Re: [PATCH] KVM: x86: Extend Spectre-v1 mitigation
From: Liran Alon
Date: Fri Nov 22 2019 - 17:05:12 EST
> On 22 Nov 2019, at 20:40, Marios Pomonis <pomonis@xxxxxxxxxx> wrote:
>
> From: Nick Finco <nifi@xxxxxxxxxx>
>
> This extends the Spectre-v1 mitigation introduced in
> commit 75f139aaf896 ("KVM: x86: Add memory barrier on vmcs field lookup")
> and commit 085331dfc6bb ("x86/kvm: Update spectre-v1 mitigation") in light
> of the Spectre-v1/L1TF combination described here:
> https://urldefense.proofpoint.com/v2/url?u=https-3A__xenbits.xen.org_xsa_advisory-2D289.html&d=DwIBaQ&c=RoP1YumCXCgaWHvlZYR8PZh8Bv7qIrMUB65eapI_JnE&r=Jk6Q8nNzkQ6LJ6g42qARkg6ryIDGQr-yKXPNGZbpTx0&m=1r1_2c3bq4nLv-MPEuf_dyQR9r4XHB3Vst3wJeE43pY&s=5GOdMhJ_AQYUkAPAveAzmHrO7hZ0qOgxKmPqKouGvnc&e=
>
> As reported in the link, an attacker can use the cache-load part of a
> Spectre-v1 gadget to bring memory into the L1 cache, then use L1TF to
> leak the loaded memory. Note that this attack is not fully mitigated by
> core scheduling; an attacker could employ L1TF on the same thread that
> loaded the memory in L1 instead of relying on neighboring hyperthreads.
This description is not accurate. The attack still relies on SMT.
If KVM flush L1D$ on every VMEntry to guest, then the only way to exploit this is via a sibling hyperthread.
Itâs true that core-scheduling does not mitigate this scenario completely, but not because the reason described above.
The attack vector that still works when hyperthreading and core-scheduling is enabled, is if on a single core,
one hyperthread runs a VMExit handler which executes a cache-load gadget (even speculatively) while
the sibling hyperthread runs inside guest and executes L1TF code sequence to leak from the coreâs L1D$.
You can find more information on this specific attack vector in the following KVM Forum 2019 slides:
https://static.sched.com/hosted_files/kvmforum2019/34/KVM%20Forum%202019%20KVM%20ASI.pdf
>
> This patch uses array_index_nospec() to prevent index computations from
> causing speculative loads into the L1 cache. These cases involve a
> bounds check followed by a memory read using the index; this is more
> common than the full Spectre-v1 pattern. In some cases, the index
> computation can be eliminated entirely by small amounts of refactoring.
>
> Signed-off-by: Nick Finco <nifi@xxxxxxxxxx>
> Signed-off-by: Marios Pomonis <pomonis@xxxxxxxxxx>
> Acked-by: Andrew Honig <ahonig@xxxxxxxxxx>
I agree with Sean that this patch should be split into a patch-series with multiple patches.
It will be easier to review and more importantly, bisect in case one of the changes here cause a regression.
In addition, I think that every such patch should have a good explanation in commit message on why it is required.
i.e. A proof that the index used as part of the cache-load gadget is attacker-controllable.
I had a brief look over the patch below and it seems that there are some changes to code-path where attacker doesnât control the index.
(I commented inline on those changes)
Therefore, I think these changes shouldnât be necessary.
