/* * Copyright (c) 2004-2006 The DragonFly Project. All rights reserved. * * This code is derived from software contributed to The DragonFly Project * by Matthew Dillon <dillon@backplane.com> * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in * the documentation and/or other materials provided with the * distribution. * 3. Neither the name of The DragonFly Project nor the names of its * contributors may be used to endorse or promote products derived * from this software without specific, prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * The journaling protocol is intended to evolve into a two-way stream * whereby transaction IDs can be acknowledged by the journaling target * when the data has been committed to hard storage. Both implicit and * explicit acknowledgement schemes will be supported, depending on the * sophistication of the journaling stream, plus resynchronization and * restart when a journaling stream is interrupted. This information will * also be made available to journaling-aware filesystems to allow better * management of their own physical storage synchronization mechanisms as * well as to allow such filesystems to take direct advantage of the kernel's * journaling layer so they don't have to roll their own. * * In addition, the worker thread will have access to much larger * spooling areas then the memory buffer is able to provide by e.g. * reserving swap space, in order to absorb potentially long interruptions * of off-site journaling streams, and to prevent 'slow' off-site linkages * from radically slowing down local filesystem operations. * * Because of the non-trivial algorithms the journaling system will be * required to support, use of a worker thread is mandatory. Efficiencies * are maintained by utilitizing the memory FIFO to batch transactions when * possible, reducing the number of gratuitous thread switches and taking * advantage of cpu caches through the use of shorter batched code paths * rather then trying to do everything in the context of the process * originating the filesystem op. In the future the memory FIFO can be * made per-cpu to remove BGL or other locking requirements. */ #include <sys/param.h> #include <sys/systm.h> #include <sys/buf.h> #include <sys/conf.h> #include <sys/kernel.h> #include <sys/queue.h> #include <sys/lock.h> #include <sys/malloc.h> #include <sys/mount.h> #include <sys/unistd.h> #include <sys/vnode.h> #include <sys/poll.h> #include <sys/mountctl.h> #include <sys/journal.h> #include <sys/file.h> #include <sys/proc.h> #include <sys/xio.h> #include <sys/socket.h> #include <sys/socketvar.h> #include <machine/limits.h> #include <vm/vm.h> #include <vm/vm_object.h> #include <vm/vm_page.h> #include <vm/vm_pager.h> #include <vm/vnode_pager.h> #include <sys/file2.h> #include <sys/mplock2.h> #include <sys/spinlock2.h> static void journal_wthread(void *info); static void journal_rthread(void *info); static void *journal_reserve(struct journal *jo, struct journal_rawrecbeg **rawpp, int16_t streamid, int bytes); static void *journal_extend(struct journal *jo, struct journal_rawrecbeg **rawpp, int truncbytes, int bytes, int *newstreamrecp); static void journal_abort(struct journal *jo, struct journal_rawrecbeg **rawpp); static void journal_commit(struct journal *jo, struct journal_rawrecbeg **rawpp, int bytes, int closeout); static void jrecord_data(struct jrecord *jrec, void *buf, int bytes, int dtype); MALLOC_DEFINE(M_JOURNAL, "journal", "Journaling structures"); MALLOC_DEFINE(M_JFIFO, "journal-fifo", "Journal FIFO"); void journal_create_threads(struct journal *jo) { jo->flags &= ~(MC_JOURNAL_STOP_REQ | MC_JOURNAL_STOP_IMM); jo->flags |= MC_JOURNAL_WACTIVE; lwkt_create(journal_wthread, jo, NULL, &jo->wthread, TDF_NOSTART, -1, "journal w:%.*s", JIDMAX, jo->id); lwkt_setpri(&jo->wthread, TDPRI_KERN_DAEMON); lwkt_schedule(&jo->wthread); if (jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) { jo->flags |= MC_JOURNAL_RACTIVE; lwkt_create(journal_rthread, jo, NULL, &jo->rthread, TDF_NOSTART, -1, "journal r:%.*s", JIDMAX, jo->id); lwkt_setpri(&jo->rthread, TDPRI_KERN_DAEMON); lwkt_schedule(&jo->rthread); } } void journal_destroy_threads(struct journal *jo, int flags) { int wcount; jo->flags |= MC_JOURNAL_STOP_REQ | (flags & MC_JOURNAL_STOP_IMM); wakeup(&jo->fifo); wcount = 0; while (jo->flags & (MC_JOURNAL_WACTIVE | MC_JOURNAL_RACTIVE)) { tsleep(jo, 0, "jwait", hz); if (++wcount % 10 == 0) { kprintf("Warning: journal %s waiting for descriptors to close\n", jo->id); } } /* * XXX SMP - threads should move to cpu requesting the restart or * termination before finishing up to properly interlock. */ tsleep(jo, 0, "jwait", hz); lwkt_free_thread(&jo->wthread); if (jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) lwkt_free_thread(&jo->rthread); } /* * The per-journal worker thread is responsible for writing out the * journal's FIFO to the target stream. */ static void journal_wthread(void *info) { struct journal *jo = info; struct journal_rawrecbeg *rawp; int error; size_t avail; size_t bytes; size_t res; /* not MPSAFE yet */ get_mplock(); for (;;) { /* * Calculate the number of bytes available to write. This buffer * area may contain reserved records so we can't just write it out * without further checks. */ bytes = jo->fifo.windex - jo->fifo.rindex; /* * sleep if no bytes are available or if an incomplete record is * encountered (it needs to be filled in before we can write it * out), and skip any pad records