op public repos

Blob

Date:: Sat Jan 12 16:04:41 2008 UTC
Message:: libdiskfs: fix ext2 directory reading bug (Mike Mammarella)
Actions:: History | Blame | Raw File
1 #include <u.h>
2 #include <libc.h>
3 #include <thread.h>
4 #include <sunrpc.h>
5 #include <nfs3.h>
6 #include <diskfs.h>
7 #include "ext2.h"
8 
9 static void parsedirent(Dirent*, uchar*);
10 static void parseinode(Inode*, uchar*);
11 static void parsegroup(Group*, uchar*);
12 static void parsesuper(Super*, uchar*);
13 
14 #define debug 0
15 
16 static int ext2sync(Fsys*);
17 static void ext2close(Fsys*);
18 static Block* ext2blockread(Fsys*, u64int);
19 
20 static Nfs3Status ext2root(Fsys*, Nfs3Handle*);
21 static Nfs3Status ext2getattr(Fsys*, SunAuthUnix *au, Nfs3Handle*, Nfs3Attr*);
22 static Nfs3Status ext2lookup(Fsys*, SunAuthUnix *au, Nfs3Handle*, char*, Nfs3Handle*);
23 static Nfs3Status ext2readfile(Fsys*, SunAuthUnix *au, Nfs3Handle*, u32int, u64int, uchar**, u32int*, u1int*);
24 static Nfs3Status ext2readlink(Fsys *fsys, SunAuthUnix *au, Nfs3Handle *h, char **link);
25 static Nfs3Status ext2readdir(Fsys *fsys, SunAuthUnix *au, Nfs3Handle *h, u32int, u64int, uchar**, u32int*, u1int*);
26 static Nfs3Status ext2access(Fsys *fsys, SunAuthUnix *au, Nfs3Handle *h, u32int want, u32int *got, Nfs3Attr *attr);
27 
28 Fsys*
29 fsysopenext2(Disk *disk)
30 {
31 	Ext2 *fs;
32 	Fsys *fsys;
33 
34 	fsys = emalloc(sizeof(Fsys));
35 	fs = emalloc(sizeof(Ext2));
36 	fs->disk = disk;
37 	fsys->priv = fs;
38 	fs->fsys = fsys;
39 	fsys->type = "ext2";
40 	fsys->_readblock = ext2blockread;
41 	fsys->_sync = ext2sync;
42 	fsys->_root = ext2root;
43 	fsys->_getattr = ext2getattr;
44 	fsys->_access = ext2access;
45 	fsys->_lookup = ext2lookup;
46 	fsys->_readfile = ext2readfile;
47 	fsys->_readlink = ext2readlink;
48 	fsys->_readdir = ext2readdir;
49 	fsys->_close = ext2close;
50 
51 	if(ext2sync(fsys) < 0)
52 		goto error;
53 
54 	return fsys;
55 
56 error:
57 	ext2close(fsys);
58 	return nil;
59 }
60 
61 static void
62 ext2close(Fsys *fsys)
63 {
64 	Ext2 *fs;
65 
66 	fs = fsys->priv;
67 	free(fs);
68 	free(fsys);
69 }
70 
71 static int
72 ext2group(Ext2 *fs, u32int i, Group *g)
73 {
74 	Block *b;
75 	u64int addr;
76 
77 	if(i >= fs->ngroup)
78 		return -1;
79 
80 	addr = fs->groupaddr + i/fs->descperblock;
81 	b = diskread(fs->disk, fs->blocksize, addr*fs->blocksize);
82 	if(b == nil)
83 		return -1;
84 	parsegroup(g, b->data+i%fs->descperblock*GroupSize);
85 	blockput(b);
86 	return 0;
87 }
88 
89 static Block*
90 ext2blockread(Fsys *fsys, u64int vbno)
91 {
92 	Block *bitb;
93 	Group g;
94 	uchar *bits;
95 	u32int bno, boff, bitblock;
96 	u64int bitpos;
97 	Ext2 *fs;
98 
99 	fs = fsys->priv;
100 	if(vbno >= fs->nblock)
101 		return nil;
102 	bno = vbno;
103 	if(bno != vbno)
104 		return nil;
105 
106 /*	
107 	if(bno < fs->firstblock)
