3 * THE REGENTS OF THE UNIVERSITY OF MICHIGAN
6 * Permission is granted to use, copy, create derivative works
7 * and redistribute this software and such derivative works
8 * for any purpose, so long as the name of The University of
9 * Michigan is not used in any advertising or publicity
10 * pertaining to the use of distribution of this software
11 * without specific, written prior authorization. If the
12 * above copyright notice or any other identification of the
13 * University of Michigan is included in any copy of any
14 * portion of this software, then the disclaimer below must
17 * THIS SOFTWARE IS PROVIDED AS IS, WITHOUT REPRESENTATION
18 * FROM THE UNIVERSITY OF MICHIGAN AS TO ITS FITNESS FOR ANY
19 * PURPOSE, AND WITHOUT WARRANTY BY THE UNIVERSITY O
20 * MICHIGAN OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING
21 * WITHOUT LIMITATION THE IMPLIED WARRANTIES OF
22 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. THE
23 * REGENTS OF THE UNIVERSITY OF MICHIGAN SHALL NOT BE LIABLE
24 * FOR ANY DAMAGES, INCLUDING SPECIAL, INDIRECT, INCIDENTAL, OR
25 * CONSEQUENTIAL DAMAGES, WITH RESPECT TO ANY CLAIM ARISING
26 * OUT OF OR IN CONNECTION WITH THE USE OF THE SOFTWARE, EVEN
27 * IF IT HAS BEEN OR IS HEREAFTER ADVISED OF THE POSSIBILITY OF
32 * Portions Copyright (c) 2008
33 * The Linux Box Corporation
36 * Permission is granted to use, copy, create derivative works
37 * and redistribute this software and such derivative works
38 * for any purpose, so long as the name of the Linux Box
39 * Corporation is not used in any advertising or publicity
40 * pertaining to the use or distribution of this software
41 * without specific, written prior authorization. If the
42 * above copyright notice or any other identification of the
43 * Linux Box Corporation is included in any copy of any
44 * portion of this software, then the disclaimer below must
47 * This software is provided as is, without representation
48 * from the Linux Box Corporation as to its fitness for any
49 * purpose, and without warranty by the Linux Box Corporation
50 * of any kind, either express or implied, including
51 * without limitation the implied warranties of
52 * merchantability and fitness for a particular purpose. The
53 * Linux Box Corporation shall not be liable for any damages,
54 * including special, indirect, incidental, or consequential
55 * damages, with respect to any claim arising out of or in
56 * connection with the use of the software, even if it has been
57 * or is hereafter advised of the possibility of such damages.
61 #include <afsconfig.h>
62 #include "afs/param.h"
63 #if defined(AFS_CACHE_BYPASS) || defined(UKERNEL)
64 #include "afs/afs_bypasscache.h"
70 #include "afs/sysincludes.h" /* Standard vendor system headers */
71 #include "afs/afsincludes.h" /* Afs-based standard headers */
72 #include "afs/afs_stats.h" /* statistics */
73 #include "afs/nfsclient.h"
74 #include "rx/rx_globals.h"
77 #define afs_min(A,B) ((A)<(B)) ? (A) : (B)
80 /* conditional GLOCK macros */
81 #define COND_GLOCK(var) \
83 var = ISAFS_GLOCK(); \
88 #define COND_RE_GUNLOCK(var) \
95 /* conditional GUNLOCK macros */
97 #define COND_GUNLOCK(var) \
99 var = ISAFS_GLOCK(); \
104 #define COND_RE_GLOCK(var) \
111 int cache_bypass_strategy = NEVER_BYPASS_CACHE;
112 afs_size_t cache_bypass_threshold = AFS_CACHE_BYPASS_DISABLED; /* file size > threshold triggers bypass */
113 int cache_bypass_prefetch = 1; /* Should we do prefetching ? */
115 extern afs_rwlock_t afs_xcbhash;
118 * This is almost exactly like the PFlush() routine in afs_pioctl.c,
119 * but that routine is static. We are about to change a file from
120 * normal caching to bypass it's caching. Therefore, we want to
121 * free up any cache space in use by the file, and throw out any
122 * existing VM pages for the file. We keep track of the number of
123 * times we go back and forth from caching to bypass.
126 afs_TransitionToBypass(struct vcache *avc,
127 afs_ucred_t *acred, int aflags)
137 if (aflags & TRANSChangeDesiredBit)
139 if (aflags & TRANSSetManualBit)
144 ObtainWriteLock(&avc->lock, 925);
146 * Someone may have beat us to doing the transition - we had no lock
147 * when we checked the flag earlier. No cause to panic, just return.
