iommu: Add a quirk for ARM Mali midgard MMU
Subsystems (2)
PATCH 1/1 iommu: Add a quirk for ARM Mali midgard MMU
ARM Mali midgard GPUs have a few differences from standard 64-bit stage 1 page tables.
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 also not set by the vendor driver, but that one doesn't seem to matter.
Add a quirk to handle all of these differences.
--- drivers/iommu/io-pgtable-arm.c | 51 ++++++++++++++++++++++------------ include/linux/io-pgtable.h | 4 +++ 2 files changed, 37 insertions(+), 18 deletions(-)
drivers/iommu/io-pgtable-arm.c +33 -18
@@ -180,11 +180,6 @@#define iopte_prot (pte) ((pte) & ARM_LPAE_PTE_ATTR_MASK) -#define iopte_leaf (pte,l) \ - (l == (ARM_LPAE_MAX_LEVELS - 1) ? \ - (iopte_type(pte,l) == ARM_LPAE_PTE_TYPE_PAGE) : \ - (iopte_type(pte,l) == ARM_LPAE_PTE_TYPE_BLOCK)) -struct arm_lpae_io_pgtable {struct io_pgtable iop;
@@ -198,6 +193,15 @@ struct arm_lpae_io_pgtable {typedef u64 arm_lpae_iopte ; +static inline bool iopte_leaf (arm_lpae_iopte pte,int l,unsigned long quirks) +{ +if ((l == (ARM_LPAE_MAX_LEVELS -1 )) && + !(quirks &IO_PGTABLE_QUIRK_ARM_MALI_MIDGARD )) +return iopte_type (pte,l) ==ARM_LPAE_PTE_TYPE_PAGE ; + +return iopte_type (pte,l) ==ARM_LPAE_PTE_TYPE_BLOCK ; +} +static arm_lpae_iopte paddr_to_iopte (phys_addr_t paddr,struct arm_lpae_io_pgtable *data) {
@@ -304,11 +308,14 @@ static void __arm_lpae_init_pte(struct arm_lpae_io_pgtable *data, pte |=ARM_LPAE_PTE_NS ;if (lvl ==ARM_LPAE_MAX_LEVELS -1 ) - pte |=ARM_LPAE_PTE_TYPE_PAGE ; + pte |= (data->iop.cfg.quirks &IO_PGTABLE_QUIRK_ARM_MALI_MIDGARD ) ? +ARM_LPAE_PTE_TYPE_BLOCK :ARM_LPAE_PTE_TYPE_PAGE ;else pte |=ARM_LPAE_PTE_TYPE_BLOCK ; - pte |=ARM_LPAE_PTE_AF |ARM_LPAE_PTE_SH_IS ; +if (!(data->iop.cfg.quirks &IO_PGTABLE_QUIRK_ARM_MALI_MIDGARD )) + pte |=ARM_LPAE_PTE_AF ; + pte |=ARM_LPAE_PTE_SH_IS ; pte |=paddr_to_iopte (paddr, data);__arm_lpae_set_pte (ptep, pte, &data->iop.cfg);
@@ -321,7 +328,7 @@ static int arm_lpae_init_pte(struct arm_lpae_io_pgtable *data, {arm_lpae_iopte pte = *ptep; -if (iopte_leaf (pte, lvl)) { +if (iopte_leaf (pte, lvl, data->iop.cfg.quirks)) {/* We require an unmap first */ WARN_ON (!selftest_running);return -EEXIST ;
@@ -409,7 +416,7 @@ static int __arm_lpae_map(struct arm_lpae_io_pgtable *data, unsigned long iova,__arm_lpae_sync_pte (ptep, cfg); } -if (pte && !iopte_leaf (pte, lvl)) { +if (pte && !iopte_leaf (pte, lvl, cfg->quirks)) { cptep =iopte_deref (pte, data); }else if (pte) {/* We require an unmap first */
