1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * Copyright 2017 Joyent, Inc.
29  * Copyright (c) 2012 by Delphix. All rights reserved.
30  */
31 
32 #ifndef _SYS_DTRACE_IMPL_H
33 #define	_SYS_DTRACE_IMPL_H
34 
35 #ifdef	__cplusplus
36 extern "C" {
37 #endif
38 
39 /*
40  * DTrace Dynamic Tracing Software: Kernel Implementation Interfaces
41  *
42  * Note: The contents of this file are private to the implementation of the
43  * Solaris system and DTrace subsystem and are subject to change at any time
44  * without notice.  Applications and drivers using these interfaces will fail
45  * to run on future releases.  These interfaces should not be used for any
46  * purpose except those expressly outlined in dtrace(4D) and libdtrace(3LIB).
47  * Please refer to the "Solaris Dynamic Tracing Guide" for more information.
48  */
49 
50 #include <sys/dtrace.h>
51 
52 /*
53  * DTrace Implementation Constants and Typedefs
54  */
55 #define	DTRACE_MAXPROPLEN		128
56 #define	DTRACE_DYNVAR_CHUNKSIZE		256
57 
58 struct dtrace_probe;
59 struct dtrace_ecb;
60 struct dtrace_predicate;
61 struct dtrace_action;
62 struct dtrace_provider;
63 struct dtrace_state;
64 
65 typedef struct dtrace_probe dtrace_probe_t;
66 typedef struct dtrace_ecb dtrace_ecb_t;
67 typedef struct dtrace_predicate dtrace_predicate_t;
68 typedef struct dtrace_action dtrace_action_t;
69 typedef struct dtrace_provider dtrace_provider_t;
70 typedef struct dtrace_meta dtrace_meta_t;
71 typedef struct dtrace_state dtrace_state_t;
72 typedef uint32_t dtrace_optid_t;
73 typedef uint32_t dtrace_specid_t;
74 typedef uint64_t dtrace_genid_t;
75 
76 /*
77  * DTrace Probes
78  *
79  * The probe is the fundamental unit of the DTrace architecture.  Probes are
80  * created by DTrace providers, and managed by the DTrace framework.  A probe
81  * is identified by a unique <provider, module, function, name> tuple, and has
82  * a unique probe identifier assigned to it.  (Some probes are not associated
83  * with a specific point in text; these are called _unanchored probes_ and have
84  * no module or function associated with them.)  Probes are represented as a
85  * dtrace_probe structure.  To allow quick lookups based on each element of the
86  * probe tuple, probes are hashed by each of provider, module, function and
87  * name.  (If a lookup is performed based on a regular expression, a
88  * dtrace_probekey is prepared, and a linear search is performed.) Each probe
89  * is additionally pointed to by a linear array indexed by its identifier.  The
90  * identifier is the provider's mechanism for indicating to the DTrace
91  * framework that a probe has fired:  the identifier is passed as the first
92  * argument to dtrace_probe(), where it is then mapped into the corresponding
93  * dtrace_probe structure.  From the dtrace_probe structure, dtrace_probe() can
94  * iterate over the probe's list of enabling control blocks; see "DTrace
95  * Enabling Control Blocks", below.)
96  */
97 struct dtrace_probe {
98 	dtrace_id_t dtpr_id;			/* probe identifier */
99 	dtrace_ecb_t *dtpr_ecb;			/* ECB list; see below */
100 	dtrace_ecb_t *dtpr_ecb_last;		/* last ECB in list */
101 	void *dtpr_arg;				/* provider argument */
102 	dtrace_cacheid_t dtpr_predcache;	/* predicate cache ID */
103 	int dtpr_aframes;			/* artificial frames */
104 	dtrace_provider_t *dtpr_provider;	/* pointer to provider */
105 	char *dtpr_mod;				/* probe's module name */
106 	char *dtpr_func;			/* probe's function name */
107 	char *dtpr_name;			/* probe's name */
108 	dtrace_probe_t *dtpr_nextmod;		/* next in module hash */
109 	dtrace_probe_t *dtpr_prevmod;		/* previous in module hash */
110 	dtrace_probe_t *dtpr_nextfunc;		/* next in function hash */
111 	dtrace_probe_t *dtpr_prevfunc;		/* previous in function hash */
112 	dtrace_probe_t *dtpr_nextname;		/* next in name hash */
113 	dtrace_probe_t *dtpr_prevname;		/* previous in name hash */
114 	dtrace_genid_t dtpr_gen;		/* probe generation ID */
115 };
116 
117 typedef int dtrace_probekey_f(const char *, const char *, int);
118 
119 typedef struct dtrace_probekey {
120 	const char *dtpk_prov;			/* provider name to match */
121 	dtrace_probekey_f *dtpk_pmatch;		/* provider matching function */
122 	const char *dtpk_mod;			/* module name to match */
123 	dtrace_probekey_f *dtpk_mmatch;		/* module matching function */
124 	const char *dtpk_func;			/* func name to match */
125 	dtrace_probekey_f *dtpk_fmatch;		/* func matching function */
126 	const char *dtpk_name;			/* name to match */
127 	dtrace_probekey_f *dtpk_nmatch;		/* name matching function */
128 	dtrace_id_t dtpk_id;			/* identifier to match */
129 } dtrace_probekey_t;
130 
131 typedef struct dtrace_hashbucket {
132 	struct dtrace_hashbucket *dthb_next;	/* next on hash chain */
133 	dtrace_probe_t *dthb_chain;		/* chain of probes */
134 	int dthb_len;				/* number of probes here */
135 } dtrace_hashbucket_t;
136 
137 typedef struct dtrace_hash {
138 	dtrace_hashbucket_t **dth_tab;		/* hash table */
139 	int dth_size;				/* size of hash table */
140 	int dth_mask;				/* mask to index into table */
141 	int dth_nbuckets;			/* total number of buckets */
142 	uintptr_t dth_nextoffs;			/* offset of next in probe */
143 	uintptr_t dth_prevoffs;			/* offset of prev in probe */
144 	uintptr_t dth_stroffs;			/* offset of str in probe */
145 } dtrace_hash_t;
146 
147 /*
148  * DTrace Enabling Control Blocks
149  *
150  * When a provider wishes to fire a probe, it calls into dtrace_probe(),
151  * passing the probe identifier as the first argument.  As described above,
152  * dtrace_probe() maps the identifier into a pointer to a dtrace_probe_t
153  * structure.  This structure contains information about the probe, and a
154  * pointer to the list of Enabling Control Blocks (ECBs).  Each ECB points to
155  * DTrace consumer state, and contains an optional predicate, and a list of
156  * actions.  (Shown schematically below.)  The ECB abstraction allows a single
157  * probe to be multiplexed across disjoint consumers, or across disjoint
158  * enablings of a single probe within one consumer.
159  *
160  *   Enabling Control Block
161  *        dtrace_ecb_t
162  * +------------------------+
163  * | dtrace_epid_t ---------+--------------> Enabled Probe ID (EPID)
164  * | dtrace_state_t * ------+--------------> State associated with this ECB
165  * | dtrace_predicate_t * --+---------+
166  * | dtrace_action_t * -----+----+    |
167  * | dtrace_ecb_t * ---+    |    |    |       Predicate (if any)
168  * +-------------------+----+    |    |       dtrace_predicate_t
169  *                     |         |    +---> +--------------------+
170  *                     |         |          | dtrace_difo_t * ---+----> DIFO
171  *                     |         |          +--------------------+
172  *                     |         |
173  *            Next ECB |         |           Action
174  *            (if any) |         |       dtrace_action_t
175  *                     :         +--> +-------------------+
176  *                     :              | dtrace_actkind_t -+------> kind
177  *                     v              | dtrace_difo_t * --+------> DIFO (if any)
178  *                                    | dtrace_recdesc_t -+------> record descr.
179  *                                    | dtrace_action_t * +------+
180  *                                    +-------------------+      |
181  *                                                               | Next action
182  *                               +-------------------------------+  (if any)
183  *                               |
184  *                               |           Action
185  *                               |       dtrace_action_t
186  *                               +--> +-------------------+
187  *                                    | dtrace_actkind_t -+------> kind
188  *                                    | dtrace_difo_t * --+------> DIFO (if any)
189  *                                    | dtrace_action_t * +------+
190  *                                    +-------------------+      |
191  *                                                               | Next action
192  *                               +-------------------------------+  (if any)
193  *                               |
194  *                               :
195  *                               v
196  *
197  *
198  * dtrace_probe() iterates over the ECB list.  If the ECB needs less space
199  * than is available in the principal buffer, the ECB is processed:  if the
200  * predicate is non-NULL, the DIF object is executed.  If the result is
201  * non-zero, the action list is processed, with each action being executed
202  * accordingly.  When the action list has been completely executed, processing
203  * advances to the next ECB. The ECB abstraction allows disjoint consumers
204  * to multiplex on single probes.
