Re: [PATCH 04/10] irqchip/gic-v3-its: Move pending table allocation to init time
Hi Julien, On 24/09/18 12:58, Julien Thierry wrote: Hi Marc, On 21/09/18 20:59, Marc Zyngier wrote: Pending tables for the redistributors are currently allocated one at a time as each CPU boots. This is causing some grief for Linux/RT (allocation from within a CPU hotplug notifier is frown upon). Let's more this allocation to take place at init time, when we only have a single CPU. It means we're allocating memory for CPUs that are not online yet, but most system will boot all of their CPUs anyway, so that's not completely wasted. Signed-off-by: Marc Zyngier --- drivers/irqchip/irq-gic-v3-its.c | 80 +++--- include/linux/irqchip/arm-gic-v3.h | 1 + 2 files changed, 53 insertions(+), 28 deletions(-) diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c index 7ef6baea2d78..462bba422189 100644 --- a/drivers/irqchip/irq-gic-v3-its.c +++ b/drivers/irqchip/irq-gic-v3-its.c @@ -173,6 +173,7 @@ static DEFINE_RAW_SPINLOCK(vmovp_lock); static DEFINE_IDA(its_vpeid_ida); #define gic_data_rdist() (raw_cpu_ptr(gic_rdists->rdist)) +#define gic_data_rdist_cpu(cpu)(per_cpu_ptr(gic_rdists->rdist, cpu)) #define gic_data_rdist_rd_base() (gic_data_rdist()->rd_base) #define gic_data_rdist_vlpi_base() (gic_data_rdist_rd_base() + SZ_128K) @@ -1625,7 +1626,7 @@ static void its_free_prop_table(struct page *prop_page) get_order(LPI_PROPBASE_SZ)); } -static int __init its_alloc_lpi_tables(void) +static int __init its_alloc_lpi_prop_table(void) A bit of a nit, but there is already a function called "its_allocate_prop_table" which I find very easy to confuse with this one. And patch 3 factored the initialization out of its_allocate_prop_table. So I was wondering whether it would not actually be better to open-code it here and get rid of that function. Otherwise I'd suggest having more distinct names. its_allocate_prop_table is also used by the VLPI code to allocate guest property tables, so I'd rather not open-code it. How about renaming this function to its_setup_lpi_prop_table? Thanks, M. -- Jazz is not dead. It just smells funny...
Re: [PATCH 04/10] irqchip/gic-v3-its: Move pending table allocation to init time
Hi Julien, On 24/09/18 12:58, Julien Thierry wrote: Hi Marc, On 21/09/18 20:59, Marc Zyngier wrote: Pending tables for the redistributors are currently allocated one at a time as each CPU boots. This is causing some grief for Linux/RT (allocation from within a CPU hotplug notifier is frown upon). Let's more this allocation to take place at init time, when we only have a single CPU. It means we're allocating memory for CPUs that are not online yet, but most system will boot all of their CPUs anyway, so that's not completely wasted. Signed-off-by: Marc Zyngier --- drivers/irqchip/irq-gic-v3-its.c | 80 +++--- include/linux/irqchip/arm-gic-v3.h | 1 + 2 files changed, 53 insertions(+), 28 deletions(-) diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c index 7ef6baea2d78..462bba422189 100644 --- a/drivers/irqchip/irq-gic-v3-its.c +++ b/drivers/irqchip/irq-gic-v3-its.c @@ -173,6 +173,7 @@ static DEFINE_RAW_SPINLOCK(vmovp_lock); static DEFINE_IDA(its_vpeid_ida); #define gic_data_rdist() (raw_cpu_ptr(gic_rdists->rdist)) +#define gic_data_rdist_cpu(cpu)(per_cpu_ptr(gic_rdists->rdist, cpu)) #define gic_data_rdist_rd_base() (gic_data_rdist()->rd_base) #define gic_data_rdist_vlpi_base() (gic_data_rdist_rd_base() + SZ_128K) @@ -1625,7 +1626,7 @@ static void its_free_prop_table(struct page *prop_page) get_order(LPI_PROPBASE_SZ)); } -static int __init its_alloc_lpi_tables(void) +static int __init its_alloc_lpi_prop_table(void) A bit of a nit, but there is already a function called "its_allocate_prop_table" which I find very easy to confuse with this one. And patch 3 factored the initialization out of its_allocate_prop_table. So I was wondering whether it would not actually be better to open-code it here and get rid of that function. Otherwise I'd suggest having more distinct names. its_allocate_prop_table is also used by the VLPI code to allocate guest property tables, so I'd rather not open-code it. How about renaming this function to its_setup_lpi_prop_table? Thanks, M. -- Jazz is not dead. It just smells funny...
Re: [PATCH 04/10] irqchip/gic-v3-its: Move pending table allocation to init time
Hi Marc,
On 21/09/18 20:59, Marc Zyngier wrote:
Pending tables for the redistributors are currently allocated
one at a time as each CPU boots. This is causing some grief
for Linux/RT (allocation from within a CPU hotplug notifier is
frown upon).
