| Directory: | cvmfs/ |
|---|---|
| File: | cvmfs/cache.h |
| Date: | 2025-10-26 02:35:25 |
| Exec | Total | Coverage | |
|---|---|---|---|
| Lines: | 28 | 37 | 75.7% |
| Branches: | 8 | 10 | 80.0% |
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /** | ||
| 2 | * This file is part of the CernVM File System. | ||
| 3 | */ | ||
| 4 | |||
| 5 | #ifndef CVMFS_CACHE_H_ | ||
| 6 | #define CVMFS_CACHE_H_ | ||
| 7 | |||
| 8 | #ifndef __STDC_FORMAT_MACROS | ||
| 9 | #define __STDC_FORMAT_MACROS | ||
| 10 | #endif | ||
| 11 | |||
| 12 | #include <stdint.h> | ||
| 13 | |||
| 14 | #include <string> | ||
| 15 | |||
| 16 | #include "compression/compression.h" | ||
| 17 | #include "crypto/hash.h" | ||
| 18 | #include "manifest.h" | ||
| 19 | #include "util/pointer.h" | ||
| 20 | |||
| 21 | |||
| 22 | class QuotaManager; | ||
| 23 | |||
| 24 | enum CacheManagerIds { | ||
| 25 | kUnknownCacheManager = 0, | ||
| 26 | kPosixCacheManager, | ||
| 27 | kRamCacheManager, | ||
| 28 | kTieredCacheManager, | ||
| 29 | kExternalCacheManager, | ||
| 30 | kStreamingCacheManager, | ||
| 31 | }; | ||
| 32 | |||
| 33 | enum CacheModes { | ||
| 34 | kCacheReadWrite = 0, | ||
| 35 | kCacheReadOnly, | ||
| 36 | }; | ||
| 37 | |||
| 38 | |||
| 39 | /** | ||
| 40 | * The Cache Manager provides (virtual) file descriptors to content-addressable | ||
| 41 | * objects in the cache. The implementation can use a POSIX file system or | ||
| 42 | * other means such as a key-value store. A file descriptor must remain | ||
| 43 | * readable until closed, no matter if it is removed from the backend storage | ||
| 44 | * or not (POSIX semantics). | ||
| 45 | * | ||
| 46 | * Writing into the cache is streamed and transactional: a file descriptor must | ||
| 47 | * be acquired from StartTxn and the object is only visible in the cache after | ||
| 48 | * CommitTxn. The state of the transaction is carried in an opque transaction | ||
| 49 | * object, which needs to be provided by the caller. The size of the object is | ||
| 50 | * returned by SizeOfTxn. This way, users of derived classes can take care of | ||
| 51 | * the storage allocation (e.g. on the stack), while the derived class | ||
| 52 | * determines the contents of the transaction object. For race-free read access | ||
| 53 | * to objects that are just being written to the cache, the OpenFromTxn routine | ||
| 54 | * is used just before the transaction is committed. | ||
| 55 | * | ||
| 56 | * Writing to the cache can be coupled to a quota manager. The quota manager | ||
| 57 | * maintains some extra information for data chunks: whether they are | ||
| 58 | * volatile, whether they are pinned, and a description (usually the path that | ||
| 59 | * corresponds to a data chunk). By default the NoopQuotaManager is used, which | ||
| 60 | * ignores all this extra information. The CtrlTxn() function is used to | ||
| 61 | * specify this extra information sometime between StartTxn() and CommitTxn(). | ||
| 62 | * | ||
| 63 | * The integer return values can be negative and, in this case, represent a | ||
| 64 | * -errno failure code. Otherwise the return value 0 indicates a success, or | ||
| 65 | * >= 0 for a file descriptor. | ||
| 66 | */ | ||
| 67 | class CacheManager : SingleCopy { | ||
| 68 | public: | ||
| 69 | /** | ||
| 70 | * Sizes of objects should be known for StartTxn(). For file catalogs we | ||
| 71 | * cannot ensure that, however, because the size field for nested catalogs | ||
| 72 | * was only recently added. | ||
| 73 | */ | ||
| 74 | static const uint64_t kSizeUnknown; | ||
| 75 | |||
| 76 | /** | ||
| 77 | * Relevant for the quota management and for downloading (URL construction). | ||
| 78 | * Used in Label::flags. | ||
| 79 | */ | ||
| 80 | static const int kLabelCatalog = 0x01; | ||
| 81 | static const int kLabelPinned = 0x02; | ||
| 82 | static const int kLabelVolatile = 0x04; | ||
| 83 | static const int kLabelExternal = 0x08; | ||
| 84 | static const int kLabelChunked = 0x10; | ||
| 85 | static const int kLabelCertificate = 0x20; | ||
| 86 | static const int kLabelMetainfo = 0x40; | ||
| 87 | static const int kLabelHistory = 0x80; | ||
| 88 | |||
| 89 | /** | ||
| 90 | * Meta-data of an object that the cache may or may not maintain/use. | ||
| 91 | * Good cache implementations should at least distinguish between volatile | ||
| 92 | * and regular objects. The data in the label are sufficient to download an | ||
| 93 | * object, if necessary. | ||
| 94 | */ | ||
| 95 | struct Label { | ||
| 96 | 15363 | Label() | |
| 97 | 15363 | : flags(0) | |
| 98 | 15363 | , size(kSizeUnknown) | |
| 99 | 15363 | , zip_algorithm(zlib::kZlibDefault) | |
| 100 | 15363 | , range_offset(-1) { } | |
| 101 | |||
| 102 | 6426 | bool IsCatalog() const { return flags & kLabelCatalog; } | |
| 103 | 3740 | bool IsPinned() const { return flags & kLabelPinned; } | |
| 104 | 952 | bool IsExternal() const { return flags & kLabelExternal; } | |
| 105 | 36 | bool IsCertificate() const { return flags & kLabelCertificate; } | |
| 106 | |||
| 107 | /** | ||
| 108 | * The description for the quota manager | ||
| 109 | */ | ||
| 110 | 7628 | std::string GetDescription() const { | |
| 111 |
2/2✓ Branch 0 taken 1013 times.
