LinuxLists.cc - [PATCH v3 0/4] Enable async page faults on s390

[permalink] [raw]

Subject: [PATCH 1/4] PF: Add FAULT_FLAG_RETRY_NOWAIT for guest fault

In case of a fault retry exit sie64() with gmap_fault indication for the
running thread set. This makes it possible to handle async page faults
without the need for mm notifiers.

Based on a patch from Martin Schwidefsky.

Signed-off-by: Dominik Dingel <[email protected]>
---
arch/s390/include/asm/pgtable.h | 2 ++
arch/s390/include/asm/processor.h | 1 +
arch/s390/kvm/kvm-s390.c | 13 +++++++++++++
arch/s390/mm/fault.c | 26 ++++++++++++++++++++++----
4 files changed, 38 insertions(+), 4 deletions(-)

diff --git a/arch/s390/include/asm/pgtable.h b/arch/s390/include/asm/pgtable.h
index 0ea4e59..4a4cc64 100644
--- a/arch/s390/include/asm/pgtable.h
+++ b/arch/s390/include/asm/pgtable.h
@@ -740,6 +740,7 @@ static inline void pgste_set_pte(pte_t *ptep, pte_t entry)
* @table: pointer to the page directory
* @asce: address space control element for gmap page table
* @crst_list: list of all crst tables used in the guest address space
+ * @pfault_enabled: defines if pfaults are applicable for the guest
*/
struct gmap {
struct list_head list;
@@ -748,6 +749,7 @@ struct gmap {
unsigned long asce;
void *private;
struct list_head crst_list;
+ unsigned long pfault_enabled;
};

/**
diff --git a/arch/s390/include/asm/processor.h b/arch/s390/include/asm/processor.h
index 6b49987..4fa96ca 100644
--- a/arch/s390/include/asm/processor.h
+++ b/arch/s390/include/asm/processor.h
@@ -77,6 +77,7 @@ struct thread_struct {
unsigned long ksp; /* kernel stack pointer */
mm_segment_t mm_segment;
unsigned long gmap_addr; /* address of last gmap fault. */
+ unsigned int gmap_pfault; /* signal of a pending guest pfault */
struct per_regs per_user; /* User specified PER registers */
struct per_event per_event; /* Cause of the last PER trap */
unsigned long per_flags; /* Flags to control debug behavior */
diff --git a/arch/s390/kvm/kvm-s390.c b/arch/s390/kvm/kvm-s390.c
index ba694d2..702daca 100644
--- a/arch/s390/kvm/kvm-s390.c
+++ b/arch/s390/kvm/kvm-s390.c
@@ -682,6 +682,15 @@ static int kvm_s390_handle_requests(struct kvm_vcpu *vcpu)
return 0;
}

+static void kvm_arch_fault_in_sync(struct kvm_vcpu *vcpu)
+{
+ hva_t fault = gmap_fault(current->thread.gmap_addr, vcpu->arch.gmap);
+ struct mm_struct *mm = current->mm;
+ down_read(&mm->mmap_sem);
+ get_user_pages(current, mm, fault, 1, 1, 0, NULL, NULL);
+ up_read(&mm->mmap_sem);
+}
+
static int __vcpu_run(struct kvm_vcpu *vcpu)
{
int rc;
@@ -715,6 +724,10 @@ static int __vcpu_run(struct kvm_vcpu *vcpu)
if (rc < 0) {
if (kvm_is_ucontrol(vcpu->kvm)) {
rc = SIE_INTERCEPT_UCONTROL;
+ } else if (current->thread.gmap_pfault) {
+ kvm_arch_fault_in_sync(vcpu);
+ current->thread.gmap_pfault = 0;
+ rc = 0;
} else {
VCPU_EVENT(vcpu, 3, "%s", "fault in sie instruction");
trace_kvm_s390_sie_fault(vcpu);
diff --git a/arch/s390/mm/fault.c b/arch/s390/mm/fault.c
index 047c3e4..7d4c4b1 100644
--- a/arch/s390/mm/fault.c
+++ b/arch/s390/mm/fault.c
@@ -50,6 +50,7 @@
#define VM_FAULT_BADMAP 0x020000
#define VM_FAULT_BADACCESS 0x040000
#define VM_FAULT_SIGNAL 0x080000
+#define VM_FAULT_PFAULT 0x100000

static unsigned long store_indication __read_mostly;

@@ -232,6 +233,7 @@ static noinline void do_fault_error(struct pt_regs *regs, int fault)
return;
}
case VM_FAULT_BADCONTEXT:
+ case VM_FAULT_PFAULT:
do_no_context(regs);
break;
case VM_FAULT_SIGNAL:
@@ -269,6 +271,9 @@ static noinline void do_fault_error(struct pt_regs *regs, int fault)
*/
static inline int do_exception(struct pt_regs *regs, int access)
{
+#ifdef CONFIG_PGSTE
+ struct gmap *gmap;
+#endif
struct task_struct *tsk;
struct mm_struct *mm;
struct vm_area_struct *vma;
@@ -307,9 +312,10 @@ static inline int do_exception(struct pt_regs *regs, int access)
down_read(&mm->mmap_sem);

#ifdef CONFIG_PGSTE
- if ((current->flags & PF_VCPU) && S390_lowcore.gmap) {
- address = __gmap_fault(address,
- (struct gmap *) S390_lowcore.gmap);
+ gmap = (struct gmap *)
+ ((current->flags & PF_VCPU) ? S390_lowcore.gmap : 0);
+ if (gmap) {
+ address = __gmap_fault(address, gmap);
if (address == -EFAULT) {
fault = VM_FAULT_BADMAP;
goto out_up;
@@ -318,6 +324,8 @@ static inline int do_exception(struct pt_regs *regs, int access)
fault = VM_FAULT_OOM;
goto out_up;
}
+ if (test_bit(1, &gmap->pfault_enabled))
+ flags |= FAULT_FLAG_RETRY_NOWAIT;
}
#endif

@@ -374,9 +382,19 @@ retry:
regs, address);
}
if (fault & VM_FAULT_RETRY) {
+#ifdef CONFIG_PGSTE
+ if (gmap && (flags & FAULT_FLAG_RETRY_NOWAIT)) {
+ /* FAULT_FLAG_RETRY_NOWAIT has been set,
+ * mmap_sem has not been released */
+ current->thread.gmap_pfault = 1;
+ fault = VM_FAULT_PFAULT;
+ goto out_up;
+ }
+#endif
/* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
* of starvation. */
- flags &= ~FAULT_FLAG_ALLOW_RETRY;
+ flags &= ~(FAULT_FLAG_ALLOW_RETRY |
+ FAULT_FLAG_RETRY_NOWAIT);
flags |= FAULT_FLAG_TRIED;
down_read(&mm->mmap_sem);
goto retry;
--
1.8.2.2

2013-07-09 13:57:08

[permalink] [raw]

Subject: [PATCH 2/4] PF: Make KVM_HVA_ERR_BAD usable on s390

Current common code uses PAGE_OFFSET to indicate a bad host virtual address.
As this check won't work on architectures that don't map kernel and user memory
into the same address space (e.g. s390), an additional implementation is made
available in the case that PAGE_OFFSET == 0.

