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 consequtively 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 "logging.h" |
26 |
|
|
#include "prng.h" |
27 |
|
|
#include "smalloc.h" |
28 |
|
|
#include "statistics.h" |
29 |
|
|
#include "util/pointer.h" |
30 |
|
|
#include "util/posix.h" |
31 |
|
|
#include "util/string.h" |
32 |
|
|
|
33 |
|
|
using namespace std; // NOLINT |
34 |
|
|
|
35 |
|
|
|
36 |
|
|
const char *NfsMapsSqlite::kSqlCreateTable = |
37 |
|
|
"CREATE TABLE IF NOT EXISTS inodes (path TEXT PRIMARY KEY);"; |
38 |
|
|
const char *NfsMapsSqlite::kSqlAddRoot = |
39 |
|
|
"INSERT INTO inodes (rowid, path) VALUES (?, \"\");"; |
40 |
|
|
const char *NfsMapsSqlite::kSqlAddInode = |
41 |
|
|
"INSERT INTO inodes VALUES (?);"; |
42 |
|
|
const char *NfsMapsSqlite::kSqlGetInode = |
43 |
|
|
"SELECT rowid FROM inodes where path = ?;"; |
44 |
|
|
const char *NfsMapsSqlite::kSqlGetPath = |
45 |
|
|
"SELECT path FROM inodes where rowid = ?;"; |
46 |
|
|
|
47 |
|
|
|
48 |
|
|
int NfsMapsSqlite::BusyHandler(void *data, int attempt) { |
49 |
|
|
BusyHandlerInfo *handler_info = static_cast<BusyHandlerInfo *>(data); |
50 |
|
|
// Reset the accumulated time if this is the start of a new request |
51 |
|
|
if (attempt == 0) |
52 |
|
|
handler_info->accumulated_ms = 0; |
53 |
|
|
LogCvmfs(kLogNfsMaps, kLogDebug, |
54 |
|
|
"busy handler, attempt %d, accumulated waiting time %u", |
55 |
|
|
attempt, handler_info->accumulated_ms); |
56 |
|
|
if (handler_info->accumulated_ms >= handler_info->kMaxWaitMs) |
57 |
|
|
return 0; |
58 |
|
|
|
59 |
|
|
const unsigned backoff_range_ms = 1 << attempt; |
60 |
|
|
unsigned backoff_ms = handler_info->prng.Next(backoff_range_ms); |
61 |
|
|
if (handler_info->accumulated_ms + backoff_ms > handler_info->kMaxWaitMs) { |
62 |
|
|
backoff_ms = handler_info->kMaxWaitMs - handler_info->accumulated_ms; |
63 |
|
|
} |
64 |
|
|
if (backoff_ms > handler_info->kMaxBackoffMs) { |
65 |
|
|
backoff_ms = handler_info->kMaxBackoffMs; |
66 |
|
|
} |
67 |
|
|
|
68 |
|
|
SafeSleepMs(backoff_ms); |
69 |
|
|
handler_info->accumulated_ms += backoff_ms; |
70 |
|
|
return 1; |
71 |
|
|
} |
72 |
|
|
|
73 |
|
|
|
74 |
|
|
NfsMapsSqlite *NfsMapsSqlite::Create( |
75 |
|
|
const string &db_dir, |
76 |
|
|
const uint64_t root_inode, |
77 |
|
|
const bool rebuild, |
78 |
|
|
perf::Statistics *statistics) |
79 |
|
|
{ |
80 |
|
|
assert(root_inode > 0); |
81 |
|
|
UniquePtr<NfsMapsSqlite> maps(new NfsMapsSqlite()); |
82 |
|
|
maps->n_db_added_ = statistics->Register( |
83 |
|
|
"nfs.sqlite.n_added", "total number of issued inode"); |
84 |
|
|
maps->n_db_seq_ = statistics->Register( |
85 |
|
|
"nfs.sqlite.n_seq", "last inode issued"); |
86 |
|
|
maps->n_db_path_found_ = statistics->Register( |
87 |
|
|
"nfs.sqlite.n_path_hit", "inode --> path hits"); |
88 |
|
|
maps->n_db_inode_found_ = statistics->Register( |
89 |
|
|
"nfs.sqlite.n_inode_hit", "path --> inode hits"); |
90 |
|
|
|
91 |
|
|
string db_path = db_dir + "/inode_maps.db"; |
92 |
|
|
|
93 |
|
|
sqlite3_stmt *stmt; |
