There was discussion about how to do 64-bit ABI conversation, and an
assertion that it might be hard to do ext2fs_block_iterate3. So I
decided give an example of how I believe the most efficient way to do
this sort of thing.
Each interface when coverted over to 64-bit should be its own patch; for
example, this one adds a new 64-bit version of ext2fs_block_iterate3.
The patch simply modified ext2fs_block_iterate2() to be
ext2fs_block_iterate3(), changing types as needed.
I then created a backwards compatibility function for
ext2fs_block_iterate2() which calls ext2fs_block_iterate3().
Total time to create the initial version of the patch was about 10
minutes. I then ran it through the (32-bit only) regression test, which
flagged some failures because I didn't quite do the first version of the
patch correctly. (I started the patch by doing a global search and
replace of blk_t with blk64_t, and that screwed up the code which
depended on the on-disk format of an indirect block really being a
32-bit blk_t.) Total time for a tested, functional patch was about 30
minutes.
- Ted
commit 54f130e7623a2a5affb0c27f5f4a42d8ace3acea
Author: Theodore Ts'o <[email protected]>
Date: Fri Oct 10 16:36:22 2008 -0400
libext2fs: Add ext2fs_block_iterate3() which has 64-bit support
Signed-off-by: "Theodore Ts'o" <[email protected]>
diff --git a/lib/ext2fs/block.c b/lib/ext2fs/block.c
index b19c450..eafb4cb 100644
--- a/lib/ext2fs/block.c
+++ b/lib/ext2fs/block.c
@@ -21,9 +21,9 @@
struct block_context {
ext2_filsys fs;
int (*func)(ext2_filsys fs,
- blk_t *blocknr,
+ blk64_t *blocknr,
e2_blkcnt_t bcount,
- blk_t ref_blk,
+ blk64_t ref_blk,
int ref_offset,
void *priv_data);
e2_blkcnt_t bcount;
@@ -62,13 +62,17 @@ static int block_iterate_ind(blk_t *ind_block, blk_t ref_block,
int ret = 0, changed = 0;
int i, flags, limit, offset;
blk_t *block_nr;
+ blk64_t blk64;
limit = ctx->fs->blocksize >> 2;
if (!(ctx->flags & BLOCK_FLAG_DEPTH_TRAVERSE) &&
- !(ctx->flags & BLOCK_FLAG_DATA_ONLY))
- ret = (*ctx->func)(ctx->fs, ind_block,
+ !(ctx->flags & BLOCK_FLAG_DATA_ONLY)) {
+ blk64 = *ind_block;
+ ret = (*ctx->func)(ctx->fs, &blk64,
BLOCK_COUNT_IND, ref_block,
ref_offset, ctx->priv_data);
+ *ind_block = blk64;
+ }
check_for_ro_violation_return(ctx, ret);
if (!*ind_block || (ret & BLOCK_ABORT)) {
ctx->bcount += limit;
@@ -91,9 +95,11 @@ static int block_iterate_ind(blk_t *ind_block, blk_t ref_block,
offset = 0;
if (ctx->flags & BLOCK_FLAG_APPEND) {
for (i = 0; i < limit; i++, ctx->bcount++, block_nr++) {
- flags = (*ctx->func)(ctx->fs, block_nr, ctx->bcount,
+ blk64 = *block_nr;
+ flags = (*ctx->func)(ctx->fs, &blk64, ctx->bcount,
*ind_block, offset,
ctx->priv_data);
+ *block_nr = blk64;
