1 | // |
---|
2 | // Cforall Version 1.0.0 Copyright (C) 2020 University of Waterloo |
---|
3 | // |
---|
4 | // The contents of this file are covered under the licence agreement in the |
---|
5 | // file "LICENCE" distributed with Cforall. |
---|
6 | // |
---|
7 | // io.cfa -- |
---|
8 | // |
---|
9 | // Author : Thierry Delisle |
---|
10 | // Created On : Thu Apr 23 17:31:00 2020 |
---|
11 | // Last Modified By : |
---|
12 | // Last Modified On : |
---|
13 | // Update Count : |
---|
14 | // |
---|
15 | |
---|
16 | #if defined(__CFA_DEBUG__) |
---|
17 | // #define __CFA_DEBUG_PRINT_IO__ |
---|
18 | #define __CFA_DEBUG_PRINT_IO_CORE__ |
---|
19 | #endif |
---|
20 | |
---|
21 | #include "kernel_private.hfa" |
---|
22 | #include "bitmanip.hfa" |
---|
23 | |
---|
24 | #if !defined(CFA_HAVE_LINUX_IO_URING_H) |
---|
25 | void __kernel_io_startup() { |
---|
26 | // Nothing to do without io_uring |
---|
27 | } |
---|
28 | |
---|
29 | void __kernel_io_shutdown() { |
---|
30 | // Nothing to do without io_uring |
---|
31 | } |
---|
32 | |
---|
33 | void ?{}(io_context & this, struct cluster & cl) {} |
---|
34 | void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) {} |
---|
35 | |
---|
36 | void ^?{}(io_context & this) {} |
---|
37 | void ^?{}(io_context & this, bool cluster_context) {} |
---|
38 | |
---|
39 | #else |
---|
40 | #define _GNU_SOURCE /* See feature_test_macros(7) */ |
---|
41 | #include <errno.h> |
---|
42 | #include <stdint.h> |
---|
43 | #include <string.h> |
---|
44 | #include <unistd.h> |
---|
45 | |
---|
46 | extern "C" { |
---|
47 | #include <sys/epoll.h> |
---|
48 | #include <sys/mman.h> |
---|
49 | #include <sys/syscall.h> |
---|
50 | |
---|
51 | #include <linux/io_uring.h> |
---|
52 | } |
---|
53 | |
---|
54 | #include "bits/signal.hfa" |
---|
55 | #include "kernel_private.hfa" |
---|
56 | #include "thread.hfa" |
---|
57 | |
---|
58 | void ?{}(io_context_params & this) { |
---|
59 | this.num_entries = 256; |
---|
60 | this.num_ready = 256; |
---|
61 | this.submit_aff = -1; |
---|
62 | this.eager_submits = false; |
---|
63 | this.poller_submits = false; |
---|
64 | this.poll_submit = false; |
---|
65 | this.poll_complete = false; |
---|
66 | } |
---|
67 | |
---|
68 | static void * __io_poller_slow( void * arg ); |
---|
69 | |
---|
70 | // Weirdly, some systems that do support io_uring don't actually define these |
---|
71 | #ifdef __alpha__ |
---|
72 | /* |
---|
73 | * alpha is the only exception, all other architectures |
---|
74 | * have common numbers for new system calls. |
---|
75 | */ |
---|
76 | #ifndef __NR_io_uring_setup |
---|
77 | #define __NR_io_uring_setup 535 |
---|
78 | #endif |
---|
79 | #ifndef __NR_io_uring_enter |
---|
80 | #define __NR_io_uring_enter 536 |
---|
81 | #endif |
---|
82 | #ifndef __NR_io_uring_register |
---|
83 | #define __NR_io_uring_register 537 |
---|
84 | #endif |
---|
85 | #else /* !__alpha__ */ |
---|
86 | #ifndef __NR_io_uring_setup |
---|
87 | #define __NR_io_uring_setup 425 |
---|
88 | #endif |
---|
89 | #ifndef __NR_io_uring_enter |
---|
90 | #define __NR_io_uring_enter 426 |
---|
91 | #endif |
---|
92 | #ifndef __NR_io_uring_register |
---|
93 | #define __NR_io_uring_register 427 |
---|
94 | #endif |
---|
95 | #endif |
---|
96 | |
---|
97 | struct __submition_data { |
---|
98 | // Head and tail of the ring (associated with array) |
---|
99 | volatile uint32_t * head; |
---|
100 | volatile uint32_t * tail; |
---|
101 | volatile uint32_t prev_head; |
---|
102 | |
---|
103 | // The actual kernel ring which uses head/tail |
---|
104 | // indexes into the sqes arrays |
---|
105 | uint32_t * array; |
---|
106 | |
---|
107 | // number of entries and mask to go with it |
---|
108 | const uint32_t * num; |
---|
109 | const uint32_t * mask; |
---|
110 | |
---|
111 | // Submission flags (Not sure what for) |
---|
112 | uint32_t * flags; |
---|
113 | |
---|
114 | // number of sqes not submitted (whatever that means) |
---|
115 | uint32_t * dropped; |
---|
116 | |
---|
117 | // Like head/tail but not seen by the kernel |
---|
118 | volatile uint32_t * ready; |
---|
119 | uint32_t ready_cnt; |
---|
120 | |
---|
121 | __spinlock_t lock; |
---|
122 | __spinlock_t release_lock; |
---|
123 | |
---|
124 | // A buffer of sqes (not the actual ring) |
---|
125 | struct io_uring_sqe * sqes; |
---|
126 | |
---|
127 | // The location and size of the mmaped area |
---|
128 | void * ring_ptr; |
---|
129 | size_t ring_sz; |
---|
130 | }; |
---|
131 | |
---|
132 | struct __completion_data { |
---|
133 | // Head and tail of the ring |
---|
134 | volatile uint32_t * head; |
---|
135 | volatile uint32_t * tail; |
---|
136 | |
---|
137 | // number of entries and mask to go with it |
---|
138 | const uint32_t * mask; |
---|
139 | const uint32_t * num; |
---|
140 | |
---|
141 | // number of cqes not submitted (whatever that means) |
---|
142 | uint32_t * overflow; |
---|
143 | |
---|
144 | // the kernel ring |
