lnet/include/linux/kp30.h

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  */
   4 #ifndef _KP30_INCLUDED
   5 #define _KP30_INCLUDED
   6
   7 #define PORTAL_DEBUG
   8
   9 #ifndef offsetof
  10 # define offsetof(typ,memb)     ((int)((char *)&(((typ *)0)->memb)))
  11 #endif
  12
  13 #define LOWEST_BIT_SET(x)       ((x) & ~((x) - 1))
  14
  15 /*
  16  *  Debugging
  17  */
  18 extern unsigned int portal_subsystem_debug;
  19 extern unsigned int portal_stack;
  20 extern unsigned int portal_debug;
  21 extern unsigned int portal_printk;
  22 extern unsigned int portal_cerror;
  23 /* Debugging subsystems (32 bits, non-overlapping) */
  24 #define S_UNDEFINED    (1 << 0)
  25 #define S_MDC          (1 << 1)
  26 #define S_MDS          (1 << 2)
  27 #define S_OSC          (1 << 3)
  28 #define S_OST          (1 << 4)
  29 #define S_CLASS        (1 << 5)
  30 #define S_LOG          (1 << 6)
  31 #define S_LLITE        (1 << 7)
  32 #define S_RPC          (1 << 8)
  33 #define S_MGMT         (1 << 9)
  34 #define S_PORTALS     (1 << 10)
  35 #define S_SOCKNAL     (1 << 11)
  36 #define S_QSWNAL      (1 << 12)
  37 #define S_PINGER      (1 << 13)
  38 #define S_FILTER      (1 << 14)
  39 #define S_PTLBD       (1 << 15)
  40 #define S_ECHO        (1 << 16)
  41 #define S_LDLM        (1 << 17)
  42 #define S_LOV         (1 << 18)
  43 #define S_GMNAL       (1 << 19)
  44 #define S_PTLROUTER   (1 << 20)
  45 #define S_COBD        (1 << 21)
  46 #define S_IBNAL       (1 << 22)
  47
  48 /* If you change these values, please keep portals/utils/debug.c
  49  * up to date! */
  50
  51 /* Debugging masks (32 bits, non-overlapping) */
  52 #define D_TRACE     (1 << 0) /* ENTRY/EXIT markers */
  53 #define D_INODE     (1 << 1)
  54 #define D_SUPER     (1 << 2)
  55 #define D_EXT2      (1 << 3) /* anything from ext2_debug */
  56 #define D_MALLOC    (1 << 4) /* print malloc, free information */
  57 #define D_CACHE     (1 << 5) /* cache-related items */
  58 #define D_INFO      (1 << 6) /* general information */
  59 #define D_IOCTL     (1 << 7) /* ioctl related information */
  60 #define D_BLOCKS    (1 << 8) /* ext2 block allocation */
  61 #define D_NET       (1 << 9) /* network communications */
  62 #define D_WARNING   (1 << 10) /* CWARN(...) == CDEBUG (D_WARNING, ...) */
  63 #define D_BUFFS     (1 << 11)
  64 #define D_OTHER     (1 << 12)
  65 #define D_DENTRY    (1 << 13)
  66 #define D_PORTALS   (1 << 14) /* ENTRY/EXIT markers */
  67 #define D_PAGE      (1 << 15) /* bulk page handling */
  68 #define D_DLMTRACE  (1 << 16)
  69 #define D_ERROR     (1 << 17) /* CERROR(...) == CDEBUG (D_ERROR, ...) */
  70 #define D_EMERG     (1 << 18) /* CEMERG(...) == CDEBUG (D_EMERG, ...) */
  71 #define D_HA        (1 << 19) /* recovery and failover */
  72 #define D_RPCTRACE  (1 << 20) /* for distributed debugging */
  73 #define D_VFSTRACE  (1 << 21)
  74 #define D_READA     (1 << 22) /* read-ahead */
  75
  76 #ifdef __KERNEL__
  77 # include <linux/sched.h> /* THREAD_SIZE */
  78 #else
  79 # ifndef THREAD_SIZE /* x86_64 has THREAD_SIZE in userspace */
  80 #  define THREAD_SIZE 8192
  81 # endif
  82 #endif
  83
  84 #define LUSTRE_TRACE_SIZE (THREAD_SIZE >> 5)
  85
  86 #ifdef __KERNEL__
  87 # ifdef  __ia64__
  88 #  define CDEBUG_STACK (THREAD_SIZE -                                      \
  89                         ((unsigned long)__builtin_dwarf_cfa() &            \
  90                          (THREAD_SIZE - 1)))
  91 # else
  92 #  define CDEBUG_STACK (THREAD_SIZE -                                      \
  93                         ((unsigned long)__builtin_frame_address(0) &       \
  94                          (THREAD_SIZE - 1)))
  95 # endif
  96
  97 #define CHECK_STACK(stack)                                                    \
  98         do {                                                                  \
  99                 if ((stack) > 3*THREAD_SIZE/4 && (stack) > portal_stack) {    \
 100                         portals_debug_msg(DEBUG_SUBSYSTEM, D_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 ll_invalidate_inode_pages(inode) invalidate_inode_pages(inode)
 188 #define ll_truncate_complete_page(page) truncate_complete_page(page)
 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) \
 206         invalidate_inode_pages((inode)->i_mapping)
 207 #define ll_truncate_complete_page(page) \
