| ... | @@ -0,0 +1,826 @@ |
| 1 | const builtin = @import("builtin"); |
| 2 | const std = @import("std"); |
| 3 | const assert = std.debug.assert; |
| 4 | const os = std.os; |
| 5 | const linux = os.linux; |
| 6 | const mem = std.mem; |
| 7 | const net = std.net; |
| 8 | const testing = std.testing; |
| 9 | |
| 10 | pub const io_uring_params = linux.io_uring_params; |
| 11 | pub const io_uring_cqe = linux.io_uring_cqe; |
| 12 | |
| 13 | // TODO Update linux.zig's definition of linux.io_uring_sqe: |
| 14 | // linux.io_uring_sqe uses numbered unions, i.e. `union1` etc. that are not future-proof and need to |
| 15 | // be re-numbered whenever new unions are interposed by the kernel. Furthermore, Zig's unions do not |
| 16 | // support assignment by any union member directly as in C, without going through the union, so the |
| 17 | // kernel adding new unions would also break existing Zig code. |
| 18 | // We therefore use a flat struct without unions to avoid these two issues. |
| 19 | // Pending https://github.com/ziglang/zig/issues/6349. |
| 20 | pub const io_uring_sqe = extern struct { |
| 21 | opcode: linux.IORING_OP, |
| 22 | flags: u8 = 0, |
| 23 | ioprio: u16 = 0, |
| 24 | fd: i32 = 0, |
| 25 | off: u64 = 0, |
| 26 | addr: u64 = 0, |
| 27 | len: u32 = 0, |
| 28 | opflags: u32 = 0, |
| 29 | user_data: u64 = 0, |
| 30 | buffer: u16 = 0, |
| 31 | personality: u16 = 0, |
| 32 | splice_fd_in: i32 = 0, |
| 33 | options: [2]u64 = [2]u64{ 0, 0 } |
| 34 | }; |
| 35 | |
| 36 | // TODO Add to zig/std/os/bits/linux.zig: |
| 37 | const IORING_SQ_CQ_OVERFLOW = 1 << 1; |
| 38 | |
| 39 | comptime { |
| 40 | assert(@sizeOf(io_uring_params) == 120); |
| 41 | assert(@sizeOf(io_uring_sqe) == 64); |
| 42 | assert(@sizeOf(io_uring_cqe) == 16); |
| 43 | |
| 44 | assert(linux.IORING_OFF_SQ_RING == 0); |
| 45 | assert(linux.IORING_OFF_CQ_RING == 0x8000000); |
| 46 | assert(linux.IORING_OFF_SQES == 0x10000000); |
| 47 | } |
| 48 | |
| 49 | pub const IO_Uring = struct { |
| 50 | fd: i32 = -1, |
| 51 | sq: SubmissionQueue, |
| 52 | cq: CompletionQueue, |
| 53 | flags: u32, |
| 54 | |
| 55 | /// A friendly way to setup an io_uring, with default io_uring_params. |
| 56 | /// `entries` must be a power of two between 1 and 4096, although the kernel will make the final |
| 57 | /// call on how many entries the submission and completion queues will ultimately have, |
| 58 | /// see https://github.com/torvalds/linux/blob/v5.8/fs/io_uring.c#L8027-L8050. |
| 59 | /// Matches the interface of io_uring_queue_init() in liburing. |
| 60 | pub fn init(entries: u32, flags: u32) !IO_Uring { |
| 61 | var params = io_uring_params { |
| 62 | .sq_entries = 0, |
| 63 | .cq_entries = 0, |
| 64 | .flags = flags, |
| 65 | .sq_thread_cpu = 0, |
| 66 | .sq_thread_idle = 1000, |
| 67 | .features = 0, |
| 68 | .wq_fd = 0, |
| 69 | .resv = [_]u32{0} ** 3, |
| 70 | .sq_off = undefined, |
| 71 | .cq_off = undefined, |
| 72 | }; |
| 73 | // The kernel will zero the memory of the sq_off and cq_off structs in io_uring_create(), |
| 74 | // see https://github.com/torvalds/linux/blob/v5.8/fs/io_uring.c#L7986-L8002. |
| 75 | return try IO_Uring.init_params(entries, &params); |
| 76 | } |
| 77 | |
| 78 | /// A powerful way to setup an io_uring, if you want to tweak io_uring_params such as submission |
| 79 | /// queue thread cpu affinity or thread idle timeout (the kernel and our default is 1 second). |
| 80 | /// `params` is passed by reference because the kernel needs to modify the parameters. |
| 81 | /// You may only set the `flags`, `sq_thread_cpu` and `sq_thread_idle` parameters. |
| 82 | /// Every other parameter belongs to the kernel and must be zeroed. |
| 83 | /// Matches the interface of io_uring_queue_init_params() in liburing. |
| 84 | pub fn init_params(entries: u32, p: *io_uring_params) !IO_Uring { |
| 85 | assert(entries >= 1 and entries <= 4096 and std.math.isPowerOfTwo(entries)); |
| 86 | assert(p.*.sq_entries == 0); |
| 87 | assert(p.*.cq_entries == 0); |
| 88 | assert(p.*.features == 0); |
| 89 | assert(p.*.wq_fd == 0); |
| 90 | assert(p.*.resv[0] == 0); |
| 91 | assert(p.*.resv[1] == 0); |
| 92 | assert(p.*.resv[2] == 0); |
| 93 | |
| 94 | const res = linux.io_uring_setup(entries, p); |
| 95 | try check_errno(res); |
| 96 | const fd = @intCast(i32, res); |
| 97 | assert(fd >= 0); |
| 98 | errdefer os.close(fd); |
| 99 | |
| 100 | // Kernel versions 5.4 and up use only one mmap() for the submission and completion queues. |
| 101 | // This is not an optional feature for us... if the kernel does it, we have to do it. |
| 102 | // The thinking on this by the kernel developers was that both the submission and the |
| 103 | // completion queue rings have sizes just over a power of two, but the submission queue ring |
