ioc_plen1 && !data->ioc_pbuf1) {
 900                 CERROR ("PORTALS ioctl: plen1 nonzero but no pbuf1 pointer\n");
 901                 return 1;
 902         }
 903         if (data->ioc_plen2 && !data->ioc_pbuf2) {
 904                 CERROR ("PORTALS ioctl: plen2 nonzero but no pbuf2 pointer\n");
 905                 return 1;
 906         }
 907         if (portal_ioctl_packlen(data) != data->ioc_len ) {
 908                 CERROR ("PORTALS ioctl: packlen != ioc_len\n");
 909                 return 1;
 910         }
 911         if (data->ioc_inllen1 &&
 912             data->ioc_bulk[data->ioc_inllen1 - 1] != '\0') {
 913                 CERROR ("PORTALS ioctl: inlbuf1 not 0 terminated\n");
 914                 return 1;
 915         }
 916         if (data->ioc_inllen2 &&
 917             data->ioc_bulk[size_round(data->ioc_inllen1) +
 918                            data->ioc_inllen2 - 1] != '\0') {
 919                 CERROR ("PORTALS ioctl: inlbuf2 not 0 terminated\n");
 920                 return 1;
 921         }
 922         return 0;
 923 }
 924
 925 #ifndef __KERNEL__
 926 static inline int portal_ioctl_pack(struct portal_ioctl_data *data, char **pbuf,
 927                                     int max)
 928 {
 929         char *ptr;
 930         struct portal_ioctl_data *overlay;
 931         data->ioc_len = portal_ioctl_packlen(data);
 932         data->ioc_version = PORTAL_IOCTL_VERSION;
 933
 934         if (*pbuf && portal_ioctl_packlen(data) > max)
 935                 return 1;
 936         if (*pbuf == NULL) {
 937                 *pbuf = malloc(data->ioc_len);
 938         }
 939         if (!*pbuf)
 940                 return 1;
 941         overlay = (struct portal_ioctl_data *)*pbuf;
 942         memcpy(*pbuf, data, sizeof(*data));
 943
 944         ptr = overlay->ioc_bulk;
 945         if (data->ioc_inlbuf1)
 946                 LOGL(data->ioc_inlbuf1, data->ioc_inllen1, ptr);
 947         if (data->ioc_inlbuf2)
 948                 LOGL(data->ioc_inlbuf2, data->ioc_inllen2, ptr);
 949         if (portal_ioctl_is_invalid(overlay))
 950                 return 1;
 951
 952         return 0;
 953 }
 954 #else
 955 #include <asm/uaccess.h>
 956
 957 /* buffer MUST be at least the size of portal_ioctl_hdr */
 958 static inline int portal_ioctl_getdata(char *buf, char *end, void *arg)
 959 {
 960         struct portal_ioctl_hdr *hdr;
 961         struct portal_ioctl_data *data;
 962         int err;
 963         ENTRY;
 964
 965         hdr = (struct portal_ioctl_hdr *)buf;
 966         data = (struct portal_ioctl_data *)buf;
 967
 968         err = copy_from_user(buf, (void *)arg, sizeof(*hdr));
 969         if ( err ) {
 970                 EXIT;
 971                 return err;
 972         }
 973
 974         if (hdr->ioc_version != PORTAL_IOCTL_VERSION) {
 975                 CERROR ("PORTALS: version mismatch kernel vs application\n");
 976                 return -EINVAL;
 977         }
 978
 979         if (hdr->ioc_len + buf >= end) {
 980                 CERROR ("PORTALS: user buffer exceeds kernel buffer\n");
 981                 return -EINVAL;
 982         }
 983
 984
 985         if (hdr->ioc_len < sizeof(struct portal_ioctl_data)) {
 986                 CERROR ("PORTALS: user buffer too small for ioctl\n");
 987                 return -EINVAL;
 988         }
 989
 990         err = copy_from_user(buf, (void *)arg, hdr->ioc_len);
 991         if ( err ) {
 992                 EXIT;
 993                 return err;
 994         }
 995
 996         if (portal_ioctl_is_invalid(data)) {
 997                 CERROR ("PORTALS: ioctl not correctly formatted\n");
 998                 return -EINVAL;
 999         }
1000
1001         if (data->ioc_inllen1) {
1002                 data->ioc_inlbuf1 = &data->ioc_bulk[0];
1003         }
1004
1005         if (data->ioc_inllen2) {
1006                 data->ioc_inlbuf2 = &data->ioc_bulk[0] +
1007                         size_round(data->ioc_inllen1);
1008         }
1009
1010         EXIT;
1011         return 0;
1012 }
1013 #endif
1014
