Hi Joerg,
This series adapts the OMAP IOMMU driver to use the DMA API to flush
the page table/directory table entries from the CPU caches instead of
the ARM assembly calls. The patches are baselined on 4.13-rc1.
Following is the patch summary:
1. Patch 1 disables the OMAP IOMMU fault interrupts instead of
disabling the MMU upon a fault, and resulted in recurring bus
errors during remoteproc recovery on OMAP4. The MMU fault itself
is triggered due to the missing PL310 L2 cache operations, and
this patch fixes the recurring bus errors.
2. The second patch makes the adaptation to the DMA API for flushing
the caches. This fixes the MMU fault triggering issues in the
first place on OMAP4.
I have tested these patches on DRA7, OMAP5, and OMAP4 platforms using
both OMAP IOMMU unit tests and some out-of-tree patches for exercising
the MMUs using the OMAP remoteproc driver.
Laurent,
Appreciate it if you can check the OMAP3ISP functionality with these
patches once.
Regards
Josue
Fernando Guzman Lugo (1):
iommu/omap: Fix disabling of MMU upon a fault
Josue Albarran (1):
iommu/omap: Use DMA-API for performing cache flushes
drivers/iommu/omap-iommu.c | 125 +++++++++++++++++++++++++++++----------------
drivers/iommu/omap-iommu.h | 1 +
2 files changed, 81 insertions(+), 45 deletions(-)
--
2.7.4
The OMAP IOMMU driver was using ARM assembly code directly for
flushing the MMU page table entries from the caches. This caused
MMU faults on OMAP4 (Cortex-A9 based SoCs) as L2 caches were not
handled due to the presence of a PL310 L2 Cache Controller. These
faults were however not seen on OMAP5/DRA7 SoCs (Cortex-A15 based
SoCs).
The OMAP IOMMU driver is adapted to use the DMA Streaming API
instead now to flush the page table/directory table entries from
the CPU caches. This ensures that the devices always see the
updated page table entries. The outer caches are now addressed
automatically with the usage of the DMA API.
Signed-off-by: Josue Albarran <[email protected]>
---
drivers/iommu/omap-iommu.c | 123 +++++++++++++++++++++++++++++----------------
drivers/iommu/omap-iommu.h | 1 +
2 files changed, 80 insertions(+), 44 deletions(-)
diff --git a/drivers/iommu/omap-iommu.c b/drivers/iommu/omap-iommu.c
index 10c9de8de45d..bd67e1b2c64e 100644
--- a/drivers/iommu/omap-iommu.c
+++ b/drivers/iommu/omap-iommu.c
@@ -11,6 +11,7 @@
* published by the Free Software Foundation.
*/
+#include <linux/dma-mapping.h>
#include <linux/err.h>
#include <linux/slab.h>
#include <linux/interrupt.h>
@@ -29,8 +30,6 @@
#include <linux/regmap.h>
#include <linux/mfd/syscon.h>
-#include <asm/cacheflush.h>
-
#include <linux/platform_data/iommu-omap.h>
#include "omap-iopgtable.h"
@@ -454,36 +453,35 @@ static void flush_iotlb_all(struct omap_iommu *obj)
/*
* H/W pagetable operations
*/
-static void flush_iopgd_range(u32 *first, u32 *last)
+static void flush_iopte_range(struct device *dev, dma_addr_t dma,
+ unsigned long offset, int num_entries)
{
- /* FIXME: L2 cache should be taken care of if it exists */
- do {
