| Line |
Branch |
Exec |
Source |
| 1 |
|
|
/** |
| 2 |
|
|
* This file is part of the CernVM File System. |
| 3 |
|
|
* |
| 4 |
|
|
* The NFS maps module maintains inode -- path relations. An inode that is |
| 5 |
|
|
* issued once by an NFS exported file system might be asked for |
| 6 |
|
|
* any time later by clients. |
| 7 |
|
|
* |
| 8 |
|
|
* In "NFS mode", cvmfs will issue inodes consecutively and reuse inodes |
| 9 |
|
|
* based on path name. The inode --> path and path --> inode maps are |
| 10 |
|
|
* handled by sqlite. This workaround is comparable to the Fuse "noforget" |
| 11 |
|
|
* option, except that the mappings are persistent and thus consistent during |
| 12 |
|
|
* cvmfs restarts. Also, sqlite allows for restricting the memory consumption. |
| 13 |
|
|
* |
| 14 |
|
|
* The maps are not accounted for by the cache quota. |
| 15 |
|
|
*/ |
| 16 |
|
|
|
| 17 |
|
|
#include "nfs_maps_sqlite.h" |
| 18 |
|
|
|
| 19 |
|
|
#include <unistd.h> |
| 20 |
|
|
|
| 21 |
|
|
#include <cassert> |
| 22 |
|
|
#include <cstddef> |
| 23 |
|
|
#include <cstdlib> |
| 24 |
|
|
|
| 25 |
|
|
#include "statistics.h" |
| 26 |
|
|
#include "util/concurrency.h" |
| 27 |
|
|
#include "util/exception.h" |
| 28 |
|
|
#include "util/logging.h" |
| 29 |
|
|
#include "util/pointer.h" |
| 30 |
|
|
#include "util/posix.h" |
| 31 |
|
|
#include "util/prng.h" |
| 32 |
|
|
#include "util/smalloc.h" |
| 33 |
|
|
#include "util/string.h" |
| 34 |
|
|
|
| 35 |
|
|
using namespace std; // NOLINT |
| 36 |
|
|
|
| 37 |
|
|
|
| 38 |
|
|
const char *NfsMapsSqlite::kSqlCreateTable = "CREATE TABLE IF NOT EXISTS " |
| 39 |
|
|
"inodes (path TEXT PRIMARY KEY);"; |
| 40 |
|
|
const char *NfsMapsSqlite:: |
| 41 |
|
|
kSqlAddRoot = "INSERT INTO inodes (oid, path) VALUES (?, ?);"; |
| 42 |
|
|
const char *NfsMapsSqlite::kSqlAddInode = "INSERT INTO inodes VALUES (?);"; |
| 43 |
|
|
const char |
| 44 |
|
|
*NfsMapsSqlite::kSqlGetInode = "SELECT rowid FROM inodes where path = ?;"; |
| 45 |
|
|
const char |
| 46 |
|
|
*NfsMapsSqlite::kSqlGetPath = "SELECT path FROM inodes where rowid = ?;"; |
| 47 |
|
|
|
| 48 |
|
|
|
| 49 |
|
✗ |
int NfsMapsSqlite::BusyHandler(void *data, int attempt) { |
| 50 |
|
✗ |
BusyHandlerInfo *handler_info = static_cast<BusyHandlerInfo *>(data); |
| 51 |
|
|
// Reset the accumulated time if this is the start of a new request |
| 52 |
|
✗ |
if (attempt == 0) |
| 53 |
|
✗ |
handler_info->accumulated_ms = 0; |
| 54 |
|
✗ |
LogCvmfs(kLogNfsMaps, kLogDebug, |
| 55 |
|
|
"busy handler, attempt %d, accumulated waiting time %u", attempt, |
| 56 |
|
|
handler_info->accumulated_ms); |
| 57 |
|
✗ |
if (handler_info->accumulated_ms >= handler_info->kMaxWaitMs) |
| 58 |
|
✗ |
return 0; |
| 59 |
|
|
|
| 60 |
|
✗ |
const unsigned backoff_range_ms = 1 << attempt; |
