AMD64 - Fix many compile-time warnings. int/ptr type mismatches, %llx, etc.
[dragonfly.git] / sys / kern / vfs_journal.c
CommitLineData
6ddb7618 1/*
f56dc967 2 * Copyright (c) 2004-2006 The DragonFly Project. All rights reserved.
6ddb7618
MD
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 *
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
0e9b9130 34 * $DragonFly: src/sys/kern/vfs_journal.c,v 1.33 2007/05/09 00:53:34 dillon Exp $
2281065e
MD
35 */
36/*
2281065e
MD
37 * The journaling protocol is intended to evolve into a two-way stream
38 * whereby transaction IDs can be acknowledged by the journaling target
39 * when the data has been committed to hard storage. Both implicit and
40 * explicit acknowledgement schemes will be supported, depending on the
41 * sophistication of the journaling stream, plus resynchronization and
42 * restart when a journaling stream is interrupted. This information will
43 * also be made available to journaling-aware filesystems to allow better
44 * management of their own physical storage synchronization mechanisms as
45 * well as to allow such filesystems to take direct advantage of the kernel's
46 * journaling layer so they don't have to roll their own.
47 *
82eaef15 48 * In addition, the worker thread will have access to much larger
2281065e
MD
49 * spooling areas then the memory buffer is able to provide by e.g.
50 * reserving swap space, in order to absorb potentially long interruptions
51 * of off-site journaling streams, and to prevent 'slow' off-site linkages
52 * from radically slowing down local filesystem operations.
53 *
54 * Because of the non-trivial algorithms the journaling system will be
55 * required to support, use of a worker thread is mandatory. Efficiencies
56 * are maintained by utilitizing the memory FIFO to batch transactions when
57 * possible, reducing the number of gratuitous thread switches and taking
58 * advantage of cpu caches through the use of shorter batched code paths
59 * rather then trying to do everything in the context of the process
82eaef15
MD
60 * originating the filesystem op. In the future the memory FIFO can be
61 * made per-cpu to remove BGL or other locking requirements.
6ddb7618 62 */
6ddb7618
MD
63#include <sys/param.h>
64#include <sys/systm.h>
65#include <sys/buf.h>
66#include <sys/conf.h>
67#include <sys/kernel.h>
82eaef15 68#include <sys/queue.h>
6ddb7618
MD
69#include <sys/lock.h>
70#include <sys/malloc.h>
71#include <sys/mount.h>
72#include <sys/unistd.h>
73#include <sys/vnode.h>
74#include <sys/poll.h>
2281065e 75#include <sys/mountctl.h>
b2f7ec6c 76#include <sys/journal.h>
2281065e 77#include <sys/file.h>
b2f7ec6c 78#include <sys/proc.h>
9578bde0 79#include <sys/msfbuf.h>
500b6a22
MD
80#include <sys/socket.h>
81#include <sys/socketvar.h>
6ddb7618
MD
82
83#include <machine/limits.h>
84
85#include <vm/vm.h>
86#include <vm/vm_object.h>
87#include <vm/vm_page.h>
88#include <vm/vm_pager.h>
89#include <vm/vnode_pager.h>
90
2281065e
MD
91#include <sys/file2.h>
92#include <sys/thread2.h>
93
432b8263
MD
94static void journal_wthread(void *info);
95static void journal_rthread(void *info);
82eaef15 96
f56dc967
MD
97static void *journal_reserve(struct journal *jo,
98 struct journal_rawrecbeg **rawpp,
99 int16_t streamid, int bytes);
82eaef15 100static void *journal_extend(struct journal *jo,
f56dc967
MD
101 struct journal_rawrecbeg **rawpp,
102 int truncbytes, int bytes, int *newstreamrecp);
103static void journal_abort(struct journal *jo,
104 struct journal_rawrecbeg **rawpp);
105static void journal_commit(struct journal *jo,
106 struct journal_rawrecbeg **rawpp,
107 int bytes, int closeout);
6ddb7618 108
2281065e 109
f56dc967
MD
110MALLOC_DEFINE(M_JOURNAL, "journal", "Journaling structures");
111MALLOC_DEFINE(M_JFIFO, "journal-fifo", "Journal FIFO");
2281065e 112
432b8263 113void
500b6a22
MD
114journal_create_threads(struct journal *jo)
115{
116 jo->flags &= ~(MC_JOURNAL_STOP_REQ | MC_JOURNAL_STOP_IMM);
117 jo->flags |= MC_JOURNAL_WACTIVE;
118 lwkt_create(journal_wthread, jo, NULL, &jo->wthread,
119 TDF_STOPREQ, -1, "journal w:%.*s", JIDMAX, jo->id);
120 lwkt_setpri(&jo->wthread, TDPRI_KERN_DAEMON);
121 lwkt_schedule(&jo->wthread);
122
123 if (jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) {
124 jo->flags |= MC_JOURNAL_RACTIVE;
125 lwkt_create(journal_rthread, jo, NULL, &jo->rthread,
126 TDF_STOPREQ, -1, "journal r:%.*s", JIDMAX, jo->id);
127 lwkt_setpri(&jo->rthread, TDPRI_KERN_DAEMON);
128 lwkt_schedule(&jo->rthread);
129 }
130}
131
f56dc967 132void
500b6a22
MD
133journal_destroy_threads(struct journal *jo, int flags)
134{
135 int wcount;
136
137 jo->flags |= MC_JOURNAL_STOP_REQ | (flags & MC_JOURNAL_STOP_IMM);
138 wakeup(&jo->fifo);
139 wcount = 0;
140 while (jo->flags & (MC_JOURNAL_WACTIVE | MC_JOURNAL_RACTIVE)) {
141 tsleep(jo, 0, "jwait", hz);
142 if (++wcount % 10 == 0) {
6ea70f76 143 kprintf("Warning: journal %s waiting for descriptors to close\n",
500b6a22
MD
144 jo->id);
145 }
146 }
147
148 /*
149 * XXX SMP - threads should move to cpu requesting the restart or
150 * termination before finishing up to properly interlock.
151 */
152 tsleep(jo, 0, "jwait", hz);
153 lwkt_free_thread(&jo->wthread);
154 if (jo->flags & MC_JOURNAL_WANT_FULLDUPLEX)
155 lwkt_free_thread(&jo->rthread);
156}
157
82eaef15
MD
158/*
159 * The per-journal worker thread is responsible for writing out the
160 * journal's FIFO to the target stream.
161 */
2281065e 162static void
432b8263 163journal_wthread(void *info)
2281065e
MD
164{
165 struct journal *jo = info;
82eaef15 166 struct journal_rawrecbeg *rawp;
2281065e 167 int error;
973c11b9
MD
168 size_t avail;
169 size_t bytes;
170 size_t res;
2281065e
MD
171
172 for (;;) {
82eaef15
MD
173 /*
174 * Calculate the number of bytes available to write. This buffer
175 * area may contain reserved records so we can't just write it out
176 * without further checks.
177 */
178 bytes = jo->fifo.windex - jo->fifo.rindex;
179
180 /*
181 * sleep if no bytes are available or if an incomplete record is
182 * encountered (it needs to be filled in before we can write it
183 * out), and skip any pad records that we encounter.
184 */
185 if (bytes == 0) {
2281065e
MD
186 if (jo->flags & MC_JOURNAL_STOP_REQ)
187 break;
82eaef15
MD
188 tsleep(&jo->fifo, 0, "jfifo", hz);
189 continue;
190 }
9578bde0
MD
191
192 /*
193 * Sleep if we can not go any further due to hitting an incomplete
194 * record. This case should occur rarely but may have to be better
195 * optimized XXX.
