source: libcfa/src/concurrency/io.cfa@ c7c178b

ADT arm-eh ast-experimental enum forall-pointer-decay jacob/cs343-translation new-ast-unique-expr pthread-emulation qualifiedEnum
Last change on this file since c7c178b was 11054eb, checked in by Thierry Delisle <tdelisle@…>, 5 years ago

Fix io to no longer use monitors since some usages aren't in threads
  • Property mode set to 100644
File size: 14.9 KB
Line 
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#define __cforall_thread__
17
18#if defined(__CFA_DEBUG__)
19 // #define __CFA_DEBUG_PRINT_IO__
20 // #define __CFA_DEBUG_PRINT_IO_CORE__
21#endif
22
23
24#if defined(CFA_HAVE_LINUX_IO_URING_H)
25 #define _GNU_SOURCE /* See feature_test_macros(7) */
26 #include <errno.h>
27 #include <signal.h>
28 #include <stdint.h>
29 #include <string.h>
30 #include <unistd.h>
31
32 extern "C" {
33 #include <sys/syscall.h>
34 #include <sys/eventfd.h>
35
36 #include <linux/io_uring.h>
37 }
38
39 #include "stats.hfa"
40 #include "kernel.hfa"
41 #include "kernel/fwd.hfa"
42 #include "io/types.hfa"
43
44 __attribute__((unused)) static const char * opcodes[] = {
45 "OP_NOP",
46 "OP_READV",
47 "OP_WRITEV",
48 "OP_FSYNC",
49 "OP_READ_FIXED",
50 "OP_WRITE_FIXED",
51 "OP_POLL_ADD",
52 "OP_POLL_REMOVE",
53 "OP_SYNC_FILE_RANGE",
54 "OP_SENDMSG",
55 "OP_RECVMSG",
56 "OP_TIMEOUT",
57 "OP_TIMEOUT_REMOVE",
58 "OP_ACCEPT",
59 "OP_ASYNC_CANCEL",
60 "OP_LINK_TIMEOUT",
61 "OP_CONNECT",
62 "OP_FALLOCATE",
63 "OP_OPENAT",
64 "OP_CLOSE",
65 "OP_FILES_UPDATE",
66 "OP_STATX",
67 "OP_READ",
68 "OP_WRITE",
69 "OP_FADVISE",
70 "OP_MADVISE",
71 "OP_SEND",
72 "OP_RECV",
73 "OP_OPENAT2",
74 "OP_EPOLL_CTL",
75 "OP_SPLICE",
76 "OP_PROVIDE_BUFFERS",
77 "OP_REMOVE_BUFFERS",
78 "OP_TEE",
79 "INVALID_OP"
80 };
81
82 static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want );
83 static void __ioarbiter_submit( $io_context * , __u32 idxs[], __u32 have, bool lazy );
84 static void __ioarbiter_flush ( $io_context & );
85 static inline void __ioarbiter_notify( $io_context & ctx );
86//=============================================================================================
87// I/O Polling
88//=============================================================================================
89 static inline unsigned __flush( struct $io_context & );
90 static inline __u32 __release_sqes( struct $io_context & );
91
92 void __cfa_io_drain( processor * proc ) {
93 /* paranoid */ verify( ! __preemption_enabled() );
94 /* paranoid */ verify( proc );
95 /* paranoid */ verify( proc->io.ctx );
96
97 // Drain the queue
98 $io_context * ctx = proc->io.ctx;
99 unsigned head = *ctx->cq.head;
100 unsigned tail = *ctx->cq.tail;
101 const __u32 mask = *ctx->cq.mask;
102
103 __u32 count = tail - head;
104 __STATS__( false, io.calls.drain++; io.calls.completed += count; )
105
106 for(i; count) {
107 unsigned idx = (head + i) & mask;
108 volatile struct io_uring_cqe & cqe = ctx->cq.cqes[idx];
109
110 /* paranoid */ verify(&cqe);
111
112 struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data;
113 __cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future );
114
115 fulfil( *future, cqe.res );
116 }
117
118 __cfadbg_print_safe(io, "Kernel I/O : %u completed\n", count);
119
120 // Mark to the kernel that the cqe has been seen
121 // Ensure that the kernel only sees the new value of the head index after the CQEs have been read.
