OpenAFS Master Repository branch, master, updated. BP-openafs-stable-1_8_x-423-gd13b647

[Gerrit Code Review]

The following commit has been merged in the master branch:
commit d13b647aa392e1d802be1023930a8e1a07fb11ab
Author: Andrew Deason <adeason@dson.org>
Date:   Sat Jul 20 23:09:27 2019 -0500

    FBSD: Give 0 'rootrefs' to vflush on unmount
    
    Currently, in afs_unmount, we give vflush a 'rootrefs' arg of 1,
    indicating that we hold 1 reference on the root vnode. But ever since
    commit 6eb1088a (freebsd: properly track vcache references), we drop
    the ref for the root vnode at the beginning of this function.
    
    What happens currently in afs_unmount for a normal successful umount
    is something like this (at least, on FreeBSD 11.2-RELEASE):
    
    - We afs_PutVCache the afs_globalVp vcache, reducing its v_usecount
      and v_holdcnt to 0, and afs_globalVp is set to NULL.
    
    - vflush calls afs_root() to get the root vnode, which sees that
      afs_globalVp is NULL, and so calls afs_GetVCache for the root fid
      and returns it (and sets afs_globalVp to that vcache), with a
      v_usecount of 1.
    
    - vflush tries to vgonel() all of our vnodes, which calls our
      afs_vop_reclaim, which calls afs_FlushVCache(). For the root vnode
      specifically, vflush() sees that v_usecount is nonzero, and so skips
      calling vgonel() at first, but later calls vgone() on it
      specifically because we gave a nonzero 'rootrefs'. The resulting
      afs_FlushVCache() for the root vnode fails, because the root vnode's
      v_usecount is still 1. Since a failure from afs_vop_reclaim would
      cause a panic, we just log a warning and try to continue on anyway.
    
    - vflush() calls vrele() on the root vnode, right before returning.
    
    All of this allows the unmount to proceed, but this means that most of
    afs_FlushVCache() doesn't actually run for the root vcache, and it
    means we always log a warning like this on unmount:
    
        afs_vop_reclaim: afs_FlushVCache failed code 16 [...]
    
    In addition, this means that setting afs_globalVp at the beginning of
    afs_unmount() is largely pointless, since it gets set to a vcache
    again near the beginning of vflush().
    
    To avoid all of this, stop lying to vflush about how many references
    to the root vnode we hold, and just say that we hold 0 references.
    
    Change-Id: Ib434c5fc48e67c3863fcad41279c3d9e0e0b8c2b
    Reviewed-on: https://gerrit.openafs.org/13709
    Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
    Tested-by: Benjamin Kaduk <kaduk@mit.edu>

 src/afs/FBSD/osi_vfsops.c |   17 ++++++++---------
 1 files changed, 8 insertions(+), 9 deletions(-)

-- 
OpenAFS Master Repository