Whamcloud - gitweb
* Applied the last patch in Bug 2306, which changes the portals router/NAL
[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 #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 PageUptodate Page_Uptodate
188 #define our_recalc_sigpending(current) recalc_sigpending(current)
189 #define num_online_cpus() smp_num_cpus
190 static inline void our_cond_resched(void)
191 {
192         if (current->need_resched)
193                schedule ();
194 }
195 #define work_struct_t       struct tq_struct
196
197 #else
198
199 #define prepare_work(wq,cb,cbdata)                                            \
200 do {                                                                          \
201         INIT_WORK((wq), (void *)(cb), (void *)(cbdata));                      \
202 } while (0)
203 #define wait_on_page wait_on_page_locked
204 #define our_recalc_sigpending(current) recalc_sigpending()
205 #define strtok(a,b) strpbrk(a, b)
206 static inline void our_cond_resched(void)
207 {
208         cond_resched();
209 }
210 #define work_struct_t      struct work_struct
211
212 #endif /* LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0) */
213
214 #ifdef PORTAL_DEBUG
215 extern void kportal_assertion_failed(char *expr, char *file, const char *func,
216                                      const int line);
217 #define LASSERT(e) ((e) ? 0 : kportal_assertion_failed( #e , __FILE__,  \
218                                                         __FUNCTION__, __LINE__))
219 /* it would be great to dump_stack() here, but some kernels
220  * export it as show_stack() and I can't be bothered to
221  * proprely engage in that dance right now */
222 #define LASSERTF(cond, fmt...)                                                \
223         do {                                                                  \
224                 if (unlikely(!(cond))) {                                      \
225                         portals_debug_msg(0, D_EMERG,  __FILE__, __FUNCTION__,\
226                                           __LINE__,  CDEBUG_STACK,            \
227                                           "ASSERTION(" #cond ") failed:" fmt);\
228                         LBUG();                                               \
229                 }                                                             \
230         } while (0)
231
232 #else
233 #define LASSERT(e)
234 #define LASSERTF(cond, fmt...) do { } while (0)
235 #endif
236
237 #ifdef __arch_um__
238 #define LBUG_WITH_LOC(file, func, line)                                 \
239 do {                                                                    \
240         CEMERG("LBUG - trying to dump log to /tmp/lustre-log\n");       \
241         portals_debug_dumplog();                                        \
242         portals_run_lbug_upcall(file, func, line);                      \
243         panic("LBUG");                                                  \
244 } while (0)
245 #else
246 #define LBUG_WITH_LOC(file, func, line)                                 \
247 do {                                                                    \
248         CEMERG("LBUG\n");                                               \
249         portals_debug_dumplog();                                        \
250         portals_run_lbug_upcall(file, func, line);                      \
251         set_task_state(current, TASK_UNINTERRUPTIBLE);                  \
252         schedule();                                                     \
253 } while (0)
254 #endif /* __arch_um__ */
255
256 #define LBUG() LBUG_WITH_LOC(__FILE__, __FUNCTION__, __LINE__)
257
258 /*
259  * Memory
260  */
261 #ifdef PORTAL_DEBUG
262 extern atomic_t portal_kmemory;
263
264 # define portal_kmem_inc(ptr, size)                                           \
265 do {                                                                          \
266         atomic_add(size, &portal_kmemory);                                    \
267 } while (0)
268
269 # define portal_kmem_dec(ptr, size) do {                                      \
270         atomic_sub(size, &portal_kmemory);                                    \
271 } while (0)
272
273 #else
274 # define portal_kmem_inc(ptr, size) do {} while (0)
275 # define portal_kmem_dec(ptr, size) do {} while (0)
276 #endif /* PORTAL_DEBUG */
277
278 #define PORTAL_VMALLOC_SIZE        16384
279
280 #define PORTAL_ALLOC_GFP(ptr, size, mask)                                 \
281 do {                                                                      \
282         LASSERT(!in_interrupt() ||                                        \
283                (size <= PORTAL_VMALLOC_SIZE && mask == GFP_ATOMIC));      \