 208         truncate_complete_page((page)->mapping, page)
 209 #define wait_on_page wait_on_page_locked
 210 #define our_recalc_sigpending(current) recalc_sigpending()
 211 #define strtok(a,b) strpbrk(a, b)
 212 static inline void our_cond_resched(void)
 213 {
 214         cond_resched();
 215 }
 216 #define work_struct_t      struct work_struct
 217
 218 #endif /* LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0) */
 219
 220 #ifdef PORTAL_DEBUG
 221 extern void kportal_assertion_failed(char *expr, char *file, const char *func,
 222                                      const int line);
 223 #define LASSERT(e) ((e) ? 0 : kportal_assertion_failed( #e , __FILE__,  \
 224                                                         __FUNCTION__, __LINE__))
 225 /* it would be great to dump_stack() here, but some kernels
 226  * export it as show_stack() and I can't be bothered to
 227  * proprely engage in that dance right now */
 228 #define LASSERTF(cond, fmt...)                                                \
 229         do {                                                                  \
 230                 if (unlikely(!(cond))) {                                      \
 231                         portals_debug_msg(0, D_EMERG,  __FILE__, __FUNCTION__,\
 232                                           __LINE__,  CDEBUG_STACK,            \
 233                                           "ASSERTION(" #cond ") failed:" fmt);\
 234                         LBUG();                                               \
 235                 }                                                             \
 236         } while (0)
 237
 238 #else
 239 #define LASSERT(e)
 240 #define LASSERTF(cond, fmt...) 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         int                  kprfd_nob;         /* # message bytes (including header) */
 372         int                  kprfd_niov;        /* # message frags (including header) */
 373         struct iovec        *kprfd_iov;         /* message fragments */
 374         void                *kprfd_router_arg;  // originating NAL's router arg
 375         kpr_fwd_callback_t   kprfd_callback;    /* completion callback */
 376         void                *kprfd_callback_arg; /* completion callback arg */
 377         kprfd_scratch_t      kprfd_scratch;    // scratchpad for routing targets
 378 } kpr_fwd_desc_t;
 379
 380 typedef void  (*kpr_fwd_t)(void *arg, kpr_fwd_desc_t *fwd);
 381 typedef void  (*kpr_notify_t)(void *arg, ptl_nid_t peer, int alive);
 382
 383 /* NAL's routing interface (Kernel Portals Routing Nal Interface) */
 384 typedef const struct {
 385         int             kprni_nalid;    /* NAL's id */
 386         void           *kprni_arg;      /* Arg to pass when calling into NAL */
 387         kpr_fwd_t       kprni_fwd;      /* NAL's forwarding entrypoint */
 388         kpr_notify_t    kprni_notify;   /* NAL's notification entrypoint */
 389 } kpr_nal_interface_t;
 390
 391 /* Router's routing interface (Kernel Portals Routing Router Interface) */
 392 typedef const struct {
 393         /* register the calling NAL with the router and get back the handle for
 394          * subsequent calls */
 395         int     (*kprri_register) (kpr_nal_interface_t *nal_interface,
 396                                    void **router_arg);
 397
 398         /* ask the router to find a gateway that forwards to 'nid' and is a
 399          * peer of the calling NAL; assume caller will send 'nob' bytes of
 400          * payload there */
 401         int     (*kprri_lookup) (void *router_arg, ptl_nid_t nid, int nob,
 402                                  ptl_nid_t *gateway_nid);
 403
 404         /* hand a packet over to the router for forwarding */
 405         kpr_fwd_t kprri_fwd_start;
 406
 407         /* hand a packet back to the router for completion */
 408         void    (*kprri_fwd_done) (void *router_arg, kpr_fwd_desc_t *fwd,
 409                                    int error);
 410
 411         /* notify the router about peer state */
 412         void    (*kprri_notify) (void *router_arg, ptl_nid_t peer,
 413                                  int alive, time_t when);
 414
 415         /* the calling NAL is shutting down */
 416         void    (*kprri_shutdown) (void *router_arg);
 417
 418         /* deregister the calling NAL with the router */
 419         void    (*kprri_deregister) (void *router_arg);
