Re: [PATCH] fs/resctrl: Slightly optimize cbm_validate()

[Dave Martin <Dave.Martin@arm.com>]

Hi,

On Mon, Nov 03, 2025 at 10:17:16AM -0800, Luck, Tony wrote:
> On Mon, Oct 27, 2025 at 11:43:49AM +0000, Dave Martin wrote:
> > Hi,
> > 
> > [Tony, I have a side question on min_cbm_bits -- see below.]
> > [...]
> > 
> > <aside>
> > 
> > Also, not directly related to this patch, but, looking at the final if
> > statement:
> > 
> > 	if ((zero_bit - first_bit) < r->cache.min_cbm_bits) {
> > 	        rdt_last_cmd_printf("Need at least %d bits in the mask\n",
> > 	                            r->cache.min_cbm_bits);
> > 	        return false;
> > 	}
> > 
> > If min_cbm_bits is two or greater, this can fail if the bitmap has
> > enough contiguous set bits but not in the first block of set bits,
> > and it can succeed if there are blocks of set bits beyond the first
> > block, that have fewer than min_cbm_bits.
> > 
> > Is that intended?  Do we ever expect arch_has_sparse_bitmasks alongside
> > min_cbm_bits > 1, or should these be mutually exclusive?
> > 
> > </aside>
> 
> There's no enumeration for the minimium number of bits in a CBM mask.
> Haswell (first to implemenent L3 cache allocation) got a quirk to
> to set it to "2". I don't expect that we'd do that again.
> 
> So safe to assume that resctrl doesn't have to handle the combination
> of min_cbm_bits > 1 with arch_has_sparse_bitmasks.
> 
> -Tony

OK.  A min_cbm_bits value > 1 seems unlikely with sparse bitmasks
anyway.  If the hardware has independent storage for each bit, there
would be no need for such a constraint...  so I would be surprised to
see this in practice.

Just wanted to check that I wasn't missing something!

In MPAM, bitmap controls always allow each bit to be controlled
independently, according to the architecture.

Cheers
---Dave