-Liran
> ---
> arch/x86/kvm/emulate.c | 11 ++++++++---
> arch/x86/kvm/hyperv.c | 10 ++++++----
> arch/x86/kvm/i8259.c | 6 +++++-
> arch/x86/kvm/ioapic.c | 15 +++++++++------
> arch/x86/kvm/lapic.c | 13 +++++++++----
> arch/x86/kvm/mtrr.c | 8 ++++++--
> arch/x86/kvm/pmu.h | 18 ++++++++++++++----
> arch/x86/kvm/vmx/pmu_intel.c | 24 ++++++++++++++++--------
> arch/x86/kvm/vmx/vmx.c | 22 ++++++++++++++++------
> arch/x86/kvm/x86.c | 18 ++++++++++++++----
> 10 files changed, 103 insertions(+), 42 deletions(-)
>
> diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
> index 952d1a4f4d7e..fcf7cdb21d60 100644
> --- a/arch/x86/kvm/emulate.c
> +++ b/arch/x86/kvm/emulate.c
> @@ -5303,10 +5303,15 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len)
> }
> break;
> case Escape:
> - if (ctxt->modrm > 0xbf)
> - opcode = opcode.u.esc->high[ctxt->modrm - 0xc0];
> - else
> + if (ctxt->modrm > 0xbf) {
> + size_t size = ARRAY_SIZE(opcode.u.esc->high);
> + u32 index = array_index_nospec(
> + ctxt->modrm - 0xc0, size);
> +
> + opcode = opcode.u.esc->high[index];
> + } else {
> opcode = opcode.u.esc->op[(ctxt->modrm >> 3) & 7];
> + }
> break;
> case InstrDual:
> if ((ctxt->modrm >> 6) == 3)
> diff --git a/arch/x86/kvm/hyperv.c b/arch/x86/kvm/hyperv.c
> index 23ff65504d7e..26408434b9bc 100644
> --- a/arch/x86/kvm/hyperv.c
> +++ b/arch/x86/kvm/hyperv.c
> @@ -809,11 +809,12 @@ static int kvm_hv_msr_get_crash_data(struct kvm_vcpu *vcpu,
> u32 index, u64 *pdata)
> {
> struct kvm_hv *hv = &vcpu->kvm->arch.hyperv;
> + size_t size = ARRAY_SIZE(hv->hv_crash_param);
>
> - if (WARN_ON_ONCE(index >= ARRAY_SIZE(hv->hv_crash_param)))
> + if (WARN_ON_ONCE(index >= size))
> return -EINVAL;
>
> - *pdata = hv->hv_crash_param[index];
> + *pdata = hv->hv_crash_param[array_index_nospec(index, size)];
> return 0;
> }
>
> @@ -852,11 +853,12 @@ static int kvm_hv_msr_set_crash_data(struct kvm_vcpu *vcpu,
> u32 index, u64 data)
> {
> struct kvm_hv *hv = &vcpu->kvm->arch.hyperv;
> + size_t size = ARRAY_SIZE(hv->hv_crash_param);
>
> - if (WARN_ON_ONCE(index >= ARRAY_SIZE(hv->hv_crash_param)))
> + if (WARN_ON_ONCE(index >= size))
> return -EINVAL;
>
> - hv->hv_crash_param[index] = data;
> + hv->hv_crash_param[array_index_nospec(index, size)] = data;
> return 0;
> }
>
> diff --git a/arch/x86/kvm/i8259.c b/arch/x86/kvm/i8259.c
> index 8b38bb4868a6..629a09ca9860 100644
> --- a/arch/x86/kvm/i8259.c
> +++ b/arch/x86/kvm/i8259.c
> @@ -460,10 +460,14 @@ static int picdev_write(struct kvm_pic *s,
> switch (addr) {
> case 0x20:
> case 0x21:
> + pic_lock(s);
> + pic_ioport_write(&s->pics[0], addr, data);
> + pic_unlock(s);
> + break;
> case 0xa0:
> case 0xa1:
> pic_lock(s);
> - pic_ioport_write(&s->pics[addr >> 7], addr, data);
> + pic_ioport_write(&s->pics[1], addr, data);
> pic_unlock(s);
> break;
> case 0x4d0:
> diff --git a/arch/x86/kvm/ioapic.c b/arch/x86/kvm/ioapic.c
> index 9fd2dd89a1c5..8aa58727045e 100644
> --- a/arch/x86/kvm/ioapic.c
> +++ b/arch/x86/kvm/ioapic.c
> @@ -36,6 +36,7 @@
> #include <linux/io.h>
> #include <linux/slab.h>
> #include <linux/export.h>
> +#include <linux/nospec.h>
> #include <asm/processor.h>
> #include <asm/page.h>
> #include <asm/current.h>
> @@ -68,13 +69,14 @@ static unsigned long ioapic_read_indirect(struct kvm_ioapic *ioapic,
> default:
> {
> u32 redir_index = (ioapic->ioregsel - 0x10) >> 1;
> - u64 redir_content;
> + u64 redir_content = ~0ULL;
>
> - if (redir_index < IOAPIC_NUM_PINS)
> - redir_content =
> - ioapic->redirtbl[redir_index].bits;
> - else
> - redir_content = ~0ULL;
> + if (redir_index < IOAPIC_NUM_PINS) {
> + u32 index = array_index_nospec(
> + redir_index, IOAPIC_NUM_PINS);
> +
> + redir_content = ioapic->redirtbl[index].bits;
> + }
>
> result = (ioapic->ioregsel & 0x1) ?
