[PATCH 4.1 42/65] kaiser: enhanced by kernel and user PCIDs

From: Pavel Tatashin
Date: Mon Mar 05 2018 - 19:33:49 EST

From: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>

Merged performance improvements to Kaiser, using distinct kernel
and user Process Context Identifiers to minimize the TLB flushing.

Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
Acked-by: Jiri Kosina <jkosina@xxxxxxx>
Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
(cherry picked from commit eb82151d0b1df53d1ad8d060ecd554ca12eb552a)
Signed-off-by: Pavel Tatashin <pasha.tatashin@xxxxxxxxxx>

Conflicts:
arch/x86/entry/entry_64.S (not in this tree)
arch/x86/kernel/entry_64.S (patched instead of that)
arch/x86/entry/entry_64_compat.S (not in this tree)
arch/x86/ia32/ia32entry.S (patched instead of that)
arch/x86/include/asm/tlbflush.h
---
arch/x86/ia32/ia32entry.S | 1 +
arch/x86/include/asm/cpufeature.h | 1 +
arch/x86/include/asm/kaiser.h | 15 ++++++++--
arch/x86/include/asm/pgtable_types.h | 26 ++++++++++++++++
arch/x86/include/asm/tlbflush.h | 37 +++++++++++++++++++++--
arch/x86/include/uapi/asm/processor-flags.h | 3 +-
arch/x86/kernel/cpu/common.c | 34 +++++++++++++++++++++
arch/x86/kernel/entry_64.S | 10 +++++--
arch/x86/kvm/x86.c | 3 +-
arch/x86/mm/kaiser.c | 7 +++++
arch/x86/mm/tlb.c | 46 +++++++++++++++++++++++++++--
11 files changed, 172 insertions(+), 11 deletions(-)

diff --git a/arch/x86/ia32/ia32entry.S b/arch/x86/ia32/ia32entry.S
index 665e2c7887fe..070470d161f6 100644
--- a/arch/x86/ia32/ia32entry.S
+++ b/arch/x86/ia32/ia32entry.S
@@ -15,6 +15,7 @@
#include <asm/irqflags.h>
#include <asm/asm.h>
#include <asm/smap.h>
+#include <asm/pgtable_types.h>
#include <asm/kaiser.h>
#include <linux/linkage.h>
#include <linux/err.h>
diff --git a/arch/x86/include/asm/cpufeature.h b/arch/x86/include/asm/cpufeature.h
index 3d6606fb97d0..80a2b9487b29 100644
--- a/arch/x86/include/asm/cpufeature.h
+++ b/arch/x86/include/asm/cpufeature.h
@@ -185,6 +185,7 @@
#define X86_FEATURE_ARAT ( 7*32+ 1) /* Always Running APIC Timer */
#define X86_FEATURE_CPB ( 7*32+ 2) /* AMD Core Performance Boost */
#define X86_FEATURE_EPB ( 7*32+ 3) /* IA32_ENERGY_PERF_BIAS support */
+#define X86_FEATURE_INVPCID_SINGLE ( 7*32+ 4) /* Effectively INVPCID && CR4.PCIDE=1 */
#define X86_FEATURE_PLN ( 7*32+ 5) /* Intel Power Limit Notification */
#define X86_FEATURE_PTS ( 7*32+ 6) /* Intel Package Thermal Status */
#define X86_FEATURE_DTHERM ( 7*32+ 7) /* Digital Thermal Sensor */
diff --git a/arch/x86/include/asm/kaiser.h b/arch/x86/include/asm/kaiser.h
index e0fc45e77aee..360ff3bc44a9 100644
--- a/arch/x86/include/asm/kaiser.h
+++ b/arch/x86/include/asm/kaiser.h
@@ -1,5 +1,8 @@
#ifndef _ASM_X86_KAISER_H
#define _ASM_X86_KAISER_H
+
+#include <uapi/asm/processor-flags.h> /* For PCID constants */
+
/*
* This file includes the definitions for the KAISER feature.
* KAISER is a counter measure against x86_64 side channel attacks on
@@ -21,13 +24,21 @@