1015 /* ioctls for manipulating snapshots 30- */
1016 #define IOC_PORTAL_TYPE                   'e'
1017 #define IOC_PORTAL_MIN_NR                 30
1018
1019 #define IOC_PORTAL_PING                    _IOWR('e', 30, long)
1020 #define IOC_PORTAL_GET_DEBUG               _IOWR('e', 31, long)
1021 #define IOC_PORTAL_CLEAR_DEBUG             _IOWR('e', 32, long)
1022 #define IOC_PORTAL_MARK_DEBUG              _IOWR('e', 33, long)
1023 #define IOC_PORTAL_PANIC                   _IOWR('e', 34, long)
1024 #define IOC_PORTAL_NAL_CMD                 _IOWR('e', 35, long)
1025 #define IOC_PORTAL_GET_NID                 _IOWR('e', 36, long)
1026 #define IOC_PORTAL_FAIL_NID                _IOWR('e', 37, long)
1027 #define IOC_PORTAL_SET_DAEMON              _IOWR('e', 38, long)
1028 #define IOC_PORTAL_LWT_CONTROL             _IOWR('e', 39, long)
1029 #define IOC_PORTAL_LWT_SNAPSHOT            _IOWR('e', 40, long)
1030 #define IOC_PORTAL_LWT_LOOKUP_STRING       _IOWR('e', 41, long)
1031 #define IOC_PORTAL_MEMHOG                  _IOWR('e', 42, long)
1032 #define IOC_PORTAL_MAX_NR                             42
1033
1034 enum {
1035         QSWNAL    = 1,
1036         SOCKNAL   = 2,
1037         GMNAL     = 3,
1038         /*          4 unused */
1039         TCPNAL    = 5,
1040         SCIMACNAL = 6,
1041         ROUTER    = 7,
1042         IBNAL     = 8,
1043         NAL_ENUM_END_MARKER
1044 };
1045
1046 #ifdef __KERNEL__
1047 extern ptl_handle_ni_t  kqswnal_ni;
1048 extern ptl_handle_ni_t  ksocknal_ni;
1049 extern ptl_handle_ni_t  kgmnal_ni;
1050 extern ptl_handle_ni_t  kibnal_ni;
1051 extern ptl_handle_ni_t  kscimacnal_ni;
1052 #endif
1053
1054 #define PTL_NALFMT_SIZE         16
1055
1056 #define NAL_MAX_NR (NAL_ENUM_END_MARKER - 1)
1057
1058 #define NAL_CMD_REGISTER_PEER_FD     100
1059 #define NAL_CMD_CLOSE_CONNECTION     101
1060 #define NAL_CMD_REGISTER_MYNID       102
1061 #define NAL_CMD_PUSH_CONNECTION      103
1062 #define NAL_CMD_GET_CONN             104
1063 #define NAL_CMD_DEL_AUTOCONN         105
1064 #define NAL_CMD_ADD_AUTOCONN         106
1065 #define NAL_CMD_GET_AUTOCONN         107
1066 #define NAL_CMD_GET_TXDESC           108
1067 #define NAL_CMD_ADD_ROUTE            109
1068 #define NAL_CMD_DEL_ROUTE            110
1069 #define NAL_CMD_GET_ROUTE            111
1070 #define NAL_CMD_NOTIFY_ROUTER        112
1071
1072 enum {
1073         DEBUG_DAEMON_START       =  1,
1074         DEBUG_DAEMON_STOP        =  2,
1075         DEBUG_DAEMON_PAUSE       =  3,
1076         DEBUG_DAEMON_CONTINUE    =  4,
1077 };
1078
1079 /* XXX remove to lustre ASAP */
1080 struct lustre_peer {
1081         ptl_nid_t       peer_nid;
1082         ptl_handle_ni_t peer_ni;
1083 };
1084
1085
1086 /* module.c */
1087 typedef int (*nal_cmd_handler_t)(struct portals_cfg *, void * private);
1088 int kportal_nal_register(int nal, nal_cmd_handler_t handler, void * private);
1089 int kportal_nal_unregister(int nal);
1090
1091 enum cfg_record_type {
1092         PORTALS_CFG_TYPE = 1,
1093         LUSTRE_CFG_TYPE = 123,
1094 };
1095
1096 typedef int (*cfg_record_cb_t)(enum cfg_record_type, int len, void *data);
1097 int kportal_nal_cmd(struct portals_cfg *);
1098
1099 ptl_handle_ni_t *kportal_get_ni (int nal);
1100 void kportal_put_ni (int nal);
1101
1102 #ifdef __CYGWIN__
1103 # ifndef BITS_PER_LONG
1104 #  if (~0UL) == 0xffffffffUL
1105 #   define BITS_PER_LONG 32
1106 #  else
1107 #   define BITS_PER_LONG 64
1108 #  endif
1109 # endif
1110 #endif
1111
1112 #if defined(__x86_64__)
1113 # define LPU64 "%Lu"
1114 # define LPD64 "%Ld"
1115 # define LPX64 "%#Lx"
1116 # define LPSZ  "%lu"
1117 # define LPSSZ "%ld"
1118 #elif (BITS_PER_LONG == 32 || __WORDSIZE == 32)
1119 # define LPU64 "%Lu"
1120 # define LPD64 "%Ld"
1121 # define LPX64 "%#Lx"
1122 # define LPSZ  "%u"
1123 # define LPSSZ "%d"
1124 #elif (BITS_PER_LONG == 64 || __WORDSIZE == 64)
1125 # define LPU64 "%lu"
1126 # define LPD64 "%ld"
1127 # define LPX64 "%#lx"
1128 # define LPSZ  "%lu"
1129 # define LPSSZ "%ld"
1130 #endif
1131 #ifndef LPU64
1132 # error "No word size defined"
1133 #endif
1134
1135 #endif