X-Git-Url: http://www.chiark.greenend.org.uk/ucgi/~ianmdlvl/git?p=elogind.git;a=blobdiff_plain;f=src%2Fjournal%2Fjournal-file.c;h=ec12e891a791a8d0b3b07f992b4bd21c350ef244;hp=fec54f31b9888730d6efb933f1a5532f92e93cc8;hb=8facc3498ed037f842891ff55d1f60fe834f4ba0;hpb=f534928ad7aaeec0bec2d653b4a50e79b0fc8418 diff --git a/src/journal/journal-file.c b/src/journal/journal-file.c index fec54f31b..ec12e891a 100644 --- a/src/journal/journal-file.c +++ b/src/journal/journal-file.c @@ -26,8 +26,9 @@ #include #include #include -#include +#include +#include "btrfs-util.h" #include "journal-def.h" #include "journal-file.h" #include "journal-authenticate.h" @@ -68,6 +69,12 @@ /* How much to increase the journal file size at once each time we allocate something new. */ #define FILE_SIZE_INCREASE (8ULL*1024ULL*1024ULL) /* 8MB */ +/* Reread fstat() of the file for detecting deletions at least this often */ +#define LAST_STAT_REFRESH_USEC (5*USEC_PER_SEC) + +/* The mmap context to use for the header we pick as one above the last defined typed */ +#define CONTEXT_HEADER _OBJECT_TYPE_MAX + static int journal_file_set_online(JournalFile *f) { assert(f); @@ -77,6 +84,9 @@ static int journal_file_set_online(JournalFile *f) { if (!(f->fd >= 0 && f->header)) return -EINVAL; + if (mmap_cache_got_sigbus(f->mmap, f->fd)) + return -EIO; + switch(f->header->state) { case STATE_ONLINE: return 0; @@ -105,8 +115,14 @@ int journal_file_set_offline(JournalFile *f) { fsync(f->fd); + if (mmap_cache_got_sigbus(f->mmap, f->fd)) + return -EIO; + f->header->state = STATE_OFFLINE; + if (mmap_cache_got_sigbus(f->mmap, f->fd)) + return -EIO; + fsync(f->fd); return 0; @@ -121,14 +137,22 @@ void journal_file_close(JournalFile *f) { journal_file_append_tag(f); #endif - /* Sync everything to disk, before we mark the file offline */ + journal_file_set_offline(f); + if (f->mmap && f->fd >= 0) mmap_cache_close_fd(f->mmap, f->fd); - journal_file_set_offline(f); + if (f->fd >= 0 && f->defrag_on_close) { - if (f->header) - munmap(f->header, PAGE_ALIGN(sizeof(Header))); + /* Be friendly to btrfs: turn COW back on again now, + * and defragment the file. We won't write to the file + * ever again, hence remove all fragmentation, and + * reenable all the good bits COW usually provides + * (such as data checksumming). */ + + (void) chattr_fd(f->fd, false, FS_NOCOW_FL); + (void) btrfs_defrag_fd(f->fd); + } safe_close(f->fd); free(f->path); @@ -195,8 +219,8 @@ static int journal_file_init_header(JournalFile *f, JournalFile *template) { } static int journal_file_refresh_header(JournalFile *f) { - int r; sd_id128_t boot_id; + int r; assert(f); @@ -213,12 +237,12 @@ static int journal_file_refresh_header(JournalFile *f) { f->header->boot_id = boot_id; - journal_file_set_online(f); + r = journal_file_set_online(f); /* Sync the online state to disk */ fsync(f->fd); - return 0; + return r; } static int journal_file_verify_header(JournalFile *f) { @@ -299,7 +323,7 @@ static int journal_file_verify_header(JournalFile *f) { } else if (state == STATE_ARCHIVED) return -ESHUTDOWN; else if (state != STATE_OFFLINE) { - log_debug("Journal file %s has unknown state %u.", f->path, state); + log_debug("Journal file %s has unknown state %i.", f->path, state); return -EBUSY; } } @@ -312,6 +336,22 @@ static int journal_file_verify_header(JournalFile *f) { return 0; } +static int journal_file_fstat(JournalFile *f) { + assert(f); + assert(f->fd >= 0); + + if (fstat(f->fd, &f->last_stat) < 0) + return -errno; + + f->last_stat_usec = now(CLOCK_MONOTONIC); + + /* Refuse appending to files that are already deleted */ + if (f->last_stat.st_nlink <= 0) + return -EIDRM; + + return 0; +} + static int journal_file_allocate(JournalFile *f, uint64_t offset, uint64_t size) { uint64_t old_size, new_size; int r; @@ -322,6 +362,9 @@ static int journal_file_allocate(JournalFile *f, uint64_t offset, uint64_t size) * for sure, since we always call posix_fallocate() * ourselves */ + if (mmap_cache_got_sigbus(f->mmap, f->fd)) + return -EIO; + old_size = le64toh(f->header->header_size) + le64toh(f->header->arena_size); @@ -330,8 +373,21 @@ static int journal_file_allocate(JournalFile *f, uint64_t offset, uint64_t size) if (new_size < le64toh(f->header->header_size)) new_size = le64toh(f->header->header_size); - if (new_size <= old_size) - return 0; + if (new_size <= old_size) { + + /* We already pre-allocated enough space, but before + * we write to it, let's check with fstat() if the + * file got deleted, in order make sure we don't throw + * away the data immediately. Don't check fstat() for + * all writes though, but only once ever 10s. */ + + if (f->last_stat_usec + LAST_STAT_REFRESH_USEC > now(CLOCK_MONOTONIC)) + return 0; + + return journal_file_fstat(f); + } + + /* Allocate more space. */ if (f->metrics.max_size > 0 && new_size > f->metrics.max_size) return -E2BIG; @@ -366,21 +422,21 @@ static int journal_file_allocate(JournalFile *f, uint64_t offset, uint64_t size) if (r != 0) return -r; - if (fstat(f->fd, &f->last_stat) < 0) - return -errno; - f->header->arena_size = htole64(new_size - le64toh(f->header->header_size)); - return 0; + return journal_file_fstat(f); } static unsigned type_to_context(ObjectType type) { /* One context for each type, plus one catch-all for the rest */ assert_cc(_OBJECT_TYPE_MAX <= MMAP_CACHE_MAX_CONTEXTS); + assert_cc(CONTEXT_HEADER < MMAP_CACHE_MAX_CONTEXTS); return type > OBJECT_UNUSED && type < _OBJECT_TYPE_MAX ? type : 0; } static int journal_file_move_to(JournalFile *f, ObjectType type, bool keep_always, uint64_t offset, uint64_t size, void **ret) { + int r; + assert(f); assert(ret); @@ -392,8 +448,11 @@ static int journal_file_move_to(JournalFile *f, ObjectType type, bool keep_alway /* Hmm, out of range? Let's refresh the fstat() data * first, before we trust that check. */ - if (fstat(f->fd, &f->last_stat) < 0 || - offset + size > (uint64_t) f->last_stat.st_size) + r = journal_file_fstat(f); + if (r < 0) + return r; + + if (offset + size > (uint64_t) f->last_stat.st_size) return -EADDRNOTAVAIL; } @@ -647,7 +706,7 @@ static int journal_file_link_field( uint64_t offset, uint64_t hash) { - uint64_t p, h; + uint64_t p, h, m; int r; assert(f); @@ -657,11 +716,14 @@ static int journal_file_link_field( if (o->object.type != OBJECT_FIELD) return -EINVAL; - /* This might alter the window we are looking at */ + m = le64toh(f->header->field_hash_table_size) / sizeof(HashItem); + if (m <= 0) + return -EBADMSG; + /* This might alter the window we are looking at */ o->field.next_hash_offset = o->field.head_data_offset = 0; - h = hash % (le64toh(f->header->field_hash_table_size) / sizeof(HashItem)); + h = hash % m; p = le64toh(f->field_hash_table[h].tail_hash_offset); if (p == 0) f->field_hash_table[h].head_hash_offset = htole64(offset); @@ -687,7 +749,7 @@ static int journal_file_link_data( uint64_t offset, uint64_t hash) { - uint64_t p, h; + uint64_t p, h, m; int r; assert(f); @@ -697,13 +759,16 @@ static int journal_file_link_data( if (o->object.type != OBJECT_DATA) return -EINVAL; - /* This might alter the window we are looking at */ + m = le64toh(f->header->data_hash_table_size) / sizeof(HashItem); + if (m <= 0) + return -EBADMSG; + /* This might alter the window we are looking at */ o->data.next_hash_offset = o->data.next_field_offset = 0; o->data.entry_offset = o->data.entry_array_offset = 0; o->data.n_entries = 0; - h = hash % (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)); + h = hash % m; p = le64toh(f->data_hash_table[h].tail_hash_offset); if (p == 0) /* Only entry in the hash table is easy */ @@ -732,7 +797,7 @@ int journal_file_find_field_object_with_hash( const void *field, uint64_t size, uint64_t hash, Object **ret, uint64_t *offset) { - uint64_t p, osize, h; + uint64_t p, osize, h, m; int r; assert(f); @@ -740,10 +805,12 @@ int journal_file_find_field_object_with_hash( osize = offsetof(Object, field.payload) + size; - if (f->header->field_hash_table_size == 0) + m = le64toh(f->header->field_hash_table_size) / sizeof(HashItem); + + if (m <= 0) return -EBADMSG; - h = hash % (le64toh(f->header->field_hash_table_size) / sizeof(HashItem)); + h = hash % m; p = le64toh(f->field_hash_table[h].head_hash_offset); while (p > 0) { @@ -793,7 +860,7 @@ int journal_file_find_data_object_with_hash( const void *data, uint64_t size, uint64_t hash, Object **ret, uint64_t *offset) { - uint64_t p, osize, h; + uint64_t p, osize, h, m; int r; assert(f); @@ -801,10 +868,11 @@ int journal_file_find_data_object_with_hash( osize = offsetof(Object, data.payload) + size; - if (f->header->data_hash_table_size == 0) + m = le64toh(f->header->data_hash_table_size) / sizeof(HashItem); + if (m <= 0) return -EBADMSG; - h = hash % (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)); + h = hash % m; p = le64toh(f->data_hash_table[h].head_hash_offset); while (p > 0) { @@ -1358,6 +1426,14 @@ int journal_file_append_entry(JournalFile *f, const dual_timestamp *ts, const st r = journal_file_append_entry_internal(f, ts, xor_hash, items, n_iovec, seqnum, ret, offset); + /* If the memory mapping triggered a SIGBUS then we return an + * IO error and ignore the error code passed down to us, since + * it is very likely just an effect of a nullified replacement + * mapping page */ + + if (mmap_cache_got_sigbus(f->mmap, f->fd)) + r = -EIO; + journal_file_post_change(f); return r; @@ -1713,7 +1789,6 @@ found: return 1; } - static int generic_array_bisect_plus_one( JournalFile *f, uint64_t extra, @@ -2342,7 +2417,7 @@ void journal_file_dump(JournalFile *f) { break; default: - printf("Type: unknown (%u)\n", o->object.type); + printf("Type: unknown (%i)\n", o->object.type); break; } @@ -2458,9 +2533,10 @@ int journal_file_open( JournalFile *template, JournalFile **ret) { + bool newly_created = false; JournalFile *f; + void *h; int r; - bool newly_created = false; assert(fname); assert(ret); @@ -2520,13 +2596,22 @@ int journal_file_open( goto fail; } - if (fstat(f->fd, &f->last_stat) < 0) { - r = -errno; + r = journal_file_fstat(f); + if (r < 0) goto fail; - } if (f->last_stat.st_size == 0 && f->writable) { - uint64_t crtime; + + /* Before we write anything, turn off COW logic. Given + * our write pattern that is quite unfriendly to COW + * file systems this should greatly improve + * performance on COW file systems, such as btrfs, at + * the expense of data integrity features (which + * shouldn't be too bad, given that we do our own + * checksumming). */ + r = chattr_fd(f->fd, true, FS_NOCOW_FL); + if (r < 