.macro _SWITCH_TO_KERNEL_CR3 reg
movq %cr3, \reg
-andq $(~KAISER_SHADOW_PGD_OFFSET), \reg
+andq $(~(X86_CR3_PCID_ASID_MASK | KAISER_SHADOW_PGD_OFFSET)), \reg
+orq X86_CR3_PCID_KERN_VAR, \reg
movq \reg, %cr3
.endm

.macro _SWITCH_TO_USER_CR3 reg
movq %cr3, \reg
-orq $(KAISER_SHADOW_PGD_OFFSET), \reg
+andq $(~(X86_CR3_PCID_ASID_MASK | KAISER_SHADOW_PGD_OFFSET)), \reg
+/*
+ * This can obviously be one instruction by putting the
+ * KAISER_SHADOW_PGD_OFFSET bit in the X86_CR3_PCID_USER_VAR.
+ * But, just leave it now for simplicity.
+ */
+orq X86_CR3_PCID_USER_VAR, \reg
+orq $(KAISER_SHADOW_PGD_OFFSET), \reg
movq \reg, %cr3
.endm

diff --git a/arch/x86/include/asm/pgtable_types.h b/arch/x86/include/asm/pgtable_types.h
index 7a5fc4410c00..b3c77e6529c2 100644
--- a/arch/x86/include/asm/pgtable_types.h
+++ b/arch/x86/include/asm/pgtable_types.h
@@ -106,6 +106,32 @@
_PAGE_SOFT_DIRTY)
#define _HPAGE_CHG_MASK (_PAGE_CHG_MASK | _PAGE_PSE)

+/* The ASID is the lower 12 bits of CR3 */
+#define X86_CR3_PCID_ASID_MASK (_AC((1<<12)-1,UL))
+
+/* Mask for all the PCID-related bits in CR3: */
+#define X86_CR3_PCID_MASK (X86_CR3_PCID_NOFLUSH | X86_CR3_PCID_ASID_MASK)
+#if defined(CONFIG_KAISER) && defined(CONFIG_X86_64)
+#define X86_CR3_PCID_ASID_KERN (_AC(0x4,UL))
+#define X86_CR3_PCID_ASID_USER (_AC(0x6,UL))
+
+#define X86_CR3_PCID_KERN_FLUSH (X86_CR3_PCID_ASID_KERN)
+#define X86_CR3_PCID_USER_FLUSH (X86_CR3_PCID_ASID_USER)
+#define X86_CR3_PCID_KERN_NOFLUSH (X86_CR3_PCID_NOFLUSH | X86_CR3_PCID_ASID_KERN)
+#define X86_CR3_PCID_USER_NOFLUSH (X86_CR3_PCID_NOFLUSH | X86_CR3_PCID_ASID_USER)
+#else
+#define X86_CR3_PCID_ASID_KERN (_AC(0x0,UL))
+#define X86_CR3_PCID_ASID_USER (_AC(0x0,UL))
+/*
+ * PCIDs are unsupported on 32-bit and none of these bits can be
+ * set in CR3:
+ */
+#define X86_CR3_PCID_KERN_FLUSH (0)
+#define X86_CR3_PCID_USER_FLUSH (0)
+#define X86_CR3_PCID_KERN_NOFLUSH (0)
+#define X86_CR3_PCID_USER_NOFLUSH (0)
+#endif
+
/*
* The cache modes defined here are used to translate between pure SW usage
* and the HW defined cache mode bits and/or PAT entries.
diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
index 09a70d8d293e..b558d1996621 100644
--- a/arch/x86/include/asm/tlbflush.h
+++ b/arch/x86/include/asm/tlbflush.h
@@ -12,7 +12,6 @@ static inline void __invpcid(unsigned long pcid, unsigned long addr,
unsigned long type)
{
struct { u64 d[2]; } desc = { { pcid, addr } };
-
/*
* The memory clobber is because the whole point is to invalidate
* stale TLB entries and, especially if we're flushing global
@@ -134,6 +133,14 @@ static inline void cr4_set_bits_and_update_boot(unsigned long mask)
static inline void __native_flush_tlb(void)
{
native_write_cr3(native_read_cr3());
+ /*
+ * We are no longer using globals with KAISER, so a
+ * "nonglobals" flush would work too. But, this is more
+ * conservative.
+ *
+ * Note, this works with CR4.PCIDE=0 or 1.
+ */
+ invpcid_flush_all();
}

