Whamcloud - gitweb
land v0.9.1 on HEAD, in preparation for a 1.0.x branch
[fs/lustre-release.git] / 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
75 #ifdef __KERNEL__
76 # include <linux/sched.h> /* THREAD_SIZE */
77 #else 
78 # ifndef THREAD_SIZE /* x86_64 has THREAD_SIZE in userspace */
79 #  define THREAD_SIZE 8192
80 # endif
81 #endif
82
83 #define LUSTRE_TRACE_SIZE (THREAD_SIZE >> 5)
84
85 #ifdef __KERNEL__
86 # ifdef  __ia64__
87 #  define CDEBUG_STACK (THREAD_SIZE -                                      \
88                         ((unsigned long)__builtin_dwarf_cfa() &            \
89                          (THREAD_SIZE - 1)))
90 # else
91 #  define CDEBUG_STACK (THREAD_SIZE -                                      \
92                         ((unsigned long)__builtin_frame_address(0) &       \
93                          (THREAD_SIZE - 1)))
94 # endif
95
96 #define CHECK_STACK(stack)                                                    \
97         do {                                                                  \
98                 if ((stack) > 3*THREAD_SIZE/4 && (stack) > portal_stack) {    \
99                         portals_debug_msg(DEBUG_SUBSYSTEM, D_ERROR,           \
100                                           __FILE__, __FUNCTION__, __LINE__,   \
101                                           (stack),                            \
102                                           "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) || ((mask) & (D_ERROR | D_EMERG)) ||                      \
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 invalidate_inode_pages
188 #define PageUptodate Page_Uptodate
189 #define our_recalc_sigpending(current) recalc_sigpending(current)
190 #define num_online_cpus() smp_num_cpus
191 static inline void our_cond_resched(void)
192 {
193         if (current->need_resched)
194                schedule ();
195 }
196 #define work_struct_t       struct tq_struct 
197
198 #else
199
200 #define prepare_work(wq,cb,cbdata)                                            \
201 do {                                                                          \
202         INIT_WORK((wq), (void *)(cb), (void *)(cbdata));                      \
203 } while (0)
204 #define ll_invalidate_inode_pages(inode) invalidate_inode_pages((inode)->i_mapping)
205 #define wait_on_page wait_on_page_locked
206 #define our_recalc_sigpending(current) recalc_sigpending()
207 #define strtok(a,b) strpbrk(a, b)
208 static inline void our_cond_resched(void)
209 {
210         cond_resched();
211 }
212 #define work_struct_t      struct work_struct
213
214 #endif /* LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0) */
215
216 #ifdef PORTAL_DEBUG
217 extern void kportal_assertion_failed(char *expr, char *file, const char *func,
218                                      const int line);
219 #define LASSERT(e) ((e) ? 0 : kportal_assertion_failed( #e , __FILE__,  \
220                                                         __FUNCTION__, __LINE__))
221 /* it would be great to dump_stack() here, but some kernels
222  * export it as show_stack() and I can't be bothered to
223  * proprely engage in that dance right now */ 
224 #define LASSERTF(cond, fmt...)                                                \
225         do {                                                                  \
226                 if (unlikely(!(cond))) {                                      \
227                         portals_debug_msg(0, D_EMERG,  __FILE__, __FUNCTION__,\
228                                           __LINE__,  CDEBUG_STACK,            \
229                                           "ASSERTION(" #cond ") failed:" fmt);\
230                         LBUG();                                               \
231                 }                                                             \
232         } while (0)
233                                 
234 #else
235 #define LASSERT(e)
236 #define LASSERTF(cond, fmt...) do { } while (0)
237 #endif
238
239 #ifdef __arch_um__
240 #define LBUG_WITH_LOC(file, func, line)                                 \
241 do {                                                                    \
242         CEMERG("LBUG - trying to dump log to /tmp/lustre-log\n");       \
243         portals_debug_dumplog();                                        \
244         portals_run_lbug_upcall(file, func, line);                      \
245         panic("LBUG");                                                  \
246 } while (0)
247 #else
248 #define LBUG_WITH_LOC(file, func, line)                                 \
249 do {                                                                    \
250         CEMERG("LBUG\n");                                               \
251         portals_debug_dumplog();                                        \
252         portals_run_lbug_upcall(file, func, line);                      \
253         set_task_state(current, TASK_UNINTERRUPTIBLE);                  \
254         schedule();                                                     \
255 } while (0)
256 #endif /* __arch_um__ */
257
258 #define LBUG() LBUG_WITH_LOC(__FILE__, __FUNCTION__, __LINE__)
259
260 /*
261  * Memory
262  */
263 #ifdef PORTAL_DEBUG
264 extern atomic_t portal_kmemory;
265
266 # define portal_kmem_inc(ptr, size)                                           \
267 do {                                                                          \
268         atomic_add(size, &portal_kmemory);                                    \
