lustre/portals/include/linux/kp30.h

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  */
   4 #ifndef _KP30_INCLUDED
   5 #define _KP30_INCLUDED
   6
   7 #define PORTAL_DEBUG
   8
   9 #ifndef offsetof
  10 # define offsetof(typ,memb)     ((int)((char *)&(((typ *)0)->memb)))
  11 #endif
  12
  13 #define LOWEST_BIT_SET(x)       ((x) & ~((x) - 1))
  14
  15 /*
  16  *  Debugging
  17  */
  18 extern unsigned int portal_subsystem_debug;
  19 extern unsigned int portal_stack;
  20 extern unsigned int portal_debug;
  21 extern unsigned int portal_printk;
  22 extern unsigned int portal_cerror;
  23 /* Debugging subsystems (32 bits, non-overlapping) */
  24 #define S_UNDEFINED    (1 << 0)
  25 #define S_MDC          (1 << 1)
  26 #define S_MDS          (1 << 2)
  27 #define S_OSC          (1 << 3)
  28 #define S_OST          (1 << 4)
  29 #define S_CLASS        (1 << 5)
  30 #define S_LOG          (1 << 6)
  31 #define S_LLITE        (1 << 7)
  32 #define S_RPC          (1 << 8)
  33 #define S_MGMT         (1 << 9)
  34 #define S_PORTALS     (1 << 10)
  35 #define S_SOCKNAL     (1 << 11)
  36 #define S_QSWNAL      (1 << 12)
  37 #define S_PINGER      (1 << 13)
  38 #define S_FILTER      (1 << 14)
  39 #define S_PTLBD       (1 << 15)
  40 #define S_ECHO        (1 << 16)
  41 #define S_LDLM        (1 << 17)
  42 #define S_LOV         (1 << 18)
  43 #define S_GMNAL       (1 << 19)
  44 #define S_PTLROUTER   (1 << 20)
  45 #define S_COBD        (1 << 21)
  46 #define S_IBNAL       (1 << 22)
  47
  48 /* If you change these values, please keep portals/utils/debug.c
  49  * up to date! */
  50
  51 /* Debugging masks (32 bits, non-overlapping) */
  52 #define D_TRACE     (1 << 0) /* ENTRY/EXIT markers */
  53 #define D_INODE     (1 << 1)
  54 #define D_SUPER     (1 << 2)
  55 #define D_EXT2      (1 << 3) /* anything from ext2_debug */
  56 #define D_MALLOC    (1 << 4) /* print malloc, free information */
  57 #define D_CACHE     (1 << 5) /* cache-related items */
  58 #define D_INFO      (1 << 6) /* general information */
  59 #define D_IOCTL     (1 << 7) /* ioctl related information */
  60 #define D_BLOCKS    (1 << 8) /* ext2 block allocation */
  61 #define D_NET       (1 << 9) /* network communications */
  62 #define D_WARNING   (1 << 10) /* CWARN(...) == CDEBUG (D_WARNING, ...) */
  63 #define D_BUFFS     (1 << 11)
  64 #define D_OTHER     (1 << 12)
  65 #define D_DENTRY    (1 << 13)
  66 #define D_PORTALS   (1 << 14) /* ENTRY/EXIT markers */
  67 #define D_PAGE      (1 << 15) /* bulk page handling */
  68 #define D_DLMTRACE  (1 << 16)
  69 #define D_ERROR     (1 << 17) /* CERROR(...) == CDEBUG (D_ERROR, ...) */
  70 #define D_EMERG     (1 << 18) /* CEMERG(...) == CDEBUG (D_EMERG, ...) */
  71 #define D_HA        (1 << 19) /* recovery and failover */
  72 #define D_RPCTRACE  (1 << 20) /* for distributed debugging */
  73 #define D_VFSTRACE  (1 << 21)
  74 #define D_READA     (1 << 22) /* read-ahead */
  75
  76 #ifdef __KERNEL__
  77 # include <linux/sched.h> /* THREAD_SIZE */
  78 #else
  79 # ifndef THREAD_SIZE /* x86_64 has THREAD_SIZE in userspace */
  80 #  define THREAD_SIZE 8192
  81 # endif
  82 #endif
  83
  84 #define LUSTRE_TRACE_SIZE (THREAD_SIZE >> 5)
  85
  86 #ifdef __KERNEL__
  87 # ifdef  __ia64__
  88 #  define CDEBUG_STACK (THREAD_SIZE -                                      \
  89                         ((unsigned long)__builtin_dwarf_cfa() &            \
  90                          (THREAD_SIZE - 1)))
  91 # else
  92 #  define CDEBUG_STACK (THREAD_SIZE -                                      \
  93                         ((unsigned long)__builtin_frame_address(0) &       \
  94                          (THREAD_SIZE - 1)))
  95 # endif
  96
  97 #define CHECK_STACK(stack)                                                    \
  98         do {                                                                  \
  99                 if ((stack) > 3*THREAD_SIZE/4 && (stack) > portal_stack) {    \
 100                         portals_debug_msg(DEBUG_SUBSYSTEM, D_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                  (size <= PORTAL_VMALLOC_SIZE &&                          \
 287                   (mask & __GFP_WAIT) == 0));                             \
 288         if ((size) > PORTAL_VMALLOC_SIZE)                                 \
 289                 (ptr) = vmalloc(size);                                    \
 290         else                                                              \
 291                 (ptr) = kmalloc((size), (mask));                          \