 420
 421 } kpr_router_interface_t;
 422
 423 /* Convenient struct for NAL to stash router interface/args */
 424 typedef struct {
 425         kpr_router_interface_t  *kpr_interface;
 426         void                    *kpr_arg;
 427 } kpr_router_t;
 428
 429 /* Router's control interface (Kernel Portals Routing Control Interface) */
 430 typedef const struct {
 431         int     (*kprci_add_route)(int gateway_nal, ptl_nid_t gateway_nid,
 432                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
 433         int     (*kprci_del_route)(int gateway_nal, ptl_nid_t gateway_nid,
 434                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
 435         int     (*kprci_get_route)(int index, int *gateway_nal,
 436                                    ptl_nid_t *gateway,
 437                                    ptl_nid_t *lo_nid, ptl_nid_t *hi_nid,
 438                                    int *alive);
 439         int     (*kprci_notify)(int gateway_nal, ptl_nid_t gateway_nid,
 440                                 int alive, time_t when);
 441 } kpr_control_interface_t;
 442
 443 extern kpr_control_interface_t  kpr_control_interface;
 444 extern kpr_router_interface_t   kpr_router_interface;
 445
 446 static inline int
 447 kpr_register (kpr_router_t *router, kpr_nal_interface_t *nalif)
 448 {
 449         int    rc;
 450
 451         router->kpr_interface = PORTAL_SYMBOL_GET (kpr_router_interface);
 452         if (router->kpr_interface == NULL)
 453                 return (-ENOENT);
 454
 455         rc = (router->kpr_interface)->kprri_register (nalif, &router->kpr_arg);
 456         if (rc != 0)
 457                 router->kpr_interface = NULL;
 458
 459         PORTAL_SYMBOL_PUT (kpr_router_interface);
 460         return (rc);
 461 }
 462
 463 static inline int
 464 kpr_routing (kpr_router_t *router)
 465 {
 466         return (router->kpr_interface != NULL);
 467 }
 468
 469 static inline int
 470 kpr_lookup (kpr_router_t *router, ptl_nid_t nid, int nob, ptl_nid_t *gateway_nid)
 471 {
 472         if (!kpr_routing (router))
 473                 return (-ENETUNREACH);
 474
 475         return (router->kpr_interface->kprri_lookup(router->kpr_arg, nid, nob,
 476                                                     gateway_nid));
 477 }
 478
 479 static inline void
 480 kpr_fwd_init (kpr_fwd_desc_t *fwd, ptl_nid_t nid,
 481               int nob, int niov, struct iovec *iov,
 482               kpr_fwd_callback_t callback, void *callback_arg)
 483 {
 484         fwd->kprfd_target_nid   = nid;
 485         fwd->kprfd_gateway_nid  = nid;
 486         fwd->kprfd_nob          = nob;
 487         fwd->kprfd_niov         = niov;
 488         fwd->kprfd_iov          = iov;
 489         fwd->kprfd_callback     = callback;
 490         fwd->kprfd_callback_arg = callback_arg;
 491 }
 492
 493 static inline void
 494 kpr_fwd_start (kpr_router_t *router, kpr_fwd_desc_t *fwd)
 495 {
 496         if (!kpr_routing (router))
 497                 fwd->kprfd_callback (fwd->kprfd_callback_arg, -ENETUNREACH);
 498         else
 499                 router->kpr_interface->kprri_fwd_start (router->kpr_arg, fwd);
 500 }
 501
 502 static inline void
 503 kpr_fwd_done (kpr_router_t *router, kpr_fwd_desc_t *fwd, int error)
 504 {
 505         LASSERT (kpr_routing (router));
 506         router->kpr_interface->kprri_fwd_done (router->kpr_arg, fwd, error);
 507 }
 508
 509 static inline void
 510 kpr_notify (kpr_router_t *router,
 511             ptl_nid_t peer, int alive, time_t when)
 512 {
 513         if (!kpr_routing (router))
 514                 return;
 515
 516         router->kpr_interface->kprri_notify(router->kpr_arg, peer, alive, when);
 517 }
 518
 519 static inline void
 520 kpr_shutdown (kpr_router_t *router)
 521 {
 522         if (kpr_routing (router))
 523                 router->kpr_interface->kprri_shutdown (router->kpr_arg);
 524 }
 525
 526 static inline void
 527 kpr_deregister (kpr_router_t *router)
 528 {
 529         if (!kpr_routing (router))
 530                 return;
 531         router->kpr_interface->kprri_deregister (router->kpr_arg);
 532         router->kpr_interface = NULL;
 533 }
 534
 535 /******************************************************************************/
 536
 537 #ifdef PORTALS_PROFILING
 538 #define prof_enum(FOO) PROF__##FOO
 539 enum {
 540         prof_enum(our_recvmsg),