269 } while (0)
270
271 # define portal_kmem_dec(ptr, size) do {                                      \
272         atomic_sub(size, &portal_kmemory);                                    \
273 } while (0)
274
275 #else
276 # define portal_kmem_inc(ptr, size) do {} while (0)
277 # define portal_kmem_dec(ptr, size) do {} while (0)
278 #endif /* PORTAL_DEBUG */
279
280 #define PORTAL_VMALLOC_SIZE        16384
281
282 #ifndef GFP_MEMALLOC
283 #define GFP_MEMALLOC 0
284 #endif
285
286 #define PORTAL_ALLOC(ptr, size)                                           \
287 do {                                                                      \
288         LASSERT (!in_interrupt());                                        \
289         if ((size) > PORTAL_VMALLOC_SIZE)                                 \
290                 (ptr) = vmalloc(size);                                    \
291         else                                                              \
292                 (ptr) = kmalloc((size), (GFP_KERNEL | GFP_MEMALLOC));     \
293         if ((ptr) == NULL)                                                \
294                 CERROR("PORTALS: out of memory at %s:%d (tried to alloc '"\
295                        #ptr "' = %d)\n", __FILE__, __LINE__, (int)(size));\
296         else {                                                            \
297                 portal_kmem_inc((ptr), (size));                           \
298                 memset((ptr), 0, (size));                                 \
299         }                                                                 \
300         CDEBUG(D_MALLOC, "kmalloced '" #ptr "': %d at %p (tot %d).\n",    \
301                (int)(size), (ptr), atomic_read (&portal_kmemory));        \
302 } while (0)
303
304 #define PORTAL_FREE(ptr, size)                                          \
305 do {                                                                    \
306         int s = (size);                                                 \
307         if ((ptr) == NULL) {                                            \
308                 CERROR("PORTALS: free NULL '" #ptr "' (%d bytes) at "   \
309                        "%s:%d\n", s, __FILE__, __LINE__);               \
310                 break;                                                  \
311         }                                                               \
312         if (s > PORTAL_VMALLOC_SIZE)                                    \
313                 vfree(ptr);                                             \
314         else                                                            \
315                 kfree(ptr);                                             \
316         portal_kmem_dec((ptr), s);                                      \
317         CDEBUG(D_MALLOC, "kfreed '" #ptr "': %d at %p (tot %d).\n",     \
318                s, (ptr), atomic_read(&portal_kmemory));                 \
319 } while (0)
320
321 #ifndef SLAB_MEMALLOC
322 #define SLAB_MEMALLOC 0
323 #endif
324
325 #define PORTAL_SLAB_ALLOC(ptr, slab, size)                                \
326 do {                                                                      \
327         LASSERT(!in_interrupt());                                         \
328         (ptr) = kmem_cache_alloc((slab), (SLAB_KERNEL | SLAB_MEMALLOC));  \
329         if ((ptr) == NULL) {                                              \
330                 CERROR("PORTALS: out of memory at %s:%d (tried to alloc"  \
331                        " '" #ptr "' from slab '" #slab "')\n", __FILE__,  \
332                        __LINE__);                                         \
333         } else {                                                          \
334                 portal_kmem_inc((ptr), (size));                           \
335                 memset((ptr), 0, (size));                                 \
336         }                                                                 \
337         CDEBUG(D_MALLOC, "kmalloced '" #ptr "': %ld at %p (tot %d).\n",   \
338                (int)(size), (ptr), atomic_read(&portal_kmemory));         \
339 } while (0)
340
341 #define PORTAL_SLAB_FREE(ptr, slab, size)                               \
342 do {                                                                    \
343         int s = (size);                                                 \
344         if ((ptr) == NULL) {                                            \
345                 CERROR("PORTALS: free NULL '" #ptr "' (%d bytes) at "   \
346                        "%s:%d\n", s, __FILE__, __LINE__);               \
347                 break;                                                  \
348         }                                                               \
349         memset((ptr), 0x5a, s);                                         \
350         kmem_cache_free((slab), ptr);                                   \
351         portal_kmem_dec((ptr), s);                                      \
352         CDEBUG(D_MALLOC, "kfreed '" #ptr "': %d at %p (tot %d).\n",     \
353                s, (ptr), atomic_read (&portal_kmemory));                \
354 } while (0)
355
356 /* ------------------------------------------------------------------- */
357
358 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
359
360 #define PORTAL_SYMBOL_REGISTER(x) inter_module_register(#x, THIS_MODULE, &x)
361 #define PORTAL_SYMBOL_UNREGISTER(x) inter_module_unregister(#x)
