[PATCH] nvme-fc: fix sleep-in-atomic-context bug caused by nvme_fc_rcv_ls_req

[James Smart]

On 10/2/2022 6:50 PM, duoming at zju.edu.cn wrote:
> Hello,
> 
> On Sun, 2 Oct 2022 10:12:15 -0700 James Smart wrote:
> 
>> On 10/1/2022 5:19 PM, Duoming Zhou wrote:
>>> The function lpfc_poll_timeout() is a timer handler that runs in an
>>> atomic context, but it calls "kzalloc(.., GFP_KERNEL)" that may sleep.
>>> As a result, the sleep-in-atomic-context bug will happen. The processes
>>> is shown below:
>>>
>>> lpfc_poll_timeout()
>>>    lpfc_sli_handle_fast_ring_event()
>>>     lpfc_sli_process_unsol_iocb()
>>>      lpfc_complete_unsol_iocb()
>>>       lpfc_nvme_unsol_ls_handler()
>>>        lpfc_nvme_handle_lsreq()
>>>         nvme_fc_rcv_ls_req()
>>>          kzalloc(sizeof(.., GFP_KERNEL) //may sleep
>>>
>>> This patch changes the gfp_t parameter of kzalloc() from GFP_KERNEL to
>>> GFP_ATOMIC in order to mitigate the bug.
>>>
>>> Fixes: 14fd1e98afaf ("nvme-fc: Add Disconnect Association Rcv support")
>>> Signed-off-by: Duoming Zhou <duoming at zju.edu.cn>
>>> ---
>>>    drivers/nvme/host/fc.c | 2 +-
>>>    1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/nvme/host/fc.c b/drivers/nvme/host/fc.c
>>> index 127abaf9ba5..36698dfc8b3 100644
>>> --- a/drivers/nvme/host/fc.c
>>> +++ b/drivers/nvme/host/fc.c
>>> @@ -1754,7 +1754,7 @@ nvme_fc_rcv_ls_req(struct nvme_fc_remote_port *portptr,
>>>    	lsop = kzalloc(sizeof(*lsop) +
>>>    			sizeof(union nvmefc_ls_requests) +
>>>    			sizeof(union nvmefc_ls_responses),
>>> -			GFP_KERNEL);
>>> +			GFP_ATOMIC);
>>>    	if (!lsop) {
>>>    		dev_info(lport->dev,
>>>    			"RCV %s LS failed: No memory\n",
>>
>> I would prefer this was fixed within lpfc rather than introducing atomic
>> allocations (1st in either host or target transport).  It was introduced
>> by lpfc change in irq handling style.
> 
> Thank your for your reply and suggestions!
> 
> Do you think change the lpfc_poll_timeout() to a delayed_work is better?
> 
> Best regards,
> Duoming Zhou

as a minimum: the lpfc_complete_unsol_iocb handler should be passing off 
the iocb to a work queue routine - so that the context changes so that 
either nvme host or nvmet ls callback routines can be called. If 
possible, it should do the axchg alloc - to avoid a GFP_ATOMIC there as 
well...

It's usually best for these nvme LS's and ELS's to be done in a slow 
path thread/work queue element. That may mean segmenting a little 
earlier in the path.

-- james