|
| 1 | +//! Reference https://github.com/asterinas/asterinas/blob/main/kernel/src/syscall/select.rs |
| 2 | +use alloc::vec::Vec; |
| 3 | +use system_error::SystemError; |
| 4 | + |
| 5 | +#[cfg(target_arch = "x86_64")] |
| 6 | +use crate::arch::syscall::nr::SYS_SELECT; |
| 7 | +use crate::{ |
| 8 | + filesystem::{ |
| 9 | + epoll::EPollEventType, |
| 10 | + poll::{do_sys_poll, poll_select_set_timeout, PollFd}, |
| 11 | + }, |
| 12 | + syscall::{ |
| 13 | + table::{FormattedSyscallParam, Syscall}, |
| 14 | + user_access::{UserBufferReader, UserBufferWriter}, |
| 15 | + }, |
| 16 | + time::{syscall::PosixTimeval, Instant}, |
| 17 | +}; |
| 18 | +// Maximum number of file descriptors in a set |
| 19 | +const FD_SETSIZE: usize = 1024; |
| 20 | +const USIZE_BITS: usize = core::mem::size_of::<usize>() * 8; |
| 21 | +/// See https://man7.org/linux/man-pages/man2/select.2.html |
| 22 | +pub struct SysSelect; |
| 23 | + |
| 24 | +impl Syscall for SysSelect { |
| 25 | + fn num_args(&self) -> usize { |
| 26 | + 5 |
| 27 | + } |
| 28 | + |
| 29 | + fn handle( |
| 30 | + &self, |
| 31 | + args: &[usize], |
| 32 | + _frame: &mut crate::arch::interrupt::TrapFrame, |
| 33 | + ) -> Result<usize, SystemError> { |
| 34 | + common_sys_select(args[0], args[1], args[2], args[3], args[4]) |
| 35 | + } |
| 36 | + |
| 37 | + fn entry_format(&self, args: &[usize]) -> Vec<crate::syscall::table::FormattedSyscallParam> { |
| 38 | + vec![ |
| 39 | + FormattedSyscallParam::new("nfds", format!("{}", args[0])), |
| 40 | + FormattedSyscallParam::new("readfds", format!("{:#x}", args[1])), |
| 41 | + FormattedSyscallParam::new("writefds", format!("{:#x}", args[2])), |
| 42 | + FormattedSyscallParam::new("exceptfds", format!("{:#x}", args[3])), |
| 43 | + FormattedSyscallParam::new("timeout", format!("{:#x}", args[4])), |
| 44 | + ] |
| 45 | + } |
| 46 | +} |
| 47 | + |
| 48 | +pub fn common_sys_select( |
| 49 | + nfds: usize, |
| 50 | + readfds_addr: usize, |
| 51 | + writefds_addr: usize, |
| 52 | + exceptfds_addr: usize, |
| 53 | + timeout_ptr: usize, |
| 54 | +) -> Result<usize, SystemError> { |
| 55 | + // log::debug!( |
| 56 | + // "common_sys_select called with nfds = {}, readfds_addr = {:#x}, writefds_addr = {:#x}, exceptfds_addr = {:#x}, timeout_ptr = {:#x}", |
| 57 | + // nfds, readfds_addr, writefds_addr, exceptfds_addr, timeout_ptr |
| 58 | + // ); |
| 59 | + let mut timeout: Option<Instant> = None; |
| 60 | + if timeout_ptr != 0 { |
| 61 | + let tsreader = UserBufferReader::new( |
| 62 | + timeout_ptr as *const PosixTimeval, |
| 63 | + size_of::<PosixTimeval>(), |
| 64 | + true, |
| 65 | + )?; |
| 66 | + let ts = *tsreader.read_one_from_user::<PosixTimeval>(0)?; |
| 67 | + let timeout_ms = ts.tv_sec * 1000 + ts.tv_usec as i64 / 1000; |
| 68 | + if timeout_ms >= 0 { |
| 69 | + timeout = poll_select_set_timeout(timeout_ms as u64); |
| 70 | + } |
| 71 | + } |
| 72 | + do_sys_select( |
| 73 | + nfds as isize, |
| 74 | + readfds_addr as *const FdSet, |
| 75 | + writefds_addr as *const FdSet, |
| 76 | + exceptfds_addr as *const FdSet, |
| 77 | + timeout, |
| 78 | + ) |
| 79 | +} |
| 80 | + |
| 81 | +fn do_sys_select( |
| 82 | + nfds: isize, |
| 83 | + readfds_addr: *const FdSet, |
| 84 | + writefds_addr: *const FdSet, |
| 85 | + exceptfds_addr: *const FdSet, |
| 86 | + timeout: Option<Instant>, |
| 87 | +) -> Result<usize, SystemError> { |
| 88 | + if nfds < 0 || nfds as usize > FD_SETSIZE { |
| 89 | + return Err(SystemError::EINVAL); |