| 61 |
|
✗ |
unsigned backoff_ms = handler_info->prng.Next(backoff_range_ms); |
| 62 |
|
✗ |
if (handler_info->accumulated_ms + backoff_ms > handler_info->kMaxWaitMs) { |
| 63 |
|
✗ |
backoff_ms = handler_info->kMaxWaitMs - handler_info->accumulated_ms; |
| 64 |
|
|
} |
| 65 |
|
✗ |
if (backoff_ms > handler_info->kMaxBackoffMs) { |
| 66 |
|
✗ |
backoff_ms = handler_info->kMaxBackoffMs; |
| 67 |
|
|
} |
| 68 |
|
|
|
| 69 |
|
✗ |
SafeSleepMs(backoff_ms); |
| 70 |
|
✗ |
handler_info->accumulated_ms += backoff_ms; |
| 71 |
|
✗ |
return 1; |
| 72 |
|
|
} |
| 73 |
|
|
|
| 74 |
|
|
|
| 75 |
|
✗ |
NfsMapsSqlite *NfsMapsSqlite::Create(const string &db_dir, |
| 76 |
|
|
const uint64_t root_inode, |
| 77 |
|
|
const bool rebuild, |
| 78 |
|
|
perf::Statistics *statistics) { |
| 79 |
|
✗ |
assert(root_inode > 0); |
| 80 |
|
✗ |
UniquePtr<NfsMapsSqlite> maps(new NfsMapsSqlite()); |
| 81 |
|
✗ |
maps->n_db_added_ = statistics->Register("nfs.sqlite.n_added", |
| 82 |
|
|
"total number of issued inode"); |
| 83 |
|
✗ |
maps->n_db_seq_ = statistics->Register("nfs.sqlite.n_seq", |
| 84 |
|
|
"last inode issued"); |
| 85 |
|
✗ |
maps->n_db_path_found_ = statistics->Register("nfs.sqlite.n_path_hit", |
| 86 |
|
|
"inode --> path hits"); |
| 87 |
|
✗ |
maps->n_db_inode_found_ = statistics->Register("nfs.sqlite.n_inode_hit", |
| 88 |
|
|
"path --> inode hits"); |
| 89 |
|
|
|
| 90 |
|
✗ |
const string db_path = db_dir + "/inode_maps.db"; |
| 91 |
|
|
|
| 92 |
|
|
sqlite3_stmt *stmt; |
| 93 |
|
✗ |
if (rebuild) { |
| 94 |
|
✗ |
LogCvmfs(kLogNfsMaps, kLogDebug | kLogSyslogWarn, |
| 95 |
|
|
"Ignoring rebuild flag as this may crash other cluster nodes."); |
| 96 |
|
|
} |
| 97 |
|
|
// We don't want the shared cache, we want minimal caching so sync is kept |
| 98 |
|
✗ |
int retval = sqlite3_enable_shared_cache(0); |
| 99 |
|
✗ |
assert(retval == SQLITE_OK); |
| 100 |
|
|
|
| 101 |
|
✗ |
retval = sqlite3_open_v2( |
| 102 |
|
✗ |
db_path.c_str(), &maps->db_, |
| 103 |
|
|
SQLITE_OPEN_NOMUTEX | SQLITE_OPEN_READWRITE | SQLITE_OPEN_CREATE, NULL); |
| 104 |
|
✗ |
if (retval != SQLITE_OK) { |
| 105 |
|
✗ |
LogCvmfs(kLogNfsMaps, kLogDebug, "Failed to create inode_maps file (%s)", |
| 106 |
|
|
db_path.c_str()); |
| 107 |
|
✗ |
return NULL; |
| 108 |
|
|
} |
| 109 |
|
|
// Be prepared to wait for up to 1 minute for transactions to complete |
| 110 |
|
|
// Being stuck for a long time is far more favorable than failing |
| 111 |
|
|
// TODO(jblomer): another busy handler. This one conflicts with SIGALRM |
| 112 |
|
✗ |
retval = sqlite3_busy_handler(maps->db_, BusyHandler, |
| 113 |
|
✗ |
&maps->busy_handler_info_); |
| 114 |
|
✗ |
assert(retval == SQLITE_OK); |
| 115 |
|
|
|
| 116 |
|
|
// Set-up the main inode table if it doesn't exist |
| 117 |
|
✗ |
