Line data Source code
1 : #ifndef HEADER_fd_src_funk_fd_funk_h
2 : #define HEADER_fd_src_funk_fd_funk_h
3 :
4 : /* Funk is a hybrid of a database and version control system designed
5 : for ultra high performance blockchain applications.
6 :
7 : The data model is a flat table of records. A record is a xid/key-val
8 : pair and records are fast O(1) indexable by their xid/key. xid is
9 : short for "transaction id" and xids have a compile time fixed size
10 : (e.g. 16 bytes). keys also have a compile time fixed size (e.g.
11 : 40 bytes). Record values can vary in length from zero to a compile
12 : time maximum size. The xid of all zeros is reserved for the "root"
13 : transaction described below. Outside this, there are no
14 : restrictions on what a record xid, key or val can be. Individual
15 : records can be created, updated, and deleted arbitrarily. They are
16 : just binary data as far as funk is concerned.
17 :
18 : The maximum number of records is practically only limited by the size
19 : of the workspace memory backing it. At present, each record requires
20 : 128 bytes of metadata (this includes records that are published and
21 : records that are in the process of being updated). In other words,
22 : about 13 GiB record metadata per hundred million records. The
23 : maximum number of records that can be held by a funk instance is set
24 : when that it was created (given the persistent and relocatable
25 : properties described below though, it is straightforward to resize
26 : this).
27 :
28 : The transaction model is richer than what is found in a regular
29 : database. A transaction is a xid-"updates to parent transaction"
30 : pair and transactions are fast O(1) indexable by xid. There is no
31 : limitation on the number of updates in a transaction. Updates to the
32 : record value are represented as the complete value record to make it
33 : trivial to apply cryptographic operations like hashing to all updated
34 : values in a transaction with file I/O, operating system calls, memory
35 : data marshalling overhead, etc.
36 :
37 : Like records, the maximum number of transactions in preparation is
38 : practically only limited by the size of the workspace memory backing
39 : it. At present, a transaction requires 96 bytes of memory. As such,
40 : it is practical to track a large number of forks during an extended
41 : period of time of consensus failure in a block chain application
42 : without using much workspace memory at all. The maximum number of
43 : transactions that can be in preparation at any given time by a funk
44 : instance is set when that it was created (as before, given the
45 : persistent and relocatable properties described below, it is
46 : straightforward to resize this).
47 :
48 : That is, a transaction is a compact representation of the entire
49 : history of _all_ the database records up to that transaction. We can
50 : trace a transaction's ancestors back to the "root" give the complete
51 : history of all database records up to that transaction. The “root”
52 : transaction is the ancestor of all transactions. The transaction
53 : history is linear from the root transaction until the "last
54 : published" transaction and cannot be modified.
55 :
56 : To start "preparing" a new transaction, we pick the new transaction's
57 : xid (ideally unique among all transactions thus far) and fork off a
58 : "parent" transaction. This operation virtually clones all database
59 : records in the parent transaction, even if the parent itself has not
60 : yet been "published". Given the above, the parent transaction can be
61 : the last published transaction or another in-preparation transaction.
62 :
63 : Record inserts, reads, removes take place within the context
64 : of a transaction, effectively isolating them to a private view of the
65 : world. If a transaction is "cancelled", the changes to a record are
66 : harmlessly discarded. Records in a transaction that has children
67 : cannot be changed ("frozen").
68 :
69 : As such, it is not possible to modify the records in transactions
70 : strictly before the last published transaction. However, it is
71 : possible to modify the records of the last published transaction if
72 : there is no transactions in preparation. This is useful, for
73 : example, loading up a transaction from a checkpointed state on
74 : startup. A common idiom at start of a block though is to fork the
75 : potential transaction of that block from its parent (freezing its
76 : parent) and then fork a child of the potential transaction that will
77 : hold updates to the block that are incrementally "merged" into the
78 : potential transaction as block processing progresses.
79 :
80 : Critically, in-preparation transactions form a tree of dependent and
81 : competing histories. This model matches blockchains, where
82 : speculative work can proceed on several blocks at once long before
83 : the blocks are finalized. When a transaction is published, all its
84 : ancestors are also published, any competing histories are
85 : cancelled, leaving only a linear history up to the published
86 : transaction. There is no practical limitation on the complexity of
87 : this tree.