196 */
82eaef15
MD
197 rawp = (void *)(jo->fifo.membase + (jo->fifo.rindex & jo->fifo.mask));
198 if (rawp->begmagic == JREC_INCOMPLETEMAGIC) {
199 tsleep(&jo->fifo, 0, "jpad", hz);
200 continue;
201 }
9578bde0
MD
202
203 /*
204 * Skip any pad records. We do not write out pad records if we can
205 * help it.
9578bde0 206 */
82eaef15 207 if (rawp->streamid == JREC_STREAMID_PAD) {
3119bac5
MD
208 if ((jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) == 0) {
209 if (jo->fifo.rindex == jo->fifo.xindex) {
210 jo->fifo.xindex += (rawp->recsize + 15) & ~15;
211 jo->total_acked += (rawp->recsize + 15) & ~15;
212 }
213 }
82eaef15 214 jo->fifo.rindex += (rawp->recsize + 15) & ~15;
9578bde0
MD
215 jo->total_acked += bytes;
216 KKASSERT(jo->fifo.windex - jo->fifo.rindex >= 0);
82eaef15
MD
217 continue;
218 }
219
220 /*
9578bde0
MD
221 * 'bytes' is the amount of data that can potentially be written out.
222 * Calculate 'res', the amount of data that can actually be written
223 * out. res is bounded either by hitting the end of the physical
224 * memory buffer or by hitting an incomplete record. Incomplete
225 * records often occur due to the way the space reservation model
226 * works.
82eaef15
MD
227 */
228 res = 0;
229 avail = jo->fifo.size - (jo->fifo.rindex & jo->fifo.mask);
230 while (res < bytes && rawp->begmagic == JREC_BEGMAGIC) {
231 res += (rawp->recsize + 15) & ~15;
232 if (res >= avail) {
233 KKASSERT(res == avail);
234 break;
235 }
9578bde0 236 rawp = (void *)((char *)rawp + ((rawp->recsize + 15) & ~15));
2281065e 237 }
82eaef15
MD
238
239 /*
240 * Issue the write and deal with any errors or other conditions.
241 * For now assume blocking I/O. Since we are record-aware the
242 * code cannot yet handle partial writes.
243 *
3119bac5
MD
244 * We bump rindex prior to issuing the write to avoid racing
245 * the acknowledgement coming back (which could prevent the ack
246 * from bumping xindex). Restarts are always based on xindex so
247 * we do not try to undo the rindex if an error occurs.
248 *
82eaef15
MD
249 * XXX EWOULDBLOCK/NBIO
250 * XXX notification on failure
9578bde0 251 * XXX permanent verses temporary failures
82eaef15
MD
252 * XXX two-way acknowledgement stream in the return direction / xindex
253 */
254 bytes = res;
3119bac5 255 jo->fifo.rindex += bytes;
82eaef15 256 error = fp_write(jo->fp,
973c11b9
MD
257 jo->fifo.membase +
258 ((jo->fifo.rindex - bytes) & jo->fifo.mask),
e7440b28 259 bytes, &res, UIO_SYSSPACE);
2281065e 260 if (error) {
6ea70f76 261 kprintf("journal_thread(%s) write, error %d\n", jo->id, error);
82eaef15 262 /* XXX */
2281065e 263 } else {
82eaef15 264 KKASSERT(res == bytes);
82eaef15
MD
265 }
266
267 /*
432b8263
MD
268 * Advance rindex. If the journal stream is not full duplex we also
269 * advance xindex, otherwise the rjournal thread is responsible for
270 * advancing xindex.
82eaef15 271 */
3119bac5 272 if ((jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) == 0) {
432b8263 273 jo->fifo.xindex += bytes;
3119bac5
MD
274 jo->total_acked += bytes;
275 }
9578bde0 276 KKASSERT(jo->fifo.windex - jo->fifo.rindex >= 0);
432b8263
MD
277 if ((jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) == 0) {
278 if (jo->flags & MC_JOURNAL_WWAIT) {
279 jo->flags &= ~MC_JOURNAL_WWAIT; /* XXX hysteresis */
280 wakeup(&jo->fifo.windex);
281 }
282 }
283 }
500b6a22 284 fp_shutdown(jo->fp, SHUT_WR);
432b8263
MD
285 jo->flags &= ~MC_JOURNAL_WACTIVE;
286 wakeup(jo);
287 wakeup(&jo->fifo.windex);
288}
289
290/*
291 * A second per-journal worker thread is created for two-way journaling
292 * streams to deal with the return acknowledgement stream.
293 */
294static void
295journal_rthread(void *info)
296{
297 struct journal_rawrecbeg *rawp;
298 struct journal_ackrecord ack;
299 struct journal *jo = info;
300 int64_t transid;
301 int error;
973c11b9
MD
302 size_t count;
303 size_t bytes;
432b8263
MD
304
305 transid = 0;
306 error = 0;
307
308 for (;;) {
309 /*
310 * We have been asked to stop
311 */
312 if (jo->flags & MC_JOURNAL_STOP_REQ)
313 break;
314
315 /*
316 * If we have no active transaction id, get one from the return
317 * stream.
318 */
319 if (transid == 0) {
e7440b28
MD
320 error = fp_read(jo->fp, &ack, sizeof(ack), &count,
321 1, UIO_SYSSPACE);
3119bac5 322#if 0
6ea70f76 323 kprintf("fp_read ack error %d count %d\n", error, count);
3119bac5
MD
324#endif
325 if (error || count != sizeof(ack))
326 break;
432b8263 327 if (error) {
6ea70f76 328 kprintf("read error %d on receive stream\n", error);
432b8263
MD
329 break;
330 }
331 if (ack.rbeg.begmagic != JREC_BEGMAGIC ||
332 ack.rend.endmagic != JREC_ENDMAGIC
333 ) {
6ea70f76 334 kprintf("bad begmagic or endmagic on receive stream\n");
432b8263
MD
335 break;
336 }
337 transid = ack.rbeg.transid;
2281065e 338 }
432b8263
MD
339
340 /*
341 * Calculate the number of unacknowledged bytes. If there are no
342 * unacknowledged bytes then unsent data was acknowledged, report,
343 * sleep a bit, and loop in that case. This should not happen
344 * normally. The ack record is thrown away.
345 */
346 bytes = jo->fifo.rindex - jo->fifo.xindex;
347
348 if (bytes == 0) {
973c11b9
MD
349 kprintf("warning: unsent data acknowledged transid %08llx\n",
350 (long long)transid);
432b8263
MD
351 tsleep(&jo->fifo.xindex, 0, "jrseq", hz);
352 transid = 0;
353 continue;
354 }
355
356 /*
3119bac5 357 * Since rindex has advanced, the record pointed to by xindex
432b8263
MD
358 * must be a valid record.
359 */
360 rawp = (void *)(jo->fifo.membase + (jo->fifo.xindex & jo->fifo.mask));
361 KKASSERT(rawp->begmagic == JREC_BEGMAGIC);
362 KKASSERT(rawp->recsize <= bytes);
363
364 /*
365 * The target can acknowledge several records at once.
