svcrdma: Fix trace point use-after-free race
authorChuck Lever <chuck.lever@oracle.com>
Mon, 30 Mar 2020 18:27:37 +0000 (14:27 -0400)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sat, 2 May 2020 06:48:48 +0000 (08:48 +0200)
commit e28b4fc652c1830796a4d3e09565f30c20f9a2cf upstream.

I hit this while testing nfsd-5.7 with kernel memory debugging
enabled on my server:

Mar 30 13:21:45 klimt kernel: BUG: unable to handle page fault for address: ffff8887e6c279a8
Mar 30 13:21:45 klimt kernel: #PF: supervisor read access in kernel mode
Mar 30 13:21:45 klimt kernel: #PF: error_code(0x0000) - not-present page
Mar 30 13:21:45 klimt kernel: PGD 3601067 P4D 3601067 PUD 87c519067 PMD 87c3e2067 PTE 800ffff8193d8060
Mar 30 13:21:45 klimt kernel: Oops: 0000 [#1] SMP DEBUG_PAGEALLOC PTI
Mar 30 13:21:45 klimt kernel: CPU: 2 PID: 1933 Comm: nfsd Not tainted 5.6.0-rc6-00040-g881e87a3c6f9 #1591
Mar 30 13:21:45 klimt kernel: Hardware name: Supermicro Super Server/X10SRL-F, BIOS 1.0c 09/09/2015
Mar 30 13:21:45 klimt kernel: RIP: 0010:svc_rdma_post_chunk_ctxt+0xab/0x284 [rpcrdma]
Mar 30 13:21:45 klimt kernel: Code: c1 83 34 02 00 00 29 d0 85 c0 7e 72 48 8b bb a0 02 00 00 48 8d 54 24 08 4c 89 e6 48 8b 07 48 8b 40 20 e8 5a 5c 2b e1 41 89 c6 <8b> 45 20 89 44 24 04 8b 05 02 e9 01 00 85 c0 7e 33 e9 5e 01 00 00
Mar 30 13:21:45 klimt kernel: RSP: 0018:ffffc90000dfbdd8 EFLAGS: 00010286
Mar 30 13:21:45 klimt kernel: RAX: 0000000000000000 RBX: ffff8887db8db400 RCX: 0000000000000030
Mar 30 13:21:45 klimt kernel: RDX: 0000000000000040 RSI: 0000000000000000 RDI: 0000000000000246
Mar 30 13:21:45 klimt kernel: RBP: ffff8887e6c27988 R08: 0000000000000000 R09: 0000000000000004
Mar 30 13:21:45 klimt kernel: R10: ffffc90000dfbdd8 R11: 00c068ef00000000 R12: ffff8887eb4e4a80
Mar 30 13:21:45 klimt kernel: R13: ffff8887db8db634 R14: 0000000000000000 R15: ffff8887fc931000
Mar 30 13:21:45 klimt kernel: FS:  0000000000000000(0000) GS:ffff88885bd00000(0000) knlGS:0000000000000000
Mar 30 13:21:45 klimt kernel: CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
Mar 30 13:21:45 klimt kernel: CR2: ffff8887e6c279a8 CR3: 000000081b72e002 CR4: 00000000001606e0
Mar 30 13:21:45 klimt kernel: Call Trace:
Mar 30 13:21:45 klimt kernel: ? svc_rdma_vec_to_sg+0x7f/0x7f [rpcrdma]
Mar 30 13:21:45 klimt kernel: svc_rdma_send_write_chunk+0x59/0xce [rpcrdma]
Mar 30 13:21:45 klimt kernel: svc_rdma_sendto+0xf9/0x3ae [rpcrdma]
Mar 30 13:21:45 klimt kernel: ? nfsd_destroy+0x51/0x51 [nfsd]
Mar 30 13:21:45 klimt kernel: svc_send+0x105/0x1e3 [sunrpc]
Mar 30 13:21:45 klimt kernel: nfsd+0xf2/0x149 [nfsd]
Mar 30 13:21:45 klimt kernel: kthread+0xf6/0xfb
Mar 30 13:21:45 klimt kernel: ? kthread_queue_delayed_work+0x74/0x74
Mar 30 13:21:45 klimt kernel: ret_from_fork+0x3a/0x50
Mar 30 13:21:45 klimt kernel: Modules linked in: ocfs2_dlmfs ocfs2_stack_o2cb ocfs2_dlm ocfs2_nodemanager ocfs2_stackglue ib_umad ib_ipoib mlx4_ib sb_edac x86_pkg_temp_thermal iTCO_wdt iTCO_vendor_support coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel aesni_intel glue_helper crypto_simd cryptd pcspkr rpcrdma i2c_i801 rdma_ucm lpc_ich mfd_core ib_iser rdma_cm iw_cm ib_cm mei_me raid0 libiscsi mei sg scsi_transport_iscsi ioatdma wmi ipmi_si ipmi_devintf ipmi_msghandler acpi_power_meter nfsd nfs_acl lockd auth_rpcgss grace sunrpc ip_tables xfs libcrc32c mlx4_en sd_mod sr_mod cdrom mlx4_core crc32c_intel igb nvme i2c_algo_bit ahci i2c_core libahci nvme_core dca libata t10_pi qedr dm_mirror dm_region_hash dm_log dm_mod dax qede qed crc8 ib_uverbs ib_core
Mar 30 13:21:45 klimt kernel: CR2: ffff8887e6c279a8
Mar 30 13:21:45 klimt kernel: ---[ end trace 87971d2ad3429424 ]---