88 :
89 : Under the hood, the database state is stored in NUMA and TLB
90 : optimized shared memory (i.e. fd_wksp) such that various database
91 : operations can be used concurrently by multiple threads distributed
92 : arbitrarily over multiple processes zero copy.
93 :
94 : Database operations are at algorithmic minimums with reasonably high
95 : performance implementations. Most are fast O(1) time and all are
96 : small O(1) space (e.g. in complex transaction tree operations, there
97 : is no use of dynamic allocation to hold temporaries and no use of
98 : recursion to bound stack utilization at trivial levels). Further,
99 : there are no explicit operating system calls and, given a well
100 : optimized workspace (i.e. the wksp pages fit within a core's TLBs) no
101 : implicit operating system calls. Critical operations (e.g. those
102 : that actually might impact transaction history) are fortified against
103 : memory corruption (e.g. robust against DoS attack by corrupting
104 : transaction metadata to create loops in transaction trees or going
105 : out of bounds in memory). Outside of record values, all memory used
106 : is preallocated. And record values are O(1) lockfree concurrent
107 : allocated via fd_alloc using the same wksp as funk (the
108 : implementation is structured in layers that are straightforward to
109 : retarget for particular applications as might be necessary).
110 :
111 : The shared memory used by a funk instance is within a workspace such
112 : that it is also persistent and remotely inspectable. For example, a
113 : process attached to a funk instance can be terminated and a new
114 : process can resume exactly where the original process left off
115 : instantly (e.g. no file I/O). Or a real-time monitor could be
116 : visualizing the ongoing activity in a database non-invasively (e.g.
117 : forks in flight, records updated by forks, etc). Or an auxiliary
118 : process could be lazily and non-invasively writing all published
119 : records to permanent storage in the background in parallel with
120 : on-going operations.
121 :
122 : The records are further stored in the workspace memory relocatably.
123 : For example, workspace memory could just be committed to a persistent
124 : memory as is (or backed by NVMe or such directly), copied to a
125 : different host, and processes on the new host could resume (indeed,
126 : though it wouldn't be space efficient, the shared memory region is
127 : usable as is as an on-disk checkpoint file). Or the workspace could
128 : be resized and what not to handle large needs than when the database
129 : was initially created and it all "just works".
130 :
131 : Limited concurrent (multithreaded) access is supported. As a
132 : general rule, transaction level operations
133 : (e.g. fd_funk_txn_cancel and fd_funk_txn_publish) have to be
134 : single-threaded. In this case, no other access is allowed at the
135 : same time. Purely record level operations are thread safe and can
136 : be arbitrarily interleaved across multiple cpus. Specifically,
137 : these are:
138 : fd_funk_rec_query_try
139 : fd_funk_rec_query_test
140 : fd_funk_rec_query_try_global
141 : fd_funk_rec_prepare
142 : fd_funk_rec_publish
143 : fd_funk_rec_cancel
144 : fd_funk_rec_remove
145 : */
146 :
147 : //#include "fd_funk_base.h" /* Includes ../util/fd_util.h */
148 : //#include "fd_funk_txn.h" /* Includes fd_funk_base.h */
149 : //#include "fd_funk_rec.h" /* Includes fd_funk_txn.h */
150 : #include "fd_funk_val.h" /* Includes fd_funk_rec.h */
151 :
152 : /* FD_FUNK_ALIGN describe the alignment needed
153 : for a funk. ALIGN should be a positive integer power of 2.