changed |= flags;
if (flags & BLOCK_ABORT) {
ret |= BLOCK_ABORT;
@@ -105,9 +111,11 @@ static int block_iterate_ind(blk_t *ind_block, blk_t ref_block,
for (i = 0; i < limit; i++, ctx->bcount++, block_nr++) {
if (*block_nr == 0)
continue;
- flags = (*ctx->func)(ctx->fs, block_nr, ctx->bcount,
+ blk64 = *block_nr;
+ flags = (*ctx->func)(ctx->fs, &blk64, ctx->bcount,
*ind_block, offset,
ctx->priv_data);
+ *block_nr = blk64;
changed |= flags;
if (flags & BLOCK_ABORT) {
ret |= BLOCK_ABORT;
@@ -125,10 +133,13 @@ static int block_iterate_ind(blk_t *ind_block, blk_t ref_block,
}
if ((ctx->flags & BLOCK_FLAG_DEPTH_TRAVERSE) &&
!(ctx->flags & BLOCK_FLAG_DATA_ONLY) &&
- !(ret & BLOCK_ABORT))
- ret |= (*ctx->func)(ctx->fs, ind_block,
+ !(ret & BLOCK_ABORT)) {
+ blk64 = *ind_block;
+ ret |= (*ctx->func)(ctx->fs, &blk64,
BLOCK_COUNT_IND, ref_block,
ref_offset, ctx->priv_data);
+ *ind_block = blk64;
+ }
check_for_ro_violation_return(ctx, ret);
return ret;
}
@@ -139,13 +150,17 @@ static int block_iterate_dind(blk_t *dind_block, blk_t ref_block,
int ret = 0, changed = 0;
int i, flags, limit, offset;
blk_t *block_nr;
+ blk64_t blk64;
limit = ctx->fs->blocksize >> 2;
if (!(ctx->flags & (BLOCK_FLAG_DEPTH_TRAVERSE |
- BLOCK_FLAG_DATA_ONLY)))
- ret = (*ctx->func)(ctx->fs, dind_block,
+ BLOCK_FLAG_DATA_ONLY))) {
+ blk64 = *dind_block;
+ ret = (*ctx->func)(ctx->fs, &blk64,
BLOCK_COUNT_DIND, ref_block,
ref_offset, ctx->priv_data);
+ *dind_block = blk64;
+ }
check_for_ro_violation_return(ctx, ret);
if (!*dind_block || (ret & BLOCK_ABORT)) {
ctx->bcount += limit*limit;
@@ -204,10 +219,13 @@ static int block_iterate_dind(blk_t *dind_block, blk_t ref_block,
}
if ((ctx->flags & BLOCK_FLAG_DEPTH_TRAVERSE) &&
!(ctx->flags & BLOCK_FLAG_DATA_ONLY) &&
- !(ret & BLOCK_ABORT))
- ret |= (*ctx->func)(ctx->fs, dind_block,
+ !(ret & BLOCK_ABORT)) {
+ blk64 = *dind_block;
+ ret |= (*ctx->func)(ctx->fs, &blk64,
BLOCK_COUNT_DIND, ref_block,
ref_offset, ctx->priv_data);
+ *dind_block = blk64;
+ }
check_for_ro_violation_return(ctx, ret);
return ret;
}
@@ -218,13 +236,17 @@ static int block_iterate_tind(blk_t *tind_block, blk_t ref_block,
int ret = 0, changed = 0;
int i, flags, limit, offset;
blk_t *block_nr;
+ blk64_t blk64;
limit = ctx->fs->blocksize >> 2;
if (!(ctx->flags & (BLOCK_FLAG_DEPTH_TRAVERSE |
- BLOCK_FLAG_DATA_ONLY)))
- ret = (*ctx->func)(ctx->fs, tind_block,
+ BLOCK_FLAG_DATA_ONLY))) {
+ blk64 = *tind_block;
+ ret = (*ctx->func)(ctx->fs, &blk64,
BLOCK_COUNT_TIND, ref_block,
ref_offset, ctx->priv_data);
+ *tind_block = blk64;
+ }
check_for_ro_violation_return(ctx, ret);
if (!*tind_block || (ret & BLOCK_ABORT)) {
ctx->bcount += limit*limit*limit;
@@ -283,22 +305,25 @@ static int block_iterate_tind(blk_t *tind_block, blk_t ref_block,
