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.
- When ->send_request() is called, rpcrdma_marshal_req() does not
set up a Reply chunk because buflen is smaller than the inline
threshold. Thus rpcrdma_convert_iovs() does not get invoked at
all and the transport's XDRBUF_SPARSE_PAGES logic is not invoked
on the receive buffer.
- During reply processing, rpcrdma_inline_fixup() tries to copy
received data into rq_rcv_buf->pages because page_len is positive.
But there are no receive pages because rpcrdma_marshal_req() never
allocated them.
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@netapp.com>
Suggested-by: Olga kornievskaia <kolga@netapp.com>
Fixes: c426fc53045e ("NFSv4.2: add the extended attribute proc functions.")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Olga kornievskaia <kolga@netapp.com>
Reviewed-by: Frank van der Linden <fllinden@amazon.com>
Tested-by: Olga kornievskaia <kolga@netapp.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
.rpc_resp = &res,
};
u32 xdrlen;
- int ret, np;
+ int ret, np, i;
+ ret = -ENOMEM;
res.scratch = alloc_page(GFP_KERNEL);
if (!res.scratch)
- return -ENOMEM;
+ goto out;
xdrlen = nfs42_listxattr_xdrsize(buflen);
if (xdrlen > server->lxasize)
np = xdrlen / PAGE_SIZE + 1;
pages = kcalloc(np, sizeof(struct page *), GFP_KERNEL);
- if (pages == NULL) {
- __free_page(res.scratch);
- return -ENOMEM;
+ if (!pages)
+ goto out_free_scratch;
+ for (i = 0; i < np; i++) {
+ pages[i] = alloc_page(GFP_KERNEL);
+ if (!pages[i])
+ goto out_free_pages;
}
arg.xattr_pages = pages;
*eofp = res.eof;
}
+out_free_pages:
while (--np >= 0) {
if (pages[np])
__free_page(pages[np]);
}
-
- __free_page(res.scratch);
kfree(pages);
-
+out_free_scratch:
+ __free_page(res.scratch);
+out:
return ret;
}
rpc_prepare_reply_pages(req, args->xattr_pages, 0, args->count,
hdr.replen);
- req->rq_rcv_buf.flags |= XDRBUF_SPARSE_PAGES;
encode_nops(&hdr);
}