Skip to contentSkip to conversation

x86/UV: GRU distributed mode mappings

PATCH In review Dimitri Sivanich 3 participants 6 messages May 23, 2013 – May 31, 2013

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/UV: GRU distributed mode mappings

Setup mappings for GRU distributed mode and include them as untracked pat ranges.

---
 arch/x86/kernel/apic/x2apic_uv_x.c |   63 +++++++++++++++++++++++++++++----
 1 file changed, 57 insertions(+), 6 deletions(-)
arch/x86/kernel/apic/x2apic_uv_x.c +57 -6
Index: linux/arch/x86/kernel/apic/x2apic_uv_x.c
===================================================================
@@ -51,6 +51,8 @@ DEFINE_PER_CPU(int, x2apic_extra_bits);
 
 static enum uv_system_type uv_system_type;
 static u64 gru_start_paddr, gru_end_paddr;
+static u64 gru_dist_base, gru_first_node_paddr = -1LL, gru_last_node_paddr;
+static u64 gru_dist_lmask, gru_dist_umask;
 static union uvh_apicid uvh_apicid;
 int uv_min_hub_revision_id;
 EXPORT_SYMBOL_GPL(uv_min_hub_revision_id);
@@ -72,7 +74,19 @@ static unsigned long __init uv_early_rea
 
 static inline bool is_GRU_range(u64 start, u64 end)
 {
-	return start >= gru_start_paddr && end <= gru_end_paddr;
+	if (gru_dist_base) {
+		u64 su = start & gru_dist_umask; /* upper (incl pnode) bits */
+		u64 sl = start & gru_dist_lmask; /* base offset bits */
+		u64 eu = end & gru_dist_umask;
+		u64 el = end & gru_dist_lmask;
+
+		/* Must reside completely within a single GRU range */
+		return (sl == gru_dist_base && el == gru_dist_base &&
+			su >= gru_first_node_paddr &&
+			su <= gru_last_node_paddr &&
+			eu == su);
+	} else
+		return start >= gru_start_paddr && end <= gru_end_paddr;
 }
 
 static bool uv_is_untracked_pat_range(u64 start, u64 end)
@@ -470,19 +484,56 @@ static __init void map_high(char *id, un
 		init_extra_mapping_wb(paddr, bytes);
 }
 
+static __init void map_gru_distributed(unsigned long c)
+{
+	union uvh_rh_gam_gru_overlay_config_mmr_u gru;
+	u64 paddr;
+	unsigned long bytes;
+	int nid;
+
+	gru.v = c;
+	/* only base bits 42:28 relevant in dist mode */
+	gru_dist_base = gru.v & 0x000007fff0000000UL;
+	if (!gru_dist_base) {
+		pr_info("UV: Map GRU_DIST base address NULL\n");
+		return;
+	}
+	bytes = 1UL << UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR_BASE_SHFT;
+	gru_dist_lmask = ((1UL << uv_hub_info->m_val) - 1) & ~(bytes - 1);
+	gru_dist_umask = ~((1UL << uv_hub_info->m_val) - 1);
+	gru_dist_base &= gru_dist_lmask; /* Clear bits above M */
+	for_each_online_node(nid) {
+		paddr = ((u64)uv_node_to_pnode(nid) << uv_hub_info->m_val) |
+				gru_dist_base;
+		init_extra_mapping_wb(paddr, bytes);
+		gru_first_node_paddr = min(paddr, gru_first_node_paddr);
+		gru_last_node_paddr = max(paddr, gru_last_node_paddr);
+	}
+	/* Save upper (63:M) bits of address only for is_GRU_range */
+	gru_first_node_paddr &= gru_dist_umask;
+	gru_last_node_paddr &= gru_dist_umask;
+	pr_info("UV: Map GRU_DIST base 0x%016llx  0x%016llx - 0x%016llx\n",
+		gru_dist_base, gru_first_node_paddr, gru_last_node_paddr);
+}
+
 static __init void map_gru_high(int max_pnode)
 {
 	union uvh_rh_gam_gru_overlay_config_mmr_u gru;
 	int shift = UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR_BASE_SHFT;
 
