source: libcfa/src/concurrency/io.cfa @ 9e27f69

ADTarm-ehast-experimentalenumforall-pointer-decayjacob/cs343-translationnew-ast-unique-exprpthread-emulationqualifiedEnum
Last change on this file since 9e27f69 was dddb3dd0, checked in by Thierry Delisle <tdelisle@…>, 3 years ago

Changed io to use ring per kernel threads.

  • Property mode set to 100644
File size: 14.6 KB
RevLine 
[ecf6b46]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
[3e2b9c9]16#define __cforall_thread__
17
[20ab637]18#if defined(__CFA_DEBUG__)
[d60d30e]19        // #define __CFA_DEBUG_PRINT_IO__
20        // #define __CFA_DEBUG_PRINT_IO_CORE__
[20ab637]21#endif
[4069faad]22
[f6660520]23
[3e2b9c9]24#if defined(CFA_HAVE_LINUX_IO_URING_H)
[31bb2e1]25        #define _GNU_SOURCE         /* See feature_test_macros(7) */
26        #include <errno.h>
[3e2b9c9]27        #include <signal.h>
[31bb2e1]28        #include <stdint.h>
29        #include <string.h>
30        #include <unistd.h>
31
[92976d9]32        extern "C" {
33                #include <sys/syscall.h>
[dddb3dd0]34                #include <sys/eventfd.h>
[92976d9]35
36                #include <linux/io_uring.h>
37        }
38
[3e2b9c9]39        #include "stats.hfa"
40        #include "kernel.hfa"
41        #include "kernel/fwd.hfa"
42        #include "io/types.hfa"
[185efe6]43
[2fab24e3]44        __attribute__((unused)) static const char * opcodes[] = {
[426f60c]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
[dddb3dd0]82        static $io_context * __ioarbiter_allocate( $io_arbiter & mutex this, processor *, __u32 idxs[], __u32 want );
83        static void __ioarbiter_submit( $io_arbiter & mutex this, $io_context * , __u32 idxs[], __u32 have, bool lazy );
84        static void __ioarbiter_flush ( $io_arbiter & mutex this, $io_context * );
85        static inline void __ioarbiter_notify( $io_context & ctx );
[92976d9]86//=============================================================================================
87// I/O Polling
88//=============================================================================================
[78da4ab]89        static inline unsigned __flush( struct $io_context & );
90        static inline __u32 __release_sqes( struct $io_context & );
[1d5e4711]91
[dddb3dd0]92        void __cfa_io_drain( processor * proc ) {
93                /* paranoid */ verify( ! __preemption_enabled() );
94                /* paranoid */ verify( proc );
95                /* paranoid */ verify( proc->io.ctx );
[6f121b8]96
[d384787]97                // Drain the queue
[dddb3dd0]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;
[92976d9]102
[4998155]103                __u32 count = tail - head;
[dddb3dd0]104                __STATS__( false, io.calls.drain++; io.calls.completed += count; )
[d60d30e]105
[d384787]106                for(i; count) {
[6f121b8]107                        unsigned idx = (head + i) & mask;
[dddb3dd0]108                        volatile struct io_uring_cqe & cqe = ctx->cq.cqes[idx];
[92976d9]109
[d384787]110                        /* paranoid */ verify(&cqe);
[92976d9]111
[78da4ab]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
[dddb3dd0]118                __cfadbg_print_safe(io, "Kernel I/O : %u completed\n", count);
[2d8f7b0]119
[92976d9]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.