static inline void __native_flush_tlb_global_irq_disabled(void)
@@ -155,6 +162,8 @@ static inline void __native_flush_tlb_global(void)
/*
* Using INVPCID is considerably faster than a pair of writes
* to CR4 sandwiched inside an IRQ flag save/restore.
+ *
+ * Note, this works with CR4.PCIDE=0 or 1.
*/
invpcid_flush_all();
return;
@@ -174,7 +183,31 @@ static inline void __native_flush_tlb_global(void)

static inline void __native_flush_tlb_single(unsigned long addr)
{
- asm volatile("invlpg (%0)" ::"r" (addr) : "memory");
+ /*
+ * SIMICS #GP's if you run INVPCID with type 2/3
+ * and X86_CR4_PCIDE clear. Shame!
+ *
+ * The ASIDs used below are hard-coded. But, we must not
+ * call invpcid(type=1/2) before CR4.PCIDE=1. Just call
+ * invpcid in the case we are called early.
+ */
+ if (!this_cpu_has(X86_FEATURE_INVPCID_SINGLE)) {
+ asm volatile("invlpg (%0)" ::"r" (addr) : "memory");
+ return;
+ }
+ /* Flush the address out of both PCIDs. */
+ /*
+ * An optimization here might be to determine addresses
+ * that are only kernel-mapped and only flush the kernel
+ * ASID. But, userspace flushes are probably much more
+ * important performance-wise.
+ *
+ * Make sure to do only a single invpcid when KAISER is
+ * disabled and we have only a single ASID.
+ */
+ if (X86_CR3_PCID_ASID_KERN != X86_CR3_PCID_ASID_USER)
+ invpcid_flush_one(X86_CR3_PCID_ASID_KERN, addr);
+ invpcid_flush_one(X86_CR3_PCID_ASID_USER, addr);
}

static inline void __flush_tlb_all(void)
diff --git a/arch/x86/include/uapi/asm/processor-flags.h b/arch/x86/include/uapi/asm/processor-flags.h
index 180a0c3c224d..bd4513b7b877 100644
--- a/arch/x86/include/uapi/asm/processor-flags.h
+++ b/arch/x86/include/uapi/asm/processor-flags.h
@@ -79,7 +79,8 @@
#define X86_CR3_PWT _BITUL(X86_CR3_PWT_BIT)
#define X86_CR3_PCD_BIT 4 /* Page Cache Disable */
#define X86_CR3_PCD _BITUL(X86_CR3_PCD_BIT)
-#define X86_CR3_PCID_MASK _AC(0x00000fff,UL) /* PCID Mask */
+#define X86_CR3_PCID_NOFLUSH_BIT 63 /* Preserve old PCID */
+#define X86_CR3_PCID_NOFLUSH _BITULL(X86_CR3_PCID_NOFLUSH_BIT)

/*
* Intel CPU features in CR4
diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
index 47194fac7eb7..6fb0d332f440 100644
--- a/arch/x86/kernel/cpu/common.c
+++ b/arch/x86/kernel/cpu/common.c
@@ -339,11 +339,45 @@ static __always_inline void setup_smap(struct cpuinfo_x86 *c)
}
}