205  *
206  * Execution of the ECB results in consuming dte_size bytes in the buffer
207  * to record data.  During execution, dte_needed bytes must be available in
208  * the buffer.  This space is used for both recorded data and tuple data.
209  */
210 struct dtrace_ecb {
211 	dtrace_epid_t dte_epid;			/* enabled probe ID */
212 	uint32_t dte_alignment;			/* required alignment */
213 	size_t dte_needed;			/* space needed for execution */
214 	size_t dte_size;			/* size of recorded payload */
215 	dtrace_predicate_t *dte_predicate;	/* predicate, if any */
216 	dtrace_action_t *dte_action;		/* actions, if any */
217 	dtrace_ecb_t *dte_next;			/* next ECB on probe */
218 	dtrace_state_t *dte_state;		/* pointer to state */
219 	uint32_t dte_cond;			/* security condition */
220 	dtrace_probe_t *dte_probe;		/* pointer to probe */
221 	dtrace_action_t *dte_action_last;	/* last action on ECB */
222 	uint64_t dte_uarg;			/* library argument */
223 };
224 
225 struct dtrace_predicate {
226 	dtrace_difo_t *dtp_difo;		/* DIF object */
227 	dtrace_cacheid_t dtp_cacheid;		/* cache identifier */
228 	int dtp_refcnt;				/* reference count */
229 };
230 
231 struct dtrace_action {
232 	dtrace_actkind_t dta_kind;		/* kind of action */
233 	uint16_t dta_intuple;			/* boolean:  in aggregation */
234 	uint32_t dta_refcnt;			/* reference count */
235 	dtrace_difo_t *dta_difo;		/* pointer to DIFO */
236 	dtrace_recdesc_t dta_rec;		/* record description */
237 	dtrace_action_t *dta_prev;		/* previous action */
238 	dtrace_action_t *dta_next;		/* next action */
239 };
240 
241 typedef struct dtrace_aggregation {
242 	dtrace_action_t dtag_action;		/* action; must be first */
243 	dtrace_aggid_t dtag_id;			/* identifier */
244 	dtrace_ecb_t *dtag_ecb;			/* corresponding ECB */
245 	dtrace_action_t *dtag_first;		/* first action in tuple */
246 	uint32_t dtag_base;			/* base of aggregation */
247 	uint8_t dtag_hasarg;			/* boolean:  has argument */
248 	uint64_t dtag_initial;			/* initial value */
249 	void (*dtag_aggregate)(uint64_t *, uint64_t, uint64_t);
250 } dtrace_aggregation_t;
251 
252 /*
253  * DTrace Buffers
254  *
255  * Principal buffers, aggregation buffers, and speculative buffers are all
256  * managed with the dtrace_buffer structure.  By default, this structure
257  * includes twin data buffers -- dtb_tomax and dtb_xamot -- that serve as the
258  * active and passive buffers, respectively.  For speculative buffers,
259  * dtb_xamot will be NULL; for "ring" and "fill" buffers, dtb_xamot will point
260  * to a scratch buffer.  For all buffer types, the dtrace_buffer structure is
261  * always allocated on a per-CPU basis; a single dtrace_buffer structure is
262  * never shared among CPUs.  (That is, there is never true sharing of the
263  * dtrace_buffer structure; to prevent false sharing of the structure, it must
264  * always be aligned to the coherence granularity -- generally 64 bytes.)
265  *
266  * One of the critical design decisions of DTrace is that a given ECB always
267  * stores the same quantity and type of data.  This is done to assure that the
268  * only metadata required for an ECB's traced data is the EPID.  That is, from
269  * the EPID, the consumer can determine the data layout.  (The data buffer
270  * layout is shown schematically below.)  By assuring that one can determine
271  * data layout from the EPID, the metadata stream can be separated from the
272  * data stream -- simplifying the data stream enormously.  The ECB always
273  * proceeds the recorded data as part of the dtrace_rechdr_t structure that
274  * includes the EPID and a high-resolution timestamp used for output ordering
275  * consistency.
276  *
277  *      base of data buffer --->  +--------+--------------------+--------+
278  *                                | rechdr | data               | rechdr |
279  *                                +--------+------+--------+----+--------+
280  *                                | data          | rechdr | data        |
281  *                                +---------------+--------+-------------+
282  *                                | data, cont.                          |
283  *                                +--------+--------------------+--------+
284  *                                | rechdr | data               |        |
285  *                                +--------+--------------------+        |
286  *                                |                ||                    |
287  *                                |                ||                    |
288  *                                |                \/                    |
289  *                                :                                      :
290  *                                .                                      .
291  *                                .                                      .
292  *                                .                                      .
293  *                                :                                      :
294  *                                |                                      |
295  *     limit of data buffer --->  +--------------------------------------+
296  *
297  * When evaluating an ECB, dtrace_probe() determines if the ECB's needs of the
298  * principal buffer (both scratch and payload) exceed the available space.  If
299  * the ECB's needs exceed available space (and if the principal buffer policy
300  * is the default "switch" policy), the ECB is dropped, the buffer's drop count
301  * is incremented, and processing advances to the next ECB.  If the ECB's needs
302  * can be met with the available space, the ECB is processed, but the offset in
303  * the principal buffer is only advanced if the ECB completes processing
304  * without error.
305  *
306  * When a buffer is to be switched (either because the buffer is the principal
307  * buffer with a "switch" policy or because it is an aggregation buffer), a
308  * cross call is issued to the CPU associated with the buffer.  In the cross
309  * call context, interrupts are disabled, and the active and the inactive
310  * buffers are atomically switched.  This involves switching the data pointers,
311  * copying the various state fields (offset, drops, errors, etc.) into their
312  * inactive equivalents, and clearing the state fields.  Because interrupts are
313  * disabled during this procedure, the switch is guaranteed to appear atomic to
314  * dtrace_probe().
315  *
316  * DTrace Ring Buffering
317  *
318  * To process a ring buffer correctly, one must know the oldest valid record.
319  * Processing starts at the oldest record in the buffer and continues until
320  * the end of the buffer is reached.  Processing then resumes starting with
321  * the record stored at offset 0 in the buffer, and continues until the
322  * youngest record is processed.  If trace records are of a fixed-length,
323  * determining the oldest record is trivial:
324  *
325  *   - If the ring buffer has not wrapped, the oldest record is the record
326  *     stored at offset 0.
327  *
328  *   - If the ring buffer has wrapped, the oldest record is the record stored
329  *     at the current offset.
330  *
331  * With variable length records, however, just knowing the current offset
332  * doesn't suffice for determining the oldest valid record:  assuming that one
333  * allows for arbitrary data, one has no way of searching forward from the
334  * current offset to find the oldest valid record.  (That is, one has no way
335  * of separating data from metadata.) It would be possible to simply refuse to
336  * process any data in the ring buffer between the current offset and the
337  * limit, but this leaves (potentially) an enormous amount of otherwise valid
338  * data unprocessed.
339  *
340  * To effect ring buffering, we track two offsets in the buffer:  the current
341  * offset and the _wrapped_ offset.  If a request is made to reserve some
342  * amount of data, and the buffer has wrapped, the wrapped offset is
343  * incremented until the wrapped offset minus the current offset is greater
344  * than or equal to the reserve request.  This is done by repeatedly looking
345  * up the ECB corresponding to the EPID at the current wrapped offset, and
346  * incrementing the wrapped offset by the size of the data payload
347  * corresponding to that ECB.  If this offset is greater than or equal to the
348  * limit of the data buffer, the wrapped offset is set to 0.  Thus, the
349  * current offset effectively "chases" the wrapped offset around the buffer.
350  * Schematically:
351  *
352  *      base of data buffer --->  +------+--------------------+------+
353  *                                | EPID | data               | EPID |
354  *                                +------+--------+------+----+------+
355  *                                | data          | EPID | data      |
356  *                                +---------------+------+-----------+
357  *                                | data, cont.                      |
358  *                                +------+---------------------------+
359  *                                | EPID | data                      |
360  *           current offset --->  +------+---------------------------+
361  *                                | invalid data                     |
362  *           wrapped offset --->  +------+--------------------+------+
363  *                                | EPID | data               | EPID |
364  *                                +------+--------+------+----+------+
365  *                                | data          | EPID | data      |
366  *                                +---------------+------+-----------+
367  *                                :                                  :
368  *                                .                                  .
