On 1/31/24 7:50 AM, Leon Romanovsky wrote:
> On Fri, Jan 26, 2024 at 04:21:23PM +0100, Daniel Vacek wrote:
>> 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
>> }
>>
>> With this patch the crashes are no longer reproducible and the machine is stable.
>>
>> Note, the header file changes are just an unrelated clean-up while I was looking
>> around trying to find the bug.
>>
>> Fixes: fd8958efe877 ("IB/hfi1: Fix sdma.h tx->num_descs off-by-one errors")
>> Cc: stable@xxxxxxxxxxxxxxx
>> Reported-by: Mats Kronberg <kronberg@xxxxxxxxxx>
>> Tested-by: Mats Kronberg <kronberg@xxxxxxxxxx>
>> Signed-off-by: Daniel Vacek <neelx@xxxxxxxxxx>
>> ---
>> drivers/infiniband/hw/hfi1/sdma.c | 2 +-
>> drivers/infiniband/hw/hfi1/sdma.h | 17 +++++++----------
>> 2 files changed, 8 insertions(+), 11 deletions(-)
>>
>> 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))) {
>
> Maybe, Dennis?
I actually have a patch that does exactly this one line change queued up to send
out.
The commit message for our fix is:
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.
>> rval = _extend_sdma_tx_descs(dd, tx);
>> if (rval) {
>> __sdma_txclean(dd, tx);
>> diff --git a/drivers/infiniband/hw/hfi1/sdma.h b/drivers/infiniband/hw/hfi1/sdma.h
>> index d77246b48434f..362815a8da267 100644
>> --- a/drivers/infiniband/hw/hfi1/sdma.h
>> +++ b/drivers/infiniband/hw/hfi1/sdma.h
>> @@ -639,13 +639,13 @@ static inline void sdma_txclean(struct hfi1_devdata *dd, struct sdma_txreq *tx)
>> static inline void _sdma_close_tx(struct hfi1_devdata *dd,
>> struct sdma_txreq *tx)
>> {
>> - u16 last_desc = tx->num_desc - 1;
>> + struct sdma_desc *desc = &tx->descp[tx->num_desc - 1];
>>
>> - tx->descp[last_desc].qw[0] |= SDMA_DESC0_LAST_DESC_FLAG;
>> - tx->descp[last_desc].qw[1] |= dd->default_desc1;
>> + desc->qw[0] |= SDMA_DESC0_LAST_DESC_FLAG;
>> + desc->qw[1] |= dd->default_desc1;
>> if (tx->flags & SDMA_TXREQ_F_URGENT)
>> - tx->descp[last_desc].qw[1] |= (SDMA_DESC1_HEAD_TO_HOST_FLAG |
>> - SDMA_DESC1_INT_REQ_FLAG);
>> + desc->qw[1] |= (SDMA_DESC1_HEAD_TO_HOST_FLAG |
>> + SDMA_DESC1_INT_REQ_FLAG);
>
> Unrelated change which doesn't change anything.
Please drop.
>
>> }
>>
>> static inline int _sdma_txadd_daddr(
>> @@ -670,13 +670,10 @@ static inline int _sdma_txadd_daddr(
>> tx->tlen -= len;
>> /* special cases for last */
>> if (!tx->tlen) {
>> - if (tx->packet_len & (sizeof(u32) - 1)) {
>> + if (tx->packet_len & (sizeof(u32) - 1))
>> rval = _pad_sdma_tx_descs(dd, tx);
>> - if (rval)
>> - return rval;
>> - } else {
>> + else
>> _sdma_close_tx(dd, tx);
>> - }
>
> Same as before, unrelated change.
Agree. Please drop.
