On Thu, May 4, 2017 at 3:10 PM, zhangyi (F) <yi.zhang@xxxxxxxxxx> wrote:
> On 2017/5/4 17:03, Amir Goldstein wrote:
>> On Thu, May 4, 2017 at 11:36 AM, zhangyi (F) <yi.zhang@xxxxxxxxxx> wrote:
>>>
>>> 2. Can we read underlying dir entry and get rid of whiteout every time for
>>> the 'impure' dircetory, not execute the merged flow ? It will save
>>> memory usage.
>>>
>> Did not understand this questions.
>>
>
> Sorry, I didn't make it clear. I mean if we set is_real = !OVL_TYPE_COPYUP()
> or is_real = !OVL_TYPE_MERGE() && !ovl_dir_impure() (with optimization),
> it will create ovl_dir_cache and ovl_cache_entry when we read COPYUP directory
> or 'impure' dircetory in ovl_iterate() flow (as merged dir does). This will
> cause bloated memory usage and some performance overhead, because this dir
> is only exist in one layer(not merged), so I think it's don't need ovl_dir_cache.
> (cost too much ?)
Does it though? cache is released when dir fd is closed.
>
> Can we find a way to avoid cache use, I find two ways(maybe too complex):
> 1. Make a wapper ovl_iterate_dir_impure(), it read the underlaying dir entry
> and get rid of whiteout entry, and use this function to deal the COPYUP or
> or 'impure' branch in ovl_iterate() like follow example:
>
This is interesting. It does cost a bit more to iterate with whiteout filter
like this because it requires getattr() while iterating char devices, but that's
not so bad I guess.
Filtering COPYUP like this is going to be more expensive.
> static int ovl_iterate(struct file *file, struct dir_context *ctx)
>
> if (od->is_real)
> return iterate_dir(od->realfile, ctx);
> + else if (!OVL_TYPE_MERGE && COPYUP)
> + return ovl_iterate_dir_impure(od->realfile, ctx);
>
> 2. Add a flag in struct dir_context, let underlying filesystem to return what we
> want to read, like follow example:
>
> #define DIR_CONTEXT_HIDE_WHITEOUT (1 << 0)
>
> struct dir_context {
> const filldir_t actor;
> + unsigned int flag;
> loff_t pos;
> };
>
> static int ovl_iterate(struct file *file, struct dir_context *ctx)
>
> if (od->is_real) /* is_real = !OVL_TYPE_MERGE() */
> return iterate_dir(od->realfile, ctx);
>
> And then let underlying filesystem deal with DIR_CONTEXT_HIDE_WHITEOUT
> flag. If take on this solution, we can cover all cases, OVL_TYPE_COPYUP
> and pure check are also not required. It changes too greate, but has
> a good expansibility (flag can be used by other demand in the future).
>
I wish filesystems would be so pleased to make these changes for overlayfs
it's less likely that you can get far with this idea ;-)
Amir.
--
To unsubscribe from this list: send the line "unsubscribe linux-unionfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