362
363 #define PORTAL_SYMBOL_GET(x) ((typeof(&x))inter_module_get(#x))
364 #define PORTAL_SYMBOL_PUT(x) inter_module_put(#x)
365
366 #define PORTAL_MODULE_USE       MOD_INC_USE_COUNT
367 #define PORTAL_MODULE_UNUSE     MOD_DEC_USE_COUNT
368 #else
369
370 #define PORTAL_SYMBOL_REGISTER(x)
371 #define PORTAL_SYMBOL_UNREGISTER(x)
372
373 #define PORTAL_SYMBOL_GET(x) symbol_get(x)
374 #define PORTAL_SYMBOL_PUT(x) symbol_put(x)
375
376 #define PORTAL_MODULE_USE       try_module_get(THIS_MODULE)
377 #define PORTAL_MODULE_UNUSE     module_put(THIS_MODULE)
378
379 #endif
380
381 /******************************************************************************/
382 /* Kernel Portals Router interface */
383
384 typedef void (*kpr_fwd_callback_t)(void *arg, int error); // completion callback
385
386 /* space for routing targets to stash "stuff" in a forwarded packet */
387 typedef union {
388         long long        _alignment;
389         void            *_space[16];            /* scale with CPU arch */
390 } kprfd_scratch_t;
391
392 /* Kernel Portals Routing Forwarded message Descriptor */
393 typedef struct {
394         struct list_head     kprfd_list;        /* stash in queues (routing target can use) */
395         ptl_nid_t            kprfd_target_nid;  /* final destination NID */
396         ptl_nid_t            kprfd_gateway_nid; /* gateway NID */
397         int                  kprfd_nob;         /* # message bytes (including header) */
398         int                  kprfd_niov;        /* # message frags (including header) */
399         struct iovec        *kprfd_iov;         /* message fragments */
400         void                *kprfd_router_arg;  // originating NAL's router arg
401         kpr_fwd_callback_t   kprfd_callback;    /* completion callback */
402         void                *kprfd_callback_arg; /* completion callback arg */
403         kprfd_scratch_t      kprfd_scratch;    // scratchpad for routing targets
404 } kpr_fwd_desc_t;
405
406 typedef void  (*kpr_fwd_t)(void *arg, kpr_fwd_desc_t *fwd);
407 typedef void  (*kpr_notify_t)(void *arg, ptl_nid_t peer, int alive);
408
409 /* NAL's routing interface (Kernel Portals Routing Nal Interface) */
410 typedef const struct {
411         int             kprni_nalid;    /* NAL's id */
412         void           *kprni_arg;      /* Arg to pass when calling into NAL */
413         kpr_fwd_t       kprni_fwd;      /* NAL's forwarding entrypoint */
414         kpr_notify_t    kprni_notify;   /* NAL's notification entrypoint */
415 } kpr_nal_interface_t;
416
417 /* Router's routing interface (Kernel Portals Routing Router Interface) */
418 typedef const struct {
419         /* register the calling NAL with the router and get back the handle for
420          * subsequent calls */
421         int     (*kprri_register) (kpr_nal_interface_t *nal_interface,
422                                    void **router_arg);
423
424         /* ask the router to find a gateway that forwards to 'nid' and is a
425          * peer of the calling NAL; assume caller will send 'nob' bytes of
426          * payload there */
427         int     (*kprri_lookup) (void *router_arg, ptl_nid_t nid, int nob,
428                                  ptl_nid_t *gateway_nid);
429
430         /* hand a packet over to the router for forwarding */
431         kpr_fwd_t kprri_fwd_start;
432
433         /* hand a packet back to the router for completion */
434         void    (*kprri_fwd_done) (void *router_arg, kpr_fwd_desc_t *fwd,
435                                    int error);
436
437         /* notify the router about peer state */
438         void    (*kprri_notify) (void *router_arg, ptl_nid_t peer,
439                                  int alive, time_t when);
440
441         /* the calling NAL is shutting down */
442         void    (*kprri_shutdown) (void *router_arg);
443
444         /* deregister the calling NAL with the router */
445         void    (*kprri_deregister) (void *router_arg);
446
447 } kpr_router_interface_t;
448
449 /* Convenient struct for NAL to stash router interface/args */
450 typedef struct {
451         kpr_router_interface_t  *kpr_interface;
452         void                    *kpr_arg;
453 } kpr_router_t;
454
455 /* Router's control interface (Kernel Portals Routing Control Interface) */
456 typedef const struct {
457         int     (*kprci_add_route)(int gateway_nal, ptl_nid_t gateway_nid,
458                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
459         int     (*kprci_del_route)(int gateway_nal, ptl_nid_t gateway_nid,
460                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
461         int     (*kprci_get_route)(int index, int *gateway_nal,
462                                    ptl_nid_t *gateway,
463                                    ptl_nid_t *lo_nid, ptl_nid_t *hi_nid,
464                                    int *alive);
465         int     (*kprci_notify)(int gateway_nal, ptl_nid_t gateway_nid, 
466                                 int alive, time_t when);
467 } kpr_control_interface_t;
468
469 extern kpr_control_interface_t  kpr_control_interface;
470 extern kpr_router_interface_t   kpr_router_interface;
471
472 static inline int
473 kpr_register (kpr_router_t *router, kpr_nal_interface_t *nalif)
474 {
475         int    rc;
476
477         router->kpr_interface = PORTAL_SYMBOL_GET (kpr_router_interface);
478         if (router->kpr_interface == NULL)
479                 return (-ENOENT);
480