| 90 | + } |
| 91 | + let get_fdset = |fdset_addr: *const FdSet| -> Result<Option<FdSet>, SystemError> { |
| 92 | + let fdset = if fdset_addr.is_null() { |
| 93 | + None |
| 94 | + } else { |
| 95 | + let fdset_buf = UserBufferReader::new(fdset_addr, size_of::<FdSet>(), true)?; |
| 96 | + let fdset = *fdset_buf.read_one_from_user::<FdSet>(0)?; |
| 97 | + Some(fdset) |
| 98 | + }; |
| 99 | + Ok(fdset) |
| 100 | + }; |
| 101 | + let mut readfds = get_fdset(readfds_addr)?; |
| 102 | + let mut writefds = get_fdset(writefds_addr)?; |
| 103 | + let mut exceptfds = get_fdset(exceptfds_addr)?; |
| 104 | + |
| 105 | + // log::debug!( |
| 106 | + // "nfds = {}, readfds = {:?}, writefds = {:?}, exceptfds = {:?}, timeout = {:?}", |
| 107 | + // nfds, |
| 108 | + // readfds, |
| 109 | + // writefds, |
| 110 | + // exceptfds, |
| 111 | + // timeout |
| 112 | + // ); |
| 113 | + |
| 114 | + let num_revents = do_select( |
| 115 | + nfds as usize, |
| 116 | + readfds.as_mut(), |
| 117 | + writefds.as_mut(), |
| 118 | + exceptfds.as_mut(), |
| 119 | + timeout, |
| 120 | + )?; |
| 121 | + |
| 122 | + let set_fdset = |fdset_addr: *const FdSet, fdset: Option<FdSet>| -> Result<(), SystemError> { |
| 123 | + if let Some(fdset) = fdset { |
| 124 | + let mut fdset_buf = |
| 125 | + UserBufferWriter::new(fdset_addr as *mut FdSet, size_of::<FdSet>(), true)?; |
| 126 | + fdset_buf.copy_one_to_user(&fdset, 0)?; |
| 127 | + } |
| 128 | + Ok(()) |
| 129 | + }; |
| 130 | + |
| 131 | + set_fdset(readfds_addr, readfds)?; |
| 132 | + set_fdset(writefds_addr, writefds)?; |
| 133 | + set_fdset(exceptfds_addr, exceptfds)?; |
| 134 | + |
| 135 | + // log::info!("num_revents = {}", num_revents); |
| 136 | + Ok(num_revents) |
| 137 | +} |
| 138 | + |
| 139 | +fn do_select( |
| 140 | + nfds: usize, |
| 141 | + mut readfds: Option<&mut FdSet>, |
| 142 | + mut writefds: Option<&mut FdSet>, |
| 143 | + mut exceptfds: Option<&mut FdSet>, |
| 144 | + timeout: Option<Instant>, |
| 145 | +) -> Result<usize, SystemError> { |
| 146 | + let mut poll_fds = { |
| 147 | + let mut poll_fds = Vec::with_capacity(nfds); |
| 148 | + for fd in 0..nfds { |
| 149 | + let events = { |
| 150 | + let readable = readfds.as_ref().is_some_and(|fds| fds.is_set(fd)); |
| 151 | + let writable = writefds.as_ref().is_some_and(|fds| fds.is_set(fd)); |
| 152 | + let except = exceptfds.as_ref().is_some_and(|fds| fds.is_set(fd)); |
| 153 | + convert_rwe_to_events(readable, writable, except) |
| 154 | + }; |
| 155 | + |
| 156 | + if events.is_empty() { |
| 157 | + continue; |
| 158 | + } |
| 159 | + |
| 160 | + let poll_fd = PollFd { |
| 161 | + fd: fd as i32, |
| 162 | + events: events.bits() as _, |
| 163 | + revents: 0, |
| 164 | + }; |
| 165 | + poll_fds.push(poll_fd); |
| 166 | + } |
| 167 | + poll_fds |
| 168 | + }; |
| 169 | + if let Some(fds) = readfds.as_mut() { |
| 170 | + fds.clear(); |
| 171 | + } |
| 172 | + if let Some(fds) = writefds.as_mut() { |
| 173 | + fds.clear(); |
| 174 | + } |
| 175 | + if let Some(fds) = exceptfds.as_mut() { |
| 176 | + fds.clear(); |
| 177 | + } |
| 178 | + |
| 179 | + // call the underlying poll syscall |
| 180 | + let num_revents = do_sys_poll(&mut poll_fds, timeout)?; |
| 181 | + if num_revents == 0 { |
| 182 | + return Ok(0); |
| 183 | + } |
| 184 | + |
| 185 | + let mut total_revents = 0; |
| 186 | + for poll_fd in &poll_fds { |
| 187 | + let fd = poll_fd.fd as usize; |
| 188 | + let revents = poll_fd.revents; |