 292         if ((ptr) == NULL) {                                              \
 293                 CERROR("PORTALS: out of memory at %s:%d (tried to alloc '"\
 294                        #ptr "' = %d)\n", __FILE__, __LINE__, (int)(size));\
 295                 CERROR("PORTALS: %d total bytes allocated by portals\n",  \
 296                        atomic_read(&portal_kmemory));                     \
 297         } else {                                                          \
 298                 portal_kmem_inc((ptr), (size));                           \
 299                 memset((ptr), 0, (size));                                 \
 300         }                                                                 \
 301         CDEBUG(D_MALLOC, "kmalloced '" #ptr "': %d at %p (tot %d).\n",    \
 302                (int)(size), (ptr), atomic_read (&portal_kmemory));        \
 303 } while (0)
 304
 305 #define PORTAL_ALLOC(ptr, size) \
 306         PORTAL_ALLOC_GFP(ptr, size, GFP_NOFS)
 307
 308 #define PORTAL_ALLOC_ATOMIC(ptr, size) \
 309         PORTAL_ALLOC_GFP(ptr, size, GFP_ATOMIC)
 310
 311 #define PORTAL_FREE(ptr, size)                                          \
 312 do {                                                                    \
 313         int s = (size);                                                 \
 314         if ((ptr) == NULL) {                                            \
 315                 CERROR("PORTALS: free NULL '" #ptr "' (%d bytes) at "   \
 316                        "%s:%d\n", s, __FILE__, __LINE__);               \
 317                 break;                                                  \
 318         }                                                               \
 319         if (s > PORTAL_VMALLOC_SIZE)                                    \
 320                 vfree(ptr);                                             \
 321         else                                                            \
 322                 kfree(ptr);                                             \
 323         portal_kmem_dec((ptr), s);                                      \
 324         CDEBUG(D_MALLOC, "kfreed '" #ptr "': %d at %p (tot %d).\n",     \
 325                s, (ptr), atomic_read(&portal_kmemory));                 \
 326 } while (0)
 327
 328 /* ------------------------------------------------------------------- */
 329
 330 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 331
 332 #define PORTAL_SYMBOL_REGISTER(x) inter_module_register(#x, THIS_MODULE, &x)
 333 #define PORTAL_SYMBOL_UNREGISTER(x) inter_module_unregister(#x)
 334
 335 #define PORTAL_SYMBOL_GET(x) ((typeof(&x))inter_module_get(#x))
 336 #define PORTAL_SYMBOL_PUT(x) inter_module_put(#x)
 337
 338 #define PORTAL_MODULE_USE       MOD_INC_USE_COUNT
 339 #define PORTAL_MODULE_UNUSE     MOD_DEC_USE_COUNT
 340 #else
 341
 342 #define PORTAL_SYMBOL_REGISTER(x)
 343 #define PORTAL_SYMBOL_UNREGISTER(x)
 344
 345 #define PORTAL_SYMBOL_GET(x) symbol_get(x)
 346 #define PORTAL_SYMBOL_PUT(x) symbol_put(x)
 347
 348 #define PORTAL_MODULE_USE       try_module_get(THIS_MODULE)
 349 #define PORTAL_MODULE_UNUSE     module_put(THIS_MODULE)
 350
 351 #endif
 352
 353 /******************************************************************************/
 354 /* Kernel Portals Router interface */
 355
 356 typedef void (*kpr_fwd_callback_t)(void *arg, int error); // completion callback
 357
 358 /* space for routing targets to stash "stuff" in a forwarded packet */
 359 typedef union {
 360         long long        _alignment;
 361         void            *_space[16];            /* scale with CPU arch */
 362 } kprfd_scratch_t;
 363
 364 /* Kernel Portals Routing Forwarded message Descriptor */
 365 typedef struct {
 366         struct list_head     kprfd_list;        /* stash in queues (routing target can use) */
 367         ptl_nid_t            kprfd_target_nid;  /* final destination NID */
 368         ptl_nid_t            kprfd_gateway_nid; /* gateway NID */
 369         int                  kprfd_nob;         /* # message bytes (including header) */
 370         int                  kprfd_niov;        /* # message frags (including header) */
 371         struct iovec        *kprfd_iov;         /* message fragments */
 372         void                *kprfd_router_arg;  // originating NAL's router arg
 373         kpr_fwd_callback_t   kprfd_callback;    /* completion callback */
 374         void                *kprfd_callback_arg; /* completion callback arg */
 375         kprfd_scratch_t      kprfd_scratch;    // scratchpad for routing targets
 376 } kpr_fwd_desc_t;
 377
