[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__ |
---|
[43784ac] | 17 | #define _GNU_SOURCE |
---|
[3e2b9c9] | 18 | |
---|
[20ab637] | 19 | #if defined(__CFA_DEBUG__) |
---|
[d60d30e] | 20 | // #define __CFA_DEBUG_PRINT_IO__ |
---|
| 21 | // #define __CFA_DEBUG_PRINT_IO_CORE__ |
---|
[20ab637] | 22 | #endif |
---|
[4069faad] | 23 | |
---|
[f6660520] | 24 | |
---|
[3e2b9c9] | 25 | #if defined(CFA_HAVE_LINUX_IO_URING_H) |
---|
[31bb2e1] | 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> |
---|
[d3605f8] | 35 | #include <sys/uio.h> |
---|
[92976d9] | 36 | |
---|
| 37 | #include <linux/io_uring.h> |
---|
| 38 | } |
---|
| 39 | |
---|
[3e2b9c9] | 40 | #include "stats.hfa" |
---|
| 41 | #include "kernel.hfa" |
---|
| 42 | #include "kernel/fwd.hfa" |
---|
[708ae38] | 43 | #include "kernel/private.hfa" |
---|
[78a580d] | 44 | #include "kernel/cluster.hfa" |
---|
[3e2b9c9] | 45 | #include "io/types.hfa" |
---|
[185efe6] | 46 | |
---|
[2fab24e3] | 47 | __attribute__((unused)) static const char * opcodes[] = { |
---|
[426f60c] | 48 | "OP_NOP", |
---|
| 49 | "OP_READV", |
---|
| 50 | "OP_WRITEV", |
---|
| 51 | "OP_FSYNC", |
---|
| 52 | "OP_READ_FIXED", |
---|
| 53 | "OP_WRITE_FIXED", |
---|
| 54 | "OP_POLL_ADD", |
---|
| 55 | "OP_POLL_REMOVE", |
---|
| 56 | "OP_SYNC_FILE_RANGE", |
---|
| 57 | "OP_SENDMSG", |
---|
| 58 | "OP_RECVMSG", |
---|
| 59 | "OP_TIMEOUT", |
---|
| 60 | "OP_TIMEOUT_REMOVE", |
---|
| 61 | "OP_ACCEPT", |
---|
| 62 | "OP_ASYNC_CANCEL", |
---|
| 63 | "OP_LINK_TIMEOUT", |
---|
| 64 | "OP_CONNECT", |
---|
| 65 | "OP_FALLOCATE", |
---|
| 66 | "OP_OPENAT", |
---|
| 67 | "OP_CLOSE", |
---|
| 68 | "OP_FILES_UPDATE", |
---|
| 69 | "OP_STATX", |
---|
| 70 | "OP_READ", |
---|
| 71 | "OP_WRITE", |
---|
| 72 | "OP_FADVISE", |
---|
| 73 | "OP_MADVISE", |
---|
| 74 | "OP_SEND", |
---|
| 75 | "OP_RECV", |
---|
| 76 | "OP_OPENAT2", |
---|
| 77 | "OP_EPOLL_CTL", |
---|
| 78 | "OP_SPLICE", |
---|
| 79 | "OP_PROVIDE_BUFFERS", |
---|
| 80 | "OP_REMOVE_BUFFERS", |
---|
| 81 | "OP_TEE", |
---|
| 82 | "INVALID_OP" |
---|
| 83 | }; |
---|
| 84 | |
---|
[8bee858] | 85 | static io_context$ * __ioarbiter_allocate( io_arbiter$ & this, __u32 idxs[], __u32 want ); |
---|
| 86 | static void __ioarbiter_submit( io_context$ * , __u32 idxs[], __u32 have, bool lazy ); |
---|
| 87 | static void __ioarbiter_flush ( io_context$ & ); |
---|
| 88 | static inline void __ioarbiter_notify( io_context$ & ctx ); |
---|
[92976d9] | 89 | //============================================================================================= |
---|
| 90 | // I/O Polling |
---|
| 91 | //============================================================================================= |
---|
[8bee858] | 92 | static inline unsigned __flush( struct io_context$ & ); |
---|
| 93 | static inline __u32 __release_sqes( struct io_context$ & ); |
---|
[24e321c] | 94 | extern void __kernel_unpark( thread$ * thrd, unpark_hint ); |
---|
[1d5e4711] | 95 | |
---|
[8bee858] | 96 | static void ioring_syscsll( struct io_context$ & ctx, unsigned int min_comp, unsigned int flags ) { |
---|
[18f7858] | 97 | __STATS__( true, io.calls.flush++; ) |
---|
[bdfd0bd] | 98 | int ret; |
---|
| 99 | for() { |
---|
| 100 | ret = syscall( __NR_io_uring_enter, ctx.fd, ctx.sq.to_submit, min_comp, flags, (sigset_t *)0p, _NSIG / 8); |
---|
| 101 | if( ret < 0 ) { |
---|
| 102 | switch((int)errno) { |
---|
| 103 | case EINTR: |
---|
| 104 | continue; |
---|
| 105 | case EAGAIN: |
---|
| 106 | case EBUSY: |
---|
| 107 | // Update statistics |
---|
| 108 | __STATS__( false, io.calls.errors.busy ++; ) |
---|
| 109 | return false; |
---|
| 110 | default: |
---|
| 111 | abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) ); |
---|
| 112 | } |
---|
[18f7858] | 113 | } |
---|
[bdfd0bd] | 114 | break; |
---|
[18f7858] | 115 | } |
---|
| 116 | |
---|
| 117 | __cfadbg_print_safe(io, "Kernel I/O : %u submitted to io_uring %d\n", ret, ctx.fd); |
---|
| 118 | __STATS__( true, io.calls.submitted += ret; ) |
---|
| 119 | /* paranoid */ verify( ctx.sq.to_submit <= *ctx.sq.num ); |
---|
| 120 | /* paranoid */ verify( ctx.sq.to_submit >= ret ); |
---|
| 121 | |
---|
| 122 | ctx.sq.to_submit -= ret; |
---|
| 123 | |
---|
| 124 | /* paranoid */ verify( ctx.sq.to_submit <= *ctx.sq.num ); |
