Re: [PATCH v4 09/13] KVM: selftests: aarch64: Add aarch64/page_fault_test

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hey Oliver,

On Tue, Jun 28, 2022 at 04:43:29PM -0700, Oliver Upton wrote:
> Hi Ricardo,
> 
> On Fri, Jun 24, 2022 at 02:32:53PM -0700, Ricardo Koller wrote:
> > Add a new test for stage 2 faults when using different combinations of
> > guest accesses (e.g., write, S1PTW), backing source type (e.g., anon)
> > and types of faults (e.g., read on hugetlbfs with a hole). The next
> > commits will add different handling methods and more faults (e.g., uffd
> > and dirty logging). This first commit starts by adding two sanity checks
> > for all types of accesses: AF setting by the hw, and accessing memslots
> > with holes.
> > 
> > Signed-off-by: Ricardo Koller <ricarkol@xxxxxxxxxx>
> > ---
> >  tools/testing/selftests/kvm/Makefile          |   1 +
> >  .../selftests/kvm/aarch64/page_fault_test.c   | 695 ++++++++++++++++++
> >  .../selftests/kvm/include/aarch64/processor.h |   6 +
> >  3 files changed, 702 insertions(+)
> >  create mode 100644 tools/testing/selftests/kvm/aarch64/page_fault_test.c
> > 
> > diff --git a/tools/testing/selftests/kvm/Makefile b/tools/testing/selftests/kvm/Makefile
> > index e4497a3a27d4..13b913225ae7 100644
> > --- a/tools/testing/selftests/kvm/Makefile
> > +++ b/tools/testing/selftests/kvm/Makefile
> > @@ -139,6 +139,7 @@ TEST_GEN_PROGS_aarch64 += aarch64/arch_timer
> >  TEST_GEN_PROGS_aarch64 += aarch64/debug-exceptions
> >  TEST_GEN_PROGS_aarch64 += aarch64/get-reg-list
> >  TEST_GEN_PROGS_aarch64 += aarch64/hypercalls
> > +TEST_GEN_PROGS_aarch64 += aarch64/page_fault_test
> >  TEST_GEN_PROGS_aarch64 += aarch64/psci_test
> >  TEST_GEN_PROGS_aarch64 += aarch64/vcpu_width_config
> >  TEST_GEN_PROGS_aarch64 += aarch64/vgic_init
> > diff --git a/tools/testing/selftests/kvm/aarch64/page_fault_test.c b/tools/testing/selftests/kvm/aarch64/page_fault_test.c
> > new file mode 100644
> > index 000000000000..bdda4e3fcdaa
> > --- /dev/null
> > +++ b/tools/testing/selftests/kvm/aarch64/page_fault_test.c
> 
> [...]
> 
> > +/* Compare and swap instruction. */
> > +static void guest_cas(void)
> > +{
> > +	uint64_t val;
> > +
> > +	GUEST_ASSERT_EQ(guest_check_lse(), 1);
> 
> Why not just GUEST_ASSERT(guest_check_lse()) ?
> 
> > +	asm volatile(".arch_extension lse\n"
> > +		     "casal %0, %1, [%2]\n"
> > +			:: "r" (0), "r" (TEST_DATA), "r" (guest_test_memory));
> > +	val = READ_ONCE(*guest_test_memory);
> > +	GUEST_ASSERT_EQ(val, TEST_DATA);
> > +}
> > +
> > +static void guest_read64(void)
> > +{
> > +	uint64_t val;
> > +
> > +	val = READ_ONCE(*guest_test_memory);
> > +	GUEST_ASSERT_EQ(val, 0);
> > +}
> > +
> > +/* Address translation instruction */
> > +static void guest_at(void)
> > +{
> > +	uint64_t par;
> > +	uint64_t paddr;
> > +
> > +	asm volatile("at s1e1r, %0" :: "r" (guest_test_memory));
> > +	par = read_sysreg(par_el1);
> 
> I believe you need explicit synchronization (an isb) before the fault
> information is guaranteed visibile in PAR_EL1.
> 
> > +	/* Bit 1 indicates whether the AT was successful */
> > +	GUEST_ASSERT_EQ(par & 1, 0);
> > +	/* The PA in bits [51:12] */
> > +	paddr = par & (((1ULL << 40) - 1) << 12);
> > +	GUEST_ASSERT_EQ(paddr, memslot[TEST].gpa);
> > +}
> > +
> > +/*
> > + * The size of the block written by "dc zva" is guaranteed to be between (2 <<
> > + * 0) and (2 << 9), which is safe in our case as we need the write to happen
> > + * for at least a word, and not more than a page.
> > + */
> > +static void guest_dc_zva(void)
> > +{
> > +	uint16_t val;
> > +
> > +	asm volatile("dc zva, %0\n"
> > +			"dsb ish\n"
> 
> nit: use the dsb() macro instead. Extremely minor, but makes it a bit
> more obvious to the reader. Or maybe I need to get my eyes checked ;-)
> 
> > +			:: "r" (guest_test_memory));
> > +	val = READ_ONCE(*guest_test_memory);
> > +	GUEST_ASSERT_EQ(val, 0);
> > +}
> > +
> > +/*
> > + * Pre-indexing loads and stores don't have a valid syndrome (ESR_EL2.ISV==0).
> > + * And that's special because KVM must take special care with those: they
> > + * should still count as accesses for dirty logging or user-faulting, but
> > + * should be handled differently on mmio.
> > + */
> > +static void guest_ld_preidx(void)
> > +{
> > +	uint64_t val;
> > +	uint64_t addr = TEST_GVA - 8;
> > +
> > +	/*
> > +	 * This ends up accessing "TEST_GVA + 8 - 8", where "TEST_GVA - 8" is
> > +	 * in a gap between memslots not backing by anything.
> > +	 */
> > +	asm volatile("ldr %0, [%1, #8]!"
> > +			: "=r" (val), "+r" (addr));
> > +	GUEST_ASSERT_EQ(val, 0);
> > +	GUEST_ASSERT_EQ(addr, TEST_GVA);
> > +}
> > +
> > +static void guest_st_preidx(void)
> > +{
> > +	uint64_t val = TEST_DATA;
> > +	uint64_t addr = TEST_GVA - 8;
> > +
> > +	asm volatile("str %0, [%1, #8]!"
> > +			: "+r" (val), "+r" (addr));
> > +
> > +	GUEST_ASSERT_EQ(addr, TEST_GVA);
> > +	val = READ_ONCE(*guest_test_memory);
> > +}
> > +
> > +static bool guest_set_ha(void)
> > +{
> > +	uint64_t mmfr1 = read_sysreg(id_aa64mmfr1_el1);
> > +	uint64_t hadbs, tcr;
> > +
> > +	/* Skip if HA is not supported. */
> > +	hadbs = FIELD_GET(ARM64_FEATURE_MASK(ID_AA64MMFR1_HADBS), mmfr1);
> > +	if (hadbs == 0)
> > +		return false;
> > +
> > +	tcr = read_sysreg(tcr_el1) | TCR_EL1_HA;
> > +	write_sysreg(tcr, tcr_el1);
> > +	isb();
> > +
> > +	return true;
> > +}
> > +
> > +static bool guest_clear_pte_af(void)
> > +{
> > +	*((uint64_t *)TEST_PTE_GVA) &= ~PTE_AF;
> > +	flush_tlb_page(TEST_PTE_GVA);
> 
> Don't you want to actually flush TEST_GVA to force the TLB fill when you
> poke the address again? This looks like you're flushing the VA of the
> *PTE* not the test address.