 378 typedef void  (*kpr_fwd_t)(void *arg, kpr_fwd_desc_t *fwd);
 379 typedef void  (*kpr_notify_t)(void *arg, ptl_nid_t peer, int alive);
 380
 381 /* NAL's routing interface (Kernel Portals Routing Nal Interface) */
 382 typedef const struct {
 383         int             kprni_nalid;    /* NAL's id */
 384         void           *kprni_arg;      /* Arg to pass when calling into NAL */
 385         kpr_fwd_t       kprni_fwd;      /* NAL's forwarding entrypoint */
 386         kpr_notify_t    kprni_notify;   /* NAL's notification entrypoint */
 387 } kpr_nal_interface_t;
 388
 389 /* Router's routing interface (Kernel Portals Routing Router Interface) */
 390 typedef const struct {
 391         /* register the calling NAL with the router and get back the handle for
 392          * subsequent calls */
 393         int     (*kprri_register) (kpr_nal_interface_t *nal_interface,
 394                                    void **router_arg);
 395
 396         /* ask the router to find a gateway that forwards to 'nid' and is a
 397          * peer of the calling NAL; assume caller will send 'nob' bytes of
 398          * payload there */
 399         int     (*kprri_lookup) (void *router_arg, ptl_nid_t nid, int nob,
 400                                  ptl_nid_t *gateway_nid);
 401
 402         /* hand a packet over to the router for forwarding */
 403         kpr_fwd_t kprri_fwd_start;
 404
 405         /* hand a packet back to the router for completion */
 406         void    (*kprri_fwd_done) (void *router_arg, kpr_fwd_desc_t *fwd,
 407                                    int error);
 408
 409         /* notify the router about peer state */
 410         void    (*kprri_notify) (void *router_arg, ptl_nid_t peer,
 411                                  int alive, time_t when);
 412
 413         /* the calling NAL is shutting down */
 414         void    (*kprri_shutdown) (void *router_arg);
 415
 416         /* deregister the calling NAL with the router */
 417         void    (*kprri_deregister) (void *router_arg);
 418
 419 } kpr_router_interface_t;
 420
 421 /* Convenient struct for NAL to stash router interface/args */
 422 typedef struct {
 423         kpr_router_interface_t  *kpr_interface;
 424         void                    *kpr_arg;
 425 } kpr_router_t;
 426
 427 /* Router's control interface (Kernel Portals Routing Control Interface) */
 428 typedef const struct {
 429         int     (*kprci_add_route)(int gateway_nal, ptl_nid_t gateway_nid,
 430                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
 431         int     (*kprci_del_route)(int gateway_nal, ptl_nid_t gateway_nid,
 432                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
 433         int     (*kprci_get_route)(int index, int *gateway_nal,
 434                                    ptl_nid_t *gateway,
 435                                    ptl_nid_t *lo_nid, ptl_nid_t *hi_nid,
 436                                    int *alive);
 437         int     (*kprci_notify)(int gateway_nal, ptl_nid_t gateway_nid,
 438                                 int alive, time_t when);
 439 } kpr_control_interface_t;
 440
 441 extern kpr_control_interface_t  kpr_control_interface;
 442 extern kpr_router_interface_t   kpr_router_interface;
 443
 444 static inline int
 445 kpr_register (kpr_router_t *router, kpr_nal_interface_t *nalif)
 446 {
 447         int    rc;
 448
 449         router->kpr_interface = PORTAL_SYMBOL_GET (kpr_router_interface);
 450         if (router->kpr_interface == NULL)
 451                 return (-ENOENT);
 452
 453         rc = (router->kpr_interface)->kprri_register (nalif, &router->kpr_arg);
 454         if (rc != 0)
 455                 router->kpr_interface = NULL;
 456
 457         PORTAL_SYMBOL_PUT (kpr_router_interface);
 458         return (rc);
 459 }
 460
 461 static inline int
 462 kpr_routing (kpr_router_t *router)
 463 {
 464         return (router->kpr_interface != NULL);
 465 }
 466
 467 static inline int
 468 kpr_lookup (kpr_router_t *router, ptl_nid_t nid, int nob, ptl_nid_t *gateway_nid)
 469 {
 470         if (!kpr_routing (router))
 471                 return (-ENETUNREACH);
 472
 473         return (router->kpr_interface->kprri_lookup(router->kpr_arg, nid, nob,
 474                                                     gateway_nid));
 475 }
 476
 477 static inline void
 478 kpr_fwd_init (kpr_fwd_desc_t *fwd, ptl_nid_t nid,
 479               int nob, int niov, struct iovec *iov,
 480               kpr_fwd_callback_t callback, void *callback_arg)
 481 {
 482         fwd->kprfd_target_nid   = nid;
 483         fwd->kprfd_gateway_nid  = nid;