---|
| 125 | |
---|
| 126 | // Release the consumed SQEs |
---|
| 127 | __release_sqes( ctx ); |
---|
| 128 | |
---|
[dddb3dd0] | 129 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
[6f121b8] | 130 | |
---|
[18f7858] | 131 | __atomic_store_n(&ctx.proc->io.pending, false, __ATOMIC_RELAXED); |
---|
| 132 | } |
---|
| 133 | |
---|
[8bee858] | 134 | static bool try_acquire( io_context$ * ctx ) __attribute__((nonnull(1))) { |
---|
[18f7858] | 135 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 136 | /* paranoid */ verify( ready_schedule_islocked() ); |
---|
[92976d9] | 137 | |
---|
[d60d30e] | 138 | |
---|
[3caf5e3] | 139 | { |
---|
| 140 | const __u32 head = *ctx->cq.head; |
---|
| 141 | const __u32 tail = *ctx->cq.tail; |
---|
| 142 | |
---|
| 143 | if(head == tail) return false; |
---|
| 144 | } |
---|
[c1c95b1] | 145 | |
---|
[3caf5e3] | 146 | // Drain the queue |
---|
[4ecc35a] | 147 | if(!__atomic_try_acquire(&ctx->cq.lock)) { |
---|
[54c1196] | 148 | __STATS__( false, io.calls.locked++; ) |
---|
[4ecc35a] | 149 | return false; |
---|
| 150 | } |
---|
| 151 | |
---|
[18f7858] | 152 | return true; |
---|
| 153 | } |
---|
| 154 | |
---|
[8bee858] | 155 | static bool __cfa_do_drain( io_context$ * ctx, cluster * cltr ) __attribute__((nonnull(1, 2))) { |
---|
[18f7858] | 156 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 157 | /* paranoid */ verify( ready_schedule_islocked() ); |
---|
| 158 | /* paranoid */ verify( ctx->cq.lock == true ); |
---|
| 159 | |
---|
| 160 | const __u32 mask = *ctx->cq.mask; |
---|
[7affcda] | 161 | const __u32 num = *ctx->cq.num; |
---|
[78a580d] | 162 | unsigned long long ts_prev = ctx->cq.ts; |
---|
[7affcda] | 163 | unsigned long long ts_next; |
---|
[78a580d] | 164 | |
---|
[7affcda] | 165 | // We might need to do this multiple times if more events completed than can fit in the queue. |
---|
| 166 | for() { |
---|
| 167 | // re-read the head and tail in case it already changed. |
---|
| 168 | const __u32 head = *ctx->cq.head; |
---|
| 169 | const __u32 tail = *ctx->cq.tail; |
---|
| 170 | const __u32 count = tail - head; |
---|
| 171 | __STATS__( false, io.calls.drain++; io.calls.completed += count; ) |
---|
[3caf5e3] | 172 | |
---|
[7affcda] | 173 | for(i; count) { |
---|
| 174 | unsigned idx = (head + i) & mask; |
---|
| 175 | volatile struct io_uring_cqe & cqe = ctx->cq.cqes[idx]; |
---|
[92976d9] | 176 | |
---|
[7affcda] | 177 | /* paranoid */ verify(&cqe); |
---|
[92976d9] | 178 | |
---|
[7affcda] | 179 | struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data; |
---|
| 180 | // __cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future ); |
---|
[78da4ab] | 181 | |
---|
[7affcda] | 182 | __kernel_unpark( fulfil( *future, cqe.res, false ), UNPARK_LOCAL ); |
---|
| 183 | } |
---|
| 184 | |
---|
| 185 | ts_next = ctx->cq.ts = rdtscl(); |
---|
[78da4ab] | 186 | |
---|
[7affcda] | 187 | // Mark to the kernel that the cqe has been seen |
---|
| 188 | // Ensure that the kernel only sees the new value of the head index after the CQEs have been read. |
---|
| 189 | __atomic_store_n( ctx->cq.head, head + count, __ATOMIC_SEQ_CST ); |
---|
| 190 | ctx->proc->idle_wctx.drain_time = ts_next; |
---|
[2d8f7b0] | 191 | |
---|
[7affcda] | 192 | if(likely(count < num)) break; |
---|
| 193 | |
---|
| 194 | ioring_syscsll( *ctx, 0, IORING_ENTER_GETEVENTS); |
---|
| 195 | } |
---|
[92976d9] | 196 | |
---|
[1e6ffb44] | 197 | __cfadbg_print_safe(io, "Kernel I/O : %u completed age %llu\n", count, ts_next); |
---|
[e9c0b4c] | 198 | /* paranoid */ verify( ready_schedule_islocked() ); |
---|
[dddb3dd0] | 199 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 200 | |
---|
[4ecc35a] | 201 | __atomic_unlock(&ctx->cq.lock); |
---|
| 202 | |
---|
[78a580d] | 203 | touch_tsc( cltr->sched.io.tscs, ctx->cq.id, ts_prev, ts_next ); |
---|
| 204 | |
---|
[c1c95b1] | 205 | return true; |
---|
[92976d9] | 206 | } |
---|
| 207 | |
---|
[4479890] | 208 | bool __cfa_io_drain( processor * proc ) { |
---|
| 209 | bool local = false; |
---|
| 210 | bool remote = false; |
---|
| 211 | |
---|
[18f7858] | 212 | ready_schedule_lock(); |
---|
| 213 | |
---|
[4479890] | 214 | cluster * const cltr = proc->cltr; |
---|