284         if ((size) > PORTAL_VMALLOC_SIZE)                                 \
285                 (ptr) = vmalloc(size);                                    \
286         else                                                              \
287                 (ptr) = kmalloc((size), (mask));                          \
288         if ((ptr) == NULL) {                                              \
289                 CERROR("PORTALS: out of memory at %s:%d (tried to alloc '"\
290                        #ptr "' = %d)\n", __FILE__, __LINE__, (int)(size));\
291                 CERROR("PORTALS: %d total bytes allocated by portals\n",  \
292                        atomic_read(&portal_kmemory));                     \
293         } else {                                                          \
294                 portal_kmem_inc((ptr), (size));                           \
295                 memset((ptr), 0, (size));                                 \
296         }                                                                 \
297         CDEBUG(D_MALLOC, "kmalloced '" #ptr "': %d at %p (tot %d).\n",    \
298                (int)(size), (ptr), atomic_read (&portal_kmemory));        \
299 } while (0)
300
301 #define PORTAL_ALLOC(ptr, size) \
302         PORTAL_ALLOC_GFP(ptr, size, GFP_NOFS)
303
304 #define PORTAL_ALLOC_ATOMIC(ptr, size) \
305         PORTAL_ALLOC_GFP(ptr, size, GFP_ATOMIC)
306
307 #define PORTAL_FREE(ptr, size)                                          \
308 do {                                                                    \
309         int s = (size);                                                 \
310         if ((ptr) == NULL) {                                            \
311                 CERROR("PORTALS: free NULL '" #ptr "' (%d bytes) at "   \
312                        "%s:%d\n", s, __FILE__, __LINE__);               \
313                 break;                                                  \
314         }                                                               \
315         if (s > PORTAL_VMALLOC_SIZE)                                    \
316                 vfree(ptr);                                             \
317         else                                                            \
318                 kfree(ptr);                                             \
319         portal_kmem_dec((ptr), s);                                      \
320         CDEBUG(D_MALLOC, "kfreed '" #ptr "': %d at %p (tot %d).\n",     \
321                s, (ptr), atomic_read(&portal_kmemory));                 \
322 } while (0)
323
324 /* ------------------------------------------------------------------- */
325
326 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
327
328 #define PORTAL_SYMBOL_REGISTER(x) inter_module_register(#x, THIS_MODULE, &x)
329 #define PORTAL_SYMBOL_UNREGISTER(x) inter_module_unregister(#x)
330
331 #define PORTAL_SYMBOL_GET(x) ((typeof(&x))inter_module_get(#x))
332 #define PORTAL_SYMBOL_PUT(x) inter_module_put(#x)
333
334 #define PORTAL_MODULE_USE       MOD_INC_USE_COUNT
335 #define PORTAL_MODULE_UNUSE     MOD_DEC_USE_COUNT
336 #else
337
338 #define PORTAL_SYMBOL_REGISTER(x)
339 #define PORTAL_SYMBOL_UNREGISTER(x)
340
341 #define PORTAL_SYMBOL_GET(x) symbol_get(x)
342 #define PORTAL_SYMBOL_PUT(x) symbol_put(x)
343
344 #define PORTAL_MODULE_USE       try_module_get(THIS_MODULE)
345 #define PORTAL_MODULE_UNUSE     module_put(THIS_MODULE)
346
347 #endif
348
349 /******************************************************************************/
350 /* Kernel Portals Router interface */
351
352 typedef void (*kpr_fwd_callback_t)(void *arg, int error); // completion callback
353
354 /* space for routing targets to stash "stuff" in a forwarded packet */
355 typedef union {
356         long long        _alignment;
357         void            *_space[16];            /* scale with CPU arch */
358 } kprfd_scratch_t;
359
360 /* Kernel Portals Routing Forwarded message Descriptor */
361 typedef struct {
362         struct list_head     kprfd_list;        /* stash in queues (routing target can use) */
363         ptl_nid_t            kprfd_target_nid;  /* final destination NID */
364         ptl_nid_t            kprfd_gateway_nid; /* gateway NID */
365         ptl_hdr_t           *kprfd_hdr;         /* header in wire byte order */
366         int                  kprfd_nob;         /* # payload bytes */
367         int                  kprfd_niov;        /* # payload frags */
368         ptl_kiov_t          *kprfd_kiov;        /* payload fragments */
369         void                *kprfd_router_arg;  /* originating NAL's router arg */