@@ -430,12 +437,19 @@ static arm_lpae_iopte arm_lpae_prot_to_pte(struct arm_lpae_io_pgtable *data, data->iop.fmt == ARM_32_LPAE_S1) { pte = ARM_LPAE_PTE_nG; -if (!(prot &IOMMU_WRITE ) && (prot &IOMMU_READ )) - pte |=ARM_LPAE_PTE_AP_RDONLY ; +if (data->iop.cfg.quirks &IO_PGTABLE_QUIRK_ARM_MALI_MIDGARD ) { +if (prot &IOMMU_WRITE ) + pte |=ARM_LPAE_PTE_AP_RDONLY ; -if (!(prot &IOMMU_PRIV )) - pte |=ARM_LPAE_PTE_AP_UNPRIV ; +if (prot &IOMMU_READ ) + pte |=ARM_LPAE_PTE_AP_UNPRIV ; + }else { +if (!(prot &IOMMU_WRITE ) && (prot &IOMMU_READ )) + pte |=ARM_LPAE_PTE_AP_RDONLY ; +if (!(prot &IOMMU_PRIV )) + pte |=ARM_LPAE_PTE_AP_UNPRIV ; + }if (prot &IOMMU_MMIO ) pte |= (ARM_LPAE_MAIR_ATTR_IDX_DEV <<ARM_LPAE_PTE_ATTRINDX_SHIFT );
@@ -511,7 +525,7 @@ static void __arm_lpae_free_pgtable(struct arm_lpae_io_pgtable *data, int lvl,while (ptep != end) {arm_lpae_iopte pte = *ptep++; -if (!pte ||iopte_leaf (pte, lvl)) +if (!pte ||iopte_leaf (pte, lvl, data->iop.cfg.quirks))continue ;__arm_lpae_free_pgtable (data, lvl +1 ,iopte_deref (pte, data));
@@ -602,7 +616,7 @@ static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,if (size ==ARM_LPAE_BLOCK_SIZE (lvl, data)) {__arm_lpae_set_pte (ptep,0 , &iop->cfg); -if (!iopte_leaf (pte, lvl)) { +if (!iopte_leaf (pte, lvl, iop->cfg.quirks)) {/* Also flush any partial walks */ io_pgtable_tlb_add_flush (iop, iova, size,ARM_LPAE_GRANULE (data), false);
@@ -621,7 +635,7 @@ static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data, }return size; - }else if (iopte_leaf (pte, lvl)) { + }else if (iopte_leaf (pte, lvl, iop->cfg.quirks)) {/* * Insert a table at the next level to map the old region, * minus the part we want to unmap
@@ -669,7 +683,7 @@ static phys_addr_t arm_lpae_iova_to_phys(struct io_pgtable_ops *ops,return 0 ;/* Leaf entry? */ -if (iopte_leaf (pte,lvl)) +if (iopte_leaf (pte,lvl, data->iop.cfg.quirks)) goto found_translation;/* Take it to the next level */
@@ -779,7 +793,8 @@ 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_ARM_MALI_MIDGARD ))return NULL ; data =arm_lpae_alloc_pgtable (cfg);
include/linux/io-pgtable.h +4 -0
@@ -75,6 +75,9 @@ struct io_pgtable_cfg {* IO_PGTABLE_QUIRK_NON_STRICT: Skip issuing synchronous leaf TLBIs * on unmap, for DMA domains using the flush queue mechanism for * delayed invalidation. +* +* IO_PGTABLE_QUIRK_ARM_MALI_MIDGARD: ARM Mali Midgard MMU has different +* mapping of access flags and PTE page bits. */ #define IO_PGTABLE_QUIRK_ARM_NS BIT(0)#define IO_PGTABLE_QUIRK_NO_PERMS BIT(1)
@@ -82,6 +85,7 @@ struct io_pgtable_cfg {#define IO_PGTABLE_QUIRK_ARM_MTK_4GB BIT(3)#define IO_PGTABLE_QUIRK_NO_DMA BIT(4)#define IO_PGTABLE_QUIRK_NON_STRICT BIT(5) +#define IO_PGTABLE_QUIRK_ARM_MALI_MIDGARD BIT(6)unsigned long quirks;unsigned long pgsize_bitmap;unsigned int ias;
Signature · 1 line
2.19.1
Robin Murphy
reviewerHi Rob,
Quoted 5 lines · Rob Herring
ARM Mali midgard GPUs have a few differences from standard 64-bit stage 1 page tables.