---|
145 | struct io_uring_cqe * cqes; |
---|
146 | |
---|
147 | // The location and size of the mmaped area |
---|
148 | void * ring_ptr; |
---|
149 | size_t ring_sz; |
---|
150 | }; |
---|
151 | |
---|
152 | struct __io_data { |
---|
153 | struct __submition_data submit_q; |
---|
154 | struct __completion_data completion_q; |
---|
155 | uint32_t ring_flags; |
---|
156 | int fd; |
---|
157 | bool eager_submits:1; |
---|
158 | bool poller_submits:1; |
---|
159 | }; |
---|
160 | |
---|
161 | //============================================================================================= |
---|
162 | // I/O Startup / Shutdown logic + Master Poller |
---|
163 | //============================================================================================= |
---|
164 | |
---|
165 | // IO Master poller loop forward |
---|
166 | static void * iopoll_loop( __attribute__((unused)) void * args ); |
---|
167 | |
---|
168 | static struct { |
---|
169 | pthread_t thrd; // pthread handle to io poller thread |
---|
170 | void * stack; // pthread stack for io poller thread |
---|
171 | int epollfd; // file descriptor to the epoll instance |
---|
172 | volatile bool run; // Whether or not to continue |
---|
173 | } iopoll; |
---|
174 | |
---|
175 | void __kernel_io_startup() { |
---|
176 | __cfaabi_dbg_print_safe( "Kernel : Creating EPOLL instance\n" ); |
---|
177 | |
---|
178 | iopoll.epollfd = epoll_create1(0); |
---|
179 | if (iopoll.epollfd == -1) { |
---|
180 | abort( "internal error, epoll_create1\n"); |
---|
181 | } |
---|
182 | |
---|
183 | __cfaabi_dbg_print_safe( "Kernel : Starting io poller thread\n" ); |
---|
184 | |
---|
185 | iopoll.run = true; |
---|
186 | iopoll.stack = __create_pthread( &iopoll.thrd, iopoll_loop, 0p ); |
---|
187 | } |
---|
188 | |
---|
189 | void __kernel_io_shutdown() { |
---|
190 | // Notify the io poller thread of the shutdown |
---|
191 | iopoll.run = false; |
---|
192 | sigval val = { 1 }; |
---|
193 | pthread_sigqueue( iopoll.thrd, SIGUSR1, val ); |
---|
194 | |
---|
195 | // Wait for the io poller thread to finish |
---|
196 | |
---|
197 | pthread_join( iopoll.thrd, 0p ); |
---|
198 | free( iopoll.stack ); |
---|
199 | |
---|
200 | int ret = close(iopoll.epollfd); |
---|
201 | if (ret == -1) { |
---|
202 | abort( "internal error, close epoll\n"); |
---|
203 | } |
---|
204 | |
---|
205 | // Io polling is now fully stopped |
---|
206 | |
---|
207 | __cfaabi_dbg_print_safe( "Kernel : IO poller stopped\n" ); |
---|
208 | } |
---|
209 | |
---|
210 | static void * iopoll_loop( __attribute__((unused)) void * args ) { |
---|
211 | __processor_id_t id; |
---|
212 | id.id = doregister(&id); |
---|
213 | __cfaabi_dbg_print_safe( "Kernel : IO poller thread starting\n" ); |
---|
214 | |
---|
215 | // Block signals to control when they arrive |
---|
216 | sigset_t mask; |
---|
217 | sigfillset(&mask); |
---|
218 | if ( pthread_sigmask( SIG_BLOCK, &mask, 0p ) == -1 ) { |
---|
219 | abort( "internal error, pthread_sigmask" ); |
---|
220 | } |
---|
221 | |
---|
222 | sigdelset( &mask, SIGUSR1 ); |
---|
223 | |
---|
224 | // Create sufficient events |
---|
225 | struct epoll_event events[10]; |
---|
226 | // Main loop |
---|
227 | while( iopoll.run ) { |
---|
228 | // Wait for events |
---|
229 | int nfds = epoll_pwait( iopoll.epollfd, events, 10, -1, &mask ); |
---|
230 | |
---|
231 | // Check if an error occured |
---|
232 | if (nfds == -1) { |
---|
233 | if( errno == EINTR ) continue; |
---|
234 | abort( "internal error, pthread_sigmask" ); |
---|
235 | } |
---|
236 | |
---|
237 | for(i; nfds) { |
---|
238 | $io_ctx_thread * io_ctx = ($io_ctx_thread *)(uintptr_t)events[i].data.u64; |
---|
239 | /* paranoid */ verify( io_ctx ); |
---|
240 | __cfadbg_print_safe(io_core, "Kernel I/O : Unparking io poller %p\n", io_ctx); |
---|
241 | #if !defined( __CFA_NO_STATISTICS__ ) |
---|
242 | kernelTLS.this_stats = io_ctx->self.curr_cluster->stats; |
---|
243 | #endif |
---|
244 | __post( io_ctx->sem, &id ); |
---|
245 | } |
---|
246 | } |
---|
247 | |
---|
248 | __cfaabi_dbg_print_safe( "Kernel : IO poller thread stopping\n" ); |
---|
249 | unregister(&id); |
---|
250 | return 0p; |
---|
251 | } |
---|
252 | |
---|
253 | //============================================================================================= |
---|
254 | // I/O Context Constrution/Destruction |
---|
255 | //============================================================================================= |
---|
256 | |
---|
257 | void ?{}($io_ctx_thread & this, struct cluster & cl) { (this.self){ "IO Poller", cl }; } |
---|
258 | void main( $io_ctx_thread & this ); |
---|
259 | static inline $thread * get_thread( $io_ctx_thread & this ) { return &this.self; } |
---|
260 | void ^?{}( $io_ctx_thread & mutex this ) {} |
---|
261 | |
---|
262 | static void __io_create ( __io_data & this, const io_context_params & params_in ); |
---|
263 | static void __io_destroy( __io_data & this ); |
