2 * Copyright (c) 2011-2014 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
18 * 3. Neither the name of The DragonFly Project nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific, prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/nlookup.h>
39 #include <sys/vnode.h>
40 #include <sys/mount.h>
41 #include <sys/fcntl.h>
44 #include <sys/vfsops.h>
45 #include <sys/sysctl.h>
46 #include <sys/socket.h>
47 #include <sys/objcache.h>
50 #include <sys/namei.h>
51 #include <sys/mountctl.h>
52 #include <sys/dirent.h>
55 #include <sys/mutex.h>
56 #include <sys/mutex2.h>
59 #include "hammer2_disk.h"
60 #include "hammer2_mount.h"
63 #include "hammer2_lz4.h"
65 #include "zlib/hammer2_zlib.h"
67 #define REPORT_REFS_ERRORS 1 /* XXX remove me */
69 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
71 struct hammer2_sync_info {
72 hammer2_trans_t trans;
77 TAILQ_HEAD(hammer2_mntlist, hammer2_mount);
78 TAILQ_HEAD(hammer2_pfslist, hammer2_pfsmount);
79 static struct hammer2_mntlist hammer2_mntlist;
80 static struct hammer2_pfslist hammer2_pfslist;
81 static struct lock hammer2_mntlk;
84 int hammer2_cluster_enable = 1;
85 int hammer2_hardlink_enable = 1;
86 int hammer2_flush_pipe = 100;
87 int hammer2_synchronous_flush = 1;
88 int hammer2_dio_count;
89 long hammer2_limit_dirty_chains;
90 long hammer2_iod_file_read;
91 long hammer2_iod_meta_read;
92 long hammer2_iod_indr_read;
93 long hammer2_iod_fmap_read;
94 long hammer2_iod_volu_read;
95 long hammer2_iod_file_write;
96 long hammer2_iod_meta_write;
97 long hammer2_iod_indr_write;
98 long hammer2_iod_fmap_write;
99 long hammer2_iod_volu_write;
100 long hammer2_ioa_file_read;
101 long hammer2_ioa_meta_read;
102 long hammer2_ioa_indr_read;
103 long hammer2_ioa_fmap_read;
104 long hammer2_ioa_volu_read;
105 long hammer2_ioa_fmap_write;
106 long hammer2_ioa_file_write;
107 long hammer2_ioa_meta_write;
108 long hammer2_ioa_indr_write;
109 long hammer2_ioa_volu_write;
111 MALLOC_DECLARE(C_BUFFER);
112 MALLOC_DEFINE(C_BUFFER, "compbuffer", "Buffer used for compression.");
114 MALLOC_DECLARE(D_BUFFER);
115 MALLOC_DEFINE(D_BUFFER, "decompbuffer", "Buffer used for decompression.");
117 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
120 &hammer2_debug, 0, "");
121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_enable, CTLFLAG_RW,
122 &hammer2_cluster_enable, 0, "");
123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, hardlink_enable, CTLFLAG_RW,
124 &hammer2_hardlink_enable, 0, "");
125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
126 &hammer2_flush_pipe, 0, "");
127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
128 &hammer2_synchronous_flush, 0, "");
129 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
130 &hammer2_limit_dirty_chains, 0, "");
131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
132 &hammer2_dio_count, 0, "");
134 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
135 &hammer2_iod_file_read, 0, "");
136 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
137 &hammer2_iod_meta_read, 0, "");
138 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
139 &hammer2_iod_indr_read, 0, "");
140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
141 &hammer2_iod_fmap_read, 0, "");
142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
143 &hammer2_iod_volu_read, 0, "");
145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
146 &hammer2_iod_file_write, 0, "");
147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
148 &hammer2_iod_meta_write, 0, "");
149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
150 &hammer2_iod_indr_write, 0, "");
151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
152 &hammer2_iod_fmap_write, 0, "");
153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
154 &hammer2_iod_volu_write, 0, "");
156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_read, CTLFLAG_RW,
157 &hammer2_ioa_file_read, 0, "");
158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_read, CTLFLAG_RW,
159 &hammer2_ioa_meta_read, 0, "");
160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_read, CTLFLAG_RW,
161 &hammer2_ioa_indr_read, 0, "");
162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_read, CTLFLAG_RW,
163 &hammer2_ioa_fmap_read, 0, "");
164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_read, CTLFLAG_RW,
165 &hammer2_ioa_volu_read, 0, "");
167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_write, CTLFLAG_RW,
168 &hammer2_ioa_file_write, 0, "");
169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_write, CTLFLAG_RW,
170 &hammer2_ioa_meta_write, 0, "");
171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_write, CTLFLAG_RW,
172 &hammer2_ioa_indr_write, 0, "");
173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_write, CTLFLAG_RW,
174 &hammer2_ioa_fmap_write, 0, "");
175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_write, CTLFLAG_RW,
176 &hammer2_ioa_volu_write, 0, "");
178 static int hammer2_vfs_init(struct vfsconf *conf);
179 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
180 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
182 static int hammer2_remount(hammer2_mount_t *, struct mount *, char *,
183 struct vnode *, struct ucred *);
184 static int hammer2_recovery(hammer2_mount_t *hmp);
185 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
186 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
187 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
189 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
191 static int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
192 ino_t ino, struct vnode **vpp);
193 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
194 struct fid *fhp, struct vnode **vpp);
195 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
196 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
197 int *exflagsp, struct ucred **credanonp);
199 static int hammer2_install_volume_header(hammer2_mount_t *hmp);
200 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
202 static void hammer2_write_thread(void *arg);
204 static void hammer2_vfs_unmount_hmp1(struct mount *mp, hammer2_mount_t *hmp);
205 static void hammer2_vfs_unmount_hmp2(struct mount *mp, hammer2_mount_t *hmp);
208 * Functions for compression in threads,
209 * from hammer2_vnops.c
211 static void hammer2_write_file_core(struct buf *bp, hammer2_trans_t *trans,
213 const hammer2_inode_data_t *ripdata,
214 hammer2_cluster_t *cparent,
215 hammer2_key_t lbase, int ioflag, int pblksize,
217 static void hammer2_compress_and_write(struct buf *bp, hammer2_trans_t *trans,
219 const hammer2_inode_data_t *ripdata,
220 hammer2_cluster_t *cparent,
221 hammer2_key_t lbase, int ioflag,
222 int pblksize, int *errorp,
223 int comp_algo, int check_algo);
224 static void hammer2_zero_check_and_write(struct buf *bp,
225 hammer2_trans_t *trans, hammer2_inode_t *ip,
226 const hammer2_inode_data_t *ripdata,
227 hammer2_cluster_t *cparent,
229 int ioflag, int pblksize, int *errorp,
231 static int test_block_zeros(const char *buf, size_t bytes);
232 static void zero_write(struct buf *bp, hammer2_trans_t *trans,
234 const hammer2_inode_data_t *ripdata,
235 hammer2_cluster_t *cparent,
238 static void hammer2_write_bp(hammer2_cluster_t *cluster, struct buf *bp,
239 int ioflag, int pblksize, int *errorp,
242 static int hammer2_rcvdmsg(kdmsg_msg_t *msg);
243 static void hammer2_autodmsg(kdmsg_msg_t *msg);
244 static int hammer2_lnk_span_reply(kdmsg_state_t *state, kdmsg_msg_t *msg);
248 * HAMMER2 vfs operations.
250 static struct vfsops hammer2_vfsops = {
251 .vfs_init = hammer2_vfs_init,
252 .vfs_uninit = hammer2_vfs_uninit,
253 .vfs_sync = hammer2_vfs_sync,
254 .vfs_mount = hammer2_vfs_mount,
255 .vfs_unmount = hammer2_vfs_unmount,
256 .vfs_root = hammer2_vfs_root,
257 .vfs_statfs = hammer2_vfs_statfs,
258 .vfs_statvfs = hammer2_vfs_statvfs,
259 .vfs_vget = hammer2_vfs_vget,
260 .vfs_vptofh = hammer2_vfs_vptofh,
261 .vfs_fhtovp = hammer2_vfs_fhtovp,
262 .vfs_checkexp = hammer2_vfs_checkexp
265 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
267 VFS_SET(hammer2_vfsops, hammer2, 0);
268 MODULE_VERSION(hammer2, 1);
272 hammer2_vfs_init(struct vfsconf *conf)
274 static struct objcache_malloc_args margs_read;
275 static struct objcache_malloc_args margs_write;
281 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
283 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
285 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
289 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
291 margs_read.objsize = 65536;
292 margs_read.mtype = D_BUFFER;
294 margs_write.objsize = 32768;
295 margs_write.mtype = C_BUFFER;
297 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
298 0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
299 objcache_malloc_free, &margs_read);
300 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
301 0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
302 objcache_malloc_free, &margs_write);
304 lockinit(&hammer2_mntlk, "mntlk", 0, 0);
305 TAILQ_INIT(&hammer2_mntlist);
306 TAILQ_INIT(&hammer2_pfslist);
308 hammer2_limit_dirty_chains = desiredvnodes / 10;
310 hammer2_trans_manage_init();
317 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
319 objcache_destroy(cache_buffer_read);
320 objcache_destroy(cache_buffer_write);
325 * Core PFS allocator. Used to allocate the pmp structure for PFS cluster
326 * mounts and the spmp structure for media (hmp) structures.
