vm: require page to be xbusied to invalidate the content

[mjg]

Summary

This will facilitate sbusy for fault handling. Also allowing invalidation while sbusied is weird to say the least.

So far did not blew up with swap testing, will ask pho later.

[kib]

sys/kern/vfs_bio.c

3046

I suspect this might recurse into busy. Imagine that we sbusied the pages for pageout, and do the writes through the buffer cache. If brelse decides to invalidate the buffer for whatever reason then we would wait for the sbusy to go against themself.

[mjg]

sys/kern/vfs_bio.c

3046

can you elaborate what's the exact codepath to end up in this state?

[mjg]

sys/kern/vfs_bio.c

3046

Now I'm even more confused:

in brelse:

if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE ||
    (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) &&
    (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 ||
    vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) {
        vfs_vmio_invalidate(bp);
        allocbuf(bp, 0);
}

this is the only caller

thus for this to happen this has to be *read* which fails.

how is read *to the page* safe to do with the page only sbusied though?

[kib]

sys/kern/vfs_bio.c

3046

The io failure is a rare case. More common is for filesystem to set e.g. B_NOCACHE to not leave the buffer constructed after the write, e.g. for directio support.