Signed-off-by: Dominik Dingel <[email protected]>
---
include/linux/kvm_host.h | 14 ++++++++++++++
1 file changed, 14 insertions(+)

diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
index a63d83e..f3c04e7 100644
--- a/include/linux/kvm_host.h
+++ b/include/linux/kvm_host.h
@@ -85,6 +85,18 @@ static inline bool is_noslot_pfn(pfn_t pfn)
return pfn == KVM_PFN_NOSLOT;
}

+#if (PAGE_OFFSET == 0)
+
+#define KVM_HVA_ERR_BAD (-1UL)
+#define KVM_HVA_ERR_RO_BAD (-1UL)
+
+static inline bool kvm_is_error_hva(unsigned long addr)
+{
+ return addr == KVM_HVA_ERR_BAD;
+}
+
+#else
+
#define KVM_HVA_ERR_BAD (PAGE_OFFSET)
#define KVM_HVA_ERR_RO_BAD (PAGE_OFFSET + PAGE_SIZE)

@@ -93,6 +105,8 @@ static inline bool kvm_is_error_hva(unsigned long addr)
return addr >= PAGE_OFFSET;
}

+#endif
+
#define KVM_ERR_PTR_BAD_PAGE (ERR_PTR(-ENOENT))

static inline bool is_error_page(struct page *page)
--
1.8.2.2

2013-07-09 13:57:38

[permalink] [raw]

Subject: [PATCH 4/4] PF: Async page fault support on s390

This patch enables async page faults for s390 kvm guests.
It provides the userspace API to enable, disable or get the status of this
feature. Also it includes the diagnose code, called by the guest to enable
async page faults.

The async page faults will use an already existing guest interface for this
purpose, as described in "CP Programming Services (SC24-6084)".

Signed-off-by: Dominik Dingel <[email protected]>
---
Documentation/s390/kvm.txt | 24 ++++++++++
arch/s390/include/asm/kvm_host.h | 22 +++++++++
arch/s390/include/uapi/asm/kvm.h | 10 ++++
arch/s390/kvm/Kconfig | 2 +
arch/s390/kvm/Makefile | 2 +-
arch/s390/kvm/diag.c | 57 ++++++++++++++++++++++
arch/s390/kvm/interrupt.c | 38 ++++++++++++---
arch/s390/kvm/kvm-s390.c | 100 ++++++++++++++++++++++++++++++++++++++-
arch/s390/kvm/kvm-s390.h | 4 ++
arch/s390/kvm/sigp.c | 2 +
include/uapi/linux/kvm.h | 2 +
11 files changed, 254 insertions(+), 9 deletions(-)

diff --git a/Documentation/s390/kvm.txt b/Documentation/s390/kvm.txt
index 85f3280..707b7e9 100644
--- a/Documentation/s390/kvm.txt
+++ b/Documentation/s390/kvm.txt
@@ -70,6 +70,30 @@ floating interrupts are:
KVM_S390_INT_VIRTIO
KVM_S390_INT_SERVICE

+ioctl: KVM_S390_APF_ENABLE:
+args: none
+This ioctl is used to enable the async page fault interface. So in a
+host page fault case the host can now submit pfault tokens to the guest.
+
+ioctl: KVM_S390_APF_DISABLE:
+args: none
+This ioctl is used to disable the async page fault interface. From this point
+on no new pfault tokens will be issued to the guest. Already existing async
+page faults are not covered by this and will be normally handled.
+
+ioctl: KVM_S390_APF_STATUS:
+args: none
+This ioctl allows the userspace to get the current status of the APF feature.
+The main purpose for this, is to ensure that no pfault tokens will be lost
+during live migration or similar management operations.
+The possible return values are:
+KVM_S390_APF_DISABLED_NON_PENDING
+KVM_S390_APF_DISABLED_PENDING
+KVM_S390_APF_ENABLED_NON_PENDING
+KVM_S390_APF_ENABLED_PENDING
+Caution: if KVM_S390_APF is enabled the PENDING status could be already changed
+as soon as the ioctl returns to userspace.
+
3. ioctl calls to the kvm-vcpu file descriptor
KVM does support the following ioctls on s390 that are common with other
architectures and do behave the same:
diff --git a/arch/s390/include/asm/kvm_host.h b/arch/s390/include/asm/kvm_host.h
index 3238d40..ae75104 100644
--- a/arch/s390/include/asm/kvm_host.h
+++ b/arch/s390/include/asm/kvm_host.h
@@ -257,6 +257,10 @@ struct kvm_vcpu_arch {
u64 stidp_data;
};
struct gmap *gmap;
+#define KVM_S390_PFAULT_TOKEN_INVALID (-1UL)
+ unsigned long pfault_token;
+ unsigned long pfault_select;
+ unsigned long pfault_compare;
};

struct kvm_vm_stat {
@@ -274,6 +278,24 @@ struct kvm_arch{
int css_support;
};

+#define ASYNC_PF_PER_VCPU 64
+struct kvm_vcpu;
+struct kvm_async_pf;
+struct kvm_arch_async_pf {
+ unsigned long pfault_token;
+};
+
+bool kvm_arch_can_inject_async_page_present(struct kvm_vcpu *vcpu);
+
+void kvm_arch_async_page_ready(struct kvm_vcpu *vcpu,
+ struct kvm_async_pf *work);
+
+void kvm_arch_async_page_not_present(struct kvm_vcpu *vcpu,
+ struct kvm_async_pf *work);
+
+void kvm_arch_async_page_present(struct kvm_vcpu *vcpu,
+ struct kvm_async_pf *work);
+
extern int sie64a(struct kvm_s390_sie_block *, u64 *);
extern char sie_exit;
#endif
diff --git a/arch/s390/include/uapi/asm/kvm.h b/arch/s390/include/uapi/asm/kvm.h
index d25da59..b6c83e0 100644
--- a/arch/s390/include/uapi/asm/kvm.h
+++ b/arch/s390/include/uapi/asm/kvm.h
@@ -57,4 +57,14 @@ struct kvm_sync_regs {
#define KVM_REG_S390_EPOCHDIFF (KVM_REG_S390 | KVM_REG_SIZE_U64 | 0x2)
#define KVM_REG_S390_CPU_TIMER (KVM_REG_S390 | KVM_REG_SIZE_U64 | 0x3)
#define KVM_REG_S390_CLOCK_COMP (KVM_REG_S390 | KVM_REG_SIZE_U64 | 0x4)
+
+/* ioctls used for setting/getting status of APF on s390x */
+#define KVM_S390_APF_ENABLE 1
+#define KVM_S390_APF_DISABLE 2
+#define KVM_S390_APF_STATUS 3
+#define KVM_S390_APF_DISABLED_NON_PENDING 0
+#define KVM_S390_APF_DISABLED_PENDING 1
+#define KVM_S390_APF_ENABLED_NON_PENDING 2
+#define KVM_S390_APF_ENABLED_PENDING 3
+
#endif
diff --git a/arch/s390/kvm/Kconfig b/arch/s390/kvm/Kconfig
index 70b46ea..4993eed 100644
--- a/arch/s390/kvm/Kconfig
+++ b/arch/s390/kvm/Kconfig
@@ -23,6 +23,8 @@ config KVM
select ANON_INODES
select HAVE_KVM_CPU_RELAX_INTERCEPT
select HAVE_KVM_EVENTFD
+ select KVM_ASYNC_PF
+ select KVM_ASYNC_PF_DIRECT
---help---
Support hosting paravirtualized guest machines using the SIE
virtualization capability on the mainframe. This should work
diff --git a/arch/s390/kvm/Makefile b/arch/s390/kvm/Makefile
index 40b4c64..63bfc28 100644
--- a/arch/s390/kvm/Makefile
+++ b/arch/s390/kvm/Makefile
@@ -7,7 +7,7 @@
# as published by the Free Software Foundation.