328 static hammer2_pfsmount_t *
329 hammer2_pfsalloc(const hammer2_inode_data_t *ripdata, hammer2_tid_t alloc_tid)
331 hammer2_pfsmount_t *pmp;
333 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
334 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
335 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
336 lockinit(&pmp->lock, "pfslk", 0, 0);
337 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
338 RB_INIT(&pmp->inum_tree);
339 TAILQ_INIT(&pmp->unlinkq);
340 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
342 pmp->alloc_tid = alloc_tid + 1; /* our first media transaction id */
343 pmp->flush_tid = pmp->alloc_tid;
345 pmp->inode_tid = ripdata->pfs_inum + 1;
346 pmp->pfs_clid = ripdata->pfs_clid;
348 mtx_init(&pmp->wthread_mtx);
349 bioq_init(&pmp->wthread_bioq);
355 * Mount or remount HAMMER2 fileystem from physical media
358 * mp mount point structure
364 * mp mount point structure
365 * path path to mount point
366 * data pointer to argument structure in user space
367 * volume volume path (device@LABEL form)
368 * hflags user mount flags
369 * cred user credentials
376 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
379 struct hammer2_mount_info info;
380 hammer2_pfsmount_t *pmp;
381 hammer2_pfsmount_t *spmp;
382 hammer2_mount_t *hmp;
383 hammer2_key_t key_next;
384 hammer2_key_t key_dummy;
387 struct nlookupdata nd;
388 hammer2_chain_t *parent;
389 hammer2_chain_t *rchain;
390 hammer2_cluster_t *cluster;
391 hammer2_cluster_t *cparent;
392 const hammer2_inode_data_t *ripdata;
393 hammer2_blockref_t bref;
395 char devstr[MNAMELEN];
413 kprintf("hammer2_mount\n");
419 bzero(&info, sizeof(info));
420 info.cluster_fd = -1;
424 * Non-root mount or updating a mount
426 error = copyin(data, &info, sizeof(info));
430 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
434 /* Extract device and label */
436 label = strchr(devstr, '@');
438 ((label + 1) - dev) > done) {
446 if (mp->mnt_flag & MNT_UPDATE) {
448 /* HAMMER2 implements NFS export via mountctl */
450 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
451 hmp = pmp->iroot->cluster.array[i]->hmp;
453 error = hammer2_remount(hmp, mp, path,
458 /*hammer2_inode_install_hidden(pmp);*/
467 * Lookup name and verify it refers to a block device.
469 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
471 error = nlookup(&nd);
473 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
477 if (vn_isdisk(devvp, &error))
478 error = vfs_mountedon(devvp);
482 * Determine if the device has already been mounted. After this
483 * check hmp will be non-NULL if we are doing the second or more
484 * hammer2 mounts from the same device.
486 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
487 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
488 if (hmp->devvp == devvp)
493 * Open the device if this isn't a secondary mount and construct
494 * the H2 device mount (hmp).
497 hammer2_chain_t *schain;
500 if (error == 0 && vcount(devvp) > 0)
504 * Now open the device
507 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
508 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
509 error = vinvalbuf(devvp, V_SAVE, 0, 0);
511 error = VOP_OPEN(devvp,
512 ronly ? FREAD : FREAD | FWRITE,
517 if (error && devvp) {
522 lockmgr(&hammer2_mntlk, LK_RELEASE);
525 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
528 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
529 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
530 RB_INIT(&hmp->iotree);
531 spin_init(&hmp->io_spin, "hm2mount_io");
532 spin_init(&hmp->list_spin, "hm2mount_list");
533 TAILQ_INIT(&hmp->flushq);
535 lockinit(&hmp->vollk, "h2vol", 0, 0);
538 * vchain setup. vchain.data is embedded.
539 * vchain.refs is initialized and will never drop to 0.
541 * NOTE! voldata is not yet loaded.
543 hmp->vchain.hmp = hmp;
544 hmp->vchain.refs = 1;
545 hmp->vchain.data = (void *)&hmp->voldata;
546 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
547 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
548 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
550 hammer2_chain_core_alloc(NULL, &hmp->vchain);
551 /* hmp->vchain.u.xxx is left NULL */
554 * fchain setup. fchain.data is embedded.
555 * fchain.refs is initialized and will never drop to 0.
557 * The data is not used but needs to be initialized to
558 * pass assertion muster. We use this chain primarily
559 * as a placeholder for the freemap's top-level RBTREE
560 * so it does not interfere with the volume's topology
563 hmp->fchain.hmp = hmp;
564 hmp->fchain.refs = 1;
565 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
566 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
567 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
568 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
569 hmp->fchain.bref.methods =
570 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
571 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
573 hammer2_chain_core_alloc(NULL, &hmp->fchain);
574 /* hmp->fchain.u.xxx is left NULL */
577 * Install the volume header and initialize fields from
580 error = hammer2_install_volume_header(hmp);
583 hammer2_vfs_unmount_hmp1(mp, hmp);
584 hammer2_vfs_unmount_hmp2(mp, hmp);
585 lockmgr(&hammer2_mntlk, LK_RELEASE);
586 hammer2_vfs_unmount(mp, MNT_FORCE);
591 * Really important to get these right or flush will get
594 hmp->spmp = hammer2_pfsalloc(NULL, hmp->voldata.mirror_tid);
595 kprintf("alloc spmp %p tid %016jx\n",
596 hmp->spmp, hmp->voldata.mirror_tid);
601 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
602 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
603 hmp->vchain.pmp = spmp;
604 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
605 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
606 hmp->fchain.pmp = spmp;
609 * First locate the super-root inode, which is key 0
610 * relative to the volume header's blockset.
612 * Then locate the root inode by scanning the directory keyspace
613 * represented by the label.
615 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
616 schain = hammer2_chain_lookup(&parent, &key_dummy,
617 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
618 &cache_index, 0, &ddflag);
619 hammer2_chain_lookup_done(parent);
620 if (schain == NULL) {
621 kprintf("hammer2_mount: invalid super-root\n");
623 hammer2_vfs_unmount_hmp1(mp, hmp);
624 hammer2_vfs_unmount_hmp2(mp, hmp);
625 lockmgr(&hammer2_mntlk, LK_RELEASE);
626 hammer2_vfs_unmount(mp, MNT_FORCE);
631 * Sanity-check schain's pmp, finish initializing spmp.
633 ripdata = &hammer2_chain_rdata(schain)->ipdata;
634 KKASSERT(schain->pmp == spmp);
635 spmp->pfs_clid = ripdata->pfs_clid;
638 * NOTE: inode_get sucks up schain's lock.
640 cluster = hammer2_cluster_from_chain(schain);
641 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster);
642 spmp->spmp_hmp = hmp;
643 hammer2_inode_ref(spmp->iroot);
644 hammer2_inode_unlock_ex(spmp->iroot, cluster);
646 /* leave spmp->iroot with one ref */
648 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
649 error = hammer2_recovery(hmp);
650 /* XXX do something with error */
655 * XXX RDONLY stuff is totally broken FIXME XXX
658 * Automatic handling of received LNK_SPAN
659 * No automatic LNK_SPAN generation - we do this ourselves
661 kdmsg_iocom_init(&hmp->iocom, hmp,
662 KDMSG_IOCOMF_AUTOCONN |
663 KDMSG_IOCOMF_AUTORXSPAN,
664 hmp->mchain, hammer2_rcvdmsg);
667 * Ref the cluster management messaging descriptor. The mount
668 * program deals with the other end of the communications pipe.
670 fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
672 hammer2_cluster_reconnect(hmp, fp);
674 kprintf("hammer2_mount: bad cluster_fd!\n");
682 * Lookup mount point under the media-localized super-root.
684 * cluster->pmp will incorrectly point to spmp and must be fixed
687 cparent = hammer2_inode_lock_ex(spmp->iroot);
688 lhc = hammer2_dirhash(label, strlen(label));
689 cluster = hammer2_cluster_lookup(cparent, &key_next,
690 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
693 if (hammer2_cluster_type(cluster) == HAMMER2_BREF_TYPE_INODE &&
695 hammer2_cluster_rdata(cluster)->ipdata.filename) == 0) {
698 cluster = hammer2_cluster_next(cparent, cluster, &key_next,
700 lhc + HAMMER2_DIRHASH_LOMASK, 0);
702 hammer2_inode_unlock_ex(spmp->iroot, cparent);
704 if (cluster == NULL) {
705 kprintf("hammer2_mount: PFS label not found\n");
706 hammer2_vfs_unmount_hmp1(mp, hmp);
707 hammer2_vfs_unmount_hmp2(mp, hmp);
708 lockmgr(&hammer2_mntlk, LK_RELEASE);
709 hammer2_vfs_unmount(mp, MNT_FORCE);
713 for (i = 0; i < cluster->nchains; ++i) {
714 rchain = cluster->array[i];
715 if (rchain->flags & HAMMER2_CHAIN_MOUNTED) {
716 kprintf("hammer2_mount: PFS label already mounted!\n");
717 hammer2_cluster_unlock(cluster);
718 hammer2_vfs_unmount_hmp1(mp, hmp);
719 hammer2_vfs_unmount_hmp2(mp, hmp);
720 lockmgr(&hammer2_mntlk, LK_RELEASE);
721 hammer2_vfs_unmount(mp, MNT_FORCE);
724 KKASSERT(rchain->pmp == NULL);
726 if (rchain->flags & HAMMER2_CHAIN_RECYCLE) {
727 kprintf("hammer2_mount: PFS label is recycling\n");
728 hammer2_cluster_unlock(cluster);
729 hammer2_vfs_unmount_hmp1(mp, hmp);
730 hammer2_vfs_unmount_hmp2(mp, hmp);
731 lockmgr(&hammer2_mntlk, LK_RELEASE);
732 hammer2_vfs_unmount(mp, MNT_FORCE);
739 * Check to see if the cluster id is already mounted at the mount
740 * point. If it is, add us to the cluster.