It's absolutely not safe to use resources pointed to by the @send_wr
argument of ib_post_send() _after_ that function returns. Those
resources are typically freed by the Send completion handler, which
can run before ib_post_send() returns.

Thus the trace points currently around ib_post_send() in the
server's RPC/RDMA transport are a hazard, even when they are
disabled. Rearrange them so that they touch the Work Request only
_before_ ib_post_send() is invoked.

Fixes: bd2abef33394 ("svcrdma: Trace key RDMA API events")
Fixes: 4201c7464753 ("svcrdma: Introduce svc_rdma_send_ctxt")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
include/trace/events/rpcrdma.h
net/sunrpc/xprtrdma/svc_rdma_rw.c
net/sunrpc/xprtrdma/svc_rdma_sendto.c

index 7fd11ec..2464311 100644 (file)
@@ -1638,17 +1638,15 @@ DECLARE_EVENT_CLASS(svcrdma_sendcomp_event,
 
 TRACE_EVENT(svcrdma_post_send,
        TP_PROTO(
-               const struct ib_send_wr *wr,
-               int status
+               const struct ib_send_wr *wr
        ),
 
-       TP_ARGS(wr, status),
+       TP_ARGS(wr),
 
        TP_STRUCT__entry(
                __field(const void *, cqe)
                __field(unsigned int, num_sge)
                __field(u32, inv_rkey)
-               __field(int, status)
        ),
 
        TP_fast_assign(
@@ -1656,12 +1654,11 @@ TRACE_EVENT(svcrdma_post_send,
                __entry->num_sge = wr->num_sge;
                __entry->inv_rkey = (wr->opcode == IB_WR_SEND_WITH_INV) ?
                                        wr->ex.invalidate_rkey : 0;
-               __entry->status = status;
        ),
 
-       TP_printk("cqe=%p num_sge=%u inv_rkey=0x%08x status=%d",
+       TP_printk("cqe=%p num_sge=%u inv_rkey=0x%08x",
                __entry->cqe, __entry->num_sge,
-               __entry->inv_rkey, __entry->status
+               __entry->inv_rkey
        )
 );
 
@@ -1726,26 +1723,23 @@ TRACE_EVENT(svcrdma_wc_receive,
 TRACE_EVENT(svcrdma_post_rw,
        TP_PROTO(
                const void *cqe,
-               int sqecount,
-               int status
+               int sqecount
        ),
 