that we encounter. */ if (bytes == 0) { if (jo->flags & MC_JOURNAL_STOP_REQ) break; tsleep(&jo->fifo, 0, "jfifo", hz); continue; } /* * Sleep if we can not go any further due to hitting an incomplete * record. This case should occur rarely but may have to be better * optimized XXX. */ rawp = (void *)(jo->fifo.membase + (jo->fifo.rindex & jo->fifo.mask)); if (rawp->begmagic == JREC_INCOMPLETEMAGIC) { tsleep(&jo->fifo, 0, "jpad", hz); continue; } /* * Skip any pad records. We do not write out pad records if we can * help it. */ if (rawp->streamid == JREC_STREAMID_PAD) { if ((jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) == 0) { if (jo->fifo.rindex == jo->fifo.xindex) { jo->fifo.xindex += (rawp->recsize + 15) & ~15; jo->total_acked += (rawp->recsize + 15) & ~15; } } jo->fifo.rindex += (rawp->recsize + 15) & ~15; jo->total_acked += bytes; KKASSERT(jo->fifo.windex - jo->fifo.rindex >= 0); continue; } /* * 'bytes' is the amount of data that can potentially be written out. * Calculate 'res', the amount of data that can actually be written * out. res is bounded either by hitting the end of the physical * memory buffer or by hitting an incomplete record. Incomplete * records often occur due to the way the space reservation model * works. */ res = 0; avail = jo->fifo.size - (jo->fifo.rindex & jo->fifo.mask); while (res < bytes && rawp->begmagic == JREC_BEGMAGIC) { res += (rawp->recsize + 15) & ~15; if (res >= avail) { KKASSERT(res == avail); break; } rawp = (void *)((char *)rawp + ((rawp->recsize + 15) & ~15)); } /* * Issue the write and deal with any errors or other conditions. * For now assume blocking I/O. Since we are record-aware the * code cannot yet handle partial writes. * * We bump rindex prior to issuing the write to avoid racing * the acknowledgement coming back (which could prevent the ack * from bumping xindex). Restarts are always based on xindex so * we do not try to undo the rindex if an error occurs. * * XXX EWOULDBLOCK/NBIO * XXX notification on failure * XXX permanent verses temporary failures * XXX two-way acknowledgement stream in the return direction / xindex */ bytes = res; jo->fifo.rindex += bytes; error = fp_write(jo->fp, jo->fifo.membase + ((jo->fifo.rindex - bytes) & jo->fifo.mask), bytes, &res, UIO_SYSSPACE); if (error) { kprintf("journal_thread(%s) write, error %d\n", jo->id, error); /* XXX */ } else { KKASSERT(res == bytes); } /* * Advance rindex. If the journal stream is not full duplex we also * advance xindex, otherwise the rjournal thread is responsible for * advancing xindex. */ if ((jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) == 0) { jo->fifo.xindex += bytes; jo->total_acked += bytes; } KKASSERT(jo->fifo.windex - jo->fifo.rindex >= 0); if ((jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) == 0) { if (jo->flags & MC_JOURNAL_WWAIT) { jo->flags &= ~MC_JOURNAL_WWAIT; /* XXX hysteresis */ wakeup(&jo->fifo.windex); } } } fp_shutdown(jo->fp, SHUT_WR); jo->flags &= ~MC_JOURNAL_WACTIVE; wakeup(jo); wakeup(&jo->fifo.windex); rel_mplock(); } /* * A second per-journal worker thread is created for two-way journaling * streams to deal with the return acknowledgement stream. */ static void journal_rthread(void *info) { struct journal_rawrecbeg *rawp; struct journal_ackrecord ack; struct journal *jo = info; int64_t transid; int error; size_t count; size_t bytes; transid = 0; error = 0; /* not MPSAFE yet */ get_mplock(); for (;;) { /* * We have been asked to stop */ if (jo->flags & MC_JOURNAL_STOP_REQ) break; /* * If we have no active transaction id, get one from the return * stream. */ if (transid == 0) { error = fp_read(jo->fp, &ack, sizeof(ack), &count, 1, UIO_SYSSPACE); #if 0 kprintf("fp_read ack error %d count %d\n", error, count); #endif if (error || count != sizeof(ack)) break; if (error) { kprintf("read error %d on receive stream\n", error); break; } if (ack.rbeg.begmagic != JREC_BEGMAGIC || ack.rend.endmagic != JREC_ENDMAGIC ) { kprintf("bad begmagic or endmagic on receive stream\n"); break; } transid = ack.rbeg.transid; } /* * Calculate the number of unacknowledged bytes. If there are no * unacknowledged bytes then unsent data was acknowledged, report, * sleep a bit, and loop in that case. This should not happen * normally. The ack record is thrown away. */ bytes = jo->fifo.rindex - jo->fifo.xindex; if (bytes == 0) { kprintf("warning: unsent data acknowledged transid %08llx\n", (long long)transid); tsleep(&jo->fifo.xindex, 0, "jrseq", hz); transid = 0; continue; } /* * Since rindex has advanced, the record pointed to by xindex * must be a valid record. */ rawp = (void *)(jo->fifo.membase + (jo->fifo.xindex & jo->fifo.mask)); KKASSERT(rawp->begmagic == JREC_BEGMAGIC); KKASSERT(rawp->recsize <= bytes); /* * The target can acknowledge several records at once. */ if (rawp->transid < transid) { #if 1 kprintf("ackskip %08llx/%08llx\n", (long long)rawp->transid, (long long)transid); #endif jo->fifo.xindex += (rawp->recsize + 15) & ~15; jo->total_acked += (rawp->recsize + 15) & ~15; if (jo->flags & MC_JOURNAL_WWAIT) { jo->flags &= ~MC_JOURNAL_WWAIT; /* XXX hysteresis */ wakeup(&jo->fifo.windex); } continue; } if (rawp->transid == transid) { #if 1 kprintf("ackskip %08llx/%08llx\n", (long long)rawp->transid, (long long)transid); #endif jo->fifo.xindex += (rawp->recsize + 15) & ~15; jo->total_acked += (rawp->recsize + 15) & ~15; if (jo->flags & MC_JOURNAL_WWAIT) { jo->flags &= ~MC_JOURNAL_WWAIT; /* XXX hysteresis */ wakeup(&jo->fifo.windex); } transid = 0; continue; } kprintf("warning: unsent data(2) acknowledged transid %08llx\n", (long long)transid); transid = 0; } jo->flags &= ~MC_JOURNAL_RACTIVE; wakeup(jo); wakeup(&jo->fifo.windex); rel_mplock(); } /* * This builds a pad record which the journaling thread will skip over. Pad * records are required when we are unable to reserve sufficient stream space * due to insufficient space at the end of the physical memory fifo. * * Even though the record is not transmitted, a normal transid must be * assigned to it so link recovery operations after a failure work properly. */ static void journal_build_pad(struct journal_rawrecbeg *rawp, int recsize, int64_t transid) { struct journal_rawrecend *rendp; KKASSERT((recsize & 15) == 0 && recsize >= 16); rawp->streamid = JREC_STREAMID_PAD; rawp->recsize = recsize; /* must be 16-byte aligned */ rawp->transid = transid; /* * WARNING, rendp may overlap rawp->transid. This is necessary to * allow PAD records to fit in 16 bytes. Use cpu_ccfence() to * hopefully cause the compiler to not make any assumptions. */ rendp = (void *)((char *)rawp + rawp->recsize - sizeof(*rendp)); rendp->endmagic = JREC_ENDMAGIC; rendp->check = 0; rendp->recsize = rawp->recsize; /* * Set the begin magic last. This is what will allow the journal * thread to write the record out. Use a store fence to prevent * compiler and cpu reordering of the writes. */ cpu_sfence(); rawp->begmagic = JREC_BEGMAGIC; } /* * Wake up the worker thread if the FIFO is more then half full or if * someone is waiting for space to be freed up. Otherwise let the * heartbeat deal with it. Being able to avoid waking up the worker * is the key to the journal's cpu performance. */ static __inline void journal_commit_wakeup(struct journal *jo) { int avail; avail = jo->fifo.size - (jo->fifo.windex - jo->fifo.xindex); KKASSERT(avail >= 0); if ((avail < (jo->fifo.size >> 1)) || (jo->flags & MC_JOURNAL_WWAIT)) wakeup(&jo->fifo); } /* * Create a new BEGIN stream record with the specified streamid and the * specified amount of payload space. *rawpp will be set to point to the * base of the new stream record and a pointer to the base of the payload * space will be returned. *rawpp does not need to be pre-NULLd prior to * making this call. The raw record header will be partially initialized. * * A stream can be extended, aborted, or committed by other API calls * below. This may result in a sequence of potentially disconnected * stream records to be output to the journaling target. The first record * (the one created by this function) will be marked JREC_STREAMCTL_BEGIN, * while the last record on commit or abort will be marked JREC_STREAMCTL_END * (and possibly also JREC_STREAMCTL_ABORTED). The last record could wind * up being the same as the first, in which case the bits are all set in * the first record. * * The stream record is created in an incomplete state by setting the begin * magic to JREC_INCOMPLETEMAGIC. This prevents the worker thread from * flushing the fifo past our record until we have finished populating it. * Other threads can reserve and operate on their own space without stalling * but the stream output will stall until we have completed operations. The * memory FIFO is intended to be large enough to absorb such situations * without stalling out other threads. */ static void * journal_reserve(struct journal *jo, struct journal_rawrecbeg **rawpp, int16_t streamid, int bytes) { struct journal_rawrecbeg *rawp; int avail; int availtoend; int req; /* * Add header and trailer overheads to the passed payload. Note that * the passed payload size need not be aligned in any way. */ bytes += sizeof(struct journal_rawrecbeg); bytes += sizeof(struct journal_rawrecend); for (;;) { /* * First, check boundary conditions. If the request would wrap around * we have to skip past the ending block and return to the beginning * of the FIFO's buffer. Calculate 'req' which is the actual number * of bytes being reserved, including wrap-around dead space. * * Neither 'bytes' or 'req' are aligned. * * Note that availtoend is not truncated to avail and so cannot be * used to determine whether the reservation is possible by itself. * Also, since all fifo ops are 16-byte aligned, we can check * the size before calculating the aligned size. */ availtoend = jo->fifo.size - (jo->fifo.windex & jo->fifo.mask); KKASSERT((availtoend & 15) == 0); if (bytes > availtoend) req = bytes + availtoend; /* add pad to end */ else req = bytes; /* * Next calculate the total available space and see if it is * sufficient. We cannot overwrite previously buffered data * past xindex because otherwise we would not be able to restart * a broken link at the target's last point of commit. */ avail = jo->fifo.size - (jo->fifo.windex - jo->fifo.xindex); KKASSERT(avail >= 0 && (avail & 15) == 0); if (avail < req) { /* XXX MC_JOURNAL_STOP_IMM */ jo->flags |= MC_JOURNAL_WWAIT; ++jo->fifostalls; tsleep(&jo->fifo.windex, 0, "jwrite", 0); continue; } /* * Create a pad record for any dead space and create an incomplete * record for the live space, then return a pointer to the * contiguous buffer space that was requested. * * NOTE: The worker thread will not flush past an incomplete * record, so the reserved space can be filled in at-will. The * journaling code must also be aware the reserved sections occuring * after this one will also not be written out even if completed * until this one is completed. * * The transaction