[8bee858] | 215 | io_context$ * const ctx = proc->io.ctx; |
---|
[4479890] | 216 | /* paranoid */ verify( cltr ); |
---|
| 217 | /* paranoid */ verify( ctx ); |
---|
| 218 | |
---|
| 219 | with(cltr->sched) { |
---|
| 220 | const size_t ctxs_count = io.count; |
---|
| 221 | |
---|
| 222 | /* paranoid */ verify( ready_schedule_islocked() ); |
---|
| 223 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 224 | /* paranoid */ verify( active_processor() == proc ); |
---|
| 225 | /* paranoid */ verify( __shard_factor.io > 0 ); |
---|
| 226 | /* paranoid */ verify( ctxs_count > 0 ); |
---|
| 227 | /* paranoid */ verify( ctx->cq.id < ctxs_count ); |
---|
| 228 | |
---|
| 229 | const unsigned this_cache = cache_id(cltr, ctx->cq.id / __shard_factor.io); |
---|
| 230 | const unsigned long long ctsc = rdtscl(); |
---|
| 231 | |
---|
[b035046] | 232 | if(proc->io.target == UINT_MAX) { |
---|
[4479890] | 233 | uint64_t chaos = __tls_rand(); |
---|
| 234 | unsigned ext = chaos & 0xff; |
---|
| 235 | unsigned other = (chaos >> 8) % (ctxs_count); |
---|
| 236 | |
---|
| 237 | if(ext < 3 || __atomic_load_n(&caches[other / __shard_factor.io].id, __ATOMIC_RELAXED) == this_cache) { |
---|
| 238 | proc->io.target = other; |
---|
| 239 | } |
---|
| 240 | } |
---|
| 241 | else { |
---|
| 242 | const unsigned target = proc->io.target; |
---|
[2af1943] | 243 | /* paranoid */ verify( io.tscs[target].t.tv != ULLONG_MAX ); |
---|
[18f7858] | 244 | HELP: if(target < ctxs_count) { |
---|
[4479890] | 245 | const unsigned long long cutoff = calc_cutoff(ctsc, ctx->cq.id, ctxs_count, io.data, io.tscs, __shard_factor.io); |
---|
[2af1943] | 246 | const unsigned long long age = moving_average(ctsc, io.tscs[target].t.tv, io.tscs[target].t.ma); |
---|
[edf247b] | 247 | __cfadbg_print_safe(io, "Kernel I/O: Help attempt on %u from %u, age %'llu vs cutoff %'llu, %s\n", target, ctx->cq.id, age, cutoff, age > cutoff ? "yes" : "no"); |
---|
[18f7858] | 248 | if(age <= cutoff) break HELP; |
---|
| 249 | |
---|
| 250 | if(!try_acquire(io.data[target])) break HELP; |
---|
| 251 | |
---|
| 252 | if(!__cfa_do_drain( io.data[target], cltr )) break HELP; |
---|
| 253 | |
---|
| 254 | remote = true; |
---|
[8a5e357] | 255 | __STATS__( true, io.calls.helped++; ) |
---|
[4479890] | 256 | } |
---|
[b035046] | 257 | proc->io.target = UINT_MAX; |
---|
[4479890] | 258 | } |
---|
| 259 | } |
---|
| 260 | |
---|
| 261 | |
---|
| 262 | // Drain the local queue |
---|
[18f7858] | 263 | if(try_acquire( proc->io.ctx )) { |
---|
| 264 | local = __cfa_do_drain( proc->io.ctx, cltr ); |
---|
| 265 | } |
---|
[4479890] | 266 | |
---|
| 267 | /* paranoid */ verify( ready_schedule_islocked() ); |
---|
| 268 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 269 | /* paranoid */ verify( active_processor() == proc ); |
---|
[18f7858] | 270 | |
---|
| 271 | ready_schedule_unlock(); |
---|
[4479890] | 272 | return local || remote; |
---|
| 273 | } |
---|
| 274 | |
---|
[18f7858] | 275 | bool __cfa_io_flush( processor * proc ) { |
---|
[dddb3dd0] | 276 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 277 | /* paranoid */ verify( proc ); |
---|
| 278 | /* paranoid */ verify( proc->io.ctx ); |
---|
[1539bbd] | 279 | |
---|
[8bee858] | 280 | io_context$ & ctx = *proc->io.ctx; |
---|
[78da4ab] | 281 | |
---|
[11054eb] | 282 | __ioarbiter_flush( ctx ); |
---|
[3c039b0] | 283 | |
---|
[18f7858] | 284 | if(ctx.sq.to_submit != 0) { |
---|
| 285 | ioring_syscsll(ctx, 0, 0); |
---|
[21a5bfb7] | 286 | |
---|
| 287 | } |
---|
[61dd73d] | 288 | |
---|
[18f7858] | 289 | return __cfa_io_drain( proc ); |
---|
[61dd73d] | 290 | } |
---|
[f6660520] | 291 | |
---|
[92976d9] | 292 | //============================================================================================= |
---|
| 293 | // I/O Submissions |
---|
| 294 | //============================================================================================= |
---|
| 295 | |
---|
[2d8f7b0] | 296 | // Submition steps : |
---|
[e46c753] | 297 | // 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones |
---|
[2d8f7b0] | 298 | // listed in sq.array are visible by the kernel. For those not listed, the kernel does not |
---|
| 299 | // offer any assurance that an entry is not being filled by multiple flags. Therefore, we |