 484         fwd->kprfd_nob          = nob;
 485         fwd->kprfd_niov         = niov;
 486         fwd->kprfd_iov          = iov;
 487         fwd->kprfd_callback     = callback;
 488         fwd->kprfd_callback_arg = callback_arg;
 489 }
 490
 491 static inline void
 492 kpr_fwd_start (kpr_router_t *router, kpr_fwd_desc_t *fwd)
 493 {
 494         if (!kpr_routing (router))
 495                 fwd->kprfd_callback (fwd->kprfd_callback_arg, -ENETUNREACH);
 496         else
 497                 router->kpr_interface->kprri_fwd_start (router->kpr_arg, fwd);
 498 }
 499
 500 static inline void
 501 kpr_fwd_done (kpr_router_t *router, kpr_fwd_desc_t *fwd, int error)
 502 {
 503         LASSERT (kpr_routing (router));
 504         router->kpr_interface->kprri_fwd_done (router->kpr_arg, fwd, error);
 505 }
 506
 507 static inline void
 508 kpr_notify (kpr_router_t *router,
 509             ptl_nid_t peer, int alive, time_t when)
 510 {
 511         if (!kpr_routing (router))
 512                 return;
 513
 514         router->kpr_interface->kprri_notify(router->kpr_arg, peer, alive, when);
 515 }
 516
 517 static inline void
 518 kpr_shutdown (kpr_router_t *router)
 519 {
 520         if (kpr_routing (router))
 521                 router->kpr_interface->kprri_shutdown (router->kpr_arg);
 522 }
 523
 524 static inline void
 525 kpr_deregister (kpr_router_t *router)
 526 {
 527         if (!kpr_routing (router))
 528                 return;
 529         router->kpr_interface->kprri_deregister (router->kpr_arg);
 530         router->kpr_interface = NULL;
 531 }
 532
 533 /******************************************************************************/
 534
 535 #ifdef PORTALS_PROFILING
 536 #define prof_enum(FOO) PROF__##FOO
 537 enum {
 538         prof_enum(our_recvmsg),
 539         prof_enum(our_sendmsg),
 540         prof_enum(socknal_recv),
 541         prof_enum(lib_parse),
 542         prof_enum(conn_list_walk),
 543         prof_enum(memcpy),
 544         prof_enum(lib_finalize),
 545         prof_enum(pingcli_time),
 546         prof_enum(gmnal_send),
 547         prof_enum(gmnal_recv),
 548         MAX_PROFS
 549 };
 550
 551 struct prof_ent {
 552         char *str;
 553         /* hrmph.  wrap-tastic. */
 554         u32       starts;
 555         u32       finishes;
 556         cycles_t  total_cycles;
 557         cycles_t  start;
 558         cycles_t  end;
 559 };
 560
 561 extern struct prof_ent prof_ents[MAX_PROFS];
 562
 563 #define PROF_START(FOO)                                         \
 564         do {                                                    \
 565                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
 566                 pe->starts++;                                   \
 567                 pe->start = get_cycles();                       \
 568         } while (0)
 569
 570 #define PROF_FINISH(FOO)                                        \
 571         do {                                                    \
 572                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
 573                 pe->finishes++;                                 \
 574                 pe->end = get_cycles();                         \
 575                 pe->total_cycles += (pe->end - pe->start);      \
 576         } while (0)
 577 #else /* !PORTALS_PROFILING */
 578 #define PROF_START(FOO) do {} while(0)
 579 #define PROF_FINISH(FOO) do {} while(0)
 580 #endif /* PORTALS_PROFILING */
 581
 582 /* debug.c */
 583 extern spinlock_t stack_backtrace_lock;
 584
 585 char *portals_debug_dumpstack(void);
 586 void portals_run_upcall(char **argv);
 587 void portals_run_lbug_upcall(char * file, const char *fn, const int line);
 588 void portals_debug_dumplog(void);
 589 int portals_debug_init(unsigned long bufsize);
 590 int portals_debug_cleanup(void);
 591 int portals_debug_clear_buffer(void);
 592 int portals_debug_mark_buffer(char *text);
 593 int portals_debug_set_daemon(unsigned int cmd, unsigned int length,
 594                              char *file, unsigned int size);
 595 __s32 portals_debug_copy_to_user(char *buf, unsigned long len);
 596 #if (__GNUC__)
 597 /* Use the special GNU C __attribute__ hack to have the compiler check the
 598  * printf style argument string against the actual argument count and
 599  * types.
 600  */
 601 #ifdef printf
 602 # warning printf has been defined as a macro...
 603 # undef printf
 604 #endif
 605 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
 606                        const int line, unsigned long stack,
 607                        char *format, ...)