108 		return diskread(fs->disk, fs->blocksize, (u64int)bno*fs->blocksize);
109 */
110 	if(bno < fs->firstblock)
111 		return nil;
112 
113 	bno -= fs->firstblock;
114 	if(ext2group(fs, bno/fs->blockspergroup, &g) < 0){
115 		if(debug)
116 			fprint(2, "loading group: %r...");
117 		return nil;
118 	}
119 /*
120 	if(debug)
121 		fprint(2, "ext2 group %d: bitblock=%ud inodebitblock=%ud inodeaddr=%ud freeblocks=%ud freeinodes=%ud useddirs=%ud\n",
122 			(int)(bno/fs->blockspergroup),
123 			g.bitblock,
124 			g.inodebitblock,
125 			g.inodeaddr,
126 			g.freeblockscount,
127 			g.freeinodescount,
128 			g.useddirscount);
129 	if(debug)
130 		fprint(2, "group %d bitblock=%d...", bno/fs->blockspergroup, g.bitblock);
131 */
132 	bitblock = g.bitblock;
133 	bitpos = (u64int)bitblock*fs->blocksize;
134 
135 	if((bitb = diskread(fs->disk, fs->blocksize, bitpos)) == nil){
136 		if(debug)
137 			fprint(2, "loading bitblock: %r...");
138 		return nil;
139 	}
140 	bits = bitb->data;
141 	boff = bno%fs->blockspergroup;
142 	if((bits[boff>>3] & (1<<(boff&7))) == 0){
143 		if(debug)
144 			fprint(2, "block %d not allocated in group %d: bitblock %d/%lld bits[%d] = %#x\n",
145 				boff, bno/fs->blockspergroup, 
146 				(int)bitblock,
147 				bitpos,
148 				boff>>3,
149 				bits[boff>>3]);
150 		blockput(bitb);
151 		return nil;
152 	}
153 	blockput(bitb);
154 
155 	bno += fs->firstblock;
156 	return diskread(fs->disk, fs->blocksize, (u64int)bno*fs->blocksize);
157 }
158 
159 static Block*
160 ext2datablock(Ext2 *fs, u32int bno, int size)
161 {
162 	USED(size);
163 	return ext2blockread(fs->fsys, bno);
164 }
165 
166 static Block*
167 ext2fileblock(Ext2 *fs, Inode *ino, u32int bno, int size)
168 {
169 	int ppb;
170 	Block *b;
171 	u32int *a;
172 	u32int obno, pbno;
173 
174 	obno = bno;
175 	if(bno < NDIRBLOCKS){
176 		if(debug)
177 			fprint(2, "fileblock %d -> %d...", 
178 				bno, ino->block[bno]);
179 		return ext2datablock(fs, ino->block[bno], size);
180 	}
181 	bno -= NDIRBLOCKS;
182 	ppb = fs->blocksize/4;
183 
184 	/* one indirect */
185 	if(bno < ppb){
186 		b = ext2datablock(fs, ino->block[INDBLOCK], fs->blocksize);
187 		if(b == nil)
188 			return nil;
189 		a = (u32int*)b->data;
190 		bno = a[bno];
191 		blockput(b);
192 		return ext2datablock(fs, bno, size);
193 	}
194 	bno -= ppb;
195 
196 	/* one double indirect */
197 	if(bno < ppb*ppb){
198 		b = ext2datablock(fs, ino->block[DINDBLOCK], fs->blocksize);
199 		if(b == nil)
200 			return nil;
201 		a = (u32int*)b->data;
202 		pbno = a[bno/ppb];
203 		bno = bno%ppb;
204 		blockput(b);
205 		b = ext2datablock(fs, pbno, fs->blocksize);
206 		if(b == nil)
207 			return nil;
208 		a = (u32int*)b->data;
209 		bno = a[bno];
210 		blockput(b);
211 		return ext2datablock(fs, bno, size);
212 	}
213 	bno -= ppb*ppb;
214 