KVM := ../../../virt/kvm
-common-objs = $(KVM)/kvm_main.o $(KVM)/eventfd.o
+common-objs = $(KVM)/kvm_main.o $(KVM)/eventfd.o $(KVM)/async_pf.o

ccflags-y := -Ivirt/kvm -Iarch/s390/kvm

diff --git a/arch/s390/kvm/diag.c b/arch/s390/kvm/diag.c
index 3074475..e54dc80 100644
--- a/arch/s390/kvm/diag.c
+++ b/arch/s390/kvm/diag.c
@@ -17,6 +17,7 @@
#include "kvm-s390.h"
#include "trace.h"
#include "trace-s390.h"
+#include "gaccess.h"

static int diag_release_pages(struct kvm_vcpu *vcpu)
{
@@ -46,6 +47,60 @@ static int diag_release_pages(struct kvm_vcpu *vcpu)
return 0;
}

+static int __diag_page_ref_service(struct kvm_vcpu *vcpu)
+{
+ struct prs_parm {
+ u16 code;
+ u16 subcode;
+ u16 parm_len;
+ u16 parm_version;
+ u64 token_addr;
+ u64 select_mask;
+ u64 compare_mask;
+ u64 zarch;
+ };
+ struct prs_parm parm;
+ int rc;
+ u16 rx = (vcpu->arch.sie_block->ipa & 0xf0) >> 4;
+ u16 ry = (vcpu->arch.sie_block->ipa & 0x0f);
+ if (copy_from_guest(vcpu, &parm, vcpu->run->s.regs.gprs[rx], sizeof(parm)))
+ return kvm_s390_inject_program_int(vcpu, PGM_ADDRESSING);
+
+ if (parm.parm_version != 2 || parm.parm_len < 0x5)
+ return kvm_s390_inject_program_int(vcpu, PGM_SPECIFICATION);
+
+ switch (parm.subcode) {
+ case 0: /* TOKEN */
+ if ((parm.zarch >> 63) != 1 || parm.token_addr & 7 ||
+ (parm.compare_mask & parm.select_mask) != parm.compare_mask)
+ return kvm_s390_inject_program_int(vcpu, PGM_SPECIFICATION);
+
+ vcpu->arch.pfault_token = parm.token_addr;
+ vcpu->arch.pfault_select = parm.select_mask;
+ vcpu->arch.pfault_compare = parm.compare_mask;
+ vcpu->run->s.regs.gprs[ry] = 0;
+ rc = 0;
+ break;
+ case 1: /* CANCEL */
+ if (vcpu->run->s.regs.gprs[rx] & 7)
+ return kvm_s390_inject_program_int(vcpu, PGM_ADDRESSING);
+
+ vcpu->run->s.regs.gprs[ry] = 0;
+
+ if (vcpu->arch.pfault_token == KVM_S390_PFAULT_TOKEN_INVALID)
+ vcpu->run->s.regs.gprs[ry] = 1;
+
+ vcpu->arch.pfault_token = KVM_S390_PFAULT_TOKEN_INVALID;
+ rc = 0;
+ break;
+ default:
+ rc = -EOPNOTSUPP;
+ break;
+ }
+
+ return rc;
+}
+
static int __diag_time_slice_end(struct kvm_vcpu *vcpu)
{
VCPU_EVENT(vcpu, 5, "%s", "diag time slice end");
@@ -143,6 +198,8 @@ int kvm_s390_handle_diag(struct kvm_vcpu *vcpu)
return __diag_time_slice_end(vcpu);
case 0x9c:
return __diag_time_slice_end_directed(vcpu);
+ case 0x258:
+ return __diag_page_ref_service(vcpu);
case 0x308:
return __diag_ipl_functions(vcpu);
case 0x500:
diff --git a/arch/s390/kvm/interrupt.c b/arch/s390/kvm/interrupt.c
index 7f35cb3..4f9dd71 100644
--- a/arch/s390/kvm/interrupt.c
+++ b/arch/s390/kvm/interrupt.c
@@ -31,7 +31,7 @@ static int is_ioint(u64 type)
return ((type & 0xfffe0000u) != 0xfffe0000u);
}

-static int psw_extint_disabled(struct kvm_vcpu *vcpu)
+int psw_extint_disabled(struct kvm_vcpu *vcpu)
{
return !(vcpu->arch.sie_block->gpsw.mask & PSW_MASK_EXT);
}
@@ -78,11 +78,8 @@ static int __interrupt_is_deliverable(struct kvm_vcpu *vcpu,
return 1;
return 0;
case KVM_S390_INT_SERVICE:
- if (psw_extint_disabled(vcpu))
- return 0;
- if (vcpu->arch.sie_block->gcr[0] & 0x200ul)
- return 1;
- return 0;
+ case KVM_S390_INT_PFAULT_INIT:
+ case KVM_S390_INT_PFAULT_DONE:
case KVM_S390_INT_VIRTIO:
if (psw_extint_disabled(vcpu))
return 0;
@@ -150,6 +147,8 @@ static void __set_intercept_indicator(struct kvm_vcpu *vcpu,
case KVM_S390_INT_EXTERNAL_CALL:
case KVM_S390_INT_EMERGENCY:
case KVM_S390_INT_SERVICE:
+ case KVM_S390_INT_PFAULT_INIT:
+ case KVM_S390_INT_PFAULT_DONE:
case KVM_S390_INT_VIRTIO:
if (psw_extint_disabled(vcpu))
__set_cpuflag(vcpu, CPUSTAT_EXT_INT);
@@ -223,6 +222,26 @@ static void __do_deliver_interrupt(struct kvm_vcpu *vcpu,
rc |= put_guest(vcpu, inti->ext.ext_params,
(u32 __user *)__LC_EXT_PARAMS);
break;
+ case KVM_S390_INT_PFAULT_INIT:
+ rc = put_guest(vcpu, 0x2603, (u16 __user *) __LC_EXT_INT_CODE);
+ rc |= put_guest(vcpu, 0x0600, (u16 __user *) __LC_EXT_CPU_ADDR);
+ rc |= copy_to_guest(vcpu, __LC_EXT_OLD_PSW,
+ &vcpu->arch.sie_block->gpsw, sizeof(psw_t));
+ rc |= copy_from_guest(vcpu, &vcpu->arch.sie_block->gpsw,
+ __LC_EXT_NEW_PSW, sizeof(psw_t));
+ rc |= put_guest(vcpu, inti->ext.ext_params2,
+ (u64 __user *) __LC_EXT_PARAMS2);
+ break;
+ case KVM_S390_INT_PFAULT_DONE:
+ rc = put_guest(vcpu, 0x2603, (u16 __user *) __LC_EXT_INT_CODE);
+ rc |= put_guest(vcpu, 0x0680, (u16 __user *) __LC_EXT_CPU_ADDR);
+ rc |= copy_to_guest(vcpu, __LC_EXT_OLD_PSW,
+ &vcpu->arch.sie_block->gpsw, sizeof(psw_t));
+ rc |= copy_from_guest(vcpu, &vcpu->arch.sie_block->gpsw,
+ __LC_EXT_NEW_PSW, sizeof(psw_t));
+ rc |= put_guest(vcpu, inti->ext.ext_params2,
+ (u64 __user *) __LC_EXT_PARAMS2);
+ break;
case KVM_S390_INT_VIRTIO:
VCPU_EVENT(vcpu, 4, "interrupt: virtio parm:%x,parm64:%llx",
inti->ext.ext_params, inti->ext.ext_params2);
@@ -357,7 +376,7 @@ static int __try_deliver_ckc_interrupt(struct kvm_vcpu *vcpu)
return 1;
}