94 |
|
|
if (rebuild) { |
95 |
|
|
LogCvmfs(kLogNfsMaps, kLogDebug | kLogSyslogWarn, |
96 |
|
|
"Ignoring rebuild flag as this may crash other cluster nodes."); |
97 |
|
|
} |
98 |
|
|
// We don't want the shared cache, we want minimal caching so sync is kept |
99 |
|
|
int retval = sqlite3_enable_shared_cache(0); |
100 |
|
|
assert(retval == SQLITE_OK); |
101 |
|
|
|
102 |
|
|
retval = sqlite3_open_v2(db_path.c_str(), &maps->db_, |
103 |
|
|
SQLITE_OPEN_NOMUTEX | SQLITE_OPEN_READWRITE |
104 |
|
|
| SQLITE_OPEN_CREATE, NULL); |
105 |
|
|
if (retval != SQLITE_OK) { |
106 |
|
|
LogCvmfs(kLogNfsMaps, kLogDebug, |
107 |
|
|
"Failed to create inode_maps file (%s)", |
108 |
|
|
db_path.c_str()); |
109 |
|
|
return NULL; |
110 |
|
|
} |
111 |
|
|
// Be prepared to wait for up to 1 minute for transactions to complete |
112 |
|
|
// Being stuck for a long time is far more favorable than failing |
113 |
|
|
// TODO(jblomer): another busy handler. This one conflicts with SIGALRM |
114 |
|
|
retval = sqlite3_busy_handler( |
115 |
|
|
maps->db_, BusyHandler, &maps->busy_handler_info_); |
116 |
|
|
assert(retval == SQLITE_OK); |
117 |
|
|
|
118 |
|
|
// Set-up the main inode table if it doesn't exist |
119 |
|
|
retval = sqlite3_prepare_v2( |
120 |
|
|
maps->db_, kSqlCreateTable, kMaxDBSqlLen, &stmt, NULL); |
121 |
|
|
if (retval != SQLITE_OK) { |
122 |
|
|
LogCvmfs(kLogNfsMaps, kLogDebug | kLogSyslogErr, |
123 |
|
|
"Failed to prepare create table statement: %s", |
124 |
|
|
sqlite3_errmsg(maps->db_)); |
125 |
|
|
return NULL; |
126 |
|
|
} |
127 |
|
|
if (sqlite3_step(stmt) != SQLITE_DONE) { |
128 |
|
|
LogCvmfs(kLogNfsMaps, kLogSyslogErr, |
129 |
|
|
"Failed to create main inode table: %s", |
130 |
|
|
sqlite3_errmsg(maps->db_)); |
131 |
|
|
sqlite3_finalize(stmt); |
132 |
|
|
return NULL; |
133 |
|
|
} |
134 |
|
|
sqlite3_finalize(stmt); |
135 |
|
|
|
136 |
|
|
// Prepare lookup and add-inode statements |
137 |
|
|
retval = sqlite3_prepare_v2( |
138 |
|
|
maps->db_, kSqlGetPath, kMaxDBSqlLen, &maps->stmt_get_path_, NULL); |
139 |
|
|
assert(retval == SQLITE_OK); |
140 |
|
|
retval = sqlite3_prepare_v2(maps->db_, kSqlGetInode, kMaxDBSqlLen, |
141 |
|
|
&maps->stmt_get_inode_, NULL); |
142 |
|
|
assert(retval == SQLITE_OK); |
143 |
|
|
retval = sqlite3_prepare_v2(maps->db_, kSqlAddInode, kMaxDBSqlLen, |
144 |
|
|
&maps->stmt_add_, NULL); |
145 |
|
|
assert(retval == SQLITE_OK); |
146 |
|
|
|
147 |
|
|
// Check the root inode exists, if not create it |
148 |
|
|
PathString rootpath("", 0); |
149 |
|
|
if (!maps->FindInode(rootpath)) { |
150 |
|
|
retval = sqlite3_prepare_v2( |
151 |
|
|
maps->db_, kSqlAddRoot, kMaxDBSqlLen, &stmt, NULL); |
152 |
|
|
assert(retval == SQLITE_OK); |
153 |
|
|
sqlite3_bind_int64(stmt, 1, root_inode); |
154 |
|
|
assert(retval == SQLITE_OK); |
155 |
|
|
if (sqlite3_step(stmt) != SQLITE_DONE) { |
156 |
|
|
LogCvmfs(kLogNfsMaps, kLogDebug | kLogSyslogErr, |
157 |
|
|
"Failed to execute CreateRoot: %s", |
158 |
|
|
sqlite3_errmsg(maps->db_)); |
159 |
|
|