> (redir_content >> 32) & 0xffffffff :
> @@ -292,6 +294,7 @@ static void ioapic_write_indirect(struct kvm_ioapic *ioapic, u32 val)
>
> if (index >= IOAPIC_NUM_PINS)
> return;
> + index = array_index_nospec(index, IOAPIC_NUM_PINS);
> e = &ioapic->redirtbl[index];
> mask_before = e->fields.mask;
> /* Preserve read-only fields */
> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
> index cf9177b4a07f..3323115f52d5 100644
> --- a/arch/x86/kvm/lapic.c
> +++ b/arch/x86/kvm/lapic.c
> @@ -1963,15 +1963,20 @@ int kvm_lapic_reg_write(struct kvm_lapic *apic, u32 reg, u32 val)
> case APIC_LVTTHMR:
> case APIC_LVTPC:
> case APIC_LVT1:
> - case APIC_LVTERR:
> + case APIC_LVTERR: {
> /* TODO: Check vector */
> + size_t size;
> + u32 index;
> +
> if (!kvm_apic_sw_enabled(apic))
> val |= APIC_LVT_MASKED;
> -
> - val &= apic_lvt_mask[(reg - APIC_LVTT) >> 4];
> + size = ARRAY_SIZE(apic_lvt_mask);
> + index = array_index_nospec(
> + (reg - APIC_LVTT) >> 4, size);
> + val &= apic_lvt_mask[index];
> kvm_lapic_set_reg(apic, reg, val);
> -
> break;
> + }
>
> case APIC_LVTT:
> if (!kvm_apic_sw_enabled(apic))
> diff --git a/arch/x86/kvm/mtrr.c b/arch/x86/kvm/mtrr.c
> index 25ce3edd1872..7f0059aa30e1 100644
> --- a/arch/x86/kvm/mtrr.c
> +++ b/arch/x86/kvm/mtrr.c
> @@ -192,11 +192,15 @@ static bool fixed_msr_to_seg_unit(u32 msr, int *seg, int *unit)
> break;
> case MSR_MTRRfix16K_80000 ... MSR_MTRRfix16K_A0000:
> *seg = 1;
> - *unit = msr - MSR_MTRRfix16K_80000;
> + *unit = array_index_nospec(
> + msr - MSR_MTRRfix16K_80000,
> + MSR_MTRRfix16K_A0000 - MSR_MTRRfix16K_80000 + 1);
> break;
> case MSR_MTRRfix4K_C0000 ... MSR_MTRRfix4K_F8000:
> *seg = 2;
> - *unit = msr - MSR_MTRRfix4K_C0000;
> + *unit = array_index_nospec(
> + msr - MSR_MTRRfix4K_C0000,
> + MSR_MTRRfix4K_F8000 - MSR_MTRRfix4K_C0000 + 1);
> break;
> default:
> return false;
> diff --git a/arch/x86/kvm/pmu.h b/arch/x86/kvm/pmu.h
> index 7ebb62326c14..13332984b6d5 100644
> --- a/arch/x86/kvm/pmu.h
> +++ b/arch/x86/kvm/pmu.h
> @@ -2,6 +2,8 @@
> #ifndef __KVM_X86_PMU_H
> #define __KVM_X86_PMU_H
>
> +#include <linux/nospec.h>
> +
> #define vcpu_to_pmu(vcpu) (&(vcpu)->arch.pmu)
> #define pmu_to_vcpu(pmu) (container_of((pmu), struct kvm_vcpu, arch.pmu))
> #define pmc_to_pmu(pmc) (&(pmc)->vcpu->arch.pmu)
> @@ -102,8 +104,12 @@ static inline bool kvm_valid_perf_global_ctrl(struct kvm_pmu *pmu,
> static inline struct kvm_pmc *get_gp_pmc(struct kvm_pmu *pmu, u32 msr,
> u32 base)
> {