369  *                                .        ... valid data ...        .
370  *                                .                                  .
371  *                                :                                  :
372  *                                +------+-------------+------+------+
373  *                                | EPID | data        | EPID | data |
374  *                                +------+------------++------+------+
375  *                                | data, cont.       | leftover     |
376  *     limit of data buffer --->  +-------------------+--------------+
377  *
378  * If the amount of requested buffer space exceeds the amount of space
379  * available between the current offset and the end of the buffer:
380  *
381  *  (1)  all words in the data buffer between the current offset and the limit
382  *       of the data buffer (marked "leftover", above) are set to
383  *       DTRACE_EPIDNONE
384  *
385  *  (2)  the wrapped offset is set to zero
386  *
387  *  (3)  the iteration process described above occurs until the wrapped offset
388  *       is greater than the amount of desired space.
389  *
390  * The wrapped offset is implemented by (re-)using the inactive offset.
391  * In a "switch" buffer policy, the inactive offset stores the offset in
392  * the inactive buffer; in a "ring" buffer policy, it stores the wrapped
393  * offset.
394  *
395  * DTrace Scratch Buffering
396  *
397  * Some ECBs may wish to allocate dynamically-sized temporary scratch memory.
398  * To accommodate such requests easily, scratch memory may be allocated in
399  * the buffer beyond the current offset plus the needed memory of the current
400  * ECB.  If there isn't sufficient room in the buffer for the requested amount
401  * of scratch space, the allocation fails and an error is generated.  Scratch
402  * memory is tracked in the dtrace_mstate_t and is automatically freed when
403  * the ECB ceases processing.  Note that ring buffers cannot allocate their
404  * scratch from the principal buffer -- lest they needlessly overwrite older,
405  * valid data.  Ring buffers therefore have their own dedicated scratch buffer
406  * from which scratch is allocated.
407  */
408 #define	DTRACEBUF_RING		0x0001		/* bufpolicy set to "ring" */
409 #define	DTRACEBUF_FILL		0x0002		/* bufpolicy set to "fill" */
410 #define	DTRACEBUF_NOSWITCH	0x0004		/* do not switch buffer */
411 #define	DTRACEBUF_WRAPPED	0x0008		/* ring buffer has wrapped */
412 #define	DTRACEBUF_DROPPED	0x0010		/* drops occurred */
413 #define	DTRACEBUF_ERROR		0x0020		/* errors occurred */
414 #define	DTRACEBUF_FULL		0x0040		/* "fill" buffer is full */
415 #define	DTRACEBUF_CONSUMED	0x0080		/* buffer has been consumed */
416 #define	DTRACEBUF_INACTIVE	0x0100		/* buffer is not yet active */
417 
418 typedef struct dtrace_buffer {
419 	uint64_t dtb_offset;			/* current offset in buffer */
420 	uint64_t dtb_size;			/* size of buffer */
421 	uint32_t dtb_flags;			/* flags */
422 	uint32_t dtb_drops;			/* number of drops */
423 	caddr_t dtb_tomax;			/* active buffer */
424 	caddr_t dtb_xamot;			/* inactive buffer */
425 	uint32_t dtb_xamot_flags;		/* inactive flags */
426 	uint32_t dtb_xamot_drops;		/* drops in inactive buffer */
427 	uint64_t dtb_xamot_offset;		/* offset in inactive buffer */
428 	uint32_t dtb_errors;			/* number of errors */
429 	uint32_t dtb_xamot_errors;		/* errors in inactive buffer */
430 #ifndef _LP64
431 	uint64_t dtb_pad1;			/* pad out to 64 bytes */
432 #endif
433 	uint64_t dtb_switched;			/* time of last switch */
434 	uint64_t dtb_interval;			/* observed switch interval */
435 	uint64_t dtb_pad2[6];			/* pad to avoid false sharing */
436 } dtrace_buffer_t;
437 
438 /*
439  * DTrace Aggregation Buffers
440  *
441  * Aggregation buffers use much of the same mechanism as described above
442  * ("DTrace Buffers").  However, because an aggregation is fundamentally a
443  * hash, there exists dynamic metadata associated with an aggregation buffer
444  * that is not associated with other kinds of buffers.  This aggregation
445  * metadata is _only_ relevant for the in-kernel implementation of
446  * aggregations; it is not actually relevant to user-level consumers.  To do
447  * this, we allocate dynamic aggregation data (hash keys and hash buckets)
448  * starting below the _limit_ of the buffer, and we allocate data from the
449  * _base_ of the buffer.  When the aggregation buffer is copied out, _only_ the
450  * data is copied out; the metadata is simply discarded.  Schematically,
451  * aggregation buffers look like:
452  *
453  *      base of data buffer --->  +-------+------+-----------+-------+
454  *                                | aggid | key  | value     | aggid |
455  *                                +-------+------+-----------+-------+
456  *                                | key                              |
457  *                                +-------+-------+-----+------------+
458  *                                | value | aggid | key | value      |
459  *                                +-------+------++-----+------+-----+
460  *                                | aggid | key  | value       |     |
461  *                                +-------+------+-------------+     |
462  *                                |                ||                |
463  *                                |                ||                |
464  *                                |                \/                |
465  *                                :                                  :
466  *                                .                                  .
467  *                                .                                  .
468  *                                .                                  .
469  *                                :                                  :
470  *                                |                /\                |
471  *                                |                ||   +------------+
472  *                                |                ||   |            |
473  *                                +---------------------+            |
474  *                                | hash keys                        |
475  *                                | (dtrace_aggkey structures)       |
476  *                                |                                  |
477  *                                +----------------------------------+
478  *                                | hash buckets                     |
479  *                                | (dtrace_aggbuffer structure)     |
480  *                                |                                  |
481  *     limit of data buffer --->  +----------------------------------+
482  *
483  *
484  * As implied above, just as we assure that ECBs always store a constant
485  * amount of data, we assure that a given aggregation -- identified by its
486  * aggregation ID -- always stores data of a constant quantity and type.
487  * As with EPIDs, this allows the aggregation ID to serve as the metadata for a
488  * given record.
489  *
490  * Note that the size of the dtrace_aggkey structure must be sizeof (uintptr_t)
491  * aligned.  (If this the structure changes such that this becomes false, an
492  * assertion will fail in dtrace_aggregate().)