 	gru.v = uv_read_local_mmr(UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR);
-	if (gru.s.enable) {
-		map_high("GRU", gru.s.base, shift, shift, max_pnode, map_wb);
-		gru_start_paddr = ((u64)gru.s.base << shift);
-		gru_end_paddr = gru_start_paddr + (1UL << shift) * (max_pnode + 1);
-	} else {
+	if (!gru.s.enable) {
 		pr_info("UV: GRU disabled\n");
+		return;
+	}
+
+	if (is_uv3_hub() && gru.s3.mode) {
+		map_gru_distributed(gru.v);
+		return;
 	}
+	map_high("GRU", gru.s.base, shift, shift, max_pnode, map_wb);
+	gru_start_paddr = ((u64)gru.s.base << shift);
+	gru_end_paddr = gru_start_paddr + (1UL << shift) * (max_pnode + 1);
 }
 
 static __init void map_mmr_high(int max_pnode)
#1 lore raw
Quoted 2 lines · Dimitri Sivanich

Setup mappings for GRU distributed mode and include them as untracked pat ranges.

Would be nice to have the customary changelog content:

" Current code does (A), this has a problem when (B).
We can improve this doing (C), because (D)."

Or so - your changelog is pretty light on the 'why' aspect.

Thanks,

Ingo

#2 lore raw
Quoted 6 lines · Ingo Molnar
Quoted 2 lines · Dimitri Sivanich

Setup mappings for GRU distributed mode and include them as untracked pat ranges.

Quoted 7 lines

Would be nice to have the customary changelog content:

" Current code does (A), this has a problem when (B).
We can improve this doing (C), because (D)."

Or so - your changelog is pretty light on the 'why' aspect.

How about the following?

UV GRU hardware will support an optional distributed mode that will allow per-node address mapping of local GRU space, as opposed to mapping all GRU hardware to the same contiguous high space.

If GRU distributed mode is selected, setup per-node page table mappings and include them as untracked pat ranges.

---
 arch/x86/kernel/apic/x2apic_uv_x.c |   63 +++++++++++++++++++++++++++++----
 1 file changed, 57 insertions(+), 6 deletions(-)
arch/x86/kernel/apic/x2apic_uv_x.c +57 -6
Index: linux/arch/x86/kernel/apic/x2apic_uv_x.c
===================================================================
@@ -51,6 +51,8 @@ DEFINE_PER_CPU(int, x2apic_extra_bits);
 
 static enum uv_system_type uv_system_type;
 static u64 gru_start_paddr, gru_end_paddr;
+static u64 gru_dist_base, gru_first_node_paddr = -1LL, gru_last_node_paddr;
+static u64 gru_dist_lmask, gru_dist_umask;
 static union uvh_apicid uvh_apicid;
 int uv_min_hub_revision_id;
 EXPORT_SYMBOL_GPL(uv_min_hub_revision_id);
@@ -72,7 +74,19 @@ static unsigned long __init uv_early_rea
 
 static inline bool is_GRU_range(u64 start, u64 end)
 {
-	return start >= gru_start_paddr && end <= gru_end_paddr;
+	if (gru_dist_base) {
+		u64 su = start & gru_dist_umask; /* upper (incl pnode) bits */
+		u64 sl = start & gru_dist_lmask; /* base offset bits */
+		u64 eu = end & gru_dist_umask;
+		u64 el = end & gru_dist_lmask;
+
+		/* Must reside completely within a single GRU range */
+		return (sl == gru_dist_base && el == gru_dist_base &&
+			su >= gru_first_node_paddr &&
+			su <= gru_last_node_paddr &&
+			eu == su);
+	} else
+		return start >= gru_start_paddr && end <= gru_end_paddr;
 }
 
 static bool uv_is_untracked_pat_range(u64 start, u64 end)
@@ -470,19 +484,56 @@ static __init void map_high(char *id, un
 		init_extra_mapping_wb(paddr, bytes);
 }
 