742 ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
743 hammer2_cluster_bref(cluster, &bref);
744 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
745 if (pmp->spmp_hmp == NULL &&
746 bcmp(&pmp->pfs_clid, &ripdata->pfs_clid,
747 sizeof(pmp->pfs_clid)) == 0) {
756 hammer2_inode_ref(pmp->iroot);
757 ccms_thread_lock(&pmp->iroot->topo_cst, CCMS_STATE_EXCLUSIVE);
759 if (pmp->iroot->cluster.nchains + cluster->nchains >
760 HAMMER2_MAXCLUSTER) {
761 kprintf("hammer2_mount: cluster full!\n");
763 ccms_thread_unlock(&pmp->iroot->topo_cst);
764 hammer2_inode_drop(pmp->iroot);
766 hammer2_cluster_unlock(cluster);
767 hammer2_vfs_unmount_hmp1(mp, hmp);
768 hammer2_vfs_unmount_hmp2(mp, hmp);
769 lockmgr(&hammer2_mntlk, LK_RELEASE);
770 hammer2_vfs_unmount(mp, MNT_FORCE);
773 kprintf("hammer2_vfs_mount: Adding pfs to existing cluster\n");
774 j = pmp->iroot->cluster.nchains;
775 for (i = 0; i < cluster->nchains; ++i) {
776 rchain = cluster->array[i];
777 KKASSERT(rchain->pmp == NULL);
779 hammer2_chain_ref(cluster->array[i]);
780 pmp->iroot->cluster.array[j] = cluster->array[i];
783 pmp->iroot->cluster.nchains = j;
784 ccms_thread_unlock(&pmp->iroot->topo_cst);
785 hammer2_inode_drop(pmp->iroot);
786 hammer2_cluster_unlock(cluster);
787 lockmgr(&hammer2_mntlk, LK_RELEASE);
790 hammer2_inode_install_hidden(pmp);
796 * Block device opened successfully, finish initializing the
799 * From this point on we have to call hammer2_unmount() on failure.
801 pmp = hammer2_pfsalloc(ripdata, bref.mirror_tid);
802 kprintf("PMP mirror_tid is %016jx\n", bref.mirror_tid);
803 for (i = 0; i < cluster->nchains; ++i) {
804 rchain = cluster->array[i];
805 KKASSERT(rchain->pmp == NULL);
807 atomic_set_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
811 ccms_domain_init(&pmp->ccms_dom);
812 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
813 lockmgr(&hammer2_mntlk, LK_RELEASE);
815 kprintf("hammer2_mount hmp=%p pmp=%p pmpcnt=%d\n",
816 hmp, pmp, hmp->pmp_count);
818 mp->mnt_flag = MNT_LOCAL;
819 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */
820 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */
823 * required mount structure initializations
825 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
826 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
828 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
829 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
834 mp->mnt_iosize_max = MAXPHYS;
835 mp->mnt_data = (qaddr_t)pmp;
839 * After this point hammer2_vfs_unmount() has visibility on hmp
840 * and manual hmp1/hmp2 calls are not needed on fatal errors.
842 pmp->iroot = hammer2_inode_get(pmp, NULL, cluster);
843 hammer2_inode_ref(pmp->iroot); /* ref for pmp->iroot */
844 hammer2_inode_unlock_ex(pmp->iroot, cluster);
847 * The logical file buffer bio write thread handles things
848 * like physical block assignment and compression.
850 * (only applicable to pfs mounts, not applicable to spmp)
852 pmp->wthread_destroy = 0;
853 lwkt_create(hammer2_write_thread, pmp,
854 &pmp->wthread_td, NULL, 0, -1, "hwrite-%s", label);
857 * With the cluster operational install ihidden.
858 * (only applicable to pfs mounts, not applicable to spmp)
860 hammer2_inode_install_hidden(pmp);
866 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
867 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
868 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
870 copyinstr(info.volume, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, &size);
871 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
872 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
873 copyinstr(path, mp->mnt_stat.f_mntonname,
874 sizeof(mp->mnt_stat.f_mntonname) - 1,
878 * Initial statfs to prime mnt_stat.
880 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
886 * Handle bioq for strategy write
890 hammer2_write_thread(void *arg)
892 hammer2_pfsmount_t *pmp;
895 hammer2_trans_t trans;
898 hammer2_cluster_t *cparent;
899 hammer2_inode_data_t *wipdata;
907 mtx_lock(&pmp->wthread_mtx);
908 while (pmp->wthread_destroy == 0) {
909 if (bioq_first(&pmp->wthread_bioq) == NULL) {
910 mtxsleep(&pmp->wthread_bioq, &pmp->wthread_mtx,
915 hammer2_trans_init(&trans, pmp, HAMMER2_TRANS_BUFCACHE);
917 while ((bio = bioq_takefirst(&pmp->wthread_bioq)) != NULL) {
919 * dummy bio for synchronization. The transaction
920 * must be reinitialized.
922 if (bio->bio_buf == NULL) {
923 bio->bio_flags |= BIO_DONE;
925 hammer2_trans_done(&trans);
926 hammer2_trans_init(&trans, pmp,
927 HAMMER2_TRANS_BUFCACHE);
932 * else normal bio processing
934 mtx_unlock(&pmp->wthread_mtx);
936 hammer2_lwinprog_drop(pmp);
944 * Inode is modified, flush size and mtime changes
945 * to ensure that the file size remains consistent
946 * with the buffers being flushed.
948 * NOTE: The inode_fsync() call only flushes the
949 * inode's meta-data state, it doesn't try
950 * to flush underlying buffers or chains.
952 cparent = hammer2_inode_lock_ex(ip);
953 if (ip->flags & (HAMMER2_INODE_RESIZED |
954 HAMMER2_INODE_MTIME)) {
955 hammer2_inode_fsync(&trans, ip, cparent);
957 wipdata = hammer2_cluster_modify_ip(&trans, ip,
959 lblksize = hammer2_calc_logical(ip, bio->bio_offset,
961 pblksize = hammer2_calc_physical(ip, wipdata, lbase);
962 hammer2_write_file_core(bp, &trans, ip, wipdata,
966 hammer2_cluster_modsync(cparent);
967 hammer2_inode_unlock_ex(ip, cparent);
969 kprintf("hammer2: error in buffer write\n");
970 bp->b_flags |= B_ERROR;
974 mtx_lock(&pmp->wthread_mtx);
976 hammer2_trans_done(&trans);
978 pmp->wthread_destroy = -1;
979 wakeup(&pmp->wthread_destroy);
981 mtx_unlock(&pmp->wthread_mtx);
985 hammer2_bioq_sync(hammer2_pfsmount_t *pmp)
989 bzero(&sync_bio, sizeof(sync_bio)); /* dummy with no bio_buf */
990 mtx_lock(&pmp->wthread_mtx);
991 if (pmp->wthread_destroy == 0 &&
992 TAILQ_FIRST(&pmp->wthread_bioq.queue)) {
993 bioq_insert_tail(&pmp->wthread_bioq, &sync_bio);
994 while ((sync_bio.bio_flags & BIO_DONE) == 0)
995 mtxsleep(&sync_bio, &pmp->wthread_mtx, 0, "h2bioq", 0);
997 mtx_unlock(&pmp->wthread_mtx);
1001 * Return a chain suitable for I/O, creating the chain if necessary
1002 * and assigning its physical block.
1006 hammer2_assign_physical(hammer2_trans_t *trans,
1007 hammer2_inode_t *ip, hammer2_cluster_t *cparent,
1008 hammer2_key_t lbase, int pblksize, int *errorp)
1010 hammer2_cluster_t *cluster;
1011 hammer2_cluster_t *dparent;
1012 hammer2_key_t key_dummy;
1013 int pradix = hammer2_getradix(pblksize);
1017 * Locate the chain associated with lbase, return a locked chain.
1018 * However, do not instantiate any data reference (which utilizes a
1019 * device buffer) because we will be using direct IO via the
1020 * logical buffer cache buffer.
1023 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
1025 dparent = hammer2_cluster_lookup_init(cparent, 0);
1026 cluster = hammer2_cluster_lookup(dparent, &key_dummy,
1028 HAMMER2_LOOKUP_NODATA, &ddflag);
1030 if (cluster == NULL) {
1032 * We found a hole, create a new chain entry.
1034 * NOTE: DATA chains are created without device backing
1035 * store (nor do we want any).
1037 *errorp = hammer2_cluster_create(trans, dparent, &cluster,
1038 lbase, HAMMER2_PBUFRADIX,
1039 HAMMER2_BREF_TYPE_DATA,
1041 if (cluster == NULL) {
1042 hammer2_cluster_lookup_done(dparent);
1043 panic("hammer2_cluster_create: par=%p error=%d\n",
1044 dparent->focus, *errorp);
1047 /*ip->delta_dcount += pblksize;*/
1049 switch (hammer2_cluster_type(cluster)) {
1050 case HAMMER2_BREF_TYPE_INODE:
1052 * The data is embedded in the inode. The
1053 * caller is responsible for marking the inode
1054 * modified and copying the data to the embedded
1058 case HAMMER2_BREF_TYPE_DATA:
1059 if (hammer2_cluster_need_resize(cluster, pblksize)) {
1060 hammer2_cluster_resize(trans, ip,
1063 HAMMER2_MODIFY_OPTDATA);
1067 * DATA buffers must be marked modified whether the
1068 * data is in a logical buffer or not. We also have
1069 * to make this call to fixup the chain data pointers
1070 * after resizing in case this is an encrypted or
1071 * compressed buffer.
