shithub: neoventi

Info  •  Files  •  Log  •  Branches

disk: fix access to empty buckets

--- a/disk.c

+++ b/disk.c

@@ -77,8 +77,18 @@

 			return 0;

 		}

 	}

-	if(s_sect->bucketmagic && U32GET(buf + 2) != s_sect->bucketmagic)

-		sysfatal("index is corrupt: invalid bucket magic: sect %ux, buck %ux", s_sect->bucketmagic, U32GET(buf + 2));

+	if(s_sect->bucketmagic && U32GET(buf + 2) != s_sect->bucketmagic){

+		if(U32GET(buf+2) != 0)

+			sysfatal("index is corrupt: invalid bucket magic: sect %ux, buck %ux", s_sect->bucketmagic, U32GET(buf + 2));

+		else{

+			cacheunlock(key, bucket & 0xffff);

+			// invoking code should not care about the

+			// distinction between "bucket is empty" and

+			// "bucket has data but not what we want."

+			werrstr("entry not found in bucket");

+			return 0;

+		}

+	}

 	if(!bucketlookup(buf, score, &entry)){

 		cacheunlock(key, bucket & 0xffff);

 		werrstr("entry not found in bucket");

--- a/notebook

+++ b/notebook

@@ -2779,3 +2779,25 @@

 Okay, so, we need bucket initialization code, too.

 Commit as is and add that, then..

+

+venti/write of empty data works, probably because libventi doesn't actually make a call for that, it just shorts it out. okay, let's try writing a single byte.

+

+With some debug prints:

+

+read 8 bytes from connection

+part of packet is missing, branching elsewhere??

+read 2 bytes from connection

+extracting packet into buffer...

+handling packet...

+computing score for write of len 2

+./7.neoventi: index is corrupt: invalid bucket magic: sect c15cf58e, buck 0

+

+Okay, so, the first read misses some of the packet and we need a second one, that's fine. Then we run into the same issue as on the read side: bucket isn't actually initialized.

+

+What's behavior of venti?

+

+On bucket read, if magic is expected but doesn't match, set entry count to zero, rather than rejecting the bucket?? Okay, then. And then presumably on write itll be set.

+

+Okay, so - we see a read (or a write) that accesses an index bucket. That bucket hasn't been initialized. Thus, we can immediately treat the read as a failure, and the write as dirty.

+

+That works fine for reads, at least, though errors aren't propagated to the client correctly.

--- a/server.c

+++ b/server.c

@@ -164,6 +164,7 @@

 		// buffer, and read more into the buffer _after_ it.

 		// Then, repeat.

 		n = read(conn.fd, buf+sz, 0x10000-sz);

+		print("read %d bytes from connection\n", n);

 		sz += n;

 		if(sz == 0)

 			// No data, and none will be coming.

@@ -174,6 +175,7 @@

 			// As long as there's a complete packet, process it.

 			u16int len = U16GET(buf + offset);

 			if(2 + offset + len > sz){

+				print("part of packet is missing, branching elsewhere??\n");

 				// missing part of packet!

 				break;

 			}

@@ -184,7 +186,9 @@

 		// anyways? Don't want allocation churn, but we can just move packetbuf

 		// into VtConn and pass both the source and destination buffer around a lot.

 		// Not important right now.

+			print("extracting packet into buffer...\n");

 			memcpy(packetbuf, buf+offset, 2+len);

+			print("handling packet...\n");

 			vtconnhandle(conn, packetbuf, len);

 			offset += 2+len;

 		}

-- 

⑨