215 	/* one triple indirect */
216 	if(bno < ppb*ppb*ppb){
217 		b = ext2datablock(fs, ino->block[TINDBLOCK], fs->blocksize);
218 		if(b == nil)
219 			return nil;
220 		a = (u32int*)b->data;
221 		pbno = a[bno/(ppb*ppb)];
222 		bno = bno%(ppb*ppb);
223 		blockput(b);
224 		b = ext2datablock(fs, pbno, fs->blocksize);
225 		if(b == nil)
226 			return nil;
227 		a = (u32int*)b->data;
228 		pbno = a[bno/ppb];
229 		bno = bno%ppb;
230 		blockput(b);
231 		b = ext2datablock(fs, pbno, fs->blocksize);
232 		if(b == nil)
233 			return nil;
234 		a = (u32int*)b->data;
235 		bno = a[bno];
236 		blockput(b);
237 		return ext2datablock(fs, bno, size);
238 	}
239 
240 	fprint(2, "ext2fileblock %ud: too big\n", obno);
241 	return nil;
242 }
243 
244 static int
245 checksuper(Super *super)
246 {
247 	if(super->magic != SUPERMAGIC){
248 		werrstr("bad magic 0x%ux wanted 0x%ux", super->magic, SUPERMAGIC);
249 		return -1;
250 	}
251 	return 0;
252 }
253 
254 static int
255 ext2sync(Fsys *fsys)
256 {
257 	int i;
258 	Group g;
259 	Block *b;
260 	Super super;
261 	Ext2 *fs;
262 	Disk *disk;
263 
264 	fs = fsys->priv;
265 	disk = fs->disk;
266 	if((b = diskread(disk, SBSIZE, SBOFF)) == nil)
267 		return -1;
268 	parsesuper(&super, b->data);
269 	blockput(b);
270 	if(checksuper(&super) < 0)
271 		return -1;
272 	fs->blocksize = MINBLOCKSIZE<<super.logblocksize;
273 	fs->nblock = super.nblock;
274 	fs->ngroup = (super.nblock+super.blockspergroup-1)
275 		/ super.blockspergroup;
276 	fs->inospergroup = super.inospergroup;
277 	fs->blockspergroup = super.blockspergroup;
278 	fs->inosperblock = fs->blocksize / InodeSize;
279 	if(fs->blocksize == SBOFF)
280 		fs->groupaddr = 2;
281 	else
282 		fs->groupaddr = 1;
283 	fs->descperblock = fs->blocksize / GroupSize;
284 	fs->firstblock = super.firstdatablock;
285 
286 	fsys->blocksize = fs->blocksize;
287 	fsys->nblock = fs->nblock;
288 	if(debug) fprint(2, "ext2 %d %d-byte blocks, first data block %d, %d groups of %d\n",
289 		fs->nblock, fs->blocksize, fs->firstblock, fs->ngroup, fs->blockspergroup);
290 
291 	if(0){
292 		for(i=0; i<fs->ngroup; i++)
293 			if(ext2group(fs, i, &g) >= 0)
294 				fprint(2, "grp %d: bitblock=%d\n", i, g.bitblock);
295 	}
296 	return 0;
297 }
298 
299 static void
300 mkhandle(Nfs3Handle *h, u64int ino)
301 {
302 	h->h[0] = ino>>24;
303 	h->h[1] = ino>>16;
304 	h->h[2] = ino>>8;
305 	h->h[3] = ino;
306 	h->len = 4;
307 }
308 
309 static u32int
310 byte2u32(uchar *p)
311 {
312 	return (p[0]<<24) | (p[1]<<16) | (p[2]<<8) | p[3];
313 }
314 
315 static Nfs3Status
316 handle2ino(Ext2 *fs, Nfs3Handle *h, u32int *pinum, Inode *ino)
317 {
318 	int i;
319 	uint ioff;
320 	u32int inum;
321 	u32int addr;
322 	Block *b;
323 	Group g;
324 
325 	if(h->len != 4)
326 		return Nfs3ErrBadHandle;
327 	inum = byte2u32(h->h);
328 	if(pinum)
329 		*pinum = inum;