| 104 | // is significantly smaller with u32 slots. By bundling both in a single mmap, the kernel |
| 105 | // gets the submission queue ring for free. |
| 106 | // See https://patchwork.kernel.org/patch/11115257 for the kernel patch. |
| 107 | // We do not support the double mmap() done before 5.4, because we want to keep the |
| 108 | // init/deinit mmap paths simple and because io_uring has had many bug fixes even since 5.4. |
| 109 | if ((p.*.features & linux.IORING_FEAT_SINGLE_MMAP) == 0) { |
| 110 | return error.IO_UringKernelNotSupported; |
| 111 | } |
| 112 | |
| 113 | // Check that the kernel has actually set params and that "impossible is nothing". |
| 114 | assert(p.*.sq_entries != 0); |
| 115 | assert(p.*.cq_entries != 0); |
| 116 | assert(p.*.cq_entries >= p.*.sq_entries); |
| 117 | |
| 118 | // From here on, we only need to read from params, so pass `p` by value for convenience. |
| 119 | // The completion queue shares the mmap with the submission queue, so pass `sq` there too. |
| 120 | var sq = try SubmissionQueue.init(fd, p.*); |
| 121 | errdefer sq.deinit(); |
| 122 | var cq = try CompletionQueue.init(fd, p.*, sq); |
| 123 | errdefer cq.deinit(); |
| 124 | |
| 125 | // Check that our starting state is as we expect. |
| 126 | assert(sq.head.* == 0); |
| 127 | assert(sq.tail.* == 0); |
| 128 | assert(sq.mask.* == p.*.sq_entries - 1); |
| 129 | // Allow flags.* to be non-zero, since the kernel may set IORING_SQ_NEED_WAKEUP at any time. |
| 130 | assert(sq.dropped.* == 0); |
| 131 | assert(sq.array.len == p.*.sq_entries); |
| 132 | assert(sq.sqes.len == p.*.sq_entries); |
| 133 | assert(sq.sqe_head == 0); |
| 134 | assert(sq.sqe_tail == 0); |
| 135 | |
| 136 | assert(cq.head.* == 0); |
| 137 | assert(cq.tail.* == 0); |
| 138 | assert(cq.mask.* == p.*.cq_entries - 1); |
| 139 | assert(cq.overflow.* == 0); |
| 140 | assert(cq.cqes.len == p.*.cq_entries); |
| 141 | |
| 142 | // Alles in Ordnung! |
| 143 | return IO_Uring { |
| 144 | .fd = fd, |
| 145 | .sq = sq, |
| 146 | .cq = cq, |
| 147 | .flags = p.*.flags |
| 148 | }; |
| 149 | } |
| 150 | |
| 151 | pub fn deinit(self: *IO_Uring) void { |
| 152 | assert(self.fd >= 0); |
| 153 | // The mmaps depend on the fd, so the order of these calls is important: |
| 154 | self.cq.deinit(); |
| 155 | self.sq.deinit(); |
| 156 | os.close(self.fd); |
| 157 | self.fd = -1; |
| 158 | } |
| 159 | |
| 160 | /// Returns a vacant SQE, or an error if the submission queue is full. |
| 161 | /// We follow the implementation (and atomics) of liburing's `io_uring_get_sqe()` exactly. |
| 162 | /// However, instead of a null we return an error to force safe handling. |
| 163 | /// Any situation where the submission queue is full tends more towards a control flow error, |
| 164 | /// and the null return in liburing is more a C idiom than anything else, for lack of a better |
| 165 | /// alternative. In Zig, we have first-class error handling... so let's use it. |
| 166 | /// Matches the implementation of io_uring_get_sqe() in liburing. |
| 167 | pub fn get_sqe(self: *IO_Uring) !*io_uring_sqe { |
| 168 | const head = @atomicLoad(u32, self.sq.head, .Acquire); |
| 169 | // Remember that these head and tail offsets wrap around every four billion operations. |
| 170 | // We must therefore use wrapping addition and subtraction to avoid a runtime crash. |
| 171 | const next = self.sq.sqe_tail +% 1; |
| 172 | if (next -% head > self.sq.sqes.len) return error.IO_UringSubmissionQueueFull; |
| 173 | var sqe = &self.sq.sqes[self.sq.sqe_tail & self.sq.mask.*]; |
| 174 | self.sq.sqe_tail = next; |
| 175 | return sqe; |
| 176 | } |
| 177 | |
| 178 | /// Submits the SQEs acquired via get_sqe() to the kernel. You can call this once after you have |
| 179 | /// called get_sqe() multiple times to setup multiple I/O requests. |
| 180 | /// Returns the number of SQEs submitted. |
| 181 | /// Matches the implementation of io_uring_submit() in liburing. |
| 182 | pub fn submit(self: *IO_Uring) !u32 { |
| 183 | return self.submit_and_wait(0); |
| 184 | } |
| 185 | |
| 186 | /// Like submit(), but allows waiting for events as well. |
| 187 | /// Returns the number of SQEs submitted. |
| 188 | /// Matches the implementation of io_uring_submit_and_wait() in liburing. |
| 189 | pub fn submit_and_wait(self: *IO_Uring, wait_nr: u32) !u32 { |
| 190 | var submitted = self.flush_sq(); |
| 191 | var flags: u32 = 0; |
| 192 | if (self.sq_ring_needs_enter(submitted, &flags) or wait_nr > 0) { |
| 193 | if (wait_nr > 0 or (self.flags & linux.IORING_SETUP_IOPOLL) > 0) { |
| 194 | flags |= linux.IORING_ENTER_GETEVENTS; |
| 195 | } |
| 196 | return try self.enter(submitted, wait_nr, flags); |
| 197 | } |
| 198 | return submitted; |
| 199 | } |
| 200 | |