| 189 | + let revents = EPollEventType::from_bits_truncate(revents as u32); |
| 190 | + let (readable, writable, except) = convert_events_to_rwe(revents)?; |
| 191 | + if let Some(ref mut fds) = readfds |
| 192 | + && readable |
| 193 | + { |
| 194 | + fds.set(fd)?; |
| 195 | + total_revents += 1; |
| 196 | + } |
| 197 | + if let Some(ref mut fds) = writefds |
| 198 | + && writable |
| 199 | + { |
| 200 | + fds.set(fd)?; |
| 201 | + total_revents += 1; |
| 202 | + } |
| 203 | + if let Some(ref mut fds) = exceptfds |
| 204 | + && except |
| 205 | + { |
| 206 | + fds.set(fd)?; |
| 207 | + total_revents += 1; |
| 208 | + } |
| 209 | + } |
| 210 | + Ok(total_revents) |
| 211 | +} |
| 212 | + |
| 213 | +/// Converts `select` RWE input to `poll` I/O event input |
| 214 | +/// according to Linux's behavior. |
| 215 | +fn convert_rwe_to_events(readable: bool, writable: bool, except: bool) -> EPollEventType { |
| 216 | + let mut events = EPollEventType::empty(); |
| 217 | + if readable { |
| 218 | + events |= EPollEventType::EPOLLIN; |
| 219 | + } |
| 220 | + if writable { |
| 221 | + events |= EPollEventType::EPOLLOUT; |
| 222 | + } |
| 223 | + if except { |
| 224 | + events |= EPollEventType::EPOLLPRI; |
| 225 | + } |
| 226 | + events |
| 227 | +} |
| 228 | + |
| 229 | +/// Converts `poll` I/O event results to `select` RWE results |
| 230 | +/// according to Linux's behavior. |
| 231 | +fn convert_events_to_rwe(events: EPollEventType) -> Result<(bool, bool, bool), SystemError> { |
| 232 | + if events.contains(EPollEventType::EPOLLNVAL) { |
| 233 | + return Err(SystemError::EBADF); |
| 234 | + } |
| 235 | + |
| 236 | + let readable = events |
| 237 | + .intersects(EPollEventType::EPOLLIN | EPollEventType::EPOLLHUP | EPollEventType::EPOLLERR); |
| 238 | + let writable = events.intersects(EPollEventType::EPOLLOUT | EPollEventType::EPOLLERR); |
| 239 | + let except = events.contains(EPollEventType::EPOLLPRI); |
| 240 | + Ok((readable, writable, except)) |
| 241 | +} |
| 242 | + |
| 243 | +#[derive(Debug, Clone, Copy)] |
| 244 | +#[repr(C)] |
| 245 | +struct FdSet { |
| 246 | + fds_bits: [usize; FD_SETSIZE / USIZE_BITS], |
| 247 | +} |
| 248 | + |
| 249 | +impl FdSet { |
| 250 | + /// Equivalent to FD_SET. |
| 251 | + pub fn set(&mut self, fd: usize) -> Result<(), SystemError> { |
| 252 | + if fd >= FD_SETSIZE { |
| 253 | + return Err(SystemError::EINVAL); |
| 254 | + } |
| 255 | + self.fds_bits[fd / USIZE_BITS] |= 1 << (fd % USIZE_BITS); |
| 256 | + Ok(()) |
| 257 | + } |
| 258 | + |
| 259 | + /// Equivalent to FD_CLR. |
| 260 | + #[expect(unused)] |
| 261 | + pub fn unset(&mut self, fd: usize) -> Result<(), SystemError> { |
| 262 | + if fd >= FD_SETSIZE { |
| 263 | + return Err(SystemError::EINVAL); |
| 264 | + } |
| 265 | + self.fds_bits[fd / USIZE_BITS] &= !(1 << (fd % USIZE_BITS)); |
| 266 | + Ok(()) |
| 267 | + } |
| 268 | + |
| 269 | + /// Equivalent to FD_ISSET. |
| 270 | + pub fn is_set(&self, fd: usize) -> bool { |
| 271 | + if fd >= FD_SETSIZE { |
| 272 | + return false; |
| 273 | + } |
| 274 | + (self.fds_bits[fd / USIZE_BITS] & (1 << (fd % USIZE_BITS))) != 0 |
| 275 | + } |
| 276 | + |
| 277 | + /// Equivalent to FD_ZERO. |
| 278 | + pub fn clear(&mut self) { |
| 279 | + for slot in self.fds_bits.iter_mut() { |
| 280 | + *slot = 0; |
| 281 | + } |
| 282 | + } |
| 283 | +} |
| 284 | +#[cfg(target_arch = "x86_64")] |
| 285 | +syscall_table_macros::declare_syscall!(SYS_SELECT, SysSelect); |
0 commit comments