330 	i = (inum-1) / fs->inospergroup;
331 	if(i >= fs->ngroup)
332 		return Nfs3ErrBadHandle;
333 	ioff = (inum-1) % fs->inospergroup;
334 	if(ext2group(fs, i, &g) < 0)
335 		return Nfs3ErrIo;
336 	addr = g.inodeaddr + ioff/fs->inosperblock;
337 	if((b = diskread(fs->disk, fs->blocksize, (u64int)addr*fs->blocksize)) == nil)
338 		return Nfs3ErrIo;
339 	parseinode(ino, b->data+InodeSize*(ioff%fs->inosperblock));
340 	blockput(b);
341 	return Nfs3Ok;
342 }
343 
344 static Nfs3Status
345 ext2root(Fsys *fsys, Nfs3Handle *h)
346 {
347 	USED(fsys);
348 	mkhandle(h, ROOTINODE);
349 	return Nfs3Ok;
350 }
351 
352 static u64int
353 inosize(Inode* ino)
354 {
355 	u64int size;
356 
357 	size = ino->size;
358 	if((ino->mode&IFMT)==IFREG)
359 		size |= (u64int)ino->diracl << 32;
360 	return size;
361 }
362 
363 static Nfs3Status
364 ino2attr(Ext2 *fs, Inode *ino, u32int inum, Nfs3Attr *attr)
365 {
366 	u32int rdev;
367 
368 	attr->type = -1;
369 	switch(ino->mode&IFMT){
370 	case IFIFO:
371 		attr->type = Nfs3FileFifo;
372 		break;
373 	case IFCHR:
374 		attr->type = Nfs3FileChar;
375 		break;
376 	case IFDIR:
377 		attr->type = Nfs3FileDir;
378 		break;
379 	case IFBLK:
380 		attr->type = Nfs3FileBlock;
381 		break;
382 	case IFREG:
383 		attr->type = Nfs3FileReg;
384 		break;
385 	case IFLNK:
386 		attr->type = Nfs3FileSymlink;
387 		break;
388 	case IFSOCK:
389 		attr->type = Nfs3FileSocket;
390 		break;
391 	case IFWHT:
392 	default:
393 		return Nfs3ErrBadHandle;
394 	}
395 
396 	attr->mode = ino->mode&07777;
397 	attr->nlink = ino->nlink;
398 	attr->uid = ino->uid;
399 	attr->gid = ino->gid;
400 	attr->size = inosize(ino);
401 	attr->used = (u64int)ino->nblock*fs->blocksize;
402 	if(attr->type==Nfs3FileBlock || attr->type==Nfs3FileChar){
403 		rdev = ino->block[0];
404 		attr->major = (rdev>>8)&0xFF;
405 		attr->minor = rdev & 0xFFFF00FF;
406 	}else{
407 		attr->major = 0;
408 		attr->minor = 0;
409 	}
410 	attr->fsid = 0;
411 	attr->fileid = inum;
412 	attr->atime.sec = ino->atime;
413 	attr->atime.nsec = 0;
414 	attr->mtime.sec = ino->mtime;
415 	attr->mtime.nsec = 0;
416 	attr->ctime.sec = ino->ctime;
417 	attr->ctime.nsec = 0;
418 	return Nfs3Ok;
419 }
420 
421 static int
422 ingroup(SunAuthUnix *au, uint gid)
423 {
424 	int i;
425 
426 	for(i=0; i<au->ng; i++)
427 		if(au->g[i] == gid)
428 			return 1;
429 	return 0;
430 }
431 
432 static Nfs3Status
433 inoperm(Inode *ino, SunAuthUnix *au, int need)
434 {
435 	int have;
436 
437 	if(allowall)
438 		return Nfs3Ok;
439 
440 	have = ino->mode&0777;
441 	if(ino->uid == au->uid)
442 		have >>= 6;
443 	else if(ino->gid == au->gid || ingroup(au, ino->gid))
444 		have >>= 3;
445 
446 	if((have&need) != need)
447 		return Nfs3ErrNotOwner;	/* really EPERM */
448 	return Nfs3Ok;
449 }
450 
451 static Nfs3Status