| 201 | // Tell the kernel we have submitted SQEs and/or want to wait for CQEs. |
| 202 | // Returns the number of SQEs submitted. |
| 203 | fn enter(self: *IO_Uring, to_submit: u32, min_complete: u32, flags: u32) !u32 { |
| 204 | assert(self.fd >= 0); |
| 205 | const res = linux.io_uring_enter(self.fd, to_submit, min_complete, flags, null); |
| 206 | try check_errno(res); |
| 207 | return @truncate(u32, res); |
| 208 | } |
| 209 | |
| 210 | // Sync internal state with kernel ring state on the SQ side. |
| 211 | // Returns the number of all pending events in the SQ ring, for the shared ring. |
| 212 | // This return value includes previously flushed SQEs, as per liburing. |
| 213 | // The reasoning for this is to suggest that an io_uring_enter() call is needed rather than not. |
| 214 | // Matches the implementation of __io_uring_flush_sq() in liburing. |
| 215 | fn flush_sq(self: *IO_Uring) u32 { |
| 216 | if (self.sq.sqe_head != self.sq.sqe_tail) { |
| 217 | // Fill in SQEs that we have queued up, adding them to the kernel ring. |
| 218 | const to_submit = self.sq.sqe_tail -% self.sq.sqe_head; |
| 219 | const mask = self.sq.mask.*; |
| 220 | var tail = self.sq.tail.*; |
| 221 | var i: usize = 0; |
| 222 | while (i < to_submit) : (i += 1) { |
| 223 | self.sq.array[tail & mask] = self.sq.sqe_head & mask; |
| 224 | tail +%= 1; |
| 225 | self.sq.sqe_head +%= 1; |
| 226 | } |
| 227 | // Ensure that the kernel can actually see the SQE updates when it sees the tail update. |
| 228 | @atomicStore(u32, self.sq.tail, tail, .Release); |
| 229 | } |
| 230 | return self.sq_ready(); |
| 231 | } |
| 232 | |
| 233 | /// Returns true if we are not using an SQ thread (thus nobody submits but us), |
| 234 | /// or if IORING_SQ_NEED_WAKEUP is set and the SQ thread must be explicitly awakened. |
| 235 | /// For the latter case, we set the SQ thread wakeup flag. |
| 236 | /// Matches the implementation of sq_ring_needs_enter() in liburing. |
| 237 | fn sq_ring_needs_enter(self: *IO_Uring, submitted: u32, flags: *u32) bool { |
| 238 | assert(flags.* == 0); |
| 239 | if ((self.flags & linux.IORING_SETUP_SQPOLL) == 0 and submitted > 0) return true; |
| 240 | if ((@atomicLoad(u32, self.sq.flags, .Unordered) & linux.IORING_SQ_NEED_WAKEUP) > 0) { |
| 241 | flags.* |= linux.IORING_ENTER_SQ_WAKEUP; |
| 242 | return true; |
| 243 | } |
| 244 | return false; |
| 245 | } |
| 246 | |
| 247 | /// Returns the number of flushed and unflushed SQEs pending in the submission queue. |
| 248 | /// In other words, this is the number of SQEs in the submission queue, i.e. its length. |
| 249 | /// These are SQEs that the kernel is yet to consume. |
| 250 | /// Matches the implementation of io_uring_sq_ready in liburing. |
| 251 | pub fn sq_ready(self: *IO_Uring) u32 { |
| 252 | // Always use the shared ring state (i.e. head and not sqe_head) to avoid going out of sync, |
| 253 | // see https://github.com/axboe/liburing/issues/92. |
| 254 | return self.sq.sqe_tail -% @atomicLoad(u32, self.sq.head, .Acquire); |
| 255 | } |
| 256 | |
| 257 | /// Returns the number of CQEs in the completion queue, i.e. its length. |
| 258 | /// These are CQEs that the application is yet to consume. |
| 259 | /// Matches the implementation of io_uring_cq_ready in liburing. |
| 260 | pub fn cq_ready(self: *IO_Uring) u32 { |
| 261 | return @atomicLoad(u32, self.cq.tail, .Acquire) -% self.cq.head.*; |
| 262 | } |
| 263 | |
| 264 | /// Copies as many CQEs as are ready, and that can fit into the destination `cqes` slice. |
| 265 | /// If none are available, enters into the kernel to wait for at most `wait_nr` CQEs. |
| 266 | /// Returns the number of CQEs copied, advancing the CQ ring. |
| 267 | /// Provides all the wait/peek methods found in liburing, but with batching and a single method. |
| 268 | /// The rationale for copying CQEs rather than copying pointers is that pointers are 8 bytes |
| 269 | /// whereas CQEs are not much more at only 16 bytes, and this provides a safer faster interface. |
| 270 | /// Safer, because you no longer need to call cqe_seen(), avoiding idempotency bugs. |
| 271 | /// Faster, because we can now amortize the atomic store release to `cq.head` across the batch. |
| 272 | /// See https://github.com/axboe/liburing/issues/103#issuecomment-686665007. |
| 273 | /// Matches the implementation of io_uring_peek_batch_cqe() in liburing, but supports waiting. |
| 274 | pub fn copy_cqes(self: *IO_Uring, cqes: []io_uring_cqe, wait_nr: u32) !u32 { |
| 275 | const count = self.copy_cqes_ready(cqes, wait_nr); |
| 276 | if (count > 0) return count; |
| 277 | if (self.cq_ring_needs_flush() or wait_nr > 0) { |
| 278 | _ = try self.enter(0, wait_nr, linux.IORING_ENTER_GETEVENTS); |
| 279 | return self.copy_cqes_ready(cqes, wait_nr); |
| 280 | } |
| 281 | return 0; |