---|
264 | |
---|
265 | void ?{}(io_context & this, struct cluster & cl, const io_context_params & params) { |
---|
266 | (this.thrd){ cl }; |
---|
267 | this.thrd.ring = malloc(); |
---|
268 | __cfadbg_print_safe(io_core, "Kernel I/O : Creating ring for io_context %p\n", &this); |
---|
269 | __io_create( *this.thrd.ring, params ); |
---|
270 | |
---|
271 | __cfadbg_print_safe(io_core, "Kernel I/O : Starting poller thread for io_context %p\n", &this); |
---|
272 | this.thrd.done = false; |
---|
273 | __thrd_start( this.thrd, main ); |
---|
274 | |
---|
275 | __cfadbg_print_safe(io_core, "Kernel I/O : io_context %p ready\n", &this); |
---|
276 | } |
---|
277 | |
---|
278 | void ?{}(io_context & this, struct cluster & cl) { |
---|
279 | io_context_params params; |
---|
280 | (this){ cl, params }; |
---|
281 | } |
---|
282 | |
---|
283 | void ^?{}(io_context & this, bool cluster_context) { |
---|
284 | __cfadbg_print_safe(io_core, "Kernel I/O : tearing down io_context %p\n", &this); |
---|
285 | |
---|
286 | // Notify the thread of the shutdown |
---|
287 | __atomic_store_n(&this.thrd.done, true, __ATOMIC_SEQ_CST); |
---|
288 | |
---|
289 | // If this is an io_context within a cluster, things get trickier |
---|
290 | $thread & thrd = this.thrd.self; |
---|
291 | if( cluster_context ) { |
---|
292 | cluster & cltr = *thrd.curr_cluster; |
---|
293 | /* paranoid */ verify( cltr.nprocessors == 0 || &cltr == mainCluster ); |
---|
294 | /* paranoid */ verify( !ready_mutate_islocked() ); |
---|
295 | |
---|
296 | // We need to adjust the clean-up based on where the thread is |
---|
297 | if( thrd.state == Ready || thrd.preempted != __NO_PREEMPTION ) { |
---|
298 | |
---|
299 | ready_schedule_lock( (struct __processor_id_t *)active_processor() ); |
---|
300 | |
---|
301 | // This is the tricky case |
---|
302 | // The thread was preempted and now it is on the ready queue |
---|
303 | // The thread should be the last on the list |
---|
304 | /* paranoid */ verify( thrd.link.next != 0p ); |
---|
305 | |
---|
306 | // Remove the thread from the ready queue of this cluster |
---|
307 | __attribute__((unused)) bool removed = remove_head( &cltr, &thrd ); |
---|
308 | /* paranoid */ verify( removed ); |
---|
309 | thrd.link.next = 0p; |
---|
310 | thrd.link.prev = 0p; |
---|
311 | __cfaabi_dbg_debug_do( thrd.unpark_stale = true ); |
---|
312 | |
---|
313 | // Fixup the thread state |
---|
314 | thrd.state = Blocked; |
---|
315 | thrd.ticket = 0; |
---|
316 | thrd.preempted = __NO_PREEMPTION; |
---|
317 | |
---|
318 | ready_schedule_unlock( (struct __processor_id_t *)active_processor() ); |
---|
319 | |
---|
320 | // Pretend like the thread was blocked all along |
---|
321 | } |
---|
322 | // !!! This is not an else if !!! |
---|
323 | if( thrd.state == Blocked ) { |
---|
324 | |
---|
325 | // This is the "easy case" |
---|
326 | // The thread is parked and can easily be moved to active cluster |
---|
327 | verify( thrd.curr_cluster != active_cluster() || thrd.curr_cluster == mainCluster ); |
---|
328 | thrd.curr_cluster = active_cluster(); |
---|
329 | |
---|
330 | // unpark the fast io_poller |
---|
331 | unpark( &thrd __cfaabi_dbg_ctx2 ); |
---|
332 | } |
---|
333 | else { |
---|
334 | |
---|
335 | // The thread is in a weird state |
---|
336 | // I don't know what to do here |
---|
337 | abort("io_context poller thread is in unexpected state, cannot clean-up correctly\n"); |
---|
338 | } |
---|
339 | } else { |
---|
340 | unpark( &thrd __cfaabi_dbg_ctx2 ); |
---|
341 | } |
---|
342 | |
---|
343 | ^(this.thrd){}; |
---|
344 | __cfadbg_print_safe(io_core, "Kernel I/O : Stopped poller thread for io_context %p\n", &this); |
---|
345 | |
---|
346 | __io_destroy( *this.thrd.ring ); |
---|
347 | __cfadbg_print_safe(io_core, "Kernel I/O : Destroyed ring for io_context %p\n", &this); |
---|
348 | |
---|
349 | free(this.thrd.ring); |
---|
350 | } |
---|
351 | |
---|
352 | void ^?{}(io_context & this) { |
---|
353 | ^(this){ false }; |
---|
354 | } |
---|
355 | |
---|
356 | static void __io_create( __io_data & this, const io_context_params & params_in ) { |
---|
357 | // Step 1 : call to setup |
---|
358 | struct io_uring_params params; |
---|
359 | memset(¶ms, 0, sizeof(params)); |
---|
360 | if( params_in.poll_submit ) params.flags |= IORING_SETUP_SQPOLL; |
---|
361 | if( params_in.poll_complete ) params.flags |= IORING_SETUP_IOPOLL; |
---|
362 | |
---|
363 | uint32_t nentries = params_in.num_entries; |
---|
364 | |
---|
365 | int fd = syscall(__NR_io_uring_setup, nentries, ¶ms ); |
---|
366 | if(fd < 0) { |
---|
367 | abort("KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno)); |
---|
368 | } |
---|
369 | |
---|
370 | // Step 2 : mmap result |
---|
371 | memset( &this, 0, sizeof(struct __io_data) ); |
---|
372 | struct __submition_data & sq = this.submit_q; |
---|
373 | struct __completion_data & cq = this.completion_q; |
---|
374 | |
