Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]
Groups > linux.kernel > #1396541 > unrolled thread
| Started by | Shanker Donthineni <shankerd@codeaurora.org> |
|---|---|
| First post | 2016-05-09 04:20 +0200 |
| Last post | 2016-05-09 06:10 +0200 |
| Articles | 2 — 1 participant |
Back to article view | Back to linux.kernel
This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by
below is the oldest one visible, not the original post.
[PATCH V2 2/2] irqchip/gicv3-its: Implement two-level(indirect) device table support Shanker Donthineni <shankerd@codeaurora.org> - 2016-05-09 04:20 +0200
Re: [PATCH V2 2/2] irqchip/gicv3-its: Implement two-level(indirect) device table support Shanker Donthineni <shankerd@codeaurora.org> - 2016-05-09 06:10 +0200
| From | Shanker Donthineni <shankerd@codeaurora.org> |
|---|---|
| Date | 2016-05-09 04:20 +0200 |
| Subject | [PATCH V2 2/2] irqchip/gicv3-its: Implement two-level(indirect) device table support |
| Message-ID | <rwIXD-2yN-1@gated-at.bofh.it> |
Since device IDs are extremely sparse, the single, a.k.a flat table is
not sufficient for the following two reasons.
1) According to ARM-GIC spec, ITS hw can access maximum of 256(pages)*
64K(pageszie) bytes. In the best case, it supports upto DEVid=21
sparse with minimum device table entry size 8bytes.
2) The maximum memory size that is possible without memblock depends on
MAX_ORDER. 4MB on 4K page size kernel with default MAX_ORDER, so it
supports DEVid range 19bits.
The two-level device table feature brings us two advantages, the first
is a very high possibility of supporting upto 32bit sparse, and the
second one is the best utilization of memory allocation.
The feature is enabled automatically during driver probe if a single
ITS page is not adequate for flat table and the hardware is capable
of two-level table walk.
Signed-off-by: Shanker Donthineni <shankerd@codeaurora.org>
---
This patch is based on Marc Zyngier's branch https://git.kernel.org/cgit/linux/kernel/git/maz/arm-platforms.git/log/?h=irq/irqchip-4.7
I have tested the Indirection feature on Qualcomm Technologies QDF2XXX server platform.
Changes since v1:
Most of this patch has been rewritten after refactoring its_alloc_tables().
Always enable device two-level if the memory requirement is more than PAGE_SIZE.
Fixed the coding bug that breaks on the BE machine.
Edited the commit text.
drivers/irqchip/irq-gic-v3-its.c | 100 ++++++++++++++++++++++++++++++++-------
1 file changed, 83 insertions(+), 17 deletions(-)
diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
index b23e00c..27be792 100644
--- a/drivers/irqchip/irq-gic-v3-its.c
+++ b/drivers/irqchip/irq-gic-v3-its.c
@@ -938,6 +938,18 @@ retry_baser:
return 0;
}
+/**
+ * Find out whether an implemented baser register supports a single, flat table
+ * or a two-level table by reading bit offset at '62' after writing '1' to it.
+ */
+static u64 its_baser_check_indirect(struct its_baser *baser)
+{
+ u64 val = GITS_BASER_InnerShareable | GITS_BASER_WaWb;
+
+ writeq_relaxed(val | GITS_BASER_INDIRECT, baser->hwreg);
+ return (readq_relaxed(baser->hwreg) & GITS_BASER_INDIRECT);
+}
+
static int its_alloc_tables(const char *node_name, struct its_node *its)
{
u64 typer = readq_relaxed(its->base + GITS_TYPER);
@@ -964,6 +976,7 @@ static int its_alloc_tables(const char *node_name, struct its_node *its)
u64 entry_size = GITS_BASER_ENTRY_SIZE(val);
int order = get_order(psz);
struct its_baser *baser = its->tables + i;
+ u64 indirect = 0;
if (type == GITS_BASER_TYPE_NONE)
continue;
@@ -977,17 +990,27 @@ static int its_alloc_tables(const char *node_name, struct its_node *its)
* Allocate as many entries as required to fit the
* range of device IDs that the ITS can grok... The ID
* space being incredibly sparse, this results in a
- * massive waste of memory.
+ * massive waste of memory if two-level device table
+ * feature is not supported by hardware.