370         kpr_fwd_callback_t   kprfd_callback;    /* completion callback */
371         void                *kprfd_callback_arg; /* completion callback arg */
372         kprfd_scratch_t      kprfd_scratch;     /* scratchpad for routing targets */
373 } kpr_fwd_desc_t;
374
375 typedef void  (*kpr_fwd_t)(void *arg, kpr_fwd_desc_t *fwd);
376 typedef void  (*kpr_notify_t)(void *arg, ptl_nid_t peer, int alive);
377
378 /* NAL's routing interface (Kernel Portals Routing Nal Interface) */
379 typedef const struct {
380         int             kprni_nalid;    /* NAL's id */
381         void           *kprni_arg;      /* Arg to pass when calling into NAL */
382         kpr_fwd_t       kprni_fwd;      /* NAL's forwarding entrypoint */
383         kpr_notify_t    kprni_notify;   /* NAL's notification entrypoint */
384 } kpr_nal_interface_t;
385
386 /* Router's routing interface (Kernel Portals Routing Router Interface) */
387 typedef const struct {
388         /* register the calling NAL with the router and get back the handle for
389          * subsequent calls */
390         int     (*kprri_register) (kpr_nal_interface_t *nal_interface,
391                                    void **router_arg);
392
393         /* ask the router to find a gateway that forwards to 'nid' and is a
394          * peer of the calling NAL; assume caller will send 'nob' bytes of
395          * payload there */
396         int     (*kprri_lookup) (void *router_arg, ptl_nid_t nid, int nob,
397                                  ptl_nid_t *gateway_nid);
398
399         /* hand a packet over to the router for forwarding */
400         kpr_fwd_t kprri_fwd_start;
401
402         /* hand a packet back to the router for completion */
403         void    (*kprri_fwd_done) (void *router_arg, kpr_fwd_desc_t *fwd,
404                                    int error);
405
406         /* notify the router about peer state */
407         void    (*kprri_notify) (void *router_arg, ptl_nid_t peer,
408                                  int alive, time_t when);
409
410         /* the calling NAL is shutting down */
411         void    (*kprri_shutdown) (void *router_arg);
412
413         /* deregister the calling NAL with the router */
414         void    (*kprri_deregister) (void *router_arg);
415
416 } kpr_router_interface_t;
417
418 /* Convenient struct for NAL to stash router interface/args */
419 typedef struct {
420         kpr_router_interface_t  *kpr_interface;
421         void                    *kpr_arg;
422 } kpr_router_t;
423
424 /* Router's control interface (Kernel Portals Routing Control Interface) */
425 typedef const struct {
426         int     (*kprci_add_route)(int gateway_nal, ptl_nid_t gateway_nid,
427                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
428         int     (*kprci_del_route)(int gateway_nal, ptl_nid_t gateway_nid,
429                                    ptl_nid_t lo_nid, ptl_nid_t hi_nid);
430         int     (*kprci_get_route)(int index, int *gateway_nal,
431                                    ptl_nid_t *gateway,
432                                    ptl_nid_t *lo_nid, ptl_nid_t *hi_nid,
433                                    int *alive);
434         int     (*kprci_notify)(int gateway_nal, ptl_nid_t gateway_nid,
435                                 int alive, time_t when);
436 } kpr_control_interface_t;
437
438 extern kpr_control_interface_t  kpr_control_interface;
439 extern kpr_router_interface_t   kpr_router_interface;
440
441 static inline int
442 kpr_register (kpr_router_t *router, kpr_nal_interface_t *nalif)
443 {
444         int    rc;
445
446         router->kpr_interface = PORTAL_SYMBOL_GET (kpr_router_interface);
447         if (router->kpr_interface == NULL)
448                 return (-ENOENT);
449
450         rc = (router->kpr_interface)->kprri_register (nalif, &router->kpr_arg);
451         if (rc != 0)
452                 router->kpr_interface = NULL;
453
454         PORTAL_SYMBOL_PUT (kpr_router_interface);
455         return (rc);
456 }
457
458 static inline int
459 kpr_routing (kpr_router_t *router)
460 {
461         return (router->kpr_interface != NULL);
462 }
463
464 static inline int
465 kpr_lookup (kpr_router_t *router, ptl_nid_t nid, int nob, ptl_nid_t *gateway_nid)
466 {
467         if (!kpr_routing (router))
468                 return (-ENETUNREACH);
469
470         return (router->kpr_interface->kprri_lookup(router->kpr_arg, nid, nob,
471                                                     gateway_nid));
472 }
473
474 static inline void
475 kpr_fwd_init (kpr_fwd_desc_t *fwd, ptl_nid_t nid, ptl_hdr_t *hdr,
476               int nob, int niov, ptl_kiov_t *kiov,
477               kpr_fwd_callback_t callback, void *callback_arg)