122 __atomic_store_n( ctx->cq.head, head + count, __ATOMIC_SEQ_CST );
123
124 /* paranoid */ verify( ! __preemption_enabled() );
125
126 return;
127 }
128
129 void __cfa_io_flush( processor * proc ) {
130 /* paranoid */ verify( ! __preemption_enabled() );
131 /* paranoid */ verify( proc );
132 /* paranoid */ verify( proc->io.ctx );
133
134 $io_context & ctx = *proc->io.ctx;
135
136 __ioarbiter_flush( ctx );
137
138 __STATS__( true, io.calls.flush++; )
139 int ret = syscall( __NR_io_uring_enter, ctx.fd, ctx.sq.to_submit, 0, 0, (sigset_t *)0p, _NSIG / 8);
140 if( ret < 0 ) {
141 switch((int)errno) {
142 case EAGAIN:
143 case EINTR:
144 case EBUSY:
145 // Update statistics
146 __STATS__( false, io.calls.errors.busy ++; )
147 return;
148 default:
149 abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
150 }
151 }
152
153 __cfadbg_print_safe(io, "Kernel I/O : %u submitted to io_uring %d\n", ret, ctx.fd);
154 __STATS__( true, io.calls.submitted += ret; )
155 /* paranoid */ verify( ctx.sq.to_submit <= *ctx.sq.num );
156 /* paranoid */ verify( ctx.sq.to_submit >= ret );
157
158 ctx.sq.to_submit -= ret;
159
160 /* paranoid */ verify( ctx.sq.to_submit <= *ctx.sq.num );
161
162 // Release the consumed SQEs
163 __release_sqes( ctx );
164
165 /* paranoid */ verify( ! __preemption_enabled() );
166
167 ctx.proc->io.pending = false;
168 }
169
170//=============================================================================================
171// I/O Submissions
172//=============================================================================================
173
174// Submition steps :
175// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
176// listed in sq.array are visible by the kernel. For those not listed, the kernel does not
177// offer any assurance that an entry is not being filled by multiple flags. Therefore, we
178// need to write an allocator that allows allocating concurrently.
179//
180// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
181//
182// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
183// needs to arrive to two concensus at the same time:
184// A - The order in which entries are listed in the array: no two threads must pick the
185// same index for their entries
186// B - When can the tail be update for the kernel. EVERY entries in the array between
187// head and tail must be fully filled and shouldn't ever be touched again.
188//
189 //=============================================================================================
190 // Allocation
191 // for user's convenience fill the sqes from the indexes
192 static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct $io_context * ctx) {
193 struct io_uring_sqe * sqes = ctx->sq.sqes;
194 for(i; want) {
195 __cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
196 out_sqes[i] = &sqes[idxs[i]];
197 }
198 }
199
200 // Try to directly allocate from the a given context
201 // Not thread-safe
202 static inline bool __alloc(struct $io_context * ctx, __u32 idxs[], __u32 want) {
203 __sub_ring_t & sq = ctx->sq;
204 const __u32 mask = *sq.mask;
205 __u32 fhead = sq.free_ring.head; // get the current head of the queue
206 __u32 ftail = sq.free_ring.tail; // get the current tail of the queue
207
208 // If we don't have enough sqes, fail
209 if((ftail - fhead) < want) { return false; }
210
211 // copy all the indexes we want from the available list
212 for(i; want) {
213 __cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
214 idxs[i] = sq.free_ring.array[(fhead + i) & mask];
215 }
216
217 // Advance the head to mark the indexes as consumed
218 __atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
219
220 // return success
221 return true;
222 }
223
224 // Allocate an submit queue entry.