+static __init void map_gru_distributed(unsigned long c)
+{
+	union uvh_rh_gam_gru_overlay_config_mmr_u gru;
+	u64 paddr;
+	unsigned long bytes;
+	int nid;
+
+	gru.v = c;
+	/* only base bits 42:28 relevant in dist mode */
+	gru_dist_base = gru.v & 0x000007fff0000000UL;
+	if (!gru_dist_base) {
+		pr_info("UV: Map GRU_DIST base address NULL\n");
+		return;
+	}
+	bytes = 1UL << UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR_BASE_SHFT;
+	gru_dist_lmask = ((1UL << uv_hub_info->m_val) - 1) & ~(bytes - 1);
+	gru_dist_umask = ~((1UL << uv_hub_info->m_val) - 1);
+	gru_dist_base &= gru_dist_lmask; /* Clear bits above M */
+	for_each_online_node(nid) {
+		paddr = ((u64)uv_node_to_pnode(nid) << uv_hub_info->m_val) |
+				gru_dist_base;
+		init_extra_mapping_wb(paddr, bytes);
+		gru_first_node_paddr = min(paddr, gru_first_node_paddr);
+		gru_last_node_paddr = max(paddr, gru_last_node_paddr);
+	}
+	/* Save upper (63:M) bits of address only for is_GRU_range */
+	gru_first_node_paddr &= gru_dist_umask;
+	gru_last_node_paddr &= gru_dist_umask;
+	pr_info("UV: Map GRU_DIST base 0x%016llx  0x%016llx - 0x%016llx\n",
+		gru_dist_base, gru_first_node_paddr, gru_last_node_paddr);
+}
+
 static __init void map_gru_high(int max_pnode)
 {
 	union uvh_rh_gam_gru_overlay_config_mmr_u gru;
 	int shift = UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR_BASE_SHFT;
 
 	gru.v = uv_read_local_mmr(UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR);
-	if (gru.s.enable) {
-		map_high("GRU", gru.s.base, shift, shift, max_pnode, map_wb);
-		gru_start_paddr = ((u64)gru.s.base << shift);
-		gru_end_paddr = gru_start_paddr + (1UL << shift) * (max_pnode + 1);
-	} else {
+	if (!gru.s.enable) {
 		pr_info("UV: GRU disabled\n");
+		return;
+	}
+
+	if (is_uv3_hub() && gru.s3.mode) {
+		map_gru_distributed(gru.v);
+		return;
 	}
+	map_high("GRU", gru.s.base, shift, shift, max_pnode, map_wb);
+	gru_start_paddr = ((u64)gru.s.base << shift);
+	gru_end_paddr = gru_start_paddr + (1UL << shift) * (max_pnode + 1);
 }
 
 static __init void map_mmr_high(int max_pnode)
#3 lore raw
Quoted 7 lines · Dimitri Sivanich
Quoted 6 lines · Ingo Molnar
Quoted 2 lines · Dimitri Sivanich

Setup mappings for GRU distributed mode and include them as untracked pat ranges.

Quoted 8 lines
Quoted 7 lines

Would be nice to have the customary changelog content:

" Current code does (A), this has a problem when (B).
We can improve this doing (C), because (D)."

Or so - your changelog is pretty light on the 'why' aspect.

How about the following?

Looks good to me.

Quoted 11 lines

UV GRU hardware will support an optional distributed mode that will allow per-node address mapping of local GRU space, as opposed to mapping all GRU hardware to the same contiguous high space.

If GRU distributed mode is selected, setup per-node page table mappings and include them as untracked pat ranges.

---
 arch/x86/kernel/apic/x2apic_uv_x.c |   63 +++++++++++++++++++++++++++++----
 1 file changed, 57 insertions(+), 6 deletions(-)

So this looks like suitable v3.11 material - not an urgent fix needed for v3.10, right?

Noticed this detail:

arch/x86/kernel/apic/x2apic_uv_x.c static unsigned long __init uv_early_rea · line 89

Quoted 2 lines
arch/x86/kernel/apic/x2apic_uv_x.c +2 -0
+	} else
+		return start >= gru_start_paddr && end <= gru_end_paddr;

in the kernel we generally try to balance curly braces - either all branches are curly or none.

arch/x86/kernel/apic/x2apic_uv_x.c static __init void map_high(char *id, un · line 516

Quoted 5 lines
arch/x86/kernel/apic/x2apic_uv_x.c +5 -0
+	/* Save upper (63:M) bits of address only for is_GRU_range */
+	gru_first_node_paddr &= gru_dist_umask;
+	gru_last_node_paddr &= gru_dist_umask;
+	pr_info("UV: Map GRU_DIST base 0x%016llx  0x%016llx - 0x%016llx\n",
+		gru_dist_base, gru_first_node_paddr, gru_last_node_paddr);

I suspect this should be pr_debug()?