---|
375 | // calculate the right ring size |
---|
376 | sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) ); |
---|
377 | cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe)); |
---|
378 | |
---|
379 | // Requires features |
---|
380 | #if defined(IORING_FEAT_SINGLE_MMAP) |
---|
381 | // adjust the size according to the parameters |
---|
382 | if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) { |
---|
383 | cq.ring_sz = sq.ring_sz = max(cq.ring_sz, sq.ring_sz); |
---|
384 | } |
---|
385 | #endif |
---|
386 | |
---|
387 | // mmap the Submit Queue into existence |
---|
388 | sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQ_RING); |
---|
389 | if (sq.ring_ptr == (void*)MAP_FAILED) { |
---|
390 | abort("KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno)); |
---|
391 | } |
---|
392 | |
---|
393 | // Requires features |
---|
394 | #if defined(IORING_FEAT_SINGLE_MMAP) |
---|
395 | // mmap the Completion Queue into existence (may or may not be needed) |
---|
396 | if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) { |
---|
397 | cq.ring_ptr = sq.ring_ptr; |
---|
398 | } |
---|
399 | else |
---|
400 | #endif |
---|
401 | { |
---|
402 | // We need multiple call to MMAP |
---|
403 | cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_CQ_RING); |
---|
404 | if (cq.ring_ptr == (void*)MAP_FAILED) { |
---|
405 | munmap(sq.ring_ptr, sq.ring_sz); |
---|
406 | abort("KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno)); |
---|
407 | } |
---|
408 | } |
---|
409 | |
---|
410 | // mmap the submit queue entries |
---|
411 | size_t size = params.sq_entries * sizeof(struct io_uring_sqe); |
---|
412 | sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQES); |
---|
413 | if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) { |
---|
414 | munmap(sq.ring_ptr, sq.ring_sz); |
---|
415 | if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz); |
---|
416 | abort("KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno)); |
---|
417 | } |
---|
418 | |
---|
419 | // Get the pointers from the kernel to fill the structure |
---|
420 | // submit queue |
---|
421 | sq.head = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.head); |
---|
422 | sq.tail = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail); |
---|
423 | sq.mask = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask); |
---|
424 | sq.num = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries); |
---|
425 | sq.flags = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags); |
---|
426 | sq.dropped = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped); |
---|
427 | sq.array = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.array); |
---|
428 | sq.prev_head = *sq.head; |
---|
429 | |
---|
430 | { |
---|
431 | const uint32_t num = *sq.num; |
---|
432 | for( i; num ) { |
---|
433 | sq.sqes[i].user_data = 0ul64; |
---|
434 | } |
---|
435 | } |
---|
436 | |
---|
437 | (sq.lock){}; |
---|
438 | (sq.release_lock){}; |
---|
439 | |
---|
440 | if( params_in.poller_submits || params_in.eager_submits ) { |
---|
441 | /* paranoid */ verify( is_pow2( params_in.num_ready ) || (params_in.num_ready < 8) ); |
---|
442 | sq.ready_cnt = max( params_in.num_ready, 8 ); |
---|
443 | sq.ready = alloc_align( 64, sq.ready_cnt ); |
---|
444 | for(i; sq.ready_cnt) { |
---|
445 | sq.ready[i] = -1ul32; |
---|
446 | } |
---|
447 | } |
---|
448 | else { |
---|
449 | sq.ready_cnt = 0; |
---|
450 | sq.ready = 0p; |
---|
451 | } |
---|
452 | |
---|
453 | // completion queue |
---|
454 | cq.head = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.head); |
---|
455 | cq.tail = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail); |
---|
456 | cq.mask = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask); |
---|
457 | cq.num = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries); |
---|
458 | cq.overflow = ( uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow); |
---|
459 | cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes); |
---|
460 | |
---|
461 | // some paranoid checks |
---|
462 | /* paranoid */ verifyf( (*cq.mask) == ((*cq.num) - 1ul32), "IO_URING Expected mask to be %u (%u entries), was %u", (*cq.num) - 1ul32, *cq.num, *cq.mask ); |
---|
463 | /* paranoid */ verifyf( (*cq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *cq.num ); |
---|
464 | /* paranoid */ verifyf( (*cq.head) == 0, "IO_URING Expected head to be 0, got %u", *cq.head ); |
---|
465 | /* paranoid */ verifyf( (*cq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *cq.tail ); |
---|
466 | |
---|
467 | /* paranoid */ verifyf( (*sq.mask) == ((*sq.num) - 1ul32), "IO_URING Expected mask to be %u (%u entries), was %u", (*sq.num) - 1ul32, *sq.num, *sq.mask ); |
---|
468 | /* paranoid */ verifyf( (*sq.num) >= nentries, "IO_URING Expected %u entries, got %u", nentries, *sq.num ); |
---|
469 | /* paranoid */ verifyf( (*sq.head) == 0, "IO_URING Expected head to be 0, got %u", *sq.head ); |