| 282 | } |
| 283 | |
| 284 | fn copy_cqes_ready(self: *IO_Uring, cqes: []io_uring_cqe, wait_nr: u32) u32 { |
| 285 | const ready = self.cq_ready(); |
| 286 | const count = std.math.min(cqes.len, ready); |
| 287 | const mask = self.cq.mask.*; |
| 288 | var head = self.cq.head.*; |
| 289 | var tail = head +% count; |
| 290 | // TODO Optimize this by using 1 or 2 memcpy's (if the tail wraps) rather than a loop. |
| 291 | var i: usize = 0; |
| 292 | // Do not use "less-than" operator since head and tail may wrap: |
| 293 | while (head != tail) { |
| 294 | cqes[i] = self.cq.cqes[head & mask]; // Copy struct by value. |
| 295 | head +%= 1; |
| 296 | i += 1; |
| 297 | } |
| 298 | self.cq_advance(count); |
| 299 | return count; |
| 300 | } |
| 301 | |
| 302 | /// Returns a copy of an I/O completion, waiting for it if necessary, and advancing the CQ ring. |
| 303 | /// A convenience method for `copy_cqes()` for when you don't need to batch or peek. |
| 304 | pub fn copy_cqe(ring: *IO_Uring) !io_uring_cqe { |
| 305 | var cqes: [1]io_uring_cqe = undefined; |
| 306 | const count = try ring.copy_cqes(&cqes, 1); |
| 307 | assert(count == 1); |
| 308 | return cqes[0]; |
| 309 | } |
| 310 | |
| 311 | // Matches the implementation of cq_ring_needs_flush() in liburing. |
| 312 | fn cq_ring_needs_flush(self: *IO_Uring) bool { |
| 313 | return (@atomicLoad(u32, self.sq.flags, .Unordered) & IORING_SQ_CQ_OVERFLOW) > 0; |
| 314 | } |
| 315 | |
| 316 | /// For advanced use cases only that implement custom completion queue methods. |
| 317 | /// If you use copy_cqes() or copy_cqe() you must not call cqe_seen() or cq_advance(). |
| 318 | /// Must be called exactly once after a zero-copy CQE has been processed by your application. |
| 319 | /// Not idempotent, calling more than once will result in other CQEs being lost. |
| 320 | /// Matches the implementation of cqe_seen() in liburing. |
| 321 | pub fn cqe_seen(self: *IO_Uring, cqe: *io_uring_cqe) void { |
| 322 | self.cq_advance(1); |
| 323 | } |
| 324 | |
| 325 | /// For advanced use cases only that implement custom completion queue methods. |
| 326 | /// Matches the implementation of cq_advance() in liburing. |
| 327 | pub fn cq_advance(self: *IO_Uring, count: u32) void { |
| 328 | if (count > 0) { |
| 329 | // Ensure the kernel only sees the new head value after the CQEs have been read. |
| 330 | @atomicStore(u32, self.cq.head, self.cq.head.* +% count, .Release); |
| 331 | } |
| 332 | } |
| 333 | |
| 334 | /// Queues (but does not submit) an SQE to perform an `accept4(2)` on a socket. |
| 335 | /// Returns a pointer to the SQE. |
| 336 | pub fn queue_accept( |
| 337 | self: *IO_Uring, |
| 338 | user_data: u64, |
| 339 | fd: os.fd_t, |
| 340 | addr: *os.sockaddr, |
| 341 | addrlen: *os.socklen_t, |
| 342 | accept_flags: u32 |
| 343 | ) !*io_uring_sqe { |
| 344 | // "sqe->fd is the file descriptor, sqe->addr holds a pointer to struct sockaddr, |
| 345 | // sqe->addr2 holds a pointer to socklen_t, and finally sqe->accept_flags holds the flags |
| 346 | // for accept(4)." - https://lwn.net/ml/linux-block/20191025173037.13486-1-axboe@kernel.dk/ |
| 347 | const sqe = try self.get_sqe(); |
| 348 | sqe.* = .{ |
| 349 | .opcode = .ACCEPT, |
| 350 | .fd = fd, |
| 351 | .off = @ptrToInt(addrlen), // `addr2` is a newer union member that maps to `off`. |
| 352 | .addr = @ptrToInt(addr), |
| 353 | .user_data = user_data, |
| 354 | .opflags = accept_flags |
| 355 | }; |
| 356 | return sqe; |
| 357 | } |
| 358 | |
| 359 | /// Queues (but does not submit) an SQE to perform an `fsync(2)`. |
| 360 | /// Returns a pointer to the SQE so that you can further modify the SQE for advanced use cases. |
| 361 | /// For example, for `fdatasync()` you can set `IORING_FSYNC_DATASYNC` in the SQE's `opflags`. |
| 362 | /// N.B. While SQEs are initiated in the order in which they appear in the submission queue, |
| 363 | /// operations execute in parallel and completions are unordered. Therefore, an application that |
| 364 | /// submits a write followed by an fsync in the submission queue cannot expect the fsync to |
| 365 | /// apply to the write, since the fsync may complete before the write is issued to the disk. |
| 366 | /// You should preferably use `link_with_next_sqe()` on a write's SQE to link it with an fsync, |
| 367 | /// or else insert a full write barrier using `drain_previous_sqes()` when queueing an fsync. |
| 368 | pub fn queue_fsync(self: *IO_Uring, user_data: u64, fd: os.fd_t) !*io_uring_sqe { |
| 369 | const sqe = try self.get_sqe(); |
| 370 | sqe.* = .{ |
| 371 | .opcode = .FSYNC, |
| 372 | .fd = fd, |
| 373 | .user_data = user_data |
| 374 | }; |
| 375 | return sqe; |
| 376 | } |
| 377 | |
| 378 | /// Queues (but does not submit) an SQE to perform a no-op. |