-static int kvm_cpu_has_interrupt(struct kvm_vcpu *vcpu)
+int kvm_cpu_has_interrupt(struct kvm_vcpu *vcpu)
{
struct kvm_s390_local_interrupt *li = &vcpu->arch.local_int;
struct kvm_s390_float_interrupt *fi = vcpu->arch.local_int.float_int;
@@ -811,6 +830,11 @@ int kvm_s390_inject_vcpu(struct kvm_vcpu *vcpu,
inti->type = s390int->type;
inti->mchk.mcic = s390int->parm64;
break;
+ case KVM_S390_INT_PFAULT_INIT:
+ case KVM_S390_INT_PFAULT_DONE:
+ inti->type = s390int->type;
+ inti->ext.ext_params2 = s390int->parm64;
+ break;
case KVM_S390_INT_VIRTIO:
case KVM_S390_INT_SERVICE:
case KVM_S390_INT_IO_MIN...KVM_S390_INT_IO_MAX:
diff --git a/arch/s390/kvm/kvm-s390.c b/arch/s390/kvm/kvm-s390.c
index 702daca..004f07d 100644
--- a/arch/s390/kvm/kvm-s390.c
+++ b/arch/s390/kvm/kvm-s390.c
@@ -145,6 +145,7 @@ int kvm_dev_ioctl_check_extension(long ext)
#ifdef CONFIG_KVM_S390_UCONTROL
case KVM_CAP_S390_UCONTROL:
#endif
+ case KVM_CAP_ASYNC_PF:
case KVM_CAP_SYNC_REGS:
case KVM_CAP_ONE_REG:
case KVM_CAP_ENABLE_CAP:
@@ -186,6 +187,33 @@ long kvm_arch_vm_ioctl(struct file *filp,
int r;

switch (ioctl) {
+ case KVM_S390_APF_ENABLE:
+ set_bit(1, &kvm->arch.gmap->pfault_enabled);
+ r = 0;
+ break;
+ case KVM_S390_APF_DISABLE:
+ clear_bit(1, &kvm->arch.gmap->pfault_enabled);
+ r = 0;
+ break;
+ case KVM_S390_APF_STATUS: {
+ bool pfaults_pending = false;
+ unsigned int i;
+ struct kvm_vcpu *vcpu;
+ r = 0;
+ if (test_bit(1, &kvm->arch.gmap->pfault_enabled))
+ r += 2;
+
+ kvm_for_each_vcpu(i, vcpu, kvm) {
+ spin_lock(&vcpu->async_pf.lock);
+ if (vcpu->async_pf.queued > 0)
+ pfaults_pending = true;
+ spin_unlock(&vcpu->async_pf.lock);
+ }
+
+ if (pfaults_pending)
+ r += 1;
+ break;
+ }
case KVM_S390_INTERRUPT: {
struct kvm_s390_interrupt s390int;

@@ -264,6 +292,7 @@ void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
{
VCPU_EVENT(vcpu, 3, "%s", "free cpu");
trace_kvm_s390_destroy_vcpu(vcpu->vcpu_id);
+ kvm_clear_async_pf_completion_queue(vcpu);
if (!kvm_is_ucontrol(vcpu->kvm)) {
clear_bit(63 - vcpu->vcpu_id,
(unsigned long *) &vcpu->kvm->arch.sca->mcn);
@@ -313,6 +342,9 @@ void kvm_arch_destroy_vm(struct kvm *kvm)
/* Section: vcpu related */
int kvm_arch_vcpu_init(struct kvm_vcpu *vcpu)
{
+ vcpu->arch.pfault_token = KVM_S390_PFAULT_TOKEN_INVALID;
+ kvm_clear_async_pf_completion_queue(vcpu);
+ kvm_async_pf_wakeup_all(vcpu);
if (kvm_is_ucontrol(vcpu->kvm)) {
vcpu->arch.gmap = gmap_alloc(current->mm);
if (!vcpu->arch.gmap)
@@ -691,10 +723,75 @@ static void kvm_arch_fault_in_sync(struct kvm_vcpu *vcpu)
up_read(&mm->mmap_sem);
}

+static void __kvm_inject_pfault_token(struct kvm_vcpu *vcpu, bool start_token,
+ unsigned long token)
+{
+ struct kvm_s390_interrupt inti;
+ inti.type = start_token ? KVM_S390_INT_PFAULT_INIT :
+ KVM_S390_INT_PFAULT_DONE;
+ inti.parm64 = token;
+ if (kvm_s390_inject_vcpu(vcpu, &inti))
+ WARN(1, "pfault interrupt injection failed");
+}
+
+void kvm_arch_async_page_not_present(struct kvm_vcpu *vcpu,
+ struct kvm_async_pf *work)
+{
+ __kvm_inject_pfault_token(vcpu, true, work->arch.pfault_token);
+}
+
+void kvm_arch_async_page_present(struct kvm_vcpu *vcpu,
+ struct kvm_async_pf *work)
+{
+ __kvm_inject_pfault_token(vcpu, false, work->arch.pfault_token);
+}
+
+void kvm_arch_async_page_ready(struct kvm_vcpu *vcpu,
+ struct kvm_async_pf *work)
+{
+ /* s390 will always inject the page directly */
+}
+
+bool kvm_arch_can_inject_async_page_present(struct kvm_vcpu *vcpu)
+{
+ /*
+ * s390 will always inject the page directly,
+ * but we still want check_async_completion to cleanup
+ */
+ return true;
+}
+
+static int kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu)
+{
+ hva_t hva = gmap_fault(current->thread.gmap_addr, vcpu->arch.gmap);
+ struct kvm_arch_async_pf arch;
+
+ if (vcpu->arch.pfault_token == KVM_S390_PFAULT_TOKEN_INVALID)
+ return 0;
+ if ((vcpu->arch.sie_block->gpsw.mask & vcpu->arch.pfault_select) !=
+ vcpu->arch.pfault_compare)
+ return 0;
+ if (psw_extint_disabled(vcpu))
+ return 0;
+ if (kvm_cpu_has_interrupt(vcpu))
+ return 0;
+ if (!(vcpu->arch.sie_block->gcr[0] & 0x200ul))
+ return 0;
+
+ if (copy_from_guest(vcpu, &arch.pfault_token, vcpu->arch.pfault_token, 8)) {
+ /* already in error case, insert the interrupt and return 0 */
+ int ign = kvm_s390_inject_program_int(vcpu, PGM_ADDRESSING);
+ return ign - ign;
+ }
+ return kvm_setup_async_pf(vcpu, current->thread.gmap_addr, hva, &arch);
+}
+
static int __vcpu_run(struct kvm_vcpu *vcpu)
{
int rc;