452 ext2getattr(Fsys *fsys, SunAuthUnix *au, Nfs3Handle *h, Nfs3Attr *attr)
453 {
454 	Inode ino;
455 	u32int inum;
456 	Ext2 *fs;
457 	Nfs3Status ok;
458 
459 	fs = fsys->priv;
460 	if((ok = handle2ino(fs, h, &inum, &ino)) != Nfs3Ok)
461 		return ok;
462 
463 	USED(au);	/* anyone can getattr */
464 	return ino2attr(fs, &ino, inum, attr);
465 }
466 
467 static Nfs3Status
468 ext2access(Fsys *fsys, SunAuthUnix *au, Nfs3Handle *h, u32int want, u32int *got, Nfs3Attr *attr)
469 {
470 	int have;
471 	Inode ino;
472 	u32int inum;
473 	Ext2 *fs;
474 	Nfs3Status ok;
475 
476 	fs = fsys->priv;
477 	if((ok = handle2ino(fs, h, &inum, &ino)) != Nfs3Ok)
478 		return ok;
479 
480 	have = ino.mode&0777;
481 	if(ino.uid == au->uid)
482 		have >>= 6;
483 	else if(ino.gid == au->gid || ingroup(au, ino.gid))
484 		have >>= 3;
485 
486 	*got = 0;
487 	if((want&Nfs3AccessRead) && (have&AREAD))
488 		*got |= Nfs3AccessRead;
489 	if((want&Nfs3AccessLookup) && (ino.mode&IFMT)==IFDIR && (have&AEXEC))
490 		*got |= Nfs3AccessLookup;
491 	if((want&Nfs3AccessExecute) && (ino.mode&IFMT)!=IFDIR && (have&AEXEC))
492 		*got |= Nfs3AccessExecute;
493 
494 	return ino2attr(fs, &ino, inum, attr);
495 }
496 
497 static Nfs3Status
498 ext2lookup(Fsys *fsys, SunAuthUnix *au, Nfs3Handle *h, char *name, Nfs3Handle *nh)
499 {
500 	u32int nblock;
501 	u32int i;
502 	uchar *p, *ep;
503 	Dirent de;
504 	Inode ino;
505 	Block *b;
506 	Ext2 *fs;
507 	Nfs3Status ok;
508 	int len, want;
509 
510 	fs = fsys->priv;
511 	if((ok = handle2ino(fs, h, nil, &ino)) != Nfs3Ok)
512 		return ok;
513 
514 	if((ino.mode&IFMT) != IFDIR)
515 		return Nfs3ErrNotDir;
516 
517 	if((ok = inoperm(&ino, au, AEXEC)) != Nfs3Ok)
518 		return ok;
519 
520 	len = strlen(name);
521 	nblock = (ino.size+fs->blocksize-1) / fs->blocksize;
522 	if(debug) fprint(2, "%d blocks in dir...", nblock);
523 	for(i=0; i<nblock; i++){
524 		if(i==nblock-1)
525 			want = ino.size % fs->blocksize;
526 		else
527 			want = fs->blocksize;
528 		b = ext2fileblock(fs, &ino, i, want);
529 		if(b == nil){
530 			if(debug) fprint(2, "empty block...");
531 			continue;
532 		}
533 		p = b->data;
534 		ep = p+b->len;
535 		while(p < ep){
536 			parsedirent(&de, p);
537 			if(de.reclen == 0){
538 				if(debug)
539 					fprint(2, "reclen 0 at offset %d of %d\n", (int)(p-b->data), b->len);
540 				break;
541 			}
542 			p += de.reclen;
543 			if(p > ep){
544 				if(debug)
545 					fprint(2, "bad len %d at offset %d of %d\n", de.reclen, (int)(p-b->data), b->len);
546 				break;
547 			}
548 			if(de.ino == 0)
549 				continue;
550 			if(4+2+2+de.namlen > de.reclen){
551 				if(debug)
552 					fprint(2, "bad namelen %d at offset %d of %d\n", de.namlen, (int)(p-b->data), b->len);
553 				break;
554 			}
555 			if(de.namlen == len && memcmp(de.name, name, len) == 0){
556 				mkhandle(nh, de.ino);
557 				blockput(b);
