On Wed, May 13, 2026 at 08:00:17PM -0700, Rick Macklem wrote: > On Wed, May 13, 2026 at 7:50 PM Konstantin Belousov <[email protected]> > wrote: > > > > On Wed, May 13, 2026 at 05:30:46PM -0700, Rick Macklem wrote: > > > Hi, > > > > > > The following KASSERT() is at the beginning of allocuio(): > > > KASSERT(iovcnt <= UIO_MAXIOV, > > > ("Requested %u iovecs exceed UIO_MAXIOV", iovcnt)); > > > > > > This fails for the NFS server if it is configured for > 1Mbyte I/O > > > size, since the number of elements (mbufs) for the VOP_READ() > > > > > > exceeds UIO_MAXIOV (1024). This shows up because ZFS > > > does a cloneuio() call which calls allocuio(). > > > > > > Since UIO_MAXIOV is used is several places, including setting > > > the limit for copyinuio() and freebsd32_copyinuio(), I don't think > > > changing the value of UIO_MAXIOV is an appropriate fix. > > > (ie. This changes the APIs, etc.) > > > > > > Now, since all that the above check does it set a sanity limit > > > on how big the allocated uio can be, do you think it is > > > reasonable to change the above KASSERT() to: > > > KASSERT(iovcnt <= 4096, > > > ("Requested %u iovecs exceed 4096", iovcnt)); > > > which would allow a 4Mbyte NFS I/O to work. > > > > > > Note that copyinuio() and freebsd32_copyinuio() check the > > > iov length for < UIO_MAXIOV before calling allocuio(), so > > > those interfaces are not broken by this. > > > > I do not quite understand how changing the assert in allocuio() would > > change anything? All callers of the function (copyinuio, cloneuio, > > and freebsd32_copyinuio) check iovcnt, and I do not see a call to > > allocuio() from the NFS server. > > > > Where specifically does the NFS server fail with too long iovec? > zfs_freebsd_write()->zfs_write()->zfs_uiocopy()->cloneuio()->allocuio() > (and zfs_freebsd_write() is ZFS's VOP_WRITE())
I would just remove the assert from allocuio() then.