481         rc = (router->kpr_interface)->kprri_register (nalif, &router->kpr_arg);
482         if (rc != 0)
483                 router->kpr_interface = NULL;
484
485         PORTAL_SYMBOL_PUT (kpr_router_interface);
486         return (rc);
487 }
488
489 static inline int
490 kpr_routing (kpr_router_t *router)
491 {
492         return (router->kpr_interface != NULL);
493 }
494
495 static inline int
496 kpr_lookup (kpr_router_t *router, ptl_nid_t nid, int nob, ptl_nid_t *gateway_nid)
497 {
498         if (!kpr_routing (router))
499                 return (-ENETUNREACH);
500
501         return (router->kpr_interface->kprri_lookup(router->kpr_arg, nid, nob,
502                                                     gateway_nid));
503 }
504
505 static inline void
506 kpr_fwd_init (kpr_fwd_desc_t *fwd, ptl_nid_t nid,
507               int nob, int niov, struct iovec *iov,
508               kpr_fwd_callback_t callback, void *callback_arg)
509 {
510         fwd->kprfd_target_nid   = nid;
511         fwd->kprfd_gateway_nid  = nid;
512         fwd->kprfd_nob          = nob;
513         fwd->kprfd_niov         = niov;
514         fwd->kprfd_iov          = iov;
515         fwd->kprfd_callback     = callback;
516         fwd->kprfd_callback_arg = callback_arg;
517 }
518
519 static inline void
520 kpr_fwd_start (kpr_router_t *router, kpr_fwd_desc_t *fwd)
521 {
522         if (!kpr_routing (router))
523                 fwd->kprfd_callback (fwd->kprfd_callback_arg, -ENETUNREACH);
524         else
525                 router->kpr_interface->kprri_fwd_start (router->kpr_arg, fwd);
526 }
527
528 static inline void
529 kpr_fwd_done (kpr_router_t *router, kpr_fwd_desc_t *fwd, int error)
530 {
531         LASSERT (kpr_routing (router));
532         router->kpr_interface->kprri_fwd_done (router->kpr_arg, fwd, error);
533 }
534
535 static inline void
536 kpr_notify (kpr_router_t *router, 
537             ptl_nid_t peer, int alive, time_t when)
538 {
539         if (!kpr_routing (router))
540                 return;
541         
542         router->kpr_interface->kprri_notify(router->kpr_arg, peer, alive, when);
543 }
544
545 static inline void
546 kpr_shutdown (kpr_router_t *router)
547 {
548         if (kpr_routing (router))
549                 router->kpr_interface->kprri_shutdown (router->kpr_arg);
550 }
551
552 static inline void
553 kpr_deregister (kpr_router_t *router)
554 {
555         if (!kpr_routing (router))
556                 return;
557         router->kpr_interface->kprri_deregister (router->kpr_arg);
558         router->kpr_interface = NULL;
559 }
560
561 /******************************************************************************/
562
563 #ifdef PORTALS_PROFILING
564 #define prof_enum(FOO) PROF__##FOO
565 enum {
566         prof_enum(our_recvmsg),
567         prof_enum(our_sendmsg),
568         prof_enum(socknal_recv),
569         prof_enum(lib_parse),
570         prof_enum(conn_list_walk),
571         prof_enum(memcpy),
572         prof_enum(lib_finalize),
573         prof_enum(pingcli_time),
574         prof_enum(gmnal_send),
575         prof_enum(gmnal_recv),
576         MAX_PROFS
577 };
578
579 struct prof_ent {
580         char *str;
581         /* hrmph.  wrap-tastic. */
582         u32       starts;
583         u32       finishes;
584         cycles_t  total_cycles;
585         cycles_t  start;
586         cycles_t  end;
587 };
588
589 extern struct prof_ent prof_ents[MAX_PROFS];
590
591 #define PROF_START(FOO)                                         \
592         do {                                                    \
593                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
594                 pe->starts++;                                   \
595                 pe->start = get_cycles();                       \
596         } while (0)
597
598 #define PROF_FINISH(FOO)                                        \
599         do {                                                    \
600                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
601                 pe->finishes++;                                 \
602                 pe->end = get_cycles();                         \
603                 pe->total_cycles += (pe->end - pe->start);      \
604         } while (0)
605 #else /* !PORTALS_PROFILING */
606 #define PROF_START(FOO) do {} while(0)
607 #define PROF_FINISH(FOO) do {} while(0)
608 #endif /* PORTALS_PROFILING */
609
610 /* debug.c */
611 extern spinlock_t stack_backtrace_lock;
612
613 char *portals_debug_dumpstack(void);
614 void portals_run_upcall(char **argv);
615 void portals_run_lbug_upcall(char * file, const char *fn, const int line);
616 void portals_debug_dumplog(void);
617 int portals_debug_init(unsigned long bufsize);
618 int portals_debug_cleanup(void);
619 int portals_debug_clear_buffer(void);
620 int portals_debug_mark_buffer(char *text);
621 int portals_debug_set_daemon(unsigned int cmd, unsigned int length,
622                              char *file, unsigned int size);
623 __s32 portals_debug_copy_to_user(char *buf, unsigned long len);
624 #if (__GNUC__)
625 /* Use the special GNU C __attribute__ hack to have the compiler check the
626  * printf style argument string against the actual argument count and
627  * types.
628  */
629 #ifdef printf
630 # warning printf has been defined as a macro...
631 # undef printf
632 #endif
633 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
634                        const int line, unsigned long stack,
635                        char *format, ...)