---|
470 | /* paranoid */ verifyf( (*sq.tail) == 0, "IO_URING Expected tail to be 0, got %u", *sq.tail ); |
---|
471 | |
---|
472 | // Update the global ring info |
---|
473 | this.ring_flags = params.flags; |
---|
474 | this.fd = fd; |
---|
475 | this.eager_submits = params_in.eager_submits; |
---|
476 | this.poller_submits = params_in.poller_submits; |
---|
477 | } |
---|
478 | |
---|
479 | void __io_destroy( __io_data & this ) { |
---|
480 | // Shutdown the io rings |
---|
481 | struct __submition_data & sq = this.submit_q; |
---|
482 | struct __completion_data & cq = this.completion_q; |
---|
483 | |
---|
484 | // unmap the submit queue entries |
---|
485 | munmap(sq.sqes, (*sq.num) * sizeof(struct io_uring_sqe)); |
---|
486 | |
---|
487 | // unmap the Submit Queue ring |
---|
488 | munmap(sq.ring_ptr, sq.ring_sz); |
---|
489 | |
---|
490 | // unmap the Completion Queue ring, if it is different |
---|
491 | if (cq.ring_ptr != sq.ring_ptr) { |
---|
492 | munmap(cq.ring_ptr, cq.ring_sz); |
---|
493 | } |
---|
494 | |
---|
495 | // close the file descriptor |
---|
496 | close(this.fd); |
---|
497 | |
---|
498 | free( this.submit_q.ready ); // Maybe null, doesn't matter |
---|
499 | } |
---|
500 | |
---|
501 | int __io_uring_enter( struct __io_data & ring, unsigned to_submit, bool get ) { |
---|
502 | bool need_sys_to_submit = false; |
---|
503 | bool need_sys_to_complete = false; |
---|
504 | unsigned flags = 0; |
---|
505 | |
---|
506 | TO_SUBMIT: |
---|
507 | if( to_submit > 0 ) { |
---|
508 | if( !(ring.ring_flags & IORING_SETUP_SQPOLL) ) { |
---|
509 | need_sys_to_submit = true; |
---|
510 | break TO_SUBMIT; |
---|
511 | } |
---|
512 | if( (*ring.submit_q.flags) & IORING_SQ_NEED_WAKEUP ) { |
---|
513 | need_sys_to_submit = true; |
---|
514 | flags |= IORING_ENTER_SQ_WAKEUP; |
---|
515 | } |
---|
516 | } |
---|
517 | |
---|
518 | if( get && !(ring.ring_flags & IORING_SETUP_SQPOLL) ) { |
---|
519 | flags |= IORING_ENTER_GETEVENTS; |
---|
520 | if( (ring.ring_flags & IORING_SETUP_IOPOLL) ) { |
---|
521 | need_sys_to_complete = true; |
---|
522 | } |
---|
523 | } |
---|
524 | |
---|
525 | int ret = 0; |
---|
526 | if( need_sys_to_submit || need_sys_to_complete ) { |
---|
527 | ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, flags, 0p, _NSIG / 8); |
---|
528 | if( ret < 0 ) { |
---|
529 | switch((int)errno) { |
---|
530 | case EAGAIN: |
---|
531 | case EINTR: |
---|
532 | ret = -1; |
---|
533 | break; |
---|
534 | default: |
---|
535 | abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) ); |
---|
536 | } |
---|
537 | } |
---|
538 | } |
---|
539 | |
---|
540 | // Memory barrier |
---|
541 | __atomic_thread_fence( __ATOMIC_SEQ_CST ); |
---|
542 | return ret; |
---|
543 | } |
---|
544 | |
---|
545 | //============================================================================================= |
---|
546 | // I/O Polling |
---|
547 | //============================================================================================= |
---|
548 | static unsigned __collect_submitions( struct __io_data & ring ); |
---|
549 | static uint32_t __release_consumed_submission( struct __io_data & ring ); |
---|
550 | |
---|
551 | static inline void process(struct io_uring_cqe & cqe ) { |
---|
552 | struct __io_user_data_t * data = (struct __io_user_data_t *)(uintptr_t)cqe.user_data; |
---|
553 | __cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", data, cqe.res, data->thrd ); |
---|
554 | |
---|
555 | data->result = cqe.res; |
---|
556 | unpark( data->thrd __cfaabi_dbg_ctx2 ); |
---|
557 | } |
---|
558 | |
---|
559 | // Process a single completion message from the io_uring |
---|
560 | // This is NOT thread-safe |
---|
561 | static [int, bool] __drain_io( & struct __io_data ring ) { |
---|
562 | /* paranoid */ verify( !kernelTLS.preemption_state.enabled ); |
---|
563 | |
---|
564 | unsigned to_submit = 0; |
---|
565 | if( ring.poller_submits ) { |
---|
566 | // If the poller thread also submits, then we need to aggregate the submissions which are ready |
---|
567 | to_submit = __collect_submitions( ring ); |
---|
568 | } |
---|
569 | |
---|
570 | int ret = __io_uring_enter(ring, to_submit, true); |
---|
571 | if( ret < 0 ) { |
---|
572 | return [0, true]; |
---|
573 | } |
---|
574 | |
---|
575 | // update statistics |
---|
576 | if (to_submit > 0) { |
---|
577 | __STATS__( true, |
---|
578 | if( to_submit > 0 ) { |
---|
579 | io.submit_q.submit_avg.rdy += to_submit; |
---|
580 | io.submit_q.submit_avg.csm += ret; |
---|
581 | io.submit_q.submit_avg.cnt += 1; |
---|
582 | } |
---|
583 | ) |
---|
584 | } |
---|
585 | |
---|
586 | // Release the consumed SQEs |
---|
587 | __release_consumed_submission( ring ); |
---|
588 | |
---|
589 | // Drain the queue |
---|
590 | unsigned head = *ring.completion_q.head; |
---|
591 | unsigned tail = *ring.completion_q.tail; |
---|
592 | const uint32_t mask = *ring.completion_q.mask; |