}
if ((ctx->flags & BLOCK_FLAG_DEPTH_TRAVERSE) &&
!(ctx->flags & BLOCK_FLAG_DATA_ONLY) &&
- !(ret & BLOCK_ABORT))
- ret |= (*ctx->func)(ctx->fs, tind_block,
+ !(ret & BLOCK_ABORT)) {
+ blk64 = *tind_block;
+ ret |= (*ctx->func)(ctx->fs, &blk64,
BLOCK_COUNT_TIND, ref_block,
ref_offset, ctx->priv_data);
+ *tind_block = blk64;
+ }
check_for_ro_violation_return(ctx, ret);
return ret;
}
-errcode_t ext2fs_block_iterate2(ext2_filsys fs,
+errcode_t ext2fs_block_iterate3(ext2_filsys fs,
ext2_ino_t ino,
int flags,
char *block_buf,
int (*func)(ext2_filsys fs,
- blk_t *blocknr,
+ blk64_t *blocknr,
e2_blkcnt_t blockcnt,
- blk_t ref_blk,
+ blk64_t ref_blk,
int ref_offset,
void *priv_data),
void *priv_data)
@@ -309,6 +334,7 @@ errcode_t ext2fs_block_iterate2(ext2_filsys fs,
errcode_t retval;
struct block_context ctx;
int limit;
+ blk64_t blk64;
EXT2_CHECK_MAGIC(fs, EXT2_ET_MAGIC_EXT2FS_FILSYS);
@@ -348,10 +374,11 @@ errcode_t ext2fs_block_iterate2(ext2_filsys fs,
if ((fs->super->s_creator_os == EXT2_OS_HURD) &&
!(flags & BLOCK_FLAG_DATA_ONLY)) {
if (inode.osd1.hurd1.h_i_translator) {
- ret |= (*ctx.func)(fs,
- &inode.osd1.hurd1.h_i_translator,
+ blk64 = inode.osd1.hurd1.h_i_translator;
+ ret |= (*ctx.func)(fs, &blk64,
BLOCK_COUNT_TRANSLATOR,
0, 0, priv_data);
+ inode.osd1.hurd1.h_i_translator = (blk_t) blk64;
if (ret & BLOCK_ABORT)
goto abort_exit;
check_for_ro_violation_goto(&ctx, ret, abort_exit);
@@ -362,7 +389,7 @@ errcode_t ext2fs_block_iterate2(ext2_filsys fs,
ext2_extent_handle_t handle;
struct ext2fs_extent extent;
e2_blkcnt_t blockcnt = 0;
- blk_t blk, new_blk;
+ blk64_t blk, new_blk;
int op = EXT2_EXTENT_ROOT;
unsigned int j;
@@ -432,7 +459,7 @@ errcode_t ext2fs_block_iterate2(ext2_filsys fs,
ctx.errcode =
ext2fs_extent_set_bmap(handle,
(blk64_t) blockcnt,
- (blk64_t) new_blk, 0);
+ new_blk, 0);
if (ctx.errcode)
break;
}
@@ -452,15 +479,17 @@ errcode_t ext2fs_block_iterate2(ext2_filsys fs,
*/
for (i = 0; i < EXT2_NDIR_BLOCKS ; i++, ctx.bcount++) {
if (inode.i_block[i] || (flags & BLOCK_FLAG_APPEND)) {
- ret |= (*ctx.func)(fs, &inode.i_block[i],
- ctx.bcount, 0, i, priv_data);
+ blk64 = inode.i_block[i];
+ ret |= (*ctx.func)(fs, &blk64, ctx.bcount, 0, i,
+ priv_data);
+ inode.i_block[i] = (blk_t) blk64;
if (ret & BLOCK_ABORT)
goto abort_exit;
}
}
check_for_ro_violation_goto(&ctx, ret, abort_exit);
if (inode.i_block[EXT2_IND_BLOCK] || (flags & BLOCK_FLAG_APPEND)) {
- ret |= block_iterate_ind(&inode.i_block[EXT2_IND_BLOCK],
+ ret |= block_iterate_ind(&inode.i_block[EXT2_IND_BLOCK],
0, EXT2_IND_BLOCK, &ctx);
if (ret & BLOCK_ABORT)
goto abort_exit;
@@ -497,6 +526,52 @@ errout:
* Emulate the old ext2fs_block_iterate function!
