thebestvibedev/tromcho.net
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Implement RDMA receive with memory copying (send remains zero-copy)

Browse Source 
This is the simplest and, as usual, the best implementation :)

100% zero-copy implementation is also possible (see rdma-zerocopy branch),
but it requires to create A LOT of queues (~128 per client) to use QPN as a 'tag'
because of the lack of receive tags and the server may simply run out of queues.
Hardware limit is 262144 on Mellanox ConnectX-4 which amounts to only 2048
'connections' per host. And even with that amount of queues it's still less optimal
than the non-zerocopy one.

In fact, newest hardware like Mellanox ConnectX-5 does have Tag Matching
support, but it's still unsuitable for us because it doesn't support scatter/gather
(tm_caps.max_sge=1).
This commit is contained in:
Vitaliy Filippov
2021-04-29 02:34:45 +03:00
parent 9e6cbc6ebc
commit 971aa4ae4f
7 changed files with 94 additions and 150 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/msgr_receive.cpp
+48 -41
View File
@@ -91,48 +91,9 @@ bool osd_messenger_t::handle_read(int result, osd_client_t *cl)
{
if (cl->read_iov.iov_base == cl->in_buf)
{
// Compose operation(s) from the buffer
int remain = result;
void *curbuf = cl->in_buf;
while (remain > 0)
if (!handle_read_buffer(cl, cl->in_buf, result))
{
if (!cl->read_op)
{
cl->read_op = new osd_op_t;
cl->read_op->peer_fd = cl->peer_fd;
cl->read_op->op_type = OSD_OP_IN;
cl->recv_list.push_back(cl->read_op->req.buf, OSD_PACKET_SIZE);
cl->read_remaining = OSD_PACKET_SIZE;
cl->read_state = CL_READ_HDR;
}
while (cl->recv_list.done < cl->recv_list.count && remain > 0)
{
iovec* cur = cl->recv_list.get_iovec();
if (cur->iov_len > remain)
{
memcpy(cur->iov_base, curbuf, remain);
cl->read_remaining -= remain;
cur->iov_len -= remain;
cur->iov_base += remain;
remain = 0;
}
else
{
memcpy(cur->iov_base, curbuf, cur->iov_len);
curbuf += cur->iov_len;
cl->read_remaining -= cur->iov_len;
remain -= cur->iov_len;
cur->iov_len = 0;
cl->recv_list.done++;
}
}
if (cl->recv_list.done >= cl->recv_list.count)
{
if (!handle_finished_read(cl))
{
goto fin;
}
}
goto fin;
}
}
else
@@ -159,6 +120,52 @@ fin:
return ret;
}
bool osd_messenger_t::handle_read_buffer(osd_client_t *cl, void *curbuf, int remain)
{
// Compose operation(s) from the buffer
while (remain > 0)
{
if (!cl->read_op)
{
cl->read_op = new osd_op_t;
cl->read_op->peer_fd = cl->peer_fd;
cl->read_op->op_type = OSD_OP_IN;
cl->recv_list.push_back(cl->read_op->req.buf, OSD_PACKET_SIZE);
cl->read_remaining = OSD_PACKET_SIZE;
cl->read_state = CL_READ_HDR;
}
while (cl->recv_list.done < cl->recv_list.count && remain > 0)
{
iovec* cur = cl->recv_list.get_iovec();
if (cur->iov_len > remain)
{
memcpy(cur->iov_base, curbuf, remain);
cl->read_remaining -= remain;
cur->iov_len -= remain;
cur->iov_base += remain;
remain = 0;
}
else
{
memcpy(cur->iov_base, curbuf, cur->iov_len);
curbuf += cur->iov_len;
cl->read_remaining -= cur->iov_len;
remain -= cur->iov_len;
cur->iov_len = 0;
cl->recv_list.done++;
}
}
if (cl->recv_list.done >= cl->recv_list.count)
{
if (!handle_finished_read(cl))
{
return false;
}
}
}
return true;
}
bool osd_messenger_t::handle_finished_read(osd_client_t *cl)
{
cl->recv_list.reset();