493  */
494 typedef struct dtrace_aggkey {
495 	uint32_t dtak_hashval;			/* hash value */
496 	uint32_t dtak_action:4;			/* action -- 4 bits */
497 	uint32_t dtak_size:28;			/* size -- 28 bits */
498 	caddr_t dtak_data;			/* data pointer */
499 	struct dtrace_aggkey *dtak_next;	/* next in hash chain */
500 } dtrace_aggkey_t;
501 
502 typedef struct dtrace_aggbuffer {
503 	uintptr_t dtagb_hashsize;		/* number of buckets */
504 	uintptr_t dtagb_free;			/* free list of keys */
505 	dtrace_aggkey_t **dtagb_hash;		/* hash table */
506 } dtrace_aggbuffer_t;
507 
508 /*
509  * DTrace Speculations
510  *
511  * Speculations have a per-CPU buffer and a global state.  Once a speculation
512  * buffer has been comitted or discarded, it cannot be reused until all CPUs
513  * have taken the same action (commit or discard) on their respective
514  * speculative buffer.  However, because DTrace probes may execute in arbitrary
515  * context, other CPUs cannot simply be cross-called at probe firing time to
516  * perform the necessary commit or discard.  The speculation states thus
517  * optimize for the case that a speculative buffer is only active on one CPU at
518  * the time of a commit() or discard() -- for if this is the case, other CPUs
519  * need not take action, and the speculation is immediately available for
520  * reuse.  If the speculation is active on multiple CPUs, it must be
521  * asynchronously cleaned -- potentially leading to a higher rate of dirty
522  * speculative drops.  The speculation states are as follows:
523  *
524  *  DTRACESPEC_INACTIVE       <= Initial state; inactive speculation
525  *  DTRACESPEC_ACTIVE         <= Allocated, but not yet speculatively traced to
526  *  DTRACESPEC_ACTIVEONE      <= Speculatively traced to on one CPU
527  *  DTRACESPEC_ACTIVEMANY     <= Speculatively traced to on more than one CPU
528  *  DTRACESPEC_COMMITTING     <= Currently being commited on one CPU
529  *  DTRACESPEC_COMMITTINGMANY <= Currently being commited on many CPUs
530  *  DTRACESPEC_DISCARDING     <= Currently being discarded on many CPUs
531  *
532  * The state transition diagram is as follows:
533  *
534  *     +----------------------------------------------------------+
535  *     |                                                          |
536  *     |                      +------------+                      |
537  *     |  +-------------------| COMMITTING |<-----------------+   |
538  *     |  |                   +------------+                  |   |
539  *     |  | copied spec.            ^             commit() on |   | discard() on
540  *     |  | into principal          |              active CPU |   | active CPU
541  *     |  |                         | commit()                |   |
542  *     V  V                         |                         |   |
543  * +----------+                 +--------+                +-----------+
544  * | INACTIVE |---------------->| ACTIVE |--------------->| ACTIVEONE |
545  * +----------+  speculation()  +--------+  speculate()   +-----------+
546  *     ^  ^                         |                         |   |
547  *     |  |                         | discard()               |   |
548  *     |  | asynchronously          |            discard() on |   | speculate()
549  *     |  | cleaned                 V            inactive CPU |   | on inactive
550  *     |  |                   +------------+                  |   | CPU
551  *     |  +-------------------| DISCARDING |<-----------------+   |
552  *     |                      +------------+                      |
553  *     | asynchronously             ^                             |
554  *     | copied spec.               |       discard()             |
555  *     | into principal             +------------------------+    |
556  *     |                                                     |    V
557  *  +----------------+             commit()              +------------+
558  *  | COMMITTINGMANY |<----------------------------------| ACTIVEMANY |
559  *  +----------------+                                   +------------+
560  */
561 typedef enum dtrace_speculation_state {
562 	DTRACESPEC_INACTIVE = 0,
563 	DTRACESPEC_ACTIVE,
564 	DTRACESPEC_ACTIVEONE,
565 	DTRACESPEC_ACTIVEMANY,
566 	DTRACESPEC_COMMITTING,
567 	DTRACESPEC_COMMITTINGMANY,
568 	DTRACESPEC_DISCARDING
569 } dtrace_speculation_state_t;
570 
571 typedef struct dtrace_speculation {
572 	dtrace_speculation_state_t dtsp_state;	/* current speculation state */
573 	int dtsp_cleaning;			/* non-zero if being cleaned */
574 	dtrace_buffer_t *dtsp_buffer;		/* speculative buffer */
575 } dtrace_speculation_t;
576 
577 /*
578  * DTrace Dynamic Variables
579  *
580  * The dynamic variable problem is obviously decomposed into two subproblems:
581  * allocating new dynamic storage, and freeing old dynamic storage.  The
582  * presence of the second problem makes the first much more complicated -- or
583  * rather, the absence of the second renders the first trivial.  This is the
584  * case with aggregations, for which there is effectively no deallocation of
585  * dynamic storage.  (Or more accurately, all dynamic storage is deallocated
586  * when a snapshot is taken of the aggregation.)  As DTrace dynamic variables
587  * allow for both dynamic allocation and dynamic deallocation, the
588  * implementation of dynamic variables is quite a bit more complicated than
589  * that of their aggregation kin.
590  *
591  * We observe that allocating new dynamic storage is tricky only because the
592  * size can vary -- the allocation problem is much easier if allocation sizes
593  * are uniform.  We further observe that in D, the size of dynamic variables is
594  * actually _not_ dynamic -- dynamic variable sizes may be determined by static
595  * analysis of DIF text.  (This is true even of putatively dynamically-sized
596  * objects like strings and stacks, the sizes of which are dictated by the
597  * "stringsize" and "stackframes" variables, respectively.)  We exploit this by
598  * performing this analysis on all DIF before enabling any probes.  For each
599  * dynamic load or store, we calculate the dynamically-allocated size plus the
600  * size of the dtrace_dynvar structure plus the storage required to key the
601  * data.  For all DIF, we take the largest value and dub it the _chunksize_.
602  * We then divide dynamic memory into two parts:  a hash table that is wide
603  * enough to have every chunk in its own bucket, and a larger region of equal
604  * chunksize units.  Whenever we wish to dynamically allocate a variable, we
605  * always allocate a single chunk of memory.  Depending on the uniformity of
606  * allocation, this will waste some amount of memory -- but it eliminates the
607  * non-determinism inherent in traditional heap fragmentation.
608  *
609  * Dynamic objects are allocated by storing a non-zero value to them; they are
610  * deallocated by storing a zero value to them.  Dynamic variables are
611  * complicated enormously by being shared between CPUs.  In particular,
612  * consider the following scenario:
613  *
614  *                 CPU A                                 CPU B
615  *  +---------------------------------+   +---------------------------------+
616  *  |                                 |   |                                 |
617  *  | allocates dynamic object a[123] |   |                                 |
618  *  | by storing the value 345 to it  |   |                                 |
619  *  |                               --------->                              |
620  *  |                                 |   | wishing to load from object     |
621  *  |                                 |   | a[123], performs lookup in      |
622  *  |                                 |   | dynamic variable space          |
623  *  |                               <---------                              |
624  *  | deallocates object a[123] by    |   |                                 |
625  *  | storing 0 to it                 |   |                                 |
626  *  |                                 |   |                                 |
627  *  | allocates dynamic object b[567] |   | performs load from a[123]       |
628  *  | by storing the value 789 to it  |   |                                 |
629  *  :                                 :   :                                 :
630  *  .                                 .   .                                 .
631  *
632  * This is obviously a race in the D program, but there are nonetheless only
633  * two valid values for CPU B's load from a[123]:  345 or 0.  Most importantly,
634  * CPU B may _not_ see the value 789 for a[123].
635  *
636  * There are essentially two ways to deal with this:
637  *
638  *  (1)  Explicitly spin-lock variables.  That is, if CPU B wishes to load
639  *       from a[123], it needs to lock a[123] and hold the lock for the
640  *       duration that it wishes to manipulate it.
641  *
642  *  (2)  Avoid reusing freed chunks until it is known that no CPU is referring
643  *       to them.
644  *
645  * The implementation of (1) is rife with complexity, because it requires the
646  * user of a dynamic variable to explicitly decree when they are done using it.
647  * Were all variables by value, this perhaps wouldn't be debilitating -- but
648  * dynamic variables of non-scalar types are tracked by reference.  That is, if
649  * a dynamic variable is, say, a string, and that variable is to be traced to,
650  * say, the principal buffer, the DIF emulation code returns to the main
651  * dtrace_probe() loop a pointer to the underlying storage, not the contents of
652  * the storage.  Further, code calling on DIF emulation would have to be aware
653  * that the DIF emulation has returned a reference to a dynamic variable that
654  * has been potentially locked.  The variable would have to be unlocked after
655  * the main dtrace_probe() loop is finished with the variable, and the main
656  * dtrace_probe() loop would have to be careful to not call any further DIF
657  * emulation while the variable is locked to avoid deadlock.  More generally,
658  * if one were to implement (1), DIF emulation code dealing with dynamic
659  * variables could only deal with one dynamic variable at a time (lest deadlock
660  * result).  To sum, (1) exports too much subtlety to the users of dynamic
661  * variables -- increasing maintenance burden and imposing serious constraints
662  * on future DTrace development.
663  *
664  * The implementation of (2) is also complex, but the complexity is more
665  * manageable.  We need to be sure that when a variable is deallocated, it is
666  * not placed on a traditional free list, but rather on a _dirty_ list.  Once a
667  * variable is on a dirty list, it cannot be found by CPUs performing a
668  * subsequent lookup of the variable -- but it may still be in use by other
669  * CPUs.  To assure that all CPUs that may be seeing the old variable have
670  * cleared out of probe context, a dtrace_sync() can be issued.  Once the
671  * dtrace_sync() has completed, it can be known that all CPUs are done
672  * manipulating the dynamic variable -- the dirty list can be atomically
673  * appended to the free list.  Unfortunately, there's a slight hiccup in this
674  * mechanism:  dtrace_sync() may not be issued from probe context.  The
675  * dtrace_sync() must be therefore issued asynchronously from non-probe
676  * context.  For this we rely on the DTrace cleaner, a cyclic that runs at the
677  * "cleanrate" frequency.  To ease this implementation, we define several chunk
678  * lists:
679  *
680  *   - Dirty.  Deallocated chunks, not yet cleaned.  Not available.
681  *
682  *   - Rinsing.  Formerly dirty chunks that are currently being asynchronously
683  *     cleaned.  Not available, but will be shortly.  Dynamic variable
684  *     allocation may not spin or block for availability, however.
685  *
686  *   - Clean.  Clean chunks, ready for allocation -- but not on the free list.