-       TP_ARGS(cqe, sqecount, status),
+       TP_ARGS(cqe, sqecount),
 
        TP_STRUCT__entry(
                __field(const void *, cqe)
                __field(int, sqecount)
-               __field(int, status)
        ),
 
        TP_fast_assign(
                __entry->cqe = cqe;
                __entry->sqecount = sqecount;
-               __entry->status = status;
        ),
 
-       TP_printk("cqe=%p sqecount=%d status=%d",
-               __entry->cqe, __entry->sqecount, __entry->status
+       TP_printk("cqe=%p sqecount=%d",
+               __entry->cqe, __entry->sqecount
        )
 );
 
@@ -1841,6 +1835,34 @@ DECLARE_EVENT_CLASS(svcrdma_sendqueue_event,
 DEFINE_SQ_EVENT(full);
 DEFINE_SQ_EVENT(retry);
 
+TRACE_EVENT(svcrdma_sq_post_err,
+       TP_PROTO(
+               const struct svcxprt_rdma *rdma,
+               int status
+       ),
+
+       TP_ARGS(rdma, status),
+
+       TP_STRUCT__entry(
+               __field(int, avail)
+               __field(int, depth)
+               __field(int, status)
+               __string(addr, rdma->sc_xprt.xpt_remotebuf)
+       ),
+
+       TP_fast_assign(
+               __entry->avail = atomic_read(&rdma->sc_sq_avail);
+               __entry->depth = rdma->sc_sq_depth;
+               __entry->status = status;
+               __assign_str(addr, rdma->sc_xprt.xpt_remotebuf);
+       ),
+
+       TP_printk("addr=%s sc_sq_avail=%d/%d status=%d",
+               __get_str(addr), __entry->avail, __entry->depth,
+               __entry->status
+       )
+);
+
 #endif /* _TRACE_RPCRDMA_H */
 
 #include <trace/define_trace.h>
index 48fe3b1..a59912e 100644 (file)
@@ -323,8 +323,6 @@ static int svc_rdma_post_chunk_ctxt(struct svc_rdma_chunk_ctxt *cc)
                if (atomic_sub_return(cc->cc_sqecount,
                                      &rdma->sc_sq_avail) > 0) {
                        ret = ib_post_send(rdma->sc_qp, first_wr, &bad_wr);
-                       trace_svcrdma_post_rw(&cc->cc_cqe,
-                                             cc->cc_sqecount, ret);
                        if (ret)
                                break;
                        return 0;
@@ -337,6 +335,7 @@ static int svc_rdma_post_chunk_ctxt(struct svc_rdma_chunk_ctxt *cc)
                trace_svcrdma_sq_retry(rdma);
        } while (1);
 
+       trace_svcrdma_sq_post_err(rdma, ret);
        set_bit(XPT_CLOSE, &xprt->xpt_flags);
 
        /* If even one was posted, there will be a completion. */
index 6fdba72..d4d9844 100644 (file)
@@ -306,15 +306,17 @@ int svc_rdma_send(struct svcxprt_rdma *rdma, struct ib_send_wr *wr)
                }
 
                svc_xprt_get(&rdma->sc_xprt);
+               trace_svcrdma_post_send(wr);
                ret = ib_post_send(rdma->sc_qp, wr, NULL);
-               trace_svcrdma_post_send(wr, ret);
-               if (ret) {
-                       set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
-                       svc_xprt_put(&rdma->sc_xprt);
-                       wake_up(&rdma->sc_send_wait);
-               }
-               break;
+               if (ret)
+                       break;
+               return 0;
        }
+
+       trace_svcrdma_sq_post_err(rdma, ret);
+       set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
+       svc_xprt_put(&rdma->sc_xprt);
+       wake_up(&rdma->sc_send_wait);
        return ret;
 }