+ kvm_check_async_pf_completion(vcpu);
+
memcpy(&vcpu->arch.sie_block->gg14, &vcpu->run->s.regs.gprs[14], 16);

if (need_resched())
@@ -725,7 +822,8 @@ static int __vcpu_run(struct kvm_vcpu *vcpu)
if (kvm_is_ucontrol(vcpu->kvm)) {
rc = SIE_INTERCEPT_UCONTROL;
} else if (current->thread.gmap_pfault) {
- kvm_arch_fault_in_sync(vcpu);
+ if (!kvm_arch_setup_async_pf(vcpu))
+ kvm_arch_fault_in_sync(vcpu);
current->thread.gmap_pfault = 0;
rc = 0;
} else {
diff --git a/arch/s390/kvm/kvm-s390.h b/arch/s390/kvm/kvm-s390.h
index 028ca9f..d0f4d2a 100644
--- a/arch/s390/kvm/kvm-s390.h
+++ b/arch/s390/kvm/kvm-s390.h
@@ -148,4 +148,8 @@ void exit_sie_sync(struct kvm_vcpu *vcpu);
/* implemented in diag.c */
int kvm_s390_handle_diag(struct kvm_vcpu *vcpu);

+/* implemented in interrupt.c */
+int kvm_cpu_has_interrupt(struct kvm_vcpu *vcpu);
+int psw_extint_disabled(struct kvm_vcpu *vcpu);
+
#endif
diff --git a/arch/s390/kvm/sigp.c b/arch/s390/kvm/sigp.c
index bec398c..768c3f3 100644
--- a/arch/s390/kvm/sigp.c
+++ b/arch/s390/kvm/sigp.c
@@ -187,6 +187,8 @@ static int __sigp_set_arch(struct kvm_vcpu *vcpu, u32 parameter)
{
int rc;

+ vcpu->arch.pfault_token = KVM_S390_PFAULT_TOKEN_INVALID;
+
switch (parameter & 0xff) {
case 0:
rc = SIGP_CC_NOT_OPERATIONAL;
diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h
index acccd08..fae432c 100644
--- a/include/uapi/linux/kvm.h
+++ b/include/uapi/linux/kvm.h
@@ -413,6 +413,8 @@ struct kvm_s390_psw {
#define KVM_S390_PROGRAM_INT 0xfffe0001u
#define KVM_S390_SIGP_SET_PREFIX 0xfffe0002u
#define KVM_S390_RESTART 0xfffe0003u
+#define KVM_S390_INT_PFAULT_INIT 0xfffe0004u
+#define KVM_S390_INT_PFAULT_DONE 0xfffe0005u
#define KVM_S390_MCHK 0xfffe1000u
#define KVM_S390_INT_VIRTIO 0xffff2603u
#define KVM_S390_INT_SERVICE 0xffff2401u
--
1.8.2.2

2013-07-09 13:58:01

[permalink] [raw]

Subject: [PATCH 3/4] PF: Provide additional direct page notification

By setting a Kconfig option, the architecture can control when
guest notifications will be presented by the apf backend.
So there is the default batch mechanism, working as before, where the vcpu thread
should pull in this information. On the other hand there is now the direct
mechanism, this will directly push the information to the guest.

Still the vcpu thread should call check_completion to cleanup leftovers,
that leaves most of the common code untouched.

Signed-off-by: Dominik Dingel <[email protected]>
---
arch/x86/kvm/mmu.c | 2 +-
include/linux/kvm_host.h | 2 +-
virt/kvm/Kconfig | 4 ++++
virt/kvm/async_pf.c | 22 +++++++++++++++++++---
4 files changed, 25 insertions(+), 5 deletions(-)

diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
index 0d094da..b8632e9 100644
--- a/arch/x86/kvm/mmu.c
+++ b/arch/x86/kvm/mmu.c
@@ -3343,7 +3343,7 @@ static int kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn)
arch.direct_map = vcpu->arch.mmu.direct_map;
arch.cr3 = vcpu->arch.mmu.get_cr3(vcpu);

- return kvm_setup_async_pf(vcpu, gva, gfn, &arch);
+ return kvm_setup_async_pf(vcpu, gva, gfn_to_hva(vcpu->kvm, gfn), &arch);
}

static bool can_do_async_pf(struct kvm_vcpu *vcpu)
diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
index f3c04e7..465c639 100644
--- a/include/linux/kvm_host.h
+++ b/include/linux/kvm_host.h
@@ -197,7 +197,7 @@ struct kvm_async_pf {

void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu);
void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu);
-int kvm_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn,
+int kvm_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, unsigned long hva,
struct kvm_arch_async_pf *arch);
int kvm_async_pf_wakeup_all(struct kvm_vcpu *vcpu);
#endif
diff --git a/virt/kvm/Kconfig b/virt/kvm/Kconfig
index 779262f..0774495 100644
--- a/virt/kvm/Kconfig
+++ b/virt/kvm/Kconfig
@@ -22,6 +22,10 @@ config KVM_MMIO
config KVM_ASYNC_PF
bool

+# Toggle to switch between direct notification and batch job
+config KVM_ASYNC_PF_SYNC
+ bool
+
config HAVE_KVM_MSI
bool

diff --git a/virt/kvm/async_pf.c b/virt/kvm/async_pf.c
index ea475cd..cfa9366 100644
--- a/virt/kvm/async_pf.c
+++ b/virt/kvm/async_pf.c
@@ -28,6 +28,21 @@
#include "async_pf.h"
#include <trace/events/kvm.h>

+static inline void kvm_async_page_present_sync(struct kvm_vcpu *vcpu,
+ struct kvm_async_pf *work)
+{
+#ifdef CONFIG_KVM_ASYNC_PF_SYNC
+ kvm_arch_async_page_present(vcpu, work);
+#endif
+}
+static inline void kvm_async_page_present_async(struct kvm_vcpu *vcpu,
+ struct kvm_async_pf *work)
+{
+#ifndef CONFIG_KVM_ASYNC_PF_SYNC
+ kvm_arch_async_page_present(vcpu, work);
+#endif
+}
+
static struct kmem_cache *async_pf_cache;

int kvm_async_pf_init(void)
@@ -70,6 +85,7 @@ static void async_pf_execute(struct work_struct *work)
down_read(&mm->mmap_sem);
get_user_pages(current, mm, addr, 1, 1, 0, &page, NULL);
up_read(&mm->mmap_sem);
+ kvm_async_page_present_sync(vcpu, apf);
unuse_mm(mm);

spin_lock(&vcpu->async_pf.lock);
@@ -134,7 +150,7 @@ void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu)

if (work->page)
kvm_arch_async_page_ready(vcpu, work);
- kvm_arch_async_page_present(vcpu, work);
+ kvm_async_page_present_async(vcpu, work);

list_del(&work->queue);
vcpu->async_pf.queued--;
@@ -144,7 +160,7 @@ void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu)
}
}

-int kvm_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn,
+int kvm_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, unsigned long hva,
struct kvm_arch_async_pf *arch)
{
struct kvm_async_pf *work;
@@ -166,7 +182,7 @@ int kvm_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn,
work->done = false;
work->vcpu = vcpu;
work->gva = gva;
- work->addr = gfn_to_hva(vcpu->kvm, gfn);
+ work->addr = hva;
work->arch = *arch;
work->mm = current->mm;
atomic_inc(&work->mm->mm_count);
--
1.8.2.2

2013-07-09 15:24:12

[permalink] [raw]

Subject: Re: [PATCH 1/4] PF: Add FAULT_FLAG_RETRY_NOWAIT for guest fault

On Tue, Jul 09, 2013 at 03:56:44PM +0200, Dominik Dingel wrote:
> In case of a fault retry exit sie64() with gmap_fault indication for the
> running thread set. This makes it possible to handle async page faults
> without the need for mm notifiers.
>
> Based on a patch from Martin Schwidefsky.
>
For that we will obviously need Christian and Cornelia ACKs. Or it can
go in via S390 tree.