478 {
479         fwd->kprfd_target_nid   = nid;
480         fwd->kprfd_gateway_nid  = nid;
481         fwd->kprfd_hdr          = hdr;
482         fwd->kprfd_nob          = nob;
483         fwd->kprfd_niov         = niov;
484         fwd->kprfd_kiov         = kiov;
485         fwd->kprfd_callback     = callback;
486         fwd->kprfd_callback_arg = callback_arg;
487 }
488
489 static inline void
490 kpr_fwd_start (kpr_router_t *router, kpr_fwd_desc_t *fwd)
491 {
492         if (!kpr_routing (router))
493                 fwd->kprfd_callback (fwd->kprfd_callback_arg, -ENETUNREACH);
494         else
495                 router->kpr_interface->kprri_fwd_start (router->kpr_arg, fwd);
496 }
497
498 static inline void
499 kpr_fwd_done (kpr_router_t *router, kpr_fwd_desc_t *fwd, int error)
500 {
501         LASSERT (kpr_routing (router));
502         router->kpr_interface->kprri_fwd_done (router->kpr_arg, fwd, error);
503 }
504
505 static inline void
506 kpr_notify (kpr_router_t *router,
507             ptl_nid_t peer, int alive, time_t when)
508 {
509         if (!kpr_routing (router))
510                 return;
511
512         router->kpr_interface->kprri_notify(router->kpr_arg, peer, alive, when);
513 }
514
515 static inline void
516 kpr_shutdown (kpr_router_t *router)
517 {
518         if (kpr_routing (router))
519                 router->kpr_interface->kprri_shutdown (router->kpr_arg);
520 }
521
522 static inline void
523 kpr_deregister (kpr_router_t *router)
524 {
525         if (!kpr_routing (router))
526                 return;
527         router->kpr_interface->kprri_deregister (router->kpr_arg);
528         router->kpr_interface = NULL;
529 }
530
531 /******************************************************************************/
532
533 #ifdef PORTALS_PROFILING
534 #define prof_enum(FOO) PROF__##FOO
535 enum {
536         prof_enum(our_recvmsg),
537         prof_enum(our_sendmsg),
538         prof_enum(socknal_recv),
539         prof_enum(lib_parse),
540         prof_enum(conn_list_walk),
541         prof_enum(memcpy),
542         prof_enum(lib_finalize),
543         prof_enum(pingcli_time),
544         prof_enum(gmnal_send),
545         prof_enum(gmnal_recv),
546         MAX_PROFS
547 };
548
549 struct prof_ent {
550         char *str;
551         /* hrmph.  wrap-tastic. */
552         u32       starts;
553         u32       finishes;
554         cycles_t  total_cycles;
555         cycles_t  start;
556         cycles_t  end;
557 };
558
559 extern struct prof_ent prof_ents[MAX_PROFS];
560
561 #define PROF_START(FOO)                                         \
562         do {                                                    \
563                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
564                 pe->starts++;                                   \
565                 pe->start = get_cycles();                       \
566         } while (0)
567
568 #define PROF_FINISH(FOO)                                        \
569         do {                                                    \
570                 struct prof_ent *pe = &prof_ents[PROF__##FOO];  \
571                 pe->finishes++;                                 \
572                 pe->end = get_cycles();                         \
573                 pe->total_cycles += (pe->end - pe->start);      \
574         } while (0)
575 #else /* !PORTALS_PROFILING */
576 #define PROF_START(FOO) do {} while(0)
577 #define PROF_FINISH(FOO) do {} while(0)
578 #endif /* PORTALS_PROFILING */
579
580 /* debug.c */
581 extern spinlock_t stack_backtrace_lock;
582
583 char *portals_debug_dumpstack(void);
584 void portals_run_upcall(char **argv);
585 void portals_run_lbug_upcall(char * file, const char *fn, const int line);
586 void portals_debug_dumplog(void);
587 int portals_debug_init(unsigned long bufsize);
588 int portals_debug_cleanup(void);
589 int portals_debug_clear_buffer(void);
590 int portals_debug_mark_buffer(char *text);
591 int portals_debug_set_daemon(unsigned int cmd, unsigned int length,
592                              char *file, unsigned int size);
593 __s32 portals_debug_copy_to_user(char *buf, unsigned long len);
594 #if (__GNUC__)
595 /* Use the special GNU C __attribute__ hack to have the compiler check the
596  * printf style argument string against the actual argument count and
597  * types.
598  */
599 #ifdef printf
600 # warning printf has been defined as a macro...
601 # undef printf
602 #endif
603 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
604                        const int line, unsigned long stack,
605                        char *format, ...)
606         __attribute__ ((format (printf, 7, 8)));
607 #else
