Re: [PATCH v2 5/6] 9p: Use a slab for allocating requests

[Greg Kurz]

On Wed, 18 Jul 2018 12:05:54 +0200
Dominique Martinet <asmadeus@xxxxxxxxxxxxx> wrote:

> +Cc Greg, I could use your opinion on this if you have a moment.
> 

Hi Dominique,

The patch is quite big and I'm not sure I can find time to review it
carefully, but I'll try to help anyway.

> Matthew Wilcox wrote on Wed, Jul 11, 2018:
> > Replace the custom batch allocation with a slab.  Use an IDR to store
> > pointers to the active requests instead of an array.  We don't try to
> > handle P9_NOTAG specially; the IDR will happily shrink all the way back
> > once the TVERSION call has completed.  
> 
> Sorry for coming back to this patch now, I just noticed something that's
> actually probably a fairly big hit on performance...
> 
> While the slab is just as good as the array for the request itself, this
> makes every single request allocate "fcalls" everytime instead of
> reusing a cached allocation.
> The default msize is 8k and these allocs probably are fairly efficient,
> but some transports like RDMA allow to increase this to up to 1MB... And

It can be even bigger with virtio:

#define VIRTQUEUE_NUM	128

	.maxsize = PAGE_SIZE * (VIRTQUEUE_NUM - 3),

On a typical ppc64 server class setup with 64KB pages, this is nearly 8MB.

> doing this kind of allocation twice for every packet is going to be very
> slow.
> (not that hogging megabytes of memory was a great practice either!)
> 
> 
> One thing is that the buffers are all going to be the same size for a
> given client (.... except virtio zc buffers, I wonder what I'm missing
> or why that didn't blow up before?)

ZC allocates a 4KB buffer, which is more than enough to hold the 7-byte 9P
header and the "dqd" part of all messages that may use ZC, ie, 16 bytes.
So I'm not sure to catch what could blow up.

> Err, that aside I was going to ask if we couldn't find a way to keep a
> pool of these somehow.
> Ideally putting them in another slab so they could be reclaimed if
> necessary, but the size could vary from one client to another, can we
> create a kmem_cache object per client? the KMEM_CACHE macro is not very
> flexible so I don't think that is encouraged... :)
> 
> 
> It's a shame because I really like that patch, I'll try to find time to
> run some light benchmark with varying msizes eventually but I'm not sure
> when I'll find time for that... Hopefully before the 4.19 merge window!
> 

Yeah, the open-coded cache we have now really obfuscates things.

Maybe have a per-client kmem_cache object for non-ZC requests with
size msize [*], and a global kmem_cache object for ZC requests with
fixed size P9_ZC_HDR_SZ.

[*] the server can require a smaller msize during version negotiation,
    so maybe we should change the kmem_cache object in this case.

Cheers,

--
Greg

> 
> >  /**
> > - * p9_tag_alloc - lookup/allocate a request by tag
> > - * @c: client session to lookup tag within
> > - * @tag: numeric id for transaction
> > - *
> > - * this is a simple array lookup, but will grow the
> > - * request_slots as necessary to accommodate transaction
> > - * ids which did not previously have a slot.
> > - *
> > - * this code relies on the client spinlock to manage locks, its
> > - * possible we should switch to something else, but I'd rather
> > - * stick with something low-overhead for the common case.
> > + * p9_req_alloc - Allocate a new request.
> > + * @c: Client session.
> > + * @type: Transaction type.
> > + * @max_size: Maximum packet size for this request.
> >   *
> > + * Context: Process context.
> > + * Return: Pointer to new request.
> >   */
> > -
> >  static struct p9_req_t *
> > -p9_tag_alloc(struct p9_client *c, u16 tag, unsigned int max_size)
> > +p9_tag_alloc(struct p9_client *c, int8_t type, unsigned int max_size)
> >  {
> > -	unsigned long flags;
> > -	int row, col;
> > -	struct p9_req_t *req;
> > +	struct p9_req_t *req = kmem_cache_alloc(p9_req_cache, GFP_NOFS);
> >  	int alloc_msize = min(c->msize, max_size);
> > +	int tag;
> >  
> > -	/* This looks up the original request by tag so we know which
> > -	 * buffer to read the data into */
> > -	tag++;
> > -
> > -	if (tag >= c->max_tag) {
> > -		spin_lock_irqsave(&c->lock, flags);
> > -		/* check again since original check was outside of lock */
> > -		while (tag >= c->max_tag) {
> > -			row = (tag / P9_ROW_MAXTAG);
> > -			c->reqs[row] = kcalloc(P9_ROW_MAXTAG,
> > -					sizeof(struct p9_req_t), GFP_ATOMIC);
> > -
> > -			if (!c->reqs[row]) {
> > -				pr_err("Couldn't grow tag array\n");
> > -				spin_unlock_irqrestore(&c->lock, flags);
> > -				return ERR_PTR(-ENOMEM);
> > -			}
> > -			for (col = 0; col < P9_ROW_MAXTAG; col++) {
> > -				req = &c->reqs[row][col];
> > -				req->status = REQ_STATUS_IDLE;
> > -				init_waitqueue_head(&req->wq);
> > -			}
> > -			c->max_tag += P9_ROW_MAXTAG;
> > -		}
> > -		spin_unlock_irqrestore(&c->lock, flags);
> > -	}
> > -	row = tag / P9_ROW_MAXTAG;
> > -	col = tag % P9_ROW_MAXTAG;
> > +	if (!req)
> > +		return NULL;
> >  
> > -	req = &c->reqs[row][col];
> > -	if (!req->tc)
> > -		req->tc = p9_fcall_alloc(alloc_msize);
> > -	if (!req->rc)
> > -		req->rc = p9_fcall_alloc(alloc_msize);
> > +	req->tc = p9_fcall_alloc(alloc_msize);
> > +	req->rc = p9_fcall_alloc(alloc_msize);
> >  	if (!req->tc || !req->rc)
> > -		goto grow_failed;
> > +		goto free;
> >  
> >  	p9pdu_reset(req->tc);
> >  	p9pdu_reset(req->rc);
> > -
> > -	req->tc->tag = tag-1;
> >  	req->status = REQ_STATUS_ALLOC;
> > +	init_waitqueue_head(&req->wq);
> > +	INIT_LIST_HEAD(&req->req_list);
> > +
> > +	idr_preload(GFP_NOFS);
> > +	spin_lock_irq(&c->lock);
> > +	if (type == P9_TVERSION)
> > +		tag = idr_alloc(&c->reqs, req, P9_NOTAG, P9_NOTAG + 1,
> > +				GFP_NOWAIT);
> > +	else
> > +		tag = idr_alloc(&c->reqs, req, 0, P9_NOTAG, GFP_NOWAIT);
> > +	req->tc->tag = tag;
> > +	spin_unlock_irq(&c->lock);
> > +	idr_preload_end();
> > +	if (tag < 0)
> > +		goto free;  
>