|
| 1 | +# distutils: language = c++ |
| 2 | +# cython: language_level=3 |
| 3 | + |
| 4 | +from cpython.mem cimport PyMem_Malloc, PyMem_Free |
| 5 | +from libc.stdint cimport uint64_t, uintptr_t |
| 6 | +from libc.string cimport memset |
| 7 | + |
| 8 | +# IMPORTANT: cimport the module by its full package path |
| 9 | +cimport opteryx.compiled.io.iouring as C |
| 10 | + |
| 11 | +cdef int _check_errno(int rc): |
| 12 | + if rc < 0: |
| 13 | + raise OSError(-rc, "io_uring error") |
| 14 | + return rc |
| 15 | + |
| 16 | +cdef class BufferPool: |
| 17 | + cdef void **ptrs |
| 18 | + cdef C.iovec *iov |
| 19 | + cdef size_t nbuf |
| 20 | + cdef size_t buf_size |
| 21 | + cdef size_t alignment |
| 22 | + |
| 23 | + def __cinit__(self, size_t nbuf, size_t buf_size, size_t alignment=4096): |
| 24 | + self.nbuf = nbuf |
| 25 | + self.buf_size = buf_size |
| 26 | + self.alignment = alignment |
| 27 | + self.ptrs = <void **>PyMem_Malloc(nbuf * sizeof(void *)) |
| 28 | + if self.ptrs == NULL: |
| 29 | + raise MemoryError("alloc ptrs") |
| 30 | + self.iov = <C.iovec *>PyMem_Malloc(nbuf * sizeof(C.iovec)) |
| 31 | + if self.iov == NULL: |
| 32 | + PyMem_Free(self.ptrs) |
| 33 | + raise MemoryError("alloc iov") |
| 34 | + for i in range(nbuf): |
| 35 | + self.ptrs[i] = NULL |
| 36 | + |
| 37 | + cdef void *p |
| 38 | + for i in range(nbuf): |
| 39 | + if C.posix_memalign(&p, alignment, buf_size) != 0: |
| 40 | + self._cleanup(i) |
| 41 | + raise MemoryError(f"posix_memalign failed for buffer {i}") |
| 42 | + memset(p, 0, buf_size) |
| 43 | + self.ptrs[i] = p |
| 44 | + self.iov[i].iov_base = p |
| 45 | + self.iov[i].iov_len = buf_size |
| 46 | + |
| 47 | + cdef void _cleanup(self, size_t upto): |
| 48 | + cdef size_t j |
| 49 | + for j in range(upto): |
| 50 | + if self.ptrs[j] != NULL: |
| 51 | + C.free(self.ptrs[j]) |
| 52 | + if self.iov != NULL: |
| 53 | + PyMem_Free(self.iov) |
| 54 | + if self.ptrs != NULL: |
| 55 | + PyMem_Free(self.ptrs) |
| 56 | + |
| 57 | + def __dealloc__(self): |
| 58 | + self._cleanup(self.nbuf) |
| 59 | + |
| 60 | + property n: |
| 61 | + def __get__(self): |
| 62 | + return self.nbuf |
| 63 | + |
| 64 | + property size: |
| 65 | + def __get__(self): |
| 66 | + return self.buf_size |
| 67 | + |
| 68 | + def addr(self, size_t idx) -> int: |
| 69 | + if idx >= self.nbuf: |
| 70 | + raise IndexError |
| 71 | + return <uintptr_t>self.ptrs[idx] |
| 72 | + |
| 73 | + def view(self, size_t idx, Py_ssize_t length): |
| 74 | + if idx >= self.nbuf or length > self.buf_size: |
| 75 | + raise IndexError |
| 76 | + cdef unsigned char[:] mv = <unsigned char[:length]> self.ptrs[idx] |
| 77 | + return mv |
| 78 | + |
| 79 | + |
| 80 | +cdef class Uring: |
| 81 | + cdef C.io_uring ring |
| 82 | + cdef BufferPool pool |
| 83 | + cdef bint buffers_registered |
| 84 | + |
| 85 | + def __cinit__(self, unsigned entries=4096, unsigned flags=0): |