608 void portals_debug_msg(int subsys, int mask, char *file, const char *fn,
609                        const int line, unsigned long stack,
610                        const char *format, ...);
611 #endif /* __GNUC__ */
612 void portals_debug_set_level(unsigned int debug_level);
613
614 # define fprintf(a, format, b...) CDEBUG(D_OTHER, format , ## b)
615 # define printf(format, b...) CDEBUG(D_OTHER, format , ## b)
616 # define time(a) CURRENT_TIME
617
618 extern void kportal_daemonize (char *name);
619 extern void kportal_blockallsigs (void);
620
621 #else  /* !__KERNEL__ */
622 # include <stdio.h>
623 # include <stdlib.h>
624 #ifndef __CYGWIN__
625 # include <stdint.h>
626 #else
627 # include <cygwin-ioctl.h>
628 #endif
629 # include <unistd.h>
630 # include <time.h>
631 # include <asm/types.h>
632 # ifndef DEBUG_SUBSYSTEM
633 #  define DEBUG_SUBSYSTEM S_UNDEFINED
634 # endif
635 # ifdef PORTAL_DEBUG
636 #  undef NDEBUG
637 #  include <assert.h>
638 #  define LASSERT(e)     assert(e)
639 #  define LASSERTF(cond, args...)     assert(cond)
640 # else
641 #  define LASSERT(e)
642 #  define LASSERTF(cond, args...) do { } while (0)
643 # endif
644 # define printk(format, args...) printf (format, ## args)
645 # define PORTAL_ALLOC(ptr, size) do { (ptr) = malloc(size); } while (0);
646 # define PORTAL_FREE(a, b) do { free(a); } while (0);
647 # define portals_debug_msg(subsys, mask, file, fn, line, stack, format, a...) \
648     printf("%02x:%06x (@%lu %s:%s,l. %d %d %lu): " format,                    \
649            (subsys), (mask), (long)time(0), file, fn, line,                   \
650            getpid() , stack, ## a);
651 #endif
652
653 /* support decl needed both by kernel and liblustre */
654 char *portals_nid2str(int nal, ptl_nid_t nid, char *str);
655
656 #ifndef CURRENT_TIME
657 # define CURRENT_TIME time(0)
658 #endif
659
660 /******************************************************************************/
661 /* Light-weight trace
662  * Support for temporary event tracing with minimal Heisenberg effect. */
663 #define LWT_SUPPORT  0
664
665 #define LWT_MEMORY   (64<<20)
666 #define LWT_MAX_CPUS 4
667
668 typedef struct {
669         cycles_t    lwte_when;
670         char       *lwte_where;
671         void       *lwte_task;
672         long        lwte_p1;
673         long        lwte_p2;
674         long        lwte_p3;
675         long        lwte_p4;
676 #if BITS_PER_LONG > 32
677         long        lwte_pad;
678 #endif
679 } lwt_event_t;
680
681 #if LWT_SUPPORT
682 #ifdef __KERNEL__
683 #define LWT_EVENTS_PER_PAGE (PAGE_SIZE / sizeof (lwt_event_t))
684
685 typedef struct _lwt_page {
686         struct list_head     lwtp_list;
687         struct page         *lwtp_page;
688         lwt_event_t         *lwtp_events;
689 } lwt_page_t;
690
691 typedef struct {
692         int                lwtc_current_index;
693         lwt_page_t        *lwtc_current_page;
694 } lwt_cpu_t;
695
696 extern int       lwt_enabled;
697 extern lwt_cpu_t lwt_cpus[];
698
699 extern int  lwt_init (void);
700 extern void lwt_fini (void);
701 extern int  lwt_lookup_string (int *size, char *knlptr,
702                                char *usrptr, int usrsize);
703 extern int  lwt_control (int enable, int clear);
704 extern int  lwt_snapshot (cycles_t *now, int *ncpu, int *total_size,
705                           void *user_ptr, int user_size);
706
707 /* Note that we _don't_ define LWT_EVENT at all if LWT_SUPPORT isn't set.
708  * This stuff is meant for finding specific problems; it never stays in
709  * production code... */
710
711 #define LWTSTR(n)       #n
712 #define LWTWHERE(f,l)   f ":" LWTSTR(l)
713
714 #define LWT_EVENT(p1, p2, p3, p4)                                       \
715 do {                                                                    \
716         unsigned long    flags;                                         \
717         lwt_cpu_t       *cpu;                                           \
718         lwt_page_t      *p;                                             \
719         lwt_event_t     *e;                                             \
720                                                                         \
721         local_irq_save (flags);                                         \
722                                                                         \
723         if (lwt_enabled) {                                              \
724                 cpu = &lwt_cpus[smp_processor_id()];                    \
725                 p = cpu->lwtc_current_page;                             \
726                 e = &p->lwtp_events[cpu->lwtc_current_index++];         \