154 : The footprint is dynamic depending on map sizes. */
155 :
156 147 : #define FD_FUNK_ALIGN (4096UL)
157 :
158 : /* The details of a fd_funk_shmem_private are exposed here to facilitate
159 : inlining various operations. */
160 :
161 42 : #define FD_FUNK_MAGIC (0xf17eda2ce7fc2c02UL) /* firedancer funk version 2 */
162 :
163 : struct __attribute__((aligned(FD_FUNK_ALIGN))) fd_funk_shmem_private {
164 :
165 : /* Metadata */
166 :
167 : ulong magic; /* ==FD_FUNK_MAGIC */
168 : ulong funk_gaddr; /* wksp gaddr of this in the backing wksp, non-zero gaddr */
169 : ulong wksp_tag; /* Tag to use for wksp allocations, positive */
170 : ulong seed; /* Seed for various hashing function used under the hood, arbitrary */
171 : ulong cycle_tag; /* Next cycle_tag to use, used internally for various data integrity checks */
172 :
173 : /* The funk transaction map stores the details about transactions
174 : in preparation and their relationships to each other. This is a
175 : fd_map_chain_para/fd_pool_para and more details are given in
176 : fd_funk_txn.h
177 :
178 : txn_max is the maximum number of transactions that can be in
179 : preparation. Due to the use of compressed map indices to reduce
180 : workspace memory footprint required, txn_max is at most
181 : FD_FUNK_TXN_IDX_NULL (currently ~4B). This should be more than
182 : ample for anticipated uses cases ... e.g. every single validator in
183 : a pool of tens of thousands Solana validator had its own fork and
184 : with no consensus ever being achieved, a funk with txn_max at the
185 : limits of a compressed index will be chug along for days to weeks
186 : before running out of indexing space. But if ever needing to
187 : support more, it is straightforward to change the code to not use
188 : index compression. Then, a funk (with a planet sized workspace
189 : backing it) would survive a similar scenario for millions of years.
190 : Presumably, if such a situation arose, in the weeks to eons while
191 : there was consensus, somebody would notice and care enough to
192 : intervene (if not it is probably irrelevant to the real world
193 : anyway).
194 :
195 : txn_map_gaddr is the wksp gaddr of the fd_funk_txn_map_t used by
196 : this funk.
197 :
198 : child_{head,tail}_cidx are compressed txn map indices. After
199 : decompression, they give the txn map index of the {oldest,youngest}
200 : child of funk (i.e. an in-preparation transaction whose parent
201 : transaction id is last_publish). FD_FUNK_TXN_IDX_NULL indicates
202 : the funk is childless. Thus, if head/tail is FD_FUNK_TXN_IDX_NULL,
203 : tail/head will be too. funk is "frozen" if it has children.
204 :
205 : last_publish is the ID of the last published transaction. It will
206 : be the root transaction if no transactions have been published.
207 : Will be the root transaction immediately after construction. */
208 :
209 : ulong txn_max; /* In [0,FD_FUNK_TXN_IDX_NULL] */
210 : ulong txn_map_gaddr; /* Non-zero wksp gaddr with tag wksp_tag
211 : seed ==fd_funk_txn_map_seed (txn_map)
212 : txn_max==fd_funk_txn_map_key_max(txn_map) */
213 : ulong txn_pool_gaddr;
214 : ulong txn_ele_gaddr;
215 :
216 : uint child_head_cidx; /* After decompression, in [0,txn_max) or FD_FUNK_TXN_IDX_NULL, FD_FUNK_TXN_IDX_NULL if txn_max 0 */
217 : uint child_tail_cidx; /* " */
218 :
219 : /* Padding to FD_FUNK_TXN_XID_ALIGN here */
220 :
221 : fd_funk_txn_xid_t root[1]; /* Always equal to the root transaction */
222 : fd_funk_txn_xid_t last_publish[1]; /* Root transaction immediately after construction, not root thereafter */
223 :
224 : /* The funk record map stores the details about all the records in
225 : the funk, including all those in the last published transaction and
226 : all those getting updated in an in-preparation translation. This
227 : is a fd_map_chain_para/fd_pool_para and more details are given in
228 : fd_funk_rec.h
229 :
230 : rec_max is the maximum number of records that can exist in this
231 : funk.
232 :
233 : rec_map_gaddr is the wksp gaddr of the fd_funk_rec_map_t used by
234 : this funk. */
235 :
236 : uint rec_max;
237 : ulong rec_map_gaddr; /* Non-zero wksp gaddr with tag wksp_tag
238 : seed ==fd_funk_rec_map_seed (rec_map)
239 : rec_max==fd_funk_rec_map_key_max(rec_map) */
240 : ulong rec_pool_gaddr;
241 : ulong rec_ele_gaddr;
242 : uint rec_head_idx; /* Record map index of the first record, FD_FUNK_REC_IDX_NULL if none (from oldest to youngest) */
243 : uint rec_tail_idx; /* " last " */
244 :
245 : /* The funk alloc is used for allocating wksp resources for record
246 : values. This is a fd_alloc and more details are given in
247 : fd_funk_val.h. Allocations from this allocator will be tagged with
248 : wksp_tag and operations on this allocator will use concurrency
249 : group 0.