 608         __attribute__ ((format (printf, 7, 8)));
 609 #else
 610 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
 611                        const int line, unsigned long stack,
 612                        const char *format, ...);
 613 #endif /* __GNUC__ */
 614 void portals_debug_set_level(unsigned int debug_level);
 615
 616 # define fprintf(a, format, b...) CDEBUG(D_OTHER, format , ## b)
 617 # define printf(format, b...) CDEBUG(D_OTHER, format , ## b)
 618 # define time(a) CURRENT_TIME
 619
 620 extern void kportal_daemonize (char *name);
 621 extern void kportal_blockallsigs (void);
 622
 623 #else  /* !__KERNEL__ */
 624 # include <stdio.h>
 625 # include <stdlib.h>
 626 #ifndef __CYGWIN__
 627 # include <stdint.h>
 628 #else
 629 # include <cygwin-ioctl.h>
 630 #endif
 631 # include <unistd.h>
 632 # include <time.h>
 633 # include <asm/types.h>
 634 # ifndef DEBUG_SUBSYSTEM
 635 #  define DEBUG_SUBSYSTEM S_UNDEFINED
 636 # endif
 637 # ifdef PORTAL_DEBUG
 638 #  undef NDEBUG
 639 #  include <assert.h>
 640 #  define LASSERT(e)     assert(e)
 641 #  define LASSERTF(cond, args...)     assert(cond)
 642 # else
 643 #  define LASSERT(e)
 644 #  define LASSERTF(cond, args...) do { } while (0)
 645 # endif
 646 # define printk(format, args...) printf (format, ## args)
 647 # define PORTAL_ALLOC(ptr, size) do { (ptr) = malloc(size); } while (0);
 648 # define PORTAL_FREE(a, b) do { free(a); } while (0);
 649 # define portals_debug_msg(subsys, mask, file, fn, line, stack, format, a...) \
 650     printf("%02x:%06x (@%lu %s:%s,l. %d %d %lu): " format,                    \
 651            (subsys), (mask), (long)time(0), file, fn, line,                   \
 652            getpid() , stack, ## a);
 653 #endif
 654
 655 /* support decl needed both by kernel and liblustre */
 656 char *portals_nid2str(int nal, ptl_nid_t nid, char *str);
 657
 658 #ifndef CURRENT_TIME
 659 # define CURRENT_TIME time(0)
 660 #endif
 661
 662 /******************************************************************************/
 663 /* Light-weight trace
 664  * Support for temporary event tracing with minimal Heisenberg effect. */
 665 #define LWT_SUPPORT  0
 666
 667 #define LWT_MEMORY   (64<<20)
 668 #define LWT_MAX_CPUS 4
 669
 670 typedef struct {
 671         cycles_t    lwte_when;
 672         char       *lwte_where;
 673         void       *lwte_task;
 674         long        lwte_p1;
 675         long        lwte_p2;
 676         long        lwte_p3;
 677         long        lwte_p4;
 678 #if BITS_PER_LONG > 32
 679         long        lwte_pad;
 680 #endif
 681 } lwt_event_t;
 682
 683 #if LWT_SUPPORT
 684 #ifdef __KERNEL__
 685 #define LWT_EVENTS_PER_PAGE (PAGE_SIZE / sizeof (lwt_event_t))
 686
 687 typedef struct _lwt_page {
 688         struct list_head     lwtp_list;
 689         struct page         *lwtp_page;
 690         lwt_event_t         *lwtp_events;
 691 } lwt_page_t;
 692
 693 typedef struct {
 694         int                lwtc_current_index;
 695         lwt_page_t        *lwtc_current_page;
 696 } lwt_cpu_t;
 697
 698 extern int       lwt_enabled;
 699 extern lwt_cpu_t lwt_cpus[];
 700
 701 extern int  lwt_init (void);
 702 extern void lwt_fini (void);
 703 extern int  lwt_lookup_string (int *size, char *knlptr,
 704                                char *usrptr, int usrsize);
 705 extern int  lwt_control (int enable, int clear);
 706 extern int  lwt_snapshot (cycles_t *now, int *ncpu, int *total_size,
 707                           void *user_ptr, int user_size);
 708
 709 /* Note that we _don't_ define LWT_EVENT at all if LWT_SUPPORT isn't set.
 710  * This stuff is meant for finding specific problems; it never stays in
 711  * production code... */
 712
 713 #define LWTSTR(n)       #n
 714 #define LWTWHERE(f,l)   f ":" LWTSTR(l)
 715
 716 #define LWT_EVENT(p1, p2, p3, p4)                                       \
 717 do {                                                                    \
 718         unsigned long    flags;                                         \
 719         lwt_cpu_t       *cpu;                                           \
 720         lwt_page_t      *p;                                             \
 721         lwt_event_t     *e;                                             \
 722                                                                         \
 723         local_irq_save (flags);                                         \
 724                                                                         \
 725         if (lwt_enabled) {                                              \
 726                 cpu = &lwt_cpus[smp_processor_id()];                    \
 727                 p = cpu->lwtc_current_page;                             \
 728                 e = &p->lwtp_events[cpu->lwtc_current_index++];         \
