On Mon, Jun 05, 2023 at 12:22:35 PM -0700, Darrick J. Wong wrote:
> On Mon, Jun 05, 2023 at 02:49:49PM +0530, Chandan Babu R wrote:
>> On Tue, May 23, 2023 at 09:48:07 AM -0700, Darrick J. Wong wrote:
>> > On Tue, May 23, 2023 at 02:30:31PM +0530, Chandan Babu R wrote:
>> >> This commit changes set_cur() to be able to read from external log
>> >> devices. This is required by a future commit which will add the ability to
>> >> dump metadata from external log devices.
>> >>
>> >> Signed-off-by: Chandan Babu R <chandan.babu@xxxxxxxxxx>
>> >> ---
>> >> db/io.c | 22 +++++++++++++++-------
>> >> db/type.c | 2 ++
>> >> db/type.h | 2 +-
>> >> 3 files changed, 18 insertions(+), 8 deletions(-)
>> >>
>> >> diff --git a/db/io.c b/db/io.c
>> >> index 3d2572364..e8c8f57e2 100644
>> >> --- a/db/io.c
>> >> +++ b/db/io.c
>> >> @@ -516,12 +516,13 @@ set_cur(
>> >> int ring_flag,
>> >> bbmap_t *bbmap)
>> >> {
>> >> - struct xfs_buf *bp;
>> >> - xfs_ino_t dirino;
>> >> - xfs_ino_t ino;
>> >> - uint16_t mode;
>> >> + struct xfs_buftarg *btargp;
>> >> + struct xfs_buf *bp;
>> >> + xfs_ino_t dirino;
>> >> + xfs_ino_t ino;
>> >> + uint16_t mode;
>> >> const struct xfs_buf_ops *ops = type ? type->bops : NULL;
>> >> - int error;
>> >> + int error;
>> >>
>> >> if (iocur_sp < 0) {
>> >> dbprintf(_("set_cur no stack element to set\n"));
>> >> @@ -534,7 +535,14 @@ set_cur(
>> >> pop_cur();
>> >> push_cur();
>> >>
>> >> + btargp = mp->m_ddev_targp;
>> >> + if (type->typnm == TYP_ELOG) {
>> >
>> > This feels like a layering violation, see below...
>> >
>> >> + ASSERT(mp->m_ddev_targp != mp->m_logdev_targp);
>> >> + btargp = mp->m_logdev_targp;
>> >> + }
>> >> +
>> >> if (bbmap) {
>> >> + ASSERT(btargp == mp->m_ddev_targp);
>> >> #ifdef DEBUG_BBMAP
>> >> int i;
>> >> printf(_("xfs_db got a bbmap for %lld\n"), (long long)blknum);
>> >> @@ -548,11 +556,11 @@ set_cur(
>> >> if (!iocur_top->bbmap)
>> >> return;
>> >> memcpy(iocur_top->bbmap, bbmap, sizeof(struct bbmap));
>> >> - error = -libxfs_buf_read_map(mp->m_ddev_targp, bbmap->b,
>> >> + error = -libxfs_buf_read_map(btargp, bbmap->b,
>> >> bbmap->nmaps, LIBXFS_READBUF_SALVAGE, &bp,
>> >> ops);
>> >> } else {
>> >> - error = -libxfs_buf_read(mp->m_ddev_targp, blknum, len,
>> >> + error = -libxfs_buf_read(btargp, blknum, len,
>> >> LIBXFS_READBUF_SALVAGE, &bp, ops);
>> >> iocur_top->bbmap = NULL;
>> >> }
>> >> diff --git a/db/type.c b/db/type.c
>> >> index efe704456..cc406ae4c 100644
>> >> --- a/db/type.c
>> >> +++ b/db/type.c
>> >> @@ -100,6 +100,7 @@ static const typ_t __typtab_crc[] = {
>> >> { TYP_INODE, "inode", handle_struct, inode_crc_hfld,
>> >> &xfs_inode_buf_ops, TYP_F_CRC_FUNC, xfs_inode_set_crc },
>> >> { TYP_LOG, "log", NULL, NULL, NULL, TYP_F_NO_CRC_OFF },
>> >> + { TYP_ELOG, "elog", NULL, NULL, NULL, TYP_F_NO_CRC_OFF },
>> >
>> > It strikes me as a little odd to create a new /metadata type/ to
>> > reference the external log. If we someday want to add a bunch of new
>> > types to xfs_db to allow us to decode/fuzz the log contents, wouldn't we
>> > have to add them twice -- once for decoding an internal log, and again
>> > to decode the external log? And the only difference between the two
>> > would be the buftarg, right? The set_cur caller needs to know the
>> > daddr already, so I don't think it's unreasonable for the caller to have
>> > to know which buftarg too.
>> >
>> > IOWs, I think set_cur ought to take the buftarg, the typ_t, and a daddr
>> > as explicit arguments. But maybe others have opinions?
>> >
>> > e.g. rename set_cur to __set_cur and make it take a buftarg, and then:
>> >
>> > int
>> > set_log_cur(
>> > const typ_t *type,
>> > xfs_daddr_t blknum,
>> > int len,
>> > int ring_flag,
>> > bbmap_t *bbmap)
>> > {
>> > if (!mp->m_logdev_targp->bt_bdev ||
>> > mp->m_logdev_targp->bt_bdev == mp->m_ddev_targp->bt_bdev) {
>> > printf(_("external log device not loaded, use -l.\n"));
>> > return ENODEV;
>> > }
>> >
>> > __set_cur(mp->m_logdev_targp, type, blknum, len, ring_flag, bbmap);
>> > return 0;
>> > }
>> >
>> > and then metadump can do something like ....