Thanks,

Ingo

#4 lore raw
Quoted 3 lines · Ingo Molnar

So this looks like suitable v3.11 material - not an urgent fix needed for v3.10, right?

Correct, v3.11 is fine.

Quoted 8 lines

Noticed this detail:

Review

Quoted 2 lines
(quoted diff) +2 -0
+	} else
+		return start >= gru_start_paddr && end <= gru_end_paddr;

in the kernel we generally try to balance curly braces - either all branches are curly or none.

Fixed, see patch below.

Quoted 9 lines

Review

Quoted 6 lines
(quoted diff) +5 -0
+	/* Save upper (63:M) bits of address only for is_GRU_range */
+	gru_first_node_paddr &= gru_dist_umask;
+	gru_last_node_paddr &= gru_dist_umask;
+	pr_info("UV: Map GRU_DIST base 0x%016llx  0x%016llx - 0x%016llx\n",
+		gru_dist_base, gru_first_node_paddr, gru_last_node_paddr);

I suspect this should be pr_debug()?

I had set it to what was being used in map_high(), but pr_debug should be OK, so I changed both instances.

GRU hardware will support an optional distributed mode that will allow per-node address mapping of local GRU space, as opposed to mapping all GRU hardware to the same contiguous high space.

If GRU distributed mode is selected, setup per-node page table mappings.

---
 arch/x86/kernel/apic/x2apic_uv_x.c |   66 +++++++++++++++++++++++++++++----
 1 file changed, 59 insertions(+), 7 deletions(-)
arch/x86/kernel/apic/x2apic_uv_x.c +59 -7
Index: linux/arch/x86/kernel/apic/x2apic_uv_x.c
===================================================================
@@ -51,6 +51,8 @@ DEFINE_PER_CPU(int, x2apic_extra_bits);
 
 static enum uv_system_type uv_system_type;
 static u64 gru_start_paddr, gru_end_paddr;
+static u64 gru_dist_base, gru_first_node_paddr = -1LL, gru_last_node_paddr;
+static u64 gru_dist_lmask, gru_dist_umask;
 static union uvh_apicid uvh_apicid;
 int uv_min_hub_revision_id;
 EXPORT_SYMBOL_GPL(uv_min_hub_revision_id);
@@ -72,7 +74,20 @@ static unsigned long __init uv_early_rea
 
 static inline bool is_GRU_range(u64 start, u64 end)
 {
-	return start >= gru_start_paddr && end <= gru_end_paddr;
+	if (gru_dist_base) {
+		u64 su = start & gru_dist_umask; /* upper (incl pnode) bits */
+		u64 sl = start & gru_dist_lmask; /* base offset bits */
+		u64 eu = end & gru_dist_umask;
+		u64 el = end & gru_dist_lmask;
+
+		/* Must reside completely within a single GRU range */
+		return (sl == gru_dist_base && el == gru_dist_base &&
+			su >= gru_first_node_paddr &&
+			su <= gru_last_node_paddr &&
+			eu == su);
+	} else {
+		return start >= gru_start_paddr && end <= gru_end_paddr;
+	}
 }
 
 static bool uv_is_untracked_pat_range(u64 start, u64 end)
@@ -463,26 +478,63 @@ static __init void map_high(char *id, un
 		pr_info("UV: Map %s_HI base address NULL\n", id);
 		return;
 	}
-	pr_info("UV: Map %s_HI 0x%lx - 0x%lx\n", id, paddr, paddr + bytes);
+	pr_debug("UV: Map %s_HI 0x%lx - 0x%lx\n", id, paddr, paddr + bytes);
 	if (map_type == map_uc)
 		init_extra_mapping_uc(paddr, bytes);
 	else
 		init_extra_mapping_wb(paddr, bytes);
 }
 
