Received-SPF: pass (google.com: domain of linux-kernel+bounces-47761-linux.lists.archive=gmail.com@vger.kernel.org designates 139.178.88.99 as permitted sender) client-ip=139.178.88.99;
From: Daniel Vacek <neelx@redhat.com>
To: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>,
	Jason Gunthorpe <jgg@ziepe.ca>,
	Leon Romanovsky <leon@kernel.org>,
	Patrick Kelsey <pat.kelsey@cornelisnetworks.com>,
	Brendan Cunningham <bcunningham@cornelisnetworks.com>
Cc: Daniel Vacek <neelx@redhat.com>,
	stable@vger.kernel.org,
	Mats Kronberg <kronberg@nsc.liu.se>,
	linux-rdma@vger.kernel.org,
	linux-kernel@vger.kernel.org
Subject: [PATCH v2] IB/hfi1: Fix sdma.h tx->num_descs off-by-one error (take two)
Date: Thu,  1 Feb 2024 09:10:08 +0100
Message-ID: <20240201081009.1109442-1-neelx@redhat.com>
In-Reply-To: <20240126152125.869509-1-neelx@redhat.com>
References: <20240126152125.869509-1-neelx@redhat.com>
Precedence: bulk
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit

Unfortunately the commit `fd8958efe877` introduced another error
causing the `descs` array to overflow. This reults in further crashes
easily reproducible by `sendmsg` system call.

[ 1080.836473] general protection fault, probably for non-canonical address 0x400300015528b00a: 0000 [#1] PREEMPT SMP PTI
[ 1080.869326] RIP: 0010:hfi1_ipoib_build_ib_tx_headers.constprop.0+0xe1/0x2b0 [hfi1]
--
[ 1080.974535] Call Trace:
[ 1080.976990]  <TASK>
[ 1081.021929]  hfi1_ipoib_send_dma_common+0x7a/0x2e0 [hfi1]
[ 1081.027364]  hfi1_ipoib_send_dma_list+0x62/0x270 [hfi1]
[ 1081.032633]  hfi1_ipoib_send+0x112/0x300 [hfi1]
[ 1081.042001]  ipoib_start_xmit+0x2a9/0x2d0 [ib_ipoib]
[ 1081.046978]  dev_hard_start_xmit+0xc4/0x210
--
[ 1081.148347]  __sys_sendmsg+0x59/0xa0

crash> ipoib_txreq 0xffff9cfeba229f00
struct ipoib_txreq {
  txreq = {
    list = {
      next = 0xffff9cfeba229f00,
      prev = 0xffff9cfeba229f00
    },
    descp = 0xffff9cfeba229f40,
    coalesce_buf = 0x0,
    wait = 0xffff9cfea4e69a48,
    complete = 0xffffffffc0fe0760 <hfi1_ipoib_sdma_complete>,
    packet_len = 0x46d,
    tlen = 0x0,
    num_desc = 0x0,
    desc_limit = 0x6,
    next_descq_idx = 0x45c,
    coalesce_idx = 0x0,
    flags = 0x0,
    descs = {{
        qw = {0x8024000120dffb00, 0x4}  # SDMA_DESC0_FIRST_DESC_FLAG (bit 63)
      }, {
        qw = {  0x3800014231b108, 0x4}
      }, {
        qw = { 0x310000e4ee0fcf0, 0x8}
      }, {
        qw = {  0x3000012e9f8000, 0x8}
      }, {
        qw = {  0x59000dfb9d0000, 0x8}
      }, {
        qw = {  0x78000e02e40000, 0x8}
      }}
  },
  sdma_hdr =  0x400300015528b000,  <<< invalid pointer in the tx request structure
  sdma_status = 0x0,                   SDMA_DESC0_LAST_DESC_FLAG (bit 62)
  complete = 0x0,
  priv = 0x0,
  txq = 0xffff9cfea4e69880,
  skb = 0xffff9d099809f400
}

If an SDMA send consists of exactly 6 descriptors and requires dword
padding (in the 7th descriptor), the sdma_txreq descriptor array
is not properly expanded and the packet will overflow into the
container structure. This results in a panic when the send completion
runs. The exact panic varies depending on what elements of the
container structure get corrupted. The fix is to use the correct
expression in _pad_sdma_tx_descs() to test the need to expand the
descriptor array.

With this patch the crashes are no longer reproducible and the machine is stable.

Fixes: fd8958efe877 ("IB/hfi1: Fix sdma.h tx->num_descs off-by-one errors")
Cc: stable@vger.kernel.org
Reported-by: Mats Kronberg <kronberg@nsc.liu.se>
Tested-by: Mats Kronberg <kronberg@nsc.liu.se>
Signed-off-by: Daniel Vacek <neelx@redhat.com>
---

Changes in v2:
 - Dropped the unrelated cleanups.
 - Improved commit message as suggested by Dennis Dalessandro

 drivers/infiniband/hw/hfi1/sdma.c |  2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/infiniband/hw/hfi1/sdma.c b/drivers/infiniband/hw/hfi1/sdma.c
index 6e5ac2023328a..b67d23b1f2862 100644
--- a/drivers/infiniband/hw/hfi1/sdma.c
+++ b/drivers/infiniband/hw/hfi1/sdma.c
@@ -3158,7 +3158,7 @@ int _pad_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx)
 {
 	int rval = 0;
 
-	if ((unlikely(tx->num_desc + 1 == tx->desc_limit))) {
+	if ((unlikely(tx->num_desc == tx->desc_limit))) {
 		rval = _extend_sdma_tx_descs(dd, tx);
 		if (rval) {
 			__sdma_txclean(dd, tx);
-- 
2.43.0