*
* For other tables, only allocate a single page.
*/
if (type == GITS_BASER_TYPE_DEVICE) {
- /*
- * 'order' was initialized earlier to the default page
- * granule of the the ITS. We can't have an allocation
- * smaller than that. If the requested allocation
- * is smaller, round up to the default page granule.
- */
+ if ((entry_size << ids) > psz)
+ indirect = its_baser_check_indirect(baser);
+
+ if (indirect) {
+ /*
+ * The size of the lvl2 table is equal to ITS
+ * page size which is 'psz'. For computing lvl1
+ * table size, subtract ID bits that sparse
+ * lvl2 table from 'ids' which is reported by
+ * ITS hardware times lvl1 table entry size.
+ */
+ ids -= ilog2(psz / entry_size);
+ entry_size = GITS_LVL1_ENTRY_SIZE;
+ }
+
order = max(get_order(entry_size << ids), order);
if (order >= MAX_ORDER) {
order = MAX_ORDER - 1;
@@ -997,7 +1020,7 @@ static int its_alloc_tables(const char *node_name, struct its_node *its)
}
}
- err = its_baser_setup(its, baser, order, 0);
+ err = its_baser_setup(its, baser, order, indirect);
if (err < 0) {
its_free_tables(its);
return err;
@@ -1187,10 +1210,60 @@ static struct its_baser *its_get_baser(struct its_node *its, u32 type)
return NULL;
}
+static bool its_alloc_device_table(struct its_node *its, u32 dev_id)
+{
+ struct its_baser *baser;
+ struct page *page;
+ u32 entry_size, idx;
+ u64 *table;
+
+ baser = its_get_baser(its, GITS_BASER_TYPE_DEVICE);
+
+ /* Don't allow 'dev_id' that exceeds ITS hardware limit */
+ if (!baser && (ilog2(dev_id) >= its->device_ids))
+ return false;
+
+ /* Don't allow 'dev_id' that exceeds single, flat table limit */
+ entry_size = GITS_BASER_ENTRY_SIZE(baser->val);
+ if (!(baser->val & GITS_BASER_INDIRECT)) {
+ if (dev_id >= (PAGE_ORDER_TO_SIZE(baser->order) / entry_size))
+ return false;
+ return true;
+ }
+
+ /* Compute Level1 table index & check if that exceeds table range */
+ idx = dev_id >> ilog2(baser->psz / entry_size);
+ if (idx >= (PAGE_ORDER_TO_SIZE(baser->order) / GITS_LVL1_ENTRY_SIZE))
+ return false;
+
+ table = baser->base;
+
+ /* Allocate memory for Level2 table */
+ if (!table[idx]) {
+ page = alloc_pages(GFP_KERNEL | __GFP_ZERO, get_order(baser->psz));
+ if (!page)
+ return false;
+
+ /* Flush memory to PoC if hardware doesn't support coherency */
+ if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
+ __flush_dcache_area(page_address(page), baser->psz);
+
+ table[idx] = cpu_to_le64(page_to_phys(page) | GITS_BASER_VALID);
+
+ /* Flush memory to PoC if hardware doesn't support coherency */
+ if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
+ __flush_dcache_area(table + idx, GITS_LVL1_ENTRY_SIZE);
+
+ /* Ensure updated table contents are visible to ITS hardware */
+ dsb(sy);
+ }
+
+ return true;
+}
+
static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
int nvecs)
{
- struct its_baser *baser;
struct its_device *dev;
unsigned long *lpi_map;
unsigned long flags;
@@ -1201,14 +1274,7 @@ static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
int nr_ites;
int sz;
- baser = its_get_baser(its, GITS_BASER_TYPE_DEVICE);
-
- /* Don't allow 'dev_id' that exceeds single, flat table limit */
- if (baser) {
- if (dev_id >= (PAGE_ORDER_TO_SIZE(baser->order) /
- GITS_BASER_ENTRY_SIZE(baser->val)))
- return NULL;
- } else if (ilog2(dev_id) >= its->device_ids)
+ if (!its_alloc_device_table(its, dev_id))
return NULL;
dev = kzalloc(sizeof(*dev), GFP_KERNEL);
--
Qualcomm Technologies, Inc. on behalf of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
[toc] | [next] | [standalone]
| From | Shanker Donthineni <shankerd@codeaurora.org> |
|---|---|
| Date | 2016-05-09 06:10 +0200 |
| Subject | Re: [PATCH V2 2/2] irqchip/gicv3-its: Implement two-level(indirect) device table support |
| Message-ID | <rwKG6-4Ch-5@gated-at.bofh.it> |
| In reply to | #1396541 |
On 05/08/2016 09:14 PM, Shanker Donthineni wrote:
> Since device IDs are extremely sparse, the single, a.k.a flat table is
> not sufficient for the following two reasons.
