From: John Spray Date: Wed, 29 Apr 2015 12:45:46 +0000 (+0100) Subject: tools/cephfs: add cephfs-data-scan X-Git-Tag: v9.0.3~38^2~1 X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=1c00b45166609e02e69c12b3a2eb7ee0760af762;p=ceph.git tools/cephfs: add cephfs-data-scan For scraping cephfs data pools for metadata in backtraces, and injecting that metadata into the cephfs metadata pool (initially offline). Signed-off-by: John Spray --- diff --git a/src/.gitignore b/src/.gitignore index 88e6bcd37d1..d8e4f891602 100644 --- a/src/.gitignore +++ b/src/.gitignore @@ -13,6 +13,7 @@ Makefile /ceph-client-debug /cephfs-journal-tool /cephfs-table-tool +/cephfs-data-scan /ceph-conf /ceph-coverage /ceph-crush-location diff --git a/src/tools/Makefile-server.am b/src/tools/Makefile-server.am index 45f943e0bfc..f3557cda9d2 100644 --- a/src/tools/Makefile-server.am +++ b/src/tools/Makefile-server.am @@ -47,6 +47,13 @@ cephfs_table_tool_SOURCES = \ cephfs_table_tool_LDADD = $(LIBMDS) $(LIBRADOS) $(CEPH_GLOBAL) bin_PROGRAMS += cephfs-table-tool +cephfs_data_scan_SOURCES = \ + tools/cephfs/cephfs-data-scan.cc \ + tools/cephfs/DataScan.cc \ + tools/cephfs/MDSUtility.cc +cephfs_data_scan_LDADD = $(LIBMDS) libcls_cephfs_client.la $(LIBRADOS) $(CEPH_GLOBAL) +bin_PROGRAMS += cephfs-data-scan + endif # WITH_RADOS endif # ENABLE_CLIENT endif # WITH_MDS diff --git a/src/tools/Makefile.am b/src/tools/Makefile.am index 7fc23b87e4f..5afedaa26d9 100644 --- a/src/tools/Makefile.am +++ b/src/tools/Makefile.am @@ -57,4 +57,6 @@ noinst_HEADERS += \ tools/RadosDump.h \ tools/rados/RadosImport.h \ tools/ceph_objectstore_tool.h \ - tools/rados/PoolDump.h + tools/rados/PoolDump.h \ + tools/cephfs/DataScan.h + diff --git a/src/tools/cephfs/DataScan.cc b/src/tools/cephfs/DataScan.cc new file mode 100644 index 00000000000..8e6bc68ffb9 --- /dev/null +++ b/src/tools/cephfs/DataScan.cc @@ -0,0 +1,1277 @@ +// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- +// vim: ts=8 sw=2 smarttab +/* + * Ceph - scalable distributed file system + * + * Copyright (C) 2015 Red Hat + * + * This is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License version 2.1, as published by the Free Software + * Foundation. See file COPYING. + * + */ + +#include "common/errno.h" +#include "common/ceph_argparse.h" +#include +#include "include/util.h" + +#include "mds/CInode.h" +#include "cls/cephfs/cls_cephfs_client.h" + +#include "DataScan.h" + +#define dout_subsys ceph_subsys_mds +#undef dout_prefix +#define dout_prefix *_dout << "datascan." << __func__ << ": " + +void DataScan::usage() +{ + std::cout << "Usage: \n" + << " cephfs-data-scan init [--force-init]\n" + << " cephfs-data-scan scan_extents [--force-pool] \n" + << " cephfs-data-scan scan_inodes [--force-pool] [--force-corrupt] \n" + << "\n" + << " --force-corrupt: overrite apparently corrupt structures\n" + << " --force-init: write root inodes even if they exist\n" + << " --force-pool: use data pool even if it is not in MDSMap\n" + << std::endl; + + generic_client_usage(); +} + +bool DataScan::parse_kwarg( + const std::vector &args, + std::vector::const_iterator &i, + int *r) +{ + if (i + 1 == args.end()) { + return false; + } + + const std::string arg(*i); + const std::string val(*(++i)); + + if (arg == std::string("--output-dir")) { + if (driver != NULL) { + derr << "Unexpected --output-dir: output already selected!" << dendl; + *r = -EINVAL; + return false; + } + driver = new LocalFileDriver(val, data_io); + return true; + } else if (arg == std::string("-n")) { + std::string err; + n = strict_strtoll(val.c_str(), 10, &err); + if (!err.empty()) { + std::cerr << "Invalid worker number '" << val << "'" << std::endl; + *r = -EINVAL; + return false; + } + return true; + } else if (arg == std::string("-m")) { + std::string err; + m = strict_strtoll(val.c_str(), 10, &err); + if (!err.empty()) { + std::cerr << "Invalid worker count '" << val << "'" << std::endl; + *r = -EINVAL; + return false; + } + return true; + } else { + return false; + } +} + +bool DataScan::parse_arg( + const std::vector &args, + std::vector::const_iterator &i) +{ + const std::string arg(*i); + if (arg == "--force-pool") { + force_pool = true; + return true; + } else if (arg == "--force-corrupt") { + force_corrupt = true; + return true; + } else if (arg == "--force-init") { + force_init = true; + return true; + } else { + return false; + } +} + +int DataScan::main(const std::vector &args) +{ + // Parse args + // ========== + if (args.size() < 1) { + usage(); + return -EINVAL; + } + + // Common RADOS init: open metadata pool + // ===================================== + librados::Rados rados; + int r = rados.init_with_context(g_ceph_context); + if (r < 0) { + derr << "RADOS unavailable" << dendl; + return r; + } + + std::string const &command = args[0]; + std::string data_pool_name; + + // Consume any known --key val or --flag arguments + for (std::vector::const_iterator i = args.begin() + 1; + i != args.end(); ++i) { + if (parse_kwarg(args, i, &r)) { + // Skip the kwarg value field + ++i; + continue; + } else if (r) { + return r; + } + + if (parse_arg(args, i)) { + continue; + } + + if (i + 1 == args.end() && + (command == "scan_inodes" || command == "scan_extents")) { + data_pool_name = *i; + continue; + } + + // Fall through: unhandled + std::cerr << "Unknown argument '" << *i << "'" << std::endl; + return -EINVAL; + } + + // Default to output to metadata pool + if (driver == NULL) { + driver = new MetadataDriver(); + driver->set_force_corrupt(force_corrupt); + driver->set_force_init(force_init); + } + + dout(4) << "connecting to RADOS..." << dendl; + rados.connect(); + r = driver->init(rados, mdsmap); + if (r < 0) { + return