- asm("mcr p15, 0, %0, c7, c10, 1 @ flush_pgd"
- : : "r" (first));
- first += L1_CACHE_BYTES / sizeof(*first);
- } while (first <= last);
-}
+ size_t size = num_entries * sizeof(u32);
-static void flush_iopte_range(u32 *first, u32 *last)
-{
- /* FIXME: L2 cache should be taken care of if it exists */
- do {
- asm("mcr p15, 0, %0, c7, c10, 1 @ flush_pte"
- : : "r" (first));
- first += L1_CACHE_BYTES / sizeof(*first);
- } while (first <= last);
+ dma_sync_single_range_for_device(dev, dma, offset, size, DMA_TO_DEVICE);
}
-static void iopte_free(u32 *iopte)
+static void iopte_free(struct omap_iommu *obj, u32 *iopte, bool dma_valid)
{
+ dma_addr_t pt_dma;
+
/* Note: freed iopte's must be clean ready for re-use */
- if (iopte)
+ if (iopte) {
+ if (dma_valid) {
+ pt_dma = virt_to_phys(iopte);
+ dma_unmap_single(obj->dev, pt_dma, IOPTE_TABLE_SIZE,
+ DMA_TO_DEVICE);
+ }
+
kmem_cache_free(iopte_cachep, iopte);
+ }
}
-static u32 *iopte_alloc(struct omap_iommu *obj, u32 *iopgd, u32 da)
+static u32 *iopte_alloc(struct omap_iommu *obj, u32 *iopgd,
+ dma_addr_t *pt_dma, u32 da)
{
u32 *iopte;
+ unsigned long offset = iopgd_index(da) * sizeof(da);
/* a table has already existed */
if (*iopgd)
@@ -500,18 +498,38 @@ static u32 *iopte_alloc(struct omap_iommu *obj, u32 *iopgd, u32 da)
if (!iopte)
return ERR_PTR(-ENOMEM);
+ *pt_dma = dma_map_single(obj->dev, iopte, IOPTE_TABLE_SIZE,
+ DMA_TO_DEVICE);
+ if (dma_mapping_error(obj->dev, *pt_dma)) {
+ dev_err(obj->dev, "DMA map error for L2 table\n");
+ iopte_free(obj, iopte, false);
+ return ERR_PTR(-ENOMEM);
+ }
+
+ /*
+ * we rely on dma address and the physical address to be
+ * the same for mapping the L2 table
+ */
+ if (WARN_ON(*pt_dma != virt_to_phys(iopte))) {
+ dev_err(obj->dev, "DMA translation error for L2 table\n");
+ dma_unmap_single(obj->dev, *pt_dma, IOPTE_TABLE_SIZE,
+ DMA_TO_DEVICE);
+ iopte_free(obj, iopte, false);
+ return ERR_PTR(-ENOMEM);
+ }
+
*iopgd = virt_to_phys(iopte) | IOPGD_TABLE;
- flush_iopgd_range(iopgd, iopgd);
+ flush_iopte_range(obj->dev, obj->pd_dma, offset, 1);
dev_vdbg(obj->dev, "%s: a new pte:%p\n", __func__, iopte);
} else {
/* We raced, free the reduniovant table */
- iopte_free(iopte);
+ iopte_free(obj, iopte, false);
}
pte_ready:
iopte = iopte_offset(iopgd, da);
-
+ *pt_dma = virt_to_phys(iopte);
dev_vdbg(obj->dev,
"%s: da:%08x pgd:%p *pgd:%08x pte:%p *pte:%08x\n",
__func__, da, iopgd, *iopgd, iopte, *iopte);
@@ -522,6 +540,7 @@ static u32 *iopte_alloc(struct omap_iommu *obj, u32 *iopgd, u32 da)
static int iopgd_alloc_section(struct omap_iommu *obj, u32 da, u32 pa, u32 prot)
{
u32 *iopgd = iopgd_offset(obj, da);
+ unsigned long offset = iopgd_index(da) * sizeof(da);
if ((da | pa) & ~IOSECTION_MASK) {
dev_err(obj->dev, "%s: %08x:%08x should aligned on %08lx\n",
@@ -530,13 +549,14 @@ static int iopgd_alloc_section(struct omap_iommu *obj, u32 da, u32 pa, u32 prot)
}
*iopgd = (pa & IOSECTION_MASK) | prot | IOPGD_SECTION;
- flush_iopgd_range(iopgd, iopgd);
+ flush_iopte_range(obj->dev, obj->pd_dma, offset, 1);