1073 hammer2_cluster_modify(trans, cluster,
1074 HAMMER2_MODIFY_OPTDATA);
1077 panic("hammer2_assign_physical: bad type");
1084 * Cleanup. If cluster wound up being the inode itself, i.e.
1085 * the DIRECTDATA case for offset 0, then we need to update cparent.
1086 * The caller expects cparent to not become stale.
1088 hammer2_cluster_lookup_done(dparent);
1089 /* dparent = NULL; safety */
1090 if (cluster && ddflag)
1091 hammer2_cluster_replace_locked(cparent, cluster);
1096 * bio queued from hammer2_vnops.c.
1098 * The core write function which determines which path to take
1099 * depending on compression settings. We also have to locate the
1100 * related clusters so we can calculate and set the check data for
1105 hammer2_write_file_core(struct buf *bp, hammer2_trans_t *trans,
1106 hammer2_inode_t *ip,
1107 const hammer2_inode_data_t *ripdata,
1108 hammer2_cluster_t *cparent,
1109 hammer2_key_t lbase, int ioflag, int pblksize,
1112 hammer2_cluster_t *cluster;
1114 switch(HAMMER2_DEC_ALGO(ripdata->comp_algo)) {
1115 case HAMMER2_COMP_NONE:
1117 * We have to assign physical storage to the buffer
1118 * we intend to dirty or write now to avoid deadlocks
1119 * in the strategy code later.
1121 * This can return NOOFFSET for inode-embedded data.
1122 * The strategy code will take care of it in that case.
1124 cluster = hammer2_assign_physical(trans, ip, cparent,
1127 hammer2_write_bp(cluster, bp, ioflag, pblksize, errorp,
1128 ripdata->check_algo);
1130 hammer2_cluster_unlock(cluster);
1132 case HAMMER2_COMP_AUTOZERO:
1134 * Check for zero-fill only
1136 hammer2_zero_check_and_write(bp, trans, ip,
1137 ripdata, cparent, lbase,
1138 ioflag, pblksize, errorp,
1139 ripdata->check_algo);
1141 case HAMMER2_COMP_LZ4:
1142 case HAMMER2_COMP_ZLIB:
1145 * Check for zero-fill and attempt compression.
1147 hammer2_compress_and_write(bp, trans, ip,
1152 ripdata->check_algo);
1158 * Generic function that will perform the compression in compression
1159 * write path. The compression algorithm is determined by the settings
1160 * obtained from inode.
1164 hammer2_compress_and_write(struct buf *bp, hammer2_trans_t *trans,
1165 hammer2_inode_t *ip, const hammer2_inode_data_t *ripdata,
1166 hammer2_cluster_t *cparent,
1167 hammer2_key_t lbase, int ioflag, int pblksize,
1168 int *errorp, int comp_algo, int check_algo)
1170 hammer2_cluster_t *cluster;
1171 hammer2_chain_t *chain;
1173 int comp_block_size;
1177 if (test_block_zeros(bp->b_data, pblksize)) {
1178 zero_write(bp, trans, ip, ripdata, cparent, lbase, errorp);
1185 KKASSERT(pblksize / 2 <= 32768);
1187 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0) {
1188 z_stream strm_compress;
1192 switch(HAMMER2_DEC_ALGO(comp_algo)) {
1193 case HAMMER2_COMP_LZ4:
1194 comp_buffer = objcache_get(cache_buffer_write,
1196 comp_size = LZ4_compress_limitedOutput(
1198 &comp_buffer[sizeof(int)],
1200 pblksize / 2 - sizeof(int));
1202 * We need to prefix with the size, LZ4
1203 * doesn't do it for us. Add the related
1206 *(int *)comp_buffer = comp_size;
1208 comp_size += sizeof(int);
1210 case HAMMER2_COMP_ZLIB:
1211 comp_level = HAMMER2_DEC_LEVEL(comp_algo);
1212 if (comp_level == 0)
1213 comp_level = 6; /* default zlib compression */
1214 else if (comp_level < 6)
1216 else if (comp_level > 9)
1218 ret = deflateInit(&strm_compress, comp_level);
1220 kprintf("HAMMER2 ZLIB: fatal error "
1221 "on deflateInit.\n");
1224 comp_buffer = objcache_get(cache_buffer_write,
1226 strm_compress.next_in = bp->b_data;
1227 strm_compress.avail_in = pblksize;
1228 strm_compress.next_out = comp_buffer;
1229 strm_compress.avail_out = pblksize / 2;
1230 ret = deflate(&strm_compress, Z_FINISH);
1231 if (ret == Z_STREAM_END) {
1232 comp_size = pblksize / 2 -
1233 strm_compress.avail_out;
1237 ret = deflateEnd(&strm_compress);
1240 kprintf("Error: Unknown compression method.\n");
1241 kprintf("Comp_method = %d.\n", comp_algo);
1246 if (comp_size == 0) {
1248 * compression failed or turned off
1250 comp_block_size = pblksize; /* safety */
1251 if (++ip->comp_heuristic > 128)
1252 ip->comp_heuristic = 8;
1255 * compression succeeded
1257 ip->comp_heuristic = 0;
1258 if (comp_size <= 1024) {
1259 comp_block_size = 1024;
1260 } else if (comp_size <= 2048) {
1261 comp_block_size = 2048;
1262 } else if (comp_size <= 4096) {
1263 comp_block_size = 4096;
1264 } else if (comp_size <= 8192) {
1265 comp_block_size = 8192;
1266 } else if (comp_size <= 16384) {
1267 comp_block_size = 16384;
1268 } else if (comp_size <= 32768) {
1269 comp_block_size = 32768;
1271 panic("hammer2: WRITE PATH: "
1272 "Weird comp_size value.");
1274 comp_block_size = pblksize;
1278 cluster = hammer2_assign_physical(trans, ip, cparent,
1279 lbase, comp_block_size,
1284 kprintf("WRITE PATH: An error occurred while "
1285 "assigning physical space.\n");
1286 KKASSERT(cluster == NULL);
1290 for (i = 0; i < cluster->nchains; ++i) {
1291 hammer2_inode_data_t *wipdata;
1295 chain = cluster->array[i]; /* XXX */
1296 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1298 switch(chain->bref.type) {
1299 case HAMMER2_BREF_TYPE_INODE:
1300 wipdata = &hammer2_chain_wdata(chain)->ipdata;
1301 KKASSERT(wipdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1302 KKASSERT(bp->b_loffset == 0);
1303 bcopy(bp->b_data, wipdata->u.data,
1304 HAMMER2_EMBEDDED_BYTES);
1306 case HAMMER2_BREF_TYPE_DATA:
1308 * Optimize out the read-before-write
1311 *errorp = hammer2_io_newnz(chain->hmp,
1312 chain->bref.data_off,
1316 hammer2_io_brelse(&dio);
1317 kprintf("hammer2: WRITE PATH: "
1318 "dbp bread error\n");
1321 bdata = hammer2_io_data(dio, chain->bref.data_off);
1324 * When loading the block make sure we don't
1325 * leave garbage after the compressed data.
1328 chain->bref.methods =
1329 HAMMER2_ENC_COMP(comp_algo) +
1330 HAMMER2_ENC_CHECK(check_algo);
1331 bcopy(comp_buffer, bdata, comp_size);
1332 if (comp_size != comp_block_size) {
1333 bzero(bdata + comp_size,
1334 comp_block_size - comp_size);
1337 chain->bref.methods =
1339 HAMMER2_COMP_NONE) +
1340 HAMMER2_ENC_CHECK(check_algo);
1341 bcopy(bp->b_data, bdata, pblksize);
1345 * The flush code doesn't calculate check codes for
1346 * file data (doing so can result in excessive I/O),
1349 hammer2_chain_setcheck(chain, bdata);
1352 * Device buffer is now valid, chain is no longer in
1353 * the initial state.
1355 * (No blockref table worries with file data)
1357 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1359 /* Now write the related bdp. */
1360 if (ioflag & IO_SYNC) {
1362 * Synchronous I/O requested.
1364 hammer2_io_bwrite(&dio);
1366 } else if ((ioflag & IO_DIRECT) &&
1367 loff + n == pblksize) {
1368 hammer2_io_bdwrite(&dio);
1370 } else if (ioflag & IO_ASYNC) {
1371 hammer2_io_bawrite(&dio);
1373 hammer2_io_bdwrite(&dio);
1377 panic("hammer2_write_bp: bad chain type %d\n",
1385 hammer2_cluster_unlock(cluster);
1387 objcache_put(cache_buffer_write, comp_buffer);
1391 * Function that performs zero-checking and writing without compression,
1392 * it corresponds to default zero-checking path.
1396 hammer2_zero_check_and_write(struct buf *bp, hammer2_trans_t *trans,
1397 hammer2_inode_t *ip, const hammer2_inode_data_t *ripdata,
1398 hammer2_cluster_t *cparent,
1399 hammer2_key_t lbase, int ioflag, int pblksize, int *errorp,
1402 hammer2_cluster_t *cluster;
1404 if (test_block_zeros(bp->b_data, pblksize)) {
1405 zero_write(bp, trans, ip, ripdata, cparent, lbase, errorp);
1407 cluster = hammer2_assign_physical(trans, ip, cparent,
1408 lbase, pblksize, errorp);
1409 hammer2_write_bp(cluster, bp, ioflag, pblksize, errorp,
1412 hammer2_cluster_unlock(cluster);
1417 * A function to test whether a block of data contains only zeros,
1418 * returns TRUE (non-zero) if the block is all zeros.
1422 test_block_zeros(const char *buf, size_t bytes)
1426 for (i = 0; i < bytes; i += sizeof(long)) {
1427 if (*(const long *)(buf + i) != 0)
1434 * Function to "write" a block that contains only zeros.