 729                                                                         \
 730                 if (cpu->lwtc_current_index >= LWT_EVENTS_PER_PAGE) {   \
 731                         cpu->lwtc_current_page =                        \
 732                                 list_entry (p->lwtp_list.next,          \
 733                                             lwt_page_t, lwtp_list);     \
 734                         cpu->lwtc_current_index = 0;                    \
 735                 }                                                       \
 736                                                                         \
 737                 e->lwte_when  = get_cycles();                           \
 738                 e->lwte_where = LWTWHERE(__FILE__,__LINE__);            \
 739                 e->lwte_task  = current;                                \
 740                 e->lwte_p1    = (long)(p1);                             \
 741                 e->lwte_p2    = (long)(p2);                             \
 742                 e->lwte_p3    = (long)(p3);                             \
 743                 e->lwte_p4    = (long)(p4);                             \
 744         }                                                               \
 745                                                                         \
 746         local_irq_restore (flags);                                      \
 747 } while (0)
 748 #else  /* __KERNEL__ */
 749 #define LWT_EVENT(p1,p2,p3,p4)     /* no userland implementation yet */
 750 #endif /* __KERNEL__ */
 751 #endif /* LWT_SUPPORT */
 752
 753 struct portals_device_userstate
 754 {
 755         int          pdu_memhog_pages;
 756         struct page *pdu_memhog_root_page;
 757 };
 758
 759 #include <linux/portals_lib.h>
 760
 761 /*
 762  * USER LEVEL STUFF BELOW
 763  */
 764
 765 #define PORTALS_CFG_VERSION 0x00010001;
 766
 767 struct portals_cfg {
 768         __u32 pcfg_version;
 769         __u32 pcfg_command;
 770
 771         __u32 pcfg_nal;
 772         __u32 pcfg_flags;
 773
 774         __u32 pcfg_gw_nal;
 775         __u64 pcfg_nid;
 776         __u64 pcfg_nid2;
 777         __u64 pcfg_nid3;
 778         __u32 pcfg_id;
 779         __u32 pcfg_misc;
 780         __u32 pcfg_fd;
 781         __u32 pcfg_count;
 782         __u32 pcfg_size;
 783         __u32 pcfg_wait;
 784
 785         __u32 pcfg_plen1; /* buffers in userspace */
 786         char *pcfg_pbuf1;
 787         __u32 pcfg_plen2; /* buffers in userspace */
 788         char *pcfg_pbuf2;
 789 };
 790
 791 #define PCFG_INIT(pcfg, cmd)                            \
 792 do {                                                    \
 793         memset(&pcfg, 0, sizeof(pcfg));                 \
 794         pcfg.pcfg_version = PORTALS_CFG_VERSION;        \
 795         pcfg.pcfg_command = (cmd);                      \
 796                                                         \
 797 } while (0)
 798
 799 #define PORTAL_IOCTL_VERSION 0x00010007
 800 #define PING_SYNC       0
 801 #define PING_ASYNC      1
 802
 803 struct portal_ioctl_data {
 804         __u32 ioc_len;
 805         __u32 ioc_version;
 806         __u64 ioc_nid;
 807         __u64 ioc_nid2;
 808         __u64 ioc_nid3;
 809         __u32 ioc_count;
 810         __u32 ioc_nal;
 811         __u32 ioc_nal_cmd;
 812         __u32 ioc_fd;
 813         __u32 ioc_id;
 814
 815         __u32 ioc_flags;
 816         __u32 ioc_size;
 817
 818         __u32 ioc_wait;
 819         __u32 ioc_timeout;
 820         __u32 ioc_misc;
 821
 822         __u32 ioc_inllen1;
 823         char *ioc_inlbuf1;
 824         __u32 ioc_inllen2;
 825         char *ioc_inlbuf2;
 826
 827         __u32 ioc_plen1; /* buffers in userspace */
 828         char *ioc_pbuf1;
 829         __u32 ioc_plen2; /* buffers in userspace */
 830         char *ioc_pbuf2;
 831
 832         char ioc_bulk[0];
 833 };
 834
 835 struct portal_ioctl_hdr {
 836         __u32 ioc_len;
 837         __u32 ioc_version;
 838 };
 839
 840 struct portals_debug_ioctl_data
 841 {
 842         struct portal_ioctl_hdr hdr;
 843         unsigned int subs;
 844         unsigned int debug;
 845 };
 846
 847 #define PORTAL_IOC_INIT(data)                           \
 848 do {                                                    \
 849         memset(&data, 0, sizeof(data));                 \
 850         data.ioc_version = PORTAL_IOCTL_VERSION;        \
 851         data.ioc_len = sizeof(data);                    \
 852 } while (0)
 853
 854 /* FIXME check conflict with lustre_lib.h */
 855 #define PTL_IOC_DEBUG_MASK             _IOWR('f', 250, long)
 856
 857 static inline int portal_ioctl_packlen(struct portal_ioctl_data *data)
 858 {
 859         int len = sizeof(*data);
 860         len += size_round(data->ioc_inllen1);
