Tom Lane <tgl@sss.pgh.pa.us> wrote:
> ITAGAKI Takahiro <itagaki.takahiro@oss.ntt.co.jp> writes:
> > - ((double) (int32) (recptr.xrecoff - ckpt_start_recptr.xrecoff)) / XLogSegSize) /
> > + ((double) recptr.xrecoff - (double) ckpt_start_recptr.xrecoff) / XLogSegSize) /
>
> Surely this makes matters worse, not better. What happens near a segment
> boundary crossing?
Here is the dumped progres information by the attached patch
(only for debug purpose).
cur prog. xlog prog. time prog.
[-400ms] 0.030503 0.019247 0.031158 (diff xlogid=0, xrecoff=82665472)
[-200ms] 0.031176 0.019957 0.031839 (diff xlogid=0, xrecoff=85712896)
[*] 0.031860 1.020706 0.032521 (diff xlogid=1, xrecoff=105709568)
> recptr.xrecoff - ckpt_start_recptr.xrecoff
recptr.xrecoff is reset to 0 or so when xlogid is bumped up. At that time,
if ckpt_start_recptr.xrecoff is greater than 2G, we cannot represent
the difference with int32 because the value is less than -2G.
Casting double after int32 does not help us.
We use ( xlogid * 255 * 16MB + xrecoff ) as a base value for the calculation.
If we interprets xrecoff=105709568 at [*] as "minus uint32", we can calcurate
it correctly as below:
without fix with fix
[-400ms] 82665472 82665472 (same)
[-200ms] 85712896 85712896 (same)
[*] 4383899648 88932353 (= 1*255*16MB - (0xFFFFFFFF - 105709568) )
Regards,
---
ITAGAKI Takahiro
NTT Open Source Software Center