GitLab

Enabling continuous read support implies several changes which must be
done atomically in order to keep the code base consistent and
bisectable.

1/ Retrieving bitflips differently

Improve the helper retrieving the number of bitflips to support the case
where many pages have been read instead of just one. In this case, if
there is one page with bitflips, we cannot know the detail and just get
the information of the maximum number of bitflips corrected in the most
corrupted chunk. Compatible Macronix flashes return:
- the ECC status for the last page read (bits 0-3),
- the amount of bitflips for the whole read operation (bits 4-7).
Hence, when reading two consecutive pages, if there was 2 bits corrected
at most in one chunk, we return this amount times (arbitrary) the number
read pages. It is probably a very pessimistic calculation in most cases,
but still less pessimistic than if we multiplied this amount by the
number of chunks. Anyway, this is just for statistics, the important
data is the maximum amount of bitflips, which leads to wear leveling.

2/ Configuring, enabling and disabling the feature

Create an init function for allocating a vendor structure. Use this
vendor structure to cache the internal continuous read state. The state
is being used to discriminate between the two bitflips retrieval
methods. Finally, helpers for enabling and disabling sequential reads
are also created.

3/ Fill the chips table

Flag all the chips supporting the feature with the ->set_cont_read()
helper.

In order to validate the changes, I modified the mtd-utils test suite
with extended versions of nandbiterrs, nanddump and flash_speed in order
to support, test and benchmark continuous reads. I also ran all the UBI
tests successfully.

The nandbiterrs tool allows to track the ECC efficiency and
feedback. Here is its default output (stripped):

Successfully corrected 0 bit errors per subpage
Read reported 1 corrected bit errors
Successfully corrected 1 bit errors per subpage
Read reported 2 corrected bit errors
Successfully corrected 2 bit errors per subpage
Read reported 3 corrected bit errors
Successfully corrected 3 bit errors per subpage
Read reported 4 corrected bit errors
Successfully corrected 4 bit errors per subpage
Read reported 5 corrected bit errors
Successfully corrected 5 bit errors per subpage
Read reported 6 corrected bit errors
Successfully corrected 6 bit errors per subpage
Read reported 7 corrected bit errors
Successfully corrected 7 bit errors per subpage
Read reported 8 corrected bit errors
Successfully corrected 8 bit errors per subpage
Failed to recover 1 bitflips
Read error after 9 bit errors per page

The output using the continuous option over two pages (the second page
is kept intact):

Successfully corrected 0 bit errors per subpage
Read reported 2 corrected bit errors
Successfully corrected 1 bit errors per subpage
Read reported 4 corrected bit errors
Successfully corrected 2 bit errors per subpage
Read reported 6 corrected bit errors
Successfully corrected 3 bit errors per subpage
Read reported 8 corrected bit errors
Successfully corrected 4 bit errors per subpage
Read reported 10 corrected bit errors
Successfully corrected 5 bit errors per subpage
Read reported 12 corrected bit errors
Successfully corrected 6 bit errors per subpage
Read reported 14 corrected bit errors
Successfully corrected 7 bit errors per subpage
Read reported 16 corrected bit errors
Successfully corrected 8 bit errors per subpage
Failed to recover 1 bitflips
Read error after 9 bit errors per page

Regarding the throughput improvements, tests have been conducted in
1-1-1 and 1-1-4 modes, reading a full block X pages at a
time, X ranging from 1 to 64 (size of a block with the tested device).
The percent value on the right is the comparison of the same test
conducted without the continuous read feature, ie. reading X pages in
one single user request, which got naturally split by the core whit the
continuous read optimization disabled into single-page reads.

* 1-1-1 result:
1 page read speed is 2634 KiB/s
2 page read speed is 2704 KiB/s (+3%)
3 page read speed is 2747 KiB/s (+5%)
4 page read speed is 2804 KiB/s (+7%)
5 page read speed is 2782 KiB/s
6 page read speed is 2826 KiB/s
7 page read speed is 2834 KiB/s
8 page read speed is 2821 KiB/s
9 page read speed is 2846 KiB/s
10 page read speed is 2819 KiB/s
11 page read speed is 2871 KiB/s (+10%)
12 page read speed is 2823 KiB/s
13 page read speed is 2880 KiB/s
14 page read speed is 2842 KiB/s
15 page read speed is 2862 KiB/s
16 page read speed is 2837 KiB/s
32 page read speed is 2879 KiB/s
64 page read speed is 2842 KiB/s

* 1-1-4 result:
1 page read speed is 7562 KiB/s
2 page read speed is 8904 KiB/s (+15%)
3 page read speed is 9655 KiB/s (+25%)
4 page read speed is 10118 KiB/s (+30%)
5 page read speed is 10084 KiB/s
6 page read speed is 10300 KiB/s
7 page read speed is 10434 KiB/s (+35%)
8 page read speed is 10406 KiB/s
9 page read speed is 10769 KiB/s (+40%)
10 page read speed is 10666 KiB/s
11 page read speed is 10757 KiB/s
12 page read speed is 10835 KiB/s
13 page read speed is 10976 KiB/s
14 page read speed is 11200 KiB/s
15 page read speed is 11009 KiB/s
16 page read speed is 11082 KiB/s
32 page read speed is 11352 KiB/s (+45%)
64 page read speed is 11403 KiB/s

This work has received support and could be achieved thanks to
Alvin Zhou <alvinzhou@mxic.com.tw>.
Signed-off-by: Miquel Raynal <miquel.raynal@bootlin.com>
Link: https://lore.kernel.org/linux-mtd/20240826101412.20644-10-miquel.raynal@bootlin.com