> - if (msr >= base && msr < base + pmu->nr_arch_gp_counters)
> - return &pmu->gp_counters[msr - base];
> + if (msr >= base && msr < base + pmu->nr_arch_gp_counters) {
> + u32 index = array_index_nospec(msr - base,
> + pmu->nr_arch_gp_counters);
> +
> + return &pmu->gp_counters[index];
> + }
>
> return NULL;
> }
> @@ -113,8 +119,12 @@ static inline struct kvm_pmc *get_fixed_pmc(struct kvm_pmu *pmu, u32 msr)
> {
> int base = MSR_CORE_PERF_FIXED_CTR0;
>
> - if (msr >= base && msr < base + pmu->nr_arch_fixed_counters)
> - return &pmu->fixed_counters[msr - base];
> + if (msr >= base && msr < base + pmu->nr_arch_fixed_counters) {
> + u32 index = array_index_nospec(msr - base,
> + pmu->nr_arch_fixed_counters);
> +
> + return &pmu->fixed_counters[index];
> + }
>
> return NULL;
> }
> diff --git a/arch/x86/kvm/vmx/pmu_intel.c b/arch/x86/kvm/vmx/pmu_intel.c
> index 7023138b1cb0..34a3a17bb6d7 100644
> --- a/arch/x86/kvm/vmx/pmu_intel.c
> +++ b/arch/x86/kvm/vmx/pmu_intel.c
> @@ -86,10 +86,14 @@ static unsigned intel_find_arch_event(struct kvm_pmu *pmu,
>
> static unsigned intel_find_fixed_event(int idx)
> {
> - if (idx >= ARRAY_SIZE(fixed_pmc_events))
> + u32 event;
> + size_t size = ARRAY_SIZE(fixed_pmc_events);
> +
> + if (idx >= size)
> return PERF_COUNT_HW_MAX;
>
> - return intel_arch_events[fixed_pmc_events[idx]].event_type;
> + event = fixed_pmc_events[array_index_nospec(idx, size)];
> + return intel_arch_events[event].event_type;
> }
>
> /* check if a PMC is enabled by comparing it with globl_ctrl bits. */
> @@ -130,16 +134,20 @@ static struct kvm_pmc *intel_rdpmc_ecx_to_pmc(struct kvm_vcpu *vcpu,
> struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
> bool fixed = idx & (1u << 30);
> struct kvm_pmc *counters;
> + unsigned int num_counters;
>
> idx &= ~(3u << 30);
> - if (!fixed && idx >= pmu->nr_arch_gp_counters)
> - return NULL;
> - if (fixed && idx >= pmu->nr_arch_fixed_counters)
> + if (fixed) {
> + counters = pmu->fixed_counters;
> + num_counters = pmu->nr_arch_fixed_counters;
> + } else {
> + counters = pmu->gp_counters;
> + num_counters = pmu->nr_arch_gp_counters;
> + }
> + if (idx >= num_counters)
> return NULL;
> - counters = fixed ? pmu->fixed_counters : pmu->gp_counters;
> *mask &= pmu->counter_bitmask[fixed ? KVM_PMC_FIXED : KVM_PMC_GP];
> -
> - return &counters[idx];
> + return &counters[array_index_nospec(idx, num_counters)];
> }
>