id must accomodate real and potential pad creation. */ rawp = (void *)(jo->fifo.membase + (jo->fifo.windex & jo->fifo.mask)); if (req != bytes) { journal_build_pad(rawp, availtoend, jo->transid); ++jo->transid; rawp = (void *)jo->fifo.membase; } rawp->begmagic = JREC_INCOMPLETEMAGIC; /* updated by abort/commit */ rawp->recsize = bytes; /* (unaligned size) */ rawp->streamid = streamid | JREC_STREAMCTL_BEGIN; rawp->transid = jo->transid; jo->transid += 2; /* * Issue a memory barrier to guarentee that the record data has been * properly initialized before we advance the write index and return * a pointer to the reserved record. Otherwise the worker thread * could accidently run past us. * * Note that stream records are always 16-byte aligned. */ cpu_sfence(); jo->fifo.windex += (req + 15) & ~15; *rawpp = rawp; return(rawp + 1); } /* not reached */ *rawpp = NULL; return(NULL); } /* * Attempt to extend the stream record by <bytes> worth of payload space. * * If it is possible to extend the existing stream record no truncation * occurs and the record is extended as specified. A pointer to the * truncation offset within the payload space is returned. * * If it is not possible to do this the existing stream record is truncated * and committed, and a new stream record of size <bytes> is created. A * pointer to the base of the new stream record's payload space is returned. * * *rawpp is set to the new reservation in the case of a new record but * the caller cannot depend on a comparison with the old rawp to determine if * this case occurs because we could end up using the same memory FIFO * offset for the new stream record. Use *newstreamrecp instead. */ static void * journal_extend(struct journal *jo, struct journal_rawrecbeg **rawpp, int truncbytes, int bytes, int *newstreamrecp) { struct journal_rawrecbeg *rawp; int16_t streamid; int availtoend; int avail; int osize; int nsize; int wbase; void *rptr; *newstreamrecp = 0; rawp = *rawpp; osize = (rawp->recsize + 15) & ~15; nsize = (rawp->recsize + bytes + 15) & ~15; wbase = (char *)rawp - jo->fifo.membase; /* * If the aligned record size does not change we can trivially adjust * the record size. */ if (nsize == osize) { rawp->recsize += bytes; return((char *)(rawp + 1) + truncbytes); } /* * If the fifo's write index hasn't been modified since we made the * reservation and we do not hit any boundary conditions, we can * trivially make the record smaller or larger. */ if ((jo->fifo.windex & jo->fifo.mask) == wbase + osize) { availtoend = jo->fifo.size - wbase; avail = jo->fifo.size - (jo->fifo.windex - jo->fifo.xindex) + osize; KKASSERT((availtoend & 15) == 0); KKASSERT((avail & 15) == 0); if (nsize <= avail && nsize <= availtoend) { jo->fifo.windex += nsize - osize; rawp->recsize += bytes; return((char *)(rawp + 1) + truncbytes); } } /* * It was not possible to extend the buffer. Commit the current * buffer and create a new one. We manually clear the BEGIN mark that * journal_reserve() creates (because this is a continuing record, not * the start of a new stream). */ streamid = rawp->streamid & JREC_STREAMID_MASK; journal_commit(jo, rawpp, truncbytes, 0); rptr = journal_reserve(jo, rawpp, streamid, bytes); rawp = *rawpp; rawp->streamid &= ~JREC_STREAMCTL_BEGIN; *newstreamrecp = 1; return(rptr); } /* * Abort a journal record. If the transaction record represents a stream * BEGIN and we can reverse the fifo's write index we can simply reverse * index the entire record, as if it were never reserved in the first place. * * Otherwise we set the JREC_STREAMCTL_ABORTED bit and commit the record * with the payload truncated to 0 bytes. */ static void journal_abort(struct journal *jo, struct journal_rawrecbeg **rawpp) { struct journal_rawrecbeg *rawp; int osize; rawp = *rawpp; osize = (rawp->recsize + 15) & ~15; if ((rawp->streamid & JREC_STREAMCTL_BEGIN) && (jo->fifo.windex & jo->fifo.mask) == (char *)rawp - jo->fifo.membase + osize) { jo->fifo.windex -= osize; *rawpp = NULL; } else { rawp->streamid |= JREC_STREAMCTL_ABORTED; journal_commit(jo, rawpp, 0, 1); } } /* * Commit a journal record and potentially truncate it to the specified * number of payload bytes. If you do not want to truncate the record, * simply pass -1 for the bytes parameter. Do not pass rawp->recsize, that * field includes header and trailer and will not be correct. Note that * passing 0 will truncate the entire data payload of the record. * * The logical stream is terminated by this function. * * If truncation occurs, and it is not possible to physically optimize the * memory FIFO due to other threads having reserved space after ours, * the remaining reserved space will be covered by a pad record. */ static void journal_commit(struct journal *jo, struct journal_rawrecbeg **rawpp, int bytes, int closeout) { struct journal_rawrecbeg *rawp; struct journal_rawrecend *rendp; int osize; int nsize; rawp = *rawpp; *rawpp = NULL; KKASSERT((char *)rawp >= jo->fifo.membase && (char *)rawp + rawp->recsize <= jo->fifo.membase + jo->fifo.size); KKASSERT(((intptr_t)rawp & 15) == 0); /* * Truncate the record if necessary. If the FIFO write index as still * at the end of our record we can optimally backindex it. Otherwise * we have to insert a pad record to cover the dead space. * * We calculate osize which is the 16-byte-aligned original recsize. * We calculate nsize which is the 16-byte-aligned new recsize. * * Due to alignment issues or in case the passed truncation bytes is * the same as the original payload, nsize may be equal to osize even * if the committed bytes is less then the originally reserved bytes. */ if (bytes >= 0) { KKASSERT(bytes >= 0 && bytes <= rawp->recsize - sizeof(struct journal_rawrecbeg) - sizeof(struct journal_rawrecend)); osize = (rawp->recsize + 15) & ~15; rawp->recsize = bytes + sizeof(struct journal_rawrecbeg) + sizeof(struct journal_rawrecend); nsize = (rawp->recsize + 15) & ~15; KKASSERT(nsize <= osize); if (osize == nsize) { /* do nothing */ } else if ((jo->fifo.windex & jo->fifo.mask) == (char *)rawp - jo->fifo.membase + osize) { /* we are able to backindex the fifo */ jo->fifo.windex -= osize - nsize; } else { /* we cannot backindex the fifo, emplace a pad in the dead space */ journal_build_pad((void *)((char *)rawp + nsize), osize - nsize, rawp->transid + 1); } } /* * Fill in the trailer. Note that unlike pad records, the trailer will * never overlap the header. */ rendp = (void *)((char *)rawp + ((rawp->recsize + 15) & ~15) - sizeof(*rendp)); rendp->endmagic = JREC_ENDMAGIC; rendp->recsize = rawp->recsize; rendp->check = 0; /* XXX check word, disabled for now */ /* * Fill in begmagic last. This will allow the worker thread to proceed. * Use a memory barrier to guarentee write ordering. Mark the stream * as terminated if closeout is set. This is the typical case. */ if (closeout) rawp->streamid |= JREC_STREAMCTL_END; cpu_sfence(); /* memory and compiler barrier */ rawp->begmagic = JREC_BEGMAGIC; journal_commit_wakeup(jo); } /************************************************************************ * TRANSACTION SUPPORT ROUTINES * ************************************************************************ * * JRECORD_*() - routines to create subrecord transactions and embed them * in the logical streams managed by the journal_*() routines. */ /* * Initialize the passed jrecord structure and start a new stream transaction * by reserving an initial build space in the journal's memory FIFO. */ void jrecord_init(struct journal *jo, struct jrecord *jrec, int16_t streamid) { bzero(jrec, sizeof(*jrec)); jrec->jo = jo; jrec->streamid = streamid; jrec->stream_residual = JREC_DEFAULTSIZE; jrec->stream_reserved = jrec->stream_residual; jrec->stream_ptr = journal_reserve(jo, &jrec->rawp, streamid, jrec->stream_reserved); } /* * Push a recursive record type. All pushes should have matching pops. * The old parent is returned and the newly pushed record becomes the * new parent. Note that the old parent's pointer may already be invalid * or may become invalid if jrecord_write() had to build a new stream * record, so the caller should not mess with the returned pointer in * any way other then to save it. */ struct journal_subrecord * jrecord_push(struct jrecord *jrec, int16_t rectype) { struct journal_subrecord *save; save = jrec->parent; jrec->parent = jrecord_write(jrec, rectype|JMASK_NESTED, 0); jrec->last = NULL; KKASSERT(jrec->parent != NULL); ++jrec->pushcount; ++jrec->pushptrgood; /* cleared on flush */ return(save); } /* * Pop a previously pushed sub-transaction. We must set JMASK_LAST * on the last record written within the subtransaction. If the last * record written is not accessible or if the subtransaction is empty, * we must write out a pad record with JMASK_LAST set before popping. * * When popping a subtransaction the parent record's recsize field * will be properly set. If the parent pointer is no longer valid * (which can occur if the data has already been flushed out to the * stream), the protocol spec allows us to leave it 0. * * The saved parent pointer which we restore may or may not be valid, * and if not valid may or may not be NULL, depending on the value * of pushptrgood. */ void jrecord_pop(struct jrecord *jrec, struct journal_subrecord *save) { struct journal_subrecord *last; KKASSERT(jrec->pushcount > 0); KKASSERT(jrec->residual == 0); /* * Set JMASK_LAST on the last record we wrote at the current * level. If last is NULL we either no longer have access to the * record or the subtransaction was empty and we must write out a pad * record. */ if ((last = jrec->last) == NULL) { jrecord_write(jrec, JLEAF_PAD|JMASK_LAST, 0); last = jrec->last; /* reload after possible flush */ } else { last->rectype |= JMASK_LAST; } /* * pushptrgood tells us how many levels of parent record pointers * are valid. The jrec only stores the current parent record pointer * (and it is only valid if pushptrgood != 0). The higher level parent * record pointers are saved by the routines calling jrecord_push() and * jrecord_pop(). These pointers may become stale and we determine * that fact by tracking the count of valid parent pointers with * pushptrgood. Pointers become invalid when their related stream * record gets pushed out. * * If no pointer is available (the data has already been pushed out), * then no fixup of e.g. the length field is possible for non-leaf * nodes. The protocol allows for this situation by placing a larger * burden on the program scanning the stream on the other end. * * [parentA] * [node X] * [parentB] * [node Y] * [node Z] * (pop B) see NOTE B * (pop A) see NOTE A * * NOTE B: This pop sets LAST in node Z if the node is still accessible, * else a PAD record is appended and LAST is set in that. * * This pop sets the record size in parentB if parentB is still * accessible, else the record size is left 0 (the scanner must * deal with that). * * This pop sets the new 'last' record to parentB, the pointer * to