---|
593 | |
---|
594 | // Nothing was new return 0 |
---|
595 | if (head == tail) { |
---|
596 | return [0, to_submit > 0]; |
---|
597 | } |
---|
598 | |
---|
599 | uint32_t count = tail - head; |
---|
600 | /* paranoid */ verify( count != 0 ); |
---|
601 | for(i; count) { |
---|
602 | unsigned idx = (head + i) & mask; |
---|
603 | struct io_uring_cqe & cqe = ring.completion_q.cqes[idx]; |
---|
604 | |
---|
605 | /* paranoid */ verify(&cqe); |
---|
606 | |
---|
607 | process( cqe ); |
---|
608 | } |
---|
609 | |
---|
610 | // Mark to the kernel that the cqe has been seen |
---|
611 | // Ensure that the kernel only sees the new value of the head index after the CQEs have been read. |
---|
612 | __atomic_thread_fence( __ATOMIC_SEQ_CST ); |
---|
613 | __atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_RELAXED ); |
---|
614 | |
---|
615 | return [count, count > 0 || to_submit > 0]; |
---|
616 | } |
---|
617 | |
---|
618 | void main( $io_ctx_thread & this ) { |
---|
619 | epoll_event ev; |
---|
620 | ev.events = EPOLLIN | EPOLLONESHOT; |
---|
621 | ev.data.u64 = (uint64_t)&this; |
---|
622 | int ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_ADD, this.ring->fd, &ev); |
---|
623 | if (ret < 0) { |
---|
624 | abort( "KERNEL ERROR: EPOLL ADD - (%d) %s\n", (int)errno, strerror(errno) ); |
---|
625 | } |
---|
626 | |
---|
627 | __cfadbg_print_safe(io_core, "Kernel I/O : IO poller %p for ring %p ready\n", &this, &this.ring); |
---|
628 | |
---|
629 | int reset = 0; |
---|
630 | // Then loop until we need to start |
---|
631 | while(!__atomic_load_n(&this.done, __ATOMIC_SEQ_CST)) { |
---|
632 | // Drain the io |
---|
633 | int count; |
---|
634 | bool again; |
---|
635 | disable_interrupts(); |
---|
636 | [count, again] = __drain_io( *this.ring ); |
---|
637 | |
---|
638 | if(!again) reset++; |
---|
639 | |
---|
640 | // Update statistics |
---|
641 | __STATS__( true, |
---|
642 | io.complete_q.completed_avg.val += count; |
---|
643 | io.complete_q.completed_avg.fast_cnt += 1; |
---|
644 | ) |
---|
645 | enable_interrupts( __cfaabi_dbg_ctx ); |
---|
646 | |
---|
647 | // If we got something, just yield and check again |
---|
648 | if(reset < 5) { |
---|
649 | yield(); |
---|
650 | } |
---|
651 | // We didn't get anything baton pass to the slow poller |
---|
652 | else { |
---|
653 | __cfadbg_print_safe(io_core, "Kernel I/O : Parking io poller %p\n", &this.self); |
---|
654 | reset = 0; |
---|
655 | |
---|
656 | // wake up the slow poller |
---|
657 | ret = epoll_ctl(iopoll.epollfd, EPOLL_CTL_MOD, this.ring->fd, &ev); |
---|
658 | if (ret < 0) { |
---|
659 | abort( "KERNEL ERROR: EPOLL REARM - (%d) %s\n", (int)errno, strerror(errno) ); |
---|
660 | } |
---|
661 | |
---|
662 | // park this thread |
---|
663 | wait( this.sem ); |
---|
664 | } |
---|
665 | } |
---|
666 | |
---|
667 | __cfadbg_print_safe(io_core, "Kernel I/O : Fast poller for ring %p stopping\n", &this.ring); |
---|
668 | } |
---|
669 | |
---|
670 | //============================================================================================= |
---|
671 | // I/O Submissions |
---|
672 | //============================================================================================= |
---|
673 | |
---|
674 | // Submition steps : |
---|
675 | // 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones |
---|
676 | // listed in sq.array are visible by the kernel. For those not listed, the kernel does not |
---|
677 | // offer any assurance that an entry is not being filled by multiple flags. Therefore, we |
---|
678 | // need to write an allocator that allows allocating concurrently. |
---|
679 | // |
---|
680 | // 2 - Actually fill the submit entry, this is the only simple and straightforward step. |
---|
681 | // |
---|
682 | // 3 - Append the entry index to the array and adjust the tail accordingly. This operation |
---|
683 | // needs to arrive to two concensus at the same time: |
---|
684 | // A - The order in which entries are listed in the array: no two threads must pick the |
---|
685 | // same index for their entries |
---|
686 | // B - When can the tail be update for the kernel. EVERY entries in the array between |
---|
687 | // head and tail must be fully filled and shouldn't ever be touched again. |
---|
688 | // |
---|
689 | |
---|
690 | [* struct io_uring_sqe, uint32_t] __submit_alloc( struct __io_data & ring, uint64_t data ) { |
---|
691 | /* paranoid */ verify( data != 0 ); |
---|
692 | |
---|
693 | // Prepare the data we need |
---|
694 | __attribute((unused)) int len = 0; |
---|
695 | __attribute((unused)) int block = 0; |
---|
696 | uint32_t cnt = *ring.submit_q.num; |
---|
697 | uint32_t mask = *ring.submit_q.mask; |
---|
698 | |
---|
699 | disable_interrupts(); |
---|
700 | uint32_t off = __tls_rand(); |
---|
701 | enable_interrupts( __cfaabi_dbg_ctx ); |
---|
702 | |
---|
703 | // Loop around looking for an available spot |
---|
704 | for() { |