636         __attribute__ ((format (printf, 7, 8)));
637 #else
638 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
639                        const int line, unsigned long stack,
640                        const char *format, ...);
641 #endif /* __GNUC__ */
642 void portals_debug_set_level(unsigned int debug_level);
643
644 # define fprintf(a, format, b...) CDEBUG(D_OTHER, format , ## b)
645 # define printf(format, b...) CDEBUG(D_OTHER, format , ## b)
646 # define time(a) CURRENT_TIME
647
648 extern void kportal_daemonize (char *name);
649 extern void kportal_blockallsigs (void);
650
651 #else  /* !__KERNEL__ */
652 # include <stdio.h>
653 # include <stdlib.h>
654 #ifndef __CYGWIN__
655 # include <stdint.h>
656 #else
657 # include <cygwin-ioctl.h>
658 #endif
659 # include <unistd.h>
660 # include <time.h>
661 # include <asm/types.h>
662 # ifndef DEBUG_SUBSYSTEM
663 #  define DEBUG_SUBSYSTEM S_UNDEFINED
664 # endif
665 # ifdef PORTAL_DEBUG
666 #  undef NDEBUG
667 #  include <assert.h>
668 #  define LASSERT(e)     assert(e)
669 #  define LASSERTF(cond, args...)     assert(cond)
670 # else
671 #  define LASSERT(e)
672 #  define LASSERTF(cond, args...) do { } while (0)
673 # endif
674 # define printk(format, args...) printf (format, ## args)
675 # define PORTAL_ALLOC(ptr, size) do { (ptr) = malloc(size); } while (0);
676 # define PORTAL_FREE(a, b) do { free(a); } while (0);
677 # define portals_debug_msg(subsys, mask, file, fn, line, stack, format, a...) \
678     printf("%02x:%06x (@%lu %s:%s,l. %d %d %lu): " format,                    \
679            (subsys), (mask), (long)time(0), file, fn, line,                   \
680            getpid() , stack, ## a);
681 #endif
682
683 /* support decl needed both by kernel and liblustre */
684 char *portals_nid2str(int nal, ptl_nid_t nid, char *str);
685
686 #ifndef CURRENT_TIME
687 # define CURRENT_TIME time(0)
688 #endif
689
690 /******************************************************************************/
691 /* Light-weight trace 
692  * Support for temporary event tracing with minimal Heisenberg effect. */
693 #define LWT_SUPPORT  1
694
695 typedef struct {
696         cycles_t    lwte_when;
697         char       *lwte_where;
698         void       *lwte_task;
699         long        lwte_p1;
700         long        lwte_p2;
701         long        lwte_p3;
702         long        lwte_p4;
703 #if BITS_PER_LONG > 32
704         long        lwte_pad;
705 #endif
706 } lwt_event_t;
707
708 #if LWT_SUPPORT
709 #ifdef __KERNEL__
710 #define LWT_EVENTS_PER_PAGE (PAGE_SIZE / sizeof (lwt_event_t))
711
712 typedef struct _lwt_page {
713         struct list_head     lwtp_list;
714         struct page         *lwtp_page;
715         lwt_event_t         *lwtp_events;
716 } lwt_page_t;
717
718 typedef struct {
719         int                lwtc_current_index;
720         lwt_page_t        *lwtc_current_page;
721 } lwt_cpu_t;
722
723 extern int       lwt_enabled;
724 extern lwt_cpu_t lwt_cpus[];
725
726 extern int  lwt_init (void);
727 extern void lwt_fini (void);
728 extern int  lwt_lookup_string (int *size, char *knlptr,
729                                char *usrptr, int usrsize);
730 extern int  lwt_control (int enable, int clear);
731 extern int  lwt_snapshot (int *ncpu, int *total_size,
732                           void *user_ptr, int user_size);
733
734 /* Note that we _don't_ define LWT_EVENT at all if LWT_SUPPORT isn't set.
735  * This stuff is meant for finding specific problems; it never stays in
736  * production code... */
737
738 #define LWTSTR(n)       #n
739 #define LWTWHERE(f,l)   f ":" LWTSTR(l)
740
741 #define LWT_EVENT(p1, p2, p3, p4)                                       \
742 do {                                                                    \
743         unsigned long    flags;                                         \
744         lwt_cpu_t       *cpu;                                           \
745         lwt_page_t      *p;                                             \
746         lwt_event_t     *e;                                             \
747                                                                         \
748         local_irq_save (flags);                                         \
749                                                                         \
750         if (lwt_enabled) {                                              \
751                 cpu = &lwt_cpus[smp_processor_id()];                    \
752                 p = cpu->lwtc_current_page;                             \
753                 e = &p->lwtp_events[cpu->lwtc_current_index++];         \