retval = sqlite3_prepare_v2(maps->db_, kSqlCreateTable, -1, &stmt, NULL); |
| 118 |
|
✗ |
if (retval != SQLITE_OK) { |
| 119 |
|
✗ |
LogCvmfs(kLogNfsMaps, kLogDebug | kLogSyslogErr, |
| 120 |
|
|
"Failed to prepare create table statement: %s", |
| 121 |
|
✗ |
sqlite3_errmsg(maps->db_)); |
| 122 |
|
✗ |
return NULL; |
| 123 |
|
|
} |
| 124 |
|
✗ |
if (sqlite3_step(stmt) != SQLITE_DONE) { |
| 125 |
|
✗ |
LogCvmfs(kLogNfsMaps, kLogSyslogErr, |
| 126 |
|
|
"Failed to create main inode table: %s", |
| 127 |
|
✗ |
sqlite3_errmsg(maps->db_)); |
| 128 |
|
✗ |
sqlite3_finalize(stmt); |
| 129 |
|
✗ |
return NULL; |
| 130 |
|
|
} |
| 131 |
|
✗ |
sqlite3_finalize(stmt); |
| 132 |
|
|
|
| 133 |
|
|
// Prepare lookup and add-inode statements |
| 134 |
|
✗ |
retval = sqlite3_prepare_v2(maps->db_, kSqlGetPath, -1, &maps->stmt_get_path_, |
| 135 |
|
|
NULL); |
| 136 |
|
✗ |
assert(retval == SQLITE_OK); |
| 137 |
|
✗ |
retval = sqlite3_prepare_v2(maps->db_, kSqlGetInode, -1, |
| 138 |
|
✗ |
&maps->stmt_get_inode_, NULL); |
| 139 |
|
✗ |
assert(retval == SQLITE_OK); |
| 140 |
|
✗ |
retval = sqlite3_prepare_v2(maps->db_, kSqlAddInode, -1, &maps->stmt_add_, |
| 141 |
|
|
NULL); |
| 142 |
|
✗ |
assert(retval == SQLITE_OK); |
| 143 |
|
|
|
| 144 |
|
|
// Check the root inode exists, if not create it |
| 145 |
|
✗ |
const PathString rootpath("", 0); |
| 146 |
|
✗ |
if (!maps->FindInode(rootpath)) { |
| 147 |
|
✗ |
retval = sqlite3_prepare_v2(maps->db_, kSqlAddRoot, -1, &stmt, NULL); |
| 148 |
|
✗ |
assert(retval == SQLITE_OK); |
| 149 |
|
✗ |
retval = sqlite3_bind_int64(stmt, 1, root_inode); |
| 150 |
|
✗ |
assert(retval == SQLITE_OK); |
| 151 |
|
✗ |
retval = sqlite3_bind_text(stmt, 2, "", 0, SQLITE_TRANSIENT); |
| 152 |
|
✗ |
assert(retval == SQLITE_OK); |
| 153 |
|
✗ |
if (sqlite3_step(stmt) != SQLITE_DONE) { |
| 154 |
|
✗ |
PANIC(kLogDebug | kLogSyslogErr, "Failed to execute CreateRoot: %s", |
| 155 |
|
|
sqlite3_errmsg(maps->db_)); |
| 156 |
|
|
} |
| 157 |
|
✗ |
sqlite3_finalize(stmt); |
| 158 |
|
|
} |
| 159 |
|
|
|
| 160 |
|
✗ |
return maps.Release(); |
| 161 |
|
|
} |
| 162 |
|
|
|
| 163 |
|
|
|
| 164 |
|
|
/** |
| 165 |
|
|
* Finds an inode by path |
| 166 |
|
|
* \return inode number, 0 if path not found |
| 167 |
|
|
*/ |
| 168 |
|
✗ |
uint64_t NfsMapsSqlite::FindInode(const PathString &path) { |
| 169 |
|
|
int sqlite_state; |
| 170 |
|
|
uint64_t inode; |
| 171 |
|
✗ |
sqlite_state = sqlite3_bind_text(stmt_get_inode_, 1, path.GetChars(), |
| 172 |
|
✗ |
path.GetLength(), SQLITE_TRANSIENT); |
| 173 |
|
✗ |
assert(sqlite_state == SQLITE_OK); |
| 174 |
|
✗ |
sqlite_state = sqlite3_step(stmt_get_inode_); |
| 175 |
|
✗ |
if (sqlite_state == SQLITE_DONE) { |
| 176 |
|
|
// Path not found in DB |
| 177 |
|
✗ |
sqlite3_reset(stmt_get_inode_); |
| 178 |
