`fd_pwrite` after setting `fdflags::append` writes from the wrong offset

[yagehu]

This is a subtle platform-specific WAIS bug. Run the following file compiled with wasi-sdk. It creates a file tmp/a and performs this sequence of actions:

1. Write 67 bytes to tmp/a.
2. Set the fdflags::append flag via fcntl().
3. Write 102 bytes at offset 0 to tmp/a with fd_pwrite.

#include <fcntl.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>

int main() {
    int f = open("tmp/a", O_CREAT | O_WRONLY);
    if (f == -1) {
        perror("open");
	return 1;
    }

    char buf[67];

    int written = write(f, buf, 67);
    printf("written %d\n", written);

    int ret = fcntl(f, F_SETFL, O_APPEND);
    if (ret == -1) {
	perror("fcntl");
	return 1;
    }


    char buf2[102];

    int written2 = pwrite(f, buf2, 102, 0);
    printf("written %d\n", written2);
}

The resulting file should be 102 bytes because fd_pwrite should write from offset 0 regardless of the append flag per specification (WASI specifies fd_pwrite to be similar to POSIX pwritev which specifies this behavior).

Wasmtime behaves correctly, producing a 102-byte file. WAMR produces a 169-byte file.

[wenyongh]

Hi, I tested the result of the gcc native (gcc -O3 -o test main, and then ./test), it also creates a 169-byte file. It is confusing whether we should comply with wasi specification or not, if yes, we may change the behavior of C application. And I found that eventually the libc pwritev (or pwrite) was called with offset 0, but it is strange that it didn't write the data from offset 0.

[yagehu]

@wenyongh Definitely a fair point. man 2 pwrite identifies this as a Linux bug. It was first raised by Micharl Kerrisk. However, you can also argue WASI aims to be platform independent. The Wasmtime team has probably fixed it by using cap-std. I leave it up to you WAMR folks to decide if you wontfix it or not. I'm just reporting this deviation from the specification.

[yagehu]

I did some digging and found that Wasmtime solves it in preview1 by managing the append flag and hiding it from the underlying syscall. It has nothing to do with cap-std, which simply forwards the call to Rust std.

Notably, here and here.

[yamt]

I did some digging and found that Wasmtime solves it in preview1 by managing the append flag and hiding it from the underlying syscall. It has nothing to do with cap-std, which simply forwards the call to Rust std.

Notably, here and here.

such a "fix" would break an important use case of O_APPEND, which is multiple writers to a single file.

[yagehu]

I did some digging and found that Wasmtime solves it in preview1 by managing the append flag and hiding it from the underlying syscall. It has nothing to do with cap-std, which simply forwards the call to Rust std. Notably, here and here.

such a "fix" would break an important use case of O_APPEND, which is multiple writers to a single file.

I don't think it would break the use case if implemented correctly considering you don't have multiple processes in WASI.

[yamt]

I did some digging and found that Wasmtime solves it in preview1 by managing the append flag and hiding it from the underlying syscall. It has nothing to do with cap-std, which simply forwards the call to Rust std. Notably, here and here.

such a "fix" would break an important use case of O_APPEND, which is multiple writers to a single file.

I don't think it would break the use case if implemented correctly considering you don't have multiple processes in WASI.

"multiple writers" can include a host process which is not related to wamr.

[yagehu]

Since this is a Linux behavior, it would be hard to fix without emulating the WASI fs layer. Closing as wontfix.