| 379 | /// Returns a pointer to the SQE so that you can further modify the SQE for advanced use cases. |
| 380 | /// A no-op is more useful than may appear at first glance. |
| 381 | /// For example, you could call `drain_previous_sqes()` on the returned SQE, to use the no-op to |
| 382 | /// know when the ring is idle before acting on a kill signal. |
| 383 | pub fn queue_nop(self: *IO_Uring, user_data: u64) !*io_uring_sqe { |
| 384 | const sqe = try self.get_sqe(); |
| 385 | sqe.* = .{ |
| 386 | .opcode = .NOP, |
| 387 | .user_data = user_data |
| 388 | }; |
| 389 | return sqe; |
| 390 | } |
| 391 | |
| 392 | /// Queues (but does not submit) an SQE to perform a `read(2)`. |
| 393 | /// Returns a pointer to the SQE. |
| 394 | pub fn queue_read( |
| 395 | self: *IO_Uring, |
| 396 | user_data: u64, |
| 397 | fd: os.fd_t, |
| 398 | buffer: []u8, |
| 399 | offset: u64 |
| 400 | ) !*io_uring_sqe { |
| 401 | const sqe = try self.get_sqe(); |
| 402 | sqe.* = .{ |
| 403 | .opcode = .READ, |
| 404 | .fd = fd, |
| 405 | .off = offset, |
| 406 | .addr = @ptrToInt(buffer.ptr), |
| 407 | .len = @truncate(u32, buffer.len), |
| 408 | .user_data = user_data |
| 409 | }; |
| 410 | return sqe; |
| 411 | } |
| 412 | |
| 413 | /// Queues (but does not submit) an SQE to perform a `write(2)`. |
| 414 | /// Returns a pointer to the SQE. |
| 415 | pub fn queue_write( |
| 416 | self: *IO_Uring, |
| 417 | user_data: u64, |
| 418 | fd: os.fd_t, |
| 419 | buffer: []const u8, |
| 420 | offset: u64 |
| 421 | ) !*io_uring_sqe { |
| 422 | const sqe = try self.get_sqe(); |
| 423 | sqe.* = .{ |
| 424 | .opcode = .WRITE, |
| 425 | .fd = fd, |
| 426 | .off = offset, |
| 427 | .addr = @ptrToInt(buffer.ptr), |
| 428 | .len = @truncate(u32, buffer.len), |
| 429 | .user_data = user_data |
| 430 | }; |
| 431 | return sqe; |
| 432 | } |
| 433 | |
| 434 | /// Queues (but does not submit) an SQE to perform a `preadv()`. |
| 435 | /// Returns a pointer to the SQE so that you can further modify the SQE for advanced use cases. |
| 436 | /// For example, if you want to do a `preadv2()` then set `opflags` on the returned SQE. |
| 437 | /// See https://linux.die.net/man/2/preadv. |
| 438 | pub fn queue_readv( |
| 439 | self: *IO_Uring, |
| 440 | user_data: u64, |
| 441 | fd: os.fd_t, |
| 442 | iovecs: []const os.iovec, |
| 443 | offset: u64 |
| 444 | ) !*io_uring_sqe { |
| 445 | const sqe = try self.get_sqe(); |
| 446 | sqe.* = .{ |
| 447 | .opcode = .READV, |
| 448 | .fd = fd, |
| 449 | .off = offset, |
| 450 | .addr = @ptrToInt(iovecs.ptr), |
| 451 | .len = @truncate(u32, iovecs.len), |
| 452 | .user_data = user_data |
| 453 | }; |
| 454 | return sqe; |
| 455 | } |
| 456 | |
| 457 | /// Queues (but does not submit) an SQE to perform a `pwritev()`. |
| 458 | /// Returns a pointer to the SQE so that you can further modify the SQE for advanced use cases. |
| 459 | /// For example, if you want to do a `pwritev2()` then set `opflags` on the returned SQE. |
| 460 | /// See https://linux.die.net/man/2/pwritev. |
| 461 | pub fn queue_writev( |
| 462 | self: *IO_Uring, |
| 463 | user_data: u64, |
| 464 | fd: os.fd_t, |
| 465 | iovecs: []const os.iovec_const, |
| 466 | offset: u64 |
| 467 | ) !*io_uring_sqe { |
| 468 | const sqe = try self.get_sqe(); |
| 469 | sqe.* = .{ |
| 470 | .opcode = .WRITEV, |
| 471 | .fd = fd, |
| 472 | .off = offset, |
| 473 | .addr = @ptrToInt(iovecs.ptr), |
| 474 | .len = @truncate(u32, iovecs.len), |
| 475 | .user_data = user_data |
| 476 | }; |
| 477 | return sqe; |
| 478 | } |
| 479 | |
| 480 | /// The next SQE will not be started until this one completes. |
| 481 | /// This can be used to chain causally dependent SQEs, and the chain can be arbitrarily long. |
| 482 | /// The tail of the chain is denoted by the first SQE that does not have this flag set. |
| 483 | /// This flag has no effect on previous SQEs, nor does it impact SQEs outside the chain. |
| 484 | /// This means that multiple chains can be executing in parallel, along with individual SQEs. |
| 485 | /// Only members inside the chain are serialized. |
| 486 | /// A chain will be broken if any SQE in the chain ends in error, where any unexpected result is |
| 487 | /// considered an error. For example, a short read will terminate the remainder of the chain. |
| 488 | pub fn link_with_next_sqe(self: *IO_Uring, sqe: *io_uring_sqe) void { |
| 489 | sqe.*.flags |= linux.IOSQE_IO_LINK; |
| 490 | } |
| 491 | |
| 492 | /// Like `link_with_next_sqe()` but stronger. |
| 493 | /// For when you don't want the chain to fail in the event of a completion result error. |
| 494 | /// For example, you may know that some commands will fail and may want the chain to continue. |
| 495 | /// Hard links are resilient to completion results, but are not resilient to submission errors. |