|
✗ |
return 0; |
| 179 |
|
|
} |
| 180 |
|
✗ |
if (sqlite_state != SQLITE_ROW) { |
| 181 |
|
✗ |
LogCvmfs(kLogNfsMaps, kLogDebug, "Error finding inode (%s): %s", |
| 182 |
|
|
path.c_str(), sqlite3_errmsg(db_)); |
| 183 |
|
✗ |
sqlite3_reset(stmt_get_inode_); |
| 184 |
|
✗ |
return 0; |
| 185 |
|
|
} |
| 186 |
|
✗ |
inode = sqlite3_column_int64(stmt_get_inode_, 0); |
| 187 |
|
✗ |
sqlite3_reset(stmt_get_inode_); |
| 188 |
|
✗ |
return inode; |
| 189 |
|
|
} |
| 190 |
|
|
|
| 191 |
|
|
|
| 192 |
|
|
/** |
| 193 |
|
|
* Adds a new inode by path |
| 194 |
|
|
* \return New inode number, 0 on error |
| 195 |
|
|
*/ |
| 196 |
|
✗ |
uint64_t NfsMapsSqlite::IssueInode(const PathString &path) { |
| 197 |
|
|
int sqlite_state; |
| 198 |
|
|
uint64_t inode; |
| 199 |
|
✗ |
sqlite_state = sqlite3_prepare_v2(db_, kSqlAddInode, -1, &stmt_add_, NULL); |
| 200 |
|
✗ |
assert(sqlite_state == SQLITE_OK); |
| 201 |
|
✗ |
sqlite_state = sqlite3_bind_text(stmt_add_, 1, path.GetChars(), |
| 202 |
|
✗ |
path.GetLength(), SQLITE_TRANSIENT); |
| 203 |
|
✗ |
if (sqlite_state != SQLITE_OK) { |
| 204 |
|
✗ |
LogCvmfs(kLogNfsMaps, kLogDebug, "Failed to bind path in IssueInode (%s)", |
| 205 |
|
|
path.c_str()); |
| 206 |
|
✗ |
sqlite3_reset(stmt_add_); |
| 207 |
|
✗ |
return 0; |
| 208 |
|
|
} |
| 209 |
|
✗ |
sqlite_state = sqlite3_step(stmt_add_); |
| 210 |
|
✗ |
if (sqlite_state != SQLITE_DONE) { |
| 211 |
|
✗ |
LogCvmfs(kLogNfsMaps, kLogDebug, |
| 212 |
|
|
"Failed to execute SQL for IssueInode (%s): %s", path.c_str(), |
| 213 |
|
|
sqlite3_errmsg(db_)); |
| 214 |
|
✗ |
sqlite3_reset(stmt_add_); |
| 215 |
|
✗ |
return 0; |
| 216 |
|
|
} |
| 217 |
|
✗ |
inode = sqlite3_last_insert_rowid(db_); |
| 218 |
|
✗ |
sqlite3_reset(stmt_add_); |
| 219 |
|
✗ |
n_db_seq_->Set(inode); |
| 220 |
|
✗ |
perf::Inc(n_db_added_); |
| 221 |
|
✗ |
return inode; |
| 222 |
|
|
} |
| 223 |
|
|
|
| 224 |
|
|
|
| 225 |
|
✗ |
uint64_t NfsMapsSqlite::RetryGetInode(const PathString &path, int attempt) { |
| 226 |
|
✗ |
if (attempt > 2) { |
| 227 |
|
|
// We have to give up eventually |
| 228 |
|
✗ |
LogCvmfs(kLogNfsMaps, kLogSyslogErr, "Failed to find & create path (%s)", |
| 229 |
|
|
path.c_str()); |
| 230 |
|
✗ |
return 0; |
| 231 |
|
|
} |
| 232 |
|
|
|
| 233 |
|
|
uint64_t inode; |
| 234 |
|
|
{ |
| 235 |
|
✗ |
const MutexLockGuard m(lock_); |
| 236 |
|
✗ |
inode = FindInode(path); |
| 237 |
|
✗ |
if (inode) { |
| 238 |
|
✗ |
perf::Inc(n_db_path_found_); |
| 239 |
|
✗ |
return inode; |
| 240 |
|
|
} |
| 241 |
|
|
// Inode not found, issue a new one |
| 242 |
|
✗ |
inode = IssueInode(path); |
| 243 |
|
|
} |
| 244 |
|
|
|
| 245 |
|
✗ |
if (!inode) { |
| 246 |
|
✗ |
inode = RetryGetInode(path, attempt + 1); |
| 247 |
|
|
} |
| 248 |
|
✗ |
return inode; |
| 249 |
|
|
} |
| 250 |