225 // The kernel cannot see these entries until they are submitted, but other threads must be
226 // able to see which entries can be used and which are already un used by an other thread
227 // for convenience, return both the index and the pointer to the sqe
228 // sqe == &sqes[idx]
229 struct $io_context * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) {
230 __cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want);
231
232 disable_interrupts();
233 processor * proc = __cfaabi_tls.this_processor;
234 $io_context * ctx = proc->io.ctx;
235 /* paranoid */ verify( __cfaabi_tls.this_processor );
236 /* paranoid */ verify( ctx );
237
238 __cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
239
240 // We can proceed to the fast path
241 if( __alloc(ctx, idxs, want) ) {
242 // Allocation was successful
243 __STATS__( true, io.alloc.fast += 1; )
244 enable_interrupts( __cfaabi_dbg_ctx );
245
246 __cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
247
248 __fill( sqes, want, idxs, ctx );
249 return ctx;
250 }
251 // The fast path failed, fallback
252 __STATS__( true, io.alloc.fail += 1; )
253
254 // Fast path failed, fallback on arbitration
255 __STATS__( true, io.alloc.slow += 1; )
256 enable_interrupts( __cfaabi_dbg_ctx );
257
258 $io_arbiter * ioarb = proc->cltr->io.arbiter;
259 /* paranoid */ verify( ioarb );
260
261 __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
262
263 struct $io_context * ret = __ioarbiter_allocate(*ioarb, idxs, want);
264
265 __cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
266
267 __fill( sqes, want, idxs,ret );
268 return ret;
269 }
270
271
272 //=============================================================================================
273 // submission
274 static inline void __submit( struct $io_context * ctx, __u32 idxs[], __u32 have, bool lazy) {
275 // We can proceed to the fast path
276 // Get the right objects
277 __sub_ring_t & sq = ctx->sq;
278 const __u32 mask = *sq.mask;
279 __u32 tail = *sq.kring.tail;
280
281 // Add the sqes to the array
282 for( i; have ) {
283 __cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
284 sq.kring.array[ (tail + i) & mask ] = idxs[i];
285 }
286
287 // Make the sqes visible to the submitter
288 __atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
289 sq.to_submit++;
290
291 ctx->proc->io.pending = true;
292 ctx->proc->io.dirty = true;
293 if(sq.to_submit > 30 || !lazy) {
294 __cfa_io_flush( ctx->proc );
295 }
296 }
297
298 void cfa_io_submit( struct $io_context * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) {
299 __cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager");
300
301 disable_interrupts();
302 processor * proc = __cfaabi_tls.this_processor;
303 $io_context * ctx = proc->io.ctx;
304 /* paranoid */ verify( __cfaabi_tls.this_processor );
305 /* paranoid */ verify( ctx );
306
307 // Can we proceed to the fast path
308 if( ctx == inctx ) // We have the right instance?
309 {
310 __submit(ctx, idxs, have, lazy);
311
312 // Mark the instance as no longer in-use, re-enable interrupts and return
313 __STATS__( true, io.submit.fast += 1; )
314 enable_interrupts( __cfaabi_dbg_ctx );
315
316 __cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
317 return;
318 }
319
320 // Fast path failed, fallback on arbitration
321 __STATS__( true, io.submit.slow += 1; )
322 enable_interrupts( __cfaabi_dbg_ctx );
323
324 __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
325
326 __ioarbiter_submit(inctx, idxs, have, lazy);
327 }
328
329 //=============================================================================================
330 // Flushing
331 // Go through the ring's submit queue and release everything that has already been consumed
332 // by io_uring
333 // This cannot be done by multiple threads
334 static __u32 __release_sqes( struct $io_context & ctx ) {
335 const __u32 mask = *ctx.sq.mask;
336
337 __attribute__((unused))
338 __u32 ctail = *ctx.sq.kring.tail; // get the current tail of the queue
339 __u32 chead = *ctx.sq.kring.head; // get the current head of the queue
340 __u32 phead = ctx.sq.kring.released; // get the head the last time we were here
341
342 __u32 ftail = ctx.sq.free_ring.tail; // get the current tail of the queue
343
344 // the 3 fields are organized like this diagram