---|
| 300 | // need to write an allocator that allows allocating concurrently. |
---|
| 301 | // |
---|
[e46c753] | 302 | // 2 - Actually fill the submit entry, this is the only simple and straightforward step. |
---|
[2d8f7b0] | 303 | // |
---|
[e46c753] | 304 | // 3 - Append the entry index to the array and adjust the tail accordingly. This operation |
---|
[2d8f7b0] | 305 | // needs to arrive to two concensus at the same time: |
---|
| 306 | // A - The order in which entries are listed in the array: no two threads must pick the |
---|
| 307 | // same index for their entries |
---|
| 308 | // B - When can the tail be update for the kernel. EVERY entries in the array between |
---|
| 309 | // head and tail must be fully filled and shouldn't ever be touched again. |
---|
| 310 | // |
---|
[78da4ab] | 311 | //============================================================================================= |
---|
| 312 | // Allocation |
---|
| 313 | // for user's convenience fill the sqes from the indexes |
---|
[8bee858] | 314 | static inline void __fill(struct io_uring_sqe * out_sqes[], __u32 want, __u32 idxs[], struct io_context$ * ctx) { |
---|
[78da4ab] | 315 | struct io_uring_sqe * sqes = ctx->sq.sqes; |
---|
| 316 | for(i; want) { |
---|
[1e6ffb44] | 317 | // __cfadbg_print_safe(io, "Kernel I/O : filling loop\n"); |
---|
[78da4ab] | 318 | out_sqes[i] = &sqes[idxs[i]]; |
---|
| 319 | } |
---|
| 320 | } |
---|
[2489d31] | 321 | |
---|
[78da4ab] | 322 | // Try to directly allocate from the a given context |
---|
| 323 | // Not thread-safe |
---|
[8bee858] | 324 | static inline bool __alloc(struct io_context$ * ctx, __u32 idxs[], __u32 want) { |
---|
[78da4ab] | 325 | __sub_ring_t & sq = ctx->sq; |
---|
| 326 | const __u32 mask = *sq.mask; |
---|
| 327 | __u32 fhead = sq.free_ring.head; // get the current head of the queue |
---|
| 328 | __u32 ftail = sq.free_ring.tail; // get the current tail of the queue |
---|
[2489d31] | 329 | |
---|
[78da4ab] | 330 | // If we don't have enough sqes, fail |
---|
| 331 | if((ftail - fhead) < want) { return false; } |
---|
[426f60c] | 332 | |
---|
[78da4ab] | 333 | // copy all the indexes we want from the available list |
---|
| 334 | for(i; want) { |
---|
[1e6ffb44] | 335 | // __cfadbg_print_safe(io, "Kernel I/O : allocating loop\n"); |
---|
[78da4ab] | 336 | idxs[i] = sq.free_ring.array[(fhead + i) & mask]; |
---|
[6f121b8] | 337 | } |
---|
[2489d31] | 338 | |
---|
[78da4ab] | 339 | // Advance the head to mark the indexes as consumed |
---|
| 340 | __atomic_store_n(&sq.free_ring.head, fhead + want, __ATOMIC_RELEASE); |
---|
[df40a56] | 341 | |
---|
[78da4ab] | 342 | // return success |
---|
| 343 | return true; |
---|
| 344 | } |
---|
[df40a56] | 345 | |
---|
[78da4ab] | 346 | // Allocate an submit queue entry. |
---|
| 347 | // The kernel cannot see these entries until they are submitted, but other threads must be |
---|
| 348 | // able to see which entries can be used and which are already un used by an other thread |
---|
| 349 | // for convenience, return both the index and the pointer to the sqe |
---|
| 350 | // sqe == &sqes[idx] |
---|
[8bee858] | 351 | struct io_context$ * cfa_io_allocate(struct io_uring_sqe * sqes[], __u32 idxs[], __u32 want) libcfa_public { |
---|
[1e6ffb44] | 352 | // __cfadbg_print_safe(io, "Kernel I/O : attempting to allocate %u\n", want); |
---|
[df40a56] | 353 | |
---|
[78da4ab] | 354 | disable_interrupts(); |
---|
| 355 | processor * proc = __cfaabi_tls.this_processor; |
---|
[8bee858] | 356 | io_context$ * ctx = proc->io.ctx; |
---|
[78da4ab] | 357 | /* paranoid */ verify( __cfaabi_tls.this_processor ); |
---|
[dddb3dd0] | 358 | /* paranoid */ verify( ctx ); |
---|
[78da4ab] | 359 | |
---|
[1e6ffb44] | 360 | // __cfadbg_print_safe(io, "Kernel I/O : attempting to fast allocation\n"); |
---|
[78da4ab] | 361 | |
---|
[dddb3dd0] | 362 | // We can proceed to the fast path |
---|
| 363 | if( __alloc(ctx, idxs, want) ) { |
---|
| 364 | // Allocation was successful |
---|
| 365 | __STATS__( true, io.alloc.fast += 1; ) |
---|
[a3821fa] | 366 | enable_interrupts(); |
---|
[df40a56] | 367 | |
---|
[1e6ffb44] | 368 | // __cfadbg_print_safe(io, "Kernel I/O : fast allocation successful from ring %d\n", ctx->fd); |