558 				return Nfs3Ok;
559 			}
560 		}
561 		blockput(b);
562 	}
563 	return Nfs3ErrNoEnt;
564 }
565 
566 static Nfs3Status
567 ext2readdir(Fsys *fsys, SunAuthUnix *au, Nfs3Handle *h, u32int count, u64int cookie, uchar **pdata, u32int *pcount, u1int *peof)
568 {
569 	u32int nblock;
570 	u32int i;
571 	int off, outofspace;
572 	uchar *data, *dp, *dep, *p, *ep, *ndp;
573 	Dirent de;
574 	Inode ino;
575 	Block *b;
576 	Ext2 *fs;
577 	Nfs3Status ok;
578 	Nfs3Entry e;
579 	int want;
580 
581 	fs = fsys->priv;
582 	if((ok = handle2ino(fs, h, nil, &ino)) != Nfs3Ok)
583 		return ok;
584 
585 	if((ino.mode&IFMT) != IFDIR)
586 		return Nfs3ErrNotDir;
587 
588 	if((ok = inoperm(&ino, au, AREAD)) != Nfs3Ok)
589 		return ok;
590 
591 	if(debug) print("readdir cookie %#llux ino.size %#llux\n", 
592 		(u64int)cookie, (u64int)ino.size);
593 
594 	if(cookie >= ino.size){
595 		*peof = 1;
596 		*pcount = 0;
597 		*pdata = 0;
598 		return Nfs3Ok;
599 	}
600 
601 	dp = malloc(count);
602 	data = dp;
603 	if(dp == nil)
604 		return Nfs3ErrNoMem;
605 	dep = dp+count;
606 	*peof = 0;
607 	nblock = (ino.size+fs->blocksize-1) / fs->blocksize;
608 	i = cookie/fs->blocksize;
609 	off = cookie%fs->blocksize;
610 	outofspace = 0;
611 	for(; i<nblock && !outofspace; i++, off=0){
612 		if(i==nblock-1)
613 			want = ino.size % fs->blocksize;
614 		else
615 			want = fs->blocksize;
616 		b = ext2fileblock(fs, &ino, i, want);
617 		if(b == nil)
618 			continue;
619 		p = b->data;
620 		ep = p+b->len;
621 		memset(&e, 0, sizeof e);
622 		while(p < ep){
623 			parsedirent(&de, p);
624 			if(de.reclen == 0){
625 				if(debug) fprint(2, "reclen 0 at offset %d of %d\n", (int)(p-b->data), b->len);
626 				break;
627 			}
628 			p += de.reclen;
629 			if(p > ep){
630 				if(debug) fprint(2, "reclen %d at offset %d of %d\n", de.reclen, (int)(p-b->data), b->len);
631 				break;
632 			}
633 			if(de.ino == 0){
634 				if(debug) fprint(2, "zero inode\n");
635 				continue;
636 			}
637 			if(4+2+2+de.namlen > de.reclen){
638 				if(debug) fprint(2, "bad namlen %d reclen %d at offset %d of %d\n", de.namlen, de.reclen, (int)(p-b->data), b->len);
639 				break;
640 			}
641 			if(debug) print("%.*s/%d ", de.namlen, de.name, (int)de.ino);
642 			if(p-de.reclen - b->data < off)
643 				continue;
644 			e.fileid = de.ino;
645 			e.name = de.name;
646 			e.namelen = de.namlen;
647 			e.cookie = (u64int)i*fs->blocksize + (p - b->data);
648 			if(debug) print("%.*s %#llux\n", utfnlen(e.name, e.namelen), e.name, (u64int)e.cookie);
649 			if(nfs3entrypack(dp, dep, &ndp, &e) < 0){
650 				outofspace = 1;
651 				break;
652 			}
653 			dp = ndp;
654 		}
655 		blockput(b);
656 	}
657 	if(i==nblock && !outofspace)
658 		*peof = 1;
659 
660 	*pcount = dp - data;
661 	*pdata = data;
662 	return Nfs3Ok;
663 }
664 
665 static Nfs3Status