345 // except it's are ring
346 // ---+--------+--------+----
347 // ---+--------+--------+----
348 // ^ ^ ^
349 // phead chead ctail
350
351 // make sure ctail doesn't wrap around and reach phead
352 /* paranoid */ verify(
353 (ctail >= chead && chead >= phead)
354 || (chead >= phead && phead >= ctail)
355 || (phead >= ctail && ctail >= chead)
356 );
357
358 // find the range we need to clear
359 __u32 count = chead - phead;
360
361 if(count == 0) {
362 return 0;
363 }
364
365 // We acquired an previous-head/current-head range
366 // go through the range and release the sqes
367 for( i; count ) {
368 __cfadbg_print_safe(io, "Kernel I/O : release loop\n");
369 __u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
370 ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
371 }
372
373 ctx.sq.kring.released = chead; // note up to were we processed
374 __atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST);
375
376 __ioarbiter_notify(ctx);
377
378 return count;
379 }
380
381//=============================================================================================
382// I/O Arbiter
383//=============================================================================================
384 static inline void block(__outstanding_io_queue & queue, __outstanding_io & item) {
385 // Lock the list, it's not thread safe
386 lock( queue.lock __cfaabi_dbg_ctx2 );
387 {
388 // Add our request to the list
389 add( queue.queue, item );
390
391 // Mark as pending
392 __atomic_store_n( &queue.empty, false, __ATOMIC_SEQ_CST );
393 }
394 unlock( queue.lock );
395
396 wait( item.sem );
397 }
398
399 static inline bool empty(__outstanding_io_queue & queue ) {
400 return __atomic_load_n( &queue.empty, __ATOMIC_SEQ_CST);
401 }
402
403 static $io_context * __ioarbiter_allocate( $io_arbiter & this, __u32 idxs[], __u32 want ) {
404 __cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
405
406 __STATS__( false, io.alloc.block += 1; )
407
408 // No one has any resources left, wait for something to finish
409 // We need to add ourself to a list of pending allocs and wait for an answer
410 __pending_alloc pa;
411 pa.idxs = idxs;
412 pa.want = want;
413
414 block(this.pending, (__outstanding_io&)pa);
415
416 return pa.ctx;
417
418 }
419
420 static void __ioarbiter_notify( $io_arbiter & this, $io_context * ctx ) {
421 /* paranoid */ verify( !empty(this.pending.queue) );
422
423 lock( this.pending.lock __cfaabi_dbg_ctx2 );
424 {
425 while( !empty(this.pending.queue) ) {
426 __cfadbg_print_safe(io, "Kernel I/O : notifying\n");
427 __u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
428 __pending_alloc & pa = (__pending_alloc&)head( this.pending.queue );
429
430 if( have > pa.want ) goto DONE;
431 drop( this.pending.queue );
432
433 /* paranoid */__attribute__((unused)) bool ret =
434
435 __alloc(ctx, pa.idxs, pa.want);
436
437 /* paranoid */ verify( ret );
438
439 pa.ctx = ctx;
440
441 post( pa.sem );
442 }
443
444 this.pending.empty = true;
445 DONE:;
446 }
447 unlock( this.pending.lock );
448 }
449
450 static void __ioarbiter_notify( $io_context & ctx ) {
451 if(!empty( ctx.arbiter->pending )) {
452 __ioarbiter_notify( *ctx.arbiter, &ctx );
453 }
454 }
455
456 // Simply append to the pending
457 static void __ioarbiter_submit( $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
458 __cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd);
459
460 __cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
461
462 __external_io ei;
463 ei.idxs = idxs;
464 ei.have = have;
465 ei.lazy = lazy;
466
467 block(ctx->ext_sq, (__outstanding_io&)ei);
468
469 __cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
470 }
471
472 static void __ioarbiter_flush( $io_context & ctx ) {
473 if(!empty( ctx.ext_sq )) {
474 __STATS__( false, io.flush.external += 1; )
475
476 __cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
477
478 lock( ctx.ext_sq.lock __cfaabi_dbg_ctx2 );
479 {
480 while( !empty(ctx.ext_sq.queue) ) {
481 __external_io & ei = (__external_io&)drop( ctx.ext_sq.queue );
482
483 __submit(&ctx, ei.idxs, ei.have, ei.lazy);
484
485 post( ei.sem );
486 }
487
488 ctx.ext_sq.empty = true;
489 }
490 unlock(ctx.ext_sq.lock );
491 }
492 }
493#endif
Note: See TracBrowser for help on using the repository browser.