149 if (avc->cachingStates & FCSBypass)
152 /* If we never cached this, just change state */
153 if (setDesire && (!(avc->cachingStates & FCSBypass))) {
154 avc->cachingStates |= FCSBypass;
158 /* cg2v, try to store any chunks not written 20071204 */
159 if (avc->execsOrWriters > 0) {
160 struct vrequest *treq = NULL;
162 code = afs_CreateReq(&treq, acred);
164 code = afs_StoreAllSegments(avc, treq, AFS_SYNC | AFS_LASTSTORE);
165 afs_DestroyReq(treq);
170 /* also cg2v, don't dequeue the callback */
171 ObtainWriteLock(&afs_xcbhash, 956);
172 afs_DequeueCallback(avc);
173 ReleaseWriteLock(&afs_xcbhash);
175 avc->f.states &= ~(CStatd | CDirty); /* next reference will re-stat */
176 /* now find the disk cache entries */
177 afs_TryToSmush(avc, acred, 1);
178 osi_dnlc_purgedp(avc);
179 if (avc->linkData && !(avc->f.states & CCore)) {
180 afs_osi_Free(avc->linkData, strlen(avc->linkData) + 1);
181 avc->linkData = NULL;
184 avc->cachingStates |= FCSBypass; /* Set the bypass flag */
186 avc->cachingStates |= FCSDesireBypass;
188 avc->cachingStates |= FCSManuallySet;
189 avc->cachingTransitions++;
192 ReleaseWriteLock(&avc->lock);
197 * This is almost exactly like the PFlush() routine in afs_pioctl.c,
198 * but that routine is static. We are about to change a file from
199 * bypassing caching to normal caching. Therefore, we want to
200 * throw out any existing VM pages for the file. We keep track of
201 * the number of times we go back and forth from caching to bypass.
204 afs_TransitionToCaching(struct vcache *avc,
214 if (aflags & TRANSChangeDesiredBit)
216 if (aflags & TRANSSetManualBit)
220 ObtainWriteLock(&avc->lock, 926);
222 * Someone may have beat us to doing the transition - we had no lock
223 * when we checked the flag earlier. No cause to panic, just return.
225 if (!(avc->cachingStates & FCSBypass))
228 /* Ok, we actually do need to flush */
229 ObtainWriteLock(&afs_xcbhash, 957);
230 afs_DequeueCallback(avc);
231 avc->f.states &= ~(CStatd | CDirty); /* next reference will re-stat cache entry */
232 ReleaseWriteLock(&afs_xcbhash);
233 /* now find the disk cache entries */
234 afs_TryToSmush(avc, acred, 1);
235 osi_dnlc_purgedp(avc);
236 if (avc->linkData && !(avc->f.states & CCore)) {
237 afs_osi_Free(avc->linkData, strlen(avc->linkData) + 1);
238 avc->linkData = NULL;
241 avc->cachingStates &= ~(FCSBypass); /* Reset the bypass flag */
243 avc->cachingStates &= ~(FCSDesireBypass);
245 avc->cachingStates |= FCSManuallySet;
246 avc->cachingTransitions++;
249 ReleaseWriteLock(&avc->lock);
253 /* In the case where there's an error in afs_NoCacheFetchProc or
254 * afs_PrefetchNoCache, all of the pages they've been passed need
258 typedef void * bypass_page_t;
260 #define unlock_and_release_pages(auio)
261 #define release_full_page(pp, pageoff)
264 typedef struct page * bypass_page_t;
266 #define unlock_and_release_pages(auio) \
268 struct iovec *ciov; \
271 afs_int32 iovno = 0; \
272 ciov = auio->uio_iov; \
273 iovmax = auio->uio_iovcnt - 1; \
274 pp = (bypass_page_t) ciov->iov_base; \
277 if (PageLocked(pp)) \
279 put_page(pp); /* decrement refcount */ \
284 ciov = (auio->uio_iov + iovno); \
285 pp = (bypass_page_t) ciov->iov_base; \
289 #define release_full_page(pp, pageoff) \
291 /* this is appropriate when no caller intends to unlock \
292 * and release the page */ \
293 SetPageUptodate(pp); \
297 afs_warn("afs_NoCacheFetchProc: page not locked!\n"); \
298 put_page(pp); /* decrement refcount */ \
303 afs_bypass_copy_page(bypass_page_t pp, int pageoff, struct iovec *rxiov,
304 int iovno, int iovoff, struct uio *auio, int curiov, int partial)
310 dolen = auio->uio_iov[curiov].iov_len - pageoff;
312 dolen = rxiov[iovno].iov_len - iovoff;
314 #if !defined(UKERNEL)
315 # if defined(KMAP_ATOMIC_TAKES_NO_KM_TYPE)
316 address = kmap_atomic(pp);
318 address = kmap_atomic(pp, KM_USER0);
323 memcpy(address + pageoff, (char *)(rxiov[iovno].iov_base) + iovoff, dolen);
324 #if !defined(UKERNEL)