Let's more this allocation to take place at init time, when we
only have a single CPU. It means we're allocating memory for CPUs
that are not online yet, but most system will boot all of their
CPUs anyway, so that's not completely wasted.
Signed-off-by: Marc Zyngier
---
drivers/irqchip/irq-gic-v3-its.c | 80 +++---
include/linux/irqchip/arm-gic-v3.h | 1 +
2 files changed, 53 insertions(+), 28 deletions(-)
diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
index 7ef6baea2d78..462bba422189 100644
--- a/drivers/irqchip/irq-gic-v3-its.c
+++ b/drivers/irqchip/irq-gic-v3-its.c
@@ -173,6 +173,7 @@ static DEFINE_RAW_SPINLOCK(vmovp_lock);
static DEFINE_IDA(its_vpeid_ida);
#define gic_data_rdist() (raw_cpu_ptr(gic_rdists->rdist))
+#define gic_data_rdist_cpu(cpu)(per_cpu_ptr(gic_rdists->rdist,
cpu))
#define gic_data_rdist_rd_base() (gic_data_rdist()->rd_base)
#define gic_data_rdist_vlpi_base()(gic_data_rdist_rd_base() + SZ_128K)
@@ -1625,7 +1626,7 @@ static void its_free_prop_table(struct page *prop_page)
get_order(LPI_PROPBASE_SZ));
}
-static int __init its_alloc_lpi_tables(void)
+static int __init its_alloc_lpi_prop_table(void)
A bit of a nit, but there is already a function called
"its_allocate_prop_table" which I find very easy to confuse with this one.
And patch 3 factored the initialization out of its_allocate_prop_table.
So I was wondering whether it would not actually be better to open-code
it here and get rid of that function. Otherwise I'd suggest having more
distinct names.
Otherwise the patch looks good.
Thanks,
{
phys_addr_t paddr;
@@ -1944,30 +1945,47 @@ static void its_free_pending_table(struct page *pt)
free_pages((unsigned long)page_address(pt), get_order(LPI_PENDBASE_SZ));
}
-static void its_cpu_init_lpis(void)
+static int __init allocate_lpi_tables(void)
{
- void __iomem *rbase = gic_data_rdist_rd_base();
- struct page *pend_page;
- u64 val, tmp;
+ int err, cpu;
- /* If we didn't allocate the pending table yet, do it now */
- pend_page = gic_data_rdist()->pend_page;
- if (!pend_page) {
- phys_addr_t paddr;
+ err = its_alloc_lpi_prop_table();
+ if (err)
+ return err;
+
+ /*
+* We allocate all the pending tables anyway, as we may have a
+* mix of RDs that have had LPIs enabled, and some that
+* don't. We'll free the unused ones as each CPU comes online.
+*/
+ for_each_possible_cpu(cpu) {
+ struct page *pend_page;
pend_page = its_allocate_pending_table(GFP_NOWAIT);
if (!pend_page) {
- pr_err("Failed to allocate PENDBASE for CPU%d\n",
- smp_processor_id());
- return;
+ pr_err("Failed to allocate PENDBASE for CPU%d\n", cpu);
+ return -ENOMEM;
}
- paddr = page_to_phys(pend_page);
- pr_info("CPU%d: using LPI pending table @%pa\n",
- smp_processor_id(), );
- gic_data_rdist()->pend_page = pend_page;
+ gic_data_rdist_cpu(cpu)->pend_page = pend_page;
}
+ return 0;
+}
+
+static void its_cpu_init_lpis(void)
+{
+ void __iomem *rbase = gic_data_rdist_rd_base();
+ struct page *pend_page;
+ phys_addr_t paddr;
+ u64 val, tmp;
+
+ if (gic_data_rdist()->lpi_enabled)
+ return;
+
+ pend_page = gic_data_rdist()->pend_page;
+ paddr = page_to_phys(pend_page);
+
/* set PROPBASE */
val = (page_to_phys(gic_rdists->prop_page) |
GICR_PROPBASER_InnerShareable |
@@ -2019,6 +2037,10 @@ static void its_cpu_init_lpis(void)
/* Make sure the GIC has seen the above */
dsb(sy);
+ gic_data_rdist()->lpi_enabled = true;
+ pr_info("GICv3: CPU%d: using LPI pending table @%pa\n",
+ smp_processor_id(),
+ );
}
static void its_cpu_init_collection(struct its_node *its)
@@ -3497,16 +3519,6 @@ static int redist_disable_lpis(void)
u64 timeout = USEC_PER_SEC;
u64 val;
- /*
-* If coming via a CPU hotplug event, we don't need to disable
-* LPIs before trying to re-enable them. They are already
-* configured and all is well in the world. Detect this case
-* by checking the allocation of the pending table for the
-* current CPU.
-*/
- if (gic_data_rdist()->pend_page)
- return 0;
-
if (!gic_rdists_supports_plpis()) {
Re: [PATCH 04/10] irqchip/gic-v3-its: Move pending table allocation to init time
Hi Marc,
On 21/09/18 20:59, Marc Zyngier wrote:
Pending tables for the redistributors are currently allocated
one at a time as each CPU boots. This is causing some grief
for Linux/RT (allocation from within a CPU hotplug notifier is
frown upon).