r; + } + + // Initialize data_io for those commands that need it + if (command == "scan_inodes" + || command == "scan_extents") { + if (data_pool_name.empty()) { + std::cerr << "Data pool not specified" << std::endl; + usage(); + return -EINVAL; + } + + data_pool_id = rados.pool_lookup(data_pool_name.c_str()); + if (data_pool_id < 0) { + std::cerr << "Data pool '" << data_pool_name << "' not found!" << std::endl; + return -ENOENT; + } else { + dout(4) << "data pool '" << data_pool_name + << "' has ID " << data_pool_id << dendl; + } + + if (!mdsmap->is_data_pool(data_pool_id)) { + std::cerr << "Warning: pool '" << data_pool_name << "' is not a " + "CephFS data pool!" << std::endl; + if (!force_pool) { + std::cerr << "Use --force-pool to continue" << std::endl; + return -EINVAL; + } + } + + dout(4) << "opening data pool '" << data_pool_name << "'" << dendl; + r = rados.ioctx_create(data_pool_name.c_str(), data_io); + if (r != 0) { + return r; + } + } + + // Finally, dispatch command + if (command == "scan_inodes") { + return scan_inodes(); + } else if (command == "scan_extents") { + return scan_extents(); + } else if (command == "init") { + return driver->init_roots(mdsmap->get_first_data_pool()); + } else { + std::cerr << "Unknown command '" << command << "'" << std::endl; + return -EINVAL; + } +} + +int MetadataDriver::inject_unlinked_inode( + inodeno_t inono, int mode, int64_t data_pool_id) +{ + const object_t oid = InodeStore::get_object_name(inono, frag_t(), ".inode"); + + // Skip if exists + bool already_exists = false; + int r = root_exists(inono, &already_exists); + if (r) { + return r; + } + if (already_exists && !force_init) { + std::cerr << "Inode 0x" << std::hex << inono << std::dec << " already" + " exists, skipping create. Use --force-init to overwrite" + " the existing object." << std::endl; + return 0; + } + + // Compose + InodeStore inode; + inode.inode.ino = inono; + inode.inode.version = 1; + inode.inode.xattr_version = 1; + inode.inode.mode = 0500 | mode; + // Fake size to 1, so that the directory doesn't appear to be empty + // (we won't actually give the *correct* size here though) + inode.inode.size = 1; + inode.inode.dirstat.nfiles = 1; + + inode.inode.ctime = + inode.inode.mtime = ceph_clock_now(g_ceph_context); + inode.inode.nlink = 1; + inode.inode.truncate_size = -1ull; + inode.inode.truncate_seq = 1; + + // Force layout to default: should we let users override this so that + // they don't have to mount the filesystem to correct it? + inode.inode.layout = g_default_file_layout; + inode.inode.layout.fl_pg_pool = data_pool_id; + + // Assume that we will get our stats wrong, and that we may + // be ignoring dirfrags that exist + inode.damage_flags |= (DAMAGE_STATS | DAMAGE_RSTATS | DAMAGE_FRAGTREE); + + // Serialize + bufferlist inode_bl; + ::encode(std::string(CEPH_FS_ONDISK_MAGIC), inode_bl); + inode.encode(inode_bl); + + // Write + r = metadata_io.write_full(oid.name, inode_bl); + if (r != 0) { + derr << "Error writing '" << oid.name << "': " << cpp_strerror(r) << dendl; + return r; + } + + return r; +} + +int MetadataDriver::root_exists(inodeno_t ino, bool *result) +{ + object_t oid = InodeStore::get_object_name(ino, frag_t(), ".inode"); + uint64_t size; + time_t mtime; + int r = metadata_io.stat(oid.name, &size, &mtime); + if (r == -ENOENT) { + *result = false; + return 0; + } else if (r < 0) { + return r; + } + + *result = true; + return 0; +} + +int MetadataDriver::init_roots(int64_t data_pool_id) +{ + int r = 0; + r = inject_unlinked_inode(MDS_INO_ROOT, S_IFDIR|0755, data_pool_id); + if (r != 0) { + return r; + } + r = inject_unlinked_inode(MDS_INO_MDSDIR(0), S_IFDIR, data_pool_id); + if (r != 0) { + return r; + } + + return 0; +} + +int MetadataDriver::check_roots(bool *result) +{ + int r; + r = root_exists(MDS_INO_ROOT, result); + if (r != 0) { + return r; + } + if (!*result) { + return 0; + } + + r = root_exists(MDS_INO_MDSDIR(0), result); + if (r != 0) { + return r; + } + if (!*result) { + return 0; + } + + return 0; +} + +/** + * Stages: + * + * SERIAL init + * 0. Create root inodes if don't exist + * PARALLEL scan_extents + * 1. Size and mtime recovery: scan ALL objects, and update 0th + * objects with max size and max mtime seen. + * PARALLEL scan_inodes + * 2. Inode recovery: scan ONLY 0th objects, and inject metadata + * into dirfrag OMAPs, creating blank dirfrags as needed. No stats + * or rstats at this stage. Inodes without backtraces go into + * lost+found + * TODO: SERIAL "recover stats" + * 3. Dirfrag statistics: depth first traverse into metadata tree, + * rebuilding dir sizes. + * TODO PARALLEL "clean up" + * 4. Cleanup; go over all 0th objects (and dirfrags if we tagged + * anything onto them) and remove any of the xattrs that we + * used for accumulating. + */ + + +int parse_oid(const std::string &oid, uint64_t *inode_no, uint64_t *obj_id) +{ + if (oid.find(".") == std::string::npos || oid.find(".") == oid.size() - 1) { + return -EINVAL; + } + + std::string err; + std::string inode_str = oid.substr(0, oid.find(".")); + *inode_no = strict_strtoll(inode_str.c_str(), 16, &err); + if (!err.empty()) { + return -EINVAL; + } + + std::string pos_string = oid.substr(oid.find(".") + 1); + *obj_id = strict_strtoll(pos_string.c_str(), 16, &err); + if (!err.empty()) { + return -EINVAL; + } + + return 0; +} + +// Pending sharded pgls & add in progress mechanism for that +#undef SHARDEDPGLS + +int DataScan::scan_extents() +{ +#ifdef SHARDED_PGLS + float progress = 0.0; + librados::NObjectIterator i = data_io.nobjects_begin(n, m); +#else + librados::NObjectIterator i = data_io.nobjects_begin(); +#endif + + librados::NObjectIterator i_end = data_io.nobjects_end(); + int r = 0; + + for (; i != i_end; ++i) { + const std::string oid = i->get_oid(); +#ifdef SHARDED_PGLS + if (i.get_progress() != progress) { + if (int(i.get_progress() * 100) / 5 != int(progress * 100) / 5) { + std::cerr << percentify(i.get_progress()) << "%" << std::endl; + } + progress = i.get_progress(); + } +#endif + + // Read size + uint64_t size; + time_t mtime; + r = data_io.stat(oid, &size, &mtime); + if (r != 0) { + dout(4) << "Cannot stat '" << oid << "': skipping" << dendl; + continue; + } + + // I need to keep track of + // * The highest object ID seen + // * The size of the highest object ID seen + // * The largest object seen + // + // Given those things, I can later infer the object chunking + // size, the offset of the last object (chunk size * highest ID seen) + // and the actual size (offset of last object + size of highest ID seen) + // + // This logic doesn't take account of striping. + uint64_t inode_no = 0; + uint64_t obj_id = 0; + r = parse_oid(oid, &inode_no, &obj_id); + if (r != 0) { + dout(4) << "Bad object name '" << oid << "' skipping" << dendl; + continue; + } + + int r = ClsCephFSClient::accumulate_inode_metadata( + data_io, + inode_no, + obj_id, + size, + mtime); + if (r < 0) { + derr << "Failed to accumulate metadata data from '" + << oid << "': " << cpp_strerror(r) << dendl; + continue; + } + } + + return 0; +} + +int DataScan::scan_inodes() +{ +#ifdef SHARDED_PGLS + float progress = 0.0; + librados::NObjectIterator i = data_io.nobjects_begin(n, m); +#else + librados::NObjectIterator i = data_io.nobjects_begin(); +#endif + librados::NObjectIterator i_end = data_io.nobjects_end(); + + bool roots_present; + int r = driver->check_roots(&roots_present); + if (r != 0) { + derr << "Unexpected error checking roots: '" + << cpp_strerror(r) << "'" << dendl; + return r; + } + + if (!roots_present) { + std::cerr << "Some or all system inodes are absent. Run 'init' from " + "one node before running 'scan_inodes'" << std::endl; + return -EIO; + } + + for (; i != i_end; ++i) { + const std::string oid = i->get_oid(); +#ifdef SHARDED_PGLS + if (i.get_progress() != progress) { + if (int(i.get_progress() * 100) / 5 != int(progress * 100) / 5) { + std::cerr << percentify(i.get_progress()) << "%" << std::endl; + } + progress = i.get_progress(); + } +#endif + + uint64_t obj_name_ino = 0; + uint64_t obj_name_offset = 0; + r = parse_oid(oid, &obj_name_ino, &obj_name_offset); + if (r != 0) { + dout(4) << "Bad object name '" << oid << "', skipping" << dendl; + continue; + } + + // We are only interested in 0th objects during this phase: we touched + // the other objects during scan_extents + if (obj_name_offset != 0) { + continue; + } + + AccumulateResult accum_res; + inode_backtrace_t backtrace; + int r = ClsCephFSClient::fetch_inode_accumulate_result( + data_io, oid, &backtrace, &accum_res); + + if (r < 0) { + dout(4) << "Unexpected error loading accumulated metadata from '" + << oid << "': " << cpp_strerror(r) << dendl; + continue; + } + + const time_t file_mtime = accum_res.max_mtime; + uint64_t file_size = 0; + uint32_t chunk_size = g_default_file_layout.fl_object_size; + bool have_backtrace = !(backtrace.ancestors.empty()); + + // Calculate file_size, guess chunk_size + if (accum_res.ceiling_obj_index > 0) { + // When there are multiple objects, the largest object probably + // indicates the chunk size. But not necessarily, because files + // can be sparse. Only make this assumption if size seen + // is a power of two, as chunk sizes typically are. + if ((accum_res.max_obj_size & (accum_res.max_obj_size - 1)) == 0) { + chunk_size = accum_res.max_obj_size; + } + + file_size = chunk_size * accum_res.ceiling_obj_index + + accum_res.ceiling_obj_size; + } else { + file_size = accum_res.ceiling_obj_size; + } + + // Santity checking backtrace ino against object name + if (have_backtrace && backtrace.ino != obj_name_ino) { + dout(4) << "Backtrace ino 0x" << std::hex << backtrace.ino + << " doesn't match object name ino 0x" << obj_name_ino + << std::dec << dendl; + have_backtrace = false; + } + + // Inject inode to the metadata pool + if (have_backtrace) { + inode_backpointer_t root_bp = *(backtrace.ancestors.rbegin()); + if (MDS_INO_IS_MDSDIR(root_bp.dirino)) { + /* Special case for strays: even if we have a good backtrace, + * don't put it in the stray dir, because while that would technically + * give it linkage it would still be invisible to the user */ + r = driver->inject_lost_and_found( + obj_name_ino, file_size, file_mtime, chunk_size, data_pool_id); + if (r < 0) { + dout(4) << "Error injecting 0x" << std::hex << backtrace.ino + << std::dec << " into lost+found: " << cpp_strerror(r) << dendl; + if (r == -EINVAL) { + dout(4) << "Use --force-corrupt to overwrite structures that " + "appear to be corrupt" << dendl; + } + } + } else { + /* Happy case: we will inject a named dentry for this inode */ + r = driver->inject_with_backtrace( + backtrace, file_size, file_mtime, chunk_size, data_pool_id); + if (r < 0) { + dout(4) << "Error injecting 0x" << std::hex << backtrace.ino + << std::dec << " with backtrace: " << cpp_strerror(r) << dendl; + if (r == -EINVAL) { + dout(4) << "Use --force-corrupt to overwrite structures that " + "appear to be corrupt" << dendl; + } + } + } + } else { + /* Backtrace-less case: we will inject a lost+found dentry */ + r = driver->inject_lost_and_found( + obj_name_ino, file_size, file_mtime, chunk_size, data_pool_id); + if (r < 0) { + dout(4) << "Error injecting 0x" << std::hex << obj_name_ino + << std::dec << " into lost+found: " << cpp_strerror(r) << dendl; + if (r == -EINVAL) { + dout(4) << "Use --force-corrupt to overwrite structures that " + "appear to be corrupt" << dendl; + } + } + } + } + + return 