325 # if defined(KMAP_ATOMIC_TAKES_NO_KM_TYPE)
326 kunmap_atomic(address);
328 kunmap_atomic(address, KM_USER0);
333 /* no-cache prefetch routine */
335 afs_NoCacheFetchProc(struct rx_call *acall,
338 afs_int32 release_pages,
343 int moredata, iovno, iovoff, iovmax, result, locked;
352 rxiov = osi_AllocSmallSpace(sizeof(struct iovec) * RX_MAXIOVECS);
353 ciov = auio->uio_iov;
354 pp = (bypass_page_t) ciov->iov_base;
355 iovmax = auio->uio_iovcnt - 1;
356 iovno = iovoff = result = 0;
359 COND_GUNLOCK(locked);
360 code = rx_Read(acall, (char *)&length, sizeof(afs_int32));
361 COND_RE_GLOCK(locked);
362 if (code != sizeof(afs_int32)) {
364 afs_warn("Preread error. code: %d instead of %d\n",
365 code, (int)sizeof(afs_int32));
366 unlock_and_release_pages(auio);
369 length = ntohl(length);
373 afs_warn("Preread error. Got length %d, which is greater than size %d\n",
375 unlock_and_release_pages(auio);
379 /* If we get a 0 length reply, time to cleanup and return */
381 unlock_and_release_pages(auio);
387 * The fetch protocol is extended for the AFS/DFS translator
388 * to allow multiple blocks of data, each with its own length,
389 * to be returned. As long as the top bit is set, there are more
392 * We do not do this for AFS file servers because they sometimes
393 * return large negative numbers as the transfer size.
395 if (avc->f.states & CForeign) {
396 moredata = length & 0x80000000;
397 length &= ~0x80000000;
402 for (curpage = 0; curpage <= iovmax; curpage++) {
404 /* properly, this should track uio_resid, not a fixed page size! */
405 while (pageoff < auio->uio_iov[curpage].iov_len) {
406 /* If no more iovs, issue new read. */
408 COND_GUNLOCK(locked);
409 bytes = rx_Readv(acall, rxiov, &nio, RX_MAXIOVECS, length);
410 COND_RE_GLOCK(locked);
412 afs_warn("afs_NoCacheFetchProc: rx_Read error. Return code was %d\n", bytes);
414 unlock_and_release_pages(auio);
416 } else if (bytes == 0) {
417 /* we failed to read the full length */
419 afs_warn("afs_NoCacheFetchProc: rx_Read returned zero. Aborting.\n");
420 unlock_and_release_pages(auio);
424 auio->uio_resid -= bytes;
427 pp = (bypass_page_t)auio->uio_iov[curpage].iov_base;
428 if (pageoff + (rxiov[iovno].iov_len - iovoff) <= auio->uio_iov[curpage].iov_len) {
429 /* Copy entire (or rest of) current iovec into current page */
431 afs_bypass_copy_page(pp, pageoff, rxiov, iovno, iovoff, auio, curpage, 0);
432 length -= (rxiov[iovno].iov_len - iovoff);
433 pageoff += rxiov[iovno].iov_len - iovoff;
437 /* Copy only what's needed to fill current page */
439 afs_bypass_copy_page(pp, pageoff, rxiov, iovno, iovoff, auio, curpage, 1);
440 length -= (auio->uio_iov[curpage].iov_len - pageoff);
441 iovoff += auio->uio_iov[curpage].iov_len - pageoff;
442 pageoff = auio->uio_iov[curpage].iov_len;
445 /* we filled a page, or this is the last page. conditionally release it */
446 if (pp && ((pageoff == auio->uio_iov[curpage].iov_len &&
447 release_pages) || (length == 0 && iovno >= nio)))
448 release_full_page(pp, pageoff);
450 if (length == 0 && iovno >= nio)
457 osi_FreeSmallSpace(rxiov);
462 /* dispatch a no-cache read request */
464 afs_ReadNoCache(struct vcache *avc,
465 struct nocache_read_request *bparms,
470 struct brequest *breq;
471 struct vrequest *areq = NULL;
475 afs_warn("afs_ReadNoCache VCache Error!\n");
480 /* the receiver will free areq */
481 code = afs_CreateReq(&areq, acred);
483 afs_warn("afs_ReadNoCache afs_CreateReq error!\n");
485 code = afs_VerifyVCache(avc, areq);
487 afs_warn("afs_ReadNoCache Failed to verify VCache!\n");
493 code = afs_CheckCode(code, areq, 11); /* failed to get it */
499 /* and queue this one */
503 breq = afs_BQueue(BOP_FETCH_NOCACHE, avc, B_DONTWAIT, 0, acred, 1, 1,
504 bparms, (void *)0, (void *)0);
509 afs_osi_Wait(10 * bcnt, 0, 0);
521 /* If there's a problem before we queue the request, we need to
522 * do everything that would normally happen when the request was
523 * processed, like unlocking the pages and freeing memory.