which may or may not still be accessible. * * NOTE A: This pop sets LAST in parentB if the node is still accessible, * else a PAD record is appended and LAST is set in that. * * This pop sets the record size in parentA if parentA is still * accessible, else the record size is left 0 (the scanner must * deal with that). * * This pop sets the new 'last' record to parentA, the pointer * to which may or may not still be accessible. * * Also note that the last record in the stream transaction, which in * the above example is parentA, does not currently have the LAST bit * set. * * The current parent becomes the last record relative to the * saved parent passed into us. It's validity is based on * whether pushptrgood is non-zero prior to decrementing. The saved * parent becomes the new parent, and its validity is based on whether * pushptrgood is non-zero after decrementing. * * The old jrec->parent may be NULL if it is no longer accessible. * If pushptrgood is non-zero, however, it is guarenteed to not * be NULL (since no flush occured). */ jrec->last = jrec->parent; --jrec->pushcount; if (jrec->pushptrgood) { KKASSERT(jrec->last != NULL && last != NULL); if (--jrec->pushptrgood == 0) { jrec->parent = NULL; /* 'save' contains garbage or NULL */ } else { KKASSERT(save != NULL); jrec->parent = save; /* 'save' must not be NULL */ } /* * Set the record size in the old parent. 'last' still points to * the original last record in the subtransaction being popped, * jrec->last points to the old parent (which became the last * record relative to the new parent being popped into). */ jrec->last->recsize = (char *)last + last->recsize - (char *)jrec->last; } else { jrec->parent = NULL; KKASSERT(jrec->last == NULL); } } /* * Write out a leaf record, including associated data. */ void jrecord_leaf(struct jrecord *jrec, int16_t rectype, void *ptr, int bytes) { jrecord_write(jrec, rectype, bytes); jrecord_data(jrec, ptr, bytes, JDATA_KERN); } void jrecord_leaf_uio(struct jrecord *jrec, int16_t rectype, struct uio *uio) { struct iovec *iov; int i; for (i = 0; i < uio->uio_iovcnt; ++i) { iov = &uio->uio_iov[i]; if (iov->iov_len == 0) continue; if (uio->uio_segflg == UIO_SYSSPACE) { jrecord_write(jrec, rectype, iov->iov_len); jrecord_data(jrec, iov->iov_base, iov->iov_len, JDATA_KERN); } else { /* UIO_USERSPACE */ jrecord_write(jrec, rectype, iov->iov_len); jrecord_data(jrec, iov->iov_base, iov->iov_len, JDATA_USER); } } } void jrecord_leaf_xio(struct jrecord *jrec, int16_t rectype, xio_t xio) { int bytes = xio->xio_npages * PAGE_SIZE; jrecord_write(jrec, rectype, bytes); jrecord_data(jrec, xio, bytes, JDATA_XIO); } /* * Write a leaf record out and return a pointer to its base. The leaf * record may contain potentially megabytes of data which is supplied * in jrecord_data() calls. The exact amount must be specified in this * call. * * THE RETURNED SUBRECORD POINTER IS ONLY VALID IMMEDIATELY AFTER THE * CALL AND MAY BECOME INVALID AT ANY TIME. ONLY THE PUSH/POP CODE SHOULD * USE THE RETURN VALUE. */ struct journal_subrecord * jrecord_write(struct jrecord *jrec, int16_t rectype, int bytes) { struct journal_subrecord *last; int pusheditout; /* * Try to catch some obvious errors. Nesting records must specify a * size of 0, and there should be no left-overs from previous operations * (such as incomplete data writeouts). */ KKASSERT(bytes == 0 || (rectype & JMASK_NESTED) == 0); KKASSERT(jrec->residual == 0); /* * Check to see if the current stream record has enough room for * the new subrecord header. If it doesn't we extend the current * stream record. * * This may have the side effect of pushing out the current stream record * and creating a new one. We must adjust our stream tracking fields * accordingly. */ if (jrec->stream_residual < sizeof(struct journal_subrecord)) { jrec->stream_ptr = journal_extend(jrec->jo, &jrec->rawp, jrec->stream_reserved - jrec->stream_residual, JREC_DEFAULTSIZE, &pusheditout); if (pusheditout) { /* * If a pushout occured, the pushed out stream record was * truncated as specified and the new record is exactly the * extension size specified. */ jrec->stream_reserved = JREC_DEFAULTSIZE; jrec->stream_residual = JREC_DEFAULTSIZE; jrec->parent = NULL; /* no longer accessible */ jrec->pushptrgood = 0; /* restored parents in pops no good */ } else { /* * If no pushout occured the stream record is NOT truncated and * IS extended. */ jrec->stream_reserved += JREC_DEFAULTSIZE; jrec->stream_residual += JREC_DEFAULTSIZE; } } last = (void *)jrec->stream_ptr; last->rectype = rectype; last->reserved = 0; /* * We may not know the record size for recursive records and the * header may become unavailable due to limited FIFO space. Write * -1 to indicate this special case. */ if ((rectype & JMASK_NESTED) && bytes == 0) last->recsize = -1; else last->recsize = sizeof(struct journal_subrecord) + bytes; jrec->last = last; jrec->residual = bytes; /* remaining data to be posted */ jrec->residual_align = -bytes & 7; /* post-data alignment required */ jrec->stream_ptr += sizeof(*last); /* current write pointer */ jrec->stream_residual -= sizeof(*last); /* space remaining in stream */ return(last); } /* * Write out the data associated with a leaf record. Any number of calls * to this routine may be made as long as the byte count adds up to the * amount originally specified in jrecord_write(). * * The act of writing out the leaf data may result in numerous stream records * being pushed out. Callers