*/
+struct xlate64 {
+ int (*func)(ext2_filsys fs,
+ blk_t *blocknr,
+ e2_blkcnt_t blockcnt,
+ blk_t ref_blk,
+ int ref_offset,
+ void *priv_data);
+ void *real_private;
+};
+
+static int xlate64_func(ext2_filsys fs, blk64_t *blocknr,
+ e2_blkcnt_t blockcnt, blk64_t ref_blk,
+ int ref_offset, void *priv_data)
+{
+ struct xlate64 *xl = (struct xlate64 *) priv_data;
+ int ret;
+ blk_t block32 = *blocknr;
+
+ ret = (*xl->func)(fs, &block32, blockcnt, (blk_t) ref_blk, ref_offset,
+ xl->real_private);
+ *blocknr = block32;
+ return ret;
+}
+
+errcode_t ext2fs_block_iterate2(ext2_filsys fs,
+ ext2_ino_t ino,
+ int flags,
+ char *block_buf,
+ int (*func)(ext2_filsys fs,
+ blk_t *blocknr,
+ e2_blkcnt_t blockcnt,
+ blk_t ref_blk,
+ int ref_offset,
+ void *priv_data),
+ void *priv_data)
+{
+ struct xlate64 xl;
+
+ xl.real_private = priv_data;
+ xl.func = func;
+
+ return ext2fs_block_iterate3(fs, ino, flags, block_buf,
+ xlate64_func, &xl);
+}
+
+
struct xlate {
int (*func)(ext2_filsys fs,
blk_t *blocknr,
diff --git a/lib/ext2fs/ext2fs.h b/lib/ext2fs/ext2fs.h
index e02a55f..aa58c23 100644
--- a/lib/ext2fs/ext2fs.h
+++ b/lib/ext2fs/ext2fs.h
@@ -703,6 +703,17 @@ errcode_t ext2fs_block_iterate2(ext2_filsys fs,
int ref_offset,
void *priv_data),
void *priv_data);
+errcode_t ext2fs_block_iterate3(ext2_filsys fs,
+ ext2_ino_t ino,
+ int flags,
+ char *block_buf,
+ int (*func)(ext2_filsys fs,
+ blk64_t *blocknr,
+ e2_blkcnt_t blockcnt,
+ blk64_t ref_blk,
+ int ref_offset,
+ void *priv_data),
+ void *priv_data);
/* bmap.c */
extern errcode_t ext2fs_bmap(ext2_filsys fs, ext2_ino_t ino,
On Fri, Oct 10, 2008 at 04:43:43PM -0400, Theodore Ts'o wrote:
>
> There was discussion about how to do 64-bit ABI conversation, and an
> assertion that it might be hard to do ext2fs_block_iterate3. So I
> decided give an example of how I believe the most efficient way to do
> this sort of thing.
I'm afraid there's been a misunderstanding; I wasn't saying it was
hard, just pointing out that different versions of interfaces are
profilerating as a result of continuing to support the old ABI at the
same time as the new 64-bit ABI - perhaps more versions of more
functions than was originally anticipated. (I wrote a similar patch
and I agree it's not hard, just more cruft to maintain.)
Eagerly awaiting ext2fs_block_iterate4,
-VAL
On Fri, Oct 10, 2008 at 06:02:49PM -0400, Valerie Aurora Henson wrote:
> On Fri, Oct 10, 2008 at 04:43:43PM -0400, Theodore Ts'o wrote:
> >
> > There was discussion about how to do 64-bit ABI conversation, and an
> > assertion that it might be hard to do ext2fs_block_iterate3. So I
> > decided give an example of how I believe the most efficient way to do
> > this sort of thing.
>
> I'm afraid there's been a misunderstanding; I wasn't saying it was
> hard, just pointing out that different versions of interfaces are
> profilerating as a result of continuing to support the old ABI at the
> same time as the new 64-bit ABI - perhaps more versions of more
> functions than was originally anticipated. (I wrote a similar patch
> and I agree it's not hard, just more cruft to maintain.)
But it's not that much more cruft to maintain. Take a look at my
patch; the original interface gets implemented on top of the old
interface, so the amount of extra cruft to maintain is minimal. The
implementation of ext2fs_block_iterate2 simply becomes:
struct xlate64 xl;
xl.real_private = priv_data;
xl.func = func;
return ext2fs_block_iterate3(fs, ino, flags, block_buf,
xlate64_func, &xl);
That wasn't that bad, was it?
> Eagerly awaiting ext2fs_block_iterate4,
If there are other changes that we need to implement at the same time,
if we haven't done a release yet, we can fold those changes into the
ext2fs_block_iterate3.
- Ted