 861         len += size_round(data->ioc_inllen2);
 862         return len;
 863 }
 864
 865 static inline int portal_ioctl_is_invalid(struct portal_ioctl_data *data)
 866 {
 867         if (data->ioc_len > (1<<30)) {
 868                 CERROR ("PORTALS ioctl: ioc_len larger than 1<<30\n");
 869                 return 1;
 870         }
 871         if (data->ioc_inllen1 > (1<<30)) {
 872                 CERROR ("PORTALS ioctl: ioc_inllen1 larger than 1<<30\n");
 873                 return 1;
 874         }
 875         if (data->ioc_inllen2 > (1<<30)) {
 876                 CERROR ("PORTALS ioctl: ioc_inllen2 larger than 1<<30\n");
 877                 return 1;
 878         }
 879         if (data->ioc_inlbuf1 && !data->ioc_inllen1) {
 880                 CERROR ("PORTALS ioctl: inlbuf1 pointer but 0 length\n");
 881                 return 1;
 882         }
 883         if (data->ioc_inlbuf2 && !data->ioc_inllen2) {
 884                 CERROR ("PORTALS ioctl: inlbuf2 pointer but 0 length\n");
 885                 return 1;
 886         }
 887         if (data->ioc_pbuf1 && !data->ioc_plen1) {
 888                 CERROR ("PORTALS ioctl: pbuf1 pointer but 0 length\n");
 889                 return 1;
 890         }
 891         if (data->ioc_pbuf2 && !data->ioc_plen2) {
 892                 CERROR ("PORTALS ioctl: pbuf2 pointer but 0 length\n");
 893                 return 1;
 894         }
 895         if (data->ioc_plen1 && !data->ioc_pbuf1) {
 896                 CERROR ("PORTALS ioctl: plen1 nonzero but no pbuf1 pointer\n");
 897                 return 1;
 898         }
 899         if (data->ioc_plen2 && !data->ioc_pbuf2) {
 900                 CERROR ("PORTALS ioctl: plen2 nonzero but no pbuf2 pointer\n");
 901                 return 1;
 902         }
 903         if (portal_ioctl_packlen(data) != data->ioc_len ) {
 904                 CERROR ("PORTALS ioctl: packlen != ioc_len\n");
 905                 return 1;
 906         }
 907         if (data->ioc_inllen1 &&
 908             data->ioc_bulk[data->ioc_inllen1 - 1] != '\0') {
 909                 CERROR ("PORTALS ioctl: inlbuf1 not 0 terminated\n");
 910                 return 1;
 911         }
 912         if (data->ioc_inllen2 &&
 913             data->ioc_bulk[size_round(data->ioc_inllen1) +
 914                            data->ioc_inllen2 - 1] != '\0') {
 915                 CERROR ("PORTALS ioctl: inlbuf2 not 0 terminated\n");
 916                 return 1;
 917         }
 918         return 0;
 919 }
 920
 921 #ifndef __KERNEL__
 922 static inline int portal_ioctl_pack(struct portal_ioctl_data *data, char **pbuf,
 923                                     int max)
 924 {
 925         char *ptr;
 926         struct portal_ioctl_data *overlay;
 927         data->ioc_len = portal_ioctl_packlen(data);
 928         data->ioc_version = PORTAL_IOCTL_VERSION;
 929
 930         if (*pbuf && portal_ioctl_packlen(data) > max)
 931                 return 1;
 932         if (*pbuf == NULL) {
 933                 *pbuf = malloc(data->ioc_len);
 934         }
 935         if (!*pbuf)
 936                 return 1;
 937         overlay = (struct portal_ioctl_data *)*pbuf;
 938         memcpy(*pbuf, data, sizeof(*data));
 939
 940         ptr = overlay->ioc_bulk;
 941         if (data->ioc_inlbuf1)
 942                 LOGL(data->ioc_inlbuf1, data->ioc_inllen1, ptr);
 943         if (data->ioc_inlbuf2)
 944                 LOGL(data->ioc_inlbuf2, data->ioc_inllen2, ptr);
 945         if (portal_ioctl_is_invalid(overlay))
 946                 return 1;
 947
 948         return 0;
 949 }
 950 #else
 951 #include <asm/uaccess.h>
 952
 953 /* buffer MUST be at least the size of portal_ioctl_hdr */
 954 static inline int portal_ioctl_getdata(char *buf, char *end, void *arg)
 955 {
 956         struct portal_ioctl_hdr *hdr;
 957         struct portal_ioctl_data *data;
 958         int err;
 959         ENTRY;
 960
 961         hdr = (struct portal_ioctl_hdr *)buf;
 962         data = (struct portal_ioctl_data *)buf;
 963
 964         err = copy_from_user(buf, (void *)arg, sizeof(*hdr));
 965         if ( err ) {
 966                 EXIT;
 967                 return err;
 968         }
 969
 970         if (hdr->ioc_version != PORTAL_IOCTL_VERSION) {
 971                 CERROR ("PORTALS: version mismatch kernel vs application\n");
 972                 return -EINVAL;
 973         }
 974
 975         if (hdr->ioc_len + buf >= end) {
 976                 CERROR ("PORTALS: user buffer exceeds kernel buffer\n");
 977                 return -EINVAL;
 978         }
 979
 980
 981         if (hdr->ioc_len < sizeof(struct portal_ioctl_data)) {
 982                 CERROR ("PORTALS: user buffer too small for ioctl\n");
 983                 return -EINVAL;
 984         }
 985