The 3rd level page entry bits are
0x1instead of0x3for page entries.
Quoted 8 lines
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 also not set by the vendor driver, but that one doesn't seem to matter.
Add a quirk to handle all of these differences.
From the look of these changes, this isn't a quirk but a distinct format. AFAICS from the mali_kbase driver, this must be "LPAE mode" rather than "AArch64 mode", so it seems unlikely that it really supports the full VMSAv8 gamut of granules, address sizes, and page sizes that this patch will happily let through.
Robin.
Quoted 171 lines
Cc Will Deacon <[email protected]>Cc Robin Murphy <[email protected]>Cc Joerg Roedel <[email protected]>Signed-off-by Rob Herring <[email protected]>--- drivers/iommu/io-pgtable-arm.c | 51 ++++++++++++++++++++++------------ include/linux/io-pgtable.h | 4 +++ 2 files changed, 37 insertions(+), 18 deletions(-)drivers/iommu/io-pgtable-arm.c +33
-18@@ -180,11 +180,6 @@#define iopte_prot (pte) ((pte) & ARM_LPAE_PTE_ATTR_MASK) -#define iopte_leaf (pte,l) \ - (l == (ARM_LPAE_MAX_LEVELS - 1) ? \ - (iopte_type(pte,l) == ARM_LPAE_PTE_TYPE_PAGE) : \ - (iopte_type(pte,l) == ARM_LPAE_PTE_TYPE_BLOCK)) -struct arm_lpae_io_pgtable {struct io_pgtable iop;@@ -198,6 +193,15 @@ struct arm_lpae_io_pgtable {typedef u64 arm_lpae_iopte ; +static inline bool iopte_leaf (arm_lpae_iopte pte,int l,unsigned long quirks) +{ +if ((l == (ARM_LPAE_MAX_LEVELS -1 )) && + !(quirks &IO_PGTABLE_QUIRK_ARM_MALI_MIDGARD )) +return iopte_type (pte,l) ==ARM_LPAE_PTE_TYPE_PAGE ; + +return iopte_type (pte,l) ==ARM_LPAE_PTE_TYPE_BLOCK ; +} +static arm_lpae_iopte paddr_to_iopte (phys_addr_t paddr,struct arm_lpae_io_pgtable *data) {@@ -304,11 +308,14 @@ static void __arm_lpae_init_pte(struct arm_lpae_io_pgtable *data, pte |=ARM_LPAE_PTE_NS ;if (lvl ==ARM_LPAE_MAX_LEVELS -1 ) - pte |=ARM_LPAE_PTE_TYPE_PAGE ; + pte |= (data->iop.cfg.quirks &IO_PGTABLE_QUIRK_ARM_MALI_MIDGARD ) ? +ARM_LPAE_PTE_TYPE_BLOCK :ARM_LPAE_PTE_TYPE_PAGE ;else pte |=ARM_LPAE_PTE_TYPE_BLOCK ; - pte |=ARM_LPAE_PTE_AF |ARM_LPAE_PTE_SH_IS ; +if (!(data->iop.cfg.quirks &IO_PGTABLE_QUIRK_ARM_MALI_MIDGARD )) + pte |=ARM_LPAE_PTE_AF ; + pte |=ARM_LPAE_PTE_SH_IS ; pte |=paddr_to_iopte (paddr, data);__arm_lpae_set_pte (ptep, pte, &data->iop.cfg);@@ -321,7 +328,7 @@ static int arm_lpae_init_pte(struct arm_lpae_io_pgtable *data, {arm_lpae_iopte pte = *ptep; -if (iopte_leaf (pte, lvl)) { +if (iopte_leaf (pte, lvl, data->iop.cfg.quirks)) {/* We require an unmap first */ WARN_ON (!selftest_running);return -EEXIST ;@@ -409,7 +416,7 @@ static int __arm_lpae_map(struct arm_lpae_io_pgtable *data, unsigned long iova,__arm_lpae_sync_pte (ptep, cfg); } -if (pte && !iopte_leaf (pte, lvl)) { +if (pte && !iopte_leaf (pte, lvl, cfg->quirks)) { cptep =iopte_deref (pte, data); }else if (pte) {/* We require an unmap first */ @@ -430,12 +437,19 @@ static arm_lpae_iopte arm_lpae_prot_to_pte(struct arm_lpae_io_pgtable *data, data->iop.fmt == ARM_32_LPAE_S1) { pte = ARM_LPAE_PTE_nG; -if (!