>
> 1) According to ARM-GIC spec, ITS hw can access maximum of 256(pages)*
> 64K(pageszie) bytes. In the best case, it supports upto DEVid=21
> sparse with minimum device table entry size 8bytes.
>
> 2) The maximum memory size that is possible without memblock depends on
> MAX_ORDER. 4MB on 4K page size kernel with default MAX_ORDER, so it
> supports DEVid range 19bits.
>
> The two-level device table feature brings us two advantages, the first
> is a very high possibility of supporting upto 32bit sparse, and the
> second one is the best utilization of memory allocation.
>
> The feature is enabled automatically during driver probe if a single
> ITS page is not adequate for flat table and the hardware is capable
> of two-level table walk.
>
> Signed-off-by: Shanker Donthineni <shankerd@codeaurora.org>
> ---
>
> This patch is based on Marc Zyngier's branch https://git.kernel.org/cgit/linux/kernel/git/maz/arm-platforms.git/log/?h=irq/irqchip-4.7
>
> I have tested the Indirection feature on Qualcomm Technologies QDF2XXX server platform.
>
> Changes since v1:
> Most of this patch has been rewritten after refactoring its_alloc_tables().
> Always enable device two-level if the memory requirement is more than PAGE_SIZE.
> Fixed the coding bug that breaks on the BE machine.
> Edited the commit text.
>
> drivers/irqchip/irq-gic-v3-its.c | 100 ++++++++++++++++++++++++++++++++-------
> 1 file changed, 83 insertions(+), 17 deletions(-)
>
> diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
> index b23e00c..27be792 100644
> --- a/drivers/irqchip/irq-gic-v3-its.c
> +++ b/drivers/irqchip/irq-gic-v3-its.c
> @@ -938,6 +938,18 @@ retry_baser:
> return 0;
> }
>
> +/**
> + * Find out whether an implemented baser register supports a single, flat table
> + * or a two-level table by reading bit offset at '62' after writing '1' to it.
> + */
> +static u64 its_baser_check_indirect(struct its_baser *baser)
> +{
> + u64 val = GITS_BASER_InnerShareable | GITS_BASER_WaWb;
> +
> + writeq_relaxed(val | GITS_BASER_INDIRECT, baser->hwreg);
> + return (readq_relaxed(baser->hwreg) & GITS_BASER_INDIRECT);
> +}
> +
> static int its_alloc_tables(const char *node_name, struct its_node *its)
> {
> u64 typer = readq_relaxed(its->base + GITS_TYPER);
> @@ -964,6 +976,7 @@ static int its_alloc_tables(const char *node_name, struct its_node *its)
> u64 entry_size = GITS_BASER_ENTRY_SIZE(val);
> int order = get_order(psz);
> struct its_baser *baser = its->tables + i;
> + u64 indirect = 0;
>
> if (type == GITS_BASER_TYPE_NONE)
> continue;
> @@ -977,17 +990,27 @@ static int its_alloc_tables(const char *node_name, struct its_node *its)
> * Allocate as many entries as required to fit the
> * range of device IDs that the ITS can grok... The ID
> * space being incredibly sparse, this results in a
> - * massive waste of memory.
> + * massive waste of memory if two-level device table
> + * feature is not supported by hardware.
> *
> * For other tables, only allocate a single page.
> */
> if (type == GITS_BASER_TYPE_DEVICE) {
> - /*
> - * 'order' was initialized earlier to the default page
> - * granule of the the ITS. We can't have an allocation
> - * smaller than that. If the requested allocation
> - * is smaller, round up to the default page granule.