366 */
367 if (rawp->transid < transid) {
3119bac5 368#if 1
973c11b9
MD
369 kprintf("ackskip %08llx/%08llx\n",
370 (long long)rawp->transid,
371 (long long)transid);
3119bac5 372#endif
432b8263 373 jo->fifo.xindex += (rawp->recsize + 15) & ~15;
3119bac5 374 jo->total_acked += (rawp->recsize + 15) & ~15;
432b8263
MD
375 if (jo->flags & MC_JOURNAL_WWAIT) {
376 jo->flags &= ~MC_JOURNAL_WWAIT; /* XXX hysteresis */
377 wakeup(&jo->fifo.windex);
378 }
379 continue;
380 }
381 if (rawp->transid == transid) {
3119bac5 382#if 1
973c11b9
MD
383 kprintf("ackskip %08llx/%08llx\n",
384 (long long)rawp->transid,
385 (long long)transid);
3119bac5 386#endif
432b8263 387 jo->fifo.xindex += (rawp->recsize + 15) & ~15;
3119bac5 388 jo->total_acked += (rawp->recsize + 15) & ~15;
432b8263
MD
389 if (jo->flags & MC_JOURNAL_WWAIT) {
390 jo->flags &= ~MC_JOURNAL_WWAIT; /* XXX hysteresis */
391 wakeup(&jo->fifo.windex);
392 }
393 transid = 0;
394 continue;
395 }
973c11b9
MD
396 kprintf("warning: unsent data(2) acknowledged transid %08llx\n",
397 (long long)transid);
432b8263 398 transid = 0;
2281065e 399 }
432b8263 400 jo->flags &= ~MC_JOURNAL_RACTIVE;
2281065e
MD
401 wakeup(jo);
402 wakeup(&jo->fifo.windex);
403}
404
9578bde0
MD
405/*
406 * This builds a pad record which the journaling thread will skip over. Pad
407 * records are required when we are unable to reserve sufficient stream space
408 * due to insufficient space at the end of the physical memory fifo.
432b8263
MD
409 *
410 * Even though the record is not transmitted, a normal transid must be
411 * assigned to it so link recovery operations after a failure work properly.
9578bde0 412 */
88c28735 413static
82eaef15 414void
432b8263 415journal_build_pad(struct journal_rawrecbeg *rawp, int recsize, int64_t transid)
2281065e 416{
82eaef15
MD
417 struct journal_rawrecend *rendp;
418
419 KKASSERT((recsize & 15) == 0 && recsize >= 16);
420
82eaef15
MD
421 rawp->streamid = JREC_STREAMID_PAD;
422 rawp->recsize = recsize; /* must be 16-byte aligned */
432b8263 423 rawp->transid = transid;
82eaef15 424 /*
9b23f2eb 425 * WARNING, rendp may overlap rawp->transid. This is necessary to
35238fa5 426 * allow PAD records to fit in 16 bytes. Use cpu_ccfence() to
82eaef15
MD
427 * hopefully cause the compiler to not make any assumptions.
428 */
82eaef15
MD
429 rendp = (void *)((char *)rawp + rawp->recsize - sizeof(*rendp));
430 rendp->endmagic = JREC_ENDMAGIC;
431 rendp->check = 0;
432 rendp->recsize = rawp->recsize;
9578bde0
MD
433
434 /*
435 * Set the begin magic last. This is what will allow the journal
35238fa5
MD
436 * thread to write the record out. Use a store fence to prevent
437 * compiler and cpu reordering of the writes.
9578bde0 438 */
35238fa5 439 cpu_sfence();
9578bde0 440 rawp->begmagic = JREC_BEGMAGIC;
2281065e
MD
441}
442
82eaef15
MD
443/*
444 * Wake up the worker thread if the FIFO is more then half full or if
445 * someone is waiting for space to be freed up. Otherwise let the
446 * heartbeat deal with it. Being able to avoid waking up the worker
9578bde0 447 * is the key to the journal's cpu performance.
82eaef15
MD
448 */
449static __inline
2281065e 450void
82eaef15 451journal_commit_wakeup(struct journal *jo)
2281065e
MD
452{
453 int avail;
454
82eaef15
MD
455 avail = jo->fifo.size - (jo->fifo.windex - jo->fifo.xindex);
456 KKASSERT(avail >= 0);
457 if ((avail < (jo->fifo.size >> 1)) || (jo->flags & MC_JOURNAL_WWAIT))
458 wakeup(&jo->fifo);
459}
460
461/*
462 * Create a new BEGIN stream record with the specified streamid and the
463 * specified amount of payload space. *rawpp will be set to point to the
464 * base of the new stream record and a pointer to the base of the payload
465 * space will be returned. *rawpp does not need to be pre-NULLd prior to
432b8263 466 * making this call. The raw record header will be partially initialized.
82eaef15
MD
467 *
468 * A stream can be extended, aborted, or committed by other API calls
469 * below. This may result in a sequence of potentially disconnected
470 * stream records to be output to the journaling target. The first record
471 * (the one created by this function) will be marked JREC_STREAMCTL_BEGIN,
472 * while the last record on commit or abort will be marked JREC_STREAMCTL_END
473 * (and possibly also JREC_STREAMCTL_ABORTED). The last record could wind
474 * up being the same as the first, in which case the bits are all set in
475 * the first record.
476 *
477 * The stream record is created in an incomplete state by setting the begin
478 * magic to JREC_INCOMPLETEMAGIC. This prevents the worker thread from
479 * flushing the fifo past our record until we have finished populating it.
480 * Other threads can reserve and operate on their own space without stalling
481 * but the stream output will stall until we have completed operations. The
482 * memory FIFO is intended to be large enough to absorb such situations
483 * without stalling out other threads.
484 */
485static
486void *
487journal_reserve(struct journal *jo, struct journal_rawrecbeg **rawpp,
488 int16_t streamid, int bytes)
489{
490 struct journal_rawrecbeg *rawp;
491 int avail;
492 int availtoend;
493 int req;
494
495 /*
496 * Add header and trailer overheads to the passed payload. Note that
497 * the passed payload size need not be aligned in any way.
498 */
499 bytes += sizeof(struct journal_rawrecbeg);
500 bytes += sizeof(struct journal_rawrecend);
501
502 for (;;) {
503 /*
504 * First, check boundary conditions. If the request would wrap around
505 * we have to skip past the ending block and return to the beginning
506 * of the FIFO's buffer. Calculate 'req' which is the actual number
507 * of bytes being reserved, including wrap-around dead space.
508 *
88c28735
MD
509 * Neither 'bytes' or 'req' are aligned.
510 *
82eaef15
MD
511 * Note that availtoend is not truncated to avail and so cannot be
512 * used to determine whether the reservation is possible by itself.
513 * Also, since all fifo ops are 16-byte aligned, we can check
514 * the size before calculating the aligned size.
515 */
516 availtoend = jo->fifo.size - (jo->fifo.windex & jo->fifo.mask);
88c28735 517 KKASSERT((availtoend & 15) == 0);
82eaef15
MD
518 if (bytes > availtoend)
519 req = bytes + availtoend; /* add pad to end */
520 else
521 req = bytes;
522
523 /*
524 * Next calculate the total available space and see if it is
525 * sufficient. We cannot overwrite previously buffered data
526 * past xindex because otherwise we would not be able to restart
527 * a broken link at the target's last point of commit.
528 */
529 avail = jo->fifo.size - (jo->fifo.windex - jo->fifo.xindex);
530 KKASSERT(avail >= 0 && (avail & 15) == 0);
531
532 if (avail < req) {
533 /* XXX MC_JOURNAL_STOP_IMM */
2281065e 534 jo->flags |= MC_JOURNAL_WWAIT;
3119bac5 535 ++jo->fifostalls;
2281065e
MD
536 tsleep(&jo->fifo.windex, 0, "jwrite", 0);
537 continue;
538 }
82eaef15
MD
539
540 /*
541 * Create a pad record for any dead space and create an incomplete
542 * record for the live space, then return a pointer to the
543 * contiguous buffer space that was requested.
544 *
545 * NOTE: The worker thread will not flush past an incomplete
546 * record, so the reserved space can be filled in at-will. The
547 * journaling code must also be aware the reserved sections occuring
548 * after this one will also not be written out even if completed
549 * until this one is completed.