 541         prof_enum(our_sendmsg),
 542         prof_enum(socknal_recv),
 543         prof_enum(lib_parse),
 544         prof_enum(conn_list_walk),
 545         prof_enum(memcpy),
 546         prof_enum(lib_finalize),
 547         prof_enum(pingcli_time),
 548         prof_enum(gmnal_send),
 549         prof_enum(gmnal_recv),
 550         MAX_PROFS
 551 };
 552
 553 struct prof_ent {
 554         char *str;
 555         /* hrmph.  wrap-tastic. */
 556         u32       starts;
 557         u32       finishes;
 558         cycles_t  total_cycles;
 559         cycles_t  start;
 560         cycles_t  end;
 561 };
 562
 563 extern struct prof_ent prof_ents[MAX_PROFS];
 564
 565 #define PROF_START(FOO)                                         \
 566         do {                                                    \
 567                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
 568                 pe->starts++;                                   \
 569                 pe->start = get_cycles();                       \
 570         } while (0)
 571
 572 #define PROF_FINISH(FOO)                                        \
 573         do {                                                    \
 574                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
 575                 pe->finishes++;                                 \
 576                 pe->end = get_cycles();                         \
 577                 pe->total_cycles += (pe->end - pe->start);      \
 578         } while (0)
 579 #else /* !PORTALS_PROFILING */
 580 #define PROF_START(FOO) do {} while(0)
 581 #define PROF_FINISH(FOO) do {} while(0)
 582 #endif /* PORTALS_PROFILING */
 583
 584 /* debug.c */
 585 extern spinlock_t stack_backtrace_lock;
 586
 587 char *portals_debug_dumpstack(void);
 588 void portals_run_upcall(char **argv);
 589 void portals_run_lbug_upcall(char * file, const char *fn, const int line);
 590 void portals_debug_dumplog(void);
 591 int portals_debug_init(unsigned long bufsize);
 592 int portals_debug_cleanup(void);
 593 int portals_debug_clear_buffer(void);
 594 int portals_debug_mark_buffer(char *text);
 595 int portals_debug_set_daemon(unsigned int cmd, unsigned int length,
 596                              char *file, unsigned int size);
 597 __s32 portals_debug_copy_to_user(char *buf, unsigned long len);
 598 #if (__GNUC__)
 599 /* Use the special GNU C __attribute__ hack to have the compiler check the
 600  * printf style argument string against the actual argument count and
 601  * types.
 602  */
 603 #ifdef printf
 604 # warning printf has been defined as a macro...
 605 # undef printf
 606 #endif
 607 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
 608                        const int line, unsigned long stack,
 609                        char *format, ...)
 610         __attribute__ ((format (printf, 7, 8)));
 611 #else
 612 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
 613                        const int line, unsigned long stack,
 614                        const char *format, ...);
 615 #endif /* __GNUC__ */
 616 void portals_debug_set_level(unsigned int debug_level);
 617
 618 # define fprintf(a, format, b...) CDEBUG(D_OTHER, format , ## b)
 619 # define printf(format, b...) CDEBUG(D_OTHER, format , ## b)
 620 # define time(a) CURRENT_TIME
 621
 622 extern void kportal_daemonize (char *name);
 623 extern void kportal_blockallsigs (void);
 624
 625 #else  /* !__KERNEL__ */
 626 # include <stdio.h>
 627 # include <stdlib.h>
 628 #ifndef __CYGWIN__
 629 # include <stdint.h>
 630 #else
 631 # include <cygwin-ioctl.h>
 632 #endif
 633 # include <unistd.h>
 634 # include <time.h>
 635 # include <asm/types.h>
 636 # ifndef DEBUG_SUBSYSTEM
 637 #  define DEBUG_SUBSYSTEM S_UNDEFINED
 638 # endif
 639 # ifdef PORTAL_DEBUG
 640 #  undef NDEBUG
 641 #  include <assert.h>
 642 #  define LASSERT(e)     assert(e)
 643 #  define LASSERTF(cond, args...)     assert(cond)
 644 # else
 645 #  define LASSERT(e)
 646 #  define LASSERTF(cond, args...) do { } while (0)
 647 # endif
 648 # define printk(format, args...) printf (format, ## args)
 649 # define PORTAL_ALLOC(ptr, size) do { (ptr) = malloc(size); } while (0);
 650 # define PORTAL_FREE(a, b) do { free(a); } while (0);
 651 # define portals_debug_msg(subsys, mask, file, fn, line, stack, format, a...) \
 652     printf("%02x:%06x (@%lu %s:%s,l. %d %d %lu): " format,                    \