[dddb3dd0]122                __atomic_store_n( ctx->cq.head, head + count, __ATOMIC_SEQ_CST );
[92976d9]123
[dddb3dd0]124                /* paranoid */ verify( ! __preemption_enabled() );
125
126                return;
[92976d9]127        }
128
[dddb3dd0]129        void __cfa_io_flush( processor * proc ) {
130                /* paranoid */ verify( ! __preemption_enabled() );
131                /* paranoid */ verify( proc );
132                /* paranoid */ verify( proc->io.ctx );
[1539bbd]133
[dddb3dd0]134                $io_context & ctx = *proc->io.ctx;
[78da4ab]135
[dddb3dd0]136                if(!ctx.ext_sq.empty) {
137                        __ioarbiter_flush( *ctx.arbiter, &ctx );
138                }
[3c039b0]139
[dddb3dd0]140                __STATS__( true, io.calls.flush++; )
141                int ret = syscall( __NR_io_uring_enter, ctx.fd, ctx.sq.to_submit, 0, 0, (sigset_t *)0p, _NSIG / 8);
142                if( ret < 0 ) {
143                        switch((int)errno) {
144                        case EAGAIN:
145                        case EINTR:
146                        case EBUSY:
147                                // Update statistics
148                                __STATS__( false, io.calls.errors.busy ++; )
149                                return;
150                        default:
151                                abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) );
[61dd73d]152                        }
[dddb3dd0]153                }
[ece0e80]154
[dddb3dd0]155                __cfadbg_print_safe(io, "Kernel I/O : %u submitted to io_uring %d\n", ret, ctx.fd);
156                __STATS__( true, io.calls.submitted += ret; )
157                /* paranoid */ verify( ctx.sq.to_submit <= *ctx.sq.num );
158                /* paranoid */ verify( ctx.sq.to_submit >= ret );
159
160                ctx.sq.to_submit -= ret;
[ece0e80]161
[dddb3dd0]162                /* paranoid */ verify( ctx.sq.to_submit <= *ctx.sq.num );
[5dadc9b]163
[dddb3dd0]164                // Release the consumed SQEs
165                __release_sqes( ctx );
[ece0e80]166
[dddb3dd0]167                /* paranoid */ verify( ! __preemption_enabled() );
[61dd73d]168
[dddb3dd0]169                ctx.proc->io.pending = false;
[61dd73d]170        }
[f6660520]171
[92976d9]172//=============================================================================================
173// I/O Submissions
174//=============================================================================================
175
[2d8f7b0]176// Submition steps :
[e46c753]177// 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones
[2d8f7b0]178//     listed in sq.array are visible by the kernel. For those not listed, the kernel does not
179//     offer any assurance that an entry is not being filled by multiple flags. Therefore, we
180//     need to write an allocator that allows allocating concurrently.
181//
[e46c753]182// 2 - Actually fill the submit entry, this is the only simple and straightforward step.
[2d8f7b0]183//
[e46c753]184// 3 - Append the entry index to the array and adjust the tail accordingly. This operation
[2d8f7b0]185//     needs to arrive to two concensus at the same time:
186//     A - The order in which entries are listed in the array: no two threads must pick the
187//         same index for their entries
188//     B - When can the tail be update for the kernel. EVERY entries in the array between
189//         head and tail must be fully filled and shouldn't ever be touched again.
190//
[78da4ab]191        //=============================================================================================
192        // Allocation
193        // for user's convenience fill the sqes from the indexes
194        static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct $io_context * ctx)  {
195                struct io_uring_sqe * sqes = ctx->sq.sqes;
196                for(i; want) {
[dddb3dd0]197                        __cfadbg_print_safe(io, "Kernel I/O : filling loop\n");
[78da4ab]198                        out_sqes[i] = &sqes[idxs[i]];
199                }
200        }
[2489d31]201
[78da4ab]202        // Try to directly allocate from the a given context
203        // Not thread-safe
204        static inline bool __alloc(struct $io_context * ctx, __u32 idxs[], __u32 want) {
205                __sub_ring_t & sq = ctx->sq;
206                const __u32 mask  = *sq.mask;
207                __u32 fhead = sq.free_ring.head;    // get the current head of the queue
208                __u32 ftail = sq.free_ring.tail;    // get the current tail of the queue
[2489d31]209
[78da4ab]210                // If we don't have enough sqes, fail
211                if((ftail - fhead) < want) { return false; }
[426f60c]212
[78da4ab]213                // copy all the indexes we want from the available list
214                for(i; want) {
[dddb3dd0]215                        __cfadbg_print_safe(io, "Kernel I/O : allocating loop\n");
[78da4ab]216                        idxs[i] = sq.free_ring.array[(fhead + i) & mask];
[6f121b8]217                }
[2489d31]218
[78da4ab]219                // Advance the head to mark the indexes as consumed
220                __atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE);
[df40a56]221
[78da4ab]222                // return success
223                return true;
224        }
[df40a56]225
[78da4ab]226        // Allocate an submit queue entry.