> Signed-off-by: Dominik Dingel <[email protected]>
> ---
> arch/s390/include/asm/pgtable.h | 2 ++
> arch/s390/include/asm/processor.h | 1 +
> arch/s390/kvm/kvm-s390.c | 13 +++++++++++++
> arch/s390/mm/fault.c | 26 ++++++++++++++++++++++----
> 4 files changed, 38 insertions(+), 4 deletions(-)
>
> diff --git a/arch/s390/include/asm/pgtable.h b/arch/s390/include/asm/pgtable.h
> index 0ea4e59..4a4cc64 100644
> --- a/arch/s390/include/asm/pgtable.h
> +++ b/arch/s390/include/asm/pgtable.h
> @@ -740,6 +740,7 @@ static inline void pgste_set_pte(pte_t *ptep, pte_t entry)
> * @table: pointer to the page directory
> * @asce: address space control element for gmap page table
> * @crst_list: list of all crst tables used in the guest address space
> + * @pfault_enabled: defines if pfaults are applicable for the guest
> */
> struct gmap {
> struct list_head list;
> @@ -748,6 +749,7 @@ struct gmap {
> unsigned long asce;
> void *private;
> struct list_head crst_list;
> + unsigned long pfault_enabled;
> };
>
> /**
> diff --git a/arch/s390/include/asm/processor.h b/arch/s390/include/asm/processor.h
> index 6b49987..4fa96ca 100644
> --- a/arch/s390/include/asm/processor.h
> +++ b/arch/s390/include/asm/processor.h
> @@ -77,6 +77,7 @@ struct thread_struct {
> unsigned long ksp; /* kernel stack pointer */
> mm_segment_t mm_segment;
> unsigned long gmap_addr; /* address of last gmap fault. */
> + unsigned int gmap_pfault; /* signal of a pending guest pfault */
> struct per_regs per_user; /* User specified PER registers */
> struct per_event per_event; /* Cause of the last PER trap */
> unsigned long per_flags; /* Flags to control debug behavior */
> diff --git a/arch/s390/kvm/kvm-s390.c b/arch/s390/kvm/kvm-s390.c
> index ba694d2..702daca 100644
> --- a/arch/s390/kvm/kvm-s390.c
> +++ b/arch/s390/kvm/kvm-s390.c
> @@ -682,6 +682,15 @@ static int kvm_s390_handle_requests(struct kvm_vcpu *vcpu)
> return 0;
> }
>
> +static void kvm_arch_fault_in_sync(struct kvm_vcpu *vcpu)
> +{
> + hva_t fault = gmap_fault(current->thread.gmap_addr, vcpu->arch.gmap);
> + struct mm_struct *mm = current->mm;
> + down_read(&mm->mmap_sem);
> + get_user_pages(current, mm, fault, 1, 1, 0, NULL, NULL);
> + up_read(&mm->mmap_sem);
> +}
> +
> static int __vcpu_run(struct kvm_vcpu *vcpu)
> {
> int rc;
> @@ -715,6 +724,10 @@ static int __vcpu_run(struct kvm_vcpu *vcpu)
> if (rc < 0) {
> if (kvm_is_ucontrol(vcpu->kvm)) {
> rc = SIE_INTERCEPT_UCONTROL;
> + } else if (current->thread.gmap_pfault) {
> + kvm_arch_fault_in_sync(vcpu);
> + current->thread.gmap_pfault = 0;
> + rc = 0;
> } else {
> VCPU_EVENT(vcpu, 3, "%s", "fault in sie instruction");
> trace_kvm_s390_sie_fault(vcpu);
> diff --git a/arch/s390/mm/fault.c b/arch/s390/mm/fault.c
> index 047c3e4..7d4c4b1 100644
> --- a/arch/s390/mm/fault.c
> +++ b/arch/s390/mm/fault.c
> @@ -50,6 +50,7 @@
> #define VM_FAULT_BADMAP 0x020000
> #define VM_FAULT_BADACCESS 0x040000
> #define VM_FAULT_SIGNAL 0x080000
> +#define VM_FAULT_PFAULT 0x100000
>
> static unsigned long store_indication __read_mostly;
>
> @@ -232,6 +233,7 @@ static noinline void do_fault_error(struct pt_regs *regs, int fault)
> return;
> }
> case VM_FAULT_BADCONTEXT:
> + case VM_FAULT_PFAULT:
> do_no_context(regs);
> break;
> case VM_FAULT_SIGNAL:
> @@ -269,6 +271,9 @@ static noinline void do_fault_error(struct pt_regs *regs, int fault)
> */
> static inline int do_exception(struct pt_regs *regs, int access)
> {
> +#ifdef CONFIG_PGSTE
> + struct gmap *gmap;
> +#endif
> struct task_struct *tsk;
> struct mm_struct *mm;
> struct vm_area_struct *vma;
> @@ -307,9 +312,10 @@ static inline int do_exception(struct pt_regs *regs, int access)
> down_read(&mm->mmap_sem);
>
> #ifdef CONFIG_PGSTE
> - if ((current->flags & PF_VCPU) && S390_lowcore.gmap) {
> - address = __gmap_fault(address,
> - (struct gmap *) S390_lowcore.gmap);
> + gmap = (struct gmap *)
> + ((current->flags & PF_VCPU) ? S390_lowcore.gmap : 0);
> + if (gmap) {
> + address = __gmap_fault(address, gmap);
> if (address == -EFAULT) {
> fault = VM_FAULT_BADMAP;
> goto out_up;
> @@ -318,6 +324,8 @@ static inline int do_exception(struct pt_regs *regs, int access)
> fault = VM_FAULT_OOM;
> goto out_up;
> }
> + if (test_bit(1, &gmap->pfault_enabled))
> + flags |= FAULT_FLAG_RETRY_NOWAIT;
> }
> #endif
>
> @@ -374,9 +382,19 @@ retry:
> regs, address);
> }
> if (fault & VM_FAULT_RETRY) {
> +#ifdef CONFIG_PGSTE
> + if (gmap && (flags & FAULT_FLAG_RETRY_NOWAIT)) {
> + /* FAULT_FLAG_RETRY_NOWAIT has been set,
> + * mmap_sem has not been released */
> + current->thread.gmap_pfault = 1;
> + fault = VM_FAULT_PFAULT;
> + goto out_up;
> + }
> +#endif
> /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
> * of starvation. */
> - flags &= ~FAULT_FLAG_ALLOW_RETRY;
> + flags &= ~(FAULT_FLAG_ALLOW_RETRY |
> + FAULT_FLAG_RETRY_NOWAIT);
> flags |= FAULT_FLAG_TRIED;
> down_read(&mm->mmap_sem);
> goto retry;
> --
> 1.8.2.2

--
Gleb.