1438 zero_write(struct buf *bp, hammer2_trans_t *trans,
1439 hammer2_inode_t *ip, const hammer2_inode_data_t *ripdata,
1440 hammer2_cluster_t *cparent,
1441 hammer2_key_t lbase, int *errorp __unused)
1443 hammer2_cluster_t *cluster;
1444 hammer2_media_data_t *data;
1445 hammer2_key_t key_dummy;
1448 cparent = hammer2_cluster_lookup_init(cparent, 0);
1449 cluster = hammer2_cluster_lookup(cparent, &key_dummy, lbase, lbase,
1450 HAMMER2_LOOKUP_NODATA, &ddflag);
1452 data = hammer2_cluster_wdata(cluster);
1455 KKASSERT(cluster->focus->flags &
1456 HAMMER2_CHAIN_MODIFIED);
1457 bzero(data->ipdata.u.data, HAMMER2_EMBEDDED_BYTES);
1458 hammer2_cluster_modsync(cluster);
1460 hammer2_cluster_delete(trans, cparent, cluster,
1461 HAMMER2_DELETE_PERMANENT);
1463 hammer2_cluster_unlock(cluster);
1465 hammer2_cluster_lookup_done(cparent);
1469 * Function to write the data as it is, without performing any sort of
1470 * compression. This function is used in path without compression and
1471 * default zero-checking path.
1475 hammer2_write_bp(hammer2_cluster_t *cluster, struct buf *bp, int ioflag,
1476 int pblksize, int *errorp, int check_algo)
1478 hammer2_chain_t *chain;
1479 hammer2_inode_data_t *wipdata;
1485 error = 0; /* XXX TODO below */
1487 for (i = 0; i < cluster->nchains; ++i) {
1488 chain = cluster->array[i]; /* XXX */
1489 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1491 switch(chain->bref.type) {
1492 case HAMMER2_BREF_TYPE_INODE:
1493 wipdata = &hammer2_chain_wdata(chain)->ipdata;
1494 KKASSERT(wipdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1495 KKASSERT(bp->b_loffset == 0);
1496 bcopy(bp->b_data, wipdata->u.data,
1497 HAMMER2_EMBEDDED_BYTES);
1500 case HAMMER2_BREF_TYPE_DATA:
1501 error = hammer2_io_newnz(chain->hmp,
1502 chain->bref.data_off,
1503 chain->bytes, &dio);
1505 hammer2_io_bqrelse(&dio);
1506 kprintf("hammer2: WRITE PATH: "
1507 "dbp bread error\n");
1510 bdata = hammer2_io_data(dio, chain->bref.data_off);
1512 chain->bref.methods = HAMMER2_ENC_COMP(
1513 HAMMER2_COMP_NONE) +
1514 HAMMER2_ENC_CHECK(check_algo);
1515 bcopy(bp->b_data, bdata, chain->bytes);
1518 * The flush code doesn't calculate check codes for
1519 * file data (doing so can result in excessive I/O),
1522 hammer2_chain_setcheck(chain, bdata);
1525 * Device buffer is now valid, chain is no longer in
1526 * the initial state.
1528 * (No blockref table worries with file data)
1530 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1532 if (ioflag & IO_SYNC) {
1534 * Synchronous I/O requested.
1536 hammer2_io_bwrite(&dio);
1538 } else if ((ioflag & IO_DIRECT) &&
1539 loff + n == pblksize) {
1540 hammer2_io_bdwrite(&dio);
1542 } else if (ioflag & IO_ASYNC) {
1543 hammer2_io_bawrite(&dio);
1545 hammer2_io_bdwrite(&dio);
1549 panic("hammer2_write_bp: bad chain type %d\n",
1555 KKASSERT(error == 0); /* XXX TODO */
1562 hammer2_remount(hammer2_mount_t *hmp, struct mount *mp, char *path,
1563 struct vnode *devvp, struct ucred *cred)
1567 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1568 error = hammer2_recovery(hmp);
1577 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1579 hammer2_pfsmount_t *pmp;
1580 hammer2_mount_t *hmp;
1581 hammer2_chain_t *rchain;
1582 hammer2_cluster_t *cluster;
1592 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1593 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
1596 * If mount initialization proceeded far enough we must flush
1599 if (mntflags & MNT_FORCE)
1604 error = vflush(mp, 0, flags);
1609 ccms_domain_uninit(&pmp->ccms_dom);
1611 if (pmp->wthread_td) {
1612 mtx_lock(&pmp->wthread_mtx);
1613 pmp->wthread_destroy = 1;
1614 wakeup(&pmp->wthread_bioq);
1615 while (pmp->wthread_destroy != -1) {
1616 mtxsleep(&pmp->wthread_destroy,
1617 &pmp->wthread_mtx, 0,
1620 mtx_unlock(&pmp->wthread_mtx);
1621 pmp->wthread_td = NULL;
1625 * Cleanup our reference on ihidden.
1628 hammer2_inode_drop(pmp->ihidden);
1629 pmp->ihidden = NULL;
1633 * Cleanup our reference on iroot. iroot is (should) not be needed
1634 * by the flush code.
1637 cluster = &pmp->iroot->cluster;
1638 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1639 rchain = pmp->iroot->cluster.array[i];
1643 hammer2_vfs_unmount_hmp1(mp, hmp);
1645 atomic_clear_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
1646 #if REPORT_REFS_ERRORS
1647 if (rchain->refs != 1)
1648 kprintf("PMP->RCHAIN %p REFS WRONG %d\n",
1649 rchain, rchain->refs);
1651 KKASSERT(rchain->refs == 1);
1653 hammer2_chain_drop(rchain);
1654 cluster->array[i] = NULL;
1655 hammer2_vfs_unmount_hmp2(mp, hmp);
1657 cluster->focus = NULL;
1659 #if REPORT_REFS_ERRORS
1660 if (pmp->iroot->refs != 1)
1661 kprintf("PMP->IROOT %p REFS WRONG %d\n",
1662 pmp->iroot, pmp->iroot->refs);
1664 KKASSERT(pmp->iroot->refs == 1);
1666 /* ref for pmp->iroot */
1667 hammer2_inode_drop(pmp->iroot);
1672 mp->mnt_data = NULL;
1674 kmalloc_destroy(&pmp->mmsg);
1675 kmalloc_destroy(&pmp->minode);
1677 kfree(pmp, M_HAMMER2);
1681 lockmgr(&hammer2_mntlk, LK_RELEASE);
1688 hammer2_vfs_unmount_hmp1(struct mount *mp, hammer2_mount_t *hmp)
1690 hammer2_mount_exlock(hmp);
1693 kprintf("hammer2_unmount hmp=%p pmpcnt=%d\n", hmp, hmp->pmp_count);
1695 kdmsg_iocom_uninit(&hmp->iocom); /* XXX chain depend deadlck? */
1698 * Cycle the volume data lock as a safety (probably not needed any
1699 * more). To ensure everything is out we need to flush at least
1700 * three times. (1) The running of the unlinkq can dirty the
1701 * filesystem, (2) A normal flush can dirty the freemap, and
1702 * (3) ensure that the freemap is fully synchronized.
1704 * The next mount's recovery scan can clean everything up but we want
1705 * to leave the filesystem in a 100% clean state on a normal unmount.