250 :
251 : TODO: Consider letting user just pass a join of alloc (and maybe
252 : the cgroup_idx to give the funk), inferring the wksp, cgroup from
253 : that and allocating exclusively from that? */
254 :
255 : ulong alloc_gaddr; /* Non-zero wksp gaddr with tag wksp tag */
256 : uchar lock; /* lock for synchronizing modifications to funk object */
257 :
258 : /* Padding to FD_FUNK_ALIGN here */
259 : };
260 :
261 : /* The details of a fd_funk_private are exposed here to facilitate
262 : inlining various operations. */
263 :
264 : #define FD_FUNK_JOIN_ALIGN 64
265 :
266 : struct __attribute__((aligned(FD_FUNK_JOIN_ALIGN))) fd_funk_private {
267 :
268 : fd_funk_shmem_t * shmem;
269 :
270 : fd_funk_txn_map_t txn_map[1];
271 : fd_funk_txn_pool_t txn_pool[1];
272 :
273 : fd_funk_rec_map_t rec_map[1];
274 : fd_funk_rec_pool_t rec_pool[1];
275 :
276 : fd_wksp_t * wksp;
277 : fd_alloc_t * alloc;
278 :
279 : };
280 :
281 : FD_PROTOTYPES_BEGIN
282 :
283 : /* Constructors */
284 :
285 : /* fd_funk_align return FD_FUNK_ALIGN. */
286 :
287 : FD_FN_CONST ulong
288 : fd_funk_align( void );
289 :
290 : /* fd_funk_footprint returns the size need for funk and all
291 : auxiliary data structures. Note that only record valus are
292 : allocated dynamically. */
293 :
294 : FD_FN_CONST ulong
295 : fd_funk_footprint( ulong txn_max,
296 : ulong rec_max );
297 :
298 : /* fd_funk_new formats an unused wksp allocation with the appropriate
299 : alignment and footprint as a funk. Caller is not joined on return.
300 : Returns shmem on success and NULL on failure (shmem NULL, shmem
301 : misaligned, zero wksp_tag, shmem is not backed by a wksp ... logs
302 : details). A workspace can be used by multiple funks concurrently.
303 : They will dynamically share the underlying workspace (along with any
304 : other non-funk usage) but will otherwise act as completely separate
305 : non-conflicting funks. To help with various diagnostics, garbage
306 : collection and what not, all allocations to the underlying wksp are
307 : tagged with the given tag (positive). Ideally, the tag used here
308 : should be distinct from all other tags used by this workspace but
309 : this is not required. */
310 :
311 : void *
312 : fd_funk_new( void * shmem,
313 : ulong wksp_tag,
314 : ulong seed,
315 : ulong txn_max,
316 : ulong rec_max );
317 :
318 : /* fd_funk_join joins the caller to a funk instance. ljoin points to a
319 : fd_funk_t compatible memory region in the caller's address space,
320 : shfunk points to the first byte of the memory region backing the funk
321 : in the caller's address space. Returns an handle to the caller's
322 : local join on success (join has ownership of the ljoin region) and
323 : NULL on failure (NULL ljoin, NULL shfunk, misaligned shfunk, shfunk
324 : is not backed by a wksp, bad magic, ... logs details). Every
325 : successful join should have a matching leave. The lifetime of the
326 : join is until the matching leave or the thread group is terminated
327 : (joins are local to a thread group). */
328 :
329 : fd_funk_t *
330 : fd_funk_join( fd_funk_t * ljoin,
331 : void * shfunk );
332 :
333 : /* fd_funk_leave leaves a funk join. Returns the memory region used for
334 : join on success (caller has ownership on return and the caller is no
335 : longer joined) and NULL on failure (logs details). Sets *opt_shfunk
336 : a pointer to the funk shm region if opt_shfunk!=NULL. */
337 :
338 : void *
339 : fd_funk_leave( fd_funk_t * funk,
340 : void ** opt_shfunk );
341 :
342 : /* fd_funk_delete unformats a wksp allocation used as a funk
343 : (additionally frees all wksp allocations used by that funk). Assumes
344 : nobody is or will be joined to the funk. Returns shmem on success
345 : and NULL on failure (logs details). Reasons for failure include
346 : shfunk is NULL, misaligned shfunk, shfunk is not backed by a
347 : workspace, etc. */
348 :
349 : void *
350 : fd_funk_delete( void * shfunk );
351 :
352 : /* fd_funk_delete_fast is an optimized verison of fd_funk_delete.