727                                                                         \
728                 if (cpu->lwtc_current_index >= LWT_EVENTS_PER_PAGE) {   \
729                         cpu->lwtc_current_page =                        \
730                                 list_entry (p->lwtp_list.next,          \
731                                             lwt_page_t, lwtp_list);     \
732                         cpu->lwtc_current_index = 0;                    \
733                 }                                                       \
734                                                                         \
735                 e->lwte_when  = get_cycles();                           \
736                 e->lwte_where = LWTWHERE(__FILE__,__LINE__);            \
737                 e->lwte_task  = current;                                \
738                 e->lwte_p1    = (long)(p1);                             \
739                 e->lwte_p2    = (long)(p2);                             \
740                 e->lwte_p3    = (long)(p3);                             \
741                 e->lwte_p4    = (long)(p4);                             \
742         }                                                               \
743                                                                         \
744         local_irq_restore (flags);                                      \
745 } while (0)
746 #else  /* __KERNEL__ */
747 #define LWT_EVENT(p1,p2,p3,p4)     /* no userland implementation yet */
748 #endif /* __KERNEL__ */
749 #endif /* LWT_SUPPORT */
750
751 struct portals_device_userstate
752 {
753         int          pdu_memhog_pages;
754         struct page *pdu_memhog_root_page;
755 };
756
757 #include <linux/portals_lib.h>
758
759 /*
760  * USER LEVEL STUFF BELOW
761  */
762
763 #define PORTALS_CFG_VERSION 0x00010001;
764
765 struct portals_cfg {
766         __u32 pcfg_version;
767         __u32 pcfg_command;
768
769         __u32 pcfg_nal;
770         __u32 pcfg_flags;
771
772         __u32 pcfg_gw_nal;
773         __u64 pcfg_nid;
774         __u64 pcfg_nid2;
775         __u64 pcfg_nid3;
776         __u32 pcfg_id;
777         __u32 pcfg_misc;
778         __u32 pcfg_fd;
779         __u32 pcfg_count;
780         __u32 pcfg_size;
781         __u32 pcfg_wait;
782
783         __u32 pcfg_plen1; /* buffers in userspace */
784         char *pcfg_pbuf1;
785         __u32 pcfg_plen2; /* buffers in userspace */
786         char *pcfg_pbuf2;
787 };
788
789 #define PCFG_INIT(pcfg, cmd)                            \
790 do {                                                    \
791         memset(&pcfg, 0, sizeof(pcfg));                 \
792         pcfg.pcfg_version = PORTALS_CFG_VERSION;        \
793         pcfg.pcfg_command = (cmd);                      \
794                                                         \
795 } while (0)
796
797 #define PORTAL_IOCTL_VERSION 0x00010007
798 #define PING_SYNC       0
799 #define PING_ASYNC      1
800
801 struct portal_ioctl_data {
802         __u32 ioc_len;
803         __u32 ioc_version;
804         __u64 ioc_nid;
805         __u64 ioc_nid2;
806         __u64 ioc_nid3;
807         __u32 ioc_count;
808         __u32 ioc_nal;
809         __u32 ioc_nal_cmd;
810         __u32 ioc_fd;
811         __u32 ioc_id;
812
813         __u32 ioc_flags;
814         __u32 ioc_size;
815
816         __u32 ioc_wait;
817         __u32 ioc_timeout;
818         __u32 ioc_misc;
819
820         __u32 ioc_inllen1;
821         char *ioc_inlbuf1;
822         __u32 ioc_inllen2;
823         char *ioc_inlbuf2;
824
825         __u32 ioc_plen1; /* buffers in userspace */
826         char *ioc_pbuf1;
827         __u32 ioc_plen2; /* buffers in userspace */
828         char *ioc_pbuf2;
829
830         char ioc_bulk[0];
831 };
832
833 struct portal_ioctl_hdr {
834         __u32 ioc_len;
835         __u32 ioc_version;
836 };
837
838 struct portals_debug_ioctl_data
839 {
840         struct portal_ioctl_hdr hdr;
841         unsigned int subs;
842         unsigned int debug;
843 };
844
845 #define PORTAL_IOC_INIT(data)                           \
846 do {                                                    \
847         memset(&data, 0, sizeof(data));                 \
848         data.ioc_version = PORTAL_IOCTL_VERSION;        \
849         data.ioc_len = sizeof(data);                    \
850 } while (0)
851
852 /* FIXME check conflict with lustre_lib.h */
853 #define PTL_IOC_DEBUG_MASK             _IOWR('f', 250, long)
854
855 static inline int portal_ioctl_packlen(struct portal_ioctl_data *data)
856 {
857         int len = sizeof(*data);
858         len += size_round(data->ioc_inllen1);
859         len += size_round(data->ioc_inllen2);
860         return len;