✓ Branch 1 taken 6615 times.
|
7628 | if (flags & kLabelCatalog) |
| 112 | 1013 | return "file catalog at " + path; | |
| 113 |
2/2✓ Branch 0 taken 179 times.
✓ Branch 1 taken 6436 times.
|
6615 | if (flags & kLabelCertificate) |
| 114 | 179 | return "certificate for " + path; | |
| 115 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6436 times.
|
6436 | if (flags & kLabelMetainfo) |
| 116 | ✗ | return "metainfo for " + path; | |
| 117 |
2/2✓ Branch 0 taken 18 times.
✓ Branch 1 taken 6418 times.
|
6436 | if (flags & kLabelHistory) |
| 118 | 18 | return "tag database for " + path; | |
| 119 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6418 times.
|
6418 | if (flags & kLabelChunked) |
| 120 | ✗ | return "Part of " + path; | |
| 121 | 6418 | return path; | |
| 122 | } | ||
| 123 | |||
| 124 | int flags; | ||
| 125 | uint64_t size; ///< unzipped size, if known | ||
| 126 | zlib::Algorithms zip_algorithm; | ||
| 127 | off_t range_offset; | ||
| 128 | /** | ||
| 129 | * The logical path on the mountpoint connected to the object. For meta- | ||
| 130 | * data objects, e.g. certificate, catalog, it's the repository name (which | ||
| 131 | * does not start with a slash). | ||
| 132 | */ | ||
| 133 | std::string path; | ||
| 134 | }; | ||
| 135 | |||
| 136 | /** | ||
| 137 | * A content hash together with the meta-data from a (partial) label. | ||
| 138 | */ | ||
| 139 | struct LabeledObject { | ||
| 140 | 4803 | explicit LabeledObject(const shash::Any &id) : id(id), label() { } | |
| 141 | 3892 | LabeledObject(const shash::Any &id, const Label &l) : id(id), label(l) { } | |
| 142 | |||
| 143 | shash::Any id; | ||
| 144 | Label label; | ||
| 145 | }; | ||
| 146 | |||
| 147 | virtual CacheManagerIds id() = 0; | ||
| 148 | /** | ||
| 149 | * Return a human readable description of the cache instance. Used in | ||
| 150 | * cvmfs_talk. | ||
| 151 | */ | ||
| 152 | virtual std::string Describe() = 0; | ||
| 153 | |||
| 154 | virtual bool AcquireQuotaManager(QuotaManager *quota_mgr) = 0; | ||
| 155 | |||
| 156 | virtual ~CacheManager(); | ||
| 157 | /** | ||
| 158 | * Opening an object might get it from a third-party source, e.g. when the | ||
| 159 | * tiered cache manager issues a copy-up operation. In this case it is | ||
| 160 | * beneficial to register the object with the accurate meta-data, in the same | ||
| 161 | * way it is done during transactions. | ||
| 162 | */ | ||
| 163 | virtual int Open(const LabeledObject &object) = 0; | ||
| 164 | virtual int64_t GetSize(int fd) = 0; | ||
| 165 | virtual int Close(int fd) = 0; | ||
| 166 | virtual int64_t Pread(int fd, void *buf, uint64_t size, uint64_t offset) = 0; | ||
| 167 | virtual int Dup(int fd) = 0; | ||
| 168 | virtual int Readahead(int fd) = 0; | ||
| 169 | |||
| 170 | virtual uint32_t SizeOfTxn() = 0; | ||
| 171 | virtual int StartTxn(const shash::Any &id, uint64_t size, void *txn) = 0; | ||
| 172 | virtual void CtrlTxn(const Label &label, | ||
| 173 | const int flags, // reserved for future use | ||
| 174 | void *txn) = 0; | ||
| 175 | virtual int64_t Write(const void *buf, uint64_t sz, void *txn) = 0; | ||
| 176 | virtual int Reset(void *txn) = 0; | ||
| 177 | virtual int AbortTxn(void *txn) = 0; | ||
| 178 | virtual int OpenFromTxn(void *txn) = 0; | ||
| 179 | virtual int CommitTxn(void *txn) = 0; | ||
| 180 | |||