353 : Unlike fd_funk_delete, makes an additional assumption that this funk
354 : was created with a wksp_tag (see fd_funk_new) that is distinct from
355 : all other tags in the workspace. Also unlike fd_funk_delete, frees
356 : wksp allocation backing the funk instance itself.
357 :
358 : WARNING: Using this function frees all wksp allocations matching the
359 : funk's wksp_tag. */
360 :
361 : void
362 : fd_funk_delete_fast( void * shfunk );
363 :
364 : /* Accessors */
365 :
366 : /* fd_funk_wksp returns the local join to the wksp backing the funk.
367 : The lifetime of the returned pointer is at least as long as the
368 : lifetime of the local join. Assumes funk is a current local join. */
369 :
370 23749272 : FD_FN_PURE static inline fd_wksp_t * fd_funk_wksp( fd_funk_t const * funk ) { return funk->wksp; }
371 :
372 : /* fd_funk_wksp_tag returns the workspace allocation tag used by the
373 : funk for its wksp allocations. Will be positive. Assumes funk is a
374 : current local join. */
375 :
376 15 : FD_FN_PURE static inline ulong fd_funk_wksp_tag( fd_funk_t * funk ) { return funk->shmem->wksp_tag; }
377 :
378 : /* fd_funk_seed returns the hash seed used by the funk for various hash
379 : functions. Arbitrary value. Assumes funk is a current local join.
380 : TODO: consider renaming hash_seed? */
381 :
382 3 : FD_FN_PURE static inline ulong fd_funk_seed( fd_funk_t * funk ) { return funk->shmem->seed; }
383 :
384 : /* fd_funk_txn_max returns maximum number of in-preparations the funk
385 : can support. Assumes funk is a current local join. Return in
386 : [0,FD_FUNK_TXN_IDX_NULL]. */
387 :
388 3 : FD_FN_PURE static inline ulong fd_funk_txn_max( fd_funk_t * funk ) { return funk->txn_pool->ele_max; }
389 :
390 : /* fd_funk_txn_map returns the funk's transaction map join. This
391 : join can copied by value and is generally stored as a stack variable. */
392 :
393 20582280 : FD_FN_PURE static inline fd_funk_txn_map_t * fd_funk_txn_map( fd_funk_t * funk ) { return funk->txn_map; }
394 :
395 : /* fd_funk_txn_pool returns the funk's transaction pool join. This
396 : join can copied by value and is generally stored as a stack variable. */
397 :
398 102009 : FD_FN_PURE static inline fd_funk_txn_pool_t * fd_funk_txn_pool( fd_funk_t * funk ) { return funk->txn_pool; }
399 :
400 : /* fd_funk_last_publish_child_{head,tail} returns a pointer in the
401 : caller's address space to {oldest,young} transaction child of root, NULL if
402 : funk is childless. All pointers are in the caller's address space.