861 }
862
863 static inline int portal_ioctl_is_invalid(struct portal_ioctl_data *data)
864 {
865         if (data->ioc_len > (1<<30)) {
866                 CERROR ("PORTALS ioctl: ioc_len larger than 1<<30\n");
867                 return 1;
868         }
869         if (data->ioc_inllen1 > (1<<30)) {
870                 CERROR ("PORTALS ioctl: ioc_inllen1 larger than 1<<30\n");
871                 return 1;
872         }
873         if (data->ioc_inllen2 > (1<<30)) {
874                 CERROR ("PORTALS ioctl: ioc_inllen2 larger than 1<<30\n");
875                 return 1;
876         }
877         if (data->ioc_inlbuf1 && !data->ioc_inllen1) {
878                 CERROR ("PORTALS ioctl: inlbuf1 pointer but 0 length\n");
879                 return 1;
880         }
881         if (data->ioc_inlbuf2 && !data->ioc_inllen2) {
882                 CERROR ("PORTALS ioctl: inlbuf2 pointer but 0 length\n");
883                 return 1;
884         }
885         if (data->ioc_pbuf1 && !data->ioc_plen1) {
886                 CERROR ("PORTALS ioctl: pbuf1 pointer but 0 length\n");
887                 return 1;
888         }
889         if (data->ioc_pbuf2 && !data->ioc_plen2) {
890                 CERROR ("PORTALS ioctl: pbuf2 pointer but 0 length\n");
891                 return 1;
892         }
893         if (data->ioc_plen1 && !data->ioc_pbuf1) {
894                 CERROR ("PORTALS ioctl: plen1 nonzero but no pbuf1 pointer\n");
895                 return 1;
896         }
897         if (data->ioc_plen2 && !data->ioc_pbuf2) {
898                 CERROR ("PORTALS ioctl: plen2 nonzero but no pbuf2 pointer\n");
899                 return 1;
900         }
901         if (portal_ioctl_packlen(data) != data->ioc_len ) {
902                 CERROR ("PORTALS ioctl: packlen != ioc_len\n");
903                 return 1;
904         }
905         if (data->ioc_inllen1 &&
906             data->ioc_bulk[data->ioc_inllen1 - 1] != '\0') {
907                 CERROR ("PORTALS ioctl: inlbuf1 not 0 terminated\n");
908                 return 1;
909         }
910         if (data->ioc_inllen2 &&
911             data->ioc_bulk[size_round(data->ioc_inllen1) +
912                            data->ioc_inllen2 - 1] != '\0') {
913                 CERROR ("PORTALS ioctl: inlbuf2 not 0 terminated\n");
914                 return 1;
915         }
916         return 0;
917 }
918
919 #ifndef __KERNEL__
920 static inline int portal_ioctl_pack(struct portal_ioctl_data *data, char **pbuf,
921                                     int max)
922 {
923         char *ptr;
924         struct portal_ioctl_data *overlay;
925         data->ioc_len = portal_ioctl_packlen(data);
926         data->ioc_version = PORTAL_IOCTL_VERSION;
927
928         if (*pbuf && portal_ioctl_packlen(data) > max)
929                 return 1;
930         if (*pbuf == NULL) {
931                 *pbuf = malloc(data->ioc_len);
932         }
933         if (!*pbuf)
934                 return 1;
935         overlay = (struct portal_ioctl_data *)*pbuf;
936         memcpy(*pbuf, data, sizeof(*data));
937
938         ptr = overlay->ioc_bulk;
939         if (data->ioc_inlbuf1)
940                 LOGL(data->ioc_inlbuf1, data->ioc_inllen1, ptr);
941         if (data->ioc_inlbuf2)
942                 LOGL(data->ioc_inlbuf2, data->ioc_inllen2, ptr);
943         if (portal_ioctl_is_invalid(overlay))
944                 return 1;
945
946         return 0;
947 }
948 #else
949 #include <asm/uaccess.h>
950
951 /* buffer MUST be at least the size of portal_ioctl_hdr */
952 static inline int portal_ioctl_getdata(char *buf, char *end, void *arg)
953 {
954         struct portal_ioctl_hdr *hdr;
955         struct portal_ioctl_data *data;
956         int err;
957         ENTRY;
958
959         hdr = (struct portal_ioctl_hdr *)buf;
960         data = (struct portal_ioctl_data *)buf;
961
962         err = copy_from_user(buf, (void *)arg, sizeof(*hdr));
963         if ( err ) {
964                 EXIT;
965                 return err;
966         }
967
968         if (hdr->ioc_version != PORTAL_IOCTL_VERSION) {
969                 CERROR ("PORTALS: version mismatch kernel vs application\n");
970                 return -EINVAL;
971         }
972
973         if (hdr->ioc_len + buf >= end) {
974                 CERROR ("PORTALS: user buffer exceeds kernel buffer\n");
975                 return -EINVAL;
976         }
977
978
979         if (hdr->ioc_len < sizeof(struct portal_ioctl_data)) {
980                 CERROR ("PORTALS: user buffer too small for ioctl\n");
981                 return -EINVAL;
982         }
983
984         err = copy_from_user(buf, (void *)arg, hdr->ioc_len);
985         if ( err ) {