---|
[2fafe7e] | 369 | |
---|
[dddb3dd0] | 370 | __fill( sqes, want, idxs, ctx ); |
---|
| 371 | return ctx; |
---|
[df40a56] | 372 | } |
---|
[dddb3dd0] | 373 | // The fast path failed, fallback |
---|
| 374 | __STATS__( true, io.alloc.fail += 1; ) |
---|
[df40a56] | 375 | |
---|
[78da4ab] | 376 | // Fast path failed, fallback on arbitration |
---|
[d60d30e] | 377 | __STATS__( true, io.alloc.slow += 1; ) |
---|
[a3821fa] | 378 | enable_interrupts(); |
---|
[78da4ab] | 379 | |
---|
[8bee858] | 380 | io_arbiter$ * ioarb = proc->cltr->io.arbiter; |
---|
[dddb3dd0] | 381 | /* paranoid */ verify( ioarb ); |
---|
| 382 | |
---|
[1e6ffb44] | 383 | // __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for allocation\n"); |
---|
[78da4ab] | 384 | |
---|
[8bee858] | 385 | struct io_context$ * ret = __ioarbiter_allocate(*ioarb, idxs, want); |
---|
[78da4ab] | 386 | |
---|
[1e6ffb44] | 387 | // __cfadbg_print_safe(io, "Kernel I/O : slow allocation completed from ring %d\n", ret->fd); |
---|
[df40a56] | 388 | |
---|
[78da4ab] | 389 | __fill( sqes, want, idxs,ret ); |
---|
| 390 | return ret; |
---|
[df40a56] | 391 | } |
---|
| 392 | |
---|
[78da4ab] | 393 | //============================================================================================= |
---|
| 394 | // submission |
---|
[8bee858] | 395 | static inline void __submit_only( struct io_context$ * ctx, __u32 idxs[], __u32 have) { |
---|
[78da4ab] | 396 | // We can proceed to the fast path |
---|
| 397 | // Get the right objects |
---|
| 398 | __sub_ring_t & sq = ctx->sq; |
---|
| 399 | const __u32 mask = *sq.mask; |
---|
[dddb3dd0] | 400 | __u32 tail = *sq.kring.tail; |
---|
[78da4ab] | 401 | |
---|
| 402 | // Add the sqes to the array |
---|
| 403 | for( i; have ) { |
---|
[1e6ffb44] | 404 | // __cfadbg_print_safe(io, "Kernel I/O : __submit loop\n"); |
---|
[78da4ab] | 405 | sq.kring.array[ (tail + i) & mask ] = idxs[i]; |
---|
[426f60c] | 406 | } |
---|
| 407 | |
---|
[78da4ab] | 408 | // Make the sqes visible to the submitter |
---|
[dddb3dd0] | 409 | __atomic_store_n(sq.kring.tail, tail + have, __ATOMIC_RELEASE); |
---|
[e8ac228] | 410 | sq.to_submit += have; |
---|
[426f60c] | 411 | |
---|
[d529ad0] | 412 | __atomic_store_n(&ctx->proc->io.pending, true, __ATOMIC_RELAXED); |
---|
| 413 | __atomic_store_n(&ctx->proc->io.dirty , true, __ATOMIC_RELAXED); |
---|
[2432e8e] | 414 | } |
---|
| 415 | |
---|
[8bee858] | 416 | static inline void __submit( struct io_context$ * ctx, __u32 idxs[], __u32 have, bool lazy) { |
---|
[2432e8e] | 417 | __sub_ring_t & sq = ctx->sq; |
---|
| 418 | __submit_only(ctx, idxs, have); |
---|
| 419 | |
---|
[70b4aeb9] | 420 | if(sq.to_submit > 30) { |
---|
| 421 | __tls_stats()->io.flush.full++; |
---|
[18f7858] | 422 | __cfa_io_flush( ctx->proc ); |
---|
[70b4aeb9] | 423 | } |
---|
| 424 | if(!lazy) { |
---|
| 425 | __tls_stats()->io.flush.eager++; |
---|
[18f7858] | 426 | __cfa_io_flush( ctx->proc ); |
---|
[dddb3dd0] | 427 | } |
---|
[78da4ab] | 428 | } |
---|
[2489d31] | 429 | |
---|
[8bee858] | 430 | void cfa_io_submit( struct io_context$ * inctx, __u32 idxs[], __u32 have, bool lazy ) __attribute__((nonnull (1))) libcfa_public { |
---|
[1e6ffb44] | 431 | // __cfadbg_print_safe(io, "Kernel I/O : attempting to submit %u (%s)\n", have, lazy ? "lazy" : "eager"); |
---|
[5dadc9b] | 432 | |
---|
[78da4ab] | 433 | disable_interrupts(); |
---|
[7ce8873] | 434 | __STATS__( true, if(!lazy) io.submit.eagr += 1; ) |
---|
[78da4ab] | 435 | processor * proc = __cfaabi_tls.this_processor; |
---|
[8bee858] | 436 | io_context$ * ctx = proc->io.ctx; |
---|
[dddb3dd0] | 437 | /* paranoid */ verify( __cfaabi_tls.this_processor ); |
---|
| 438 | /* paranoid */ verify( ctx ); |
---|
[e46c753] | 439 | |
---|
[78da4ab] | 440 | // Can we proceed to the fast path |
---|
[dddb3dd0] | 441 | if( ctx == inctx ) // We have the right instance? |
---|
[78da4ab] | 442 | { |
---|
[dddb3dd0] | 443 | __submit(ctx, idxs, have, lazy); |
---|
[e46c753] | 444 | |
---|
[78da4ab] | 445 | // Mark the instance as no longer in-use, re-enable interrupts and return |
---|
[d60d30e] | 446 | __STATS__( true, io.submit.fast += 1; ) |
---|
[a3821fa] | 447 | enable_interrupts(); |
---|
[ece0e80] | 448 | |
---|
