[PATCH RFC] svcrdma: Preserve CB send buffer across retransmits

[Chuck Lever <chuck.lever@oracle.com>]

During each NFSv4 callback Call, an RDMA Send completion frees the
page that contains the RPC Call message. If the upper layer determines
that a retransmit is necessary, this is too soon.

One possible symptom: after a GARBAGE_ARGS response an NFSv4.1
callback request, the following BUG fires on the NFS server:

kernel: BUG: Bad page state in process kworker/0:2H  pfn:7d3ce2
kernel: page:ffffea001f4f3880 count:-2 mapcount:0 mapping:          (null) index:0x0
kernel: flags: 0x2fffff80000000()
kernel: raw: 002fffff80000000 0000000000000000 0000000000000000 fffffffeffffffff
kernel: raw: dead000000000100 dead000000000200 0000000000000000 0000000000000000
kernel: page dumped because: nonzero _refcount
kernel: Modules linked in: cts rpcsec_gss_krb5 ocfs2_dlmfs ocfs2_stack_o2cb ocfs2_dlm
ocfs2_nodemanager ocfs2_stackglue rpcrdm a ib_ipoib rdma_ucm ib_ucm ib_uverbs ib_umad
rdma_cm ib_cm iw_cm x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel
kvm irqbypass crct10dif_pc lmul crc32_pclmul ghash_clmulni_intel pcbc iTCO_wdt
iTCO_vendor_support aesni_intel crypto_simd glue_helper cryptd pcspkr lpc_ich i2c_i801
mei_me mf d_core mei raid0 sg wmi ioatdma ipmi_si ipmi_devintf ipmi_msghandler shpchp
acpi_power_meter acpi_pad nfsd nfs_acl lockd auth_rpcgss grace sunrpc ip_tables xfs
libcrc32c mlx4_en mlx4_ib mlx5_ib ib_core sd_mod sr_mod cdrom ast drm_kms_helper
syscopyarea sysfillrect sysimgblt fb_sys_fops ttm ahci crc32c_intel libahci drm
mlx5_core igb libata mlx4_core dca i2c_algo_bit i2c_core nvme
kernel: ptp nvme_core pps_core dm_mirror dm_region_hash dm_log dm_mod dax
kernel: CPU: 0 PID: 11495 Comm: kworker/0:2H Not tainted 4.14.0-rc3-00001-g577ce48 #811
kernel: Hardware name: Supermicro Super Server/X10SRL-F, BIOS 1.0c 09/09/2015
kernel: Workqueue: ib-comp-wq ib_cq_poll_work [ib_core]
kernel: Call Trace:
kernel: dump_stack+0x62/0x80
kernel: bad_page+0xfe/0x11a
kernel: free_pages_check_bad+0x76/0x78
kernel: free_pcppages_bulk+0x364/0x441
kernel: ? ttwu_do_activate.isra.61+0x71/0x78
kernel: free_hot_cold_page+0x1c5/0x202
kernel: __put_page+0x2c/0x36
kernel: svc_rdma_put_context+0xd9/0xe4 [rpcrdma]
kernel: svc_rdma_wc_send+0x50/0x98 [rpcrdma]

This issue exists all the way back to v4.5, but refactoring and code
re-organization prevents this simple patch from applying to kernels
older than v4.12. The fix is the same, however, if someone needs to
backport it.

Reported-by: Ben Coddington <bcodding@redhat.com>
Fixes: 5d252f90a800 ('svcrdma: Add class for RDMA backwards ... ')
Cc: stable@vger.kernel.org # v4.12
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
---
 net/sunrpc/xprtrdma/svc_rdma_backchannel.c |    6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

Hi Ben, Jeff-

I was able to reproduce the problem here with a rigged Linux client.
This is my proposed fix. Fix also tested with a prototype Solaris
client similar to the one that exposed the problem.


diff --git a/net/sunrpc/xprtrdma/svc_rdma_backchannel.c b/net/sunrpc/xprtrdma/svc_rdma_backchannel.c
index ec37ad8..1854db2 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_backchannel.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_backchannel.c
@@ -132,6 +132,10 @@ static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma,
 	if (ret)
 		goto out_err;
 
+	/* Bump page refcnt so Send completion doesn't release
+	 * the rq_buffer before all retransmits are complete.
+	 */
+	get_page(virt_to_page(rqst->rq_buffer));
 	ret = svc_rdma_post_send_wr(rdma, ctxt, 1, 0);
 	if (ret)
 		goto out_unmap;
@@ -164,7 +168,6 @@ static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma,
 		return -EINVAL;
 	}
 
-	/* svc_rdma_sendto releases this page */
 	page = alloc_page(RPCRDMA_DEF_GFP);
 	if (!page)
 		return -ENOMEM;
@@ -183,6 +186,7 @@ static int svc_rdma_bc_sendto(struct svcxprt_rdma *rdma,
 {
 	struct rpc_rqst *rqst = task->tk_rqstp;
 
+	put_page(virt_to_page(rqst->rq_buffer));
 	kfree(rqst->rq_rbuffer);
 }