return 0;
}
static int iopgd_alloc_super(struct omap_iommu *obj, u32 da, u32 pa, u32 prot)
{
u32 *iopgd = iopgd_offset(obj, da);
+ unsigned long offset = iopgd_index(da) * sizeof(da);
int i;
if ((da | pa) & ~IOSUPER_MASK) {
@@ -547,20 +567,22 @@ static int iopgd_alloc_super(struct omap_iommu *obj, u32 da, u32 pa, u32 prot)
for (i = 0; i < 16; i++)
*(iopgd + i) = (pa & IOSUPER_MASK) | prot | IOPGD_SUPER;
- flush_iopgd_range(iopgd, iopgd + 15);
+ flush_iopte_range(obj->dev, obj->pd_dma, offset, 16);
return 0;
}
static int iopte_alloc_page(struct omap_iommu *obj, u32 da, u32 pa, u32 prot)
{
u32 *iopgd = iopgd_offset(obj, da);
- u32 *iopte = iopte_alloc(obj, iopgd, da);
+ dma_addr_t pt_dma;
+ u32 *iopte = iopte_alloc(obj, iopgd, &pt_dma, da);
+ unsigned long offset = iopte_index(da) * sizeof(da);
if (IS_ERR(iopte))
return PTR_ERR(iopte);
*iopte = (pa & IOPAGE_MASK) | prot | IOPTE_SMALL;
- flush_iopte_range(iopte, iopte);
+ flush_iopte_range(obj->dev, pt_dma, offset, 1);
dev_vdbg(obj->dev, "%s: da:%08x pa:%08x pte:%p *pte:%08x\n",
__func__, da, pa, iopte, *iopte);
@@ -571,7 +593,9 @@ static int iopte_alloc_page(struct omap_iommu *obj, u32 da, u32 pa, u32 prot)
static int iopte_alloc_large(struct omap_iommu *obj, u32 da, u32 pa, u32 prot)
{
u32 *iopgd = iopgd_offset(obj, da);
- u32 *iopte = iopte_alloc(obj, iopgd, da);
+ dma_addr_t pt_dma;
+ u32 *iopte = iopte_alloc(obj, iopgd, &pt_dma, da);
+ unsigned long offset = iopte_index(da) * sizeof(da);
int i;
if ((da | pa) & ~IOLARGE_MASK) {
@@ -585,7 +609,7 @@ static int iopte_alloc_large(struct omap_iommu *obj, u32 da, u32 pa, u32 prot)
for (i = 0; i < 16; i++)
*(iopte + i) = (pa & IOLARGE_MASK) | prot | IOPTE_LARGE;
- flush_iopte_range(iopte, iopte + 15);
+ flush_iopte_range(obj->dev, pt_dma, offset, 16);
return 0;
}
@@ -674,6 +698,9 @@ static size_t iopgtable_clear_entry_core(struct omap_iommu *obj, u32 da)
size_t bytes;
u32 *iopgd = iopgd_offset(obj, da);
int nent = 1;
+ dma_addr_t pt_dma;
+ unsigned long pd_offset = iopgd_index(da) * sizeof(da);
+ unsigned long pt_offset = iopte_index(da) * sizeof(da);
if (!*iopgd)
return 0;
@@ -690,7 +717,8 @@ static size_t iopgtable_clear_entry_core(struct omap_iommu *obj, u32 da)
}
bytes *= nent;
memset(iopte, 0, nent * sizeof(*iopte));
- flush_iopte_range(iopte, iopte + (nent - 1) * sizeof(*iopte));
+ pt_dma = virt_to_phys(iopte);
+ flush_iopte_range(obj->dev, pt_dma, pt_offset, nent);
/*
* do table walk to check if this table is necessary or not
@@ -700,7 +728,7 @@ static size_t iopgtable_clear_entry_core(struct omap_iommu *obj, u32 da)
if (iopte[i])
goto out;
- iopte_free(iopte);
+ iopte_free(obj, iopte, true);
nent = 1; /* for the next L1 entry */
} else {
bytes = IOPGD_SIZE;
@@ -712,7 +740,7 @@ static size_t iopgtable_clear_entry_core(struct omap_iommu *obj, u32 da)
bytes *= nent;
}
memset(iopgd, 0, nent * sizeof(*iopgd));
- flush_iopgd_range(iopgd, iopgd + (nent - 1) * sizeof(*iopgd));
+ flush_iopte_range(obj->dev, obj->pd_dma, pd_offset, nent);
out:
return bytes;
}