 653            (subsys), (mask), (long)time(0), file, fn, line,                   \
 654            getpid() , stack, ## a);
 655 #endif
 656
 657 /* support decl needed both by kernel and liblustre */
 658 char *portals_nid2str(int nal, ptl_nid_t nid, char *str);
 659
 660 #ifndef CURRENT_TIME
 661 # define CURRENT_TIME time(0)
 662 #endif
 663
 664 /******************************************************************************/
 665 /* Light-weight trace
 666  * Support for temporary event tracing with minimal Heisenberg effect. */
 667 #define LWT_SUPPORT  0
 668
 669 #define LWT_MEMORY   (64<<20)
 670 #define LWT_MAX_CPUS 4
 671
 672 typedef struct {
 673         cycles_t    lwte_when;
 674         char       *lwte_where;
 675         void       *lwte_task;
 676         long        lwte_p1;
 677         long        lwte_p2;
 678         long        lwte_p3;
 679         long        lwte_p4;
 680 #if BITS_PER_LONG > 32
 681         long        lwte_pad;
 682 #endif
 683 } lwt_event_t;
 684
 685 #if LWT_SUPPORT
 686 #ifdef __KERNEL__
 687 #define LWT_EVENTS_PER_PAGE (PAGE_SIZE / sizeof (lwt_event_t))
 688
 689 typedef struct _lwt_page {
 690         struct list_head     lwtp_list;
 691         struct page         *lwtp_page;
 692         lwt_event_t         *lwtp_events;
 693 } lwt_page_t;
 694
 695 typedef struct {
 696         int                lwtc_current_index;
 697         lwt_page_t        *lwtc_current_page;
 698 } lwt_cpu_t;
 699
 700 extern int       lwt_enabled;
 701 extern lwt_cpu_t lwt_cpus[];
 702
 703 extern int  lwt_init (void);
 704 extern void lwt_fini (void);
 705 extern int  lwt_lookup_string (int *size, char *knlptr,
 706                                char *usrptr, int usrsize);
 707 extern int  lwt_control (int enable, int clear);
 708 extern int  lwt_snapshot (cycles_t *now, int *ncpu, int *total_size,
 709                           void *user_ptr, int user_size);
 710
 711 /* Note that we _don't_ define LWT_EVENT at all if LWT_SUPPORT isn't set.
 712  * This stuff is meant for finding specific problems; it never stays in
 713  * production code... */
 714
 715 #define LWTSTR(n)       #n
 716 #define LWTWHERE(f,l)   f ":" LWTSTR(l)
 717
 718 #define LWT_EVENT(p1, p2, p3, p4)                                       \
 719 do {                                                                    \
 720         unsigned long    flags;                                         \
 721         lwt_cpu_t       *cpu;                                           \
 722         lwt_page_t      *p;                                             \
 723         lwt_event_t     *e;                                             \
 724                                                                         \
 725         local_irq_save (flags);                                         \
 726                                                                         \
 727         if (lwt_enabled) {                                              \
 728                 cpu = &lwt_cpus[smp_processor_id()];                    \
 729                 p = cpu->lwtc_current_page;                             \
 730                 e = &p->lwtp_events[cpu->lwtc_current_index++];         \
 731                                                                         \
 732                 if (cpu->lwtc_current_index >= LWT_EVENTS_PER_PAGE) {   \
 733                         cpu->lwtc_current_page =                        \
 734                                 list_entry (p->lwtp_list.next,          \
 735                                             lwt_page_t, lwtp_list);     \
 736                         cpu->lwtc_current_index = 0;                    \
 737                 }                                                       \
 738                                                                         \
 739                 e->lwte_when  = get_cycles();                           \
 740                 e->lwte_where = LWTWHERE(__FILE__,__LINE__);            \
 741                 e->lwte_task  = current;                                \
 742                 e->lwte_p1    = (long)(p1);                             \
 743                 e->lwte_p2    = (long)(p2);                             \
 744                 e->lwte_p3    = (long)(p3);                             \
 745                 e->lwte_p4    = (long)(p4);                             \