2013-07-09 15:36:57

[permalink] [raw]

Subject: Re: [PATCH 1/4] PF: Add FAULT_FLAG_RETRY_NOWAIT for guest fault

On 09/07/13 17:23, Gleb Natapov wrote:
> On Tue, Jul 09, 2013 at 03:56:44PM +0200, Dominik Dingel wrote:
>> In case of a fault retry exit sie64() with gmap_fault indication for the
>> running thread set. This makes it possible to handle async page faults
>> without the need for mm notifiers.
>>
>> Based on a patch from Martin Schwidefsky.
>>
> For that we will obviously need Christian and Cornelia ACKs. Or it can
> go in via S390 tree.
>

>> Signed-off-by: Dominik Dingel <[email protected]>
Acked-by: Christian Borntraeger <[email protected]>

Do you want me or Conny to apply these patches add a signoff and resend them?
Otherwise I will review the s390 specific patches and ack them individually.

Christian

2013-07-09 15:43:54

[permalink] [raw]

Subject: Re: [PATCH 1/4] PF: Add FAULT_FLAG_RETRY_NOWAIT for guest fault

On Tue, Jul 09, 2013 at 05:36:46PM +0200, Christian Borntraeger wrote:
> On 09/07/13 17:23, Gleb Natapov wrote:
> > On Tue, Jul 09, 2013 at 03:56:44PM +0200, Dominik Dingel wrote:
> >> In case of a fault retry exit sie64() with gmap_fault indication for the
> >> running thread set. This makes it possible to handle async page faults
> >> without the need for mm notifiers.
> >>
> >> Based on a patch from Martin Schwidefsky.
> >>
> > For that we will obviously need Christian and Cornelia ACKs. Or it can
> > go in via S390 tree.
> >
>
> >> Signed-off-by: Dominik Dingel <[email protected]>
> Acked-by: Christian Borntraeger <[email protected]>
>
> Do you want me or Conny to apply these patches add a signoff and resend them?
> Otherwise I will review the s390 specific patches and ack them individually.
>
If your are OK with me merging this one through kvm.git then just
ack/review all others. I have a small comment on patch 2, but otherwise
the patch series looks OK to me from KVM perspective.

--
Gleb.

2013-07-09 15:47:43

[permalink] [raw]

Subject: Re: [PATCH 2/4] PF: Make KVM_HVA_ERR_BAD usable on s390

On Tue, Jul 09, 2013 at 03:56:45PM +0200, Dominik Dingel wrote:
> Current common code uses PAGE_OFFSET to indicate a bad host virtual address.
> As this check won't work on architectures that don't map kernel and user memory
> into the same address space (e.g. s390), an additional implementation is made
> available in the case that PAGE_OFFSET == 0.
>
> Signed-off-by: Dominik Dingel <[email protected]>
> ---
> include/linux/kvm_host.h | 14 ++++++++++++++
> 1 file changed, 14 insertions(+)
>
> diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
> index a63d83e..f3c04e7 100644
> --- a/include/linux/kvm_host.h
> +++ b/include/linux/kvm_host.h
> @@ -85,6 +85,18 @@ static inline bool is_noslot_pfn(pfn_t pfn)
> return pfn == KVM_PFN_NOSLOT;
> }
>
> +#if (PAGE_OFFSET == 0)
> +
Please do it like I described it in the previous review. Lets not
rely on arch low level mmu details like that here.

> +#define KVM_HVA_ERR_BAD (-1UL)
> +#define KVM_HVA_ERR_RO_BAD (-1UL)
> +
> +static inline bool kvm_is_error_hva(unsigned long addr)
> +{
> + return addr == KVM_HVA_ERR_BAD;
> +}
> +
> +#else
> +
> #define KVM_HVA_ERR_BAD (PAGE_OFFSET)
> #define KVM_HVA_ERR_RO_BAD (PAGE_OFFSET + PAGE_SIZE)
>
> @@ -93,6 +105,8 @@ static inline bool kvm_is_error_hva(unsigned long addr)
> return addr >= PAGE_OFFSET;
> }
>
> +#endif
> +
> #define KVM_ERR_PTR_BAD_PAGE (ERR_PTR(-ENOENT))
>
> static inline bool is_error_page(struct page *page)
> --
> 1.8.2.2

--
Gleb.

2013-07-09 16:01:54

[permalink] [raw]

Subject: Re: [PATCH 3/4] PF: Provide additional direct page notification

On 09/07/13 15:56, Dominik Dingel wrote:
> By setting a Kconfig option, the architecture can control when
> guest notifications will be presented by the apf backend.
> So there is the default batch mechanism, working as before, where the vcpu thread
> should pull in this information. On the other hand there is now the direct
> mechanism, this will directly push the information to the guest.
>
> Still the vcpu thread should call check_completion to cleanup leftovers,
> that leaves most of the common code untouched.
>
> Signed-off-by: Dominik Dingel <[email protected]>

Acked-by: Christian Borntraeger <[email protected]>
for the "why". We want to use the existing architectured interface.

2013-07-10 08:18:21

[permalink] [raw]

Subject: Re: [PATCH 4/4] PF: Async page fault support on s390

On 09/07/13 15:56, Dominik Dingel wrote:
> This patch enables async page faults for s390 kvm guests.
> It provides the userspace API to enable, disable or get the status of this
> feature. Also it includes the diagnose code, called by the guest to enable
> async page faults.
>
> The async page faults will use an already existing guest interface for this
> purpose, as described in "CP Programming Services (SC24-6084)".
>
> Signed-off-by: Dominik Dingel <[email protected]>

Re-reading this patch again, found some thing that you should take
care of (nothing major, just small details). Sorry for not seeing them
earlier.

[...]
> + case 1: /* CANCEL */
> + if (vcpu->run->s.regs.gprs[rx] & 7)
> + return kvm_s390_inject_program_int(vcpu, PGM_ADDRESSING);
> +
> + vcpu->run->s.regs.gprs[ry] = 0;
> +
> + if (vcpu->arch.pfault_token == KVM_S390_PFAULT_TOKEN_INVALID)
> + vcpu->run->s.regs.gprs[ry] = 1;
> +
> + vcpu->arch.pfault_token = KVM_S390_PFAULT_TOKEN_INVALID;
> + rc = 0;
> + break;

Dont we need a kvm_clear_async_pf_completion_queue(vcpu) or similar here as well?
The cancel function is supposed to purge all outstanding requests (those were no
completion signal was made pending yet)

[...]
> @@ -264,6 +292,7 @@ void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
> {
> VCPU_EVENT(vcpu, 3, "%s", "free cpu");
> trace_kvm_s390_destroy_vcpu(vcpu->vcpu_id);
> + kvm_clear_async_pf_completion_queue(vcpu);
> if (!kvm_is_ucontrol(vcpu->kvm)) {
> clear_bit(63 - vcpu->vcpu_id,
> (unsigned long *) &vcpu->kvm->arch.sca->mcn);
> @@ -313,6 +342,9 @@ void kvm_arch_destroy_vm(struct kvm *kvm)
> /* Section: vcpu related */
> int kvm_arch_vcpu_init(struct kvm_vcpu *vcpu)
> {
> + vcpu->arch.pfault_token = KVM_S390_PFAULT_TOKEN_INVALID;
> + kvm_clear_async_pf_completion_queue(vcpu);
> + kvm_async_pf_wakeup_all(vcpu);
> if (kvm_is_ucontrol(vcpu->kvm)) {
> vcpu->arch.gmap = gmap_alloc(current->mm);
> if (!vcpu->arch.gmap)

We should also reset pfault handling for CPU reset, no?