+/*
+ * These can have bit 63 set, so we can not just use a plain "or"
+ * instruction to get their value or'd into CR3. It would take
+ * another register. So, we use a memory reference to these
+ * instead.
+ *
+ * This is also handy because systems that do not support
+ * PCIDs just end up or'ing a 0 into their CR3, which does
+ * no harm.
+ */
+__aligned(PAGE_SIZE) unsigned long X86_CR3_PCID_KERN_VAR = 0;
+__aligned(PAGE_SIZE) unsigned long X86_CR3_PCID_USER_VAR = 0;
+
static void setup_pcid(struct cpuinfo_x86 *c)
{
if (cpu_has(c, X86_FEATURE_PCID)) {
if (cpu_has(c, X86_FEATURE_PGE)) {
cr4_set_bits(X86_CR4_PCIDE);
+ /*
+ * These variables are used by the entry/exit
+ * code to change PCIDs.
+ */
+#ifdef CONFIG_KAISER
+ X86_CR3_PCID_KERN_VAR = X86_CR3_PCID_KERN_NOFLUSH;
+ X86_CR3_PCID_USER_VAR = X86_CR3_PCID_USER_NOFLUSH;
+#endif
+ /*
+ * INVPCID has two "groups" of types:
+ * 1/2: Invalidate an individual address
+ * 3/4: Invalidate all contexts
+ *
+ * 1/2 take a PCID, but 3/4 do not. So, 3/4
+ * ignore the PCID argument in the descriptor.
+ * But, we have to be careful not to call 1/2
+ * with an actual non-zero PCID in them before
+ * we do the above cr4_set_bits().
+ */
+ if (cpu_has(c, X86_FEATURE_INVPCID))
+ set_cpu_cap(c, X86_FEATURE_INVPCID_SINGLE);
} else {
/*
* flush_tlb_all(), as currently implemented, won't
diff --git a/arch/x86/kernel/entry_64.S b/arch/x86/kernel/entry_64.S
index d2b2372e01c6..0f7bba928264 100644
--- a/arch/x86/kernel/entry_64.S
+++ b/arch/x86/kernel/entry_64.S
@@ -1541,7 +1541,10 @@ ENTRY(nmi)
/* %rax is saved above, so OK to clobber here */
movq %cr3, %rax
pushq %rax
- andq $(~KAISER_SHADOW_PGD_OFFSET), %rax
+ /* mask off "user" bit of pgd address and 12 PCID bits: */
+ andq $(~(X86_CR3_PCID_ASID_MASK | KAISER_SHADOW_PGD_OFFSET)), %rax
+ /* Add back kernel PCID and "no flush" bit */
+ orq X86_CR3_PCID_KERN_VAR, %rax
movq %rax, %cr3
#endif
call do_nmi
@@ -1777,7 +1780,10 @@ end_repeat_nmi:
/* %rax is saved above, so OK to clobber here */
movq %cr3, %rax
pushq %rax
- andq $(~KAISER_SHADOW_PGD_OFFSET), %rax
+ /* mask off "user" bit of pgd address and 12 PCID bits: */
+ andq $(~(X86_CR3_PCID_ASID_MASK | KAISER_SHADOW_PGD_OFFSET)), %rax
+ /* Add back kernel PCID and "no flush" bit */
+ orq X86_CR3_PCID_KERN_VAR, %rax
movq %rax, %cr3
#endif
DEFAULT_FRAME 0 /* XXX: Do we need this? */
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index e4e7d45fd551..153afbb6c080 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -733,7 +733,8 @@ int kvm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
return 1;

/* PCID can not be enabled when cr3[11:0]!=000H or EFER.LMA=0 */
- if ((kvm_read_cr3(vcpu) & X86_CR3_PCID_MASK) || !is_long_mode(vcpu))
+ if ((kvm_read_cr3(vcpu) & X86_CR3_PCID_ASID_MASK) ||
+ !is_long_mode(vcpu))
return 1;
}

diff --git a/arch/x86/mm/kaiser.c b/arch/x86/mm/kaiser.c
index 50d650799f39..91968328ccdf 100644
--- a/arch/x86/mm/kaiser.c
+++ b/arch/x86/mm/kaiser.c
@@ -240,6 +240,8 @@ static void __init kaiser_init_all_pgds(void)
} while (0)

