Hi all,
I've been seeing the same sort of issues repoted by Ionut and as addressed by this patch: http://lists.denx.de/pipermail/u-boot/2014-January/171459.html
That patch was added in commit fc0fc50 and reverted in commit 715b56f.
It no longer applies cleanly, and when I tried to resurrect it, I saw errors traversing directories and perhaps something went wrong with my merge.
Ionut, do you have a current version of this patch?
When I looked a little further, I found that a read of a ~150 MiB was around 30x slower than typical, and that the **same** 8 blocks accounted for the majority of the reads.
In fact, these same blocks were read back-to-back.
The following is a quick picture of the output from a simple printf in mmc_bread of the block number and count during a load of the problem file.
~$ uniq -c < mmc_bread.log | sort -n 1 mmc_bread: 0/1 1 mmc_bread: 2293760/1 1 mmc_bread: 2293762/2 1 mmc_bread: 2293768/1 1 mmc_bread: 2293768/1 1 mmc_bread: 2293768/1 1 mmc_bread: 2293768/1 1 mmc_bread: 2295264/1 1 mmc_bread: 2295270/1 1 mmc_bread: 2295290/1 1 mmc_bread: 2295645/1 1 mmc_bread: 7536640/131072 1 mmc_bread: 7667712/32768 1 mmc_bread: 7700480/16384 1 mmc_bread: 7729152/65536 1 mmc_bread: 7798784/130722 1 mmc_bread: 7929506/1 1 mmc_init: 0, time 129 2 mmc_bread: 0/1 6 mmc_bread: 2359120/1 10 mmc_bread: 2358752/1 34 mmc_bread: 2358808/1 2048 mmc_bread: 2557936/8 4096 mmc_bread: 2557936/8 8193 mmc_bread: 2557936/8 16340 mmc_bread: 2557936/8 16384 mmc_bread: 2557936/8
~$ sort < mmc_bread.log | uniq -c | sort -n | tail -n 1 47061 mmc_bread: 2557936/8
In English, the 8 blocks starting at 2557936 are read 47061 times, and back to back in large (2k/4k/16k) bunches, so a very simple **single** block cache of the last read will fix the speed issue and I hacked something up to verify that.
Is anybody else working on things in this area?
I think this is something that's probably easier to fix at the block device level rather than within the ext4 filesystem code.
That said, the 2k/4k/16k bunches above may also indicate a simpler fix in the ext4 code.
Please chime in with your thoughts.
Regards,
Eric Nelson