754                                                                         \
755                 if (cpu->lwtc_current_index >= LWT_EVENTS_PER_PAGE) {   \
756                         cpu->lwtc_current_page =                        \
757                                 list_entry (p->lwtp_list.next,          \
758                                             lwt_page_t, lwtp_list);     \
759                         cpu->lwtc_current_index = 0;                    \
760                 }                                                       \
761                                                                         \
762                 e->lwte_when  = get_cycles();                           \
763                 e->lwte_where = LWTWHERE(__FILE__,__LINE__);            \
764                 e->lwte_task  = current;                                \
765                 e->lwte_p1    = (long)(p1);                             \
766                 e->lwte_p2    = (long)(p2);                             \
767                 e->lwte_p3    = (long)(p3);                             \
768                 e->lwte_p4    = (long)(p4);                             \
769         }                                                               \
770                                                                         \
771         local_irq_restore (flags);                                      \
772 } while (0)
773 #else  /* __KERNEL__ */
774 #define LWT_EVENT(p1,p2,p3,p4)     /* no userland implementation yet */
775 #endif /* __KERNEL__ */
776 #endif /* LWT_SUPPORT */
777
778
779 #include <linux/portals_lib.h>
780
781 /*
782  * USER LEVEL STUFF BELOW
783  */
784
785 #define PORTALS_CFG_VERSION 0x00010001;
786
787 struct portals_cfg {
788         __u32 pcfg_version;
789         __u32 pcfg_command;
790
791         __u32 pcfg_nal;
792         __u32 pcfg_flags;
793
794         __u32 pcfg_gw_nal;
795         __u64 pcfg_nid;
796         __u64 pcfg_nid2;
797         __u64 pcfg_nid3;
798         __u32 pcfg_id;
799         __u32 pcfg_misc;
800         __u32 pcfg_fd;
801         __u32 pcfg_count;
802         __u32 pcfg_size;
803         __u32 pcfg_wait;
804
805         __u32 pcfg_plen1; /* buffers in userspace */
806         char *pcfg_pbuf1;
807         __u32 pcfg_plen2; /* buffers in userspace */
808         char *pcfg_pbuf2;
809 };
810
811 #define PCFG_INIT(pcfg, cmd)                            \
812 do {                                                    \
813         memset(&pcfg, 0, sizeof(pcfg));                 \
814         pcfg.pcfg_version = PORTALS_CFG_VERSION;        \
815         pcfg.pcfg_command = (cmd);                      \
816                                                         \
817 } while (0)
818
819 #define PORTAL_IOCTL_VERSION 0x00010007
820 #define PING_SYNC       0
821 #define PING_ASYNC      1
822
823 struct portal_ioctl_data {
824         __u32 ioc_len;
825         __u32 ioc_version;
826         __u64 ioc_nid;
827         __u64 ioc_nid2;
828         __u64 ioc_nid3;
829         __u32 ioc_count;
830         __u32 ioc_nal;
831         __u32 ioc_nal_cmd;
832         __u32 ioc_fd;
833         __u32 ioc_id;
834
835         __u32 ioc_flags;
836         __u32 ioc_size;
837
838         __u32 ioc_wait;
839         __u32 ioc_timeout;
840         __u32 ioc_misc;
841
842         __u32 ioc_inllen1;
843         char *ioc_inlbuf1;
844         __u32 ioc_inllen2;
845         char *ioc_inlbuf2;
846
847         __u32 ioc_plen1; /* buffers in userspace */
848         char *ioc_pbuf1;
849         __u32 ioc_plen2; /* buffers in userspace */
850         char *ioc_pbuf2;
851
852         char ioc_bulk[0];
853 };
854
855 struct portal_ioctl_hdr {
856         __u32 ioc_len;
857         __u32 ioc_version;
858 };
859
860 struct portals_debug_ioctl_data
861 {
862         struct portal_ioctl_hdr hdr;
863         unsigned int subs;
864         unsigned int debug;
865 };
866
867 #define PORTAL_IOC_INIT(data)                           \
868 do {                                                    \
869         memset(&data, 0, sizeof(data));                 \
870         data.ioc_version = PORTAL_IOCTL_VERSION;        \
871         data.ioc_len = sizeof(data);                    \
872 } while (0)
873
874 /* FIXME check conflict with lustre_lib.h */
875 #define PTL_IOC_DEBUG_MASK             _IOWR('f', 250, long)
876
877 static inline int portal_ioctl_packlen(struct portal_ioctl_data *data)
878 {
879         int len = sizeof(*data);
880         len += size_round(data->ioc_inllen1);
881         len += size_round(data->ioc_inllen2);
882         return len;
883 }
884