(prot &IOMMU_WRITE ) && (prot &IOMMU_READ )) - pte |=ARM_LPAE_PTE_AP_RDONLY ; +if (data->iop.cfg.quirks &IO_PGTABLE_QUIRK_ARM_MALI_MIDGARD ) { +if (prot &IOMMU_WRITE ) + pte |=ARM_LPAE_PTE_AP_RDONLY ; -if (!(prot &IOMMU_PRIV )) - pte |=ARM_LPAE_PTE_AP_UNPRIV ; +if (prot &IOMMU_READ ) + pte |=ARM_LPAE_PTE_AP_UNPRIV ; + }else { +if (!(prot &IOMMU_WRITE ) && (prot &IOMMU_READ )) + pte |=ARM_LPAE_PTE_AP_RDONLY ; +if (!(prot &IOMMU_PRIV )) + pte |=ARM_LPAE_PTE_AP_UNPRIV ; + }if (prot &IOMMU_MMIO ) pte |= (ARM_LPAE_MAIR_ATTR_IDX_DEV <<ARM_LPAE_PTE_ATTRINDX_SHIFT );@@ -511,7 +525,7 @@ static void __arm_lpae_free_pgtable(struct arm_lpae_io_pgtable *data, int lvl,while (ptep != end) {arm_lpae_iopte pte = *ptep++; -if (!pte ||iopte_leaf (pte, lvl)) +if (!pte ||iopte_leaf (pte, lvl, data->iop.cfg.quirks))continue ;__arm_lpae_free_pgtable (data, lvl +1 ,iopte_deref (pte, data));@@ -602,7 +616,7 @@ static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,if (size ==ARM_LPAE_BLOCK_SIZE (lvl, data)) {__arm_lpae_set_pte (ptep,0 , &iop->cfg); -if (!iopte_leaf (pte, lvl)) { +if (!iopte_leaf (pte, lvl, iop->cfg.quirks)) {/* Also flush any partial walks */ io_pgtable_tlb_add_flush (iop, iova, size,ARM_LPAE_GRANULE (data), false);@@ -621,7 +635,7 @@ static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data, }return size; - }else if (iopte_leaf (pte, lvl)) { + }else if (iopte_leaf (pte, lvl, iop->cfg.quirks)) {/* * Insert a table at the next level to map the old region, * minus the part we want to unmap @@ -669,7 +683,7 @@ static phys_addr_t arm_lpae_iova_to_phys(struct io_pgtable_ops *ops,return 0 ;/* Leaf entry? */ -if (iopte_leaf (pte,lvl)) +if (iopte_leaf (pte,lvl, data->iop.cfg.quirks)) goto found_translation;/* Take it to the next level */ @@ -779,7 +793,8 @@ 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_ARM_MALI_MIDGARD ))return NULL ; data =arm_lpae_alloc_pgtable (cfg);include/linux/io-pgtable.h +4
-0@@ -75,6 +75,9 @@ struct io_pgtable_cfg {* IO_PGTABLE_QUIRK_NON_STRICT: Skip issuing synchronous leaf TLBIs * on unmap, for DMA domains using the flush queue mechanism for * delayed invalidation. +* +* IO_PGTABLE_QUIRK_ARM_MALI_MIDGARD: ARM Mali Midgard MMU has different +* mapping of access flags and PTE page bits. */ #define IO_PGTABLE_QUIRK_ARM_NS BIT(0)#define IO_PGTABLE_QUIRK_NO_PERMS BIT(1)@@ -82,6 +85,7 @@ struct io_pgtable_cfg {#define IO_PGTABLE_QUIRK_ARM_MTK_4GB BIT(3)#define IO_PGTABLE_QUIRK_NO_DMA BIT(4)#define IO_PGTABLE_QUIRK_NON_STRICT BIT(5) +#define IO_PGTABLE_QUIRK_ARM_MALI_MIDGARD BIT(6)unsigned long quirks;unsigned long pgsize_bitmap;unsigned int ias;
Quoted 16 lines · Robin Murphy
Hi Rob,
Quoted 12 lines · Rob Herring