432b8263
MD
550 *
551 * The transaction id must accomodate real and potential pad creation.
82eaef15
MD
552 */
553 rawp = (void *)(jo->fifo.membase + (jo->fifo.windex & jo->fifo.mask));
554 if (req != bytes) {
432b8263
MD
555 journal_build_pad(rawp, availtoend, jo->transid);
556 ++jo->transid;
82eaef15
MD
557 rawp = (void *)jo->fifo.membase;
558 }
559 rawp->begmagic = JREC_INCOMPLETEMAGIC; /* updated by abort/commit */
560 rawp->recsize = bytes; /* (unaligned size) */
561 rawp->streamid = streamid | JREC_STREAMCTL_BEGIN;
432b8263
MD
562 rawp->transid = jo->transid;
563 jo->transid += 2;
82eaef15
MD
564
565 /*
566 * Issue a memory barrier to guarentee that the record data has been
567 * properly initialized before we advance the write index and return
568 * a pointer to the reserved record. Otherwise the worker thread
569 * could accidently run past us.
570 *
571 * Note that stream records are always 16-byte aligned.
572 */
35238fa5 573 cpu_sfence();
82eaef15
MD
574 jo->fifo.windex += (req + 15) & ~15;
575 *rawpp = rawp;
576 return(rawp + 1);
577 }
578 /* not reached */
579 *rawpp = NULL;
580 return(NULL);
581}
582
583/*
143c4f15
MD
584 * Attempt to extend the stream record by <bytes> worth of payload space.
585 *
586 * If it is possible to extend the existing stream record no truncation
587 * occurs and the record is extended as specified. A pointer to the
588 * truncation offset within the payload space is returned.
82eaef15 589 *
143c4f15
MD
590 * If it is not possible to do this the existing stream record is truncated
591 * and committed, and a new stream record of size <bytes> is created. A
592 * pointer to the base of the new stream record's payload space is returned.
82eaef15 593 *
143c4f15
MD
594 * *rawpp is set to the new reservation in the case of a new record but
595 * the caller cannot depend on a comparison with the old rawp to determine if
596 * this case occurs because we could end up using the same memory FIFO
597 * offset for the new stream record. Use *newstreamrecp instead.
82eaef15
MD
598 */
599static void *
600journal_extend(struct journal *jo, struct journal_rawrecbeg **rawpp,
601 int truncbytes, int bytes, int *newstreamrecp)
602{
603 struct journal_rawrecbeg *rawp;
604 int16_t streamid;
605 int availtoend;
606 int avail;
607 int osize;
608 int nsize;
609 int wbase;
610 void *rptr;
611
612 *newstreamrecp = 0;
613 rawp = *rawpp;
614 osize = (rawp->recsize + 15) & ~15;
615 nsize = (rawp->recsize + bytes + 15) & ~15;
616 wbase = (char *)rawp - jo->fifo.membase;
617
618 /*
143c4f15
MD
619 * If the aligned record size does not change we can trivially adjust
620 * the record size.
82eaef15
MD
621 */
622 if (nsize == osize) {
623 rawp->recsize += bytes;
143c4f15 624 return((char *)(rawp + 1) + truncbytes);
82eaef15
MD
625 }
626
627 /*
628 * If the fifo's write index hasn't been modified since we made the
629 * reservation and we do not hit any boundary conditions, we can
143c4f15 630 * trivially make the record smaller or larger.
82eaef15
MD
631 */
632 if ((jo->fifo.windex & jo->fifo.mask) == wbase + osize) {
633 availtoend = jo->fifo.size - wbase;
634 avail = jo->fifo.size - (jo->fifo.windex - jo->fifo.xindex) + osize;
635 KKASSERT((availtoend & 15) == 0);
636 KKASSERT((avail & 15) == 0);
637 if (nsize <= avail && nsize <= availtoend) {
638 jo->fifo.windex += nsize - osize;
639 rawp->recsize += bytes;
143c4f15 640 return((char *)(rawp + 1) + truncbytes);
82eaef15
MD
641 }
642 }
643
644 /*
645 * It was not possible to extend the buffer. Commit the current
646 * buffer and create a new one. We manually clear the BEGIN mark that
647 * journal_reserve() creates (because this is a continuing record, not
648 * the start of a new stream).
649 */
650 streamid = rawp->streamid & JREC_STREAMID_MASK;
651 journal_commit(jo, rawpp, truncbytes, 0);
652 rptr = journal_reserve(jo, rawpp, streamid, bytes);
653 rawp = *rawpp;
654 rawp->streamid &= ~JREC_STREAMCTL_BEGIN;
655 *newstreamrecp = 1;
656 return(rptr);
657}
658
659/*
660 * Abort a journal record. If the transaction record represents a stream
661 * BEGIN and we can reverse the fifo's write index we can simply reverse
662 * index the entire record, as if it were never reserved in the first place.
663 *
664 * Otherwise we set the JREC_STREAMCTL_ABORTED bit and commit the record
665 * with the payload truncated to 0 bytes.
666 */
667static void
668journal_abort(struct journal *jo, struct journal_rawrecbeg **rawpp)
669{
670 struct journal_rawrecbeg *rawp;
671 int osize;
672
673 rawp = *rawpp;
674 osize = (rawp->recsize + 15) & ~15;
675
676 if ((rawp->streamid & JREC_STREAMCTL_BEGIN) &&
677 (jo->fifo.windex & jo->fifo.mask) ==
678 (char *)rawp - jo->fifo.membase + osize)
679 {
680 jo->fifo.windex -= osize;
681 *rawpp = NULL;
682 } else {
683 rawp->streamid |= JREC_STREAMCTL_ABORTED;
684 journal_commit(jo, rawpp, 0, 1);
685 }
686}
687
688/*
689 * Commit a journal record and potentially truncate it to the specified
690 * number of payload bytes. If you do not want to truncate the record,
691 * simply pass -1 for the bytes parameter. Do not pass rawp->recsize, that
692 * field includes header and trailer and will not be correct. Note that
693 * passing 0 will truncate the entire data payload of the record.
694 *
695 * The logical stream is terminated by this function.
696 *
697 * If truncation occurs, and it is not possible to physically optimize the
698 * memory FIFO due to other threads having reserved space after ours,
699 * the remaining reserved space will be covered by a pad record.
700 */
701static void
702journal_commit(struct journal *jo, struct journal_rawrecbeg **rawpp,
703 int bytes, int closeout)
704{
705 struct journal_rawrecbeg *rawp;
706 struct journal_rawrecend *rendp;
707 int osize;
708 int nsize;
709
710 rawp = *rawpp;
711 *rawpp = NULL;
712
713 KKASSERT((char *)rawp >= jo->fifo.membase &&
714 (char *)rawp + rawp->recsize <= jo->fifo.membase + jo->fifo.size);
715 KKASSERT(((intptr_t)rawp & 15) == 0);
716
717 /*
88c28735 718 * Truncate the record if necessary. If the FIFO write index as still
82eaef15 719 * at the end of our record we can optimally backindex it. Otherwise
88c28735 720 * we have to insert a pad record to cover the dead space.
82eaef15
MD
721 *
722 * We calculate osize which is the 16-byte-aligned original recsize.
723 * We calculate nsize which is the 16-byte-aligned new recsize.
724 *
725 * Due to alignment issues or in case the passed truncation bytes is
88c28735
MD
726 * the same as the original payload, nsize may be equal to osize even
727 * if the committed bytes is less then the originally reserved bytes.