> - */
> + if ((entry_size << ids) > psz)
> + indirect = its_baser_check_indirect(baser);
> +
> + if (indirect) {
> + /*
> + * The size of the lvl2 table is equal to ITS
> + * page size which is 'psz'. For computing lvl1
> + * table size, subtract ID bits that sparse
> + * lvl2 table from 'ids' which is reported by
> + * ITS hardware times lvl1 table entry size.
> + */
> + ids -= ilog2(psz / entry_size);
> + entry_size = GITS_LVL1_ENTRY_SIZE;
> + }
> +
> order = max(get_order(entry_size << ids), order);
> if (order >= MAX_ORDER) {
> order = MAX_ORDER - 1;
> @@ -997,7 +1020,7 @@ static int its_alloc_tables(const char *node_name, struct its_node *its)
> }
> }
>
> - err = its_baser_setup(its, baser, order, 0);
> + err = its_baser_setup(its, baser, order, indirect);
> if (err < 0) {
> its_free_tables(its);
> return err;
> @@ -1187,10 +1210,60 @@ static struct its_baser *its_get_baser(struct its_node *its, u32 type)
> return NULL;
> }
>
> +static bool its_alloc_device_table(struct its_node *its, u32 dev_id)
> +{
> + struct its_baser *baser;
> + struct page *page;
> + u32 entry_size, idx;
> + u64 *table;
> +
> + baser = its_get_baser(its, GITS_BASER_TYPE_DEVICE);
> +
> + /* Don't allow 'dev_id' that exceeds ITS hardware limit */
> + if (!baser && (ilog2(dev_id) >= its->device_ids))
> + return false;
> +
I have missed this change, will be fixed in v3 patchset.
if (!baser) {
if (ilog2(dev_id) >= its->device_ids))
return false;
return true;
}
> + /* Don't allow 'dev_id' that exceeds single, flat table limit */
> + entry_size = GITS_BASER_ENTRY_SIZE(baser->val);
> + if (!(baser->val & GITS_BASER_INDIRECT)) {
> + if (dev_id >= (PAGE_ORDER_TO_SIZE(baser->order) / entry_size))
> + return false;
> + return true;
> + }
> +
> + /* Compute Level1 table index & check if that exceeds table range */
> + idx = dev_id >> ilog2(baser->psz / entry_size);
> + if (idx >= (PAGE_ORDER_TO_SIZE(baser->order) / GITS_LVL1_ENTRY_SIZE))
> + return false;
> +
> + table = baser->base;
> +
> + /* Allocate memory for Level2 table */
> + if (!table[idx]) {
> + page = alloc_pages(GFP_KERNEL | __GFP_ZERO, get_order(baser->psz));
> + if (!page)
> + return false;
> +
> + /* Flush memory to PoC if hardware doesn't support coherency */
> + if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
> + __flush_dcache_area(page_address(page), baser->psz);
> +
> + table[idx] = cpu_to_le64(page_to_phys(page) | GITS_BASER_VALID);
> +
> + /* Flush memory to PoC if hardware doesn't support coherency */
> + if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
> + __flush_dcache_area(table + idx, GITS_LVL1_ENTRY_SIZE);
> +
> + /* Ensure updated table contents are visible to ITS hardware */
> + dsb(sy);
> + }
> +
> + return true;
> +}
> +
> static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
> int nvecs)
> {
> - struct its_baser *baser;
> struct its_device *dev;
> unsigned long *lpi_map;
> unsigned long flags;
> @@ -1201,14 +1274,7 @@ static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
> int nr_ites;
> int sz;
>
> - baser = its_get_baser(its, GITS_BASER_TYPE_DEVICE);
> -
> - /* Don't allow 'dev_id' that exceeds single, flat table limit */
> - if (baser) {
> - if (dev_id >= (PAGE_ORDER_TO_SIZE(baser->order) /
> - GITS_BASER_ENTRY_SIZE(baser->val)))
> - return NULL;
> - } else if (ilog2(dev_id) >= its->device_ids)
> + if (!its_alloc_device_table(its, dev_id))
> return NULL;
>
> dev = kzalloc(sizeof(*dev), GFP_KERNEL);
--
Shanker Donthineni
Qualcomm Technologies, Inc. on behalf of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project
[toc] | [prev] | [standalone]
Back to top | Article view | linux.kernel
csiph-web