Hi Laurentiu,

Thanks a lot for your comments!

-----Original Message----- From: Laurentiu Tudor laurentiu.tudor@nxp.com Sent: 2021年3月5日 13:49 To: Z.q. Hou zhiqiang.hou@nxp.com; u-boot@lists.denx.de Cc: sjg@chromium.org; rayagonda.kokatanur@broadcom.com; Priyanka Jain priyanka.jain@nxp.com; Vladimir Oltean vladimir.oltean@nxp.com Subject: Re: [PATCHv2] arm64: gic-v3-its: Clear the Pending table before enabling LPIs

Hi Zhiqiang,

On 3/5/2021 4:21 AM, Zhiqiang Hou wrote:

...

From: Hou Zhiqiang Zhiqiang.Hou@nxp.com

The GICv3 RM requires "The first 1KB of memory for the LPI Pending tables must contain only zeros on initial allocation, and this must be visible to the Redistributors, or else the effect is UNPREDICTABLE".

And as the following statement, we here clear the whole Pending tables instead of the first 1KB. "An LPI Pending table that contains only zeros, including in the first 1KB, indicates that there are no pending LPIs. The first 1KB of the LPI Pending table is IMPLEMENTATION DEFINED. However, if the first 1KB of the LPI Pending table and the rest of the table contain only zeros, this must indicate that there are no pending LPIs."

And there isn't any pending LPI under U-Boot, so it's unnecessary to load the contents of the Pending table during the enablement, then set the GICR_PENDBASER.PTZ flag.

Signed-off-by: Hou Zhiqiang Zhiqiang.Hou@nxp.com

V2:

• Clear the Pending tables using virtual address.
• Correct some typos in the change log.

arch/arm/lib/gic-v3-its.c | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/arch/arm/lib/gic-v3-its.c b/arch/arm/lib/gic-v3-its.c index f5a921b3d1..2dadc48a50 100644 --- a/arch/arm/lib/gic-v3-its.c +++ b/arch/arm/lib/gic-v3-its.c @@ -3,6 +3,7 @@

• Copyright 2019 Broadcom.

*/ #include <common.h> +#include <cpu_func.h> #include <dm.h> #include <regmap.h> #include <syscon.h> @@ -108,6 +109,8 @@ int gic_lpi_tables_init(void) int i; u64 redist_lpi_base; u64 pend_base;

• ulong pend_tab_total_sz;

• void *pend_tab_va;

  if (gic_v3_its_get_gic_addr(&priv)) return -EINVAL;

@@ -160,7 +163,12 @@ int gic_lpi_tables_init(void) } }

• redist_lpi_base = priv.lpi_base + LPI_PROPBASE_SZ;

I don't think you want to drop this.

Yes, it's a mistake, will fix in v3.

...

• pend_tab_total_sz = priv.num_redist * LPI_PENDBASE_SZ;
• pend_tab_va = map_physmem(redist_lpi_base, pend_tab_total_sz,

• 		  MAP_NOCACHE);
  

• memset(pend_tab_va, 0, pend_tab_total_sz);
• flush_cache((ulong)pend_tab_va, pend_tab_total_sz);

Given that the memory is mapped as non-cacheable, is this flush necessary?

In theory, the flush is unnecessary with the MAP_NOCACHE flag, but on arm platforms as the map_physmem() just does a phys_to_virt address conversion and ignores the given flag, the flag doesn't take effect currently. To ensure the Pending tables in main memory has been cleared before the LPI enablement of each RD, I keep the flush here.

Thanks, Zhiqiang

---

Best Regards, Laurentiu

...

• unmap_physmem(pend_tab_va, MAP_NOCACHE);

  pend_base = priv.gicr_base + GICR_PENDBASER; for (i = 0; i < priv.num_redist; i++) { @@ -168,7 +176,8 @@ int

gic_lpi_tables_init(void)

val = ((redist_lpi_base + (i * LPI_PENDBASE_SZ)) |
	GICR_PENDBASER_INNERSHAREABLE |

• 	GICR_PENDBASER_RAWAWB);
  

• 	GICR_PENDBASER_RAWAWB |
  

• 	GICR_PENDBASER_PTZ);
  
  

  writeq(val, (uintptr_t)(pend_base + offset)); tmp = readq((uintptr_t)(pend_base + offset));