82eaef15
MD
728 */
729 if (bytes >= 0) {
730 KKASSERT(bytes >= 0 && bytes <= rawp->recsize - sizeof(struct journal_rawrecbeg) - sizeof(struct journal_rawrecend));
731 osize = (rawp->recsize + 15) & ~15;
732 rawp->recsize = bytes + sizeof(struct journal_rawrecbeg) +
733 sizeof(struct journal_rawrecend);
734 nsize = (rawp->recsize + 15) & ~15;
88c28735 735 KKASSERT(nsize <= osize);
82eaef15
MD
736 if (osize == nsize) {
737 /* do nothing */
738 } else if ((jo->fifo.windex & jo->fifo.mask) == (char *)rawp - jo->fifo.membase + osize) {
739 /* we are able to backindex the fifo */
740 jo->fifo.windex -= osize - nsize;
741 } else {
742 /* we cannot backindex the fifo, emplace a pad in the dead space */
432b8263
MD
743 journal_build_pad((void *)((char *)rawp + nsize), osize - nsize,
744 rawp->transid + 1);
82eaef15 745 }
2281065e 746 }
82eaef15
MD
747
748 /*
749 * Fill in the trailer. Note that unlike pad records, the trailer will
750 * never overlap the header.
751 */
752 rendp = (void *)((char *)rawp +
753 ((rawp->recsize + 15) & ~15) - sizeof(*rendp));
754 rendp->endmagic = JREC_ENDMAGIC;
755 rendp->recsize = rawp->recsize;
756 rendp->check = 0; /* XXX check word, disabled for now */
757
758 /*
759 * Fill in begmagic last. This will allow the worker thread to proceed.
760 * Use a memory barrier to guarentee write ordering. Mark the stream
761 * as terminated if closeout is set. This is the typical case.
762 */
763 if (closeout)
764 rawp->streamid |= JREC_STREAMCTL_END;
35238fa5 765 cpu_sfence(); /* memory and compiler barrier */
82eaef15
MD
766 rawp->begmagic = JREC_BEGMAGIC;
767
768 journal_commit_wakeup(jo);
769}
770
771/************************************************************************
772 * TRANSACTION SUPPORT ROUTINES *
773 ************************************************************************
774 *
775 * JRECORD_*() - routines to create subrecord transactions and embed them
776 * in the logical streams managed by the journal_*() routines.
777 */
778
82eaef15
MD
779/*
780 * Initialize the passed jrecord structure and start a new stream transaction
781 * by reserving an initial build space in the journal's memory FIFO.
782 */
f56dc967 783void
82eaef15
MD
784jrecord_init(struct journal *jo, struct jrecord *jrec, int16_t streamid)
785{
786 bzero(jrec, sizeof(*jrec));
787 jrec->jo = jo;
82eaef15
MD
788 jrec->streamid = streamid;
789 jrec->stream_residual = JREC_DEFAULTSIZE;
790 jrec->stream_reserved = jrec->stream_residual;
791 jrec->stream_ptr =
792 journal_reserve(jo, &jrec->rawp, streamid, jrec->stream_reserved);
793}
794
795/*
796 * Push a recursive record type. All pushes should have matching pops.
797 * The old parent is returned and the newly pushed record becomes the
798 * new parent. Note that the old parent's pointer may already be invalid
799 * or may become invalid if jrecord_write() had to build a new stream
800 * record, so the caller should not mess with the returned pointer in
801 * any way other then to save it.
802 */
82eaef15
MD
803struct journal_subrecord *
804jrecord_push(struct jrecord *jrec, int16_t rectype)
805{
806 struct journal_subrecord *save;
807
808 save = jrec->parent;
809 jrec->parent = jrecord_write(jrec, rectype|JMASK_NESTED, 0);
810 jrec->last = NULL;
811 KKASSERT(jrec->parent != NULL);
812 ++jrec->pushcount;
813 ++jrec->pushptrgood; /* cleared on flush */
814 return(save);
815}
816
817/*
818 * Pop a previously pushed sub-transaction. We must set JMASK_LAST
819 * on the last record written within the subtransaction. If the last
820 * record written is not accessible or if the subtransaction is empty,
821 * we must write out a pad record with JMASK_LAST set before popping.
822 *
823 * When popping a subtransaction the parent record's recsize field
824 * will be properly set. If the parent pointer is no longer valid
825 * (which can occur if the data has already been flushed out to the
826 * stream), the protocol spec allows us to leave it 0.
827 *
828 * The saved parent pointer which we restore may or may not be valid,
829 * and if not valid may or may not be NULL, depending on the value
830 * of pushptrgood.
831 */
f56dc967 832void
82eaef15
MD
833jrecord_pop(struct jrecord *jrec, struct journal_subrecord *save)
834{
835 struct journal_subrecord *last;
836
837 KKASSERT(jrec->pushcount > 0);
838 KKASSERT(jrec->residual == 0);
839
840 /*
841 * Set JMASK_LAST on the last record we wrote at the current
842 * level. If last is NULL we either no longer have access to the
843 * record or the subtransaction was empty and we must write out a pad
844 * record.
845 */
846 if ((last = jrec->last) == NULL) {
847 jrecord_write(jrec, JLEAF_PAD|JMASK_LAST, 0);
848 last = jrec->last; /* reload after possible flush */
849 } else {
850 last->rectype |= JMASK_LAST;
851 }
852
853 /*
854 * pushptrgood tells us how many levels of parent record pointers
855 * are valid. The jrec only stores the current parent record pointer
856 * (and it is only valid if pushptrgood != 0). The higher level parent
857 * record pointers are saved by the routines calling jrecord_push() and
858 * jrecord_pop(). These pointers may become stale and we determine
859 * that fact by tracking the count of valid parent pointers with
860 * pushptrgood. Pointers become invalid when their related stream
861 * record gets pushed out.
862 *
b2f7ec6c
MD
863 * If no pointer is available (the data has already been pushed out),
864 * then no fixup of e.g. the length field is possible for non-leaf
865 * nodes. The protocol allows for this situation by placing a larger
866 * burden on the program scanning the stream on the other end.
867 *
82eaef15
MD
868 * [parentA]
869 * [node X]
870 * [parentB]
871 * [node Y]
872 * [node Z]
873 * (pop B) see NOTE B
874 * (pop A) see NOTE A
875 *
876 * NOTE B: This pop sets LAST in node Z if the node is still accessible,
877 * else a PAD record is appended and LAST is set in that.
878 *
879 * This pop sets the record size in parentB if parentB is still
880 * accessible, else the record size is left 0 (the scanner must
881 * deal with that).
882 *
883 * This pop sets the new 'last' record to parentB, the pointer
884 * to which may or may not still be accessible.
885 *
886 * NOTE A: This pop sets LAST in parentB if the node is still accessible,
887 * else a PAD record is appended and LAST is set in that.
888 *
889 * This pop sets the record size in parentA if parentA is still
890 * accessible, else the record size is left 0 (the scanner must
891 * deal with that).
892 *
893 * This pop sets the new 'last' record to parentA, the pointer
894 * to which may or may not still be accessible.
895 *
896 * Also note that the last record in the stream transaction, which in
897 * the above example is parentA, does not currently have the LAST bit
898 * set.
899 *
900 * The current parent becomes the last record relative to the
901 * saved parent passed into us. It's validity is based on
902 * whether pushptrgood is non-zero prior to decrementing. The saved
903 * parent becomes the new parent, and its validity is based on whether
904 * pushptrgood is non-zero after decrementing.
905 *
906 * The old jrec->parent may be NULL if it is no longer accessible.
907 * If pushptrgood is non-zero, however, it is guarenteed to not
908 * be NULL (since no flush occured).