 986         err = copy_from_user(buf, (void *)arg, hdr->ioc_len);
 987         if ( err ) {
 988                 EXIT;
 989                 return err;
 990         }
 991
 992         if (portal_ioctl_is_invalid(data)) {
 993                 CERROR ("PORTALS: ioctl not correctly formatted\n");
 994                 return -EINVAL;
 995         }
 996
 997         if (data->ioc_inllen1) {
 998                 data->ioc_inlbuf1 = &data->ioc_bulk[0];
 999         }
1000
1001         if (data->ioc_inllen2) {
1002                 data->ioc_inlbuf2 = &data->ioc_bulk[0] +
1003                         size_round(data->ioc_inllen1);
1004         }
1005
1006         EXIT;
1007         return 0;
1008 }
1009 #endif
1010
1011 /* ioctls for manipulating snapshots 30- */
1012 #define IOC_PORTAL_TYPE                   'e'
1013 #define IOC_PORTAL_MIN_NR                 30
1014
1015 #define IOC_PORTAL_PING                    _IOWR('e', 30, long)
1016 #define IOC_PORTAL_GET_DEBUG               _IOWR('e', 31, long)
1017 #define IOC_PORTAL_CLEAR_DEBUG             _IOWR('e', 32, long)
1018 #define IOC_PORTAL_MARK_DEBUG              _IOWR('e', 33, long)
1019 #define IOC_PORTAL_PANIC                   _IOWR('e', 34, long)
1020 #define IOC_PORTAL_NAL_CMD                 _IOWR('e', 35, long)
1021 #define IOC_PORTAL_GET_NID                 _IOWR('e', 36, long)
1022 #define IOC_PORTAL_FAIL_NID                _IOWR('e', 37, long)
1023 #define IOC_PORTAL_SET_DAEMON              _IOWR('e', 38, long)
1024 #define IOC_PORTAL_LWT_CONTROL             _IOWR('e', 39, long)
1025 #define IOC_PORTAL_LWT_SNAPSHOT            _IOWR('e', 40, long)
1026 #define IOC_PORTAL_LWT_LOOKUP_STRING       _IOWR('e', 41, long)
1027 #define IOC_PORTAL_MEMHOG                  _IOWR('e', 42, long)
1028 #define IOC_PORTAL_MAX_NR                             42
1029
1030 enum {
1031         QSWNAL    = 1,
1032         SOCKNAL   = 2,
1033         GMNAL     = 3,
1034         /*          4 unused */
1035         TCPNAL    = 5,
1036         SCIMACNAL = 6,
1037         ROUTER    = 7,
1038         IBNAL     = 8,
1039         NAL_ENUM_END_MARKER
1040 };
1041
1042 #ifdef __KERNEL__
1043 extern ptl_handle_ni_t  kqswnal_ni;
1044 extern ptl_handle_ni_t  ksocknal_ni;
1045 extern ptl_handle_ni_t  kgmnal_ni;
1046 extern ptl_handle_ni_t  kibnal_ni;
1047 extern ptl_handle_ni_t  kscimacnal_ni;
1048 #endif
1049
1050 #define PTL_NALFMT_SIZE         16
1051
1052 #define NAL_MAX_NR (NAL_ENUM_END_MARKER - 1)
1053
1054 #define NAL_CMD_REGISTER_PEER_FD     100
1055 #define NAL_CMD_CLOSE_CONNECTION     101
1056 #define NAL_CMD_REGISTER_MYNID       102
1057 #define NAL_CMD_PUSH_CONNECTION      103
1058 #define NAL_CMD_GET_CONN             104
1059 #define NAL_CMD_DEL_AUTOCONN         105
1060 #define NAL_CMD_ADD_AUTOCONN         106
1061 #define NAL_CMD_GET_AUTOCONN         107
1062 #define NAL_CMD_GET_TXDESC           108
1063 #define NAL_CMD_ADD_ROUTE            109
1064 #define NAL_CMD_DEL_ROUTE            110
1065 #define NAL_CMD_GET_ROUTE            111
1066 #define NAL_CMD_NOTIFY_ROUTER        112
1067
1068 enum {
1069         DEBUG_DAEMON_START       =  1,
1070         DEBUG_DAEMON_STOP        =  2,
1071         DEBUG_DAEMON_PAUSE       =  3,
1072         DEBUG_DAEMON_CONTINUE    =  4,
1073 };
1074
1075 /* XXX remove to lustre ASAP */
1076 struct lustre_peer {
1077         ptl_nid_t       peer_nid;
1078         ptl_handle_ni_t peer_ni;
1079 };
1080
1081
1082 /* module.c */
1083 typedef int (*nal_cmd_handler_t)(struct portals_cfg *, void * private);
1084 int kportal_nal_register(int nal, nal_cmd_handler_t handler, void * private);
1085 int kportal_nal_unregister(int nal);
1086
1087 enum cfg_record_type {
1088         PORTALS_CFG_TYPE = 1,
1089         LUSTRE_CFG_TYPE = 123,
1090 };
1091
1092 typedef int (*cfg_record_cb_t)(enum cfg_record_type, int len, void *data);
1093 int kportal_nal_cmd(struct portals_cfg *);
1094
1095 ptl_handle_ni_t *kportal_get_ni (int nal);
1096 void kportal_put_ni (int nal);
1097
1098 #ifdef __CYGWIN__
1099 # ifndef BITS_PER_LONG
1100 #  if (~0UL) == 0xffffffffUL
1101 #   define BITS_PER_LONG 32
1102 #  else
1103 #   define BITS_PER_LONG 64
1104 #  endif
1105 # endif
1106 #endif
1107
1108 #if defined(__x86_64__)
1109 # define LPU64 "%Lu"
1110 # define LPD64 "%Ld"
1111 # define LPX64 "%#Lx"
1112 # define LPSZ  "%lu"
1113 # define LPSSZ "%ld"
1114 #elif (BITS_PER_LONG == 32 || __WORDSIZE == 32)
1115 # define LPU64 "%Lu"
1116 # define LPD64 "%Ld"
1117 # define LPX64 "%#Lx"
1118 # define LPSZ  "%u"
1119 # define LPSSZ "%d"
1120 #elif (BITS_PER_LONG == 64 || __WORDSIZE == 64)
1121 # define LPU64 "%lu"
1122 # define LPD64 "%ld"
1123 # define LPX64 "%#lx"
1124 # define LPSZ  "%lu"
1125 # define LPSSZ "%ld"
1126 #endif
1127 #ifndef LPU64
1128 # error "No word size defined"
1129 #endif
1130
1131 #endif