885 static inline int portal_ioctl_is_invalid(struct portal_ioctl_data *data)
886 {
887         if (data->ioc_len > (1<<30)) {
888                 CERROR ("PORTALS ioctl: ioc_len larger than 1<<30\n");
889                 return 1;
890         }
891         if (data->ioc_inllen1 > (1<<30)) {
892                 CERROR ("PORTALS ioctl: ioc_inllen1 larger than 1<<30\n");
893                 return 1;
894         }
895         if (data->ioc_inllen2 > (1<<30)) {
896                 CERROR ("PORTALS ioctl: ioc_inllen2 larger than 1<<30\n");
897                 return 1;
898         }
899         if (data->ioc_inlbuf1 && !data->ioc_inllen1) {
900                 CERROR ("PORTALS ioctl: inlbuf1 pointer but 0 length\n");
901                 return 1;
902         }
903         if (data->ioc_inlbuf2 && !data->ioc_inllen2) {
904                 CERROR ("PORTALS ioctl: inlbuf2 pointer but 0 length\n");
905                 return 1;
906         }
907         if (data->ioc_pbuf1 && !data->ioc_plen1) {
908                 CERROR ("PORTALS ioctl: pbuf1 pointer but 0 length\n");
909                 return 1;
910         }
911         if (data->ioc_pbuf2 && !data->ioc_plen2) {
912                 CERROR ("PORTALS ioctl: pbuf2 pointer but 0 length\n");
913                 return 1;
914         }
915         if (data->ioc_plen1 && !data->ioc_pbuf1) {
916                 CERROR ("PORTALS ioctl: plen1 nonzero but no pbuf1 pointer\n");
917                 return 1;
918         }
919         if (data->ioc_plen2 && !data->ioc_pbuf2) {
920                 CERROR ("PORTALS ioctl: plen2 nonzero but no pbuf2 pointer\n");
921                 return 1;
922         }
923         if (portal_ioctl_packlen(data) != data->ioc_len ) {
924                 CERROR ("PORTALS ioctl: packlen != ioc_len\n");
925                 return 1;
926         }
927         if (data->ioc_inllen1 &&
928             data->ioc_bulk[data->ioc_inllen1 - 1] != '\0') {
929                 CERROR ("PORTALS ioctl: inlbuf1 not 0 terminated\n");
930                 return 1;
931         }
932         if (data->ioc_inllen2 &&
933             data->ioc_bulk[size_round(data->ioc_inllen1) +
934                            data->ioc_inllen2 - 1] != '\0') {
935                 CERROR ("PORTALS ioctl: inlbuf2 not 0 terminated\n");
936                 return 1;
937         }
938         return 0;
939 }
940
941 #ifndef __KERNEL__
942 static inline int portal_ioctl_pack(struct portal_ioctl_data *data, char **pbuf,
943                                     int max)
944 {
945         char *ptr;
946         struct portal_ioctl_data *overlay;
947         data->ioc_len = portal_ioctl_packlen(data);
948         data->ioc_version = PORTAL_IOCTL_VERSION;
949
950         if (*pbuf && portal_ioctl_packlen(data) > max)
951                 return 1;
952         if (*pbuf == NULL) {
953                 *pbuf = malloc(data->ioc_len);
954         }
955         if (!*pbuf)
956                 return 1;
957         overlay = (struct portal_ioctl_data *)*pbuf;
958         memcpy(*pbuf, data, sizeof(*data));
959
960         ptr = overlay->ioc_bulk;
961         if (data->ioc_inlbuf1)
962                 LOGL(data->ioc_inlbuf1, data->ioc_inllen1, ptr);
963         if (data->ioc_inlbuf2)
964                 LOGL(data->ioc_inlbuf2, data->ioc_inllen2, ptr);
965         if (portal_ioctl_is_invalid(overlay))
966                 return 1;
967
968         return 0;
969 }
970 #else
971 #include <asm/uaccess.h>
972
973 /* buffer MUST be at least the size of portal_ioctl_hdr */
974 static inline int portal_ioctl_getdata(char *buf, char *end, void *arg)
975 {
976         struct portal_ioctl_hdr *hdr;
977         struct portal_ioctl_data *data;
978         int err;
979         ENTRY;
980
981         hdr = (struct portal_ioctl_hdr *)buf;
982         data = (struct portal_ioctl_data *)buf;
983
984         err = copy_from_user(buf, (void *)arg, sizeof(*hdr));
985         if ( err ) {
986                 EXIT;
987                 return err;
988         }
989
990         if (hdr->ioc_version != PORTAL_IOCTL_VERSION) {
991                 CERROR ("PORTALS: version mismatch kernel vs application\n");
992                 return -EINVAL;
993         }
994
995         if (hdr->ioc_len + buf >= end) {
996                 CERROR ("PORTALS: user buffer exceeds kernel buffer\n");
997                 return -EINVAL;
998         }
999
1000
1001         if (hdr->ioc_len < sizeof(struct portal_ioctl_data)) {
1002                 CERROR ("PORTALS: user buffer too small for ioctl\n");
1003                 return -EINVAL;
1004         }
1005
1006         err = copy_from_user(buf, (void *)arg, hdr->ioc_len);
1007         if ( err ) {