909 */
910 jrec->last = jrec->parent;
911 --jrec->pushcount;
912 if (jrec->pushptrgood) {
913 KKASSERT(jrec->last != NULL && last != NULL);
914 if (--jrec->pushptrgood == 0) {
915 jrec->parent = NULL; /* 'save' contains garbage or NULL */
916 } else {
917 KKASSERT(save != NULL);
918 jrec->parent = save; /* 'save' must not be NULL */
919 }
920
921 /*
922 * Set the record size in the old parent. 'last' still points to
923 * the original last record in the subtransaction being popped,
924 * jrec->last points to the old parent (which became the last
925 * record relative to the new parent being popped into).
926 */
927 jrec->last->recsize = (char *)last + last->recsize - (char *)jrec->last;
928 } else {
929 jrec->parent = NULL;
930 KKASSERT(jrec->last == NULL);
931 }
932}
933
934/*
b2f7ec6c
MD
935 * Write out a leaf record, including associated data.
936 */
b2f7ec6c
MD
937void
938jrecord_leaf(struct jrecord *jrec, int16_t rectype, void *ptr, int bytes)
939{
940 jrecord_write(jrec, rectype, bytes);
941 jrecord_data(jrec, ptr, bytes);
b2f7ec6c
MD
942}
943
944/*
82eaef15
MD
945 * Write a leaf record out and return a pointer to its base. The leaf
946 * record may contain potentially megabytes of data which is supplied
947 * in jrecord_data() calls. The exact amount must be specified in this
948 * call.
b2f7ec6c
MD
949 *
950 * THE RETURNED SUBRECORD POINTER IS ONLY VALID IMMEDIATELY AFTER THE
951 * CALL AND MAY BECOME INVALID AT ANY TIME. ONLY THE PUSH/POP CODE SHOULD
952 * USE THE RETURN VALUE.
82eaef15 953 */
82eaef15
MD
954struct journal_subrecord *
955jrecord_write(struct jrecord *jrec, int16_t rectype, int bytes)
956{
957 struct journal_subrecord *last;
958 int pusheditout;
959
960 /*
961 * Try to catch some obvious errors. Nesting records must specify a
962 * size of 0, and there should be no left-overs from previous operations
963 * (such as incomplete data writeouts).
964 */
965 KKASSERT(bytes == 0 || (rectype & JMASK_NESTED) == 0);
966 KKASSERT(jrec->residual == 0);
967
968 /*
969 * Check to see if the current stream record has enough room for
970 * the new subrecord header. If it doesn't we extend the current
971 * stream record.
972 *
973 * This may have the side effect of pushing out the current stream record
974 * and creating a new one. We must adjust our stream tracking fields
975 * accordingly.
976 */
977 if (jrec->stream_residual < sizeof(struct journal_subrecord)) {
978 jrec->stream_ptr = journal_extend(jrec->jo, &jrec->rawp,
979 jrec->stream_reserved - jrec->stream_residual,
980 JREC_DEFAULTSIZE, &pusheditout);
981 if (pusheditout) {
143c4f15
MD
982 /*
983 * If a pushout occured, the pushed out stream record was
984 * truncated as specified and the new record is exactly the
985 * extension size specified.
986 */
82eaef15
MD
987 jrec->stream_reserved = JREC_DEFAULTSIZE;
988 jrec->stream_residual = JREC_DEFAULTSIZE;
989 jrec->parent = NULL; /* no longer accessible */
990 jrec->pushptrgood = 0; /* restored parents in pops no good */
991 } else {
143c4f15
MD
992 /*
993 * If no pushout occured the stream record is NOT truncated and
994 * IS extended.
995 */
82eaef15
MD
996 jrec->stream_reserved += JREC_DEFAULTSIZE;
997 jrec->stream_residual += JREC_DEFAULTSIZE;
998 }
999 }
1000 last = (void *)jrec->stream_ptr;
1001 last->rectype = rectype;
1002 last->reserved = 0;
b7ef558f
MD
1003
1004 /*
1005 * We may not know the record size for recursive records and the
1006 * header may become unavailable due to limited FIFO space. Write
1007 * -1 to indicate this special case.
1008 */
1009 if ((rectype & JMASK_NESTED) && bytes == 0)
1010 last->recsize = -1;
1011 else
1012 last->recsize = sizeof(struct journal_subrecord) + bytes;
82eaef15
MD
1013 jrec->last = last;
1014 jrec->residual = bytes; /* remaining data to be posted */
1015 jrec->residual_align = -bytes & 7; /* post-data alignment required */
143c4f15
MD
1016 jrec->stream_ptr += sizeof(*last); /* current write pointer */
1017 jrec->stream_residual -= sizeof(*last); /* space remaining in stream */
82eaef15
MD
1018 return(last);
1019}
1020
1021/*
1022 * Write out the data associated with a leaf record. Any number of calls
1023 * to this routine may be made as long as the byte count adds up to the
1024 * amount originally specified in jrecord_write().
1025 *
1026 * The act of writing out the leaf data may result in numerous stream records
1027 * being pushed out. Callers should be aware that even the associated
1028 * subrecord header may become inaccessible due to stream record pushouts.
1029 */
f56dc967 1030void
82eaef15
MD
1031jrecord_data(struct jrecord *jrec, const void *buf, int bytes)
1032{
1033 int pusheditout;
1034 int extsize;
1035
1036 KKASSERT(bytes >= 0 && bytes <= jrec->residual);
1037
1038 /*
1039 * Push out stream records as long as there is insufficient room to hold
1040 * the remaining data.
1041 */
1042 while (jrec->stream_residual < bytes) {
1043 /*
1044 * Fill in any remaining space in the current stream record.
1045 */
1046 bcopy(buf, jrec->stream_ptr, jrec->stream_residual);
1047 buf = (const char *)buf + jrec->stream_residual;
1048 bytes -= jrec->stream_residual;
1049 /*jrec->stream_ptr += jrec->stream_residual;*/
82eaef15 1050 jrec->residual -= jrec->stream_residual;
9578bde0 1051 jrec->stream_residual = 0;
82eaef15
MD
1052
1053 /*
1054 * Try to extend the current stream record, but no more then 1/4
1055 * the size of the FIFO.
1056 */
1057 extsize = jrec->jo->fifo.size >> 2;
1058 if (extsize > bytes)
1059 extsize = (bytes + 15) & ~15;
1060
1061 jrec->stream_ptr = journal_extend(jrec->jo, &jrec->rawp,
1062 jrec->stream_reserved - jrec->stream_residual,
1063 extsize, &pusheditout);
1064 if (pusheditout) {
1065 jrec->stream_reserved = extsize;
1066 jrec->stream_residual = extsize;
1067 jrec->parent = NULL; /* no longer accessible */
1068 jrec->last = NULL; /* no longer accessible */
1069 jrec->pushptrgood = 0; /* restored parents in pops no good */
1070 } else {
1071 jrec->stream_reserved += extsize;
1072 jrec->stream_residual += extsize;
1073 }
1074 }
1075
1076 /*
1077 * Push out any remaining bytes into the current stream record.
1078 */
1079 if (bytes) {
1080 bcopy(buf, jrec->stream_ptr, bytes);
1081 jrec->stream_ptr += bytes;
1082 jrec->stream_residual -= bytes;
1083 jrec->residual -= bytes;
1084 }
1085
1086 /*
1087 * Handle data alignment requirements for the subrecord. Because the
1088 * stream record's data space is more strictly aligned, it must already
1089 * have sufficient space to hold any subrecord alignment slop.