should be aware that even the associated * subrecord header may become inaccessible due to stream record pushouts. */ static void jrecord_data(struct jrecord *jrec, void *buf, int bytes, int dtype) { int pusheditout; int extsize; int xio_offset = 0; KKASSERT(bytes >= 0 && bytes <= jrec->residual); /* * Push out stream records as long as there is insufficient room to hold * the remaining data. */ while (jrec->stream_residual < bytes) { /* * Fill in any remaining space in the current stream record. */ switch (dtype) { case JDATA_KERN: bcopy(buf, jrec->stream_ptr, jrec->stream_residual); break; case JDATA_USER: copyin(buf, jrec->stream_ptr, jrec->stream_residual); break; case JDATA_XIO: xio_copy_xtok((xio_t)buf, xio_offset, jrec->stream_ptr, jrec->stream_residual); xio_offset += jrec->stream_residual; break; } if (dtype != JDATA_XIO) buf = (char *)buf + jrec->stream_residual; bytes -= jrec->stream_residual; /*jrec->stream_ptr += jrec->stream_residual;*/ jrec->residual -= jrec->stream_residual; jrec->stream_residual = 0; /* * Try to extend the current stream record, but no more then 1/4 * the size of the FIFO. */ extsize = jrec->jo->fifo.size >> 2; if (extsize > bytes) extsize = (bytes + 15) & ~15; jrec->stream_ptr = journal_extend(jrec->jo, &jrec->rawp, jrec->stream_reserved - jrec->stream_residual, extsize, &pusheditout); if (pusheditout) { jrec->stream_reserved = extsize; jrec->stream_residual = extsize; jrec->parent = NULL; /* no longer accessible */ jrec->last = NULL; /* no longer accessible */ jrec->pushptrgood = 0; /* restored parents in pops no good */ } else { jrec->stream_reserved += extsize; jrec->stream_residual += extsize; } } /* * Push out any remaining bytes into the current stream record. */ if (bytes) { switch (dtype) { case JDATA_KERN: bcopy(buf, jrec->stream_ptr, bytes); break; case JDATA_USER: copyin(buf, jrec->stream_ptr, bytes); break; case JDATA_XIO: xio_copy_xtok((xio_t)buf, xio_offset, jrec->stream_ptr, bytes); break; } jrec->stream_ptr += bytes; jrec->stream_residual -= bytes; jrec->residual -= bytes; } /* * Handle data alignment requirements for the subrecord. Because the * stream record's data space is more strictly aligned, it must already * have sufficient space to hold any subrecord alignment slop. */ if (jrec->residual == 0 && jrec->residual_align) { KKASSERT(jrec->residual_align <= jrec->stream_residual); bzero(jrec->stream_ptr, jrec->residual_align); jrec->stream_ptr += jrec->residual_align; jrec->stream_residual -= jrec->residual_align; jrec->residual_align = 0; } } /* * We are finished with the transaction. This closes the transaction created * by jrecord_init(). * * NOTE: If abortit is not set then we must be at the top level with no * residual subrecord data left to output. * * If abortit is set then we can be in any state, all pushes will be * popped and it is ok for there to be residual data. This works * because the virtual stream itself is truncated. Scanners must deal * with this situation. * * The stream record will be committed or aborted as specified and jrecord * resources will be cleaned up. */ void jrecord_done(struct jrecord *jrec, int abortit) { KKASSERT(jrec->rawp != NULL); if (abortit) { journal_abort(jrec->jo, &jrec->rawp); } else { KKASSERT(jrec->pushcount == 0 && jrec->residual == 0); journal_commit(jrec->jo, &jrec->rawp, jrec->stream_reserved - jrec->stream_residual, 1); } /* * jrec should not be used beyond this point without another init, * but clean up some fields to ensure that we panic if it is. * * Note that jrec->rawp is NULLd out by journal_abort/journal_commit. */ jrec->jo = NULL; jrec->stream_ptr = NULL; } /************************************************************************ * LOW LEVEL RECORD SUPPORT ROUTINES * ************************************************************************ * * These routine create low level recursive and leaf subrecords representing * common filesystem structures. */ /* * Write out a filename path relative to the base of the mount point. * rectype is typically JLEAF_PATH{1,2,3,4}. */ void jrecord_write_path(struct jrecord *jrec, int16_t rectype, struct namecache *ncp) { char buf[64]; /* local buffer if it fits, else malloced */ char *base; int pathlen; int index; struct namecache *scan; /* * Pass 1 - figure out the number of bytes required. Include terminating * \0 on last element and '/' separator on other elements. * * The namecache topology terminates at the root of the filesystem * (the normal lookup code would then continue by using the mount * structure to figure out what it was mounted on). */ again: pathlen = 0; for (scan = ncp; scan; scan = scan->nc_parent) { if (scan->nc_nlen > 0) pathlen += scan->nc_nlen + 1; } if (pathlen <= sizeof(buf)) base = buf; else base = kmalloc(pathlen, M_TEMP, M_INTWAIT); /* * Pass 2 - generate the path buffer */ index = pathlen; for (scan = ncp; scan; scan = scan->nc_parent) { if (scan->nc_nlen == 0) continue; if (scan->nc_nlen >= index) { if (base != buf) kfree(base, M_TEMP); goto again; } if (index == pathlen) base[--index] = 0; else base[--index] = '/'; index -= scan->nc_nlen; bcopy(scan->nc_name, base + index, scan->nc_nlen); } jrecord_leaf(jrec, rectype, base + index, pathlen - index); if (base != buf) kfree(base, M_TEMP); } /* * Write out a file attribute structure. While somewhat inefficient, using * a recursive data structure is the most portable and extensible way. */ void jrecord_write_vattr(struct jrecord *jrec, struct