Re: [PATCH v2 2/8] s390/cio: introduce DMA pools to cio

[Cornelia Huck <cohuck@redhat.com>]

On Mon, 27 May 2019 13:47:55 +0200
Halil Pasic <pasic@linux.ibm.com> wrote:

> On Mon, 27 May 2019 08:57:18 +0200
> Cornelia Huck <cohuck@redhat.com> wrote:
> 
> > On Thu, 23 May 2019 18:22:03 +0200
> > Michael Mueller <mimu@linux.ibm.com> wrote:
> >   
> > > From: Halil Pasic <pasic@linux.ibm.com>
> > > 
> > > To support protected virtualization cio will need to make sure the
> > > memory used for communication with the hypervisor is DMA memory.
> > > 
> > > Let us introduce one global cio, and some tools for pools seated  
> > 
> > "one global pool for cio"?
> >   
> 
> Nod.
> 
> > > at individual devices.
> > > 
> > > Our DMA pools are implemented as a gen_pool backed with DMA pages. The
> > > idea is to avoid each allocation effectively wasting a page, as we
> > > typically allocate much less than PAGE_SIZE.
> > > 
> > > Signed-off-by: Halil Pasic <pasic@linux.ibm.com>
> > > ---
> > >  arch/s390/Kconfig           |   1 +
> > >  arch/s390/include/asm/cio.h |  11 +++++
> > >  drivers/s390/cio/css.c      | 110 ++++++++++++++++++++++++++++++++++++++++++++
> > >  3 files changed, 122 insertions(+)
> > >   
> > 
> > (...)
> >   
> > > @@ -1018,6 +1024,109 @@ static struct notifier_block css_power_notifier = {
> > >  	.notifier_call = css_power_event,
> > >  };
> > >  
> > > +#define POOL_INIT_PAGES 1
> > > +static struct gen_pool *cio_dma_pool;
> > > +/* Currently cio supports only a single css */  
> > 
> > This comment looks misplaced.  
> 
> Right! Move to ...
> 
> >   
> > > +#define  CIO_DMA_GFP (GFP_KERNEL | __GFP_ZERO)
> > > +
> > > +  
> 
> ... here?

Yes :)

> 
> > > +struct device *cio_get_dma_css_dev(void)
> > > +{
> > > +	return &channel_subsystems[0]->device;
> > > +}
> > > +
> > > +struct gen_pool *cio_gp_dma_create(struct device *dma_dev, int nr_pages)
> > > +{
> > > +	struct gen_pool *gp_dma;
> > > +	void *cpu_addr;
> > > +	dma_addr_t dma_addr;
> > > +	int i;
> > > +
> > > +	gp_dma = gen_pool_create(3, -1);
> > > +	if (!gp_dma)
> > > +		return NULL;
> > > +	for (i = 0; i < nr_pages; ++i) {
> > > +		cpu_addr = dma_alloc_coherent(dma_dev, PAGE_SIZE, &dma_addr,
> > > +					      CIO_DMA_GFP);
> > > +		if (!cpu_addr)
> > > +			return gp_dma;  
> > 
> > So, you may return here with no memory added to the pool at all (or
> > less than requested), but for the caller that is indistinguishable from
> > an allocation that went all right. May that be a problem?
> >   
> 
> I do not think it can cause a problem: cio_gp_dma_zalloc() is going to
> try to allocate the memory required and put it in the pool. If that
> fails as well, we return a NULL pointer like kmalloc(). So I think we
> are clean.
> 
> > > +		gen_pool_add_virt(gp_dma, (unsigned long) cpu_addr,
> > > +				  dma_addr, PAGE_SIZE, -1);
> > > +	}
> > > +	return gp_dma;
> > > +}
> > > +  
> > 
> > (...)
> >   
> > > +static void __init cio_dma_pool_init(void)
> > > +{
> > > +	/* No need to free up the resources: compiled in */
> > > +	cio_dma_pool = cio_gp_dma_create(cio_get_dma_css_dev(), 1);  
> > 
> > Does it make sense to continue if you did not get a pool here? I don't
> > think that should happen unless things were really bad already?
> >   
> 
> I agree, this should not fail under any sane circumstances. I don't
> think it makes sense to continue. Shall we simply call panic()?

Can we continue without the common I/O layer? Probably not. It might
really be an 'oh crap, let's panic' situation.

> 
> > > +}
> > > +
> > > +void *cio_gp_dma_zalloc(struct gen_pool *gp_dma, struct device *dma_dev,
> > > +			size_t size)
> > > +{
> > > +	dma_addr_t dma_addr;
> > > +	unsigned long addr;
> > > +	size_t chunk_size;
> > > +
> > > +	addr = gen_pool_alloc(gp_dma, size);
> > > +	while (!addr) {
> > > +		chunk_size = round_up(size, PAGE_SIZE);
> > > +		addr = (unsigned long) dma_alloc_coherent(dma_dev,
> > > +					 chunk_size, &dma_addr, CIO_DMA_GFP);
> > > +		if (!addr)
> > > +			return NULL;
> > > +		gen_pool_add_virt(gp_dma, addr, dma_addr, chunk_size, -1);
> > > +		addr = gen_pool_alloc(gp_dma, size);
> > > +	}
> > > +	return (void *) addr;
> > > +}
> > > +
> > > +void cio_gp_dma_free(struct gen_pool *gp_dma, void *cpu_addr, size_t size)
> > > +{
> > > +	if (!cpu_addr)
> > > +		return;
> > > +	memset(cpu_addr, 0, size);
> > > +	gen_pool_free(gp_dma, (unsigned long) cpu_addr, size);
> > > +}
> > > +
> > > +/**
> > > + * Allocate dma memory from the css global pool. Intended for memory not
> > > + * specific to any single device within the css. The allocated memory
> > > + * is not guaranteed to be 31-bit addressable.
> > > + *
> > > + * Caution: Not suitable for early stuff like console.
> > > + *
> > > + */
> > > +void *cio_dma_zalloc(size_t size)
> > > +{
> > > +	return cio_gp_dma_zalloc(cio_dma_pool, cio_get_dma_css_dev(), size);  
> > 
> > Ok, that looks like the failure I mentioned above should be
> > accommodated by the code. Still, I think it's a bit odd.
> >   
> 
> I think the behavior is reasonable: if client code wants pre-allocate n
> page sized chunks we pre-allocate as may as we can. If we can't
> pre-allocate all n, it ain't necessarily bad. There is no guarantee we
> will hit a wall in a non-recoverable fashion.

It's not necessarily broken, but there are two things that feel a bit
weird to me:
- The caller doesn't know if the requested pre-allocation worked or not.
- If we can't get memory in this early init phase, is it likely that we
  can get memory later on?

> 
> But if you insist, I can get rid of the pre-allocation or fail create and
> do a rollback if it fails.
> 
> Thanks for having a look!
> 
> Regards,
> Halil
> 
> > > +}  
> >   
>