abort(); |
160 |
|
|
} |
161 |
|
|
sqlite3_finalize(stmt); |
162 |
|
|
} |
163 |
|
|
|
164 |
|
|
return maps.Release(); |
165 |
|
|
} |
166 |
|
|
|
167 |
|
|
|
168 |
|
|
/** |
169 |
|
|
* Finds an inode by path |
170 |
|
|
* \return inode number, 0 if path not found |
171 |
|
|
*/ |
172 |
|
|
uint64_t NfsMapsSqlite::FindInode(const PathString &path) { |
173 |
|
|
int sqlite_state; |
174 |
|
|
uint64_t inode; |
175 |
|
|
sqlite_state = sqlite3_bind_text(stmt_get_inode_, 1, path.GetChars(), |
176 |
|
|
path.GetLength(), SQLITE_TRANSIENT); |
177 |
|
|
assert(sqlite_state == SQLITE_OK); |
178 |
|
|
sqlite_state = sqlite3_step(stmt_get_inode_); |
179 |
|
|
if (sqlite_state == SQLITE_DONE) { |
180 |
|
|
// Path not found in DB |
181 |
|
|
sqlite3_reset(stmt_get_inode_); |
182 |
|
|
return 0; |
183 |
|
|
} |
184 |
|
|
if (sqlite_state != SQLITE_ROW) { |
185 |
|
|
LogCvmfs(kLogNfsMaps, kLogDebug, "Error finding inode (%s): %s", |
186 |
|
|
path.c_str(), sqlite3_errmsg(db_)); |
187 |
|
|
sqlite3_reset(stmt_get_inode_); |
188 |
|
|
return 0; |
189 |
|
|
} |
190 |
|
|
inode = sqlite3_column_int64(stmt_get_inode_, 0); |
191 |
|
|
sqlite3_reset(stmt_get_inode_); |
192 |
|
|
return inode; |
193 |
|
|
} |
194 |
|
|
|
195 |
|
|
|
196 |
|
|
/** |
197 |
|
|
* Adds a new inode by path |
198 |
|
|
* \return New inode number, 0 on error |
199 |
|
|
*/ |
200 |
|
|
uint64_t NfsMapsSqlite::IssueInode(const PathString &path) { |
201 |
|
|
int sqlite_state; |
202 |
|
|
uint64_t inode; |
203 |
|
|
sqlite_state = sqlite3_prepare_v2(db_, kSqlAddInode, kMaxDBSqlLen, |
204 |
|
|
&stmt_add_, NULL); |
205 |
|
|
assert(sqlite_state == SQLITE_OK); |
206 |
|
|
sqlite_state = sqlite3_bind_text(stmt_add_, 1, path.GetChars(), |
207 |
|
|
path.GetLength(), SQLITE_TRANSIENT); |
208 |
|
|
if (sqlite_state != SQLITE_OK) { |
209 |
|
|
LogCvmfs(kLogNfsMaps, kLogDebug, |
210 |
|
|
"Failed to bind path in IssueInode (%s)", path.c_str()); |
211 |
|
|
sqlite3_reset(stmt_add_); |
212 |
|
|
return 0; |
213 |
|
|
} |
214 |
|
|
sqlite_state = sqlite3_step(stmt_add_); |
215 |
|
|
if (sqlite_state != SQLITE_DONE) { |
216 |
|
|
LogCvmfs(kLogNfsMaps, kLogDebug, |
217 |
|
|
"Failed to execute SQL for IssueInode (%s): %s", |
218 |
|
|
path.c_str(), sqlite3_errmsg(db_)); |
219 |
|
|
sqlite3_reset(stmt_add_); |
220 |
|
|
return 0; |
221 |
|
|
} |
222 |
|
|
inode = sqlite3_last_insert_rowid(db_); |
223 |
|
|
sqlite3_reset(stmt_add_); |
224 |
|
|
n_db_seq_->Set(inode); |
225 |
|
|
perf::Inc(n_db_added_); |
226 |
|
|
return inode; |
227 |
|
|
} |
228 |
|
|
|
229 |
|
|
|
230 |
|
|
uint64_t NfsMapsSqlite::RetryGetInode(const PathString &path, int attempt) { |
231 |
|
|
if (attempt > 2) { |
232 |
|
|
// We have to give up eventually |
233 |
|
|
LogCvmfs(kLogNfsMaps, kLogSyslogErr, "Failed to find & create path (%s)", |
234 |
|
|
path.c_str()); |
235 |
|
|
return 0; |
236 |
|
|
} |
237 |
|
|
|
238 |
|
|
uint64_t inode; |
239 |
|
|
pthread_mutex_lock(lock_); |
240 |
|
|
inode = FindInode(path); |
241 |
|
|
if (inode) { |
242 |
|
|
perf::Inc(n_db_path_found_); |