---|
705 | // Look through the list starting at some offset |
---|
706 | for(i; cnt) { |
---|
707 | uint64_t expected = 0; |
---|
708 | uint32_t idx = (i + off) & mask; |
---|
709 | struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx]; |
---|
710 | volatile uint64_t * udata = &sqe->user_data; |
---|
711 | |
---|
712 | if( *udata == expected && |
---|
713 | __atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) |
---|
714 | { |
---|
715 | // update statistics |
---|
716 | __STATS__( false, |
---|
717 | io.submit_q.alloc_avg.val += len; |
---|
718 | io.submit_q.alloc_avg.block += block; |
---|
719 | io.submit_q.alloc_avg.cnt += 1; |
---|
720 | ) |
---|
721 | |
---|
722 | |
---|
723 | // Success return the data |
---|
724 | return [sqe, idx]; |
---|
725 | } |
---|
726 | verify(expected != data); |
---|
727 | |
---|
728 | len ++; |
---|
729 | } |
---|
730 | |
---|
731 | block++; |
---|
732 | yield(); |
---|
733 | } |
---|
734 | } |
---|
735 | |
---|
736 | static inline uint32_t __submit_to_ready_array( struct __io_data & ring, uint32_t idx, const uint32_t mask ) { |
---|
737 | /* paranoid */ verify( idx <= mask ); |
---|
738 | /* paranoid */ verify( idx != -1ul32 ); |
---|
739 | |
---|
740 | // We need to find a spot in the ready array |
---|
741 | __attribute((unused)) int len = 0; |
---|
742 | __attribute((unused)) int block = 0; |
---|
743 | uint32_t ready_mask = ring.submit_q.ready_cnt - 1; |
---|
744 | |
---|
745 | disable_interrupts(); |
---|
746 | uint32_t off = __tls_rand(); |
---|
747 | enable_interrupts( __cfaabi_dbg_ctx ); |
---|
748 | |
---|
749 | uint32_t picked; |
---|
750 | LOOKING: for() { |
---|
751 | for(i; ring.submit_q.ready_cnt) { |
---|
752 | picked = (i + off) & ready_mask; |
---|
753 | uint32_t expected = -1ul32; |
---|
754 | if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) { |
---|
755 | break LOOKING; |
---|
756 | } |
---|
757 | verify(expected != idx); |
---|
758 | |
---|
759 | len ++; |
---|
760 | } |
---|
761 | |
---|
762 | block++; |
---|
763 | if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) { |
---|
764 | __release_consumed_submission( ring ); |
---|
765 | unlock( ring.submit_q.lock ); |
---|
766 | } |
---|
767 | else { |
---|
768 | yield(); |
---|
769 | } |
---|
770 | } |
---|
771 | |
---|
772 | // update statistics |
---|
773 | __STATS__( false, |
---|
774 | io.submit_q.look_avg.val += len; |
---|
775 | io.submit_q.look_avg.block += block; |
---|
776 | io.submit_q.look_avg.cnt += 1; |
---|
777 | ) |
---|
778 | |
---|
779 | return picked; |
---|
780 | } |
---|
781 | |
---|
782 | void __submit( struct io_context * ctx, uint32_t idx ) __attribute__((nonnull (1))) { |
---|
783 | __io_data & ring = *ctx->thrd.ring; |
---|
784 | // Get now the data we definetely need |
---|
785 | uint32_t * const tail = ring.submit_q.tail; |
---|
786 | const uint32_t mask = *ring.submit_q.mask; |
---|
787 | |
---|
788 | // There are 2 submission schemes, check which one we are using |
---|
789 | if( ring.poller_submits ) { |
---|
790 | // If the poller thread submits, then we just need to add this to the ready array |
---|
791 | __submit_to_ready_array( ring, idx, mask ); |
---|
792 | |
---|
793 | post( ctx->thrd.sem ); |
---|
794 | |
---|
795 | __cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() ); |
---|
796 | } |
---|
797 | else if( ring.eager_submits ) { |
---|
798 | uint32_t picked = __submit_to_ready_array( ring, idx, mask ); |
---|
799 | |
---|
800 | for() { |
---|
801 | yield(); |
---|
802 | |
---|
803 | // If some one else collected our index, we are done |
---|
804 | #warning ABA problem |
---|
805 | if( ring.submit_q.ready[picked] != idx ) { |
---|
806 | __STATS__( false, |
---|
807 | io.submit_q.helped += 1; |
---|
808 | ) |
---|
809 | return; |
---|
810 | } |
---|
811 | |
---|
812 | if( try_lock(ring.submit_q.lock __cfaabi_dbg_ctx2) ) { |
---|
813 | __STATS__( false, |
---|
814 | io.submit_q.leader += 1; |
---|
815 | ) |
---|
816 | break; |
---|
817 | } |
---|
818 | |
---|
819 | __STATS__( false, |
---|
820 | io.submit_q.busy += 1; |
---|
821 | ) |
---|
822 | } |
---|
823 | |
---|
824 | // We got the lock |
---|
825 | unsigned to_submit = __collect_submitions( ring ); |
---|
826 | int ret = __io_uring_enter( ring, to_submit, false ); |
---|
827 | if( ret < 0 ) { |
---|
828 | unlock(ring.submit_q.lock); |
---|
829 | return; |
---|
830 | } |
---|
831 | |
---|
832 | /* paranoid */ verify( ret > 0 || (ring.ring_flags & IORING_SETUP_SQPOLL) ); |
---|
833 | |
---|
834 | // Release the consumed SQEs |
---|
835 | __release_consumed_submission( ring ); |
---|
836 | |
---|
837 | // update statistics |
---|
838 | __STATS__( true, |
---|
839 | io.submit_q.submit_avg.rdy += to_submit; |
---|
840 | io.submit_q.submit_avg.csm += ret; |
---|
841 | io.submit_q.submit_avg.cnt += 1; |
---|
842 | ) |
---|
843 | |
---|
844 | unlock(ring.submit_q.lock); |
---|
845 | } |
---|
846 | else { |
---|