Let's more this allocation to take place at init time, when we
only have a single CPU. It means we're allocating memory for CPUs
that are not online yet, but most system will boot all of their
CPUs anyway, so that's not completely wasted.
Signed-off-by: Marc Zyngier
---
drivers/irqchip/irq-gic-v3-its.c | 80 +++---
include/linux/irqchip/arm-gic-v3.h | 1 +
2 files changed, 53 insertions(+), 28 deletions(-)
diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
index 7ef6baea2d78..462bba422189 100644
--- a/drivers/irqchip/irq-gic-v3-its.c
+++ b/drivers/irqchip/irq-gic-v3-its.c
@@ -173,6 +173,7 @@ static DEFINE_RAW_SPINLOCK(vmovp_lock);
static DEFINE_IDA(its_vpeid_ida);
#define gic_data_rdist() (raw_cpu_ptr(gic_rdists->rdist))
+#define gic_data_rdist_cpu(cpu)(per_cpu_ptr(gic_rdists->rdist,
cpu))
#define gic_data_rdist_rd_base() (gic_data_rdist()->rd_base)
#define gic_data_rdist_vlpi_base()(gic_data_rdist_rd_base() + SZ_128K)
@@ -1625,7 +1626,7 @@ static void its_free_prop_table(struct page *prop_page)
get_order(LPI_PROPBASE_SZ));
}
-static int __init its_alloc_lpi_tables(void)
+static int __init its_alloc_lpi_prop_table(void)
A bit of a nit, but there is already a function called
"its_allocate_prop_table" which I find very easy to confuse with this one.
And patch 3 factored the initialization out of its_allocate_prop_table.
So I was wondering whether it would not actually be better to open-code
it here and get rid of that function. Otherwise I'd suggest having more
distinct names.
Otherwise the patch looks good.
Thanks,
{
phys_addr_t paddr;
@@ -1944,30 +1945,47 @@ static void its_free_pending_table(struct page *pt)
free_pages((unsigned long)page_address(pt), get_order(LPI_PENDBASE_SZ));
}
-static void its_cpu_init_lpis(void)
+static int __init allocate_lpi_tables(void)
{
- void __iomem *rbase = gic_data_rdist_rd_base();
- struct page *pend_page;
- u64 val, tmp;
+ int err, cpu;
- /* If we didn't allocate the pending table yet, do it now */
- pend_page = gic_data_rdist()->pend_page;
- if (!pend_page) {
- phys_addr_t paddr;
+ err = its_alloc_lpi_prop_table();
+ if (err)
+ return err;
+
+ /*
+* We allocate all the pending tables anyway, as we may have a
+* mix of RDs that have had LPIs enabled, and some that
+* don't. We'll free the unused ones as each CPU comes online.
+*/
+ for_each_possible_cpu(cpu) {
+ struct page *pend_page;
pend_page = its_allocate_pending_table(GFP_NOWAIT);
if (!pend_page) {
- pr_err("Failed to allocate PENDBASE for CPU%d\n",
- smp_processor_id());
- return;
+ pr_err("Failed to allocate PENDBASE for CPU%d\n", cpu);
+ return -ENOMEM;
}
- paddr = page_to_phys(pend_page);
- pr_info("CPU%d: using LPI pending table @%pa\n",
- smp_processor_id(), );
- gic_data_rdist()->pend_page = pend_page;
+ gic_data_rdist_cpu(cpu)->pend_page = pend_page;
}
+ return 0;
+}
+
+static void its_cpu_init_lpis(void)
+{
+ void __iomem *rbase = gic_data_rdist_rd_base();
+ struct page *pend_page;
+ phys_addr_t paddr;
+ u64 val, tmp;
+
+ if (gic_data_rdist()->lpi_enabled)
+ return;
+
+ pend_page = gic_data_rdist()->pend_page;
+ paddr = page_to_phys(pend_page);
+
/* set PROPBASE */
val = (page_to_phys(gic_rdists->prop_page) |
GICR_PROPBASER_InnerShareable |
@@ -2019,6 +2037,10 @@ static void its_cpu_init_lpis(void)
/* Make sure the GIC has seen the above */
dsb(sy);
+ gic_data_rdist()->lpi_enabled = true;
+ pr_info("GICv3: CPU%d: using LPI pending table @%pa\n",
+ smp_processor_id(),
+ );
}
static void its_cpu_init_collection(struct its_node *its)
@@ -3497,16 +3519,6 @@ static int redist_disable_lpis(void)
u64 timeout = USEC_PER_SEC;
u64 val;
- /*
-* If coming via a CPU hotplug event, we don't need to disable
-* LPIs before trying to re-enable them. They are already
-* configured and all is well in the world. Detect this case
-* by checking the allocation of the pending table for the
-* current CPU.
-*/
- if (gic_data_rdist()->pend_page)
- return 0;
-
if (!gic_rdists_supports_plpis()) {