687  *
688  *   - Free.  Available for allocation.
689  *
690  * Moreover, to avoid absurd contention, _each_ of these lists is implemented
691  * on a per-CPU basis.  This is only for performance, not correctness; chunks
692  * may be allocated from another CPU's free list.  The algorithm for allocation
693  * then is this:
694  *
695  *   (1)  Attempt to atomically allocate from current CPU's free list.  If list
696  *        is non-empty and allocation is successful, allocation is complete.
697  *
698  *   (2)  If the clean list is non-empty, atomically move it to the free list,
699  *        and reattempt (1).
700  *
701  *   (3)  If the dynamic variable space is in the CLEAN state, look for free
702  *        and clean lists on other CPUs by setting the current CPU to the next
703  *        CPU, and reattempting (1).  If the next CPU is the current CPU (that
704  *        is, if all CPUs have been checked), atomically switch the state of
705  *        the dynamic variable space based on the following:
706  *
707  *        - If no free chunks were found and no dirty chunks were found,
708  *          atomically set the state to EMPTY.
709  *
710  *        - If dirty chunks were found, atomically set the state to DIRTY.
711  *
712  *        - If rinsing chunks were found, atomically set the state to RINSING.
713  *
714  *   (4)  Based on state of dynamic variable space state, increment appropriate
715  *        counter to indicate dynamic drops (if in EMPTY state) vs. dynamic
716  *        dirty drops (if in DIRTY state) vs. dynamic rinsing drops (if in
717  *        RINSING state).  Fail the allocation.
718  *
719  * The cleaning cyclic operates with the following algorithm:  for all CPUs
720  * with a non-empty dirty list, atomically move the dirty list to the rinsing
721  * list.  Perform a dtrace_sync().  For all CPUs with a non-empty rinsing list,
722  * atomically move the rinsing list to the clean list.  Perform another
723  * dtrace_sync().  By this point, all CPUs have seen the new clean list; the
724  * state of the dynamic variable space can be restored to CLEAN.
725  *
726  * There exist two final races that merit explanation.  The first is a simple
727  * allocation race:
728  *
729  *                 CPU A                                 CPU B
730  *  +---------------------------------+   +---------------------------------+
731  *  |                                 |   |                                 |
732  *  | allocates dynamic object a[123] |   | allocates dynamic object a[123] |
733  *  | by storing the value 345 to it  |   | by storing the value 567 to it  |
734  *  |                                 |   |                                 |
735  *  :                                 :   :                                 :
736  *  .                                 .   .                                 .
737  *
738  * Again, this is a race in the D program.  It can be resolved by having a[123]
739  * hold the value 345 or a[123] hold the value 567 -- but it must be true that
740  * a[123] have only _one_ of these values.  (That is, the racing CPUs may not
741  * put the same element twice on the same hash chain.)  This is resolved
742  * simply:  before the allocation is undertaken, the start of the new chunk's
743  * hash chain is noted.  Later, after the allocation is complete, the hash
744  * chain is atomically switched to point to the new element.  If this fails
745  * (because of either concurrent allocations or an allocation concurrent with a
746  * deletion), the newly allocated chunk is deallocated to the dirty list, and
747  * the whole process of looking up (and potentially allocating) the dynamic
748  * variable is reattempted.
749  *
750  * The final race is a simple deallocation race:
751  *
752  *                 CPU A                                 CPU B
753  *  +---------------------------------+   +---------------------------------+
754  *  |                                 |   |                                 |
755  *  | deallocates dynamic object      |   | deallocates dynamic object      |
756  *  | a[123] by storing the value 0   |   | a[123] by storing the value 0   |
757  *  | to it                           |   | to it                           |
758  *  |                                 |   |                                 |
759  *  :                                 :   :                                 :
760  *  .                                 .   .                                 .
761  *
762  * Once again, this is a race in the D program, but it is one that we must
763  * handle without corrupting the underlying data structures.  Because
764  * deallocations require the deletion of a chunk from the middle of a hash
765  * chain, we cannot use a single-word atomic operation to remove it.  For this,
766  * we add a spin lock to the hash buckets that is _only_ used for deallocations
767  * (allocation races are handled as above).  Further, this spin lock is _only_
768  * held for the duration of the delete; before control is returned to the DIF
769  * emulation code, the hash bucket is unlocked.
770  */
771 typedef struct dtrace_key {
772 	uint64_t dttk_value;			/* data value or data pointer */
773 	uint64_t dttk_size;			/* 0 if by-val, >0 if by-ref */
774 } dtrace_key_t;
775 
776 typedef struct dtrace_tuple {
777 	uint32_t dtt_nkeys;			/* number of keys in tuple */
778 	uint32_t dtt_pad;			/* padding */
779 	dtrace_key_t dtt_key[1];		/* array of tuple keys */
780 } dtrace_tuple_t;
781 
782 typedef struct dtrace_dynvar {
783 	uint64_t dtdv_hashval;			/* hash value -- 0 if free */
784 	struct dtrace_dynvar *dtdv_next;	/* next on list or hash chain */
785 	void *dtdv_data;			/* pointer to data */
786 	dtrace_tuple_t dtdv_tuple;		/* tuple key */
787 } dtrace_dynvar_t;
788 
789 typedef enum dtrace_dynvar_op {
790 	DTRACE_DYNVAR_ALLOC,
791 	DTRACE_DYNVAR_NOALLOC,
792 	DTRACE_DYNVAR_DEALLOC
793 } dtrace_dynvar_op_t;
794 
795 typedef struct dtrace_dynhash {
796 	dtrace_dynvar_t *dtdh_chain;		/* hash chain for this bucket */
797 	uintptr_t dtdh_lock;			/* deallocation lock */
798 #ifdef _LP64
799 	uintptr_t dtdh_pad[6];			/* pad to avoid false sharing */
800 #else
801 	uintptr_t dtdh_pad[14];			/* pad to avoid false sharing */
802 #endif
803 } dtrace_dynhash_t;
804 
805 typedef struct dtrace_dstate_percpu {
806 	dtrace_dynvar_t *dtdsc_free;		/* free list for this CPU */
807 	dtrace_dynvar_t *dtdsc_dirty;		/* dirty list for this CPU */
808 	dtrace_dynvar_t *dtdsc_rinsing;		/* rinsing list for this CPU */
809 	dtrace_dynvar_t *dtdsc_clean;		/* clean list for this CPU */
810 	uint64_t dtdsc_drops;			/* number of capacity drops */
811 	uint64_t dtdsc_dirty_drops;		/* number of dirty drops */
812 	uint64_t dtdsc_rinsing_drops;		/* number of rinsing drops */
813 #ifdef _LP64
814 	uint64_t dtdsc_pad;			/* pad to avoid false sharing */
815 #else
816 	uint64_t dtdsc_pad[2];			/* pad to avoid false sharing */
817 #endif
818 } dtrace_dstate_percpu_t;
819 
820 typedef enum dtrace_dstate_state {
821 	DTRACE_DSTATE_CLEAN = 0,
822 	DTRACE_DSTATE_EMPTY,
823 	DTRACE_DSTATE_DIRTY,
824 	DTRACE_DSTATE_RINSING
825 } dtrace_dstate_state_t;
826 
827 typedef struct dtrace_dstate {
828 	void *dtds_base;			/* base of dynamic var. space */
829 	size_t dtds_size;			/* size of dynamic var. space */
830 	size_t dtds_hashsize;			/* number of buckets in hash */
831 	size_t dtds_chunksize;			/* size of each chunk */
832 	dtrace_dynhash_t *dtds_hash;		/* pointer to hash table */
833 	dtrace_dstate_state_t dtds_state;	/* current dynamic var. state */
834 	dtrace_dstate_percpu_t *dtds_percpu;	/* per-CPU dyn. var. state */
835 } dtrace_dstate_t;
836 
837 /*
838  * DTrace Variable State
839  *
840  * The DTrace variable state tracks user-defined variables in its dtrace_vstate
841  * structure.  Each DTrace consumer has exactly one dtrace_vstate structure,
842  * but some dtrace_vstate structures may exist without a corresponding DTrace
843  * consumer (see "DTrace Helpers", below).  As described in <sys/dtrace.h>,
844  * user-defined variables can have one of three scopes:
845  *
846  *  DIFV_SCOPE_GLOBAL  =>  global scope
847  *  DIFV_SCOPE_THREAD  =>  thread-local scope (i.e. "self->" variables)
848  *  DIFV_SCOPE_LOCAL   =>  clause-local scope (i.e. "this->" variables)
849  *
850  * The variable state tracks variables by both their scope and their allocation
851  * type:
852  *
853  *  - The dtvs_globals and dtvs_locals members each point to an array of
854  *    dtrace_statvar structures.  These structures contain both the variable
855  *    metadata (dtrace_difv structures) and the underlying storage for all
856  *    statically allocated variables, including statically allocated
857  *    DIFV_SCOPE_GLOBAL variables and all DIFV_SCOPE_LOCAL variables.
