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

[Cornelia Huck <cohuck@redhat.com>]

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"?

> 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.

> +#define  CIO_DMA_GFP (GFP_KERNEL | __GFP_ZERO)
> +
> +
> +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?

> +		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?

> +}
> +
> +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.

> +}