| 181 | virtual void Spawn() = 0; | ||
| 182 | |||
| 183 | int ChecksumFd(int fd, shash::Any *id); | ||
| 184 | int OpenPinned(const LabeledObject &object); | ||
| 185 | bool Open2Mem(const LabeledObject &object, unsigned char **buffer, | ||
| 186 | uint64_t *size); | ||
| 187 | bool CommitFromMem(const LabeledObject &object, | ||
| 188 | const unsigned char *buffer, | ||
| 189 | const uint64_t size); | ||
| 190 | |||
| 191 | 903 | QuotaManager *quota_mgr() { return quota_mgr_; } | |
| 192 | |||
| 193 | // Rescue the open file table during reload of the fuse module. For the | ||
| 194 | // POSIX cache, nothing needs to be done because the table is keep in the | ||
| 195 | // kernel for the process. Other cache managers need to do it manually. | ||
| 196 | void *SaveState(const int fd_progress); | ||
| 197 | /** | ||
| 198 | * When RestoreState is called, the cache has already exactly one file | ||
| 199 | * descriptor open: the root file catalog. This file descriptor might be | ||
| 200 | * remapped to another number. A return value of -1 means no action needs | ||
| 201 | * to take place. A smaller value indicates an error. | ||
| 202 | */ | ||
| 203 | int RestoreState(const int fd_progress, void *state); | ||
| 204 | void FreeState(const int fd_progress, void *state); | ||
| 205 | ✗ | CacheManagerIds PeekState(void *state) { | |
| 206 | ✗ | return static_cast<State *>(state)->manager_type; | |
| 207 | } | ||
| 208 | |||
| 209 | /** | ||
| 210 | * While not strictly necessary, cache managers often have a directory | ||
| 211 | * associated with them. This directory is currently used to find the | ||
| 212 | * cached manifest copy, the cvmfschecksum.$reponame file. This is important | ||
| 213 | * to make pre-loaded alien caches work, even in a tiered setup. | ||
| 214 | */ | ||
| 215 | 350 | virtual manifest::Breadcrumb LoadBreadcrumb(const std::string & /*fqrn*/) { | |
| 216 | 350 | return manifest::Breadcrumb(); | |
| 217 | } | ||
| 218 | ✗ | virtual bool StoreBreadcrumb(const manifest::Manifest & /*manifest*/) { | |
| 219 | ✗ | return false; | |
| 220 | } | ||
| 221 | |||
| 222 | protected: | ||
| 223 | CacheManager(); | ||
| 224 | |||
| 225 | // Unless overwritten, Saving/Restoring states will crash the Fuse module | ||
| 226 | ✗ | virtual void *DoSaveState() { return NULL; } | |
| 227 | ✗ | virtual int DoRestoreState(void *data) { return false; } | |
| 228 | ✗ | virtual bool DoFreeState(void *data) { return false; } | |
| 229 | |||
| 230 | /** | ||
| 231 | * Never NULL but defaults to NoopQuotaManager. | ||
| 232 | */ | ||
| 233 | QuotaManager *quota_mgr_; | ||
| 234 | |||
| 235 | private: | ||
| 236 | static const unsigned kStateVersion = 0; | ||
| 237 | |||
| 238 | /** | ||
| 239 | * Wraps around the concrete cache manager's state block in memory. The | ||
| 240 | * state pointer is used in DoSaveState, DoRestoreState, DoFreeState. | ||
| 241 | */ | ||
| 242 | struct State : SingleCopy { | ||
| 243 | 66 | State() | |
| 244 | 132 | : version(kStateVersion) | |
| 245 | 66 | , manager_type(kUnknownCacheManager) | |
| 246 | 66 | , concrete_state(NULL) { } | |
| 247 | |||
| 248 | unsigned version; | ||
| 249 | CacheManagerIds manager_type; | ||
| 250 | void *concrete_state; | ||
| 251 | }; | ||
| 252 | }; // class CacheManager | ||
| 253 | |||
| 254 | #endif // CVMFS_CACHE_H_ | ||
| 255 |