0; +} + +int MetadataDriver::read_fnode( + inodeno_t ino, frag_t frag, fnode_t *fnode, + uint64_t *last_version) +{ + assert(fnode != NULL); + + object_t frag_oid = InodeStore::get_object_name(ino, frag, ""); + bufferlist fnode_bl; + int r = metadata_io.omap_get_header(frag_oid.name, &fnode_bl); + *last_version = metadata_io.get_last_version(); + if (r < 0) { + return r; + } + + bufferlist::iterator old_fnode_iter = fnode_bl.begin(); + try { + (*fnode).decode(old_fnode_iter); + } catch (const buffer::error &err) { + return -EINVAL; + } + + return 0; +} + +int MetadataDriver::read_dentry(inodeno_t parent_ino, frag_t frag, + const std::string &dname, InodeStore *inode) +{ + assert(inode != NULL); + + + std::string key; + dentry_key_t dn_key(CEPH_NOSNAP, dname.c_str()); + dn_key.encode(key); + + std::set keys; + keys.insert(key); + std::map vals; + object_t frag_oid = InodeStore::get_object_name(parent_ino, frag, ""); + int r = metadata_io.omap_get_vals_by_keys(frag_oid.name, keys, &vals); + dout(20) << "oid=" << frag_oid.name + << " dname=" << dname + << " frag=" << frag + << ", r=" << r << dendl; + if (r < 0) { + return r; + } + + if (vals.find(key) == vals.end()) { + dout(20) << key << " not found in result" << dendl; + return -ENOENT; + } + + try { + bufferlist::iterator q = vals[key].begin(); + snapid_t dnfirst; + ::decode(dnfirst, q); + char dentry_type; + ::decode(dentry_type, q); + if (dentry_type == 'I') { + inode->decode_bare(q); + return 0; + } else { + dout(20) << "dentry type '" << dentry_type << "': cannot" + "read an inode out of that" << dendl; + return -EINVAL; + } + } catch (const buffer::error &err) { + dout(20) << "encoding error in dentry 0x" << std::hex << parent_ino + << std::dec << "/" << dname << dendl; + return -EINVAL; + } + + return 0; +} + +int MetadataDriver::inject_lost_and_found(inodeno_t ino, uint64_t file_size, + time_t file_mtime, uint32_t chunk_size, int64_t data_pool_id) +{ + // Create lost+found if doesn't exist + bool created = false; + int r = find_or_create_dirfrag(CEPH_INO_ROOT, frag_t(), &created); + if (r < 0) { + return r; + } + InodeStore lf_ino; + r = read_dentry(CEPH_INO_ROOT, frag_t(), "lost+found", &lf_ino); + if (r == -ENOENT || r == -EINVAL) { + if (r == -EINVAL && !force_corrupt) { + return r; + } + // Inject dentry + lf_ino.inode.mode = 0755 | S_IFDIR; + // Set nfiles to something non-zero, to fool any other code + // that tries to ignore 'empty' directories. This won't be + // accurate, but it should avoid functional issues. + lf_ino.inode.dirstat.nfiles = 1; + lf_ino.inode.size = 1; + lf_ino.inode.nlink = 1; + lf_ino.inode.ino = CEPH_INO_LOST_AND_FOUND; + lf_ino.inode.version = 1; + lf_ino.inode.backtrace_version = 1; + r = inject_linkage(CEPH_INO_ROOT, "lost+found", frag_t(), lf_ino); + if (r < 0) { + return r; + } + } else { + if (!(lf_ino.inode.mode & S_IFDIR)) { + derr << "lost+found exists but is not a directory!" << dendl; + // In this case we error out, and the user should do something about + // this problem. + return -EINVAL; + } + } + + r = find_or_create_dirfrag(CEPH_INO_LOST_AND_FOUND, frag_t(), &created); + if (r < 0) { + return r; + } + + InodeStore recovered_ino; + recovered_ino.inode.mode = 0500 | S_IFREG; + recovered_ino.inode.size = file_size; + recovered_ino.inode.max_size_ever = file_size; + recovered_ino.inode.mtime.tv.tv_sec = file_mtime; + recovered_ino.inode.atime.tv.tv_sec = file_mtime; + recovered_ino.inode.ctime.tv.tv_sec = file_mtime; + + recovered_ino.inode.layout = g_default_file_layout; + recovered_ino.inode.layout.fl_object_size = chunk_size; + recovered_ino.inode.layout.fl_stripe_unit = chunk_size; + recovered_ino.inode.layout.fl_pg_pool = data_pool_id; + + recovered_ino.inode.truncate_seq = 1; + recovered_ino.inode.truncate_size = -1ull; + + recovered_ino.inode.inline_data.version = CEPH_INLINE_NONE; + + recovered_ino.inode.nlink = 1; + recovered_ino.inode.ino = ino; + recovered_ino.inode.version = 1; + recovered_ino.inode.backtrace_version = 1; + + const std::string dname = lost_found_dname(ino); + + // Write dentry into lost+found dirfrag + return inject_linkage(lf_ino.inode.ino, dname, frag_t(), recovered_ino); +} + + +int MetadataDriver::get_frag_of( + inodeno_t dirino, + const std::string &target_dname, + frag_t *result_ft) +{ + object_t root_frag_oid = InodeStore::get_object_name(dirino, frag_t(), ""); + + dout(20) << "dirino=" << dirino << " target_dname=" << target_dname << dendl; + + // Find and load fragtree if existing dirfrag + // ========================================== + bool have_backtrace = false; + bufferlist parent_bl; + int r = metadata_io.getxattr(root_frag_oid.name, "parent", parent_bl); + if (r == -ENODATA) { + dout(10) << "No backtrace on '" << root_frag_oid << "'" << dendl; + } else if (r < 0) { + dout(4) << "Unexpected error on '" << root_frag_oid << "': " + << cpp_strerror(r) << dendl; + return r; + } + + // Deserialize backtrace + inode_backtrace_t backtrace; + if (parent_bl.length()) { + try { + bufferlist::iterator q = parent_bl.begin(); + backtrace.decode(q); + have_backtrace = true; + } catch (buffer::error &e) { + dout(4) << "Corrupt backtrace on '" << root_frag_oid << "': " << e << dendl; + } + } + + if (!