Yes, you are right, this was supposed to be:
flush_tlb_page(TEST_GVA);
(I could swear this was TEST_GVA at one time)

> 
> > +	return true;
> > +}
> > +
> > +static void guest_check_pte_af(void)
> 
> nit: call this guest_test_pte_af(). You use the guest_check_* pattern
> for test preconditions (like guest_check_lse()).
> 
> > +{
> > +	flush_tlb_page(TEST_PTE_GVA);
> 
> What is the purpose of this flush? I believe you are actually depending
> on a dsb(ish) between the hardware PTE update and the load below. Or,
> that's at least what I gleaned from the jargon of DDI0487H.a D5.4.13 
> 'Ordering of hardware updates to the translation tables'.

This was also supposed to be: flush_tlb_page(TEST_GVA)
But will removed based on D5.4.13, as it's indeed saying that the DSB
should be enough.

> 
> > +	GUEST_ASSERT_EQ(*((uint64_t *)TEST_PTE_GVA) & PTE_AF, PTE_AF);
> > +}
> 
> [...]
> 
> > +static void sync_stats_from_guest(struct kvm_vm *vm)
> > +{
> > +	struct event_cnt *ec = addr_gva2hva(vm, (uint64_t)&events);
> > +
> > +	events.aborts += ec->aborts;
> > +}
> 
> I believe you can use sync_global_from_guest() instead of this.
> 
> > +void fail_vcpu_run_no_handler(int ret)
> > +{
> > +	TEST_FAIL("Unexpected vcpu run failure\n");
> > +}
> > +
> > +extern unsigned char __exec_test;
> > +
> > +void noinline __return_0x77(void)
> > +{
> > +	asm volatile("__exec_test: mov x0, #0x77\n"
> > +			"ret\n");
> > +}
> > +
> > +static void load_exec_code_for_test(void)
> > +{
> > +	uint64_t *code, *c;
> > +
> > +	assert(TEST_EXEC_GVA - TEST_GVA);
> > +	code = memslot[TEST].hva + 8;
> > +
> > +	/*
> > +	 * We need the cast to be separate in order for the compiler to not
> > +	 * complain with: "‘memcpy’ forming offset [1, 7] is out of the bounds
> > +	 * [0, 1] of object ‘__exec_test’ with type ‘unsigned char’"
> > +	 */
> > +	c = (uint64_t *)&__exec_test;
> > +	memcpy(code, c, 8);
> 
> Don't you need to sync D$ and I$?

This is done before running the VM for the first time, and it's only
ever written this one time. I think KVM itself is doing the sync when
mapping new pages for the first time, which would be this case.

> 
> --
> Thanks,
> Oliver

ACK on all the other points, will fix accordingly.

Thanks for the review!



[Index of Archives]     [KVM ARM]     [KVM ia64]     [KVM ppc]     [Virtualization Tools]     [Spice Development]     [Libvirt]     [Libvirt Users]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite Questions]     [Linux Kernel]     [Linux SCSI]     [XFree86]

  Powered by Linux