 746         }                                                               \
 747                                                                         \
 748         local_irq_restore (flags);                                      \
 749 } while (0)
 750 #else  /* __KERNEL__ */
 751 #define LWT_EVENT(p1,p2,p3,p4)     /* no userland implementation yet */
 752 #endif /* __KERNEL__ */
 753 #endif /* LWT_SUPPORT */
 754
 755 struct portals_device_userstate
 756 {
 757         int          pdu_memhog_pages;
 758         struct page *pdu_memhog_root_page;
 759 };
 760
 761 #include <linux/portals_lib.h>
 762
 763 /*
 764  * USER LEVEL STUFF BELOW
 765  */
 766
 767 #define PORTALS_CFG_VERSION 0x00010001;
 768
 769 struct portals_cfg {
 770         __u32 pcfg_version;
 771         __u32 pcfg_command;
 772
 773         __u32 pcfg_nal;
 774         __u32 pcfg_flags;
 775
 776         __u32 pcfg_gw_nal;
 777         __u64 pcfg_nid;
 778         __u64 pcfg_nid2;
 779         __u64 pcfg_nid3;
 780         __u32 pcfg_id;
 781         __u32 pcfg_misc;
 782         __u32 pcfg_fd;
 783         __u32 pcfg_count;
 784         __u32 pcfg_size;
 785         __u32 pcfg_wait;
 786
 787         __u32 pcfg_plen1; /* buffers in userspace */
 788         char *pcfg_pbuf1;
 789         __u32 pcfg_plen2; /* buffers in userspace */
 790         char *pcfg_pbuf2;
 791 };
 792
 793 #define PCFG_INIT(pcfg, cmd)                            \
 794 do {                                                    \
 795         memset(&pcfg, 0, sizeof(pcfg));                 \
 796         pcfg.pcfg_version = PORTALS_CFG_VERSION;        \
 797         pcfg.pcfg_command = (cmd);                      \
 798                                                         \
 799 } while (0)
 800
 801 #define PORTAL_IOCTL_VERSION 0x00010007
 802 #define PING_SYNC       0
 803 #define PING_ASYNC      1
 804
 805 struct portal_ioctl_data {
 806         __u32 ioc_len;
 807         __u32 ioc_version;
 808         __u64 ioc_nid;
 809         __u64 ioc_nid2;
 810         __u64 ioc_nid3;
 811         __u32 ioc_count;
 812         __u32 ioc_nal;
 813         __u32 ioc_nal_cmd;
 814         __u32 ioc_fd;
 815         __u32 ioc_id;
 816
 817         __u32 ioc_flags;
 818         __u32 ioc_size;
 819
 820         __u32 ioc_wait;
 821         __u32 ioc_timeout;
 822         __u32 ioc_misc;
 823
 824         __u32 ioc_inllen1;
 825         char *ioc_inlbuf1;
 826         __u32 ioc_inllen2;
 827         char *ioc_inlbuf2;
 828
 829         __u32 ioc_plen1; /* buffers in userspace */
 830         char *ioc_pbuf1;
 831         __u32 ioc_plen2; /* buffers in userspace */
 832         char *ioc_pbuf2;
 833
 834         char ioc_bulk[0];
 835 };
 836
 837 struct portal_ioctl_hdr {
 838         __u32 ioc_len;
 839         __u32 ioc_version;
 840 };
 841
 842 struct portals_debug_ioctl_data
 843 {
 844         struct portal_ioctl_hdr hdr;
 845         unsigned int subs;
 846         unsigned int debug;
 847 };
 848
 849 #define PORTAL_IOC_INIT(data)                           \
 850 do {                                                    \
 851         memset(&data, 0, sizeof(data));                 \
 852         data.ioc_version = PORTAL_IOCTL_VERSION;        \
 853         data.ioc_len = sizeof(data);                    \
 854 } while (0)
 855
 856 /* FIXME check conflict with lustre_lib.h */
 857 #define PTL_IOC_DEBUG_MASK             _IOWR('f', 250, long)
 858
 859 static inline int portal_ioctl_packlen(struct portal_ioctl_data *data)
 860 {
 861         int len = sizeof(*data);
 862         len += size_round(data->ioc_inllen1);
 863         len += size_round(data->ioc_inllen2);
 864         return len;
 865 }
 866
 867 static inline int portal_ioctl_is_invalid(struct portal_ioctl_data *data)
 868 {
 869         if (data->ioc_len > (1<<30)) {
 870                 CERROR ("PORTALS ioctl: ioc_len larger than 1<<30\n");
 871                 return 1;
 872         }
 873         if (data->ioc_inllen1 > (1<<30)) {
 874                 CERROR ("PORTALS ioctl: ioc_inllen1 larger than 1<<30\n");
 875                 return 1;
 876         }
 877         if (data->ioc_inllen2 > (1<<30)) {
 878                 CERROR ("PORTALS ioctl: ioc_inllen2 larger than 1<<30\n");
 879                 return 1;
 880         }
 881         if (data->ioc_inlbuf1 && !data->ioc_inllen1) {
 882                 CERROR ("PORTALS ioctl: inlbuf1 pointer but 0 length\n");