243 |
|
|
pthread_mutex_unlock(lock_); |
244 |
|
|
return inode; |
245 |
|
|
} |
246 |
|
|
// Inode not found, issue a new one |
247 |
|
|
inode = IssueInode(path); |
248 |
|
|
pthread_mutex_unlock(lock_); |
249 |
|
|
if (!inode) { |
250 |
|
|
inode = RetryGetInode(path, attempt + 1); |
251 |
|
|
} |
252 |
|
|
return inode; |
253 |
|
|
} |
254 |
|
|
|
255 |
|
|
|
256 |
|
|
/** |
257 |
|
|
* Finds the inode for path or issues a new inode. |
258 |
|
|
*/ |
259 |
|
|
uint64_t NfsMapsSqlite::GetInode(const PathString &path) { |
260 |
|
|
return RetryGetInode(path, 0); |
261 |
|
|
} |
262 |
|
|
|
263 |
|
|
|
264 |
|
|
/** |
265 |
|
|
* Finds the path that belongs to an inode. This must be successful. The |
266 |
|
|
* inode input comes from the file system, i.e. it must have been issued |
267 |
|
|
* before. |
268 |
|
|
* \return false if not found |
269 |
|
|
*/ |
270 |
|
|
bool NfsMapsSqlite::GetPath(const uint64_t inode, PathString *path) { |
271 |
|
|
int sqlite_state; |
272 |
|
|
pthread_mutex_lock(lock_); |
273 |
|
|
sqlite_state = sqlite3_bind_int64(stmt_get_path_, 1, inode); |
274 |
|
|
assert(sqlite_state == SQLITE_OK); |
275 |
|
|
sqlite_state = sqlite3_step(stmt_get_path_); |
276 |
|
|
if (sqlite_state == SQLITE_DONE) { |
277 |
|
|
// Success, but inode not found! |
278 |
|
|
sqlite3_reset(stmt_get_path_); |
279 |
|
|
pthread_mutex_unlock(lock_); |
280 |
|
|
return false; |
281 |
|
|
} |
282 |
|
|
if (sqlite_state != SQLITE_ROW) { |
283 |
|
|
LogCvmfs(kLogNfsMaps, kLogSyslogErr, |
284 |
|
|
"Failed to execute SQL for GetPath (%" PRIu64 "): %s", |
285 |
|
|
inode, sqlite3_errmsg(db_)); |
286 |
|
|
pthread_mutex_unlock(lock_); |
287 |
|
|
abort(); |
288 |
|
|
} |
289 |
|
|
const char *raw_path = (const char *)sqlite3_column_text(stmt_get_path_, 0); |
290 |
|
|
path->Assign(raw_path, strlen(raw_path)); |
291 |
|
|
sqlite3_reset(stmt_get_path_); |
292 |
|
|
pthread_mutex_unlock(lock_); |
293 |
|
|
perf::Inc(n_db_inode_found_); |
294 |
|
|
return true; |
295 |
|
|
} |
296 |
|
|
|
297 |
|
|
|
298 |
|
|
NfsMapsSqlite::NfsMapsSqlite() |
299 |
|
|
: db_(NULL) |
300 |
|
|
, stmt_get_path_(NULL) |
301 |
|
|
, stmt_get_inode_(NULL) |
302 |
|
|
, stmt_add_(NULL) |
303 |
|
|
, lock_(NULL) |
304 |
|
|
, n_db_seq_(NULL) |
305 |
|
|
, n_db_added_(NULL) |
306 |
|
|
, n_db_path_found_(NULL) |
307 |
|
|
, n_db_inode_found_(NULL) |
308 |
|
|
{ |
309 |
|
|
lock_ = reinterpret_cast<pthread_mutex_t *>(smalloc(sizeof(pthread_mutex_t))); |
310 |
|
|
int retval = pthread_mutex_init(lock_, NULL); |
311 |
|
|
assert(retval == 0); |
312 |
|
|
} |
313 |
|
|
|
314 |
|
|
|
315 |
|
|
NfsMapsSqlite::~NfsMapsSqlite() { |
316 |
|
|
if (stmt_add_) sqlite3_finalize(stmt_add_); |
317 |
|
|
if (stmt_get_path_) sqlite3_finalize(stmt_get_path_); |
318 |
|
|
if (stmt_get_inode_) sqlite3_finalize(stmt_get_inode_); |
319 |
|
|
// Close the handles, it is explicitly OK to call close with NULL |
320 |
|
|
sqlite3_close_v2(db_); |
321 |
|
|
pthread_mutex_destroy(lock_); |
322 |
|
|
free(lock_); |
323 |
|
|
} |