Re: [PATCH AUTOSEL 4.14 32/58] bcache: at least try to shrink 1 node in bch_mca_scan()

[John Stoffel]

>>>>> "Coly" == Coly Li <colyli@xxxxxxx> writes:

Coly> On 2019/12/12 11:48 äå, John Stoffel wrote:
>>>>>>> "Sasha" == Sasha Levin <sashal@xxxxxxxxxx> writes:
>> 
Sasha> From: Coly Li <colyli@xxxxxxx>
Sasha> [ Upstream commit 9fcc34b1a6dd4b8e5337e2b6ef45e428897eca6b ]
>> 
Sasha> In bch_mca_scan(), the number of shrinking btree node is calculated
Sasha> by code like this,
Sasha> unsigned long nr = sc->nr_to_scan;
>> 
Sasha> nr /= c->btree_pages;
Sasha> nr = min_t(unsigned long, nr, mca_can_free(c));
Sasha> variable sc->nr_to_scan is number of objects (here is bcache B+tree
Sasha> nodes' number) to shrink, and pointer variable sc is sent from memory
Sasha> management code as parametr of a callback.
>> 
Sasha> If sc->nr_to_scan is smaller than c->btree_pages, after the above
Sasha> calculation, variable 'nr' will be 0 and nothing will be shrunk. It is
Sasha> frequeently observed that only 1 or 2 is set to sc->nr_to_scan and make
Sasha> nr to be zero. Then bch_mca_scan() will do nothing more then acquiring
Sasha> and releasing mutex c->bucket_lock.
>> 
Sasha> This patch checkes whether nr is 0 after the above calculation, if 0
Sasha> is the result then set 1 to variable 'n'. Then at least bch_mca_scan()
Sasha> will try to shrink a single B+tree node.
>> 
Sasha> nr /= c->btree_pages;
Sasha> +	if (nr == 0)
Sasha> +		nr = 1;
>> 
>> 
>> Wouldn't it be even more clear with:
>> 
>> nr /= c->bree_pages || 1;
>> 
>> instead?

Coly> No, it is not more clear. At least to me, I may confuse does it mean,
Coly> - nr = (nr / c->btree_pages) || 1;
Coly> - or nr = nr / (c->btree_pages || 1)

Coly> If I don't check C manual, I am not able to tell the correct
Coly> calculate at first time.

You're right, it's not quite as clear, it needs proper parenthesis.
But maybe instead of a (possibly) expensive division all the time, why
not just shift and assume you have it shrink a node, or try to.

I honestly haven't looked closely enough at the code to figure out the
best shift to use here.  But isn't this calculation wrong anyway?  If
you have lots of c->bree_pages, wouldn't you want to do more freeing?

I'd need to read the code better, but I'm heading to bed now.  Sorry.

John