(have_backtrace && backtrace.ancestors.size())) { + // Can't work out fragtree without a backtrace + dout(4) << "No backtrace on '" << root_frag_oid + << "': cannot determine fragtree" << dendl; + return -ENOENT; + } + + // The parentage of dirino + const inode_backpointer_t &bp = *(backtrace.ancestors.begin()); + + // The inode of dirino's parent + const inodeno_t parent_ino = bp.dirino; + + // The dname of dirino in its parent. + const std::string &parent_dname = bp.dname; + + dout(20) << "got backtrace parent " << parent_ino << "/" + << parent_dname << dendl; + + // The primary dentry for dirino + InodeStore existing_dentry; + + // See if we can find ourselves in dirfrag zero of the parent: this + // is a fast path that avoids needing to go further up the tree + // if the parent isn't fragmented (worst case we would have to + // go all the way to the root) + r = read_dentry(parent_ino, frag_t(), parent_dname, &existing_dentry); + if (r >= 0) { + // Great, fast path: return the fragtree from here + if (existing_dentry.inode.ino != dirino) { + dout(4) << "Unexpected inode in dentry! 0x" << std::hex + << existing_dentry.inode.ino + << " vs expected 0x" << dirino << std::dec << dendl; + return -ENOENT; + } + dout(20) << "fast path, fragtree is " + << existing_dentry.dirfragtree << dendl; + *result_ft = existing_dentry.pick_dirfrag(target_dname); + dout(20) << "frag is " << *result_ft << dendl; + return 0; + } else if (r != -ENOENT) { + // Dentry not present in 0th frag, must read parent's fragtree + frag_t parent_frag; + r = get_frag_of(parent_ino, parent_dname, &parent_frag); + if (r == 0) { + // We have the parent fragtree, so try again to load our dentry + r = read_dentry(parent_ino, parent_frag, parent_dname, &existing_dentry); + if (r >= 0) { + // Got it! + *result_ft = existing_dentry.pick_dirfrag(target_dname); + dout(20) << "resolved via parent, frag is " << *result_ft << dendl; + return 0; + } else { + if (r == -EINVAL || r == -ENOENT) { + return -ENOENT; // dentry missing or corrupt, so frag is missing + } else { + return r; + } + } + } else { + // Couldn't resolve parent fragtree, so can't find ours. + return r; + } + } else if (r == -EINVAL) { + // Unreadable dentry, can't know the fragtree. + return -ENOENT; + } else { + // Unexpected error, raise it + return r; + } +} + + +int MetadataDriver::inject_with_backtrace( + const inode_backtrace_t &backtrace, uint64_t file_size, time_t file_mtime, + uint32_t chunk_size, int64_t data_pool_id) + +{ + + // On dirfrags + // =========== + // In order to insert something into a directory, we first (ideally) + // need to know the fragtree for the directory. Sometimes we can't + // get that, in which case we just go ahead and insert it into + // fragment zero for a good chance of that being the right thing + // anyway (most moderate-sized dirs aren't fragmented!) + + // On ancestry + // =========== + // My immediate ancestry should be correct, so if we can find that + // directory's dirfrag then go inject it there. This works well + // in the case that this inode's dentry was somehow lost and we + // are recreating it, because the rest of the hierarchy + // will probably still exist. + // + // It's more of a "better than nothing" approach when rebuilding + // a whole tree, as backtraces will in general not be up to date + // beyond the first parent, if anything in the trace was ever + // moved after the file was created. + + // On inode numbers + // ================ + // The backtrace tells us inodes for each of the parents. If we are + // creating those parent dirfrags, then there is a risk that somehow + // the inode indicated here was also used for data (not a dirfrag) at + // some stage. That would be a zany situation, and we don't check + // for it here, because to do so would require extra IOs for everything + // we inject, and anyway wouldn't guarantee that the inode number + // wasn't in use in some dentry elsewhere in the metadata tree that + // just happened not to have any data objects. + + // On multiple workers touching the same traces + // ============================================ + // When creating linkage for a directory, *only* create it if we are + // also creating the object. That way, we might not manage to get the + // *right* linkage for a directory, but at least we won't multiply link + // it. We assume that if a root dirfrag exists for a directory, then + // it is linked somewhere (i.e. that the metadata pool is not already + // inconsistent). + // + // Making sure *that* is true is someone else's job! Probably someone + // who is not going to run in parallel, so that they can self-consistently + // look at versions and move things around as they go. + // Note this isn't 100% safe: if we die immediately after creating dirfrag + // object, next run will fail to create linkage for the dirfrag object + // and leave it orphaned. + + inodeno_t ino = backtrace.ino; + dout(10) << " inode: 0x" << std::hex << ino << std::dec << dendl; + for (std::vector::const_iterator i = backtrace.ancestors.begin(); + i != backtrace.ancestors.end(); ++i) { + const inode_backpointer_t &backptr = *i; + dout(10) << " backptr: 0x" << std::hex << backptr.dirino << std::dec + << "/" << backptr.dname << dendl; + + // Examine root dirfrag for parent + const inodeno_t parent_ino = backptr.dirino; + const std::string dname = backptr.dname; + + frag_t fragment; + int r = get_frag_of(parent_ino, dname, &fragment); + if (r == -ENOENT) { + // Don't know fragment, fall back to assuming root + dout(20) << "don't know fragment for 0x" << std::hex << + parent_ino << std::dec << "/" << dname << ", will insert to root" + << dendl; + } + + // Find or create dirfrag + // ====================== + bool created_dirfrag; + r = find_or_create_dirfrag(parent_ino, fragment, &created_dirfrag); + if (r < 0) { + return r; + } + + // Check if dentry already exists + // ============================== + InodeStore existing_dentry; + r = read_dentry(parent_ino, fragment, dname, &existing_dentry); + bool write_dentry = false; + if (r == -ENOENT || r == -EINVAL) { + if (r == -EINVAL && !force_corrupt) { + return r; + } + // Missing or corrupt dentry + write_dentry = true; + } else if (r < 0) { + derr << "Unexpected error reading dentry 0x" << std::hex + << parent_ino << std::dec << "/" + << dname << ": " << cpp_strerror(r) << dendl; + break; + } else { + // Dentry already present, does it link to me? + if (existing_dentry.inode.ino == ino) { + dout(20) << "Dentry 0x" << std::hex + << parent_ino << std::dec << "/" + << dname << " already exists and points to me" << dendl; + } else { + derr << "Dentry 0x" << std::hex + << parent_ino << std::dec << "/" + << dname << " already exists but points to 0x" + << std::hex << existing_dentry.inode.ino << std::dec << dendl; + // Fall back to lost+found! + return inject_lost_and_found(backtrace.ino, file_size, file_mtime, + chunk_size, data_pool_id); + } + } + + // Inject linkage + // ============== + if (write_dentry) { + InodeStore dentry; + if (i == backtrace.ancestors.begin()) { + // This is the linkage for a file + dentry.inode.mode = 0500 | S_IFREG; + dout(10) << "Linking inode 0x" << std::hex << ino + << " at 0x" << parent_ino << "/" << dname << std::dec + << " with size=" << file_size << " bytes" << dendl; + + // The file size and mtime we learned by scanning globally + dentry.inode.size = file_size; + dentry.inode.max_size_ever = file_size; + dentry.inode.mtime.tv.tv_sec = file_mtime; + dentry.inode.atime.tv.tv_sec = file_mtime; + dentry.inode.ctime.tv.tv_sec = file_mtime; + + dentry.inode.layout = g_default_file_layout; + dentry.inode.layout.fl_object_size = chunk_size; + dentry.inode.layout.fl_stripe_unit = chunk_size; + dentry.inode.layout.fl_pg_pool = data_pool_id; + + dentry.inode.truncate_seq = 1; + dentry.inode.truncate_size = -1ull; + + dentry.inode.inline_data.version = CEPH_INLINE_NONE; + } else { + // This is the linkage for a directory + dentry.inode.mode = 0755 | S_IFDIR; + + // Set nfiles to something non-zero, to fool any other code + // that tries to ignore 'empty' directories. This won't be + // accurate, but it should avoid functional issues. + dentry.inode.dirstat.nfiles = 1; + dentry.inode.size = 1; + + } + dentry.inode.nlink = 1; + dentry.inode.ino = ino; + dentry.inode.version = 1; + dentry.inode.backtrace_version = 1; + r = inject_linkage(parent_ino, dname, fragment, dentry); + if (r < 0) { + return r; + } + } + + if (!created_dirfrag) { + // If the parent dirfrag already existed, then stop traversing the + // backtrace: assume that the other ancestors already exist too. This + // is an assumption rather than a truth, but it's a convenient way + // to avoid the risk of creating multiply-linked directories while + // injecting data. If there are in fact missing ancestors, this + // should be fixed up using a separate tool scanning the metadata + // pool. + break; + } else { + // Proceed up the backtrace, creating parents + ino = parent_ino; + } + } + + return 0; +} + +int MetadataDriver::find_or_create_dirfrag( + inodeno_t ino, + frag_t fragment, + bool *created) +{ + assert(created != NULL); + + fnode_t existing_fnode; + *created = false; + + uint64_t read_version = 0; + int r = read_fnode(ino, fragment, &existing_fnode, &read_version); + dout(10) << "read_version = " << read_version << dendl; + + if (r == -ENOENT || r == -EINVAL) { + if (r == -EINVAL && !force_corrupt) { + return r; + } + + // Missing or corrupt fnode, create afresh + bufferlist fnode_bl; + fnode_t blank_fnode; + blank_fnode.version = 1; + blank_fnode.damage_flags |= (DAMAGE_RSTATS | DAMAGE_RSTATS); + blank_fnode.encode(fnode_bl); + + + librados::ObjectWriteOperation op; + + if (read_version) { + assert(r == -EINVAL); + // Case A: We must assert that the version isn't changed since we saw the object + // was unreadable, to avoid the possibility of two data-scan processes + // both creating the frag. + op.assert_version(read_version); + } else { + assert(r == -ENOENT); + // Case B: The object didn't exist in read_fnode, so while creating it we must + // use an exclusive create to correctly populate *creating with + // whether we created it ourselves or someone beat us to it. + op.create(true); + } + + object_t frag_oid = InodeStore::get_object_name(ino, fragment, ""); + op.omap_set_header(fnode_bl); + r = metadata_io.operate(frag_oid.name, &op); + if (r == -EOVERFLOW) { + // Someone else wrote it (see case A above) + dout(10) << "Dirfrag creation race: 0x" << std::hex + << ino << " " << fragment << std::dec << dendl; + *created = false; + return 0; + } else if (r < 0) { + // We were unable to create or write it, error out + derr << "Failed to create dirfrag 0x" << std::hex + << ino << std::dec << ": " << cpp_strerror(r) << dendl; + return r; + } else { + // Success: the dirfrag object now exists with a value header + dout(10) << "Created dirfrag: 0x" << std::hex + << ino << std::dec << dendl; + *created = true; + } + } else if (r < 0) { + derr << "Unexpected error reading dirfrag 0x" << std::hex + << ino << std::dec << " : " << cpp_strerror(r) << dendl; + return r; + } else { + dout(20) << "Dirfrag already exists: 0x" << std::hex + << ino << " " << fragment << std::dec << dendl; + } + + return 0; +} + +int MetadataDriver::inject_linkage( + inodeno_t dir_ino, const std::string &dname, + const frag_t fragment, const InodeStore &inode) +{ + // We have no information about snapshots, so everything goes + // in as CEPH_NOSNAP + snapid_t snap = CEPH_NOSNAP; + + object_t frag_oid = InodeStore::get_object_name(dir_ino, fragment, ""); + + std::string key; + dentry_key_t dn_key(snap, dname.c_str()); + dn_key.encode(key); + + bufferlist dentry_bl; + ::encode(snap, dentry_bl); + ::encode('I', dentry_bl); + inode.encode_bare(dentry_bl); + + // Write out + std::map vals; + vals[key] = dentry_bl; + int r = metadata_io.omap_set(frag_oid.name, vals); + if (r != 0) { + derr << "Error writing dentry 0x" << std::hex + << dir_ino << std::dec << "/" + << dname << ": " << cpp_strerror(r) << dendl; + return r; + } else { + dout(20) << "Injected dentry 0x" << std::hex + << dir_ino << "/" << dname << " pointing to 0x" + << inode.inode.ino << std::dec << dendl; + return 0; + } +} + + +int MetadataDriver::init(librados::Rados &rados, const MDSMap *mdsmap) +{ + int const metadata_pool_id = mdsmap->get_metadata_pool(); + + dout(4) << "resolving metadata pool " << metadata_pool_id << dendl; + std::string metadata_pool_name; + int r = rados.pool_reverse_lookup(metadata_pool_id, &metadata_pool_name); + if (r < 0) { + derr << "Pool " << metadata_pool_id + << " identified in MDS map not found in RADOS!" << dendl; + return r; + } + dout(4) << "found metadata pool '" << metadata_pool_name << "'" << dendl; + return rados.ioctx_create(metadata_pool_name.c_str(), metadata_io); +} + +int LocalFileDriver::init(librados::Rados &rados, const MDSMap *mdsmap) +{ + return 0; +} + +int LocalFileDriver::inject_data( + const std::string &file_path, + uint64_t size, + uint32_t chunk_size, + inodeno_t ino) +{ + // Scrape the file contents out of the data pool and into the + // local filesystem + std::fstream f; + f.open(file_path.c_str(), std::fstream::out | std::fstream::binary); + + for (uint64_t offset = 0; offset < size; offset += chunk_size) { + bufferlist bl; + + char buf[32]; + snprintf(buf, sizeof(buf), + "%llx.%08llx", + (unsigned long long)ino, + (unsigned long long)(offset / chunk_size)); + std::string oid(buf); + + int r = data_io.read(oid, bl, chunk_size, 0); + + if (r <= 0 && r != -ENOENT) { + derr << "error reading data object '" << oid << "': " + << cpp_strerror(r) << dendl; + f.close(); + return r; + } else if (r >=0) { + + f.seekp(offset); + bl.write_stream(f); + } + } + f.close(); + + return 0; +} + + +int LocalFileDriver::inject_with_backtrace( + const inode_backtrace_t &bt, + uint64_t size, + time_t mtime, + uint32_t chunk_size, + int64_t data_pool_id) +{ + std::string path_builder = path; + + // Iterate through backtrace creating directory parents + std::vector::const_reverse_iterator i; + for (i = bt.ancestors.rbegin(); + i != bt.ancestors.rend(); ++i) { + + const inode_backpointer_t &backptr = *i; + path_builder += "/"; + path_builder += backptr.dname; + + // Last entry is the filename itself + bool is_file = (i + 1 == bt.ancestors.rend()); + if (is_file) { + inject_data(path_builder, size, chunk_size, bt.ino); + } else { + int r = mkdir(path_builder.c_str(), 0755); + if (r != 0 && r != -EPERM) { + derr << "error creating directory: '" << path_builder << "': " + << cpp_strerror(r) << dendl; + return r; + } + } + } + + return 0; +} + +int LocalFileDriver::inject_lost_and_found( + inodeno_t ino, + uint64_t size, + time_t mtime, + uint32_t chunk_size, + int64_t data_pool_id) +{ + std::string lf_path = path + "/lost+found"; + int r = mkdir(lf_path.c_str(), 0755); + if (r != 0 && r != -EPERM) { + derr << "error creating directory: '" << lf_path << "': " + << cpp_strerror(r) << dendl; + return r; + } + + std::string file_path = lf_path + "/" + lost_found_dname(ino); + return inject_data(file_path, size, chunk_size, ino); +} + +int LocalFileDriver::init_roots(int64_t data_pool_id) +{ + // Ensure that the path exists and is a directory + bool exists; + int r = check_roots(&exists); + if (r != 0) { + return r; + } + + if (exists) { + return 0; + } else { + return ::mkdir(path.c_str(), 0755); + } +} + +int LocalFileDriver::check_roots(bool *result) +{ + // Check if the path exists and is a directory + DIR *d = ::opendir(path.c_str()); + if (d == NULL) { + *result = false; + } else { + int r = closedir(d); + if (r != 0) { + // Weird, but maybe possible with e.g. stale FD on NFS mount? + *result = false; + } else { + *result = true; + } + } + + return 0; +} + diff --git a/src/tools/cephfs/DataScan.h b/src/tools/cephfs/DataScan.h new file mode 100644 index 00000000000..054ce81edce --- /dev/null +++ b/src/tools/cephfs/DataScan.h @@ -0,0 +1,278 @@ +// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- +// vim: ts=8 sw=2 smarttab +/* + * Ceph - scalable distributed file system + * + * Copyright (C) 2015 Red Hat + * + * This is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License version 2.1, as published by the Free Software + * Foundation. See file COPYING. + * + */ + + +#include "MDSUtility.h" +#include "include/rados/librados.hpp" + +class InodeStore; + +class RecoveryDriver { + protected: + // If true, overwrite structures that generate decoding errors. + bool force_corrupt; + + // If true, overwrite root objects during init_roots even if they + // exist + bool force_init; + + public: + virtual int init(librados::Rados &rados, const MDSMap *mdsmap) = 0; + + void set_force_corrupt(const bool val) + { + force_corrupt = val; + } + + void set_force_init(const bool val) + { + force_init = val; + } + + + /** + * Inject an inode + dentry parents into the metadata pool, + * based on a backtrace recovered from the data pool + */ + virtual int inject_with_backtrace( + const inode_backtrace_t &bt, + uint64_t size, + time_t mtime, + uint32_t chunk_size, + int64_t data_pool_id) = 0; + + /** + * Inject an inode + dentry into the lost+found directory, + * when all we know about a file is its inode. + */ + virtual int inject_lost_and_found( + inodeno_t ino, + uint64_t size, + time_t mtime, + uint32_t chunk_size, + int64_t data_pool_id) = 0; + + /** + * Create any missing roots (i.e. mydir, strays, root inode) + */ + virtual int init_roots( + int64_t data_pool_id) = 0; + + /** + * Pre-injection check that all the roots are present in + * the metadata pool. Used to avoid parallel