| 86 | + if flags == 0: |
| 87 | + flags = C.IORING_SETUP_CLAMP | C.IORING_SETUP_COOP_TASKRUN | C.IORING_SETUP_SINGLE_ISSUER |
| 88 | + _check_errno(C.io_uring_queue_init(entries, &self.ring, flags)) |
| 89 | + self.pool = None |
| 90 | + self.buffers_registered = False |
| 91 | + |
| 92 | + def __dealloc__(self): |
| 93 | + try: |
| 94 | + if self.buffers_registered: |
| 95 | + C.io_uring_unregister_buffers(&self.ring) |
| 96 | + except Exception: |
| 97 | + pass |
| 98 | + C.io_uring_queue_exit(&self.ring) |
| 99 | + |
| 100 | + def register_buffers(self, BufferPool pool): |
| 101 | + if pool is None: |
| 102 | + raise ValueError("pool is None") |
| 103 | + _check_errno(C.io_uring_register_buffers(&self.ring, pool.iov, <unsigned>pool.nbuf)) |
| 104 | + self.pool = pool |
| 105 | + self.buffers_registered = True |
| 106 | + |
| 107 | + def submit_read_fixed(self, int fd, size_t buf_index, size_t nbytes, long long offset, uint64_t user_data=0): |
| 108 | + if not self.buffers_registered: |
| 109 | + raise RuntimeError("buffers not registered") |
| 110 | + if buf_index >= self.pool.nbuf: |
| 111 | + raise IndexError |
| 112 | + if nbytes > self.pool.buf_size: |
| 113 | + raise ValueError("nbytes > buffer size") |
| 114 | + |
| 115 | + cdef C.io_uring_sqe* sqe = C.io_uring_get_sqe(&self.ring) |
| 116 | + if sqe == NULL: |
| 117 | + raise RuntimeError("no available SQE (ring full)") |
| 118 | + |
| 119 | + C.io_uring_prep_read_fixed(sqe, fd, self.pool.ptrs[buf_index], |
| 120 | + <unsigned>nbytes, offset, <int>buf_index) |
| 121 | + # Use helper instead of touching struct fields |
| 122 | + C.io_uring_sqe_set_data64(sqe, user_data) |
| 123 | + |
| 124 | + def submit(self) -> int: |
| 125 | + return _check_errno(C.io_uring_submit(&self.ring)) |
| 126 | + |
| 127 | + def wait_cqe(self): |
| 128 | + cdef C.io_uring_cqe* cqe |
| 129 | + _check_errno(C.io_uring_wait_cqe(&self.ring, &cqe)) |
| 130 | + res = cqe.res |
| 131 | + ud = C.io_uring_cqe_get_data64(cqe) |
| 132 | + C.io_uring_cqe_seen(&self.ring, cqe) |
| 133 | + return res, ud |
| 134 | + |
| 135 | + def peek_cqe(self): |
| 136 | + cdef C.io_uring_cqe* cqe |
| 137 | + rc = C.io_uring_peek_cqe(&self.ring, &cqe) |
| 138 | + if rc == 0 and cqe != NULL: |
| 139 | + res = cqe.res |
| 140 | + ud = C.io_uring_cqe_get_data64(cqe) |
| 141 | + C.io_uring_cqe_seen(&self.ring, cqe) |
| 142 | + return res, ud |
| 143 | + return None |
| 144 | + |
| 145 | + |
| 146 | +def open_direct(path: bytes) -> int: |
| 147 | + """Open O_RDONLY|O_DIRECT|O_CLOEXEC. Caller must close(fd).""" |
| 148 | + cdef int fd = C.open(<const char*>path, C.O_RDONLY | C.O_DIRECT | C.O_CLOEXEC) |
| 149 | + if fd < 0: |
| 150 | + raise OSError(C.errno, "open(O_DIRECT) failed") |
| 151 | + return fd |
| 152 | + |
| 153 | + |
| 154 | +def close_fd(int fd): |
| 155 | + if C.close(fd) != 0: |
| 156 | + raise OSError(C.errno, "close failed") |
0 commit comments