1707 hammer2_voldata_lock(hmp);
1708 hammer2_voldata_unlock(hmp);
1710 hammer2_vfs_sync(mp, MNT_WAIT);
1711 hammer2_vfs_sync(mp, MNT_WAIT);
1712 hammer2_vfs_sync(mp, MNT_WAIT);
1715 if (hmp->pmp_count == 0) {
1716 if ((hmp->vchain.flags | hmp->fchain.flags) &
1717 HAMMER2_CHAIN_FLUSH_MASK) {
1718 kprintf("hammer2_unmount: chains left over "
1719 "after final sync\n");
1720 kprintf(" vchain %08x\n", hmp->vchain.flags);
1721 kprintf(" fchain %08x\n", hmp->fchain.flags);
1723 if (hammer2_debug & 0x0010)
1724 Debugger("entered debugger");
1731 hammer2_vfs_unmount_hmp2(struct mount *mp, hammer2_mount_t *hmp)
1733 hammer2_pfsmount_t *spmp;
1734 struct vnode *devvp;
1736 int ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1739 * If no PFS's left drop the master hammer2_mount for the
1742 if (hmp->pmp_count == 0) {
1744 * Clean up SPMP and the super-root inode
1749 hammer2_inode_drop(spmp->iroot);
1753 kmalloc_destroy(&spmp->mmsg);
1754 kmalloc_destroy(&spmp->minode);
1755 kfree(spmp, M_HAMMER2);
1759 * Finish up with the device vnode
1761 if ((devvp = hmp->devvp) != NULL) {
1762 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1763 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1765 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1772 * Clear vchain/fchain flags that might prevent final cleanup
1775 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1776 atomic_clear_int(&hmp->vchain.flags,
1777 HAMMER2_CHAIN_MODIFIED);
1778 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1779 hammer2_chain_drop(&hmp->vchain);
1781 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1782 atomic_clear_int(&hmp->vchain.flags,
1783 HAMMER2_CHAIN_UPDATE);
1784 hammer2_chain_drop(&hmp->vchain);
1787 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1788 atomic_clear_int(&hmp->fchain.flags,
1789 HAMMER2_CHAIN_MODIFIED);
1790 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1791 hammer2_chain_drop(&hmp->fchain);
1793 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1794 atomic_clear_int(&hmp->fchain.flags,
1795 HAMMER2_CHAIN_UPDATE);
1796 hammer2_chain_drop(&hmp->fchain);
1800 * Final drop of embedded freemap root chain to
1801 * clean up fchain.core (fchain structure is not
1802 * flagged ALLOCATED so it is cleaned out and then
1805 hammer2_chain_drop(&hmp->fchain);
1808 * Final drop of embedded volume root chain to clean
1809 * up vchain.core (vchain structure is not flagged
1810 * ALLOCATED so it is cleaned out and then left to
1814 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1816 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1817 hammer2_mount_unlock(hmp);
1818 hammer2_chain_drop(&hmp->vchain);
1820 hammer2_io_cleanup(hmp, &hmp->iotree);
1821 if (hmp->iofree_count) {
1822 kprintf("io_cleanup: %d I/O's left hanging\n",
1826 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1827 kmalloc_destroy(&hmp->mchain);
1828 kfree(hmp, M_HAMMER2);
1830 hammer2_mount_unlock(hmp);
1836 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1837 ino_t ino, struct vnode **vpp)
1839 kprintf("hammer2_vget\n");
1840 return (EOPNOTSUPP);
1845 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1847 hammer2_pfsmount_t *pmp;
1848 hammer2_cluster_t *cparent;
1853 if (pmp->iroot == NULL) {
1857 cparent = hammer2_inode_lock_sh(pmp->iroot);
1858 vp = hammer2_igetv(pmp->iroot, cparent, &error);
1859 hammer2_inode_unlock_sh(pmp->iroot, cparent);
1862 kprintf("vnodefail\n");
1871 * XXX incorporate ipdata->inode_quota and data_quota
1875 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1877 hammer2_pfsmount_t *pmp;
1878 hammer2_mount_t *hmp;
1881 KKASSERT(pmp->iroot->cluster.nchains >= 1);
1882 hmp = pmp->iroot->cluster.focus->hmp; /* XXX */
1884 mp->mnt_stat.f_files = pmp->inode_count;
1885 mp->mnt_stat.f_ffree = 0;
1886 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1887 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1888 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1890 *sbp = mp->mnt_stat;
1896 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1898 hammer2_pfsmount_t *pmp;
1899 hammer2_mount_t *hmp;
1902 KKASSERT(pmp->iroot->cluster.nchains >= 1);
1903 hmp = pmp->iroot->cluster.focus->hmp; /* XXX */
1905 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1906 mp->mnt_vstat.f_files = pmp->inode_count;
1907 mp->mnt_vstat.f_ffree = 0;
1908 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1909 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1910 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1912 *sbp = mp->mnt_vstat;
1917 * Mount-time recovery (RW mounts)
1919 * Updates to the free block table are allowed to lag flushes by one
1920 * transaction. In case of a crash, then on a fresh mount we must do an
1921 * incremental scan of the last committed transaction id and make sure that
1922 * all related blocks have been marked allocated.
1924 * The super-root topology and each PFS has its own transaction id domain,
1925 * so we must track PFS boundary transitions.
1927 struct hammer2_recovery_elm {
1928 TAILQ_ENTRY(hammer2_recovery_elm) entry;
1929 hammer2_chain_t *chain;
1930 hammer2_tid_t sync_tid;
1933 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1935 struct hammer2_recovery_info {
1936 struct hammer2_recovery_list list;
1940 static int hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1941 hammer2_chain_t *parent,
1942 struct hammer2_recovery_info *info,
1943 hammer2_tid_t sync_tid);
1945 #define HAMMER2_RECOVERY_MAXDEPTH 10
1949 hammer2_recovery(hammer2_mount_t *hmp)
1951 hammer2_trans_t trans;
1952 struct hammer2_recovery_info info;
1953 struct hammer2_recovery_elm *elm;
1954 hammer2_chain_t *parent;
1955 hammer2_tid_t sync_tid;
1957 int cumulative_error = 0;
1959 hammer2_trans_init(&trans, hmp->spmp, 0);
1962 TAILQ_INIT(&info.list);
1964 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1965 cumulative_error = hammer2_recovery_scan(&trans, hmp, parent,
1967 hammer2_chain_lookup_done(parent);
1969 while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1970 TAILQ_REMOVE(&info.list, elm, entry);
1971 parent = elm->chain;
1972 sync_tid = elm->sync_tid;
1973 kfree(elm, M_HAMMER2);
1975 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
1976 HAMMER2_RESOLVE_NOREF);
1977 error = hammer2_recovery_scan(&trans, hmp, parent,
1979 hammer2_chain_unlock(parent);
1981 cumulative_error = error;
1983 hammer2_trans_done(&trans);
1985 return cumulative_error;
1990 hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1991 hammer2_chain_t *parent,
1992 struct hammer2_recovery_info *info,
1993 hammer2_tid_t sync_tid)
1995 const hammer2_inode_data_t *ripdata;
1996 hammer2_chain_t *chain;
1998 int cumulative_error = 0;
1999 int pfs_boundary = 0;
2003 * Adjust freemap to ensure that the block(s) are marked allocated.
2005 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2006 hammer2_freemap_adjust(trans, hmp, &parent->bref,
2007 HAMMER2_FREEMAP_DORECOVER);
2011 * Check type for recursive scan
2013 switch(parent->bref.type) {
2014 case HAMMER2_BREF_TYPE_VOLUME:
2015 /* data already instantiated */
2017 case HAMMER2_BREF_TYPE_INODE:
2019 * Must instantiate data for DIRECTDATA test and also
2022 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2023 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2024 if (ripdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2025 /* not applicable to recovery scan */
2026 hammer2_chain_unlock(parent);
2029 if ((ripdata->op_flags & HAMMER2_OPFLAG_PFSROOT) &&
2032 sync_tid = parent->bref.mirror_tid - 1;
2034 hammer2_chain_unlock(parent);
2036 case HAMMER2_BREF_TYPE_INDIRECT:
2038 * Must instantiate data for recursion
2040 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2041 hammer2_chain_unlock(parent);
2043 case HAMMER2_BREF_TYPE_DATA:
2044 case HAMMER2_BREF_TYPE_FREEMAP:
2045 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2046 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2047 /* not applicable to recovery scan */
2055 * Defer operation if depth limit reached or if we are crossing a
2058 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH || pfs_boundary) {
2059 struct hammer2_recovery_elm *elm;
2061 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2062 elm->chain = parent;
2063 elm->sync_tid = sync_tid;
2064 hammer2_chain_ref(parent);
2065 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2066 /* unlocked by caller */
2073 * Recursive scan of the last flushed transaction only. We are
2074 * doing this without pmp assignments so don't leave the chains
2075 * hanging around after we are done with them.
2078 chain = hammer2_chain_scan(parent, NULL, &cache_index,
2079 HAMMER2_LOOKUP_NODATA);
2081 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2082 if (chain->bref.mirror_tid >= sync_tid) {
2084 error = hammer2_recovery_scan(trans, hmp, chain,
2088 cumulative_error = error;
2090 chain = hammer2_chain_scan(parent, chain, &cache_index,
2091 HAMMER2_LOOKUP_NODATA);
2094 return cumulative_error;
2098 * Sync the entire filesystem; this is called from the filesystem syncer
2099 * process periodically and whenever a user calls sync(1) on the hammer
2102 * Currently is actually called from the syncer! \o/
2104 * This task will have to snapshot the state of the dirty inode chain.
2105 * From that, it will have to make sure all of the inodes on the dirty
2106 * chain have IO initiated. We make sure that io is initiated for the root
2109 * If waitfor is set, we wait for media to acknowledge the new rootblock.
2111 * THINKS: side A vs side B, to have sync not stall all I/O?
2114 hammer2_vfs_sync(struct mount *mp, int waitfor)
2116 struct hammer2_sync_info info;
2117 hammer2_inode_t *iroot;
2118 hammer2_chain_t *chain;
2119 hammer2_chain_t *parent;
2120 hammer2_pfsmount_t *pmp;
2121 hammer2_mount_t *hmp;
2132 KKASSERT(iroot->pmp == pmp);
2135 * We can't acquire locks on existing vnodes while in a transaction
2136 * without risking a deadlock. This assumes that vfsync() can be
2137 * called without the vnode locked (which it can in DragonFly).
2138 * Otherwise we'd have to implement a multi-pass or flag the lock
2139 * failures and retry.
2141 * The reclamation code interlocks with the sync list's token
2142 * (by removing the vnode from the scan list) before unlocking
2143 * the inode, giving us time to ref the inode.
2145 /*flags = VMSC_GETVP;*/
2147 if (waitfor & MNT_LAZY)
2148 flags |= VMSC_ONEPASS;
2151 * Start our flush transaction. This does not return until all
2152 * concurrent transactions have completed and will prevent any
2153 * new transactions from running concurrently, except for the
2154 * buffer cache transactions.
2156 * For efficiency do an async pass before making sure with a
2157 * synchronous pass on all related buffer cache buffers. It
2158 * should theoretically not be possible for any new file buffers
2159 * to be instantiated during this sequence.