1090 */
1091 if (jrec->residual == 0 && jrec->residual_align) {
1092 KKASSERT(jrec->residual_align <= jrec->stream_residual);
1093 bzero(jrec->stream_ptr, jrec->residual_align);
1094 jrec->stream_ptr += jrec->residual_align;
1095 jrec->stream_residual -= jrec->residual_align;
1096 jrec->residual_align = 0;
1097 }
1098}
1099
1100/*
9578bde0
MD
1101 * We are finished with the transaction. This closes the transaction created
1102 * by jrecord_init().
1103 *
1104 * NOTE: If abortit is not set then we must be at the top level with no
1105 * residual subrecord data left to output.
1106 *
1107 * If abortit is set then we can be in any state, all pushes will be
1108 * popped and it is ok for there to be residual data. This works
1109 * because the virtual stream itself is truncated. Scanners must deal
1110 * with this situation.
82eaef15
MD
1111 *
1112 * The stream record will be committed or aborted as specified and jrecord
1113 * resources will be cleaned up.
1114 */
f56dc967 1115void
82eaef15
MD
1116jrecord_done(struct jrecord *jrec, int abortit)
1117{
1118 KKASSERT(jrec->rawp != NULL);
1119
1120 if (abortit) {
1121 journal_abort(jrec->jo, &jrec->rawp);
1122 } else {
1123 KKASSERT(jrec->pushcount == 0 && jrec->residual == 0);
1124 journal_commit(jrec->jo, &jrec->rawp,
1125 jrec->stream_reserved - jrec->stream_residual, 1);
1126 }
1127
1128 /*
1129 * jrec should not be used beyond this point without another init,
1130 * but clean up some fields to ensure that we panic if it is.
1131 *
1132 * Note that jrec->rawp is NULLd out by journal_abort/journal_commit.
1133 */
1134 jrec->jo = NULL;
1135 jrec->stream_ptr = NULL;
1136}
1137
1138/************************************************************************
b2f7ec6c 1139 * LOW LEVEL RECORD SUPPORT ROUTINES *
82eaef15
MD
1140 ************************************************************************
1141 *
b2f7ec6c
MD
1142 * These routine create low level recursive and leaf subrecords representing
1143 * common filesystem structures.
82eaef15
MD
1144 */
1145
b2f7ec6c
MD
1146/*
1147 * Write out a filename path relative to the base of the mount point.
1148 * rectype is typically JLEAF_PATH{1,2,3,4}.
1149 */
f56dc967 1150void
82eaef15
MD
1151jrecord_write_path(struct jrecord *jrec, int16_t rectype, struct namecache *ncp)
1152{
b2f7ec6c
MD
1153 char buf[64]; /* local buffer if it fits, else malloced */
1154 char *base;
1155 int pathlen;
1156 int index;
1157 struct namecache *scan;
1158
1159 /*
1160 * Pass 1 - figure out the number of bytes required. Include terminating
1161 * \0 on last element and '/' separator on other elements.
28623bf9
MD
1162 *
1163 * The namecache topology terminates at the root of the filesystem
1164 * (the normal lookup code would then continue by using the mount
1165 * structure to figure out what it was mounted on).
b2f7ec6c
MD
1166 */
1167again:
1168 pathlen = 0;
28623bf9 1169 for (scan = ncp; scan; scan = scan->nc_parent) {
78552ca0
MD
1170 if (scan->nc_nlen > 0)
1171 pathlen += scan->nc_nlen + 1;
b2f7ec6c
MD
1172 }
1173
1174 if (pathlen <= sizeof(buf))
1175 base = buf;
1176 else
efda3bd0 1177 base = kmalloc(pathlen, M_TEMP, M_INTWAIT);
b2f7ec6c
MD
1178
1179 /*
1180 * Pass 2 - generate the path buffer
1181 */
1182 index = pathlen;
28623bf9 1183 for (scan = ncp; scan; scan = scan->nc_parent) {
78552ca0
MD
1184 if (scan->nc_nlen == 0)
1185 continue;
b2f7ec6c
MD
1186 if (scan->nc_nlen >= index) {
1187 if (base != buf)
efda3bd0 1188 kfree(base, M_TEMP);
b2f7ec6c
MD
1189 goto again;
1190 }
1191 if (index == pathlen)
1192 base[--index] = 0;
1193 else
1194 base[--index] = '/';
1195 index -= scan->nc_nlen;
1196 bcopy(scan->nc_name, base + index, scan->nc_nlen);
1197 }
1198 jrecord_leaf(jrec, rectype, base + index, pathlen - index);
1199 if (base != buf)
efda3bd0 1200 kfree(base, M_TEMP);
82eaef15
MD
1201}
1202
b2f7ec6c
MD
1203/*
1204 * Write out a file attribute structure. While somewhat inefficient, using
1205 * a recursive data structure is the most portable and extensible way.
1206 */
f56dc967 1207void
82eaef15
MD
1208jrecord_write_vattr(struct jrecord *jrec, struct vattr *vat)
1209{
b2f7ec6c
MD
1210 void *save;
1211
1212 save = jrecord_push(jrec, JTYPE_VATTR);
1213 if (vat->va_type != VNON)
432b8263 1214 jrecord_leaf(jrec, JLEAF_VTYPE, &vat->va_type, sizeof(vat->va_type));
d0887c34 1215 if (vat->va_mode != (mode_t)VNOVAL)
432b8263 1216 jrecord_leaf(jrec, JLEAF_MODES, &vat->va_mode, sizeof(vat->va_mode));
b2f7ec6c
MD
1217 if (vat->va_nlink != VNOVAL)
1218 jrecord_leaf(jrec, JLEAF_NLINK, &vat->va_nlink, sizeof(vat->va_nlink));
1219 if (vat->va_uid != VNOVAL)
1220 jrecord_leaf(jrec, JLEAF_UID, &vat->va_uid, sizeof(vat->va_uid));
1221 if (vat->va_gid != VNOVAL)
1222 jrecord_leaf(jrec, JLEAF_GID, &vat->va_gid, sizeof(vat->va_gid));
1223 if (vat->va_fsid != VNOVAL)
1224 jrecord_leaf(jrec, JLEAF_FSID, &vat->va_fsid, sizeof(vat->va_fsid));
1225 if (vat->va_fileid != VNOVAL)
1226 jrecord_leaf(jrec, JLEAF_INUM, &vat->va_fileid, sizeof(vat->va_fileid));
1227 if (vat->va_size != VNOVAL)
1228 jrecord_leaf(jrec, JLEAF_SIZE, &vat->va_size, sizeof(vat->va_size));
1229 if (vat->va_atime.tv_sec != VNOVAL)
1230 jrecord_leaf(jrec, JLEAF_ATIME, &vat->va_atime, sizeof(vat->va_atime));
1231 if (vat->va_mtime.tv_sec != VNOVAL)
1232 jrecord_leaf(jrec, JLEAF_MTIME, &vat->va_mtime, sizeof(vat->va_mtime));
1233 if (vat->va_ctime.tv_sec != VNOVAL)
1234 jrecord_leaf(jrec, JLEAF_CTIME, &vat->va_ctime, sizeof(vat->va_ctime));
1235 if (vat->va_gen != VNOVAL)
1236 jrecord_leaf(jrec, JLEAF_GEN, &vat->va_gen, sizeof(vat->va_gen));
1237 if (vat->va_flags != VNOVAL)
1238 jrecord_leaf(jrec, JLEAF_FLAGS, &vat->va_flags, sizeof(vat->va_flags));
0e9b9130
MD
1239 if (vat->va_rmajor != VNOVAL) {
1240 udev_t rdev = makeudev(vat->va_rmajor, vat->va_rminor);
1241 jrecord_leaf(jrec, JLEAF_UDEV, &rdev, sizeof(rdev));
1242 jrecord_leaf(jrec, JLEAF_UMAJOR, &vat->va_rmajor, sizeof(vat->va_rmajor));
1243 jrecord_leaf(jrec, JLEAF_UMINOR, &vat->va_rminor, sizeof(vat->va_rminor));
1244 }
b2f7ec6c
MD
1245#if 0
1246 if (vat->va_filerev != VNOVAL)
1247 jrecord_leaf(jrec, JLEAF_FILEREV, &vat->va_filerev, sizeof(vat->va_filerev));
1248#endif
1249 jrecord_pop(jrec, save);
b2f7ec6c
MD
1250}
1251
1252/*
1253 * Write out the creds used to issue a file operation. If a process is
1254 * available write out additional tracking information related to the
1255 * process.