+static __init void map_gru_distributed(unsigned long c)
+{
+	union uvh_rh_gam_gru_overlay_config_mmr_u gru;
+	u64 paddr;
+	unsigned long bytes;
+	int nid;
+
+	gru.v = c;
+	/* only base bits 42:28 relevant in dist mode */
+	gru_dist_base = gru.v & 0x000007fff0000000UL;
+	if (!gru_dist_base) {
+		pr_info("UV: Map GRU_DIST base address NULL\n");
+		return;
+	}
+	bytes = 1UL << UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR_BASE_SHFT;
+	gru_dist_lmask = ((1UL << uv_hub_info->m_val) - 1) & ~(bytes - 1);
+	gru_dist_umask = ~((1UL << uv_hub_info->m_val) - 1);
+	gru_dist_base &= gru_dist_lmask; /* Clear bits above M */
+	for_each_online_node(nid) {
+		paddr = ((u64)uv_node_to_pnode(nid) << uv_hub_info->m_val) |
+				gru_dist_base;
+		init_extra_mapping_wb(paddr, bytes);
+		gru_first_node_paddr = min(paddr, gru_first_node_paddr);
+		gru_last_node_paddr = max(paddr, gru_last_node_paddr);
+	}
+	/* Save upper (63:M) bits of address only for is_GRU_range */
+	gru_first_node_paddr &= gru_dist_umask;
+	gru_last_node_paddr &= gru_dist_umask;
+	pr_debug("UV: Map GRU_DIST base 0x%016llx  0x%016llx - 0x%016llx\n",
+		gru_dist_base, gru_first_node_paddr, gru_last_node_paddr);
+}
+
 static __init void map_gru_high(int max_pnode)
 {
 	union uvh_rh_gam_gru_overlay_config_mmr_u gru;
 	int shift = UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR_BASE_SHFT;
 
 	gru.v = uv_read_local_mmr(UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR);
-	if (gru.s.enable) {
-		map_high("GRU", gru.s.base, shift, shift, max_pnode, map_wb);
-		gru_start_paddr = ((u64)gru.s.base << shift);
-		gru_end_paddr = gru_start_paddr + (1UL << shift) * (max_pnode + 1);
-	} else {
+	if (!gru.s.enable) {
 		pr_info("UV: GRU disabled\n");
+		return;
+	}
+
+	if (is_uv3_hub() && gru.s3.mode) {
+		map_gru_distributed(gru.v);
+		return;
 	}
+	map_high("GRU", gru.s.base, shift, shift, max_pnode, map_wb);
+	gru_start_paddr = ((u64)gru.s.base << shift);
+	gru_end_paddr = gru_start_paddr + (1UL << shift) * (max_pnode + 1);
 }
 
 static __init void map_mmr_high(int max_pnode)
#5 lore raw

[tip:x86/uv] x86/UV: Add GRU distributed mode mappings

x86/UV: Add GRU distributed mode mappings
Commit-ID:  879d5ad0dc3b096170c8f985b1465de5dca88de8
Gitweb:     http://git.kernel.org/tip/879d5ad0dc3b096170c8f985b1465de5dca88de8
Author:     Dimitri Sivanich <[email protected]>
AuthorDate: Wed, 29 May 2013 10:56:09 -0500
Committer:  Ingo Molnar <[email protected]>
CommitDate: Thu, 30 May 2013 09:46:09 +0200

x86/UV: Add GRU distributed mode mappings

GRU hardware will support an optional distributed mode that will allow per-node address mapping of local GRU space, as opposed to mapping all GRU hardware to the same contiguous high space.

If GRU distributed mode is selected, setup per-node page table mappings.

---
 arch/x86/kernel/apic/x2apic_uv_x.c | 66 ++++++++++++++++++++++++++++++++++----
 1 file changed, 59 insertions(+), 7 deletions(-)
arch/x86/kernel/apic/x2apic_uv_x.c +59 -7
index 794f6eb..39cc7f7 100644
@@ -51,6 +51,8 @@ DEFINE_PER_CPU(int, x2apic_extra_bits);
 
 static enum uv_system_type uv_system_type;
 static u64 gru_start_paddr, gru_end_paddr;
+static u64 gru_dist_base, gru_first_node_paddr = -1LL, gru_last_node_paddr;
+static u64 gru_dist_lmask, gru_dist_umask;
 static union uvh_apicid uvh_apicid;
 int uv_min_hub_revision_id;
 EXPORT_SYMBOL_GPL(uv_min_hub_revision_id);
@@ -72,7 +74,20 @@ static unsigned long __init uv_early_read_mmr(unsigned long addr)
 