2161 hammer2_trans_init(&info.trans, pmp, HAMMER2_TRANS_ISFLUSH |
2162 HAMMER2_TRANS_PREFLUSH);
2163 hammer2_run_unlinkq(&info.trans, pmp);
2166 info.waitfor = MNT_NOWAIT;
2167 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2168 info.waitfor = MNT_WAIT;
2169 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2172 * Clear PREFLUSH. This prevents (or asserts on) any new logical
2173 * buffer cache flushes which occur during the flush. Device buffers
2178 if (info.error == 0 && (waitfor & MNT_WAIT)) {
2179 info.waitfor = waitfor;
2180 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2184 hammer2_bioq_sync(info.trans.pmp);
2185 atomic_clear_int(&info.trans.flags, HAMMER2_TRANS_PREFLUSH);
2190 * Flush all storage elements making up the cluster
2192 * We must also flush any deleted siblings because the super-root
2193 * flush won't do it for us. They all must be staged or the
2194 * super-root flush will not be able to update its block table
2197 * XXX currently done serially instead of concurrently
2199 for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
2200 chain = iroot->cluster.array[i];
2202 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
2203 hammer2_flush(&info.trans, chain);
2204 hammer2_chain_unlock(chain);
2208 hammer2_trans_done(&info.trans);
2212 * Flush all volume roots to synchronize PFS flushes with the
2213 * storage media. Use a super-root transaction for each one.
2215 * The flush code will detect super-root -> pfs-root chain
2216 * transitions using the last pfs-root flush.
2218 for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
2219 chain = iroot->cluster.array[i];
2226 * We only have to flush each hmp once
2228 for (j = i - 1; j >= 0; --j) {
2229 if (iroot->cluster.array[j] &&
2230 iroot->cluster.array[j]->hmp == hmp)
2235 hammer2_trans_spmp(&info.trans, hmp->spmp);
2238 * Force an update of the XID from the PFS root to the
2239 * topology root. We couldn't do this from the PFS
2240 * transaction because a SPMP transaction is needed.
2241 * This does not modify blocks, instead what it does is
2242 * allow the flush code to find the transition point and
2243 * then update on the way back up.
2245 parent = chain->parent;
2246 KKASSERT(chain->pmp != parent->pmp);
2247 hammer2_chain_setflush(&info.trans, parent);
2250 * Media mounts have two 'roots', vchain for the topology
2251 * and fchain for the free block table. Flush both.
2253 * Note that the topology and free block table are handled
2254 * independently, so the free block table can wind up being
2255 * ahead of the topology. We depend on the bulk free scan
2256 * code to deal with any loose ends.
2258 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2259 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2260 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2262 * This will also modify vchain as a side effect,
2263 * mark vchain as modified now.
2265 hammer2_voldata_modify(hmp);
2266 chain = &hmp->fchain;
2267 hammer2_flush(&info.trans, chain);
2268 KKASSERT(chain == &hmp->fchain);
2270 hammer2_chain_unlock(&hmp->fchain);
2271 hammer2_chain_unlock(&hmp->vchain);
2273 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2274 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2275 chain = &hmp->vchain;
2276 hammer2_flush(&info.trans, chain);
2277 KKASSERT(chain == &hmp->vchain);
2282 hammer2_chain_unlock(&hmp->vchain);
2285 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2286 if ((hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) ||
2288 /* this will also modify vchain as a side effect */
2289 chain = &hmp->fchain;
2290 hammer2_flush(&info.trans, chain);
2291 KKASSERT(chain == &hmp->fchain);
2293 hammer2_chain_unlock(&hmp->fchain);
2299 * We can't safely flush the volume header until we have
2300 * flushed any device buffers which have built up.
2302 * XXX this isn't being incremental
2304 vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
2305 error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
2306 vn_unlock(hmp->devvp);
2309 * The flush code sets CHAIN_VOLUMESYNC to indicate that the
2310 * volume header needs synchronization via hmp->volsync.
2312 * XXX synchronize the flag & data with only this flush XXX
2315 (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
2319 * Synchronize the disk before flushing the volume
2323 bp->b_bio1.bio_offset = 0;
2326 bp->b_cmd = BUF_CMD_FLUSH;
2327 bp->b_bio1.bio_done = biodone_sync;
2328 bp->b_bio1.bio_flags |= BIO_SYNC;
2329 vn_strategy(hmp->devvp, &bp->b_bio1);
2330 biowait(&bp->b_bio1, "h2vol");
2334 * Then we can safely flush the version of the
2335 * volume header synchronized by the flush code.
2337 i = hmp->volhdrno + 1;
2338 if (i >= HAMMER2_NUM_VOLHDRS)
2340 if (i * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
2341 hmp->volsync.volu_size) {
2344 kprintf("sync volhdr %d %jd\n",
2345 i, (intmax_t)hmp->volsync.volu_size);
2346 bp = getblk(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2347 HAMMER2_PBUFSIZE, 0, 0);
2348 atomic_clear_int(&hmp->vchain.flags,
2349 HAMMER2_CHAIN_VOLUMESYNC);
2350 bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
2355 total_error = error;
2358 hammer2_trans_done(&info.trans);
2361 hammer2_trans_done(&info.trans);
2363 return (total_error);
2370 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2372 struct hammer2_sync_info *info = data;
2373 hammer2_inode_t *ip;
2382 if (vp->v_type == VNON || vp->v_type == VBAD) {
2386 if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2387 RB_EMPTY(&vp->v_rbdirty_tree)) {
2393 * VOP_FSYNC will start a new transaction so replicate some code
2394 * here to do it inline (see hammer2_vop_fsync()).
2396 * WARNING: The vfsync interacts with the buffer cache and might
2397 * block, we can't hold the inode lock at that time.
2398 * However, we MUST ref ip before blocking to ensure that
2399 * it isn't ripped out from under us (since we do not
2400 * hold a lock on the vnode).
2402 hammer2_inode_ref(ip);
2403 atomic_clear_int(&ip->flags, HAMMER2_INODE_MODIFIED);
2405 vfsync(vp, MNT_NOWAIT, 1, NULL, NULL);
2407 hammer2_inode_drop(ip);
2411 info->error = error;
2418 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2425 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2426 struct fid *fhp, struct vnode **vpp)
2433 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2434 int *exflagsp, struct ucred **credanonp)
2440 * Support code for hammer2_vfs_mount(). Read, verify, and install the volume
2441 * header into the HMP
2443 * XXX read four volhdrs and use the one with the highest TID whos CRC
2448 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2449 * nonexistant locations.
2451 * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2455 hammer2_install_volume_header(hammer2_mount_t *hmp)
2457 hammer2_volume_data_t *vd;
2459 hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2471 * There are up to 4 copies of the volume header (syncs iterate
2472 * between them so there is no single master). We don't trust the
2473 * volu_size field so we don't know precisely how large the filesystem
2474 * is, so depend on the OS to return an error if we go beyond the
2475 * block device's EOF.
2477 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2478 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2479 HAMMER2_VOLUME_BYTES, &bp);
2486 vd = (struct hammer2_volume_data *) bp->b_data;
2487 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2488 (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2494 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2495 /* XXX: Reversed-endianness filesystem */
2496 kprintf("hammer2: reverse-endian filesystem detected");
2502 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2503 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2504 HAMMER2_VOLUME_ICRC0_SIZE);
2505 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2506 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2507 HAMMER2_VOLUME_ICRC1_SIZE);
2508 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2509 kprintf("hammer2 volume header crc "
2510 "mismatch copy #%d %08x/%08x\n",
2517 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2526 hmp->volsync = hmp->voldata;
2528 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2529 kprintf("hammer2: using volume header #%d\n",
2534 kprintf("hammer2: no valid volume headers found!\n");
2540 * Reconnect using the passed file pointer. The caller must ref the
2544 hammer2_cluster_reconnect(hammer2_mount_t *hmp, struct file *fp)
2547 const char *name = "disk-volume";
2550 * Closes old comm descriptor, kills threads, cleans up
2551 * states, then installs the new descriptor and creates
2554 kdmsg_iocom_reconnect(&hmp->iocom, fp, "hammer2");
2557 * Setup LNK_CONN fields for autoinitiated state machine. We
2558 * will use SPANs to advertise multiple PFSs so only pass the
2559 * fsid and HAMMER2_PFSTYPE_SUPROOT for the AUTOCONN.
2561 * We are not initiating a LNK_SPAN so we do not have to set-up
2562 * iocom.auto_lnk_span.
2564 bzero(&hmp->iocom.auto_lnk_conn.pfs_clid,
2565 sizeof(hmp->iocom.auto_lnk_conn.pfs_clid));
2566 hmp->iocom.auto_lnk_conn.pfs_fsid = hmp->voldata.fsid;
2567 hmp->iocom.auto_lnk_conn.pfs_type = HAMMER2_PFSTYPE_SUPROOT;
2568 hmp->iocom.auto_lnk_conn.proto_version = DMSG_SPAN_PROTO_1;
2570 hmp->iocom.auto_lnk_conn.peer_type = hmp->voldata.peer_type;
2572 hmp->iocom.auto_lnk_conn.peer_type = DMSG_PEER_HAMMER2;
2575 * Filter adjustment. Clients do not need visibility into other
2576 * clients (otherwise millions of clients would present a serious
2577 * problem). The fs_label also serves to restrict the namespace.
2579 hmp->iocom.auto_lnk_conn.peer_mask = 1LLU << DMSG_PEER_HAMMER2;
2580 hmp->iocom.auto_lnk_conn.pfs_mask = (uint64_t)-1;
2583 switch (ipdata->pfs_type) {
2584 case DMSG_PFSTYPE_CLIENT:
2585 hmp->iocom.auto_lnk_conn.peer_mask &=
2586 ~(1LLU << DMSG_PFSTYPE_CLIENT);
2593 name_len = strlen(name);
2594 if (name_len >= sizeof(hmp->iocom.auto_lnk_conn.fs_label))
2595 name_len = sizeof(hmp->iocom.auto_lnk_conn.fs_label) - 1;
2596 bcopy(name, hmp->iocom.auto_lnk_conn.fs_label, name_len);
2597 hmp->iocom.auto_lnk_conn.fs_label[name_len] = 0;
2599 kdmsg_iocom_autoinitiate(&hmp->iocom, hammer2_autodmsg);
2603 hammer2_rcvdmsg(kdmsg_msg_t *msg)
2605 kprintf("RCVMSG %08x\n", msg->tcmd);
2608 case DMSG_DBG_SHELL:
2611 * Execute shell command (not supported atm)
2613 kdmsg_msg_result(msg, DMSG_ERR_NOSUPP);
2615 case DMSG_DBG_SHELL | DMSGF_REPLY:
2619 if (msg->aux_data) {
2620 msg->aux_data[msg->aux_size - 1] = 0;
2621 kprintf("HAMMER2 DBG: %s\n", msg->aux_data);
2626 * Unsupported message received. We only need to
2627 * reply if it's a transaction in order to close our end.
