[PATCH/RFC] HW ECC optimization
Vitaly Wool
vwool at ru.mvista.com
Thu Dec 15 04:28:42 EST 2005
Hi,
currently the whole page is being written for nand_write_oob in HW ECC case.
This is done in order to support some HW ECC generators which need a whole page to be written to generate proper ECC. However, for most of those generators it's not necessary and it results in perfrormance penalty during OOB writes.
Well, the patch inlined below presents a proof-of-concept approach for speeding this up. The main problem is how to distinguish between different types of HW ECC generators? I'm using SYNDROME flag for this purpose, although it may make sense to create a separate one.
Anyway, here's some measurements with and w/o this patch applied on my 2k-page NAND.
layout {data, oob, ecc, data, oob, ecc, data, oob, ecc, data, oob, ecc}, w/ the patch:
root at 192.168.3.101:~# time flash_eraseall -j /dev/mtd/3
Erasing 128 Kibyte @ 12e0000 -- 99 % complete. Cleanmarker written at 12e0000.
real 0m0.709s
user 0m0.010s
sys 0m0.350s
layout {data, oob, ecc, data, oob, ecc, data, oob, ecc, data, oob, ecc}, w/o the patch:
root at 192.168.3.101:~# time flash_eraseall -j /dev/mtd/3
Erasing 128 Kibyte @ 12e0000 -- 99 % complete. Cleanmarker written at 12e0000.
real 0m0.811s
user 0m0.000s
sys 0m0.370s
layout {data, oob, ecc, oob, ecc, oob, ecc, oob, ecc}, w/o the patch:
root at 192.168.3.101:~# ./xx.sh
root at 192.168.3.101:~# time flash_eraseall -j /dev/mtd/3
Erasing 128 Kibyte @ 12e0000 -- 99 % complete. Cleanmarker written at 12e0000.
real 0m0.849s
user 0m0.000s
sys 0m0.330s
layout {data, oob, ecc, oob, ecc, oob, ecc, oob, ecc}, w/ the patch:
root at 192.168.3.101:~# time flash_eraseall -j /dev/mtd/3
Erasing 128 Kibyte @ 12e0000 -- 99 % complete. Cleanmarker written at 12e0000.
real 0m0.709s
user 0m0.000s
sys 0m0.340s
Any comments are welcome.
Vitaly
--- drivers/mtd/nand/nand_base.c 2005-12-14 15:29:42.000000000 +0300
+++ drivers/mtd/nand/nand_base.c 2005-12-14 21:37:47.324960872 +0300
@@ -1948,7 +1948,7 @@
}
} else {
int i = 0, j = 0;
- int fflen = 0, ooblen = 0;
+ int fflen = 0, old_fflen = 0, ooblen = 0;
/* Write out desired data */
this->cmdfunc (mtd, NAND_CMD_SEQIN, 0, page & this->pagemask);
@@ -1957,9 +1957,27 @@
for (j = 0; this->layout[j].length; j++) {
switch (this->layout[j].type) {
case ITEM_TYPE_DATA:
- this->enable_hwecc(mtd, NAND_ECC_WRITE);
- this->write_buf(mtd, ffchars, this->layout[j].length);
- fflen += this->layout[j].length;
+ if (this->options & NAND_HWECC_SYNDROME) {
+ this->enable_hwecc(mtd, NAND_ECC_WRITE);
+ this->write_buf(mtd, ffchars, this->layout[j].length);
+ fflen += this->layout[j].length;
+ } else {
+ if (old_fflen < fflen) {
+ this->cmdfunc (mtd, NAND_CMD_PAGEPROG, -1, -1);
+ status = this->waitfunc (mtd, this, FL_WRITING);
+ if (status & NAND_STATUS_FAIL) {
+ DEBUG (MTD_DEBUG_LEVEL0, "%s: Failed write, page 0x%08x\n", __FUNCTION__, page);
+ ret = -EIO;
+ goto out;
+ }
+ }
+ fflen += this->layout[j].length;
+ if (this->options & NAND_BUSWIDTH_16 && (fflen + ooblen) & 1)
+ this->cmdfunc (mtd, NAND_CMD_SEQIN, fflen + ooblen - 1, page & this->pagemask);
+ else
+ this->cmdfunc (mtd, NAND_CMD_SEQIN, fflen + ooblen, page & this->pagemask);
+ old_fflen = fflen;
+ }
break;
case ITEM_TYPE_ECC:
More information about the linux-mtd
mailing list