858  *
859  *  - The dtvs_tlocals member points to an array of dtrace_difv structures for
860  *    DIFV_SCOPE_THREAD variables.  As such, this array tracks _only_ the
861  *    variable metadata for DIFV_SCOPE_THREAD variables; the underlying storage
862  *    is allocated out of the dynamic variable space.
863  *
864  *  - The dtvs_dynvars member is the dynamic variable state associated with the
865  *    variable state.  The dynamic variable state (described in "DTrace Dynamic
866  *    Variables", above) tracks all DIFV_SCOPE_THREAD variables and all
867  *    dynamically-allocated DIFV_SCOPE_GLOBAL variables.
868  */
869 typedef struct dtrace_statvar {
870 	uint64_t dtsv_data;			/* data or pointer to it */
871 	size_t dtsv_size;			/* size of pointed-to data */
872 	int dtsv_refcnt;			/* reference count */
873 	dtrace_difv_t dtsv_var;			/* variable metadata */
874 } dtrace_statvar_t;
875 
876 typedef struct dtrace_vstate {
877 	dtrace_state_t *dtvs_state;		/* back pointer to state */
878 	dtrace_statvar_t **dtvs_globals;	/* statically-allocated glbls */
879 	int dtvs_nglobals;			/* number of globals */
880 	dtrace_difv_t *dtvs_tlocals;		/* thread-local metadata */
881 	int dtvs_ntlocals;			/* number of thread-locals */
882 	dtrace_statvar_t **dtvs_locals;		/* clause-local data */
883 	int dtvs_nlocals;			/* number of clause-locals */
884 	dtrace_dstate_t dtvs_dynvars;		/* dynamic variable state */
885 } dtrace_vstate_t;
886 
887 /*
888  * DTrace Machine State
889  *
890  * In the process of processing a fired probe, DTrace needs to track and/or
891  * cache some per-CPU state associated with that particular firing.  This is
892  * state that is always discarded after the probe firing has completed, and
893  * much of it is not specific to any DTrace consumer, remaining valid across
894  * all ECBs.  This state is tracked in the dtrace_mstate structure.
895  */
896 #define	DTRACE_MSTATE_ARGS		0x00000001
897 #define	DTRACE_MSTATE_PROBE		0x00000002
898 #define	DTRACE_MSTATE_EPID		0x00000004
899 #define	DTRACE_MSTATE_TIMESTAMP		0x00000008
900 #define	DTRACE_MSTATE_STACKDEPTH	0x00000010
901 #define	DTRACE_MSTATE_CALLER		0x00000020
902 #define	DTRACE_MSTATE_IPL		0x00000040
903 #define	DTRACE_MSTATE_FLTOFFS		0x00000080
904 #define	DTRACE_MSTATE_WALLTIMESTAMP	0x00000100
905 #define	DTRACE_MSTATE_USTACKDEPTH	0x00000200
906 #define	DTRACE_MSTATE_UCALLER		0x00000400
907 
908 typedef struct dtrace_mstate {
909 	uintptr_t dtms_scratch_base;		/* base of scratch space */
910 	uintptr_t dtms_scratch_ptr;		/* current scratch pointer */
911 	size_t dtms_scratch_size;		/* scratch size */
912 	uint32_t dtms_present;			/* variables that are present */
913 	uint64_t dtms_arg[5];			/* cached arguments */
914 	dtrace_epid_t dtms_epid;		/* current EPID */
915 	uint64_t dtms_timestamp;		/* cached timestamp */
916 	hrtime_t dtms_walltimestamp;		/* cached wall timestamp */
917 	int dtms_stackdepth;			/* cached stackdepth */
918 	int dtms_ustackdepth;			/* cached ustackdepth */
919 	struct dtrace_probe *dtms_probe;	/* current probe */
920 	uintptr_t dtms_caller;			/* cached caller */
921 	uint64_t dtms_ucaller;			/* cached user-level caller */
922 	int dtms_ipl;				/* cached interrupt pri lev */
923 	int dtms_fltoffs;			/* faulting DIFO offset */
924 	uintptr_t dtms_strtok;			/* saved strtok() pointer */
925 	uintptr_t dtms_strtok_limit;		/* upper bound of strtok ptr */
926 	uint32_t dtms_access;			/* memory access rights */
927 	dtrace_difo_t *dtms_difo;		/* current dif object */
928 	file_t *dtms_getf;			/* cached rval of getf() */
929 } dtrace_mstate_t;
930 
931 #define	DTRACE_COND_OWNER	0x1
932 #define	DTRACE_COND_USERMODE	0x2
933 #define	DTRACE_COND_ZONEOWNER	0x4
934 
935 #define	DTRACE_PROBEKEY_MAXDEPTH	8	/* max glob recursion depth */
936 
937 /*
938  * Access flag used by dtrace_mstate.dtms_access.
939  */
940 #define	DTRACE_ACCESS_KERNEL	0x1		/* the priv to read kmem */
941 #define	DTRACE_ACCESS_PROC	0x2		/* the priv for proc state */
942 #define	DTRACE_ACCESS_ARGS	0x4		/* the priv to examine args */
943 
944 /*
945  * DTrace Activity
946  *
947  * Each DTrace consumer is in one of several states, which (for purposes of
948  * avoiding yet-another overloading of the noun "state") we call the current
949  * _activity_.  The activity transitions on dtrace_go() (from DTRACIOCGO), on
950  * dtrace_stop() (from DTRACIOCSTOP) and on the exit() action.  Activities may
951  * only transition in one direction; the activity transition diagram is a
952  * directed acyclic graph.  The activity transition diagram is as follows:
953  *
954  *
955  * +----------+                   +--------+                   +--------+
956  * | INACTIVE |------------------>| WARMUP |------------------>| ACTIVE |
957  * +----------+   dtrace_go(),    +--------+   dtrace_go(),    +--------+
958  *                before BEGIN        |        after BEGIN       |  |  |
959  *                                    |                          |  |  |
960  *                      exit() action |                          |  |  |
961  *                     from BEGIN ECB |                          |  |  |
962  *                                    |                          |  |  |
963  *                                    v                          |  |  |
964  *                               +----------+     exit() action  |  |  |
965  * +-----------------------------| DRAINING |<-------------------+  |  |
966  * |                             +----------+                       |  |
967  * |                                  |                             |  |
968  * |                   dtrace_stop(), |                             |  |
969  * |                     before END   |                             |  |
970  * |                                  |                             |  |
971  * |                                  v                             |  |
972  * | +---------+                 +----------+                       |  |
973  * | | STOPPED |<----------------| COOLDOWN |<----------------------+  |
974  * | +---------+  dtrace_stop(), +----------+     dtrace_stop(),       |
975  * |                after END                       before END         |
976  * |                                                                   |
977  * |                              +--------+                           |
978  * +----------------------------->| KILLED |<--------------------------+
979  *       deadman timeout or       +--------+     deadman timeout or
980  *        killed consumer                         killed consumer
981  *
982  * Note that once a DTrace consumer has stopped tracing, there is no way to
983  * restart it; if a DTrace consumer wishes to restart tracing, it must reopen
984  * the DTrace pseudodevice.
985  */
986 typedef enum dtrace_activity {
987 	DTRACE_ACTIVITY_INACTIVE = 0,		/* not yet running */
988 	DTRACE_ACTIVITY_WARMUP,			/* while starting */
989 	DTRACE_ACTIVITY_ACTIVE,			/* running */
990 	DTRACE_ACTIVITY_DRAINING,		/* before stopping */
991 	DTRACE_ACTIVITY_COOLDOWN,		/* while stopping */
992 	DTRACE_ACTIVITY_STOPPED,		/* after stopping */
993 	DTRACE_ACTIVITY_KILLED			/* killed */
994 } dtrace_activity_t;
995 
996 /*
997  * DTrace Helper Implementation
998  *
999  * A description of the helper architecture may be found in <sys/dtrace.h>.
