Re: [PATCH] opensm/osm_sa_path_record.c: livelock in pr_rcv_get_path_parms

[Line Holen <Line.Holen-xsfywfwIY+M@public.gmane.org>]

On 04/19/10 05:34 PM, Sasha Khapyorsky wrote:
> On 11:15 Mon 19 Apr     , Line Holen wrote:
>> SA path request handling can end up in a livelock in pr_rcv_get_path_parms().
>> This can happen if a path request is handled while LFT updates to the fabric
>> are in progress. 
>> The LFT of the switch data structure is updated as part of the LFT response 
>> processing. So while the SM is busy pushing the LFT updates, some switches have
>> up to date LFT info while others are not yet updated and contains the LFT of
>> the previous routing. For a (short) time interval there is a potential for 
>> loops in the fabric. The livelock occurs if a path request is received during
>> this time interval.
>> Both LFT response handling and path request processing needs the SM lock.
>> When the livelock occurs the LFT response handling blocks forever waiting for 
>> the lock to be released.
>>
>> The suggested fix is simply to introduce a max number of hops that should
>> be traversed while handling the path request. If this max is reached then
>> the request will return with NO_RECORD response and release the SM lock.
>> This way the LFT processing will be able to complete.
>>
>> Signed-off-by: Line Holen <Line.Holen-xsfywfwIY+M@public.gmane.org>
> 
> Applied. Thanks. See minor question/note below.
> 
>> ---
>>
>> diff --git a/opensm/opensm/osm_sa_path_record.c b/opensm/opensm/osm_sa_path_record.c
>> index c4c3f86..b399b70 100644
>> --- a/opensm/opensm/osm_sa_path_record.c
>> +++ b/opensm/opensm/osm_sa_path_record.c
>> @@ -4,6 +4,7 @@
>>   * Copyright (c) 1996-2003 Intel Corporation. All rights reserved.
>>   * Copyright (c) 2008 Xsigo Systems Inc. All rights reserved.
>>   * Copyright (c) 2009 HNR Consulting. All rights reserved.
>> + * Copyright (c) 2010 Sun Microsystems, Inc. All rights reserved.
>>   *
>>   * This software is available to you under a choice of one of two
>>   * licenses.  You may choose to be licensed under the terms of the GNU
>> @@ -69,6 +70,9 @@
>>  #include <opensm/osm_prefix_route.h>
>>  #include <opensm/osm_ucast_lash.h>
>>  
>> +
>> +#define MAX_HOPS 128
> 
> IB spec defines maximal number of hops for a fabric which is 64. Would
> it be netter to use this value here?
> 
> Sasha

The value of 128 was chosen as 2x max DR path allowing the SM to be in
the middle of a fabric. But I have no problem lowering to 64.

Line

> 
>> +
>>  typedef struct osm_pr_item {
>>  	cl_list_item_t list_item;
>>  	ib_path_rec_t path_rec;
>> @@ -178,6 +182,7 @@ static ib_api_status_t pr_rcv_get_path_parms(IN osm_sa_t * sa,
>>  	osm_qos_level_t *p_qos_level = NULL;
>>  	uint16_t valid_sl_mask = 0xffff;
>>  	int is_lash;
>> +	int hops = 0;
>>  
>>  	OSM_LOG_ENTER(sa->p_log);
>>  
>> @@ -369,6 +374,25 @@ static ib_api_status_t pr_rcv_get_path_parms(IN osm_sa_t * sa,
>>  				goto Exit;
>>  			}
>>  		}
>> +
>> +		/* update number of hops traversed */
>> +		hops++;
>> +		if (hops > MAX_HOPS) {
>> +
>> +			OSM_LOG(sa->p_log, OSM_LOG_ERROR,
>> +			    "Path from GUID 0x%016" PRIx64 " (%s) to lid %u GUID 0x%016"
>> +			    PRIx64 " (%s) needs more than %d hops, "
>> +			    "max %d hops allowed\n",
>> +			    cl_ntoh64(osm_physp_get_port_guid(p_src_physp)),
>> +			    p_src_physp->p_node->print_desc,
>> +			    dest_lid_ho,
>> +			    cl_ntoh64(osm_physp_get_port_guid(p_dest_physp)),
>> +			    p_dest_physp->p_node->print_desc,
>> +			    hops, MAX_HOPS);
>> +
>> +			status = IB_NOT_FOUND;
>> +			goto Exit;
>> +		}
>>  	}
>>  
>>  	/*
>>

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html