0) + log_warning_errno(errno, "Failed to set file attributes: %m"); /* Let's attach the creation time to the journal file, * so that the vacuuming code knows the age of this @@ -2538,8 +2623,7 @@ int journal_file_open( * attributes are not supported we'll just skip this, * and rely solely on mtime/atime/ctime of the file. */ - crtime = htole64((uint64_t) now(CLOCK_REALTIME)); - fsetxattr(f->fd, "user.crtime_usec", &crtime, sizeof(crtime), XATTR_CREATE); + fd_setcrtime(f->fd, 0); #ifdef HAVE_GCRYPT /* Try to load the FSPRG state, and if we can't, then @@ -2555,10 +2639,9 @@ int journal_file_open( if (r < 0) goto fail; - if (fstat(f->fd, &f->last_stat) < 0) { - r = -errno; + r = journal_file_fstat(f); + if (r < 0) goto fail; - } newly_created = true; } @@ -2568,13 +2651,14 @@ int journal_file_open( goto fail; } - f->header = mmap(NULL, PAGE_ALIGN(sizeof(Header)), prot_from_flags(flags), MAP_SHARED, f->fd, 0); - if (f->header == MAP_FAILED) { - f->header = NULL; + r = mmap_cache_get(f->mmap, f->fd, f->prot, CONTEXT_HEADER, true, 0, PAGE_ALIGN(sizeof(Header)), &f->last_stat, &h); + if (r < 0) { r = -errno; goto fail; } + f->header = h; + if (!newly_created) { r = journal_file_verify_header(f); if (r < 0) @@ -2631,10 +2715,18 @@ int journal_file_open( if (r < 0) goto fail; + if (mmap_cache_got_sigbus(f->mmap, f->fd)) { + r = -EIO; + goto fail; + } + *ret = f; return 0; fail: + if (f->fd >= 0 && mmap_cache_got_sigbus(f->mmap, f->fd)) + r = -EIO; + journal_file_close(f); return r; @@ -2666,12 +2758,20 @@ int journal_file_rotate(JournalFile **f, bool compress, bool seal) { if (r < 0) return -ENOMEM; + /* Try to rename the file to the archived version. If the file + * already was deleted, we'll get ENOENT, let's ignore that + * case. */ r = rename(old_file->path, p); - if (r < 0) + if (r < 0 && errno != ENOENT) return -errno; old_file->header->state = STATE_ARCHIVED; + /* Currently, btrfs is not very good with out write patterns + * and fragments heavily. Let's defrag our journal files when + * we archive them */ + old_file->defrag_on_close = true; + r = journal_file_open(old_file->path, old_file->flags, old_file->mode, compress, seal, NULL, old_file->mmap, old_file, &new_file); journal_file_close(old_file); @@ -2701,7 +2801,9 @@ int journal_file_open_reliably( r != -EHOSTDOWN && /* other machine */ r != -EPROTONOSUPPORT && /* incompatible feature */ r != -EBUSY && /* unclean shutdown */ - r != -ESHUTDOWN /* already archived */) + r != -ESHUTDOWN && /* already archived */ + r != -EIO && /* IO error, including SIGBUS on mmap */ + r != -EIDRM /* File has been deleted */) return r; if ((flags & O_ACCMODE) == O_RDONLY) @@ -2726,6 +2828,12 @@ int journal_file_open_reliably( if (r < 0) return -errno; + /* btrfs doesn't cope well with our write pattern and + * fragments heavily. Let's defrag all files we rotate */ + + (void) chattr_path(p, false, FS_NOCOW_FL); + (void) btrfs_defrag(p); + log_warning("File %s corrupted or uncleanly shut down, renaming and replacing.", fname); return journal_file_open(fname, flags, mode, compress, seal, @@ -2808,7 +2916,12 @@ int journal_file_copy_entry(JournalFile *from, JournalFile *to, Object *o, uint6 return r; } - return journal_file_append_entry_internal(to, &ts, xor_hash, items, n, seqnum, ret, offset); + r = journal_file_append_entry_internal(to, &ts, xor_hash, items, n, seqnum, ret, offset); + + if (mmap_cache_got_sigbus(to->mmap, to->fd)) + return -EIO; + + return r; } void journal_default_metrics(JournalMetrics *m, int fd) {