workers interfering + * with one another, by cueing the user to go run 'init' on a + * single node before running a parallel scan. + * + * @param result: set to true if roots are present, else set to false + * @returns 0 on no unexpected errors, else error code. Missing objects + * are not considered an unexpected error: check *result for + * this case. + */ + virtual int check_roots(bool *result) = 0; + + /** + * Helper to compose dnames for links to lost+found + * inodes. + */ + std::string lost_found_dname(inodeno_t ino) + { + char s[20]; + snprintf(s, sizeof(s), "%llx", (unsigned long long)ino); + return std::string(s); + } + + RecoveryDriver() + : force_corrupt(false) + {} + + virtual ~RecoveryDriver() {} +}; + +class LocalFileDriver : public RecoveryDriver +{ + protected: + const std::string path; + librados::IoCtx &data_io; + + int inject_data( + const std::string &file_path, + uint64_t size, + uint32_t chunk_size, + inodeno_t ino); + public: + + LocalFileDriver(const std::string &path_, librados::IoCtx &data_io_) + : RecoveryDriver(), path(path_), data_io(data_io_) + {} + + // Implement RecoveryDriver interface + int init(librados::Rados &rados, const MDSMap *mdsmap); + + int inject_with_backtrace( + const inode_backtrace_t &bt, + uint64_t size, + time_t mtime, + uint32_t chunk_size, + int64_t data_pool_id); + + int inject_lost_and_found( + inodeno_t ino, + uint64_t size, + time_t mtime, + uint32_t chunk_size, + int64_t data_pool_id); + + int init_roots(int64_t data_pool_id); + + int check_roots(bool *result); +}; + +/** + * A class that knows how to manipulate CephFS metadata pools + */ +class MetadataDriver : public RecoveryDriver +{ + protected: + + librados::IoCtx metadata_io; + + /** + * Create a .inode object, i.e. root or mydir + */ + int inject_unlinked_inode(inodeno_t inono, int mode, int64_t data_pool_id); + + /** + * Check for existence of .inode objects, before + * trying to go ahead and inject metadata. + */ + int root_exists(inodeno_t ino, bool *result); + + /** + * Try and read an fnode from a dirfrag + */ + int read_fnode(inodeno_t ino, frag_t frag, + fnode_t *fnode, uint64_t *read_version); + + /** + * Try and read a dentry from a dirfrag + */ + int read_dentry(inodeno_t parent_ino, frag_t frag, + const std::string &dname, InodeStore *inode); + + int find_or_create_dirfrag( + inodeno_t ino, + frag_t fragment, + bool *created); + + int inject_linkage( + inodeno_t dir_ino, const std::string &dname, + const frag_t fragment, const InodeStore &inode); + + /** + * Work out which fragment of a directory should contain a named + * dentry, recursing up the trace as necessary to retrieve + * fragtrees. + */ + int get_frag_of( + inodeno_t dirino, + const std::string &dname, + frag_t *result_ft); + + public: + + // Implement RecoveryDriver interface + int init(librados::Rados &rados, const MDSMap *mdsmap); + + int inject_with_backtrace( + const inode_backtrace_t &bt, + uint64_t size, + time_t mtime, + uint32_t chunk_size, + int64_t data_pool_id); + + int inject_lost_and_found( + inodeno_t ino, + uint64_t size, + time_t mtime, + uint32_t chunk_size, + int64_t data_pool_id); + + int init_roots(int64_t data_pool_id); + + int check_roots(bool *result); +}; + +class DataScan : public MDSUtility +{ + protected: + RecoveryDriver *driver; + + // IoCtx for data pool (where we scrape backtraces from) + librados::IoCtx data_io; + // Remember the data pool ID for use in layouts + int64_t data_pool_id; + + uint32_t n; + uint32_t m; + + /** + * Scan data pool for backtraces, and inject inodes to metadata pool + */ + int scan_inodes(); + + /** + * Scan data pool for file sizes and mtimes + */ + int scan_extents(); + + // Accept pools which are not in the MDSMap + bool force_pool; + // Respond to decode errors by overwriting + bool force_corrupt; + // Overwrite root objects even if they exist + bool force_init; + + /** + * @param r set to error on valid key with invalid value + * @return true if argument consumed, else false + */ + bool parse_kwarg( + const std::vector &args, + std::vector::const_iterator &i, + int *r); + + /** + * @return true if argument consumed, else false + */ + bool parse_arg( + const std::vector &arg, + std::vector::const_iterator &i); + + public: + void usage(); + int main(const std::vector &args); + + DataScan() + : driver(NULL), data_pool_id(-1), n(0), m(1), + force_pool(false) + { + } + + ~DataScan() + { + delete driver; + } +}; + diff --git a/src/tools/cephfs/cephfs-data-scan.cc b/src/tools/cephfs/cephfs-data-scan.cc new file mode 100644 index 00000000000..5e9c05a0717 --- /dev/null +++ b/src/tools/cephfs/cephfs-data-scan.cc @@ -0,0 +1,45 @@ + +#include "include/types.h" +#include "common/config.h" +#include "common/ceph_argparse.h" +#include "common/errno.h" +#include "global/global_init.h" + +#include "DataScan.h" + + +int main(int argc, const char **argv) +{ + vector args; + argv_to_vec(argc, argv, args); + env_to_vec(args); + + global_init(NULL, args, CEPH_ENTITY_TYPE_CLIENT, CODE_ENVIRONMENT_UTILITY, 0); + common_init_finish(g_ceph_context); + + DataScan data_scan; + + // Handle --help before calling init() so we don't depend on network. + if (args.empty() || (args.size() == 1 && (std::string(args[0]) == "--help" || std::string(args[0]) == "-h"))) { + data_scan.usage(); + return 0; + } + + // Connect to mon cluster, download MDS map etc + int rc = data_scan.init(); + if (rc != 0) { + std::cerr << "Error in initialization: " << cpp_strerror(rc) << std::endl; + return rc; + } + + // Finally, execute the user's commands + rc = data_scan.main(args); + if (rc != 0) { + std::cerr << "Error (" << cpp_strerror(rc) << ")" << std::endl; + } + + data_scan.shutdown(); + + return rc; +} +