On Thu, Nov 26, 2020 at 12:10:21PM -0500, Chuck Lever wrote:
>
>
> > On Nov 25, 2020, at 7:21 PM, Frank van der Linden <fllinden@xxxxxxxxxx> wrote:
> >
> > On Tue, Nov 24, 2020 at 10:40:25PM +0000, Kornievskaia, Olga wrote:
> >>
> >>
> >> On 11/24/20, 4:20 PM, "Frank van der Linden" <fllinden@xxxxxxxxxx> wrote:
> >>
> >> On Tue, Nov 24, 2020 at 08:50:36PM +0000, Kornievskaia, Olga wrote:
> >>>
> >>>
> >>> On 11/24/20, 3:06 PM, "Frank van der Linden" <fllinden@xxxxxxxxxx> wrote:
> >>>
> >>> On Tue, Nov 24, 2020 at 12:26:32PM -0500, Chuck Lever wrote:
> >>>>
> >>>>
> >>>>
> >>>> By switching to an XFS-backed export, I am able to reproduce the
> >>>> ibcomp worker crash on my client with xfstests generic/013.
> >>>>
> >>>> For the failing LISTXATTRS operation, xdr_inline_pages() is called
> >>>> with page_len=12 and buflen=128. Then:
> >>>>
> >>>> - Because buflen is small, rpcrdma_marshal_req will not set up a
> >>>> Reply chunk and the rpcrdma's XDRBUF_SPARSE_PAGES logic does not
> >>>> get invoked at all.
> >>>>
> >>>> - Because page_len is non-zero, rpcrdma_inline_fixup() tries to
> >>>> copy received data into rq_rcv_buf->pages, but they're missing.
> >>>>
> >>>> The result is that the ibcomp worker faults and dies. Sometimes that
> >>>> causes a visible crash, and sometimes it results in a transport
> >>>> hang without other symptoms.
> >>>>
> >>>> RPC/RDMA's XDRBUF_SPARSE_PAGES support is not entirely correct, and
> >>>> should eventually be fixed or replaced. However, my preference is
> >>>> that upper-layer operations should explicitly allocate their receive
> >>>> buffers (using GFP_KERNEL) when possible, rather than relying on
> >>>> XDRBUF_SPARSE_PAGES.
> >>>>
> >>>> Reported-by: Olga kornievskaia <kolga@xxxxxxxxxx>
> >>>> Suggested-by: Olga kornievskaia <kolga@xxxxxxxxxx>
> >>>> Signed-off-by: Chuck Lever <chuck.lever@xxxxxxxxxx>
> >>>> ---
> >>>> fs/nfs/nfs42proc.c | 17 ++++++++++-------
> >>>> fs/nfs/nfs42xdr.c | 1 -
> >>>> 2 files changed, 10 insertions(+), 8 deletions(-)
> >>>>
> >>>> Hi-
> >>>>
> >>>> I like Olga's proposed approach. What do you think of this patch?
> >>>>
> >>>>
> >>>> diff --git a/fs/nfs/nfs42proc.c b/fs/nfs/nfs42proc.c
> >>>> index 2b2211d1234e..24810305ec1c 100644
> >>>> --- a/fs/nfs/nfs42proc.c
> >>>> +++ b/fs/nfs/nfs42proc.c
> >>>> @@ -1241,7 +1241,7 @@ static ssize_t _nfs42_proc_listxattrs(struct inode *inode, void *buf,
> >>>> .rpc_resp = &res,
> >>>> };
> >>>> u32 xdrlen;
> >>>> - int ret, np;
> >>>> + int ret, np, i;
> >>>>
> >>>>
> >>>> res.scratch = alloc_page(GFP_KERNEL);
> >>>> @@ -1253,10 +1253,14 @@ static ssize_t _nfs42_proc_listxattrs(struct inode *inode, void *buf,
> >>>> xdrlen = server->lxasize;
> >>>> np = xdrlen / PAGE_SIZE + 1;
> >>>>
> >>>> + ret = -ENOMEM;
> >>>> pages = kcalloc(np, sizeof(struct page *), GFP_KERNEL);
> >>>> - if (pages == NULL) {
> >>>> - __free_page(res.scratch);
> >>>> - return -ENOMEM;
> >>>> + if (pages == NULL)
> >>>> + goto out_free;
> >>>> + for (i = 0; i < np; i++) {
> >>>> + pages[i] = alloc_page(GFP_KERNEL);
> >>>> + if (!pages[i])
> >>>> + goto out_free;
> >>>> }
> >>>>
> >>>> arg.xattr_pages = pages;
> >>>> @@ -1271,14 +1275,13 @@ static ssize_t _nfs42_proc_listxattrs(struct inode *inode, void *buf,
> >>>> *eofp = res.eof;
> >>>> }
> >>>>
> >>>> +out_free:
> >>>> while (--np >= 0) {