1256 *
1257 * XXX additional tracking info
1258 * XXX tty line info
1259 */
f56dc967 1260void
b2f7ec6c
MD
1261jrecord_write_cred(struct jrecord *jrec, struct thread *td, struct ucred *cred)
1262{
1263 void *save;
1264 struct proc *p;
1265
1266 save = jrecord_push(jrec, JTYPE_CRED);
1267 jrecord_leaf(jrec, JLEAF_UID, &cred->cr_uid, sizeof(cred->cr_uid));
1268 jrecord_leaf(jrec, JLEAF_GID, &cred->cr_gid, sizeof(cred->cr_gid));
1269 if (td && (p = td->td_proc) != NULL) {
1270 jrecord_leaf(jrec, JLEAF_PID, &p->p_pid, sizeof(p->p_pid));
1271 jrecord_leaf(jrec, JLEAF_COMM, p->p_comm, sizeof(p->p_comm));
1272 }
1273 jrecord_pop(jrec, save);
b2f7ec6c
MD
1274}
1275
1276/*
1277 * Write out information required to identify a vnode
143c4f15
MD
1278 *
1279 * XXX this needs work. We should write out the inode number as well,
1280 * and in fact avoid writing out the file path for seqential writes
1281 * occuring within e.g. a certain period of time.
b2f7ec6c 1282 */
f56dc967 1283void
b2f7ec6c
MD
1284jrecord_write_vnode_ref(struct jrecord *jrec, struct vnode *vp)
1285{
143c4f15
MD
1286 struct namecache *ncp;
1287
1288 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
1289 if ((ncp->nc_flag & (NCF_UNRESOLVED|NCF_DESTROYED)) == 0)
1290 break;
1291 }
1292 if (ncp)
1293 jrecord_write_path(jrec, JLEAF_PATH_REF, ncp);
1294}
1295
f56dc967 1296void
f4659a6c
MD
1297jrecord_write_vnode_link(struct jrecord *jrec, struct vnode *vp,
1298 struct namecache *notncp)
1299{
1300 struct namecache *ncp;
1301
1302 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
1303 if (ncp == notncp)
1304 continue;
1305 if ((ncp->nc_flag & (NCF_UNRESOLVED|NCF_DESTROYED)) == 0)
1306 break;
1307 }
1308 if (ncp)
1309 jrecord_write_path(jrec, JLEAF_PATH_REF, ncp);
1310}
1311
143c4f15
MD
1312/*
1313 * Write out the data represented by a pagelist
1314 */
f56dc967 1315void
143c4f15
MD
1316jrecord_write_pagelist(struct jrecord *jrec, int16_t rectype,
1317 struct vm_page **pglist, int *rtvals, int pgcount,
1318 off_t offset)
1319{
1320 struct msf_buf *msf;
1321 int error;
1322 int b;
1323 int i;
1324
1325 i = 0;
1326 while (i < pgcount) {
1327 /*
1328 * Find the next valid section. Skip any invalid elements
1329 */
1330 if (rtvals[i] != VM_PAGER_OK) {
1331 ++i;
1332 offset += PAGE_SIZE;
1333 continue;
1334 }
1335
1336 /*
1337 * Figure out how big the valid section is, capping I/O at what the
1338 * MSFBUF can represent.
1339 */
1340 b = i;
1341 while (i < pgcount && i - b != XIO_INTERNAL_PAGES &&
1342 rtvals[i] == VM_PAGER_OK
1343 ) {
1344 ++i;
1345 }
1346
1347 /*
1348 * And write it out.
1349 */
1350 if (i - b) {
1351 error = msf_map_pagelist(&msf, pglist + b, i - b, 0);
1352 if (error == 0) {
6ea70f76 1353 kprintf("RECORD PUTPAGES %d\n", msf_buf_bytes(msf));
143c4f15
MD
1354 jrecord_leaf(jrec, JLEAF_SEEKPOS, &offset, sizeof(offset));
1355 jrecord_leaf(jrec, rectype,
1356 msf_buf_kva(msf), msf_buf_bytes(msf));
1357 msf_buf_free(msf);
1358 } else {
6ea70f76 1359 kprintf("jrecord_write_pagelist: mapping failure\n");
143c4f15
MD
1360 }
1361 offset += (off_t)(i - b) << PAGE_SHIFT;
1362 }
1363 }
b2f7ec6c
MD
1364}
1365
1366/*
9578bde0 1367 * Write out the data represented by a UIO.
b2f7ec6c 1368 */
9578bde0
MD
1369struct jwuio_info {
1370 struct jrecord *jrec;
1371 int16_t rectype;
1372};
1373
1374static int jrecord_write_uio_callback(void *info, char *buf, int bytes);
1375
f56dc967 1376void
b2f7ec6c
MD
1377jrecord_write_uio(struct jrecord *jrec, int16_t rectype, struct uio *uio)
1378{
9578bde0
MD
1379 struct jwuio_info info = { jrec, rectype };
1380 int error;
1381
143c4f15
MD
1382 if (uio->uio_segflg != UIO_NOCOPY) {
1383 jrecord_leaf(jrec, JLEAF_SEEKPOS, &uio->uio_offset,
1384 sizeof(uio->uio_offset));
1385 error = msf_uio_iterate(uio, jrecord_write_uio_callback, &info);
1386 if (error)
6ea70f76 1387 kprintf("XXX warning uio iterate failed %d\n", error);
143c4f15 1388 }
9578bde0
MD
1389}
1390
1391static int
1392jrecord_write_uio_callback(void *info_arg, char *buf, int bytes)
1393{
1394 struct jwuio_info *info = info_arg;
1395
9578bde0
MD
1396 jrecord_leaf(info->jrec, info->rectype, buf, bytes);
1397 return(0);
2281065e
MD
1398}
1399
f56dc967 1400void
26e603ed
MD
1401jrecord_file_data(struct jrecord *jrec, struct vnode *vp,
1402 off_t off, off_t bytes)
1403{
1404 const int bufsize = 8192;
1405 char *buf;
1406 int error;
1407 int n;
1408
efda3bd0 1409 buf = kmalloc(bufsize, M_JOURNAL, M_WAITOK);
26e603ed
MD
1410 jrecord_leaf(jrec, JLEAF_SEEKPOS, &off, sizeof(off));
1411 while (bytes) {
1412 n = (bytes > bufsize) ? bufsize : (int)bytes;
1413 error = vn_rdwr(UIO_READ, vp, buf, n, off, UIO_SYSSPACE, IO_NODELOCKED,
87de5057 1414 proc0.p_ucred, NULL);
26e603ed
MD
1415 if (error) {
1416 jrecord_leaf(jrec, JLEAF_ERROR, &error, sizeof(error));
1417 break;
1418 }
1419 jrecord_leaf(jrec, JLEAF_FILEDATA, buf, n);
1420 bytes -= n;
1421 off += n;
1422 }
efda3bd0 1423 kfree(buf, M_JOURNAL);
26e603ed
MD
1424}
1425