> static bool intel_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index d39475e2d44e..84def8e46d10 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -753,7 +753,9 @@ static bool vmx_segment_cache_test_set(struct vcpu_vmx *vmx, unsigned seg,
>
> static u16 vmx_read_guest_seg_selector(struct vcpu_vmx *vmx, unsigned seg)
> {
> - u16 *p = &vmx->segment_cache.seg[seg].selector;
> + size_t size = ARRAY_SIZE(vmx->segment_cache.seg);
> + size_t index = array_index_nospec(seg, size);
> + u16 *p = &vmx->segment_cache.seg[index].selector;
>
> if (!vmx_segment_cache_test_set(vmx, seg, SEG_FIELD_SEL))
> *p = vmcs_read16(kvm_vmx_segment_fields[seg].selector);
> @@ -762,7 +764,9 @@ static u16 vmx_read_guest_seg_selector(struct vcpu_vmx *vmx, unsigned seg)
>
> static ulong vmx_read_guest_seg_base(struct vcpu_vmx *vmx, unsigned seg)
> {
> - ulong *p = &vmx->segment_cache.seg[seg].base;
> + size_t size = ARRAY_SIZE(vmx->segment_cache.seg);
> + size_t index = array_index_nospec(seg, size);
> + ulong *p = &vmx->segment_cache.seg[index].base;
>
> if (!vmx_segment_cache_test_set(vmx, seg, SEG_FIELD_BASE))
> *p = vmcs_readl(kvm_vmx_segment_fields[seg].base);
> @@ -771,7 +775,9 @@ static ulong vmx_read_guest_seg_base(struct vcpu_vmx *vmx, unsigned seg)
>
> static u32 vmx_read_guest_seg_limit(struct vcpu_vmx *vmx, unsigned seg)
> {
> - u32 *p = &vmx->segment_cache.seg[seg].limit;
> + size_t size = ARRAY_SIZE(vmx->segment_cache.seg);
> + size_t index = array_index_nospec(seg, size);
> + u32 *p = &vmx->segment_cache.seg[index].limit;
>
> if (!vmx_segment_cache_test_set(vmx, seg, SEG_FIELD_LIMIT))
> *p = vmcs_read32(kvm_vmx_segment_fields[seg].limit);
> @@ -780,7 +786,9 @@ static u32 vmx_read_guest_seg_limit(struct vcpu_vmx *vmx, unsigned seg)
>
> static u32 vmx_read_guest_seg_ar(struct vcpu_vmx *vmx, unsigned seg)
> {
> - u32 *p = &vmx->segment_cache.seg[seg].ar;
> + size_t size = ARRAY_SIZE(vmx->segment_cache.seg);
> + size_t index = array_index_nospec(seg, size);
> + u32 *p = &vmx->segment_cache.seg[index].ar;
>
> if (!vmx_segment_cache_test_set(vmx, seg, SEG_FIELD_AR))
> *p = vmcs_read32(kvm_vmx_segment_fields[seg].ar_bytes);
> @@ -5828,6 +5836,8 @@ static int vmx_handle_exit(struct kvm_vcpu *vcpu)
> {
> struct vcpu_vmx *vmx = to_vmx(vcpu);
> u32 exit_reason = vmx->exit_reason;
> + u32 bounded_exit_reason = array_index_nospec(exit_reason,
> + kvm_vmx_max_exit_handlers);
Unlike the rest of this patch changes, exit_reason is not attacker-controllable.
Therefore, I donât think we need this change to vmx_handle_exit().