2628 * Ignore any one-way messages or any further messages
2629 * associated with the transaction.
2631 * NOTE: This case also includes DMSG_LNK_ERROR messages
2632 * which might be one-way, replying to those would
2633 * cause an infinite ping-pong.
2635 if (msg->any.head.cmd & DMSGF_CREATE)
2636 kdmsg_msg_reply(msg, DMSG_ERR_NOSUPP);
2643 * This function is called after KDMSG has automatically handled processing
2644 * of a LNK layer message (typically CONN, SPAN, or CIRC).
2646 * We tag off the LNK_CONN to trigger our LNK_VOLCONF messages which
2647 * advertises all available hammer2 super-root volumes.
2649 static void hammer2_update_spans(hammer2_mount_t *hmp, kdmsg_state_t *state);
2652 hammer2_autodmsg(kdmsg_msg_t *msg)
2654 hammer2_mount_t *hmp = msg->state->iocom->handle;
2657 kprintf("RCAMSG %08x\n", msg->tcmd);
2660 case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_REPLY:
2661 case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_DELETE | DMSGF_REPLY:
2662 if (msg->any.head.cmd & DMSGF_CREATE) {
2663 kprintf("HAMMER2: VOLDATA DUMP\n");
2666 * Dump the configuration stored in the volume header.
2667 * This will typically be import/export access rights,
2668 * master encryption keys (encrypted), etc.
2670 hammer2_voldata_lock(hmp);
2672 while (copyid < HAMMER2_COPYID_COUNT) {
2673 if (hmp->voldata.copyinfo[copyid].copyid)
2674 hammer2_volconf_update(hmp, copyid);
2677 hammer2_voldata_unlock(hmp);
2679 kprintf("HAMMER2: INITIATE SPANs\n");
2680 hammer2_update_spans(hmp, msg->state);
2682 if ((msg->any.head.cmd & DMSGF_DELETE) &&
2683 msg->state && (msg->state->txcmd & DMSGF_DELETE) == 0) {
2684 kprintf("HAMMER2: CONN WAS TERMINATED\n");
2693 * Update LNK_SPAN state
2696 hammer2_update_spans(hammer2_mount_t *hmp, kdmsg_state_t *state)
2698 const hammer2_inode_data_t *ripdata;
2699 hammer2_cluster_t *cparent;
2700 hammer2_cluster_t *cluster;
2701 hammer2_pfsmount_t *spmp;
2702 hammer2_key_t key_next;
2708 * Lookup mount point under the media-localized super-root.
2710 * cluster->pmp will incorrectly point to spmp and must be fixed
2714 cparent = hammer2_inode_lock_ex(spmp->iroot);
2715 cluster = hammer2_cluster_lookup(cparent, &key_next,
2720 if (hammer2_cluster_type(cluster) != HAMMER2_BREF_TYPE_INODE)
2722 ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
2723 kprintf("UPDATE SPANS: %s\n", ripdata->filename);
2725 rmsg = kdmsg_msg_alloc(state, DMSG_LNK_SPAN | DMSGF_CREATE,
2726 hammer2_lnk_span_reply, NULL);
2727 rmsg->any.lnk_span.pfs_clid = ripdata->pfs_clid;
2728 rmsg->any.lnk_span.pfs_fsid = ripdata->pfs_fsid;
2729 rmsg->any.lnk_span.pfs_type = ripdata->pfs_type;
2730 rmsg->any.lnk_span.peer_type = DMSG_PEER_HAMMER2;
2731 rmsg->any.lnk_span.proto_version = DMSG_SPAN_PROTO_1;
2732 name_len = ripdata->name_len;
2733 if (name_len >= sizeof(rmsg->any.lnk_span.fs_label))
2734 name_len = sizeof(rmsg->any.lnk_span.fs_label) - 1;
2735 bcopy(ripdata->filename, rmsg->any.lnk_span.fs_label, name_len);
2737 kdmsg_msg_write(rmsg);
2739 cluster = hammer2_cluster_next(cparent, cluster,
2745 hammer2_inode_unlock_ex(spmp->iroot, cparent);
2750 hammer2_lnk_span_reply(kdmsg_state_t *state, kdmsg_msg_t *msg)
2752 if ((state->txcmd & DMSGF_DELETE) == 0 &&
2753 (msg->any.head.cmd & DMSGF_DELETE)) {
2754 kdmsg_msg_reply(msg, 0);
2760 * Volume configuration updates are passed onto the userland service
2761 * daemon via the open LNK_CONN transaction.
2764 hammer2_volconf_update(hammer2_mount_t *hmp, int index)
2768 /* XXX interlock against connection state termination */
2769 kprintf("volconf update %p\n", hmp->iocom.conn_state);
2770 if (hmp->iocom.conn_state) {
2771 kprintf("TRANSMIT VOLCONF VIA OPEN CONN TRANSACTION\n");
2772 msg = kdmsg_msg_alloc(hmp->iocom.conn_state,
2773 DMSG_LNK_HAMMER2_VOLCONF,
2775 H2_LNK_VOLCONF(msg)->copy = hmp->voldata.copyinfo[index];
2776 H2_LNK_VOLCONF(msg)->mediaid = hmp->voldata.fsid;
2777 H2_LNK_VOLCONF(msg)->index = index;
2778 kdmsg_msg_write(msg);
2783 * This handles hysteresis on regular file flushes. Because the BIOs are
2784 * routed to a thread it is possible for an excessive number to build up
2785 * and cause long front-end stalls long before the runningbuffspace limit
2786 * is hit, so we implement hammer2_flush_pipe to control the
2789 * This is a particular problem when compression is used.
2792 hammer2_lwinprog_ref(hammer2_pfsmount_t *pmp)
2794 atomic_add_int(&pmp->count_lwinprog, 1);
2798 hammer2_lwinprog_drop(hammer2_pfsmount_t *pmp)
2802 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2803 if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2804 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2805 atomic_clear_int(&pmp->count_lwinprog,
2806 HAMMER2_LWINPROG_WAITING);
2807 wakeup(&pmp->count_lwinprog);
2812 hammer2_lwinprog_wait(hammer2_pfsmount_t *pmp)
2817 lwinprog = pmp->count_lwinprog;
2819 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2821 tsleep_interlock(&pmp->count_lwinprog, 0);
2822 atomic_set_int(&pmp->count_lwinprog, HAMMER2_LWINPROG_WAITING);
2823 lwinprog = pmp->count_lwinprog;
2824 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2826 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2831 * Manage excessive memory resource use for chain and related
2835 hammer2_pfs_memory_wait(hammer2_pfsmount_t *pmp)
2845 * Atomic check condition and wait. Also do an early speedup of
2846 * the syncer to try to avoid hitting the wait.
2849 waiting = pmp->inmem_dirty_chains;
2851 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2853 limit = pmp->mp->mnt_nvnodelistsize / 10;
2854 if (limit < hammer2_limit_dirty_chains)
2855 limit = hammer2_limit_dirty_chains;
2860 if ((int)(ticks - zzticks) > hz) {
2862 kprintf("count %ld %ld\n", count, limit);
2867 * Block if there are too many dirty chains present, wait
2868 * for the flush to clean some out.
2870 if (count > limit) {
2871 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2872 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2874 waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2875 speedup_syncer(pmp->mp);
2876 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2879 continue; /* loop on success or fail */
2883 * Try to start an early flush before we are forced to block.
2885 if (count > limit * 7 / 10)
2886 speedup_syncer(pmp->mp);
2892 hammer2_pfs_memory_inc(hammer2_pfsmount_t *pmp)
2895 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2900 hammer2_pfs_memory_wakeup(hammer2_pfsmount_t *pmp)
2908 waiting = pmp->inmem_dirty_chains;
2910 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2913 ~HAMMER2_DIRTYCHAIN_WAITING)) {
2918 if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2919 wakeup(&pmp->inmem_dirty_chains);
2926 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2928 hammer2_chain_t *scan;
2929 hammer2_chain_t *parent;
2933 kprintf("%*.*s...\n", tab, tab, "");
2938 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2940 chain, chain->bref.type,
2941 chain->bref.key, chain->bref.keybits,
2942 chain->bref.mirror_tid);
2944 kprintf("%*.*s [%08x] (%s) refs=%d\n",
2947 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2948 chain->data) ? (char *)chain->data->ipdata.filename : "?"),
2951 kprintf("%*.*s core [%08x]",
2955 parent = chain->parent;
2957 kprintf("\n%*.*s p=%p [pflags %08x prefs %d",
2959 parent, parent->flags, parent->refs);
2960 if (RB_EMPTY(&chain->core.rbtree)) {
2964 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2965 hammer2_dump_chain(scan, tab + 4, countp, 'a');
2966 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2967 kprintf("%*.*s}(%s)\n", tab, tab, "",
2968 chain->data->ipdata.filename);
2970 kprintf("%*.*s}\n", tab, tab, "");