 883                 return 1;
 884         }
 885         if (data->ioc_inlbuf2 && !data->ioc_inllen2) {
 886                 CERROR ("PORTALS ioctl: inlbuf2 pointer but 0 length\n");
 887                 return 1;
 888         }
 889         if (data->ioc_pbuf1 && !data->ioc_plen1) {
 890                 CERROR ("PORTALS ioctl: pbuf1 pointer but 0 length\n");
 891                 return 1;
 892         }
 893         if (data->ioc_pbuf2 && !data->ioc_plen2) {
 894                 CERROR ("PORTALS ioctl: pbuf2 pointer but 0 length\n");
 895                 return 1;
 896         }
 897         if (data->ioc_plen1 && !data->ioc_pbuf1) {
 898                 CERROR ("PORTALS ioctl: plen1 nonzero but no pbuf1 pointer\n");
 899                 return 1;
 900         }
 901         if (data->ioc_plen2 && !data->ioc_pbuf2) {
 902                 CERROR ("PORTALS ioctl: plen2 nonzero but no pbuf2 pointer\n");
 903                 return 1;
 904         }
 905         if (portal_ioctl_packlen(data) != data->ioc_len ) {
 906                 CERROR ("PORTALS ioctl: packlen != ioc_len\n");
 907                 return 1;
 908         }
 909         if (data->ioc_inllen1 &&
 910             data->ioc_bulk[data->ioc_inllen1 - 1] != '\0') {
 911                 CERROR ("PORTALS ioctl: inlbuf1 not 0 terminated\n");
 912                 return 1;
 913         }
 914         if (data->ioc_inllen2 &&
 915             data->ioc_bulk[size_round(data->ioc_inllen1) +
 916                            data->ioc_inllen2 - 1] != '\0') {
 917                 CERROR ("PORTALS ioctl: inlbuf2 not 0 terminated\n");
 918                 return 1;
 919         }
 920         return 0;
 921 }
 922
 923 #ifndef __KERNEL__
 924 static inline int portal_ioctl_pack(struct portal_ioctl_data *data, char **pbuf,
 925                                     int max)
 926 {
 927         char *ptr;
 928         struct portal_ioctl_data *overlay;
 929         data->ioc_len = portal_ioctl_packlen(data);
 930         data->ioc_version = PORTAL_IOCTL_VERSION;
 931
 932         if (*pbuf && portal_ioctl_packlen(data) > max)
 933                 return 1;
 934         if (*pbuf == NULL) {
 935                 *pbuf = malloc(data->ioc_len);
 936         }
 937         if (!*pbuf)
 938                 return 1;
 939         overlay = (struct portal_ioctl_data *)*pbuf;
 940         memcpy(*pbuf, data, sizeof(*data));
 941
 942         ptr = overlay->ioc_bulk;
 943         if (data->ioc_inlbuf1)
 944                 LOGL(data->ioc_inlbuf1, data->ioc_inllen1, ptr);
 945         if (data->ioc_inlbuf2)
 946                 LOGL(data->ioc_inlbuf2, data->ioc_inllen2, ptr);
 947         if (portal_ioctl_is_invalid(overlay))
 948                 return 1;
 949
 950         return 0;
 951 }
 952 #else
 953 #include <asm/uaccess.h>
 954
 955 /* buffer MUST be at least the size of portal_ioctl_hdr */
 956 static inline int portal_ioctl_getdata(char *buf, char *end, void *arg)
 957 {
 958         struct portal_ioctl_hdr *hdr;
 959         struct portal_ioctl_data *data;
 960         int err;
 961         ENTRY;
 962
 963         hdr = (struct portal_ioctl_hdr *)buf;
 964         data = (struct portal_ioctl_data *)buf;
 965
 966         err = copy_from_user(buf, (void *)arg, sizeof(*hdr));
 967         if ( err ) {
 968                 EXIT;
 969                 return err;
 970         }
 971
 972         if (hdr->ioc_version != PORTAL_IOCTL_VERSION) {
 973                 CERROR ("PORTALS: version mismatch kernel vs application\n");
 974                 return -EINVAL;
 975         }
 976
 977         if (hdr->ioc_len + buf >= end) {
 978                 CERROR ("PORTALS: user buffer exceeds kernel buffer\n");
 979                 return -EINVAL;
 980         }
 981
 982
 983         if (hdr->ioc_len < sizeof(struct portal_ioctl_data)) {
 984                 CERROR ("PORTALS: user buffer too small for ioctl\n");
 985                 return -EINVAL;
 986         }
 987
 988         err = copy_from_user(buf, (void *)arg, hdr->ioc_len);
 989         if ( err ) {
 990                 EXIT;
 991                 return err;
 992         }
 993
 994         if (portal_ioctl_is_invalid(data)) {
 995                 CERROR ("PORTALS: ioctl not correctly formatted\n");
 996                 return -EINVAL;
 997         }
 998
 999         if (data->ioc_inllen1) {
