Re: [PATCH 2/4] RAS: ATL: Expand helpers for adding and removing base and hole

[John Allen]

On Mon, Mar 18, 2024 at 11:28:48AM -0400, Yazen Ghannam wrote:
> On 3/14/24 12:35, John Allen wrote:
> > Data fabric 4.5 denormalization will need to frequently add and remove
> 
> More specifically, the non-power-of-2 cases will need this.
> 
> > the base and the legacy MMIO hole. Modify existing helpers to improve DF
> > 4.5 denormalization flow and add helper to remove the base and hole.
> 
> Please write the what/context, why/issue, and how/fix information as
> separate paragraphs even if they're just a single sentence each. I think
> this helps to find the details more easily.
> 
> > 
> > Signed-off-by: John Allen <john.allen@xxxxxxx>
> > ---
> >   drivers/ras/amd/atl/core.c     | 43 ++++++++++++++++++++++------------
> >   drivers/ras/amd/atl/internal.h |  3 +++
> >   2 files changed, 31 insertions(+), 15 deletions(-)
> > 
> > diff --git a/drivers/ras/amd/atl/core.c b/drivers/ras/amd/atl/core.c
> > index c1710d233adb..cafdfc57d929 100644
> > --- a/drivers/ras/amd/atl/core.c
> > +++ b/drivers/ras/amd/atl/core.c
> > @@ -49,15 +49,26 @@ static bool legacy_hole_en(struct addr_ctx *ctx)
> >   	return FIELD_GET(DF_LEGACY_MMIO_HOLE_EN, reg);
> >   }
> > -static int add_legacy_hole(struct addr_ctx *ctx)
> > +static u64 add_legacy_hole(struct addr_ctx *ctx, u64 addr)
> >   {
> >   	if (!legacy_hole_en(ctx))
> > -		return 0;
> > +		return addr;
> > -	if (ctx->addr >= df_cfg.dram_hole_base)
> > -		ctx->addr += (BIT_ULL(32) - df_cfg.dram_hole_base);
> > +	if (addr >= df_cfg.dram_hole_base)
> > +		addr += (BIT_ULL(32) - df_cfg.dram_hole_base);
> > -	return 0;
> > +	return addr;
> > +}
> > +
> > +static u64 remove_legacy_hole(struct addr_ctx *ctx, u64 addr)
> > +{
> > +	if (!legacy_hole_en(ctx))
> > +		return addr;
> > +
> > +	if (addr >= df_cfg.dram_hole_base)
> > +		addr -= (BIT_ULL(32) - df_cfg.dram_hole_base);
> > +
> > +	return addr;
> >   }
> >   static u64 get_base_addr(struct addr_ctx *ctx)
> > @@ -72,14 +83,16 @@ static u64 get_base_addr(struct addr_ctx *ctx)
> >   	return base_addr << DF_DRAM_BASE_LIMIT_LSB;
> >   }
> > -static int add_base_and_hole(struct addr_ctx *ctx)
> > +u64 add_base_and_hole(struct addr_ctx *ctx, u64 addr)
> >   {
> > -	ctx->ret_addr += get_base_addr(ctx);
> > -
> > -	if (add_legacy_hole(ctx))
> > -		return -EINVAL;
> > +	addr += get_base_addr(ctx);
> > +	return add_legacy_hole(ctx, addr);
> > +}
> > -	return 0;
> > +u64 remove_base_and_hole(struct addr_ctx *ctx, u64 addr)
> > +{
> > +	addr -= get_base_addr(ctx);
> > +	return remove_legacy_hole(ctx, addr);
> 
> This should be the inverse of the "add" operation, I think. So remove
> the legacy hole first, then remove the base address.
> 
> >   }
> >   static bool late_hole_remove(struct addr_ctx *ctx)
> > @@ -123,14 +136,14 @@ unsigned long norm_to_sys_addr(u8 socket_id, u8 die_id, u8 coh_st_inst_id, unsig
> >   	if (denormalize_address(&ctx))
> >   		return -EINVAL;
> > -	if (!late_hole_remove(&ctx) && add_base_and_hole(&ctx))
> > -		return -EINVAL;
> > +	if (!late_hole_remove(&ctx))
> > +		ctx.ret_addr = add_base_and_hole(&ctx, ctx.ret_addr);
> >   	if (dehash_address(&ctx))
> >   		return -EINVAL;
> > -	if (late_hole_remove(&ctx) && add_base_and_hole(&ctx))
> > -		return -EINVAL;
> > +	if (late_hole_remove(&ctx))
> > +		ctx.ret_addr = add_base_and_hole(&ctx, ctx.ret_addr);
> >   	if (addr_over_limit(&ctx))
> >   		return -EINVAL;
> > diff --git a/drivers/ras/amd/atl/internal.h b/drivers/ras/amd/atl/internal.h
> > index 1413c8ddc6c5..05b870fcb24e 100644
> > --- a/drivers/ras/amd/atl/internal.h
> > +++ b/drivers/ras/amd/atl/internal.h
> > @@ -236,6 +236,9 @@ int dehash_address(struct addr_ctx *ctx);
> >   unsigned long norm_to_sys_addr(u8 socket_id, u8 die_id, u8 coh_st_inst_id, unsigned long addr);
> >   unsigned long convert_umc_mca_addr_to_sys_addr(struct atl_err *err);
> > +u64 add_base_and_hole(struct addr_ctx *ctx, u64 addr);
> > +u64 remove_base_and_hole(struct addr_ctx *ctx, u64 addr);
> 
> remove_base_and_hole() is only used in denormalize.c, correct? So why
> not define it there as static? Other than trying to keep the code
> together and symmetrical, I mean.

In addition to keeping the two inverse functions together,
remove_base_and_hole depends on other functions in core.c. So if we
don't expose remove_base_and_hole in the header, then we would need to
expose get_base_addr and remove_legacy_hole in the header.
Alternatively, we could move remove_legacy_hole to denormalize.c and
expose get_base_addr and legacy_hole_en in the header instead. So
exposing one function that's the inverse of the other just looks better
to me than exposing two.

Thanks,
John