227        // The kernel cannot see these entries until they are submitted, but other threads must be
228        // able to see which entries can be used and which are already un used by an other thread
229        // for convenience, return both the index and the pointer to the sqe
230        // sqe == &sqes[idx]
231        struct $io_context * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) {
232                __cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want);
[df40a56]233
[78da4ab]234                disable_interrupts();
235                processor * proc = __cfaabi_tls.this_processor;
[dddb3dd0]236                $io_context * ctx = proc->io.ctx;
[78da4ab]237                /* paranoid */ verify( __cfaabi_tls.this_processor );
[dddb3dd0]238                /* paranoid */ verify( ctx );
[78da4ab]239
[dddb3dd0]240                __cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n");
[78da4ab]241
[dddb3dd0]242                // We can proceed to the fast path
243                if( __alloc(ctx, idxs, want) ) {
244                        // Allocation was successful
245                        __STATS__( true, io.alloc.fast += 1; )
246                        enable_interrupts( __cfaabi_dbg_ctx );
[df40a56]247
[dddb3dd0]248                        __cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd);
[2fafe7e]249
[dddb3dd0]250                        __fill( sqes, want, idxs, ctx );
251                        return ctx;
[df40a56]252                }
[dddb3dd0]253                // The fast path failed, fallback
254                __STATS__( true, io.alloc.fail += 1; )
[df40a56]255
[78da4ab]256                // Fast path failed, fallback on arbitration
[d60d30e]257                __STATS__( true, io.alloc.slow += 1; )
[78da4ab]258                enable_interrupts( __cfaabi_dbg_ctx );
259
[dddb3dd0]260                $io_arbiter * ioarb = proc->cltr->io.arbiter;
261                /* paranoid */ verify( ioarb );
262
[78da4ab]263                __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n");
264
265                struct $io_context * ret = __ioarbiter_allocate(*ioarb, proc, idxs, want);
266
[dddb3dd0]267                __cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd);
[df40a56]268
[78da4ab]269                __fill( sqes, want, idxs,ret );
270                return ret;
[df40a56]271        }
272
[426f60c]273
[78da4ab]274        //=============================================================================================
275        // submission
[dddb3dd0]276        static inline void __submit( struct $io_context * ctx, __u32 idxs[], __u32 have, bool lazy) {
[78da4ab]277                // We can proceed to the fast path
278                // Get the right objects
279                __sub_ring_t & sq = ctx->sq;
280                const __u32 mask  = *sq.mask;
[dddb3dd0]281                __u32 tail = *sq.kring.tail;
[78da4ab]282
283                // Add the sqes to the array
284                for( i; have ) {
[dddb3dd0]285                        __cfadbg_print_safe(io, "Kernel I/O : __submit loop\n");
[78da4ab]286                        sq.kring.array[ (tail + i) & mask ] = idxs[i];
[426f60c]287                }
288
[78da4ab]289                // Make the sqes visible to the submitter
[dddb3dd0]290                __atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE);
291                sq.to_submit++;
[426f60c]292
[dddb3dd0]293                ctx->proc->io.pending = true;
294                ctx->proc->io.dirty   = true;
295                if(sq.to_submit > 30 || !lazy) {
296                        __cfa_io_flush( ctx->proc );
297                }
[78da4ab]298        }
[2489d31]299
[dddb3dd0]300        void cfa_io_submit( struct $io_context * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) {
301                __cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager");
[5dadc9b]302
[78da4ab]303                disable_interrupts();
304                processor * proc = __cfaabi_tls.this_processor;
305                $io_context * ctx = proc->io.ctx;
[dddb3dd0]306                /* paranoid */ verify( __cfaabi_tls.this_processor );
307                /* paranoid */ verify( ctx );
[e46c753]308
[78da4ab]309                // Can we proceed to the fast path
[dddb3dd0]310                if( ctx == inctx )              // We have the right instance?
[78da4ab]311                {
[dddb3dd0]312                        __submit(ctx, idxs, have, lazy);
[e46c753]313
[78da4ab]314                        // Mark the instance as no longer in-use, re-enable interrupts and return
[d60d30e]315                        __STATS__( true, io.submit.fast += 1; )
[78da4ab]316                        enable_interrupts( __cfaabi_dbg_ctx );
[ece0e80]317
[78da4ab]318                        __cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n");
319                        return;
[e46c753]320                }
[d384787]321
[78da4ab]322                // Fast path failed, fallback on arbitration
[d60d30e]323                __STATS__( true, io.submit.slow += 1; )
[78da4ab]324                enable_interrupts( __cfaabi_dbg_ctx );
[5dadc9b]325
[78da4ab]326                __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n");
[426f60c]327
[dddb3dd0]328                __ioarbiter_submit(*inctx->arbiter, inctx, idxs, have, lazy);
[78da4ab]329        }
[2fab24e3]330
[78da4ab]331        //=============================================================================================
332        // Flushing
[426f60c]333        // Go through the ring's submit queue and release everything that has already been consumed
334        // by io_uring
[78da4ab]335        // This cannot be done by multiple threads
336        static __u32 __release_sqes( struct $io_context & ctx ) {
337                const __u32 mask = *ctx.sq.mask;