403 : These are all a fast O(1) but not fortified against memory data
404 : corruption. */
405 :
406 : FD_FN_PURE static inline fd_funk_txn_t *
407 : fd_funk_last_publish_child_head( fd_funk_t * funk,
408 5538 : fd_funk_txn_pool_t * pool ) {
409 5538 : ulong idx = fd_funk_txn_idx( funk->shmem->child_head_cidx );
410 5538 : if( fd_funk_txn_idx_is_null( idx ) ) return NULL; /* TODO: Consider branchless? */
411 5538 : return pool->ele + idx;
412 5538 : }
413 :
414 : FD_FN_PURE static inline fd_funk_txn_t *
415 : fd_funk_last_publish_child_tail( fd_funk_t * funk,
416 5538 : fd_funk_txn_pool_t * pool ) {
417 5538 : ulong idx = fd_funk_txn_idx( funk->shmem->child_tail_cidx );
418 5538 : if( fd_funk_txn_idx_is_null( idx ) ) return NULL; /* TODO: Consider branchless? */
419 5538 : return pool->ele + idx;
420 5538 : }
421 :
422 : /* fd_funk_root returns a pointer in the caller's address space to the
423 : transaction id of the root transaction. Assumes funk is a current
424 : local join. Lifetime of the returned pointer is the lifetime of the
425 : current local join. The value at this pointer will always be the
426 : root transaction id. */
427 :
428 206560 : FD_FN_CONST static inline fd_funk_txn_xid_t const * fd_funk_root( fd_funk_t * funk ) { return funk->shmem->root; }
429 :
430 : /* fd_funk_last_publish returns a pointer in the caller's address space
431 : to transaction id of the last published transaction. Assumes funk is
432 : a current local join. Lifetime of the returned pointer is the
433 : lifetime of the current local join. The value at this pointer will
434 : be constant until the next transaction is published. */
435 :
436 3145734 : FD_FN_CONST static inline fd_funk_txn_xid_t const * fd_funk_last_publish( fd_funk_t * funk ) { return funk->shmem->last_publish; }
437 :
438 : /* fd_funk_is_frozen returns 1 if the records of the last published
439 : transaction are frozen (i.e. the funk has children) and 0 otherwise
440 : (i.e. the funk is childless). Assumes funk is a current local join. */
441 :
442 : FD_FN_PURE static inline int
443 189318884 : fd_funk_last_publish_is_frozen( fd_funk_t const * funk ) {
444 189318884 : return fd_funk_txn_idx( funk->shmem->child_head_cidx )!=FD_FUNK_TXN_IDX_NULL;
445 189318884 : }
446 :
447 : /* fd_funk_rec_max returns maximum number of records that can be held
448 : in the funk. This includes both records of the last published
449 : transaction and records for transactions that are in-flight. */
450 :
451 0 : FD_FN_PURE static inline ulong fd_funk_rec_max( fd_funk_t * funk ) { return funk->rec_pool->ele_max; }
452 :
453 : /* fd_funk_rec_map returns the funk's record map join. This
454 : join can copied by value and is generally stored as a stack variable. */
455 :
456 23740431 : FD_FN_PURE static inline fd_funk_rec_map_t * fd_funk_rec_map( fd_funk_t * funk ) { return funk->rec_map; }
457 :
458 : /* fd_funk_rec_pool returns the funk's record pool join. This
459 : join can copied by value and is generally stored as a stack variable. */
460 :
461 0 : FD_FN_PURE static inline fd_funk_rec_pool_t * fd_funk_rec_pool( fd_funk_t * funk ) { return funk->rec_pool; }
462 :
463 : /* fd_funk_alloc returns a pointer in the caller's address space to
464 : the funk's allocator. */
465 :
466 23444997 : FD_FN_PURE static inline fd_alloc_t * fd_funk_alloc( fd_funk_t * funk ) { return funk->alloc; }
467 :
468 : /* fd_funk_rec_is_full returns 1 if no more records can be allocated
469 : and 0 otherwise. */
470 :
471 : static inline int
472 24223038 : fd_funk_rec_is_full( fd_funk_t * funk ) {
473 24223038 : return fd_funk_rec_pool_is_empty( funk->rec_pool );
474 24223038 : }
475 :
476 : /* fd_funk_txn_is_full returns true if the transaction map is
477 : full. No more in-preparation transactions are allowed. */
478 :
479 : static inline int
480 836880 : fd_funk_txn_is_full( fd_funk_t * funk ) {
481 836880 : return fd_funk_txn_pool_is_empty( funk->txn_pool );
482 836880 : }
483 :
484 : /* Misc */
485 :
486 : /* fd_funk_verify verifies the integrity of funk. Returns
487 : FD_FUNK_SUCCESS if funk appears to be intact and FD_FUNK_ERR_INVAL
488 : otherwise (logs details). Assumes funk is a current local join (NULL
489 : returns FD_FUNK_ERR_INVAL and logs details.) */
490 :
491 : int
492 : fd_funk_verify( fd_funk_t * funk );
493 :
494 : FD_PROTOTYPES_END
495 :
496 : #endif /* HEADER_fd_src_funk_fd_funk_h */
|
