Message-ID: <a5a11a56-46dc-ab44-564b-d2216f32451e@arm.com>
Date:   Fri, 5 Apr 2019 10:51:38 +0100
From:   Robin Murphy <robin.murphy@....com>
To:     Steven Price <steven.price@....com>, Rob Herring <robh@...nel.org>,
        dri-devel@...ts.freedesktop.org
Cc:     Sean Paul <sean@...rly.run>,
        Maxime Ripard <maxime.ripard@...tlin.com>,
        Neil Armstrong <narmstrong@...libre.com>,
        Will Deacon <will.deacon@....com>,
        linux-kernel@...r.kernel.org, David Airlie <airlied@...ux.ie>,
        iommu@...ts.linux-foundation.org,
        Alyssa Rosenzweig <alyssa@...enzweig.io>,
        linux-arm-kernel@...ts.infradead.org
Subject: Re: [PATCH v3 1/3] iommu: io-pgtable: Add ARM Mali midgard MMU page
 table format

Hi Steve,

On 05/04/2019 10:42, Steven Price wrote:
> First let me say congratulations to everyone working on Panfrost - it's
> an impressive achievement!
> 
> Full disclosure: I used to work on the Mali kbase driver. And have been
> playing around with running the Mali user-space blob with the Panfrost
> kernel driver.
> 
> On 01/04/2019 08:47, Rob Herring wrote:
>> ARM Mali midgard GPU is similar to standard 64-bit stage 1 page tables, but
>> have a few differences. Add a new format type to represent the format. The
>> input address size is 48-bits and the output address size is 40-bits (and
>> possibly less?). Note that the later bifrost GPUs follow the standard
>> 64-bit stage 1 format.
>>
>> The differences in the format compared to 64-bit stage 1 format are:
>>
>> The 3rd level page entry bits are 0x1 instead of 0x3 for page entries.
>>
>> The access flags are not read-only and unprivileged, but read and write.
>> This is similar to stage 2 entries, but the memory attributes field matches
>> stage 1 being an index.
>>
>> The nG bit is not set by the vendor driver. This one didn't seem to matter,
>> but we'll keep it aligned to the vendor driver.
> 
> The nG bit should be ignored by the hardware.
> 
> The MMU in Midgard/Bifrost has a quirk similar to
> IO_PGTABLE_QUIRK_TLBI_ON_MAP - you must perform a cache flush for the
> GPU to (reliably) pick up new page table mappings.
> 
> You may not have seen this because of the use of the JS_CONFIG_START_MMU
> bit - this effectively performs a cache flush and TLB invalidate before
> starting a job, however when using a GPU like T760 (e.g. on the Firefly
> RK3288) this bit isn't being set. In my testing on the Firefly board I
> saw GPU page faults because of this.
> 
> There's two options for fixing this - a patch like below adds the quirk
> mode to the MMU. Or alternatively always set JS_CONFIG_START_MMU on
> jobs. In my testing both options solve the page faults.
> 
> To be honest I don't know the reasoning behind kbase making the
> JS_CONFIG_START_MMU bit conditional - I'm not aware of any reason why it
> can't always be set. My guess is performance, but I haven't benchmarked
> the difference between this and JS_CONFIG_START_MMU.
> 
> -----8<----------
>  From e3f75c7f04e43238dfc579029b8c11fb6b4a0c18 Mon Sep 17 00:00:00 2001
> From: Steven Price <steven.price@....com>
> Date: Thu, 4 Apr 2019 15:53:17 +0100
> Subject: [PATCH] iommu: io-pgtable: IO_PGTABLE_QUIRK_TLBI_ON_MAP for LPAE
> 
> Midgard/Bifrost GPUs require a TLB invalidation when mapping pages,
> implement IO_PGTABLE_QUIRK_TLBI_ON_MAP for LPAE iommu page table
> formats and add the quirk bit to Panfrost.
> 
> Signed-off-by: Steven Price <steven.price@....com>
> ---
>   drivers/gpu/drm/panfrost/panfrost_mmu.c |  1 +
>   drivers/iommu/io-pgtable-arm.c          | 11 +++++++++--
>   2 files changed, 10 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/panfrost/panfrost_mmu.c
> b/drivers/gpu/drm/panfrost/panfrost_mmu.c
> index f3aad8591cf4..094312074d66 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_mmu.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_mmu.c
> @@ -343,6 +343,7 @@ int panfrost_mmu_init(struct panfrost_device *pfdev)
>   	mmu_write(pfdev, MMU_INT_MASK, ~0);
> 
>   	pfdev->mmu->pgtbl_cfg = (struct io_pgtable_cfg) {
> +		.quirks		= IO_PGTABLE_QUIRK_TLBI_ON_MAP,
>   		.pgsize_bitmap	= SZ_4K, // | SZ_2M | SZ_1G),
>   		.ias		= 48,
>   		.oas		= 40,	/* Should come from dma mask? */
> diff --git a/drivers/iommu/io-pgtable-arm.c b/drivers/iommu/io-pgtable-arm.c
> index 84beea1f47a7..45fd7bbdf9aa 100644
> --- a/drivers/iommu/io-pgtable-arm.c
> +++ b/drivers/iommu/io-pgtable-arm.c
> @@ -505,7 +505,13 @@ static int arm_lpae_map(struct io_pgtable_ops *ops,
> unsigned long iova,
>   	 * Synchronise all PTE updates for the new mapping before there's
>   	 * a chance for anything to kick off a table walk for the new iova.
>   	 */
> -	wmb();
> +	if (data->iop.cfg.quirks & IO_PGTABLE_QUIRK_TLBI_ON_MAP) {
> +		io_pgtable_tlb_add_flush(&data->iop, iova, size,
> +					 ARM_LPAE_BLOCK_SIZE(2, data), false);

For correctness (in case this ever ends up used for something with 
VMSA-like invalidation behaviour), the granule would need to be "size" 
here, rather than effectively hard-coded.

However, since Mali's invalidations appear to operate on arbitrary 
ranges, it would probably be a lot more efficient for the driver to 
handle this case directly, by just issuing a single big invalidation 
after the for_each_sg() loop in panfrost_mmu_map().

Robin.

> +		io_pgtable_tlb_sync(&data->iop);
> +	} else {
> +		wmb();
> +	}
> 
>   	return ret;
>   }
> @@ -800,7 +806,8 @@ arm_64_lpae_alloc_pgtable_s1(struct io_pgtable_cfg
> *cfg, void *cookie)
>   	struct arm_lpae_io_pgtable *data;
> 
>   	if (cfg->quirks & ~(IO_PGTABLE_QUIRK_ARM_NS | IO_PGTABLE_QUIRK_NO_DMA |
> -			    IO_PGTABLE_QUIRK_NON_STRICT))
> +			    IO_PGTABLE_QUIRK_NON_STRICT |
> +			    IO_PGTABLE_QUIRK_TLBI_ON_MAP))
>   		return NULL;
> 
>   	data = arm_lpae_alloc_pgtable(cfg);
> 

Powered by blists - more mailing lists