1000  * Each process contains a pointer to its helpers in its p_dtrace_helpers
1001  * member.  This is a pointer to a dtrace_helpers structure, which contains an
1002  * array of pointers to dtrace_helper structures, helper variable state (shared
1003  * among a process's helpers) and a generation count.  (The generation count is
1004  * used to provide an identifier when a helper is added so that it may be
1005  * subsequently removed.)  The dtrace_helper structure is self-explanatory,
1006  * containing pointers to the objects needed to execute the helper.  Note that
1007  * helpers are _duplicated_ across fork(2), and destroyed on exec(2).  No more
1008  * than dtrace_helpers_max are allowed per-process.
1009  */
1010 #define	DTRACE_HELPER_ACTION_USTACK	0
1011 #define	DTRACE_NHELPER_ACTIONS		1
1012 
1013 typedef struct dtrace_helper_action {
1014 	int dtha_generation;			/* helper action generation */
1015 	int dtha_nactions;			/* number of actions */
1016 	dtrace_difo_t *dtha_predicate;		/* helper action predicate */
1017 	dtrace_difo_t **dtha_actions;		/* array of actions */
1018 	struct dtrace_helper_action *dtha_next;	/* next helper action */
1019 } dtrace_helper_action_t;
1020 
1021 typedef struct dtrace_helper_provider {
1022 	int dthp_generation;			/* helper provider generation */
1023 	uint32_t dthp_ref;			/* reference count */
1024 	dof_helper_t dthp_prov;			/* DOF w/ provider and probes */
1025 } dtrace_helper_provider_t;
1026 
1027 typedef struct dtrace_helpers {
1028 	dtrace_helper_action_t **dthps_actions;	/* array of helper actions */
1029 	dtrace_vstate_t dthps_vstate;		/* helper action var. state */
1030 	dtrace_helper_provider_t **dthps_provs;	/* array of providers */
1031 	uint_t dthps_nprovs;			/* count of providers */
1032 	uint_t dthps_maxprovs;			/* provider array size */
1033 	int dthps_generation;			/* current generation */
1034 	pid_t dthps_pid;			/* pid of associated proc */
1035 	int dthps_deferred;			/* helper in deferred list */
1036 	struct dtrace_helpers *dthps_next;	/* next pointer */
1037 	struct dtrace_helpers *dthps_prev;	/* prev pointer */
1038 } dtrace_helpers_t;
1039 
1040 /*
1041  * DTrace Helper Action Tracing
1042  *
1043  * Debugging helper actions can be arduous.  To ease the development and
1044  * debugging of helpers, DTrace contains a tracing-framework-within-a-tracing-
1045  * framework: helper tracing.  If dtrace_helptrace_enabled is non-zero (which
1046  * it is by default on DEBUG kernels), all helper activity will be traced to a
1047  * global, in-kernel ring buffer.  Each entry includes a pointer to the specific
1048  * helper, the location within the helper, and a trace of all local variables.
1049  * The ring buffer may be displayed in a human-readable format with the
1050  * ::dtrace_helptrace mdb(1) dcmd.
1051  */
1052 #define	DTRACE_HELPTRACE_NEXT	(-1)
1053 #define	DTRACE_HELPTRACE_DONE	(-2)
1054 #define	DTRACE_HELPTRACE_ERR	(-3)
1055 
1056 typedef struct dtrace_helptrace {
1057 	dtrace_helper_action_t	*dtht_helper;	/* helper action */
1058 	int dtht_where;				/* where in helper action */
1059 	int dtht_nlocals;			/* number of locals */
1060 	int dtht_fault;				/* type of fault (if any) */
1061 	int dtht_fltoffs;			/* DIF offset */
1062 	uint64_t dtht_illval;			/* faulting value */
1063 	uint64_t dtht_locals[1];		/* local variables */
1064 } dtrace_helptrace_t;
1065 
1066 /*
1067  * DTrace Credentials
1068  *
1069  * In probe context, we have limited flexibility to examine the credentials
1070  * of the DTrace consumer that created a particular enabling.  We use
1071  * the Least Privilege interfaces to cache the consumer's cred pointer and
1072  * some facts about that credential in a dtrace_cred_t structure. These
1073  * can limit the consumer's breadth of visibility and what actions the
1074  * consumer may take.
1075  */
1076 #define	DTRACE_CRV_ALLPROC		0x01
1077 #define	DTRACE_CRV_KERNEL		0x02
1078 #define	DTRACE_CRV_ALLZONE		0x04
1079 
1080 #define	DTRACE_CRV_ALL		(DTRACE_CRV_ALLPROC | DTRACE_CRV_KERNEL | \
1081 	DTRACE_CRV_ALLZONE)
1082 
1083 #define	DTRACE_CRA_PROC				0x0001
1084 #define	DTRACE_CRA_PROC_CONTROL			0x0002
1085 #define	DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER	0x0004
1086 #define	DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE	0x0008
1087 #define	DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG	0x0010
1088 #define	DTRACE_CRA_KERNEL			0x0020
1089 #define	DTRACE_CRA_KERNEL_DESTRUCTIVE		0x0040
1090 
1091 #define	DTRACE_CRA_ALL		(DTRACE_CRA_PROC | \
1092 	DTRACE_CRA_PROC_CONTROL | \
1093 	DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER | \
1094 	DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE | \
1095 	DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG | \
1096 	DTRACE_CRA_KERNEL | \
1097 	DTRACE_CRA_KERNEL_DESTRUCTIVE)
1098 
1099 typedef struct dtrace_cred {
1100 	cred_t			*dcr_cred;
1101 	uint8_t			dcr_destructive;
1102 	uint8_t			dcr_visible;
1103 	uint16_t		dcr_action;
1104 } dtrace_cred_t;
1105 
1106 /*
1107  * DTrace Consumer State
1108  *
1109  * Each DTrace consumer has an associated dtrace_state structure that contains
1110  * its in-kernel DTrace state -- including options, credentials, statistics and
1111  * pointers to ECBs, buffers, speculations and formats.  A dtrace_state
1112  * structure is also allocated for anonymous enablings.  When anonymous state
1113  * is grabbed, the grabbing consumers dts_anon pointer is set to the grabbed
1114  * dtrace_state structure.
1115  */
1116 struct dtrace_state {
1117 	dev_t dts_dev;				/* device */
1118 	int dts_necbs;				/* total number of ECBs */
1119 	dtrace_ecb_t **dts_ecbs;		/* array of ECBs */
1120 	dtrace_epid_t dts_epid;			/* next EPID to allocate */
1121 	size_t dts_needed;			/* greatest needed space */
1122 	struct dtrace_state *dts_anon;		/* anon. state, if grabbed */
1123 	dtrace_activity_t dts_activity;		/* current activity */
1124 	dtrace_vstate_t dts_vstate;		/* variable state */
1125 	dtrace_buffer_t *dts_buffer;		/* principal buffer */
1126 	dtrace_buffer_t *dts_aggbuffer;		/* aggregation buffer */
1127 	dtrace_speculation_t *dts_speculations;	/* speculation array */
1128 	int dts_nspeculations;			/* number of speculations */
1129 	int dts_naggregations;			/* number of aggregations */
1130 	dtrace_aggregation_t **dts_aggregations; /* aggregation array */
1131 	vmem_t *dts_aggid_arena;		/* arena for aggregation IDs */
1132 	uint64_t dts_errors;			/* total number of errors */
1133 	uint32_t dts_speculations_busy;		/* number of spec. busy */
1134 	uint32_t dts_speculations_unavail;	/* number of spec unavail */
1135 	uint32_t dts_stkstroverflows;		/* stack string tab overflows */
1136 	uint32_t dts_dblerrors;			/* errors in ERROR probes */
1137 	uint32_t dts_reserve;			/* space reserved for END */
1138 	hrtime_t dts_laststatus;		/* time of last status */
1139 	cyclic_id_t dts_cleaner;		/* cleaning cyclic */
1140 	cyclic_id_t dts_deadman;		/* deadman cyclic */
1141 	hrtime_t dts_alive;			/* time last alive */
1142 	char dts_speculates;			/* boolean: has speculations */
1143 	char dts_destructive;			/* boolean: has dest. actions */
1144 	int dts_nformats;			/* number of formats */
1145 	char **dts_formats;			/* format string array */
1146 	dtrace_optval_t dts_options[DTRACEOPT_MAX]; /* options */
1147 	dtrace_cred_t dts_cred;			/* credentials */
1148 	size_t dts_nretained;			/* number of retained enabs */
1149 	int dts_getf;				/* number of getf() calls */
1150 };
1151 
1152 struct dtrace_provider {
1153 	dtrace_pattr_t dtpv_attr;		/* provider attributes */
1154 	dtrace_ppriv_t dtpv_priv;		/* provider privileges */
1155 	dtrace_pops_t dtpv_pops;		/* provider operations */
1156 	char *dtpv_name;			/* provider name */
1157 	void *dtpv_arg;				/* provider argument */
1158 	hrtime_t dtpv_defunct;			/* when made defunct */
1159 	struct dtrace_provider *dtpv_next;	/* next provider */
1160 };
1161 
1162 struct dtrace_meta {
1163 	dtrace_mops_t dtm_mops;			/* meta provider operations */
1164 	char *dtm_name;				/* meta provider name */
1165 	void *dtm_arg;				/* meta provider user arg */
1166 	uint64_t dtm_count;			/* no. of associated provs. */
1167 };
1168 
1169 /*
1170  * DTrace Enablings
1171  *
1172  * A dtrace_enabling structure is used to track a collection of ECB
1173  * descriptions -- before they have been turned into actual ECBs.  This is
1174  * created as a result of DOF processing, and is generally used to generate
1175  * ECBs immediately thereafter.  However, enablings are also generally
1176  * retained should the probes they describe be created at a later time; as
1177  * each new module or provider registers with the framework, the retained
1178  * enablings are reevaluated, with any new match resulting in new ECBs.  To
1179  * prevent probes from being matched more than once, the enabling tracks the
1180  * last probe generation matched, and only matches probes from subsequent
1181  * generations.