| 496 | pub fn hardlink_with_next_sqe(self: *IO_Uring, sqe: *io_uring_sqe) void { |
| 497 | sqe.*.flags |= linux.IOSQE_IO_HARDLINK; |
| 498 | } |
| 499 | |
| 500 | /// This creates a full pipeline barrier in the submission queue. |
| 501 | /// This SQE will not be started until previous SQEs complete. |
| 502 | /// Subsequent SQEs will not be started until this SQE completes. |
| 503 | /// In other words, this stalls the entire submission queue. |
| 504 | /// You should first consider using link_with_next_sqe() for more granular SQE sequence control. |
| 505 | pub fn drain_previous_sqes(self: *IO_Uring, sqe: *io_uring_sqe) void { |
| 506 | sqe.*.flags |= linux.IOSQE_IO_DRAIN; |
| 507 | } |
| 508 | |
| 509 | /// Registers an array of file descriptors. |
| 510 | /// Every time a file descriptor is put in an SQE and submitted to the kernel, the kernel must |
| 511 | /// retrieve a reference to the file, and once I/O has completed the file reference must be |
| 512 | /// dropped. The atomic nature of this file reference can be a slowdown for high IOPS workloads. |
| 513 | /// This slowdown can be avoided by pre-registering file descriptors. |
| 514 | /// To refer to a registered file descriptor, IOSQE_FIXED_FILE must be set in the SQE's flags, |
| 515 | /// and the SQE's fd must be set to the index of the file descriptor in the registered array. |
| 516 | /// Registering file descriptors will wait for the ring to idle. |
| 517 | /// Files are automatically unregistered by the kernel when the ring is torn down. |
| 518 | /// An application need unregister only if it wants to register a new array of file descriptors. |
| 519 | pub fn register_files(self: *IO_Uring, fds: []const i32) !void { |
| 520 | assert(self.fd >= 0); |
| 521 | const res = linux.io_uring_register( |
| 522 | self.fd, |
| 523 | .REGISTER_FILES, |
| 524 | fds.ptr, |
| 525 | @truncate(u32, fds.len) |
| 526 | ); |
| 527 | try check_errno(res); |
| 528 | } |
| 529 | |
| 530 | /// Changes the semantics of the SQE's `fd` to refer to a pre-registered file descriptor. |
| 531 | pub fn use_registered_fd(self: *IO_Uring, sqe: *io_uring_sqe) void { |
| 532 | sqe.*.flags |= linux.IOSQE_FIXED_FILE; |
| 533 | } |
| 534 | |
| 535 | /// Unregisters all registered file descriptors previously associated with the ring. |
| 536 | pub fn unregister_files(self: *IO_Uring) !void { |
| 537 | assert(self.fd >= 0); |
| 538 | const res = linux.io_uring_register(self.fd, .UNREGISTER_FILES, null, 0); |
| 539 | try check_errno(res); |
| 540 | } |
| 541 | }; |
| 542 | |
| 543 | |
| 544 | pub const SubmissionQueue = struct { |
| 545 | head: *u32, |
| 546 | tail: *u32, |
| 547 | mask: *u32, |
| 548 | flags: *u32, |
| 549 | dropped: *u32, |
| 550 | array: []u32, |
| 551 | sqes: []io_uring_sqe, |
| 552 | mmap: []align(std.mem.page_size) u8, |
| 553 | mmap_sqes: []align(std.mem.page_size) u8, |
| 554 | |
| 555 | // We use `sqe_head` and `sqe_tail` in the same way as liburing: |
| 556 | // We increment `sqe_tail` (but not `tail`) for each call to `get_sqe()`. |
| 557 | // We then set `tail` to `sqe_tail` once, only when these events are actually submitted. |
| 558 | // This allows us to amortize the cost of the @atomicStore to `tail` across multiple SQEs. |
| 559 | sqe_head: u32 = 0, |
| 560 | sqe_tail: u32 = 0, |
| 561 | |
| 562 | pub fn init(fd: i32, p: io_uring_params) !SubmissionQueue { |
| 563 | assert(fd >= 0); |
| 564 | assert((p.features & linux.IORING_FEAT_SINGLE_MMAP) > 0); |
| 565 | const size = std.math.max( |
| 566 | p.sq_off.array + p.sq_entries * @sizeOf(u32), |
| 567 | p.cq_off.cqes + p.cq_entries * @sizeOf(io_uring_cqe) |
| 568 | ); |
| 569 | const mmap = try os.mmap( |
| 570 | null, |
| 571 | size, |
| 572 | os.PROT_READ | os.PROT_WRITE, |
| 573 | os.MAP_SHARED | os.MAP_POPULATE, |
| 574 | fd, |
| 575 | linux.IORING_OFF_SQ_RING, |
| 576 | ); |
| 577 | errdefer os.munmap(mmap); |
| 578 | assert(mmap.len == size); |
| 579 | |
| 580 | // The motivation for the `sqes` and `array` indirection is to make it possible for the |
| 581 | // application to preallocate static io_uring_sqe entries and then replay them when needed. |
| 582 | const size_sqes = p.sq_entries * @sizeOf(io_uring_sqe); |
| 583 | const mmap_sqes = try os.mmap( |
| 584 | null, |
| 585 | size_sqes, |
| 586 | os.PROT_READ | os.PROT_WRITE, |
| 587 | os.MAP_SHARED | os.MAP_POPULATE, |
| 588 | fd, |
| 589 | linux.IORING_OFF_SQES, |
| 590 | ); |
| 591 | errdefer os.munmap(mmap_sqes); |
| 592 | assert(mmap_sqes.len == size_sqes); |
| 593 | |
| 594 | const array = @ptrCast([*]u32, @alignCast(@alignOf(u32), &mmap[p.sq_off.array])); |
| 595 | const sqes = @ptrCast([*]io_uring_sqe, @alignCast(@alignOf(io_uring_sqe), &mmap_sqes[0])); |
