---------------------
PatchSet 10341 
Date: 2007/12/29 05:36:13
Author: adri
Branch: s27_adri
Tag: (none) 
Log:
More things to do ..

Members: 
	ADRIAN_TODO:1.1.2.17->1.1.2.18 

Index: squid/ADRIAN_TODO
===================================================================
RCS file: /cvsroot/squid-sf//squid/Attic/ADRIAN_TODO,v
retrieving revision 1.1.2.17
retrieving revision 1.1.2.18
diff -u -r1.1.2.17 -r1.1.2.18
--- squid/ADRIAN_TODO	20 Dec 2007 10:37:04 -0000	1.1.2.17
+++ squid/ADRIAN_TODO	29 Dec 2007 05:36:13 -0000	1.1.2.18
@@ -97,3 +97,31 @@
   (benchmark, test, etc!) and then eliminate these and possibly other messy
   strCDup()s.
 
+
+* The disk.c code and all the store disk write callbacks expect a FREE * pointer to free
+  the written buffer. This is not enough to write and free an stmem ref. Gah, that hack
+  is/was ugly! The storeWrite() path has an assert in front of it now; please go through
+  and add completion and buffer free callbacks; rework disk.c to do the buffer free
+  callbacks "right" (just to keep src/fs/ufs happy for now) and then redo all the FSes.
+
+* Eventually! Collapse ufs/aufs/diskd into "ufs", with different IO op vectors.
+  Like Robert's Squid-3 work, but "better" - make it work with disk.c for gods sake,
+  then see if its feasable to do use file_write() for the log files again with stuff
+  like POSIX aio/diskd/asyncio doing disk io. It might not be..
+
+  (in any case, bring sanity to the disk code!)
+
+* Audit all the storeClientRef() users (again!) and see if they're able to handle buffers
+  larger than SM_PAGE_SIZE. The stmem stuff can now handle arbitrary-sized buffers and
+  these should be read/written as much as possible.
+
+* .. then, add a "header size" swap meta type (if one doesn't exist already) and then
+  modify the store client code to read in -that- much data (in multiple reads if needed)
+  before parsing the headers. Yes, at this point we can handle >4k reply status+headers..
+
+* Hm, what else? Ah, buf_fill() does a read() where it should do an FD_READ_METHOD
+  and update the comm counters. Look at moving buf_fill() into src/ so it can access
+  the FD_READ_METHOD and associated gunk (don't make buf depend on comm and comm depend
+  on buf; have a "netbuf" library which depends on both. No cyclic dependencies!)
+
+