extern char __per_cpu_user_mapped_start[], __per_cpu_user_mapped_end[];
+extern unsigned long X86_CR3_PCID_KERN_VAR;
+extern unsigned long X86_CR3_PCID_USER_VAR;
/*
* If anything in here fails, we will likely die on one of the
* first kernel->user transitions and init will die. But, we
@@ -290,6 +292,11 @@ void __init kaiser_init(void)
kaiser_add_user_map_early(&debug_idt_table,
sizeof(gate_desc) * NR_VECTORS,
__PAGE_KERNEL);
+
+ kaiser_add_user_map_early(&X86_CR3_PCID_KERN_VAR, PAGE_SIZE,
+ __PAGE_KERNEL);
+ kaiser_add_user_map_early(&X86_CR3_PCID_USER_VAR, PAGE_SIZE,
+ __PAGE_KERNEL);
}

/* Add a mapping to the shadow mapping, and synchronize the mappings */
diff --git a/arch/x86/mm/tlb.c b/arch/x86/mm/tlb.c
index 95624d903c0e..7176496112ff 100644
--- a/arch/x86/mm/tlb.c
+++ b/arch/x86/mm/tlb.c
@@ -34,6 +34,46 @@ struct flush_tlb_info {
unsigned long flush_end;
};

+static void load_new_mm_cr3(pgd_t *pgdir)
+{
+ unsigned long new_mm_cr3 = __pa(pgdir);
+
+ /*
+ * KAISER, plus PCIDs needs some extra work here. But,
+ * if either of features is not present, we need no
+ * PCIDs here and just do a normal, full TLB flush with
+ * the write_cr3()
+ */
+ if (!IS_ENABLED(CONFIG_KAISER) ||
+ !cpu_feature_enabled(X86_FEATURE_PCID))
+ goto out_set_cr3;
+ /*
+ * We reuse the same PCID for different tasks, so we must
+ * flush all the entires for the PCID out when we change
+ * tasks.
+ */
+ new_mm_cr3 = X86_CR3_PCID_KERN_FLUSH | __pa(pgdir);
+
+ /*
+ * The flush from load_cr3() may leave old TLB entries
+ * for userspace in place. We must flush that context
+ * separately. We can theoretically delay doing this
+ * until we actually load up the userspace CR3, but
+ * that's a bit tricky. We have to have the "need to
+ * flush userspace PCID" bit per-cpu and check it in the
+ * exit-to-userspace paths.
+ */
+ invpcid_flush_single_context(X86_CR3_PCID_ASID_USER);
+
+out_set_cr3:
+ /*
+ * Caution: many callers of this function expect
+ * that load_cr3() is serializing and orders TLB
+ * fills with respect to the mm_cpumask writes.
+ */
+ write_cr3(new_mm_cr3);
+}
+
/*
* We cannot call mmdrop() because we are in interrupt context,
* instead update mm->cpu_vm_mask.
@@ -45,7 +85,7 @@ void leave_mm(int cpu)
BUG();
if (cpumask_test_cpu(cpu, mm_cpumask(active_mm))) {
cpumask_clear_cpu(cpu, mm_cpumask(active_mm));
- load_cr3(swapper_pg_dir);
+ load_new_mm_cr3(swapper_pg_dir);
/*
* This gets called in the idle path where RCU
* functions differently. Tracing normally
@@ -105,7 +145,7 @@ void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
* ordering guarantee we need.
*
*/
- load_cr3(next->pgd);
+ load_new_mm_cr3(next->pgd);

trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);

@@ -150,7 +190,7 @@ void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
* As above, load_cr3() is serializing and orders TLB
* fills with respect to the mm_cpumask write.
*/
- load_cr3(next->pgd);
+ load_new_mm_cr3(next->pgd);
trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
load_mm_cr4(next);
load_mm_ldt(next);
--
2.16.2