1000                 data->ioc_inlbuf1 = &data->ioc_bulk[0];
1001         }
1002
1003         if (data->ioc_inllen2) {
1004                 data->ioc_inlbuf2 = &data->ioc_bulk[0] +
1005                         size_round(data->ioc_inllen1);
1006         }
1007
1008         EXIT;
1009         return 0;
1010 }
1011 #endif
1012
1013 /* ioctls for manipulating snapshots 30- */
1014 #define IOC_PORTAL_TYPE                   'e'
1015 #define IOC_PORTAL_MIN_NR                 30
1016
1017 #define IOC_PORTAL_PING                    _IOWR('e', 30, long)
1018 #define IOC_PORTAL_GET_DEBUG               _IOWR('e', 31, long)
1019 #define IOC_PORTAL_CLEAR_DEBUG             _IOWR('e', 32, long)
1020 #define IOC_PORTAL_MARK_DEBUG              _IOWR('e', 33, long)
1021 #define IOC_PORTAL_PANIC                   _IOWR('e', 34, long)
1022 #define IOC_PORTAL_NAL_CMD                 _IOWR('e', 35, long)
1023 #define IOC_PORTAL_GET_NID                 _IOWR('e', 36, long)
1024 #define IOC_PORTAL_FAIL_NID                _IOWR('e', 37, long)
1025 #define IOC_PORTAL_SET_DAEMON              _IOWR('e', 38, long)
1026 #define IOC_PORTAL_LWT_CONTROL             _IOWR('e', 39, long)
1027 #define IOC_PORTAL_LWT_SNAPSHOT            _IOWR('e', 40, long)
1028 #define IOC_PORTAL_LWT_LOOKUP_STRING       _IOWR('e', 41, long)
1029 #define IOC_PORTAL_MEMHOG                  _IOWR('e', 42, long)
1030 #define IOC_PORTAL_MAX_NR                             42
1031
1032 enum {
1033         QSWNAL    = 1,
1034         SOCKNAL   = 2,
1035         GMNAL     = 3,
1036         /*          4 unused */
1037         TCPNAL    = 5,
1038         SCIMACNAL = 6,
1039         ROUTER    = 7,
1040         IBNAL     = 8,
1041         NAL_ENUM_END_MARKER
1042 };
1043
1044 #ifdef __KERNEL__
1045 extern ptl_handle_ni_t  kqswnal_ni;
1046 extern ptl_handle_ni_t  ksocknal_ni;
1047 extern ptl_handle_ni_t  kgmnal_ni;
1048 extern ptl_handle_ni_t  kibnal_ni;
1049 extern ptl_handle_ni_t  kscimacnal_ni;
1050 #endif
1051
1052 #define PTL_NALFMT_SIZE         16
1053
1054 #define NAL_MAX_NR (NAL_ENUM_END_MARKER - 1)
1055
1056 #define NAL_CMD_REGISTER_PEER_FD     100
1057 #define NAL_CMD_CLOSE_CONNECTION     101
1058 #define NAL_CMD_REGISTER_MYNID       102
1059 #define NAL_CMD_PUSH_CONNECTION      103
1060 #define NAL_CMD_GET_CONN             104
1061 #define NAL_CMD_DEL_AUTOCONN         105
1062 #define NAL_CMD_ADD_AUTOCONN         106
1063 #define NAL_CMD_GET_AUTOCONN         107
1064 #define NAL_CMD_GET_TXDESC           108
1065 #define NAL_CMD_ADD_ROUTE            109
1066 #define NAL_CMD_DEL_ROUTE            110
1067 #define NAL_CMD_GET_ROUTE            111
1068 #define NAL_CMD_NOTIFY_ROUTER        112
1069
1070 enum {
1071         DEBUG_DAEMON_START       =  1,
1072         DEBUG_DAEMON_STOP        =  2,
1073         DEBUG_DAEMON_PAUSE       =  3,
1074         DEBUG_DAEMON_CONTINUE    =  4,
1075 };
1076
1077 /* XXX remove to lustre ASAP */
1078 struct lustre_peer {
1079         ptl_nid_t       peer_nid;
1080         ptl_handle_ni_t peer_ni;
1081 };
1082
1083
1084 /* module.c */
1085 typedef int (*nal_cmd_handler_t)(struct portals_cfg *, void * private);
1086 int kportal_nal_register(int nal, nal_cmd_handler_t handler, void * private);
1087 int kportal_nal_unregister(int nal);
1088
1089 enum cfg_record_type {
1090         PORTALS_CFG_TYPE = 1,
1091         LUSTRE_CFG_TYPE = 123,
1092 };
1093
1094 typedef int (*cfg_record_cb_t)(enum cfg_record_type, int len, void *data);
1095 int kportal_nal_cmd(struct portals_cfg *);
1096
1097 ptl_handle_ni_t *kportal_get_ni (int nal);
1098 void kportal_put_ni (int nal);
1099
1100 #ifdef __CYGWIN__
1101 # ifndef BITS_PER_LONG
1102 #  if (~0UL) == 0xffffffffUL
1103 #   define BITS_PER_LONG 32
1104 #  else
1105 #   define BITS_PER_LONG 64
1106 #  endif
1107 # endif
1108 #endif
1109
1110 #if defined(__x86_64__)
1111 # define LPU64 "%Lu"
1112 # define LPD64 "%Ld"
1113 # define LPX64 "%#Lx"
1114 # define LPSZ  "%lu"
1115 # define LPSSZ "%ld"
1116 #elif (BITS_PER_LONG == 32 || __WORDSIZE == 32)
1117 # define LPU64 "%Lu"
1118 # define LPD64 "%Ld"
1119 # define LPX64 "%#Lx"
1120 # define LPSZ  "%u"
1121 # define LPSSZ "%d"
1122 #elif (BITS_PER_LONG == 64 || __WORDSIZE == 64)
1123 # define LPU64 "%lu"
1124 # define LPD64 "%ld"
1125 # define LPX64 "%#lx"
1126 # define LPSZ  "%lu"
1127 # define LPSSZ "%ld"
1128 #endif
1129 #ifndef LPU64
1130 # error "No word size defined"
1131 #endif
1132
1133 #endif