> u32 vectoring_info = vmx->idt_vectoring_info;
>
> trace_kvm_exit(exit_reason, vcpu, KVM_ISA_VMX);
> @@ -5911,7 +5921,7 @@ static int vmx_handle_exit(struct kvm_vcpu *vcpu)
> }
>
> if (exit_reason < kvm_vmx_max_exit_handlers
> - && kvm_vmx_exit_handlers[exit_reason]) {
> + && kvm_vmx_exit_handlers[bounded_exit_reason]) {
> #ifdef CONFIG_RETPOLINE
> if (exit_reason == EXIT_REASON_MSR_WRITE)
> return kvm_emulate_wrmsr(vcpu);
> @@ -5926,7 +5936,7 @@ static int vmx_handle_exit(struct kvm_vcpu *vcpu)
> else if (exit_reason == EXIT_REASON_EPT_MISCONFIG)
> return handle_ept_misconfig(vcpu);
> #endif
> - return kvm_vmx_exit_handlers[exit_reason](vcpu);
> + return kvm_vmx_exit_handlers[bounded_exit_reason](vcpu);
> } else {
> vcpu_unimpl(vcpu, "vmx: unexpected exit reason 0x%x\n",
> exit_reason);
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index a256e09f321a..9a2789652231 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -1057,9 +1057,11 @@ static u64 kvm_dr6_fixed(struct kvm_vcpu *vcpu)
>
> static int __kvm_set_dr(struct kvm_vcpu *vcpu, int dr, unsigned long val)
> {
> + size_t size = ARRAY_SIZE(vcpu->arch.db);
> +
Why is this change needed?
âdrâ shouldnât be attacker-controllable.
Itâs value is read from vmcs->exit_qualification or vmcb->control.exit_code on DR-access intercept.
Relevant also for kvm_get_dr().
> switch (dr) {
> case 0 ... 3:
> - vcpu->arch.db[dr] = val;
> + vcpu->arch.db[array_index_nospec(dr, size)] = val;
> if (!(vcpu->guest_debug & KVM_GUESTDBG_USE_HW_BP))
> vcpu->arch.eff_db[dr] = val;
> break;
> @@ -1096,9 +1098,11 @@ EXPORT_SYMBOL_GPL(kvm_set_dr);
>
> int kvm_get_dr(struct kvm_vcpu *vcpu, int dr, unsigned long *val)
> {
> + size_t size = ARRAY_SIZE(vcpu->arch.db);
> +
> switch (dr) {
> case 0 ... 3:
> - *val = vcpu->arch.db[dr];
> + *val = vcpu->arch.db[array_index_nospec(dr, size)];
> break;
> case 4:
> /* fall through */
> @@ -2496,7 +2500,10 @@ static int set_msr_mce(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
> default:
> if (msr >= MSR_IA32_MC0_CTL &&
> msr < MSR_IA32_MCx_CTL(bank_num)) {
> - u32 offset = msr - MSR_IA32_MC0_CTL;
> + u32 offset = array_index_nospec(
> + msr - MSR_IA32_MC0_CTL,
> + MSR_IA32_MCx_CTL(bank_num) - MSR_IA32_MC0_CTL);
> +
> /* only 0 or all 1s can be written to IA32_MCi_CTL
> * some Linux kernels though clear bit 10 in bank 4 to
> * workaround a BIOS/GART TBL issue on AMD K8s, ignore
> @@ -2937,7 +2944,10 @@ static int get_msr_mce(struct kvm_vcpu *vcpu, u32 msr, u64 *pdata, bool host)
> default:
> if (msr >= MSR_IA32_MC0_CTL &&
> msr < MSR_IA32_MCx_CTL(bank_num)) {
> - u32 offset = msr - MSR_IA32_MC0_CTL;
> + u32 offset = array_index_nospec(
> + msr - MSR_IA32_MC0_CTL,
> + MSR_IA32_MCx_CTL(bank_num) - MSR_IA32_MC0_CTL);
> +
> data = vcpu->arch.mce_banks[offset];
> break;
> }
> --
> 2.24.0.432.g9d3f5f5b63-goog
>