Linux 管道到底能有多快？

% ./fizzbuzz | pv >/dev/null
 422GiB 0:00:16 [36.2GiB/s

int main() {
  size_t buf_size = 1 << 18; // 256KiB
  char* buf = (char*) malloc(buf_size);
  memset((void*)buf, 'X', buf_size); // output Xs
  while (true) {
    size_t remaining = buf_size;
    while (remaining > 0) {
      // Keep invoking `write` until we've written the entirety
      // of the buffer. Remember that write returns how much
      // it could write into the destination -- in this case,
      // our pipe.
      ssize_t written = write(
        STDOUT_FILENO, buf + (buf_size - remaining), remaining
      );
      remaining -= written;
    }
  }
}

% ./write | ./read
3.7GiB/s, 256KiB buffer, 40960 iterations (10GiB piped)

% perf record -g sh -c './write | ./read'
3.2GiB/s, 256KiB buffer, 40960 iterations (10GiB piped)
[ perf record: Woken up 6 times to write data ]
[ perf record: Captured and wrote 2.851 MB perf.data (21201 samples) ]

% perf report -g --symbol-filter=write
-   48.05%     0.05%  write    libc-2.33.so       [.] __GI___libc_write
   - 48.04% __GI___libc_write
      - 47.69% entry_SYSCALL_64_after_hwframe
         - do_syscall_64
            - 47.54% ksys_write
               - 47.40% vfs_write
                  - 47.23% new_sync_write
                     - pipe_write
                        + 24.08% copy_page_from_iter
                        + 11.76% __alloc_pages
                        + 4.32% schedule
                        + 2.98% __wake_up_common_lock
                          0.95% _raw_spin_lock_irq
                          0.74% alloc_pages
                          0.66% prepare_to_wait_event

struct pipe_inode_info {
  unsigned int head;
  unsigned int tail;
  struct pipe_buffer *bufs;
};

struct pipe_buffer {
  struct page *page;
  unsigned int offset, len;
};

% sysbench memory --memory-block-size=1G --memory-oper=read --threads=1 run
...
102400.00 MiB transferred (15921.22 MiB/sec)

struct iovec {
  void  *iov_base; // Starting address
  size_t iov_len;  // Number of bytes
};

// Returns how much we've spliced into the pipe
ssize_t vmsplice(
  int fd, const struct iovec *iov, size_t nr_segs, unsigned int flags
);

int main() {
  size_t buf_size = 1 << 18; // 256KiB
  char* buf = malloc(buf_size);
  memset((void*)buf, 'X', buf_size); // output Xs
  char* bufs[2] = { buf, buf + buf_size/2 };
  int buf_ix = 0;
  // Flip between the two buffers, splicing until we're done.
  while (true) {
    struct iovec bufvec = {
      .iov_base = bufs[buf_ix],
      .iov_len = buf_size/2
    };
    buf_ix = (buf_ix + 1) % 2;
    while (bufvec.iov_len > 0) {
      ssize_t ret = vmsplice(STDOUT_FILENO, &bufvec, 1, 0);
      bufvec.iov_base = (void*) (((char*) bufvec.iov_base) + ret);
      bufvec.iov_len -= ret;
    }
  }
}

% ./write --write_with_vmsplice | ./read
12.7GiB/s, 256KiB buffer, 40960 iterations (10GiB piped)

% ./write --write_with_vmsplice | ./read --read_with_splice
32.8GiB/s, 256KiB buffer, 40960 iterations (10GiB piped)

% perf record -g sh -c './write --write_with_vmsplice | ./read --read_with_splice'
33.4GiB/s, 256KiB buffer, 40960 iterations (10GiB piped)
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.305 MB perf.data (2413 samples) ]
% perf report --symbol-filter=vmsplice
-   49.59%     0.38%  write    libc-2.33.so       [.] vmsplice
   - 49.46% vmsplice
      - 45.17% entry_SYSCALL_64_after_hwframe
         - do_syscall_64
            - 44.30% __do_sys_vmsplice
               + 17.88% iov_iter_get_pages
               + 16.57% __mutex_lock.constprop.0
                 3.89% add_to_pipe
                 1.17% iov_iter_advance
                 0.82% mutex_unlock
                 0.75% pipe_lock
        2.01% __entry_text_start
        1.45% syscall_return_via_sysret

void* buf = malloc(10000);
printf("%p\n", buf);          // 0x6f42430

struct pipe_inode_info {
  unsigned int head;
  unsigned int tail;
  struct pipe_buffer *bufs;
};

struct pipe_buffer {
  struct page *page;
  unsigned int offset, len;
};

ssize_t iov_iter_get_pages(
  struct iov_iter *i,  // input: a sized buffer in virtual memory
  struct page **pages, // output: the list of pages which back the input buffers
  size_t maxsize,      // maximum number of bytes to get
  unsigned maxpages,   // maximum number of pages to get
  size_t *start        // offset into first page, if the input buffer wasn't page-aligned
);

% perf report -g --symbol-filter=iov_iter_get_pages
-   17.08%     0.17%  write    [kernel.kallsyms]  [k] iov_iter_get_pages
   - 16.91% iov_iter_get_pages
      - 16.88% internal_get_user_pages_fast
           11.22% try_grab_compound_head

int get_user_pages_fast(
  // virtual address, page aligned
  unsigned long start,
  // number of pages to retrieve
  int nr_pages,
  // flags, the meaning of which we won't get into
  unsigned int gup_flags,
  // output physical pages
  struct page **pages
)

void* buf = aligned_alloc(1 << 21, size);
madvise(buf, size, MADV_HUGEPAGE)

% ./write --write_with_vmsplice --huge_page | ./read --read_with_splice
51.0GiB/s, 256KiB buffer, 40960 iterations (10GiB piped)

-   46.91%     0.38%  write    libc-2.33.so       [.] vmsplice
   - 46.84% vmsplice
      - 43.15% entry_SYSCALL_64_after_hwframe
         - do_syscall_64
            - 41.80% __do_sys_vmsplice
               + 14.90% wait_for_space
               + 8.27% __wake_up_common_lock
                 4.40% add_to_pipe
               + 4.24% iov_iter_get_pages
               + 3.92% __mutex_lock.constprop.0
                 1.81% iov_iter_advance
               + 0.55% import_iovec
            + 0.76% syscall_exit_to_user_mode
        1.54% syscall_return_via_sysret
        1.49% __entry_text_start

...
// SPLICE_F_NONBLOCK will cause `vmsplice` to return immediately
// if we can't write to the pipe, returning EAGAIN
ssize_t ret = vmsplice(STDOUT_FILENO, &bufvec, 1, SPLICE_F_NONBLOCK);
if (ret < 0 && errno == EAGAIN) {
  continue; // busy loop if not ready to write
}
...

% ./write --write_with_vmsplice --huge_page --busy_loop | ./read --read_with_splice --busy_loop
62.5GiB/s, 256KiB buffer, 40960 iterations (10GiB piped)

boot.kernelPatches = [{
  name = "disable-psi";
  patch = null;
  extraConfig = '' 
    PSI n 
  '';
}];

% ./write --write_with_vmsplice --dont_touch_pages | ./read --read_with_splice
25.0GiB/s, 256KiB buffer, 40960 iterations (10GiB piped)