Re: [PATCH 1/2] mtd: mtdconcat: Judge callback function existence getting from master for each partition

[Zhihao Cheng]

在 2021/8/7 18:32, Miquel Raynal 写道:
Hi Miquel,
> Hi Zhihao,
>
> Zhihao Cheng <chengzhihao1@xxxxxxxxxx> wrote on Sat, 7 Aug 2021
> 10:15:46 +0800:
>
> > 在 2021/8/7 3:28, Miquel Raynal 写道:
> > Hi Miquel,
> > > Hi Zhihao,
> > >
> > > Zhihao Cheng <chengzhihao1@xxxxxxxxxx> wrote on Sat, 31 Jul 2021
> > > 10:32:42 +0800:
> > > @@ -721,14 +724,15 @@ struct mtd_info *mtd_concat_create(struct mtd_info *subdev[],	/* subdevices to c
> > >    				    subdev[i]->flags & MTD_WRITEABLE;
> > >    		}
> > >    > +		subdev_master = mtd_get_master(subdev[i]);
> > >    		concat->mtd.size += subdev[i]->size;
> > >    		concat->mtd.ecc_stats.badblocks +=
> > >    			subdev[i]->ecc_stats.badblocks;
> > >    		if (concat->mtd.writesize   !=  subdev[i]->writesize ||
> > >    		    concat->mtd.subpage_sft != subdev[i]->subpage_sft ||
> > >    		    concat->mtd.oobsize    !=  subdev[i]->oobsize ||
> > > -		    !concat->mtd._read_oob  != !subdev[i]->_read_oob ||
> > > -		    !concat->mtd._write_oob != !subdev[i]->_write_oob) {
> > > +		    !concat->mtd._read_oob  != !subdev_master->_read_oob ||
> > > +		    !concat->mtd._write_oob != !subdev_master->_write_oob) {
> > > Do you really need this change?
> > I think both "!concat->mtd._read_oob != !subdev[i]->_read_oob" and "!concat->mtd._write_oob != !subdev[i]->_write_oob" need to be modified otherwise concatenating goes failure.
> >
> > I thought there exists two problems:
> >
> >     1. Wrong callback fetching in mtd partition device
> >
> >     2. Warning for existence of _read and _read_oob at the same time
> >
> > so I solved them in two steps to make history commit logs a bit clear.
> >
> > Though these two patches can be combined to one.
> No please keep the split.
>
> What I mean here is that I don't think your fix is valid. Maybe we
> should propagate these callbacks as well instead of trying to hack into
> this condition. I don't see why you should check against subdev[i] for
> half of the callbacks and check for subdev_master for the last two.

I have the following understanding of mtd:

1. The existence of mtd partition device's callbacks(what can mtd do, 
_read, _write, _read_oob, etc.) is decided by mtd master device. All mtd 
common functions (mtd_read, mtd_read_oob) pass master mtd device rather 
than partition mtd device as parameter, because nand_chip(initialized as 
master mtd device) process requests.  So there are two steps in mtd 
common function: fetch master mtd device; invoke master mtd devices's 
callback and pass in master mtd device.

  Propogating callbacks to partition mtd device may bring some 
imperfections:

  A. No adaptions to mtd common functions: partition mtd device's 
callbacks will never be invoked, they are only used to judge whether 
assigin concatenated device's callback while concatenating. Looks weird.

  @@ -86,6 +86,61 @@ static struct mtd_info *allocate_partition(struct 
mtd_info *parent,
        child->part.offset = part->offset;
        INIT_LIST_HEAD(&child->partitions);

+       if (parent->_read)
+               child->_read = parent->_read;
+       if (parent->_write)
+               child->_write = parent->_write;
[...]
+       if (parent->_read_oob)
+               child->_read_oob = parent->_read_oob;
+       if (parent->_write_oob)


  B. Re-import removed partition mtd device's callbacks and adapt mtd 
common functions: Current implemention is simplier than the version 
before 46b5889cc2c54("mtd: implement proper partition handling"). 
Adapting mtd common functions is a risky thing, which could effect other 
modules, should we do that?

+static int part_read(struct mtd_info *mtd, loff_t from, size_t len,
+               size_t *retlen, u_char *buf)
+{
+       struct mtd_part *part = mtd_to_part(mtd);
+       struct mtd_ecc_stats stats;
+       int res;
+
+      stats = part->parent->ecc_stats;
+       res = part->parent->_read(part->parent, from + part->offset, len,
+                                 retlen, buf);
+       if (unlikely(mtd_is_eccerr(res)))
+               mtd->ecc_stats.failed +=
+                       part->parent->ecc_stats.failed - stats.failed;
+       else
+               mtd->ecc_stats.corrected +=
+                       part->parent->ecc_stats.corrected - stats.corrected;
+       return res;
+}

 static int mtd_read_oob_std(struct mtd_info *mtd, loff_t from,
                            struct mtd_oob_ops *ops)
 {
-       struct mtd_info *master = mtd_get_master(mtd);
        int ret;

-       from = mtd_get_master_ofs(mtd, from);
-       if (master->_read_oob)
-               ret = master->_read_oob(master, from, ops);
+       if (mtd->_read_oob)
+               ret = mtd->_read_oob(mtd, from, ops);
        else
-               ret = master->_read(master, from, ops->len, &ops->retlen,
+               ret = mtd->_read(mtd, from, ops->len, &ops->retlen,
                                    ops->datbuf);


2. Checking against subdev[i] for data members and check against 
subdev_master for the callbacks looks weird. I modified it based on the 
assumption that partition mtd device' callbacks inherit from master mtd 
device but data members(name, size) may not. Maybe I should add some 
comment to explain why checking against subdev[i] for data members and 
checking against subdev_master for the callbacks.


So, which method is better? Any other method?