Re: [Patch] chunkd: use port xxx82 to build

[Pete Zaitcev <zaitcev@redhat.com>]

On Fri, 7 Aug 2009 19:36:28 -0400 (EDT), Rick Peralta <fbp@tiac.net> wrote:

> I was looking from net_write() & net_read() to see where they were
> called and what was passed to them.  struct chunksrv_req & struct
> chunksrv_resp_get seem to be the primary structures.

I didn't know how serious you were about it, so I went ahead and
looked at the endian stuff. The immediate issue - the hang of
basic-object - was addressed with the following fix:

diff --git a/server/object.c b/server/object.c
index 4eb62f7..6ea1500 100644
--- a/server/object.c
+++ b/server/object.c
@@ -290,7 +290,7 @@ bool object_get(struct client *cli, bool want_body)
 
 	cli->in_len = obj->size;
 
-	resp->req.data_len = GUINT32_TO_LE(obj->size);
+	resp->req.data_len = GUINT64_TO_LE((uint64_t)obj->size);
 	memcpy(resp->req.checksum, obj->hashstr, sizeof(obj->hashstr));
 	resp->req.checksum[sizeof(obj->hashstr)] = 0;
 	resp->mtime = GUINT64_TO_LE(obj->mtime);
diff --git a/server/server.c b/server/server.c
index 2ad3db2..161dc67 100644
--- a/server/server.c
+++ b/server/server.c
@@ -639,7 +639,7 @@ static bool cli_resp_xml(struct client *cli, GList *content)
 {
 	int rc;
 	bool rcb;
-	int content_len = strlist_len(content);
+	size_t content_len = strlist_len(content);
 	struct chunksrv_req *resp = NULL;
 
 	resp = malloc(sizeof(*resp));
@@ -650,7 +650,7 @@ static bool cli_resp_xml(struct client *cli, GList *content)
 
 	memcpy(resp, &cli->creq, sizeof(cli->creq));
 
-	resp->data_len = GUINT64_TO_LE(content_len);
+	resp->data_len = GUINT64_TO_LE((uint64_t)content_len);
 
 	cli->state = evt_recycle;
 

We have two problems above:

 #1 the result of byteswap must not be allowed to be type-promoted,
    sign-extended, or otherwise changed by the language. Only bit-to-bit
    assignment will do. Therefore, we must only use 64-bit conversions
    when assigning to req->data_len.

 #2 for some reason swapping of a 32-bit value does not work either,
    I suspect clipping when shifting. So I had to fix that too.

With the above, basic-object passes, test advance further, and then
large-object fails like so:

test failed on line 134

        rcb = stc_get_start(stc, key, &rfd, &len);
        OK(len == (N_BUFS * BUFSZ));   <------- fails. how much is len?

Curiously, it fails on 32-bit PPC only and works on ppc64.

With this, I'm going to quit stepping on your toes and go work on
CLD interface for Chunkd, then data replication in tabled.
Sorry for the interference.

Good luck,
-- Pete