> >>>> if (pages[np])
> >>>> __free_page(pages[np]);
> >>>> }
> >>>> -
> >>>> - __free_page(res.scratch);
> >>>> kfree(pages);
> >>>> -
> >>>> + __free_page(res.scratch);
> >>>> return ret;
> >>>>
> >>>> }
> >>>> diff --git a/fs/nfs/nfs42xdr.c b/fs/nfs/nfs42xdr.c
> >>>> index 6e060a88f98c..8432bd6b95f0 100644
> >>>> --- a/fs/nfs/nfs42xdr.c
> >>>> +++ b/fs/nfs/nfs42xdr.c
> >>>> @@ -1528,7 +1528,6 @@ static void nfs4_xdr_enc_listxattrs(struct rpc_rqst *req,
> >>>>
> >>>> rpc_prepare_reply_pages(req, args->xattr_pages, 0, args->count,
> >>>> hdr.replen);
> >>>> - req->rq_rcv_buf.flags |= XDRBUF_SPARSE_PAGES;
> >>>>
> >>>> encode_nops(&hdr);
> >>>> }
> >>>>
> >>>>
> >>>
> >>> I can see why this is the simplest and most pragmatic solution, so it's
> >>> fine with me.
> >>>
> >>> Why doesn't this happen with getxattr? Do we need to convert that too?
> >>>
> >>> [olga] I don't know if GETXATTR/SETXATTR works. I'm not sure what tests exercise those operations. I just ran into the fact that generic/013 wasn't passing. And I don't see that it's an xattr specific tests. I'm not sure how it ends up triggering is usage of xattr.
> >>
> >> I'm attaching the test program I used, it should give things a better workout.
> >>
> >> [olga] I'm not sure if I'm doing something wrong but there are only 2 GETXATTR call on the network trace from running this application and both calls are returning an error (ERR_NOXATTR). Which btw might explain why no problems are seen since no decoding of data is happening. There are lots of SETXATTRs and REMOVEXATTR and there is a LISTXATTR (which btw network trace is marking as malformed so there might something bad there). Anyway...
> >>
> >> This is my initial report: no real exercise of the GETXATTR code as far as I can tell.
> >
> > True, the test is heavier on the setxattr / listxattr side. And with caching,
> > you're not going to see a lot of GETXATTR calls. I used the same test program
> > with caching off, and it works fine, though.
>
> I unintentionally broke GETXATTR while developing the LISTXATTRS fix,
> and generic/013 rather aggressively informed me that GETXATTR was no
> longer working. There is some test coverage there, fwiw.
Oh, the coverage was good - in my testing I also used a collection of
small unit test programs, and I was the one who made the xattr tests
in xfstests work on NFS.
>
>
> > In any case, after converting GETXATTR to pre-allocate pages, I noticed that,
> > when I disabled caching, I was getting EIO instead of ERANGE back from
> > calls that test the case of calling getxattr() with a buffer length that
> > is insufficient.
>
> Was TCP the underlying RPC transport?
Yes, this was TCP. I have set up rdma over rxe, which I'll test too if I
can get this figured out. It might be a long standing bug in xdr_inline_pages,
as listxattr / getxattr seem to be simply the first ones to pass in a
length that is not page aligned.
It does make some sense to round the length up to PAGE_SIZE, and just check if
the received data fits when decoding, like other calls do. It improves your
chances of getting a result that you can still cache, even if it's too big for
the length that was asked for. E.g. if the result is > requested_length, but
< ROUND_UP(requested_length, PAGE_SIZE), you can cache it, even though you
have to return -ERANGE to the caller.
- Frank