vattr *vat) { void *save; save = jrecord_push(jrec, JTYPE_VATTR); if (vat->va_type != VNON) jrecord_leaf(jrec, JLEAF_VTYPE, &vat->va_type, sizeof(vat->va_type)); if (vat->va_mode != (mode_t)VNOVAL) jrecord_leaf(jrec, JLEAF_MODES, &vat->va_mode, sizeof(vat->va_mode)); if (vat->va_nlink != VNOVAL) jrecord_leaf(jrec, JLEAF_NLINK, &vat->va_nlink, sizeof(vat->va_nlink)); if (vat->va_uid != VNOVAL) jrecord_leaf(jrec, JLEAF_UID, &vat->va_uid, sizeof(vat->va_uid)); if (vat->va_gid != VNOVAL) jrecord_leaf(jrec, JLEAF_GID, &vat->va_gid, sizeof(vat->va_gid)); if (vat->va_fsid != VNOVAL) jrecord_leaf(jrec, JLEAF_FSID, &vat->va_fsid, sizeof(vat->va_fsid)); if (vat->va_fileid != VNOVAL) jrecord_leaf(jrec, JLEAF_INUM, &vat->va_fileid, sizeof(vat->va_fileid)); if (vat->va_size != VNOVAL) jrecord_leaf(jrec, JLEAF_SIZE, &vat->va_size, sizeof(vat->va_size)); if (vat->va_atime.tv_sec != VNOVAL) jrecord_leaf(jrec, JLEAF_ATIME, &vat->va_atime, sizeof(vat->va_atime)); if (vat->va_mtime.tv_sec != VNOVAL) jrecord_leaf(jrec, JLEAF_MTIME, &vat->va_mtime, sizeof(vat->va_mtime)); if (vat->va_ctime.tv_sec != VNOVAL) jrecord_leaf(jrec, JLEAF_CTIME, &vat->va_ctime, sizeof(vat->va_ctime)); if (vat->va_gen != VNOVAL) jrecord_leaf(jrec, JLEAF_GEN, &vat->va_gen, sizeof(vat->va_gen)); if (vat->va_flags != VNOVAL) jrecord_leaf(jrec, JLEAF_FLAGS, &vat->va_flags, sizeof(vat->va_flags)); if (vat->va_rmajor != VNOVAL) { dev_t rdev = makeudev(vat->va_rmajor, vat->va_rminor); jrecord_leaf(jrec, JLEAF_UDEV, &rdev, sizeof(rdev)); jrecord_leaf(jrec, JLEAF_UMAJOR, &vat->va_rmajor, sizeof(vat->va_rmajor)); jrecord_leaf(jrec, JLEAF_UMINOR, &vat->va_rminor, sizeof(vat->va_rminor)); } #if 0 if (vat->va_filerev != VNOVAL) jrecord_leaf(jrec, JLEAF_FILEREV, &vat->va_filerev, sizeof(vat->va_filerev)); #endif jrecord_pop(jrec, save); } /* * Write out the creds used to issue a file operation. If a process is * available write out additional tracking information related to the * process. * * XXX additional tracking info * XXX tty line info */ void jrecord_write_cred(struct jrecord *jrec, struct thread *td, struct ucred *cred) { void *save; struct proc *p; save = jrecord_push(jrec, JTYPE_CRED); jrecord_leaf(jrec, JLEAF_UID, &cred->cr_uid, sizeof(cred->cr_uid)); jrecord_leaf(jrec, JLEAF_GID, &cred->cr_gid, sizeof(cred->cr_gid)); if (td && (p = td->td_proc) != NULL) { jrecord_leaf(jrec, JLEAF_PID, &p->p_pid, sizeof(p->p_pid)); jrecord_leaf(jrec, JLEAF_COMM, p->p_comm, sizeof(p->p_comm)); } jrecord_pop(jrec, save); } /* * Write out information required to identify a vnode * * XXX this needs work. We should write out the inode number as well, * and in fact avoid writing out the file path for seqential writes * occuring within e.g. a certain period of time. */ void jrecord_write_vnode_ref(struct jrecord *jrec, struct vnode *vp) { struct nchandle nch; nch.mount = vp->v_mount; spin_lock(&vp->v_spin); TAILQ_FOREACH(nch.ncp, &vp->v_namecache, nc_vnode) { if ((nch.ncp->nc_flag & (NCF_UNRESOLVED|NCF_DESTROYED)) == 0) break; } if (nch.ncp) { cache_hold(&nch); spin_unlock(&vp->v_spin); jrecord_write_path(jrec, JLEAF_PATH_REF, nch.ncp); cache_drop(&nch); } else { spin_unlock(&vp->v_spin); } } void jrecord_write_vnode_link(struct jrecord *jrec, struct vnode *vp, struct namecache *notncp) { struct nchandle nch; nch.mount = vp->v_mount; spin_lock(&vp->v_spin); TAILQ_FOREACH(nch.ncp, &vp->v_namecache, nc_vnode) { if (nch.ncp == notncp) continue; if ((nch.ncp->nc_flag & (NCF_UNRESOLVED|NCF_DESTROYED)) == 0) break; } if (nch.ncp) { cache_hold(&nch); spin_unlock(&vp->v_spin); jrecord_write_path(jrec, JLEAF_PATH_REF, nch.ncp); cache_drop(&nch); } else { spin_unlock(&vp->v_spin); } } /* * Write out the data represented by a pagelist */ void jrecord_write_pagelist(struct jrecord *jrec, int16_t rectype, struct vm_page **pglist, int *rtvals, int pgcount, off_t offset) { struct xio xio; int error; int b; int i; i = 0; xio_init(&xio); while (i < pgcount) { /* * Find the next valid section. Skip any invalid elements */ if (rtvals[i] != VM_PAGER_OK) { ++i; offset += PAGE_SIZE; continue; } /* * Figure out how big the valid section is, capping I/O at what the * MSFBUF can represent. */ b = i; while (i < pgcount && i - b != XIO_INTERNAL_PAGES && rtvals[i] == VM_PAGER_OK ) { ++i; } /* * And write it out. */ if (i - b) { error = xio_init_pages(&xio, pglist + b, i - b, XIOF_READ); if (error == 0) { jrecord_leaf(jrec, JLEAF_SEEKPOS, &offset, sizeof(offset)); jrecord_leaf_xio(jrec, rectype, &xio); } else { kprintf("jrecord_write_pagelist: xio init failure\n"); } xio_release(&xio); offset += (off_t)(i - b) << PAGE_SHIFT; } } } /* * Write out the data represented by a UIO. */ void jrecord_write_uio(struct jrecord *jrec, int16_t rectype, struct uio *uio) { if (uio->uio_segflg != UIO_NOCOPY) { jrecord_leaf(jrec, JLEAF_SEEKPOS, &uio->uio_offset, sizeof(uio->uio_offset)); jrecord_leaf_uio(jrec, rectype, uio); } } void jrecord_file_data(struct jrecord *jrec, struct vnode *vp, off_t off, off_t bytes) { const int bufsize = 8192; char *buf; int error; int n; buf = kmalloc(bufsize, M_JOURNAL, M_WAITOK); jrecord_leaf(jrec, JLEAF_SEEKPOS, &off, sizeof(off)); while (bytes) { n = (bytes > bufsize) ? bufsize : (int)bytes; error = vn_rdwr(UIO_READ, vp, buf, n, off, UIO_SYSSPACE, IO_NODELOCKED, proc0.p_ucred, NULL); if (error) { jrecord_leaf(jrec, JLEAF_ERROR, &error, sizeof(error)); break; } jrecord_leaf(jrec, JLEAF_FILEDATA, buf, n); bytes -= n; off += n; } kfree(buf, M_JOURNAL); }