[1e6ffb44] | 449 | // __cfadbg_print_safe(io, "Kernel I/O : submitted on fast path\n"); |
---|
[78da4ab] | 450 | return; |
---|
[e46c753] | 451 | } |
---|
[d384787] | 452 | |
---|
[78da4ab] | 453 | // Fast path failed, fallback on arbitration |
---|
[d60d30e] | 454 | __STATS__( true, io.submit.slow += 1; ) |
---|
[a3821fa] | 455 | enable_interrupts(); |
---|
[5dadc9b] | 456 | |
---|
[1e6ffb44] | 457 | // __cfadbg_print_safe(io, "Kernel I/O : falling back on arbiter for submission\n"); |
---|
[426f60c] | 458 | |
---|
[11054eb] | 459 | __ioarbiter_submit(inctx, idxs, have, lazy); |
---|
[78da4ab] | 460 | } |
---|
[2fab24e3] | 461 | |
---|
[78da4ab] | 462 | //============================================================================================= |
---|
| 463 | // Flushing |
---|
[426f60c] | 464 | // Go through the ring's submit queue and release everything that has already been consumed |
---|
| 465 | // by io_uring |
---|
[78da4ab] | 466 | // This cannot be done by multiple threads |
---|
[8bee858] | 467 | static __u32 __release_sqes( struct io_context$ & ctx ) { |
---|
[78da4ab] | 468 | const __u32 mask = *ctx.sq.mask; |
---|
[732b406] | 469 | |
---|
[426f60c] | 470 | __attribute__((unused)) |
---|
[78da4ab] | 471 | __u32 ctail = *ctx.sq.kring.tail; // get the current tail of the queue |
---|
| 472 | __u32 chead = *ctx.sq.kring.head; // get the current head of the queue |
---|
| 473 | __u32 phead = ctx.sq.kring.released; // get the head the last time we were here |
---|
| 474 | |
---|
| 475 | __u32 ftail = ctx.sq.free_ring.tail; // get the current tail of the queue |
---|
[732b406] | 476 | |
---|
[426f60c] | 477 | // the 3 fields are organized like this diagram |
---|
| 478 | // except it's are ring |
---|
| 479 | // ---+--------+--------+---- |
---|
| 480 | // ---+--------+--------+---- |
---|
| 481 | // ^ ^ ^ |
---|
| 482 | // phead chead ctail |
---|
| 483 | |
---|
| 484 | // make sure ctail doesn't wrap around and reach phead |
---|
| 485 | /* paranoid */ verify( |
---|
| 486 | (ctail >= chead && chead >= phead) |
---|
| 487 | || (chead >= phead && phead >= ctail) |
---|
| 488 | || (phead >= ctail && ctail >= chead) |
---|
| 489 | ); |
---|
| 490 | |
---|
| 491 | // find the range we need to clear |
---|
[4998155] | 492 | __u32 count = chead - phead; |
---|
[426f60c] | 493 | |
---|
[78da4ab] | 494 | if(count == 0) { |
---|
| 495 | return 0; |
---|
| 496 | } |
---|
| 497 | |
---|
[426f60c] | 498 | // We acquired an previous-head/current-head range |
---|
| 499 | // go through the range and release the sqes |
---|
[34b61882] | 500 | for( i; count ) { |
---|
[1e6ffb44] | 501 | // __cfadbg_print_safe(io, "Kernel I/O : release loop\n"); |
---|
[78da4ab] | 502 | __u32 idx = ctx.sq.kring.array[ (phead + i) & mask ]; |
---|
| 503 | ctx.sq.free_ring.array[ (ftail + i) & mask ] = idx; |
---|
[34b61882] | 504 | } |
---|
[78da4ab] | 505 | |
---|
| 506 | ctx.sq.kring.released = chead; // note up to were we processed |
---|
| 507 | __atomic_store_n(&ctx.sq.free_ring.tail, ftail + count, __ATOMIC_SEQ_CST); |
---|
| 508 | |
---|
| 509 | __ioarbiter_notify(ctx); |
---|
| 510 | |
---|
[34b61882] | 511 | return count; |
---|
| 512 | } |
---|
[35285fd] | 513 | |
---|
[78da4ab] | 514 | //============================================================================================= |
---|
| 515 | // I/O Arbiter |
---|
| 516 | //============================================================================================= |
---|
[9f5a71eb] | 517 | static inline bool enqueue(__outstanding_io_queue & queue, __outstanding_io & item) { |
---|
| 518 | bool was_empty; |
---|
| 519 | |
---|
[11054eb] | 520 | // Lock the list, it's not thread safe |
---|
| 521 | lock( queue.lock __cfaabi_dbg_ctx2 ); |
---|
| 522 | { |
---|
[9f5a71eb] | 523 | was_empty = empty(queue.queue); |
---|
| 524 | |
---|
[11054eb] | 525 | // Add our request to the list |
---|
| 526 | add( queue.queue, item ); |
---|
| 527 | |
---|
| 528 | // Mark as pending |
---|
| 529 | __atomic_store_n( &queue.empty, false, __ATOMIC_SEQ_CST ); |
---|
| 530 | } |
---|
| 531 | unlock( queue.lock ); |
---|
| 532 | |
---|
[9f5a71eb] | 533 | return was_empty; |
---|
[11054eb] | 534 | } |
---|
| 535 | |
---|