 static inline bool is_GRU_range(u64 start, u64 end)
 {
-	return start >= gru_start_paddr && end <= gru_end_paddr;
+	if (gru_dist_base) {
+		u64 su = start & gru_dist_umask; /* upper (incl pnode) bits */
+		u64 sl = start & gru_dist_lmask; /* base offset bits */
+		u64 eu = end & gru_dist_umask;
+		u64 el = end & gru_dist_lmask;
+
+		/* Must reside completely within a single GRU range */
+		return (sl == gru_dist_base && el == gru_dist_base &&
+			su >= gru_first_node_paddr &&
+			su <= gru_last_node_paddr &&
+			eu == su);
+	} else {
+		return start >= gru_start_paddr && end <= gru_end_paddr;
+	}
 }
 
 static bool uv_is_untracked_pat_range(u64 start, u64 end)
@@ -463,26 +478,63 @@ static __init void map_high(char *id, unsigned long base, int pshift,
 		pr_info("UV: Map %s_HI base address NULL\n", id);
 		return;
 	}
-	pr_info("UV: Map %s_HI 0x%lx - 0x%lx\n", id, paddr, paddr + bytes);
+	pr_debug("UV: Map %s_HI 0x%lx - 0x%lx\n", id, paddr, paddr + bytes);
 	if (map_type == map_uc)
 		init_extra_mapping_uc(paddr, bytes);
 	else
 		init_extra_mapping_wb(paddr, bytes);
 }
 
+static __init void map_gru_distributed(unsigned long c)
+{
+	union uvh_rh_gam_gru_overlay_config_mmr_u gru;
+	u64 paddr;
+	unsigned long bytes;
+	int nid;
+
+	gru.v = c;
+	/* only base bits 42:28 relevant in dist mode */
+	gru_dist_base = gru.v & 0x000007fff0000000UL;
+	if (!gru_dist_base) {
+		pr_info("UV: Map GRU_DIST base address NULL\n");
+		return;
+	}
+	bytes = 1UL << UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR_BASE_SHFT;
+	gru_dist_lmask = ((1UL << uv_hub_info->m_val) - 1) & ~(bytes - 1);
+	gru_dist_umask = ~((1UL << uv_hub_info->m_val) - 1);
+	gru_dist_base &= gru_dist_lmask; /* Clear bits above M */
+	for_each_online_node(nid) {
+		paddr = ((u64)uv_node_to_pnode(nid) << uv_hub_info->m_val) |
+				gru_dist_base;
+		init_extra_mapping_wb(paddr, bytes);
+		gru_first_node_paddr = min(paddr, gru_first_node_paddr);
+		gru_last_node_paddr = max(paddr, gru_last_node_paddr);
+	}
+	/* Save upper (63:M) bits of address only for is_GRU_range */
+	gru_first_node_paddr &= gru_dist_umask;
+	gru_last_node_paddr &= gru_dist_umask;
+	pr_debug("UV: Map GRU_DIST base 0x%016llx  0x%016llx - 0x%016llx\n",
+		gru_dist_base, gru_first_node_paddr, gru_last_node_paddr);
+}
+
 static __init void map_gru_high(int max_pnode)
 {
 	union uvh_rh_gam_gru_overlay_config_mmr_u gru;
 	int shift = UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR_BASE_SHFT;
 
 	gru.v = uv_read_local_mmr(UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR);
-	if (gru.s.enable) {
-		map_high("GRU", gru.s.base, shift, shift, max_pnode, map_wb);
-		gru_start_paddr = ((u64)gru.s.base << shift);
-		gru_end_paddr = gru_start_paddr + (1UL << shift) * (max_pnode + 1);
-	} else {
+	if (!gru.s.enable) {
 		pr_info("UV: GRU disabled\n");
+		return;
+	}
+
+	if (is_uv3_hub() && gru.s3.mode) {
+		map_gru_distributed(gru.v);
+		return;
 	}
+	map_high("GRU", gru.s.base, shift, shift, max_pnode, map_wb);
+	gru_start_paddr = ((u64)gru.s.base << shift);
+	gru_end_paddr = gru_start_paddr + (1UL << shift) * (max_pnode + 1);
 }
 
 static __init void map_mmr_high(int max_pnode)