ARM Mali midgard GPUs have a few differences from standard 64-bit stage 1 page tables.
The 3rd level page entry bits are
0x1instead of0x3for 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 also not set by the vendor driver, but that one doesn't seem to matter.
Quoted 7 lines
Quoted 1 line
Add a quirk to handle all of these differences.
From the look of these changes, this isn't a quirk but a distinct
format. AFAICS from themali_kbasedriver, this must be "LPAE mode" rather than "AArch64 mode", so it seems unlikely that it really supports the full VMSAv8 gamut of granules, address sizes, and page sizes that this patch will happily let through.
Right, but the page size bitmap and the in and out address sizes in the config struct should be enough to restrict those.
What do you propose? Add another init function which hardcodes all those or add some checks of allowed settings?
Rob
Robin Murphy
reviewerQuoted 25 lines · Rob Herring
Quoted 16 lines · Robin Murphy
Hi Rob,
Quoted 12 lines · Rob Herring
ARM Mali midgard GPUs have a few differences from standard 64-bit stage 1 page tables.
The 3rd level page entry bits are
0x1instead of0x3for 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 also not set by the vendor driver, but that one doesn't seem to matter.
Quoted 7 lines
Quoted 1 line
Add a quirk to handle all of these differences.
From the look of these changes, this isn't a quirk but a distinct
format. AFAICS from themali_kbasedriver, this must be "LPAE mode" rather than "AArch64 mode", so it seems unlikely that it really supports the full VMSAv8 gamut of granules, address sizes, and page sizes that this patch will happily let through.
Quoted 5 lines
Right, but the page size bitmap and the in and out address sizes in the config struct should be enough to restrict those.
What do you propose? Add another init function which hardcodes all those or add some checks of allowed settings?
Or rather somewhere in between, i.e. add a new io_pgtable_fmt with an init function which sanity checks/constrains the request appropriately (like the existing ones do), and which could at least help make arm_lpae_pte_to_prot() less of a mess.
We don't treat v7 LPAE as a quirk of v8 AArch64, even though we implement it as a strict subset, and neither do we treat stage 2 as a quirk of stage 1 (or vice versa). This Midgard "LPAE" format appears more different from any of the VMSA long-descriptor formats than they are from each other - it certainly doesn't seem like a case of someone aiming to implement AArch64 stage 1 but getting one or two details slightly wrong.
(I know I implemented Mediatek's Frankein-short-descriptor as a set of quirks rather than introducing multiple v7s formats, but that's largely because it's the only real user of that code, and most of them are generic enough that they could well be common to other formats in future)
Robin.