847 | // get mutual exclusion |
---|
848 | lock(ring.submit_q.lock __cfaabi_dbg_ctx2); |
---|
849 | |
---|
850 | /* paranoid */ verifyf( ring.submit_q.sqes[ idx ].user_data != 0, |
---|
851 | /* paranoid */ "index %u already reclaimed\n" |
---|
852 | /* paranoid */ "head %u, prev %u, tail %u\n" |
---|
853 | /* paranoid */ "[-0: %u,-1: %u,-2: %u,-3: %u]\n", |
---|
854 | /* paranoid */ idx, |
---|
855 | /* paranoid */ *ring.submit_q.head, ring.submit_q.prev_head, *tail |
---|
856 | /* paranoid */ ,ring.submit_q.array[ ((*ring.submit_q.head) - 0) & (*ring.submit_q.mask) ] |
---|
857 | /* paranoid */ ,ring.submit_q.array[ ((*ring.submit_q.head) - 1) & (*ring.submit_q.mask) ] |
---|
858 | /* paranoid */ ,ring.submit_q.array[ ((*ring.submit_q.head) - 2) & (*ring.submit_q.mask) ] |
---|
859 | /* paranoid */ ,ring.submit_q.array[ ((*ring.submit_q.head) - 3) & (*ring.submit_q.mask) ] |
---|
860 | /* paranoid */ ); |
---|
861 | |
---|
862 | // Append to the list of ready entries |
---|
863 | |
---|
864 | /* paranoid */ verify( idx <= mask ); |
---|
865 | ring.submit_q.array[ (*tail) & mask ] = idx; |
---|
866 | __atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST); |
---|
867 | |
---|
868 | // Submit however, many entries need to be submitted |
---|
869 | int ret = __io_uring_enter( ring, 1, false ); |
---|
870 | if( ret < 0 ) { |
---|
871 | switch((int)errno) { |
---|
872 | default: |
---|
873 | abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) ); |
---|
874 | } |
---|
875 | } |
---|
876 | |
---|
877 | // update statistics |
---|
878 | __STATS__( false, |
---|
879 | io.submit_q.submit_avg.csm += 1; |
---|
880 | io.submit_q.submit_avg.cnt += 1; |
---|
881 | ) |
---|
882 | |
---|
883 | // Release the consumed SQEs |
---|
884 | __release_consumed_submission( ring ); |
---|
885 | |
---|
886 | unlock(ring.submit_q.lock); |
---|
887 | |
---|
888 | __cfadbg_print_safe( io, "Kernel I/O : Performed io_submit for %p, returned %d\n", active_thread(), ret ); |
---|
889 | } |
---|
890 | } |
---|
891 | |
---|
892 | static unsigned __collect_submitions( struct __io_data & ring ) { |
---|
893 | /* paranoid */ verify( ring.submit_q.ready != 0p ); |
---|
894 | /* paranoid */ verify( ring.submit_q.ready_cnt > 0 ); |
---|
895 | |
---|
896 | unsigned to_submit = 0; |
---|
897 | uint32_t tail = *ring.submit_q.tail; |
---|
898 | const uint32_t mask = *ring.submit_q.mask; |
---|
899 | |
---|
900 | // Go through the list of ready submissions |
---|
901 | for( i; ring.submit_q.ready_cnt ) { |
---|
902 | // replace any submission with the sentinel, to consume it. |
---|
903 | uint32_t idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED); |
---|
904 | |
---|
905 | // If it was already the sentinel, then we are done |
---|
906 | if( idx == -1ul32 ) continue; |
---|
907 | |
---|
908 | // If we got a real submission, append it to the list |
---|
909 | ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask; |
---|
910 | to_submit++; |
---|
911 | } |
---|
912 | |
---|
913 | // Increment the tail based on how many we are ready to submit |
---|
914 | __atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST); |
---|
915 | |
---|
916 | return to_submit; |
---|
917 | } |
---|
918 | |
---|
919 | static uint32_t __release_consumed_submission( struct __io_data & ring ) { |
---|
920 | const uint32_t smask = *ring.submit_q.mask; |
---|
921 | |
---|
922 | if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0; |
---|
923 | uint32_t chead = *ring.submit_q.head; |
---|
924 | uint32_t phead = ring.submit_q.prev_head; |
---|
925 | ring.submit_q.prev_head = chead; |
---|
926 | unlock(ring.submit_q.release_lock); |
---|
927 | |
---|
928 | uint32_t count = chead - phead; |
---|
929 | for( i; count ) { |
---|
930 | uint32_t idx = ring.submit_q.array[ (phead + i) & smask ]; |
---|
931 | ring.submit_q.sqes[ idx ].user_data = 0; |
---|
932 | } |
---|
933 | return count; |
---|
934 | } |
---|
935 | |
---|
936 | //============================================================================================= |
---|
937 | // I/O Submissions |
---|
938 | //============================================================================================= |
---|
939 | |
---|
940 | void register_fixed_files( io_context & ctx, int * files, unsigned count ) { |
---|
941 | int ret = syscall( __NR_io_uring_register, ctx.thrd.ring->fd, IORING_REGISTER_FILES, files, count ); |
---|
942 | if( ret < 0 ) { |
---|
943 | abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) ); |
---|
944 | } |
---|
945 | |
---|
946 | __cfadbg_print_safe( io_core, "Kernel I/O : Performed io_register for %p, returned %d\n", active_thread(), ret ); |
---|
947 | } |
---|
948 | |
---|
949 | void register_fixed_files( cluster & cltr, int * files, unsigned count ) { |
---|
950 | for(i; cltr.io.cnt) { |
---|
951 | register_fixed_files( cltr.io.ctxs[i], files, count ); |
---|
952 | } |
---|
953 | } |
---|
954 | #endif |
---|