|
|
|
| 251 |
|
|
|
| 252 |
|
|
/** |
| 253 |
|
|
* Finds the inode for path or issues a new inode. |
| 254 |
|
|
*/ |
| 255 |
|
✗ |
uint64_t NfsMapsSqlite::GetInode(const PathString &path) { |
| 256 |
|
✗ |
return RetryGetInode(path, 0); |
| 257 |
|
|
} |
| 258 |
|
|
|
| 259 |
|
|
|
| 260 |
|
|
/** |
| 261 |
|
|
* Finds the path that belongs to an inode. This must be successful. The |
| 262 |
|
|
* inode input comes from the file system, i.e. it must have been issued |
| 263 |
|
|
* before. |
| 264 |
|
|
* \return false if not found |
| 265 |
|
|
*/ |
| 266 |
|
✗ |
bool NfsMapsSqlite::GetPath(const uint64_t inode, PathString *path) { |
| 267 |
|
|
int sqlite_state; |
| 268 |
|
✗ |
const MutexLockGuard m(lock_); |
| 269 |
|
|
|
| 270 |
|
✗ |
sqlite_state = sqlite3_bind_int64(stmt_get_path_, 1, inode); |
| 271 |
|
✗ |
assert(sqlite_state == SQLITE_OK); |
| 272 |
|
✗ |
sqlite_state = sqlite3_step(stmt_get_path_); |
| 273 |
|
✗ |
if (sqlite_state == SQLITE_DONE) { |
| 274 |
|
|
// Success, but inode not found! |
| 275 |
|
✗ |
sqlite3_reset(stmt_get_path_); |
| 276 |
|
✗ |
return false; |
| 277 |
|
|
} |
| 278 |
|
✗ |
if (sqlite_state != SQLITE_ROW) { |
| 279 |
|
✗ |
PANIC(kLogSyslogErr, "Failed to execute SQL for GetPath (%" PRIu64 "): %s", |
| 280 |
|
|
inode, sqlite3_errmsg(db_)); |
| 281 |
|
|
} |
| 282 |
|
✗ |
const char *raw_path = (const char *)sqlite3_column_text(stmt_get_path_, 0); |
| 283 |
|
✗ |
path->Assign(raw_path, strlen(raw_path)); |
| 284 |
|
✗ |
sqlite3_reset(stmt_get_path_); |
| 285 |
|
✗ |
perf::Inc(n_db_inode_found_); |
| 286 |
|
✗ |
return true; |
| 287 |
|
|
} |
| 288 |
|
|
|
| 289 |
|
|
|
| 290 |
|
✗ |
NfsMapsSqlite::NfsMapsSqlite() |
| 291 |
|
✗ |
: db_(NULL) |
| 292 |
|
✗ |
, stmt_get_path_(NULL) |
| 293 |
|
✗ |
, stmt_get_inode_(NULL) |
| 294 |
|
✗ |
, stmt_add_(NULL) |
| 295 |
|
✗ |
, lock_(NULL) |
| 296 |
|
✗ |
, n_db_seq_(NULL) |
| 297 |
|
✗ |
, n_db_added_(NULL) |
| 298 |
|
✗ |
, n_db_path_found_(NULL) |
| 299 |
|
✗ |
, n_db_inode_found_(NULL) { |
| 300 |
|
✗ |
lock_ = reinterpret_cast<pthread_mutex_t *>(smalloc(sizeof(pthread_mutex_t))); |
| 301 |
|
✗ |
const int retval = pthread_mutex_init(lock_, NULL); |
| 302 |
|
✗ |
assert(retval == 0); |
| 303 |
|
|
} |
| 304 |
|
|
|
| 305 |
|
|
|
| 306 |
|
✗ |
NfsMapsSqlite::~NfsMapsSqlite() { |
| 307 |
|
✗ |
if (stmt_add_) |
| 308 |
|
✗ |
sqlite3_finalize(stmt_add_); |
| 309 |
|
✗ |
if (stmt_get_path_) |
| 310 |
|
✗ |
sqlite3_finalize(stmt_get_path_); |
| 311 |
|
✗ |
if (stmt_get_inode_) |
| 312 |
|
✗ |
sqlite3_finalize(stmt_get_inode_); |
| 313 |
|
|
// Close the handles, it is explicitly OK to call close with NULL |
| 314 |
|
✗ |
sqlite3_close_v2(db_); |
| 315 |
|
✗ |
pthread_mutex_destroy(lock_); |
| 316 |
|
✗ |
free(lock_); |
| 317 |
|
|
} |
| 318 |
|
|
|