666 ext2readfile(Fsys *fsys, SunAuthUnix *au, Nfs3Handle *h, u32int count,
667 	u64int offset, uchar **pdata, u32int *pcount, u1int *peof)
668 {
669 	uchar *data;
670 	Block *b;
671 	Ext2 *fs;
672 	int skip1, tot, want, fragcount;
673 	Inode ino;
674 	Nfs3Status ok;
675 	u64int size;
676 
677 	fs = fsys->priv;
678 	if((ok = handle2ino(fs, h, nil, &ino)) != Nfs3Ok)
679 		return ok;
680 
681 	if((ok = inoperm(&ino, au, AREAD)) != Nfs3Ok)
682 		return ok;
683 
684 	size = inosize(&ino);
685 	if(offset >= size){
686 		*pdata = 0;
687 		*pcount = 0;
688 		*peof = 1;
689 		return Nfs3Ok;
690 	}
691 	if(offset+count > size)
692 		count = size-offset;
693 
694 	data = malloc(count);
695 	if(data == nil)
696 		return Nfs3ErrNoMem;
697 	memset(data, 0, count);
698 
699 	skip1 = offset%fs->blocksize;
700 	offset -= skip1;
701 	want = skip1+count;
702 
703 	/*
704 	 * have to read multiple blocks if we get asked for a big read.
705 	 * Linux NFS client assumes that if you ask for 8k and only get 4k
706 	 * back, the remaining 4k is zeros.
707 	 */
708 	for(tot=0; tot<want; tot+=fragcount){
709 		b = ext2fileblock(fs, &ino, (offset+tot)/fs->blocksize, fs->blocksize);
710 		fragcount = fs->blocksize;
711 		if(b == nil)
712 			continue;
713 		if(tot+fragcount > want)
714 			fragcount = want - tot;
715 		if(tot == 0)
716 			memmove(data, b->data+skip1, fragcount-skip1);
717 		else
718 			memmove(data+tot-skip1, b->data, fragcount);
719 		blockput(b);
720 	}
721 	count = tot - skip1;
722 
723 	*peof = (offset+count == size);
724 	*pcount = count;
725 	*pdata = data;
726 	return Nfs3Ok;
727 }
728 
729 static Nfs3Status
730 ext2readlink(Fsys *fsys, SunAuthUnix *au, Nfs3Handle *h, char **link)
731 {
732 	Ext2 *fs;
733 	Nfs3Status ok;
734 	int len;
735 	Inode ino;
736 	Block *b;
737 
738 	fs = fsys->priv;
739 	if((ok = handle2ino(fs, h, nil, &ino)) != Nfs3Ok)
740 		return ok;
741 	if((ok = inoperm(&ino, au, AREAD)) != Nfs3Ok)
742 		return ok;
743 
744 	if(ino.size > 1024)
745 		return Nfs3ErrIo;
746 	len = ino.size;
747 
748 	if(ino.nblock != 0){
749 		/* BUG: assumes symlink fits in one block */
750 		b = ext2fileblock(fs, &ino, 0, len);
751 		if(b == nil)
752 			return Nfs3ErrIo;
753 		if(memchr(b->data, 0, len) != nil){
754 			blockput(b);
755 			return Nfs3ErrIo;
756 		}
757 		*link = malloc(len+1);
758 		if(*link == 0){
759 			blockput(b);
760 			return Nfs3ErrNoMem;
761 		}
762 		memmove(*link, b->data, len);
763 		(*link)[len] = 0;
764 		blockput(b);
765 		return Nfs3Ok;
766 	}
767 
768 	if(len > sizeof ino.block)
769 		return Nfs3ErrIo;
770 
771 	*link = malloc(len+1);
772 	if(*link == 0)
773 		return Nfs3ErrNoMem;
774 	memmove(*link, ino.block, ino.size);
775 	(*link)[len] = 0;
776 	return Nfs3Ok;
777 }
778 
779 /*
780  * Ext2 is always little-endian, even on big-endian machines.
