[dpdk-dev] [PATCH v3 2/5] eal: don't process IPC messages before init finished

Burakov, Anatoly anatoly.burakov at intel.com
Wed Feb 28 10:45:54 CET 2018


On 28-Feb-18 1:09 AM, Tan, Jianfeng wrote:
> 
> 
>> -----Original Message-----
>> From: Burakov, Anatoly
>> Sent: Tuesday, February 27, 2018 10:36 PM
>> To: dev at dpdk.org
>> Cc: Tan, Jianfeng
>> Subject: [PATCH v3 2/5] eal: don't process IPC messages before init finished
>>
>> It is not possible for a primary process to receive any messages
>> while initializing, because RTE_MAGIC value is not set in the
>> shared config, and hence no secondary process can ever spin up
>> during that time.
>>
>> However, it is possible for a secondary process to receive messages
>> from the primary during initialization. We can't just drop the
>> messages as they may be important, and also we might need to process
>> replies to our own requests (e.g. VFIO) during initialization.
>>
>> Therefore, add a tailq for incoming messages, and queue them up
>> until initialization is complete, and process them in order they
>> arrived.
>>
>> Signed-off-by: Anatoly Burakov <anatoly.burakov at intel.com>
>> ---
>>
>> Notes:
>>      v3: check for init_complete after receiving message
>>
>>      v2: no changes
>>
>>   lib/librte_eal/common/eal_common_proc.c | 52
>> +++++++++++++++++++++++++++++----
>>   1 file changed, 47 insertions(+), 5 deletions(-)
>>
>> diff --git a/lib/librte_eal/common/eal_common_proc.c
>> b/lib/librte_eal/common/eal_common_proc.c
>> index 3a1088e..a6e24e6 100644
>> --- a/lib/librte_eal/common/eal_common_proc.c
>> +++ b/lib/librte_eal/common/eal_common_proc.c
>> @@ -25,6 +25,7 @@
>>   #include <rte_errno.h>
>>   #include <rte_lcore.h>
>>   #include <rte_log.h>
>> +#include <rte_tailq.h>
>>
>>   #include "eal_private.h"
>>   #include "eal_filesystem.h"
>> @@ -58,6 +59,18 @@ struct mp_msg_internal {
>>   	struct rte_mp_msg msg;
>>   };
>>
>> +struct message_queue_entry {
>> +	TAILQ_ENTRY(message_queue_entry) next;
>> +	struct mp_msg_internal msg;
>> +	struct sockaddr_un sa;
>> +};
>> +
>> +/** Double linked list of received messages. */
>> +TAILQ_HEAD(message_queue, message_queue_entry);
>> +
>> +static struct message_queue message_queue =
>> +	TAILQ_HEAD_INITIALIZER(message_queue);
>> +
>>   struct sync_request {
>>   	TAILQ_ENTRY(sync_request) next;
>>   	int reply_received;
>> @@ -276,12 +289,41 @@ process_msg(struct mp_msg_internal *m, struct
>> sockaddr_un *s)
>>   static void *
>>   mp_handle(void *arg __rte_unused)
>>   {
>> -	struct mp_msg_internal msg;
>> -	struct sockaddr_un sa;
>> -
>> +	struct message_queue_entry *cur_msg, *next_msg, *new_msg =
>> NULL;
>>   	while (1) {
>> -		if (read_msg(&msg, &sa) == 0)
>> -			process_msg(&msg, &sa);
>> +		/* we want to process all messages in order of their arrival,
>> +		 * but status of init_complete may change while we're
>> iterating
>> +		 * the tailq. so, store it here and check once every iteration.
>> +		 */
>> +		int init_complete;
>> +
>> +		if (new_msg == NULL)
>> +			new_msg = malloc(sizeof(*new_msg));
>> +		if (read_msg(&new_msg->msg, &new_msg->sa) == 0) {
> 
> Suppose a case that: req and msg received but init not completed, so we enqueue all of them in the tailq; and from now on, no req/rep/msg comes. Then mp thread will hang here for reading new message.
> In such a case, we might need the master thread to signal mp thread to wake up when init is completed?

True. Will have to think about how to do that.

-- 
Thanks,
Anatoly


More information about the dev mailing list