On 2021/2/4 18:18, Dr. David Alan Gilbert wrote:
> * Chuan Zheng (zhengch...@huawei.com) wrote:
>> Signed-off-by: Zhimin Feng <fengzhim...@huawei.com>
>> Signed-off-by: Chuan Zheng <zhengch...@huawei.com>
>> ---
>> migration/rdma.c | 65
>> ++++++++++++++++++++++++++++++++++++++++++++++++++++----
>> 1 file changed, 61 insertions(+), 4 deletions(-)
>>
>> diff --git a/migration/rdma.c b/migration/rdma.c
>> index 2097839..c19a91f 100644
>> --- a/migration/rdma.c
>> +++ b/migration/rdma.c
>> @@ -2002,6 +2002,20 @@ static int qemu_rdma_write_one(QEMUFile *f,
>> RDMAContext *rdma,
>> .repeat = 1,
>> };
>>
>> + /* use multifd to send data */
>> + if (migrate_use_multifd()) {
>> + int channel = get_multifd_RDMA_channel();
>> + int ret = 0;
>> + MultiFDSendParams *multifd_send_param = NULL;
>> + ret = get_multifd_send_param(channel, &multifd_send_param);
>> + if (ret) {
>> + error_report("rdma: error getting multifd_send_param(%d)",
>> channel);
>> + return -EINVAL;
>> + }
>> + rdma = (RDMAContext *)multifd_send_param->rdma;
>> + block = &(rdma->local_ram_blocks.block[current_index]);
>> + }
>> +
>> retry:
>> sge.addr = (uintptr_t)(block->local_host_addr +
>> (current_addr - block->offset));
>> @@ -2197,6 +2211,27 @@ retry:
>> return 0;
>> }
>>
>> +static int multifd_rdma_write_flush(void)
>> +{
>> + /* The multifd RDMA threads send data */
>> + MultiFDSendParams *multifd_send_param = NULL;
>> + RDMAContext *rdma = NULL;
>> + MigrationState *s = migrate_get_current();
>> + int ret = 0;
>> +
>> + ret = get_multifd_send_param(s->rdma_channel,
>> + &multifd_send_param);
>> + if (ret) {
>> + error_report("rdma: error getting multifd_send_param(%d)",
>> + s->rdma_channel);
>
> Do we need these error_report's for get_multifd_send_param calls - how
> can they fail in practice?
>
Maybe we do not need it.
The s->rdma_channel should not exceed the migrate_multifd_channels and should
not negative.
>> + return ret;
>> + }
>> + rdma = (RDMAContext *)(multifd_send_param->rdma);
>> + rdma->nb_sent++;
>> +
>> + return ret;
>
> But this doesn't actually 'flush' anything?
>
Yes, it just use to increase the nb_sent. we need to choose a more suitable
function name.
>> +}
>> +
>> /*
>> * Push out any unwritten RDMA operations.
>> *
>> @@ -2219,8 +2254,15 @@ static int qemu_rdma_write_flush(QEMUFile *f,
>> RDMAContext *rdma)
>> }
>>
>> if (ret == 0) {
>> - rdma->nb_sent++;
>> - trace_qemu_rdma_write_flush(rdma->nb_sent);
>> + if (migrate_use_multifd()) {
>> + ret = multifd_rdma_write_flush();
>> + if (ret) {
>> + return ret;
>> + }
>> + } else {
>> + rdma->nb_sent++;
>> + trace_qemu_rdma_write_flush(rdma->nb_sent);
>> + }
>> }
>>
>> rdma->current_length = 0;
>> @@ -4062,6 +4104,7 @@ wait_reg_complete:
>> }
>>
>> qemu_sem_post(&multifd_send_param->sem_sync);
>> + qemu_sem_wait(&multifd_send_param->sem);
>
> why?
>
The multifd send thread would post sem signal after finishing sending data.
The main thread need wait for multifd RDMA send threads to poll the CQE.
>> }
>> }
>>
>> @@ -4443,6 +4486,7 @@ static void *multifd_rdma_send_thread(void *opaque)
>> Error *local_err = NULL;
>> int ret = 0;
>> RDMAControlHeader head = { .len = 0, .repeat = 1 };
>> + RDMAContext *rdma = p->rdma;
>>
>> trace_multifd_send_thread_start(p->id);
>> if (multifd_send_initial_packet(p, &local_err) < 0) {
>> @@ -4451,7 +4495,7 @@ static void *multifd_rdma_send_thread(void *opaque)
>>
>> /* wait for semaphore notification to register memory */
>> qemu_sem_wait(&p->sem_sync);
>> - if (qemu_rdma_registration(p->rdma) < 0) {
>> + if (qemu_rdma_registration(rdma) < 0) {
>> goto out;
>> }
>> /*
>> @@ -4466,12 +4510,25 @@ static void *multifd_rdma_send_thread(void *opaque)
>> break;
>> }
>> }
>> + /* To complete polling(CQE) */
>> + while (rdma->nb_sent) {
>
> Where is nb_sent decremented?
>
the nb_sent is decreased in qemu_rdma_poll which is called by
qemu_rdma_block_for_wrid.
>> + ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE,
>> NULL);
>> + if (ret < 0) {
>> + error_report("multifd RDMA migration: "
>> + "complete polling error!");
>> + return NULL;
>> + }
>> + }
>> /* Send FINISHED to the destination */
>> head.type = RDMA_CONTROL_REGISTER_FINISHED;
>> - ret = qemu_rdma_exchange_send(p->rdma, &head, NULL, NULL, NULL,
>> NULL);
>> + ret = qemu_rdma_exchange_send(rdma, &head, NULL, NULL, NULL, NULL);
>> if (ret < 0) {
>> + error_report("multifd RDMA migration: "
>> + "sending remote error!");
>> return NULL;
>> }
>> + /* sync main thread */
>> + qemu_sem_post(&p->sem);
>> }
>>
>> out:
>> --
>> 1.8.3.1
>>
--
Regards.
Chuan