986                 EXIT;
987                 return err;
988         }
989
990         if (portal_ioctl_is_invalid(data)) {
991                 CERROR ("PORTALS: ioctl not correctly formatted\n");
992                 return -EINVAL;
993         }
994
995         if (data->ioc_inllen1) {
996                 data->ioc_inlbuf1 = &data->ioc_bulk[0];
997         }
998
999         if (data->ioc_inllen2) {
1000                 data->ioc_inlbuf2 = &data->ioc_bulk[0] +
1001                         size_round(data->ioc_inllen1);
1002         }
1003
1004         EXIT;
1005         return 0;
1006 }
1007 #endif
1008
1009 /* ioctls for manipulating snapshots 30- */
1010 #define IOC_PORTAL_TYPE                   'e'
1011 #define IOC_PORTAL_MIN_NR                 30
1012
1013 #define IOC_PORTAL_PING                    _IOWR('e', 30, long)
1014 #define IOC_PORTAL_GET_DEBUG               _IOWR('e', 31, long)
1015 #define IOC_PORTAL_CLEAR_DEBUG             _IOWR('e', 32, long)
1016 #define IOC_PORTAL_MARK_DEBUG              _IOWR('e', 33, long)
1017 #define IOC_PORTAL_PANIC                   _IOWR('e', 34, long)
1018 #define IOC_PORTAL_NAL_CMD                 _IOWR('e', 35, long)
1019 #define IOC_PORTAL_GET_NID                 _IOWR('e', 36, long)
1020 #define IOC_PORTAL_FAIL_NID                _IOWR('e', 37, long)
1021 #define IOC_PORTAL_SET_DAEMON              _IOWR('e', 38, long)
1022 #define IOC_PORTAL_LWT_CONTROL             _IOWR('e', 39, long)
1023 #define IOC_PORTAL_LWT_SNAPSHOT            _IOWR('e', 40, long)
1024 #define IOC_PORTAL_LWT_LOOKUP_STRING       _IOWR('e', 41, long)
1025 #define IOC_PORTAL_MEMHOG                  _IOWR('e', 42, long)
1026 #define IOC_PORTAL_MAX_NR                             42
1027
1028 enum {
1029         QSWNAL    = 1,
1030         SOCKNAL   = 2,
1031         GMNAL     = 3,
1032         /*          4 unused */
1033         TCPNAL    = 5,
1034         SCIMACNAL = 6,
1035         ROUTER    = 7,
1036         IBNAL     = 8,
1037         NAL_ENUM_END_MARKER
1038 };
1039
1040 #ifdef __KERNEL__
1041 extern ptl_handle_ni_t  kqswnal_ni;
1042 extern ptl_handle_ni_t  ksocknal_ni;
1043 extern ptl_handle_ni_t  kgmnal_ni;
1044 extern ptl_handle_ni_t  kibnal_ni;
1045 extern ptl_handle_ni_t  kscimacnal_ni;
1046 #endif
1047
1048 #define PTL_NALFMT_SIZE         16
1049
1050 #define NAL_MAX_NR (NAL_ENUM_END_MARKER - 1)
1051
1052 #define NAL_CMD_REGISTER_PEER_FD     100
1053 #define NAL_CMD_CLOSE_CONNECTION     101
1054 #define NAL_CMD_REGISTER_MYNID       102
1055 #define NAL_CMD_PUSH_CONNECTION      103
1056 #define NAL_CMD_GET_CONN             104
1057 #define NAL_CMD_DEL_AUTOCONN         105
1058 #define NAL_CMD_ADD_AUTOCONN         106
1059 #define NAL_CMD_GET_AUTOCONN         107
1060 #define NAL_CMD_GET_TXDESC           108
1061 #define NAL_CMD_ADD_ROUTE            109
1062 #define NAL_CMD_DEL_ROUTE            110
1063 #define NAL_CMD_GET_ROUTE            111
1064 #define NAL_CMD_NOTIFY_ROUTER        112
1065
1066 enum {
1067         DEBUG_DAEMON_START       =  1,
1068         DEBUG_DAEMON_STOP        =  2,
1069         DEBUG_DAEMON_PAUSE       =  3,
1070         DEBUG_DAEMON_CONTINUE    =  4,
1071 };
1072
1073 /* XXX remove to lustre ASAP */
1074 struct lustre_peer {
1075         ptl_nid_t       peer_nid;
1076         ptl_handle_ni_t peer_ni;
1077 };
1078
1079
1080 /* module.c */
1081 typedef int (*nal_cmd_handler_t)(struct portals_cfg *, void * private);
1082 int kportal_nal_register(int nal, nal_cmd_handler_t handler, void * private);
1083 int kportal_nal_unregister(int nal);
1084
1085 enum cfg_record_type {
1086         PORTALS_CFG_TYPE = 1,
1087         LUSTRE_CFG_TYPE = 123,
1088 };
1089
1090 typedef int (*cfg_record_cb_t)(enum cfg_record_type, int len, void *data);
1091 int kportal_nal_cmd(struct portals_cfg *);
1092
1093 ptl_handle_ni_t *kportal_get_ni (int nal);
1094 void kportal_put_ni (int nal);
1095
1096 #ifdef __CYGWIN__
1097 # ifndef BITS_PER_LONG
1098 #  if (~0UL) == 0xffffffffUL
1099 #   define BITS_PER_LONG 32
1100 #  else
1101 #   define BITS_PER_LONG 64
1102 #  endif
1103 # endif
1104 #endif
1105
1106 #if defined(__x86_64__)
1107 # define LPU64 "%Lu"
1108 # define LPD64 "%Ld"
1109 # define LPX64 "%#Lx"
1110 # define LPSZ  "%lu"
1111 # define LPSSZ "%ld"
1112 #elif (BITS_PER_LONG == 32 || __WORDSIZE == 32)
1113 # define LPU64 "%Lu"
1114 # define LPD64 "%Ld"
1115 # define LPX64 "%#Lx"
1116 # define LPSZ  "%u"
1117 # define LPSSZ "%d"
1118 #elif (BITS_PER_LONG == 64 || __WORDSIZE == 64)
1119 # define LPU64 "%lu"
1120 # define LPD64 "%ld"
1121 # define LPX64 "%#lx"
1122 # define LPSZ  "%lu"
1123 # define LPSSZ "%ld"
1124 #endif
1125 #ifndef LPU64
1126 # error "No word size defined"
1127 #endif
1128
1129 #endif