| 536 | static inline bool empty(__outstanding_io_queue & queue ) { |
---|
| 537 | return __atomic_load_n( &queue.empty, __ATOMIC_SEQ_CST); |
---|
| 538 | } |
---|
| 539 | |
---|
[8bee858] | 540 | static io_context$ * __ioarbiter_allocate( io_arbiter$ & this, __u32 idxs[], __u32 want ) { |
---|
[1e6ffb44] | 541 | // __cfadbg_print_safe(io, "Kernel I/O : arbiter allocating\n"); |
---|
[78da4ab] | 542 | |
---|
[d60d30e] | 543 | __STATS__( false, io.alloc.block += 1; ) |
---|
| 544 | |
---|
[78da4ab] | 545 | // No one has any resources left, wait for something to finish |
---|
[11054eb] | 546 | // We need to add ourself to a list of pending allocs and wait for an answer |
---|
| 547 | __pending_alloc pa; |
---|
| 548 | pa.idxs = idxs; |
---|
| 549 | pa.want = want; |
---|
[78da4ab] | 550 | |
---|
[9f5a71eb] | 551 | enqueue(this.pending, (__outstanding_io&)pa); |
---|
| 552 | |
---|
| 553 | wait( pa.sem ); |
---|
[78da4ab] | 554 | |
---|
[11054eb] | 555 | return pa.ctx; |
---|
[dddb3dd0] | 556 | |
---|
[78da4ab] | 557 | } |
---|
| 558 | |
---|
[8bee858] | 559 | static void __ioarbiter_notify( io_arbiter$ & this, io_context$ * ctx ) { |
---|
[11054eb] | 560 | /* paranoid */ verify( !empty(this.pending.queue) ); |
---|
[78da4ab] | 561 | |
---|
[11054eb] | 562 | lock( this.pending.lock __cfaabi_dbg_ctx2 ); |
---|
| 563 | { |
---|
| 564 | while( !empty(this.pending.queue) ) { |
---|
| 565 | __cfadbg_print_safe(io, "Kernel I/O : notifying\n"); |
---|
| 566 | __u32 have = ctx->sq.free_ring.tail - ctx->sq.free_ring.head; |
---|
| 567 | __pending_alloc & pa = (__pending_alloc&)head( this.pending.queue ); |
---|
[78da4ab] | 568 | |
---|
[11054eb] | 569 | if( have > pa.want ) goto DONE; |
---|
| 570 | drop( this.pending.queue ); |
---|
[78da4ab] | 571 | |
---|
[11054eb] | 572 | /* paranoid */__attribute__((unused)) bool ret = |
---|
[78da4ab] | 573 | |
---|
[11054eb] | 574 | __alloc(ctx, pa.idxs, pa.want); |
---|
| 575 | |
---|
| 576 | /* paranoid */ verify( ret ); |
---|
| 577 | |
---|
| 578 | pa.ctx = ctx; |
---|
| 579 | |
---|
| 580 | post( pa.sem ); |
---|
| 581 | } |
---|
| 582 | |
---|
| 583 | this.pending.empty = true; |
---|
| 584 | DONE:; |
---|
| 585 | } |
---|
| 586 | unlock( this.pending.lock ); |
---|
[78da4ab] | 587 | } |
---|
| 588 | |
---|
[8bee858] | 589 | static void __ioarbiter_notify( io_context$ & ctx ) { |
---|
[11054eb] | 590 | if(!empty( ctx.arbiter->pending )) { |
---|
[78da4ab] | 591 | __ioarbiter_notify( *ctx.arbiter, &ctx ); |
---|
| 592 | } |
---|
| 593 | } |
---|
| 594 | |
---|
| 595 | // Simply append to the pending |
---|
[8bee858] | 596 | static void __ioarbiter_submit( io_context$ * ctx, __u32 idxs[], __u32 have, bool lazy ) { |
---|
[78da4ab] | 597 | __cfadbg_print_safe(io, "Kernel I/O : submitting %u from the arbiter to context %u\n", have, ctx->fd); |
---|
| 598 | |
---|
| 599 | __cfadbg_print_safe(io, "Kernel I/O : waiting to submit %u\n", have); |
---|
| 600 | |
---|
[11054eb] | 601 | __external_io ei; |
---|
| 602 | ei.idxs = idxs; |
---|
| 603 | ei.have = have; |
---|
| 604 | ei.lazy = lazy; |
---|
[78da4ab] | 605 | |
---|
[9f5a71eb] | 606 | bool we = enqueue(ctx->ext_sq, (__outstanding_io&)ei); |
---|
| 607 | |
---|
[d529ad0] | 608 | __atomic_store_n(&ctx->proc->io.pending, true, __ATOMIC_SEQ_CST); |
---|
[9f5a71eb] | 609 | |
---|
| 610 | if( we ) { |
---|
| 611 | sigval_t value = { PREEMPT_IO }; |
---|
[428adbc] | 612 | real_pthread_sigqueue(ctx->proc->kernel_thread, SIGUSR1, value); |
---|
[9f5a71eb] | 613 | } |
---|
| 614 | |
---|
| 615 | wait( ei.sem ); |
---|
[78da4ab] | 616 | |
---|
| 617 | __cfadbg_print_safe(io, "Kernel I/O : %u submitted from arbiter\n", have); |
---|
| 618 | } |
---|
| 619 | |
---|
[8bee858] | 620 | static void __ioarbiter_flush( io_context$ & ctx ) { |
---|
[11054eb] | 621 | if(!empty( ctx.ext_sq )) { |
---|
| 622 | __STATS__( false, io.flush.external += 1; ) |
---|
[78da4ab] | 623 | |
---|
[11054eb] | 624 | __cfadbg_print_safe(io, "Kernel I/O : arbiter flushing\n"); |
---|
[d60d30e] | 625 | |
---|
[11054eb] | 626 | lock( ctx.ext_sq.lock __cfaabi_dbg_ctx2 ); |
---|
| 627 | { |
---|
| 628 | while( !empty(ctx.ext_sq.queue) ) { |
---|
| 629 | __external_io & ei = (__external_io&)drop( ctx.ext_sq.queue ); |
---|
[78da4ab] | 630 | |
---|
