* Michael R. Hines (mrhi...@linux.vnet.ibm.com) wrote:
> On 06/11/2015 12:17 PM, Dr. David Alan Gilbert (git) wrote:
> >From: "Dr. David Alan Gilbert" <dgilb...@redhat.com>
> >
> >The 'offset' field in RDMACompress and 'current_addr' field
> >in RDMARegister are commented as being offsets within a particular
> >RAMBlock, however they appear to actually be offsets within the
> >ram_addr_t space.
> >
> >The code currently assumes that the offsets on the source/destination
> >match, this change removes the need for the assumption for these
> >structures by translating the addresses into the ram_addr_t space of
> >the destination host.
> >
> >Note: An alternative would be to change the fields to actually
> >take the data they're commented for; this would potentially be
> >simpler but would break stream compatibility for those cases
> >that currently work.
> >
> >Signed-off-by: Dr. David Alan Gilbert <dgilb...@redhat.com>
> >---
> > migration/rdma.c | 31 ++++++++++++++++++++++++-------
> > 1 file changed, 24 insertions(+), 7 deletions(-)
> >
> >diff --git a/migration/rdma.c b/migration/rdma.c
> >index 9532461..cb66721 100644
> >--- a/migration/rdma.c
> >+++ b/migration/rdma.c
> >@@ -411,7 +411,7 @@ static void network_to_control(RDMAControlHeader
> >*control)
> > */
> > typedef struct QEMU_PACKED {
> > union QEMU_PACKED {
> >- uint64_t current_addr; /* offset into the ramblock of the chunk */
> >+ uint64_t current_addr; /* offset into the ram_addr_t space */
> > uint64_t chunk; /* chunk to lookup if unregistering */
> > } key;
> > uint32_t current_index; /* which ramblock the chunk belongs to */
> >@@ -419,8 +419,19 @@ typedef struct QEMU_PACKED {
> > uint64_t chunks; /* how many sequential chunks to register
> > */
> > } RDMARegister;
> >
> >-static void register_to_network(RDMARegister *reg)
> >+static void register_to_network(RDMAContext *rdma, RDMARegister *reg)
> > {
> >+ RDMALocalBlock *local_block;
> >+ local_block = &rdma->local_ram_blocks.block[reg->current_index];
> >+
> >+ if (local_block->is_ram_block) {
> >+ /*
> >+ * current_addr as passed in is an address in the local ram_addr_t
> >+ * space, we need to translate this for the destination
> >+ */
> >+ reg->key.current_addr -= local_block->offset;
> >+ reg->key.current_addr +=
> >rdma->dest_blocks[reg->current_index].offset;
> >+ }
> > reg->key.current_addr = htonll(reg->key.current_addr);
> > reg->current_index = htonl(reg->current_index);
> > reg->chunks = htonll(reg->chunks);
> >@@ -436,13 +447,19 @@ static void network_to_register(RDMARegister *reg)
> > typedef struct QEMU_PACKED {
> > uint32_t value; /* if zero, we will madvise() */
> > uint32_t block_idx; /* which ram block index */
> >- uint64_t offset; /* where in the remote ramblock this chunk */
> >+ uint64_t offset; /* Address in remote ram_addr_t space */
> > uint64_t length; /* length of the chunk */
> > } RDMACompress;
> >
> >-static void compress_to_network(RDMACompress *comp)
> >+static void compress_to_network(RDMAContext *rdma, RDMACompress *comp)
> > {
> > comp->value = htonl(comp->value);
> >+ /*
> >+ * comp->offset as passed in is an address in the local ram_addr_t
> >+ * space, we need to translate this for the destination
> >+ */
> >+ comp->offset -= rdma->local_ram_blocks.block[comp->block_idx].offset;
> >+ comp->offset += rdma->dest_blocks[comp->block_idx].offset;
> > comp->block_idx = htonl(comp->block_idx);
> > comp->offset = htonll(comp->offset);
> > comp->length = htonll(comp->length);
>
> So, why add the destination block's offset on the source side
> just for it to be re-adjusted again when it gets to the destination side?
>
> Can you just stop at this:
>
> + reg->key.current_addr -= local_block->offset;
>
> Without this:
>
> + reg->key.current_addr +=
> rdma->dest_blocks[reg->current_index].offset;
>
> ... on the source, followed by this on the destionation:
>
> + comp->offset -= rdma->local_ram_blocks.block[comp->block_idx].offset;
>
> Without this:
>
> + comp->offset += rdma->dest_blocks[comp->block_idx].offset;
>
> Did I follow correctly?
Aren't both of those conversions happening on the source?
Anyway, I think what you're saying is that we change the value sent over
the network to be an offset within the block instead of an offset in
the whole ram_addr_t space (i.e. that's what happens if you don't
add back on the dest_blocks[].offset). As I commented in the commit
message, that would work but it would break compatibility with existing
RDMA migrations since the offset field would now have a different meaning.
Dave
>
> >@@ -1288,7 +1305,7 @@ static int qemu_rdma_unregister_waiting(RDMAContext
> >*rdma)
> > rdma->total_registrations--;
> >
> > reg.key.chunk = chunk;
> >- register_to_network(®);
> >+ register_to_network(rdma, ®);
> > ret = qemu_rdma_exchange_send(rdma, &head, (uint8_t *) ®,
> > &resp, NULL, NULL);
> > if (ret < 0) {
> >@@ -1909,7 +1926,7 @@ retry:
> > trace_qemu_rdma_write_one_zero(chunk, sge.length,
> > current_index,
> > current_addr);
> >
> >- compress_to_network(&comp);
> >+ compress_to_network(rdma, &comp);
> > ret = qemu_rdma_exchange_send(rdma, &head,
> > (uint8_t *) &comp, NULL, NULL, NULL);
> >
> >@@ -1936,7 +1953,7 @@ retry:
> > trace_qemu_rdma_write_one_sendreg(chunk, sge.length,
> > current_index,
> > current_addr);
> >
> >- register_to_network(®);
> >+ register_to_network(rdma, ®);
> > ret = qemu_rdma_exchange_send(rdma, &head, (uint8_t *) ®,
> > &resp, ®_result_idx, NULL);
> > if (ret < 0) {
>
--
Dr. David Alan Gilbert / dgilb...@redhat.com / Manchester, UK