[732b406]338
[426f60c]339                __attribute__((unused))
[78da4ab]340                __u32 ctail = *ctx.sq.kring.tail;    // get the current tail of the queue
341                __u32 chead = *ctx.sq.kring.head;        // get the current head of the queue
342                __u32 phead = ctx.sq.kring.released; // get the head the last time we were here
343
344                __u32 ftail = ctx.sq.free_ring.tail;  // get the current tail of the queue
[732b406]345
[426f60c]346                // the 3 fields are organized like this diagram
347                // except it's are ring
348                // ---+--------+--------+----
349                // ---+--------+--------+----
350                //    ^        ^        ^
351                // phead    chead    ctail
352
353                // make sure ctail doesn't wrap around and reach phead
354                /* paranoid */ verify(
355                           (ctail >= chead && chead >= phead)
356                        || (chead >= phead && phead >= ctail)
357                        || (phead >= ctail && ctail >= chead)
358                );
359
360                // find the range we need to clear
[4998155]361                __u32 count = chead - phead;
[426f60c]362
[78da4ab]363                if(count == 0) {
364                        return 0;
365                }
366
[426f60c]367                // We acquired an previous-head/current-head range
368                // go through the range and release the sqes
[34b61882]369                for( i; count ) {
[dddb3dd0]370                        __cfadbg_print_safe(io, "Kernel I/O : release loop\n");
[78da4ab]371                        __u32 idx = ctx.sq.kring.array[ (phead + i) & mask ];
372                        ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx;
[34b61882]373                }
[78da4ab]374
375                ctx.sq.kring.released = chead;          // note up to were we processed
376                __atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST);
377
378                __ioarbiter_notify(ctx);
379
[34b61882]380                return count;
381        }
[35285fd]382
[78da4ab]383//=============================================================================================
384// I/O Arbiter
385//=============================================================================================
386        static $io_context * __ioarbiter_allocate( $io_arbiter & mutex this, processor * proc, __u32 idxs[], __u32 want ) {
387                __cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n");
388
[d60d30e]389                __STATS__( false, io.alloc.block += 1; )
390
[78da4ab]391                // No one has any resources left, wait for something to finish
392                // Mark as pending
393                __atomic_store_n( &this.pending.flag, true, __ATOMIC_SEQ_CST );
394
395                // Wait for our turn to submit
396                wait( this.pending.blocked, want );
397
398                __attribute((unused)) bool ret =
399                __alloc( this.pending.ctx, idxs, want);
400                /* paranoid */ verify( ret );
401
402                return this.pending.ctx;
[dddb3dd0]403
[78da4ab]404        }
405
406        static void __ioarbiter_notify( $io_arbiter & mutex this, $io_context * ctx ) {
407                /* paranoid */ verify( !is_empty(this.pending.blocked) );
408                this.pending.ctx = ctx;
409
410                while( !is_empty(this.pending.blocked) ) {
[dddb3dd0]411                        __cfadbg_print_safe(io, "Kernel I/O : notifying\n");
[78da4ab]412                        __u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head;
413                        __u32 want = front( this.pending.blocked );
414
415                        if( have > want ) return;
416
417                        signal_block( this.pending.blocked );
418                }
419
420                this.pending.flag = false;
421        }
422
423        static void __ioarbiter_notify( $io_context & ctx ) {
424                if(__atomic_load_n( &ctx.arbiter->pending.flag, __ATOMIC_SEQ_CST)) {
425                        __ioarbiter_notify( *ctx.arbiter, &ctx );
426                }
427        }
428
429        // Simply append to the pending
[dddb3dd0]430        static void __ioarbiter_submit( $io_arbiter & mutex this, $io_context * ctx, __u32 idxs[], __u32 have, bool lazy ) {
[78da4ab]431                __cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd);
432
433                /* paranoid */ verify( &this == ctx->arbiter );
434
435                // Mark as pending
436                __atomic_store_n( &ctx->ext_sq.empty, false, __ATOMIC_SEQ_CST );
437
438                __cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have);
439
440                // Wait for our turn to submit
441                wait( ctx->ext_sq.blocked );
442
443                // Submit our indexes
[dddb3dd0]444                __submit(ctx, idxs, have, lazy);
[78da4ab]445
446                __cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have);
447        }
448
449        static void __ioarbiter_flush( $io_arbiter & mutex this, $io_context * ctx ) {
450                /* paranoid */ verify( &this == ctx->arbiter );
451
[d60d30e]452                __STATS__( false, io.flush.external += 1; )
453
[78da4ab]454                __cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n");
455
456                condition & blcked = ctx->ext_sq.blocked;
457                /* paranoid */ verify( ctx->ext_sq.empty == is_empty( blcked ) );
458                while(!is_empty( blcked )) {
459                        signal_block( blcked );
460                }
461
462                ctx->ext_sq.empty = true;
463        }
[47746a2]464#endif
Note: See TracBrowser for help on using the repository browser.