1008                 EXIT;
1009                 return err;
1010         }
1011
1012         if (portal_ioctl_is_invalid(data)) {
1013                 CERROR ("PORTALS: ioctl not correctly formatted\n");
1014                 return -EINVAL;
1015         }
1016
1017         if (data->ioc_inllen1) {
1018                 data->ioc_inlbuf1 = &data->ioc_bulk[0];
1019         }
1020
1021         if (data->ioc_inllen2) {
1022                 data->ioc_inlbuf2 = &data->ioc_bulk[0] +
1023                         size_round(data->ioc_inllen1);
1024         }
1025
1026         EXIT;
1027         return 0;
1028 }
1029 #endif
1030
1031 /* ioctls for manipulating snapshots 30- */
1032 #define IOC_PORTAL_TYPE                   'e'
1033 #define IOC_PORTAL_MIN_NR                 30
1034
1035 #define IOC_PORTAL_PING                    _IOWR('e', 30, long)
1036 #define IOC_PORTAL_GET_DEBUG               _IOWR('e', 31, long)
1037 #define IOC_PORTAL_CLEAR_DEBUG             _IOWR('e', 32, long)
1038 #define IOC_PORTAL_MARK_DEBUG              _IOWR('e', 33, long)
1039 #define IOC_PORTAL_PANIC                   _IOWR('e', 34, long)
1040 #define IOC_PORTAL_NAL_CMD                 _IOWR('e', 35, long)
1041 #define IOC_PORTAL_GET_NID                 _IOWR('e', 36, long)
1042 #define IOC_PORTAL_FAIL_NID                _IOWR('e', 37, long)
1043 #define IOC_PORTAL_SET_DAEMON              _IOWR('e', 38, long)
1044 #define IOC_PORTAL_LWT_CONTROL             _IOWR('e', 39, long)
1045 #define IOC_PORTAL_LWT_SNAPSHOT            _IOWR('e', 40, long)
1046 #define IOC_PORTAL_LWT_LOOKUP_STRING       _IOWR('e', 41, long)
1047 #define IOC_PORTAL_MAX_NR                             41
1048
1049 enum {
1050         QSWNAL  =  1,
1051         SOCKNAL,
1052         GMNAL,
1053         TOENAL,
1054         TCPNAL,
1055         SCIMACNAL,
1056         ROUTER,
1057         IBNAL,
1058         NAL_ENUM_END_MARKER
1059 };
1060
1061 #ifdef __KERNEL__
1062 extern ptl_handle_ni_t  kqswnal_ni;
1063 extern ptl_handle_ni_t  ksocknal_ni;
1064 extern ptl_handle_ni_t  ktoenal_ni;
1065 extern ptl_handle_ni_t  kgmnal_ni;
1066 extern ptl_handle_ni_t  kibnal_ni;
1067 extern ptl_handle_ni_t  kscimacnal_ni;
1068 #endif
1069
1070 #define PTL_NALFMT_SIZE         16
1071
1072 #define NAL_MAX_NR (NAL_ENUM_END_MARKER - 1)
1073
1074 #define NAL_CMD_REGISTER_PEER_FD     100
1075 #define NAL_CMD_CLOSE_CONNECTION     101
1076 #define NAL_CMD_REGISTER_MYNID       102
1077 #define NAL_CMD_PUSH_CONNECTION      103
1078 #define NAL_CMD_GET_CONN             104
1079 #define NAL_CMD_DEL_AUTOCONN         105
1080 #define NAL_CMD_ADD_AUTOCONN         106
1081 #define NAL_CMD_GET_AUTOCONN         107
1082 #define NAL_CMD_GET_TXDESC           108
1083 #define NAL_CMD_ADD_ROUTE            109
1084 #define NAL_CMD_DEL_ROUTE            110
1085 #define NAL_CMD_GET_ROUTE            111
1086 #define NAL_CMD_NOTIFY_ROUTER        112
1087
1088 enum {
1089         DEBUG_DAEMON_START       =  1,
1090         DEBUG_DAEMON_STOP        =  2,
1091         DEBUG_DAEMON_PAUSE       =  3,
1092         DEBUG_DAEMON_CONTINUE    =  4,
1093 };
1094
1095 /* XXX remove to lustre ASAP */
1096 struct lustre_peer {
1097         ptl_nid_t       peer_nid;
1098         ptl_handle_ni_t peer_ni;
1099 };
1100
1101
1102 /* module.c */
1103 typedef int (*nal_cmd_handler_t)(struct portals_cfg *, void * private);
1104 int kportal_nal_register(int nal, nal_cmd_handler_t handler, void * private);
1105 int kportal_nal_unregister(int nal);
1106
1107 enum cfg_record_type {
1108         PORTALS_CFG_TYPE = 1,
1109         LUSTRE_CFG_TYPE = 123,
1110 };
1111
1112 typedef int (*cfg_record_cb_t)(enum cfg_record_type, int len, void *data);
1113 int kportal_nal_cmd(struct portals_cfg *);
1114
1115 ptl_handle_ni_t *kportal_get_ni (int nal);
1116 void kportal_put_ni (int nal);
1117
1118 #ifdef __CYGWIN__
1119 # ifndef BITS_PER_LONG
1120 #  if (~0UL) == 0xffffffffUL
1121 #   define BITS_PER_LONG 32
1122 #  else
1123 #   define BITS_PER_LONG 64
1124 #  endif
1125 # endif
1126 #endif
1127
1128 #if defined(__x86_64__)
1129 # define LPU64 "%Lu"
1130 # define LPD64 "%Ld"
1131 # define LPX64 "%#Lx"
1132 # define LPSZ  "%lu"
1133 # define LPSSZ "%ld"
1134 #elif (BITS_PER_LONG == 32 || __WORDSIZE == 32)
1135 # define LPU64 "%Lu"
1136 # define LPD64 "%Ld"
1137 # define LPX64 "%#Lx"
1138 # define LPSZ  "%u"
1139 # define LPSSZ "%d"
1140 #elif (BITS_PER_LONG == 64 || __WORDSIZE == 64)
1141 # define LPU64 "%lu"
1142 # define LPD64 "%ld"
1143 # define LPX64 "%#lx"
1144 # define LPSZ  "%lu"
1145 # define LPSSZ "%ld"
1146 #endif
1147 #ifndef LPU64
1148 # error "No word size defined"
1149 #endif
1150
1151 #endif