Re: [RFC] mballoc patches

[Alex Tomas <alex@clusterfs.com>]

Eric Sandeen wrote:
> +/*
> + * default stripe size = 1MB
> + */
> +#define MB_DEFAULT_STRIPE              256

agree, though seems we'd better make it blocksize-insensitive

> 
> Units?  Doesn't seem to matter anyway as it's never referenced.
> 
> +       /* tunables */
> +       unsigned long s_mb_factor;
> +       unsigned long s_stripe;
> +       unsigned long s_mb_small_req;
> +       unsigned long s_mb_large_req;
> +       unsigned long s_mb_max_to_scan;
> +       unsigned long s_mb_min_to_scan;
> 
> could we get some comments here as to what these are, and what units?

OK, I'll do as well as policy (and tunning) description.

> 
> Same is true many places... for example
> 
> +static int mb_find_extent(struct ext4_buddy *e3b, int order, int block,
> +                               int needed, struct ext4_free_extent *ex)
> 
> how many "what" are needed?

well, blocks :)

> And perhaps an addition of the new mount options to
> Documentation/fs/ext4.txt would be good.
> 
> +#define EXT4_MB_BITMAP(e3b)    ((e3b)->bd_bitmap)
> +#define EXT4_MB_BUDDY(e3b)     ((e3b)->bd_buddy)
> 
> For the sake of consistency should these (and others) be e4b?

OK

> 
> Also there are a *lot* of BUGs and BUG_ONs added in this patch... are
> none of these recoverable?

well, I'll review the code in this regard again, but most of them are not.
not that I like kernel panics, but BUG_ON() are very helpful to maintain
code, especially in long-term.

thanks, Alex