>> >
>> > error = set_log_cur(&typtab[TYP_LOG], 0,
>> > mp->m_sb.sb_logblocks * blkbb, DB_RING_IGN, NULL);
>> >
>>
>> Darrick, How about implementing the following instead,
>>
>> static void
>> __set_cur(
>> struct xfs_buftarg *btargp,
>> const typ_t *type,
>> xfs_daddr_t blknum,
>> int len,
>> int ring_flag,
>> bbmap_t *bbmap)
>> {
>> struct xfs_buf *bp;
>> xfs_ino_t dirino;
>> xfs_ino_t ino;
>> uint16_t mode;
>> const struct xfs_buf_ops *ops = type ? type->bops : NULL;
>> int error;
>>
>> if (iocur_sp < 0) {
>> dbprintf(_("set_cur no stack element to set\n"));
>> return;
>> }
>>
>> ino = iocur_top->ino;
>> dirino = iocur_top->dirino;
>> mode = iocur_top->mode;
>> pop_cur();
>> push_cur();
>>
>> if (bbmap) {
>> #ifdef DEBUG_BBMAP
>> int i;
>> printf(_("xfs_db got a bbmap for %lld\n"), (long long)blknum);
>> printf(_("\tblock map"));
>> for (i = 0; i < bbmap->nmaps; i++)
>> printf(" %lld:%d", (long long)bbmap->b[i].bm_bn,
>> bbmap->b[i].bm_len);
>> printf("\n");
>> #endif
>> iocur_top->bbmap = malloc(sizeof(struct bbmap));
>> if (!iocur_top->bbmap)
>> return;
>> memcpy(iocur_top->bbmap, bbmap, sizeof(struct bbmap));
>> error = -libxfs_buf_read_map(btargp, bbmap->b,
>> bbmap->nmaps, LIBXFS_READBUF_SALVAGE, &bp,
>> ops);
>> } else {
>> error = -libxfs_buf_read(btargp, blknum, len,
>> LIBXFS_READBUF_SALVAGE, &bp, ops);
>> iocur_top->bbmap = NULL;
>> }
>>
>> /*
>> * Salvage mode means that we still get a buffer even if the verifier
>> * says the metadata is corrupt. Therefore, the only errors we should
>> * get are for IO errors or runtime errors.
>> */
>> if (error)
>> return;
>> iocur_top->buf = bp->b_addr;
>> iocur_top->bp = bp;
>> if (!ops) {
>> bp->b_ops = NULL;
>> bp->b_flags |= LIBXFS_B_UNCHECKED;
>> }
>>
>> iocur_top->bb = blknum;
>> iocur_top->blen = len;
>> iocur_top->boff = 0;
>> iocur_top->data = iocur_top->buf;
>> iocur_top->len = BBTOB(len);
>> iocur_top->off = blknum << BBSHIFT;
>> iocur_top->typ = cur_typ = type;
>> iocur_top->ino = ino;
>> iocur_top->dirino = dirino;
>> iocur_top->mode = mode;
>> iocur_top->ino_buf = 0;
>> iocur_top->dquot_buf = 0;
>>
>> /* store location in ring */
>> if (ring_flag)
>> ring_add();
>> }
>>
>> void
>> set_cur(
>> const typ_t *type,
>> xfs_daddr_t blknum,
>> int len,
>> int ring_flag,
>> bbmap_t *bbmap)
>> {
>> struct xfs_buftarg *btargp = mp->m_ddev_targp;
>>
>> if (type->typnm == TYP_LOG &&
>> mp->m_logdev_targp->bt_bdev != mp->m_ddev_targp->bt_bdev) {
>> ASSERT(mp->m_sb.sb_logstart == 0);
>> btargp = mp->m_logdev_targp;
>> }
>>
>> __set_cur(btargp, type, blknum, len, ring_flag, bbmap);
>> }
>>
>> i.e. We continue to have just one type for the log and set_cur() will
>> internally decide which buftarg to pass to __set_cur(). Please let me know
>> your opinion on this approach.
>
> If I'm understanding this correctly, you're proposing to push the
> buftarg decision down into set_cur instead of encoding it in the typ_t
> information?
>
> I still don't like this, because that decision should be made by the
> callers of set_*cur, not down in the io cursor handling code.
>
> Take a look at the users of set_log_cur and set_rt_cur in the 'dblock'
> command as of djwong-wtf:
> https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfsprogs-dev.git/tree/db/block.c?h=djwong-wtf_2023-06-05#n217
>
> Notice this bit here:
>
> static inline bool
> is_rtfile(
> struct xfs_dinode *dip)
> {
> return dip->di_flags & cpu_to_be16(XFS_DIFLAG_REALTIME);
> }
>
> static int
> dblock_f(...)
> {
> ...
>
> if (is_rtfile(iocur_top->data))
> set_rt_cur(&typtab[type], (int64_t)XFS_FSB_TO_DADDR(mp, dfsbno),
> nb * blkbb, DB_RING_ADD,
> nex > 1 ? &bbmap : NULL);
> else
> set_cur(&typtab[type], (int64_t)XFS_FSB_TO_DADDR(mp, dfsbno),
> nb * blkbb, DB_RING_ADD,
> nex > 1 ? &bbmap : NULL);
>
> xfs_db can now access the data blocks of realtime files, because we have
> the high level logic to decide which buftarg based on the di_flags set
> in the inode core. TYP_DATA doesn't know anything at all about inodes
> or data blocks or whatever -- down at the level of "data block" we don't
> actually have the context we need to select a device.
>
Ok. Now I understand. Thanks for the explaination.
--
chandan