@@ -738,6 +766,7 @@ static size_t iopgtable_clear_entry(struct omap_iommu *obj, u32 da)
static void iopgtable_clear_entry_all(struct omap_iommu *obj)
{
+ unsigned long offset;
int i;
spin_lock(&obj->page_table_lock);
@@ -748,15 +777,16 @@ static void iopgtable_clear_entry_all(struct omap_iommu *obj)
da = i << IOPGD_SHIFT;
iopgd = iopgd_offset(obj, da);
+ offset = iopgd_index(da) * sizeof(da);
if (!*iopgd)
continue;
if (iopgd_is_table(*iopgd))
- iopte_free(iopte_offset(iopgd, 0));
+ iopte_free(obj, iopte_offset(iopgd, 0), true);
*iopgd = 0;
- flush_iopgd_range(iopgd, iopgd);
+ flush_iopte_range(obj->dev, obj->pd_dma, offset, 1);
}
flush_iotlb_all(obj);
@@ -815,10 +845,18 @@ static int omap_iommu_attach(struct omap_iommu *obj, u32 *iopgd)
spin_lock(&obj->iommu_lock);
+ obj->pd_dma = dma_map_single(obj->dev, iopgd, IOPGD_TABLE_SIZE,
+ DMA_TO_DEVICE);
+ if (dma_mapping_error(obj->dev, obj->pd_dma)) {
+ dev_err(obj->dev, "DMA map error for L1 table\n");
+ err = -ENOMEM;
+ goto out_err;
+ }
+
obj->iopgd = iopgd;
err = iommu_enable(obj);
if (err)
- goto err_enable;
+ goto out_err;
flush_iotlb_all(obj);
spin_unlock(&obj->iommu_lock);
@@ -827,7 +865,7 @@ static int omap_iommu_attach(struct omap_iommu *obj, u32 *iopgd)
return 0;
-err_enable:
+out_err:
spin_unlock(&obj->iommu_lock);
return err;
@@ -844,7 +882,10 @@ static void omap_iommu_detach(struct omap_iommu *obj)
spin_lock(&obj->iommu_lock);
+ dma_unmap_single(obj->dev, obj->pd_dma, IOPGD_TABLE_SIZE,
+ DMA_TO_DEVICE);
iommu_disable(obj);
+ obj->pd_dma = 0;
obj->iopgd = NULL;
spin_unlock(&obj->iommu_lock);
@@ -1008,11 +1049,6 @@ static struct platform_driver omap_iommu_driver = {
},
};
-static void iopte_cachep_ctor(void *iopte)
-{
- clean_dcache_area(iopte, IOPTE_TABLE_SIZE);
-}
-
static u32 iotlb_init_entry(struct iotlb_entry *e, u32 da, u32 pa, int pgsz)
{
memset(e, 0, sizeof(*e));
@@ -1159,7 +1195,6 @@ static struct iommu_domain *omap_iommu_domain_alloc(unsigned type)
if (WARN_ON(!IS_ALIGNED((long)omap_domain->pgtable, IOPGD_TABLE_SIZE)))
goto fail_align;
- clean_dcache_area(omap_domain->pgtable, IOPGD_TABLE_SIZE);
spin_lock_init(&omap_domain->lock);
omap_domain->domain.geometry.aperture_start = 0;
@@ -1347,7 +1382,7 @@ static int __init omap_iommu_init(void)
of_node_put(np);
p = kmem_cache_create("iopte_cache", IOPTE_TABLE_SIZE, align, flags,
- iopte_cachep_ctor);
+ NULL);
if (!p)
return -ENOMEM;
iopte_cachep = p;
diff --git a/drivers/iommu/omap-iommu.h b/drivers/iommu/omap-iommu.h
index 6e70515e6038..a675af29a6ec 100644
--- a/drivers/iommu/omap-iommu.h
+++ b/drivers/iommu/omap-iommu.h
@@ -61,6 +61,7 @@ struct omap_iommu {
*/
u32 *iopgd;
spinlock_t page_table_lock; /* protect iopgd */
+ dma_addr_t pd_dma;
int nr_tlb_entries;
--
2.7.4
From: Fernando Guzman Lugo <[email protected]>
The IOMMU framework lets its client users be notified on a
MMU fault and allows them to either handle the interrupt by
dynamic reloading of an appropriate TLB/PTE for the offending
fault address or to completely restart/recovery the device
and its IOMMU.