[2432e8e] | 631 | __submit_only(&ctx, ei.idxs, ei.have); |
---|
[78da4ab] | 632 | |
---|
[11054eb] | 633 | post( ei.sem ); |
---|
| 634 | } |
---|
| 635 | |
---|
| 636 | ctx.ext_sq.empty = true; |
---|
| 637 | } |
---|
| 638 | unlock(ctx.ext_sq.lock ); |
---|
| 639 | } |
---|
[78da4ab] | 640 | } |
---|
[7ef162b2] | 641 | |
---|
[d3605f8] | 642 | #if defined(CFA_WITH_IO_URING_IDLE) |
---|
| 643 | bool __kernel_read(processor * proc, io_future_t & future, iovec & iov, int fd) { |
---|
[8bee858] | 644 | io_context$ * ctx = proc->io.ctx; |
---|
[6ddef36] | 645 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 646 | /* paranoid */ verify( proc == __cfaabi_tls.this_processor ); |
---|
| 647 | /* paranoid */ verify( ctx ); |
---|
[7ef162b2] | 648 | |
---|
[6ddef36] | 649 | __u32 idx; |
---|
| 650 | struct io_uring_sqe * sqe; |
---|
[7ef162b2] | 651 | |
---|
[6ddef36] | 652 | // We can proceed to the fast path |
---|
[010636f] | 653 | if( !__alloc(ctx, &idx, 1) ) { |
---|
| 654 | /* paranoid */ verify( false ); // for now check if this happens, next time just abort the sleep. |
---|
| 655 | return false; |
---|
| 656 | } |
---|
[6ddef36] | 657 | |
---|
| 658 | // Allocation was successful |
---|
| 659 | __fill( &sqe, 1, &idx, ctx ); |
---|
| 660 | |
---|
| 661 | sqe->user_data = (uintptr_t)&future; |
---|
| 662 | sqe->flags = 0; |
---|
[a1f3d93] | 663 | sqe->fd = fd; |
---|
[6ddef36] | 664 | sqe->off = 0; |
---|
[d3605f8] | 665 | sqe->ioprio = 0; |
---|
[6ddef36] | 666 | sqe->fsync_flags = 0; |
---|
| 667 | sqe->__pad2[0] = 0; |
---|
| 668 | sqe->__pad2[1] = 0; |
---|
| 669 | sqe->__pad2[2] = 0; |
---|
[d3605f8] | 670 | |
---|
| 671 | #if defined(CFA_HAVE_IORING_OP_READ) |
---|
| 672 | sqe->opcode = IORING_OP_READ; |
---|
| 673 | sqe->addr = (uint64_t)iov.iov_base; |
---|
| 674 | sqe->len = iov.iov_len; |
---|
| 675 | #elif defined(CFA_HAVE_READV) && defined(CFA_HAVE_IORING_OP_READV) |
---|
| 676 | sqe->opcode = IORING_OP_READV; |
---|
| 677 | sqe->addr = (uintptr_t)&iov; |
---|
| 678 | sqe->len = 1; |
---|
| 679 | #else |
---|
| 680 | #error CFA_WITH_IO_URING_IDLE but none of CFA_HAVE_READV, CFA_HAVE_IORING_OP_READV or CFA_HAVE_IORING_OP_READ defined |
---|
| 681 | #endif |
---|
[6ddef36] | 682 | |
---|
| 683 | asm volatile("": : :"memory"); |
---|
| 684 | |
---|
| 685 | /* paranoid */ verify( sqe->user_data == (uintptr_t)&future ); |
---|
[010636f] | 686 | __submit_only( ctx, &idx, 1 ); |
---|
[6ddef36] | 687 | |
---|
| 688 | /* paranoid */ verify( proc == __cfaabi_tls.this_processor ); |
---|
| 689 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
[078fb05] | 690 | |
---|
| 691 | return true; |
---|
[6ddef36] | 692 | } |
---|
[18f7858] | 693 | |
---|
| 694 | void __cfa_io_idle( processor * proc ) { |
---|
| 695 | iovec iov; |
---|
| 696 | __atomic_acquire( &proc->io.ctx->cq.lock ); |
---|
| 697 | |
---|
[262fafd9] | 698 | __attribute__((used)) volatile bool was_reset = false; |
---|
| 699 | |
---|
[d5cdbed] | 700 | with( proc->idle_wctx) { |
---|
[18f7858] | 701 | |
---|
[37a3aa23] | 702 | // Do we already have a pending read |
---|
| 703 | if(available(*ftr)) { |
---|
| 704 | // There is no pending read, we need to add one |
---|
| 705 | reset(*ftr); |
---|
| 706 | |
---|
| 707 | iov.iov_base = rdbuf; |
---|
| 708 | iov.iov_len = sizeof(eventfd_t); |
---|
| 709 | __kernel_read(proc, *ftr, iov, evfd ); |
---|
[262fafd9] | 710 | ftr->result = 0xDEADDEAD; |
---|
| 711 | *((eventfd_t *)rdbuf) = 0xDEADDEADDEADDEAD; |
---|
| 712 | was_reset = true; |
---|
[37a3aa23] | 713 | } |
---|
[18f7858] | 714 | } |
---|
| 715 | |
---|
[010636f] | 716 | if( !__atomic_load_n( &proc->do_terminate, __ATOMIC_SEQ_CST ) ) { |
---|
| 717 | __ioarbiter_flush( *proc->io.ctx ); |
---|
[262fafd9] | 718 | proc->idle_wctx.sleep_time = rdtscl(); |
---|
[010636f] | 719 | ioring_syscsll( *proc->io.ctx, 1, IORING_ENTER_GETEVENTS); |
---|
| 720 | } |
---|
[18f7858] | 721 | |
---|
[d5cdbed] | 722 | ready_schedule_lock(); |
---|
[18f7858] | 723 | __cfa_do_drain( proc->io.ctx, proc->cltr ); |
---|
[d5cdbed] | 724 | ready_schedule_unlock(); |
---|
[262fafd9] | 725 | |
---|
| 726 | asm volatile ("" :: "m" (was_reset)); |
---|
[18f7858] | 727 | } |
---|
[6ddef36] | 728 | #endif |
---|
[47746a2] | 729 | #endif |
---|