| 596 | // We expect the kernel copies p.sq_entries to the u32 pointed to by p.sq_off.ring_entries, |
| 597 | // see https://github.com/torvalds/linux/blob/v5.8/fs/io_uring.c#L7843-L7844. |
| 598 | assert( |
| 599 | p.sq_entries == |
| 600 | @ptrCast(*u32, @alignCast(@alignOf(u32), &mmap[p.sq_off.ring_entries])).* |
| 601 | ); |
| 602 | return SubmissionQueue { |
| 603 | .head = @ptrCast(*u32, @alignCast(@alignOf(u32), &mmap[p.sq_off.head])), |
| 604 | .tail = @ptrCast(*u32, @alignCast(@alignOf(u32), &mmap[p.sq_off.tail])), |
| 605 | .mask = @ptrCast(*u32, @alignCast(@alignOf(u32), &mmap[p.sq_off.ring_mask])), |
| 606 | .flags = @ptrCast(*u32, @alignCast(@alignOf(u32), &mmap[p.sq_off.flags])), |
| 607 | .dropped = @ptrCast(*u32, @alignCast(@alignOf(u32), &mmap[p.sq_off.dropped])), |
| 608 | .array = array[0..p.sq_entries], |
| 609 | .sqes = sqes[0..p.sq_entries], |
| 610 | .mmap = mmap, |
| 611 | .mmap_sqes = mmap_sqes |
| 612 | }; |
| 613 | } |
| 614 | |
| 615 | pub fn deinit(self: *SubmissionQueue) void { |
| 616 | os.munmap(self.mmap_sqes); |
| 617 | os.munmap(self.mmap); |
| 618 | } |
| 619 | }; |
| 620 | |
| 621 | pub const CompletionQueue = struct { |
| 622 | head: *u32, |
| 623 | tail: *u32, |
| 624 | mask: *u32, |
| 625 | overflow: *u32, |
| 626 | cqes: []io_uring_cqe, |
| 627 | |
| 628 | pub fn init(fd: i32, p: io_uring_params, sq: SubmissionQueue) !CompletionQueue { |
| 629 | assert(fd >= 0); |
| 630 | assert((p.features & linux.IORING_FEAT_SINGLE_MMAP) > 0); |
| 631 | const mmap = sq.mmap; |
| 632 | const cqes = @ptrCast( |
| 633 | [*]io_uring_cqe, |
| 634 | @alignCast(@alignOf(io_uring_cqe), &mmap[p.cq_off.cqes]) |
| 635 | ); |
| 636 | assert( |
| 637 | p.cq_entries == |
| 638 | @ptrCast(*u32, @alignCast(@alignOf(u32), &mmap[p.cq_off.ring_entries])).* |
| 639 | ); |
| 640 | return CompletionQueue { |
| 641 | .head = @ptrCast(*u32, @alignCast(@alignOf(u32), &mmap[p.cq_off.head])), |
| 642 | .tail = @ptrCast(*u32, @alignCast(@alignOf(u32), &mmap[p.cq_off.tail])), |
| 643 | .mask = @ptrCast(*u32, @alignCast(@alignOf(u32), &mmap[p.cq_off.ring_mask])), |
| 644 | .overflow = @ptrCast(*u32, @alignCast(@alignOf(u32), &mmap[p.cq_off.overflow])), |
| 645 | .cqes = cqes[0..p.cq_entries] |
| 646 | }; |
| 647 | } |
| 648 | |
| 649 | pub fn deinit(self: *CompletionQueue) void { |
| 650 | // A no-op since we now share the mmap with the submission queue. |
| 651 | // Here for symmetry with the submission queue, and for any future feature support. |
| 652 | } |
| 653 | }; |
| 654 | |
| 655 | inline fn check_errno(res: usize) !void { |
| 656 | const errno = linux.getErrno(res); |
| 657 | if (errno != 0) return os.unexpectedErrno(errno); |
| 658 | } |
| 659 | |
| 660 | test "queue_nop" { |
| 661 | if (builtin.os.tag != .linux) return error.SkipZigTest; |
| 662 | |
| 663 | var ring = try IO_Uring.init(1, 0); |
| 664 | defer { |
| 665 | ring.deinit(); |
| 666 | testing.expectEqual(@as(i32, -1), ring.fd); |
| 667 | } |
| 668 | |
| 669 | var sqe = try ring.queue_nop(@intCast(u64, 0xaaaaaaaa)); |
| 670 | testing.expectEqual(io_uring_sqe { |
| 671 | .opcode = .NOP, |
| 672 | .flags = 0, |
| 673 | .ioprio = 0, |
| 674 | .fd = 0, |
| 675 | .off = 0, |
| 676 | .addr = 0, |
| 677 | .len = 0, |
| 678 | .opflags = 0, |
| 679 | .user_data = @intCast(u64, 0xaaaaaaaa), |
| 680 | .buffer = 0, |
| 681 | .personality = 0, |
| 682 | .splice_fd_in = 0, |
| 683 | .options = [2]u64{ 0, 0 } |
| 684 | }, sqe.*); |
| 685 | |
| 686 | testing.expectEqual(@as(u32, 0), ring.sq.sqe_head); |
| 687 | testing.expectEqual(@as(u32, 1), ring.sq.sqe_tail); |
| 688 | testing.expectEqual(@as(u32, 0), ring.sq.tail.*); |
| 689 | testing.expectEqual(@as(u32, 0), ring.cq.head.*); |
| 690 | testing.expectEqual(@as(u32, 1), ring.sq_ready()); |
| 691 | testing.expectEqual(@as(u32, 0), ring.cq_ready()); |
| 692 | |
| 693 | testing.expectEqual(@as(u32, 1), try ring.submit()); |
| 694 | testing.expectEqual(@as(u32, 1), ring.sq.sqe_head); |
| 695 | testing.expectEqual(@as(u32, 1), ring.sq.sqe_tail); |
| 696 | testing.expectEqual(@as(u32, 1), ring.sq.tail.*); |
| 697 | testing.expectEqual(@as(u32, 0), ring.cq.head.*); |
| 698 | testing.expectEqual(@as(u32, 0), ring.sq_ready()); |
| 699 | |
| 700 | testing.expectEqual(io_uring_cqe { |
| 701 | .user_data = 0xaaaaaaaa, |
| 702 | .res = 0, |
| 703 | .flags = 0 |
| 704 | }, try ring.copy_cqe()); |
| 705 | testing.expectEqual(@as(u32, 1), ring.cq.head.*); |
| 706 | testing.expectEqual(@as(u32, 0), ring.cq_ready()); |
| 707 | |
| 708 | var sqe_barrier = try ring.queue_nop(@intCast(u64, 0xbbbbbbbb)); |
| 709 | ring.drain_previous_sqes(sqe_barrier); |
| 710 | testing.expectEqual(@as(u8, linux.IOSQE_IO_DRAIN), sqe_barrier.*.flags); |
| 711 | testing.expectEqual(@as(u32, 1), try ring.submit()); |