The OMAP remoteproc driver performs the latter option, and
does so after unwinding the previous mappings. The OMAP IOMMU
fault handler however disables the MMU and cuts off the clock
upon a MMU fault at present, resulting in an interconnect abort
during any subsequent operation that touches the MMU registers.
So, disable the IP-level fault interrupts instead of disabling
the MMU, to allow continued MMU register operations as well as
to avoid getting interrupted again.
Signed-off-by: Fernando Guzman Lugo <[email protected]>
[[email protected]: add commit description]
Signed-off-by: Suman Anna <[email protected]>
Signed-off-by: Josue Albarran <[email protected]>
---
drivers/iommu/omap-iommu.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/iommu/omap-iommu.c b/drivers/iommu/omap-iommu.c
index 641e035cf866..10c9de8de45d 100644
--- a/drivers/iommu/omap-iommu.c
+++ b/drivers/iommu/omap-iommu.c
@@ -786,7 +786,7 @@ static irqreturn_t iommu_fault_handler(int irq, void *data)
if (!report_iommu_fault(domain, obj->dev, da, 0))
return IRQ_HANDLED;
- iommu_disable(obj);
+ iommu_write_reg(obj, 0, MMU_IRQENABLE);
iopgd = iopgd_offset(obj, da);
--
2.7.4
On 07/28/2017 03:49 PM, Josue Albarran wrote:
> The OMAP IOMMU driver was using ARM assembly code directly for
> flushing the MMU page table entries from the caches. This caused
> MMU faults on OMAP4 (Cortex-A9 based SoCs) as L2 caches were not
> handled due to the presence of a PL310 L2 Cache Controller. These
> faults were however not seen on OMAP5/DRA7 SoCs (Cortex-A15 based
> SoCs).
>
> The OMAP IOMMU driver is adapted to use the DMA Streaming API
> instead now to flush the page table/directory table entries from
> the CPU caches. This ensures that the devices always see the
> updated page table entries. The outer caches are now addressed
> automatically with the usage of the DMA API.
>
> Signed-off-by: Josue Albarran <[email protected]>
Thanks for fixing this,
Acked-by: Suman Anna <[email protected]>
[snip]
On Fri, Jul 28, 2017 at 03:49:12PM -0500, Josue Albarran wrote:
> Fernando Guzman Lugo (1):
> iommu/omap: Fix disabling of MMU upon a fault
>
> Josue Albarran (1):
> iommu/omap: Use DMA-API for performing cache flushes
>
> drivers/iommu/omap-iommu.c | 125 +++++++++++++++++++++++++++++----------------
> drivers/iommu/omap-iommu.h | 1 +
> 2 files changed, 81 insertions(+), 45 deletions(-)
Applied, thanks.
Hi Josue,
Thank you for the patches.
On Friday 28 Jul 2017 15:49:12 Josue Albarran wrote:
> Hi Joerg,
>
> This series adapts the OMAP IOMMU driver to use the DMA API to flush
> the page table/directory table entries from the CPU caches instead of
> the ARM assembly calls. The patches are baselined on 4.13-rc1.
>
> Following is the patch summary:
> 1. Patch 1 disables the OMAP IOMMU fault interrupts instead of
> disabling the MMU upon a fault, and resulted in recurring bus
> errors during remoteproc recovery on OMAP4. The MMU fault itself
> is triggered due to the missing PL310 L2 cache operations, and
> this patch fixes the recurring bus errors.
> 2. The second patch makes the adaptation to the DMA API for flushing
> the caches. This fixes the MMU fault triggering issues in the
> first place on OMAP4.
>
> I have tested these patches on DRA7, OMAP5, and OMAP4 platforms using
> both OMAP IOMMU unit tests and some out-of-tree patches for exercising
> the MMUs using the OMAP remoteproc driver.
>
> Laurent,
> Appreciate it if you can check the OMAP3ISP functionality with these
> patches once.
I apologize for the delay, I had to resurrect my Beagleboard-xM, which
involved updating and then debugging U-Boot.
Tested-by: Laurent Pinchart <[email protected]>
--
Regards,
Laurent Pinchart