525 unlock_and_release_pages(bparms->auio);
527 afs_DestroyReq(areq);
529 osi_Free(bparms->auio->uio_iov,
530 bparms->auio->uio_iovcnt * sizeof(struct iovec));
531 osi_Free(bparms->auio, sizeof(struct uio));
532 osi_Free(bparms, sizeof(struct nocache_read_request));
537 /* Cannot have static linkage--called from BPrefetch (afs_daemons) */
539 afs_PrefetchNoCache(struct vcache *avc,
541 struct nocache_read_request *bparms)
545 struct iovec *iovecp;
547 struct vrequest *areq;
549 struct rx_connection *rxconn;
550 #ifdef AFS_64BIT_CLIENT
551 afs_int32 length_hi, bytes, locked;
555 struct rx_call *tcall;
557 struct AFSVolSync tsync;
558 struct AFSFetchStatus OutStatus;
559 struct AFSCallBack CallBack;
561 struct tlocal1 *tcallspec;
566 iovecp = auio->uio_iov;
569 tcallspec = osi_Alloc(sizeof(struct tlocal1));
571 tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK /* ignored */, &rxconn);
573 avc->callback = tc->parent->srvr->server;
574 tcall = rx_NewCall(rxconn);
575 #ifdef AFS_64BIT_CLIENT
576 if (!afs_serverHasNo64Bit(tc)) {
577 code = StartRXAFS_FetchData64(tcall,
578 (struct AFSFid *) &avc->f.fid.Fid,
582 COND_GUNLOCK(locked);
583 bytes = rx_Read(tcall, (char *)&length_hi,
585 COND_RE_GLOCK(locked);
587 if (bytes != sizeof(afs_int32)) {
589 COND_GUNLOCK(locked);
590 code = rx_EndCall(tcall, RX_PROTOCOL_ERROR);
591 COND_RE_GLOCK(locked);
595 } /* afs_serverHasNo64Bit */
596 if (code == RXGEN_OPCODE || afs_serverHasNo64Bit(tc)) {
597 if (auio->uio_offset > 0x7FFFFFFF) {
601 pos = auio->uio_offset;
602 COND_GUNLOCK(locked);
604 tcall = rx_NewCall(rxconn);
605 code = StartRXAFS_FetchData(tcall,
606 (struct AFSFid *) &avc->f.fid.Fid,
607 pos, bparms->length);
608 COND_RE_GLOCK(locked);
610 afs_serverSetNo64Bit(tc);
613 code = StartRXAFS_FetchData(tcall,
614 (struct AFSFid *) &avc->f.fid.Fid,
615 auio->uio_offset, bparms->length);
618 code = afs_NoCacheFetchProc(tcall, avc, auio,
619 1 /* release_pages */,
622 afs_warn("BYPASS: StartRXAFS_FetchData failed: %d\n", code);
623 unlock_and_release_pages(auio);
624 afs_PutConn(tc, rxconn, SHARED_LOCK);
628 code = EndRXAFS_FetchData(tcall, &tcallspec->OutStatus,
629 &tcallspec->CallBack,
632 afs_warn("BYPASS: NoCacheFetchProc failed: %d\n", code);
634 code = rx_EndCall(tcall, code);
636 afs_warn("BYPASS: No connection.\n");
638 unlock_and_release_pages(auio);
641 } while (afs_Analyze(tc, rxconn, code, &avc->f.fid, areq,
642 AFS_STATS_FS_RPCIDX_FETCHDATA,
646 * Copy appropriate fields into vcache
650 afs_ProcessFS(avc, &tcallspec->OutStatus, areq);
652 osi_Free(areq, sizeof(struct vrequest));
653 osi_Free(tcallspec, sizeof(struct tlocal1));
654 osi_Free(bparms, sizeof(struct nocache_read_request));
656 /* in UKERNEL, the "pages" are passed in */
657 osi_Free(iovecp, auio->uio_iovcnt * sizeof(struct iovec));
658 osi_Free(auio, sizeof(struct uio));