ref: 466cf20d3524b8e42edc333a6d2df2a01e99a95b
dir: /sys/src/cmd/cwfs/doc/changes/
changes to Ken's file server to make this 63-bit file server
Geoff Collyer
July—October 2004
note: 2⁶³=9,223,372,036,854,775,808 or 8EB (9.2×10¹⁸)
• identified longs that refer to offsets, sizes and block numbers, and
changed them to type Off (vlong); fixed all print formats to match.
fixed byte-swapping for the 'x' config to match.
• fixed VLONG 9p1 message packing and unpacking macros to actually
handle 64-bit offsets and sizes.
• implemented triple-indirect blocks. affected code in
dev/cw.c port/con.c port/dentry.c port/sub.c
port/chk.c port/console.c port/portdat.h
• Fri Aug 6 16:50:59 PDT 2004
; ./sizes
Plan 9 v4 63-bit file server
sizeof(Dentry) = 124
sizeof(Cache) = 88
• added long(er) file name components (56 bytes), long enough for all but one
name in my /.longnames file (68-byte .xml name).
• Fri Aug 6 21:43:41 PDT 2004
; ./sizes
Plan 9 v4 63-bit file server sizes
sizeof(Dentry) = 160
sizeof(Cache) = 88
• touched up lib.h (from libc.h) to bring it up to date with formatting
functions, verbs & flags.
• check now reports stack usage: 320 bytes upon entry to fsck first time,
92 bytes of stack per recursion. given 16000 bytes of stack,
that's 170 recursions maximum.
• booted xtc (terminal) from fs64 (used fs64 as main file system)
note: current file server with triple-indirect blocks at 4k block size
has a maximum file size of ~505GB (5.42×10¹¹).
with quadruple-indirect blocks, max would be ~275TB @ 4k block size.
• got igbe fs driver working (a couple small changes)
• eliminated some gotos (started with 580, down to 454)
• added quadruple indirect blocks: lets us reach 2⁶³ with a 32kB block size
• got igbe boot driver & pxe booting working
• on-disk qid paths are now Offs, but 9p1 qids on the wire are still ulongs
• generalised & parameterised indirect block implementation
• tested with plain w0 fs, cached fake worm on w0, cw jukebox (hp 160fx)
• ip directories in fs & fs64 are identical except for whitespace and
goto-elimination
• replaced most of nemo's ide code with newer ide code lifted from 9load,
then from cpu kernel (sdata.c & support). this brings us dma, rwm & lba48,
finds ide controllers by itself, even pci ones, & copes with dead drives
(i.e., doesn't panic).
• fixed long-standing bug that caused a 5-second delay before each console
prompt on systems without a serial console.
• further type parameterisation: Userid (short), Timet (long), Devsize (vlong).
Comment on v7 kernel portability work, quoting scj & dmr from BSTJ v57
#6 part 2., p. 2038: ``The important data types used within the
system were identified and specified using typedef: disk offsets,
absolute times, internal device names, and the like. This effort was
carried out by K. Thompson.''
• corrected compat.h dependencies in mkfiles
• eliminated all warnings
• implemented truncation via wstat