1182  */
1183 typedef struct dtrace_enabling {
1184 	dtrace_ecbdesc_t **dten_desc;		/* all ECB descriptions */
1185 	int dten_ndesc;				/* number of ECB descriptions */
1186 	int dten_maxdesc;			/* size of ECB array */
1187 	dtrace_vstate_t *dten_vstate;		/* associated variable state */
1188 	dtrace_genid_t dten_probegen;		/* matched probe generation */
1189 	dtrace_ecbdesc_t *dten_current;		/* current ECB description */
1190 	int dten_error;				/* current error value */
1191 	int dten_primed;			/* boolean: set if primed */
1192 	struct dtrace_enabling *dten_prev;	/* previous enabling */
1193 	struct dtrace_enabling *dten_next;	/* next enabling */
1194 } dtrace_enabling_t;
1195 
1196 /*
1197  * DTrace Anonymous Enablings
1198  *
1199  * Anonymous enablings are DTrace enablings that are not associated with a
1200  * controlling process, but rather derive their enabling from DOF stored as
1201  * properties in the dtrace.conf file.  If there is an anonymous enabling, a
1202  * DTrace consumer state and enabling are created on attach.  The state may be
1203  * subsequently grabbed by the first consumer specifying the "grabanon"
1204  * option.  As long as an anonymous DTrace enabling exists, dtrace(4D) will
1205  * refuse to unload.
1206  */
1207 typedef struct dtrace_anon {
1208 	dtrace_state_t *dta_state;		/* DTrace consumer state */
1209 	dtrace_enabling_t *dta_enabling;	/* pointer to enabling */
1210 	processorid_t dta_beganon;		/* which CPU BEGIN ran on */
1211 } dtrace_anon_t;
1212 
1213 /*
1214  * DTrace Error Debugging
1215  */
1216 #ifdef DEBUG
1217 #define	DTRACE_ERRDEBUG
1218 #endif
1219 
1220 #ifdef DTRACE_ERRDEBUG
1221 
1222 typedef struct dtrace_errhash {
1223 	const char	*dter_msg;	/* error message */
1224 	int		dter_count;	/* number of times seen */
1225 } dtrace_errhash_t;
1226 
1227 #define	DTRACE_ERRHASHSZ	256	/* must be > number of err msgs */
1228 
1229 #endif	/* DTRACE_ERRDEBUG */
1230 
1231 /*
1232  * DTrace Toxic Ranges
1233  *
1234  * DTrace supports safe loads from probe context; if the address turns out to
1235  * be invalid, a bit will be set by the kernel indicating that DTrace
1236  * encountered a memory error, and DTrace will propagate the error to the user
1237  * accordingly.  However, there may exist some regions of memory in which an
1238  * arbitrary load can change system state, and from which it is impossible to
1239  * recover from such a load after it has been attempted.  Examples of this may
1240  * include memory in which programmable I/O registers are mapped (for which a
1241  * read may have some implications for the device) or (in the specific case of
1242  * UltraSPARC-I and -II) the virtual address hole.  The platform is required
1243  * to make DTrace aware of these toxic ranges; DTrace will then check that
1244  * target addresses are not in a toxic range before attempting to issue a
1245  * safe load.
1246  */
1247 typedef struct dtrace_toxrange {
1248 	uintptr_t	dtt_base;		/* base of toxic range */
1249 	uintptr_t	dtt_limit;		/* limit of toxic range */
1250 } dtrace_toxrange_t;
1251 
1252 extern uint64_t dtrace_getarg(int, int);
1253 extern greg_t dtrace_getfp(void);
1254 extern int dtrace_getipl(void);
1255 extern uintptr_t dtrace_caller(int);
1256 extern uint32_t dtrace_cas32(uint32_t *, uint32_t, uint32_t);
1257 extern void *dtrace_casptr(void *, void *, void *);
1258 extern void dtrace_copyin(uintptr_t, uintptr_t, size_t, volatile uint16_t *);
1259 extern void dtrace_copyinstr(uintptr_t, uintptr_t, size_t, volatile uint16_t *);
1260 extern void dtrace_copyout(uintptr_t, uintptr_t, size_t, volatile uint16_t *);
1261 extern void dtrace_copyoutstr(uintptr_t, uintptr_t, size_t,
1262     volatile uint16_t *);
1263 extern void dtrace_getpcstack(pc_t *, int, int, uint32_t *);
1264 extern ulong_t dtrace_getreg(struct regs *, uint_t);
1265 extern void dtrace_setreg(struct regs *, uint_t, ulong_t);
1266 extern uint64_t dtrace_getvmreg(uint_t, volatile uint16_t *);
1267 extern int dtrace_getstackdepth(int);
1268 extern void dtrace_getupcstack(uint64_t *, int);
1269 extern void dtrace_getufpstack(uint64_t *, uint64_t *, int);
1270 extern int dtrace_getustackdepth(void);
1271 extern uintptr_t dtrace_fulword(void *);
1272 extern uint8_t dtrace_fuword8(void *);
1273 extern uint16_t dtrace_fuword16(void *);
1274 extern uint32_t dtrace_fuword32(void *);
1275 extern uint64_t dtrace_fuword64(void *);
1276 extern void dtrace_probe_error(dtrace_state_t *, dtrace_epid_t, int, int,
1277     int, uintptr_t);
1278 extern int dtrace_assfail(const char *, const char *, int);
1279 extern int dtrace_attached(void);
1280 extern hrtime_t dtrace_gethrestime();
1281 
1282 #ifdef __sparc
1283 extern void dtrace_flush_windows(void);
1284 extern void dtrace_flush_user_windows(void);
1285 extern uint_t dtrace_getotherwin(void);
1286 extern uint_t dtrace_getfprs(void);
1287 #else
1288 extern void dtrace_copy(uintptr_t, uintptr_t, size_t);
1289 extern void dtrace_copystr(uintptr_t, uintptr_t, size_t, volatile uint16_t *);
1290 #endif
1291 
1292 /*
1293  * DTrace Assertions
1294  *
1295  * DTrace calls ASSERT and VERIFY from probe context.  To assure that a failed
1296  * ASSERT or VERIFY does not induce a markedly more catastrophic failure (e.g.,
1297  * one from which a dump cannot be gleaned), DTrace must define its own ASSERT
1298  * and VERIFY macros to be ones that may safely be called from probe context.
1299  * This header file must thus be included by any DTrace component that calls
1300  * ASSERT and/or VERIFY from probe context, and _only_ by those components.
1301  * (The only exception to this is kernel debugging infrastructure at user-level
1302  * that doesn't depend on calling ASSERT.)
1303  */
1304 #undef ASSERT
1305 #undef VERIFY
1306 #define	VERIFY(EX)	((void)((EX) || \
1307 			dtrace_assfail(#EX, __FILE__, __LINE__)))
1308 #ifdef DEBUG
1309 #define	ASSERT(EX)	((void)((EX) || \
1310 			dtrace_assfail(#EX, __FILE__, __LINE__)))
1311 #else
1312 #define	ASSERT(X)	((void)0)
1313 #endif
1314 
1315 #ifdef	__cplusplus
1316 }
1317 #endif
1318 
1319 #endif /* _SYS_DTRACE_IMPL_H */
1320