781  */
782 
783 static u32int
784 l32(uchar *p)
785 {
786 	return p[0] | (p[1]<<8) | (p[2]<<16) | (p[3]<<24);
787 }
788 
789 static u16int
790 l16(uchar *p)
791 {
792 	return p[0] | (p[1]<<8);
793 }
794 
795 static u8int
796 l8(uchar *p)
797 {
798 	return p[0];
799 }
800 
801 static void
802 parsedirent(Dirent *de, uchar *p)
803 {
804 	de->ino = l32(p);
805 	de->reclen = l16(p+4);
806 	de->namlen = l8(p+6);
807 	/* 1 byte pad */
808 	de->name = (char*)p+8;
809 }
810 
811 static void
812 parseinode(Inode *ino, uchar *p)
813 {
814 	int i;
815 
816 	ino->mode = l16(p);
817 	ino->uid = l16(p+2);
818 	ino->size = l32(p+4);
819 	ino->atime = l32(p+8);
820 	ino->ctime = l32(p+12);
821 	ino->mtime = l32(p+16);
822 	ino->dtime = l32(p+20);
823 	ino->gid = l16(p+24);
824 	ino->nlink = l16(p+26);
825 	ino->nblock = l32(p+28);
826 	ino->flags = l32(p+32);
827 	/* 4 byte osd1 */
828 	for(i=0; i<NBLOCKS; i++)
829 		ino->block[i] = l32(p+40+i*4);
830 	ino->version = l32(p+100);
831 	ino->fileacl = l32(p+104);
832 	ino->diracl = l32(p+108);
833 	ino->faddr = l32(p+112);
834 	/* 12 byte osd2 */
835 }
836 
837 static void
838 parsegroup(Group *g, uchar *p)
839 {
840 	g->bitblock = l32(p);
841 	g->inodebitblock = l32(p+4);
842 	g->inodeaddr = l32(p+8);
843 	g->freeblockscount = l16(p+12);
844 	g->freeinodescount = l16(p+14);
845 	g->useddirscount = l16(p+16);
846 	/* 2 byte pad */
847 	/* 12 byte reserved */
848 }
849 
850 static void
851 parsesuper(Super *s, uchar *p)
852 {
853 	s->ninode = l32(p);
854 	s->nblock = l32(p+4);
855 	s->rblockcount = l32(p+8);
856 	s->freeblockcount = l32(p+12);
857 	s->freeinodecount = l32(p+16);
858 	s->firstdatablock = l32(p+20);
859 	s->logblocksize = l32(p+24);
860 	s->logfragsize = l32(p+28);
861 	s->blockspergroup = l32(p+32);
862 	s->fragpergroup = l32(p+36);
863 	s->inospergroup = l32(p+40);
864 	s->mtime = l32(p+44);
865 	s->wtime = l32(p+48);
866 	s->mntcount = l16(p+52);
867 	s->maxmntcount = l16(p+54);
868 	s->magic = l16(p+56);
869 	s->state = l16(p+58);
870 	s->errors = l16(p+60);
871 	/* 2 byte pad */
872 	s->lastcheck = l32(p+64);
873 	s->checkinterval = l32(p+68);
874 	s->creatoros = l32(p+72);
875 	s->revlevel = l32(p+76);
876 	s->defresuid = l16(p+80);
877 	s->defresgid = l16(p+82);
878 	/* 940 byte reserved */
879 }