| 712 | testing.expectEqual(io_uring_cqe { |
| 713 | .user_data = 0xbbbbbbbb, |
| 714 | .res = 0, |
| 715 | .flags = 0 |
| 716 | }, try ring.copy_cqe()); |
| 717 | testing.expectEqual(@as(u32, 2), ring.sq.sqe_head); |
| 718 | testing.expectEqual(@as(u32, 2), ring.sq.sqe_tail); |
| 719 | testing.expectEqual(@as(u32, 2), ring.sq.tail.*); |
| 720 | testing.expectEqual(@as(u32, 2), ring.cq.head.*); |
| 721 | } |
| 722 | |
| 723 | test "queue_readv" { |
| 724 | if (builtin.os.tag != .linux) return error.SkipZigTest; |
| 725 | |
| 726 | var ring = try IO_Uring.init(1, 0); |
| 727 | defer ring.deinit(); |
| 728 | |
| 729 | const fd = try os.openZ("/dev/zero", os.O_RDONLY | os.O_CLOEXEC, 0); |
| 730 | defer os.close(fd); |
| 731 | |
| 732 | var registered_fds = [_]i32{-1} ** 10; |
| 733 | const fd_index = 9; |
| 734 | registered_fds[fd_index] = fd; |
| 735 | try ring.register_files(registered_fds[0..]); |
| 736 | |
| 737 | var buffer = [_]u8{42} ** 128; |
| 738 | var iovecs = [_]os.iovec{ os.iovec { .iov_base = &buffer, .iov_len = buffer.len } }; |
| 739 | var sqe = try ring.queue_readv(0xcccccccc, fd_index, iovecs[0..], 0); |
| 740 | ring.use_registered_fd(sqe); |
| 741 | testing.expectEqual(@as(u8, linux.IOSQE_FIXED_FILE), sqe.*.flags); |
| 742 | |
| 743 | testing.expectError(error.IO_UringSubmissionQueueFull, ring.queue_nop(0)); |
| 744 | testing.expectEqual(@as(u32, 1), try ring.submit()); |
| 745 | testing.expectEqual(linux.io_uring_cqe { |
| 746 | .user_data = 0xcccccccc, |
| 747 | .res = buffer.len, |
| 748 | .flags = 0, |
| 749 | }, try ring.copy_cqe()); |
| 750 | testing.expectEqualSlices(u8, &([_]u8{0} ** buffer.len), buffer[0..]); |
| 751 | |
| 752 | try ring.unregister_files(); |
| 753 | } |
| 754 | |
| 755 | test "queue_writev/queue_fsync" { |
| 756 | if (builtin.os.tag != .linux) return error.SkipZigTest; |
| 757 | |
| 758 | var ring = try IO_Uring.init(2, 0); |
| 759 | defer ring.deinit(); |
| 760 | |
| 761 | const path = "test_io_uring_queue_writev"; |
| 762 | const file = try std.fs.cwd().createFile(path, .{ .truncate = true }); |
| 763 | defer file.close(); |
| 764 | defer std.fs.cwd().deleteFile(path) catch {}; |
| 765 | const fd = file.handle; |
| 766 | |
| 767 | var buffer = [_]u8{42} ** 128; |
| 768 | var iovecs = [_]os.iovec_const { |
| 769 | os.iovec_const { .iov_base = &buffer, .iov_len = buffer.len } |
| 770 | }; |
| 771 | var sqe_writev = try ring.queue_writev(0xdddddddd, fd, iovecs[0..], 0); |
| 772 | ring.link_with_next_sqe(sqe_writev); |
| 773 | testing.expectEqual(@as(u8, linux.IOSQE_IO_LINK), sqe_writev.*.flags); |
| 774 | |
| 775 | var sqe_fsync = try ring.queue_fsync(0xeeeeeeee, fd); |
| 776 | testing.expectEqual(fd, sqe_fsync.*.fd); |
| 777 | |
| 778 | testing.expectEqual(@as(u32, 2), ring.sq_ready()); |
| 779 | testing.expectEqual(@as(u32, 2), try ring.submit_and_wait(2)); |
| 780 | testing.expectEqual(@as(u32, 0), ring.sq_ready()); |
| 781 | testing.expectEqual(@as(u32, 2), ring.cq_ready()); |
| 782 | testing.expectEqual(linux.io_uring_cqe { |
| 783 | .user_data = 0xdddddddd, |
| 784 | .res = buffer.len, |
| 785 | .flags = 0, |
| 786 | }, try ring.copy_cqe()); |
| 787 | testing.expectEqual(@as(u32, 1), ring.cq_ready()); |
| 788 | testing.expectEqual(linux.io_uring_cqe { |
| 789 | .user_data = 0xeeeeeeee, |
| 790 | .res = 0, |
| 791 | .flags = 0, |
| 792 | }, try ring.copy_cqe()); |
| 793 | testing.expectEqual(@as(u32, 0), ring.cq_ready()); |
| 794 | } |
| 795 | |
| 796 | test "queue_write/queue_read" { |
| 797 | if (builtin.os.tag != .linux) return error.SkipZigTest; |
| 798 | // This test may require newer kernel versions. |
| 799 | |
| 800 | var ring = try IO_Uring.init(2, 0); |
| 801 | defer ring.deinit(); |
| 802 | |
| 803 | const path = "test_io_uring_queue_write"; |
| 804 | const file = try std.fs.cwd().createFile(path, .{ .read = true, .truncate = true }); |
| 805 | defer file.close(); |
| 806 | defer std.fs.cwd().deleteFile(path) catch {}; |
| 807 | const fd = file.handle; |
| 808 | |
| 809 | var buffer_write = [_]u8{97} ** 20; |
| 810 | var buffer_read = [_]u8{98} ** 20; |
| 811 | var sqe_write = try ring.queue_write(123, fd, buffer_write[0..], 10); |
| 812 | ring.link_with_next_sqe(sqe_write); |
| 813 | var sqe_read = try ring.queue_read(456, fd, buffer_read[0..], 10); |
| 814 | testing.expectEqual(@as(u32, 2), try ring.submit()); |
| 815 | testing.expectEqual(linux.io_uring_cqe { |
| 816 | .user_data = 123, |
| 817 | .res = buffer_write.len, |
| 818 | .flags = 0, |
| 819 | }, try ring.copy_cqe()); |
| 820 | testing.expectEqual(linux.io_uring_cqe { |
| 821 | .user_data = 456, |
| 822 | .res = buffer_read.len, |
| 823 | .flags = 0, |
| 824 | }, try ring.copy_cqe()); |
| 825 | testing.expectEqualSlices(u8, buffer_write[0..], buffer_read[0..]); |
| 826 | } |