> @@ -691,10 +723,75 @@ static void kvm_arch_fault_in_sync(struct kvm_vcpu *vcpu)
> up_read(&mm->mmap_sem);
> }
>
> +static void __kvm_inject_pfault_token(struct kvm_vcpu *vcpu, bool start_token,
> + unsigned long token)
> +{
> + struct kvm_s390_interrupt inti;
> + inti.type = start_token ? KVM_S390_INT_PFAULT_INIT :
> + KVM_S390_INT_PFAULT_DONE;
> + inti.parm64 = token;
> + if (kvm_s390_inject_vcpu(vcpu, &inti))
> + WARN(1, "pfault interrupt injection failed");
> +}

The PFAULT_DONE is architectured as a floating interrupt (can happen
on other CPUs).

[...]
> --- a/arch/s390/kvm/sigp.c
> +++ b/arch/s390/kvm/sigp.c
> @@ -187,6 +187,8 @@ static int __sigp_set_arch(struct kvm_vcpu *vcpu, u32 parameter)
> {
> int rc;
>
> + vcpu->arch.pfault_token = KVM_S390_PFAULT_TOKEN_INVALID;
> +

sigp set architecture affects all cpus, so we must reset pfault via
kvm_for_each_vcpu, I guess.

Otherwise patch looks good. I guess only one more iteration
Christian

2013-07-10 10:39:09

[permalink] [raw]

Subject: Re: [PATCH 3/4] PF: Provide additional direct page notification

On 09.07.2013, at 18:01, Christian Borntraeger wrote:

> On 09/07/13 15:56, Dominik Dingel wrote:
>> By setting a Kconfig option, the architecture can control when
>> guest notifications will be presented by the apf backend.
>> So there is the default batch mechanism, working as before, where the vcpu thread
>> should pull in this information. On the other hand there is now the direct
>> mechanism, this will directly push the information to the guest.
>>
>> Still the vcpu thread should call check_completion to cleanup leftovers,
>> that leaves most of the common code untouched.
>>
>> Signed-off-by: Dominik Dingel <[email protected]>
>
> Acked-by: Christian Borntraeger <[email protected]>
> for the "why". We want to use the existing architectured interface.

Shouldn't this be a runtime option?

Alex

2013-07-10 10:43:59

[permalink] [raw]

Subject: Re: [PATCH 3/4] PF: Provide additional direct page notification

On Wed, Jul 10, 2013 at 12:39:01PM +0200, Alexander Graf wrote:
>
> On 09.07.2013, at 18:01, Christian Borntraeger wrote:
>
> > On 09/07/13 15:56, Dominik Dingel wrote:
> >> By setting a Kconfig option, the architecture can control when
> >> guest notifications will be presented by the apf backend.
> >> So there is the default batch mechanism, working as before, where the vcpu thread
> >> should pull in this information. On the other hand there is now the direct
> >> mechanism, this will directly push the information to the guest.
> >>
> >> Still the vcpu thread should call check_completion to cleanup leftovers,
> >> that leaves most of the common code untouched.
> >>
> >> Signed-off-by: Dominik Dingel <[email protected]>
> >
> > Acked-by: Christian Borntraeger <[email protected]>
> > for the "why". We want to use the existing architectured interface.
>
> Shouldn't this be a runtime option?
>
Why? What is the advantage of using sync delivery when HW can do it
async?

--
Gleb.

2013-07-10 10:46:08

[permalink] [raw]

Subject: Re: [PATCH 3/4] PF: Provide additional direct page notification

On 10.07.2013, at 12:42, Gleb Natapov wrote:

> On Wed, Jul 10, 2013 at 12:39:01PM +0200, Alexander Graf wrote:
>>
>> On 09.07.2013, at 18:01, Christian Borntraeger wrote:
>>
>>> On 09/07/13 15:56, Dominik Dingel wrote:
>>>> By setting a Kconfig option, the architecture can control when
>>>> guest notifications will be presented by the apf backend.
>>>> So there is the default batch mechanism, working as before, where the vcpu thread
>>>> should pull in this information. On the other hand there is now the direct
>>>> mechanism, this will directly push the information to the guest.
>>>>
>>>> Still the vcpu thread should call check_completion to cleanup leftovers,
>>>> that leaves most of the common code untouched.
>>>>
>>>> Signed-off-by: Dominik Dingel <[email protected]>
>>>
>>> Acked-by: Christian Borntraeger <[email protected]>
>>> for the "why". We want to use the existing architectured interface.
>>
>> Shouldn't this be a runtime option?
>>
> Why? What is the advantage of using sync delivery when HW can do it
> async?

What's the advantage of having an option at all then? Who selects it?

Alex

2013-07-10 10:49:27

[permalink] [raw]

Subject: Re: [PATCH 3/4] PF: Provide additional direct page notification

On 10/07/13 12:39, Alexander Graf wrote:
>
> On 09.07.2013, at 18:01, Christian Borntraeger wrote:
>
>> On 09/07/13 15:56, Dominik Dingel wrote:
>>> By setting a Kconfig option, the architecture can control when
>>> guest notifications will be presented by the apf backend.
>>> So there is the default batch mechanism, working as before, where the vcpu thread
>>> should pull in this information. On the other hand there is now the direct
>>> mechanism, this will directly push the information to the guest.
>>>
>>> Still the vcpu thread should call check_completion to cleanup leftovers,
>>> that leaves most of the common code untouched.
>>>
>>> Signed-off-by: Dominik Dingel <[email protected]>
>>
>> Acked-by: Christian Borntraeger <[email protected]>
>> for the "why". We want to use the existing architectured interface.
>
> Shouldn't this be a runtime option?

This is an a) or b) depending on the architecture. So making this a kconfig
option is the most sane approach no?

Christian

2013-07-10 10:50:43

[permalink] [raw]

Subject: Re: [PATCH 3/4] PF: Provide additional direct page notification

On Wed, Jul 10, 2013 at 12:45:59PM +0200, Alexander Graf wrote:
>
> On 10.07.2013, at 12:42, Gleb Natapov wrote:
>
> > On Wed, Jul 10, 2013 at 12:39:01PM +0200, Alexander Graf wrote:
> >>
> >> On 09.07.2013, at 18:01, Christian Borntraeger wrote:
> >>
> >>> On 09/07/13 15:56, Dominik Dingel wrote:
> >>>> By setting a Kconfig option, the architecture can control when
> >>>> guest notifications will be presented by the apf backend.
> >>>> So there is the default batch mechanism, working as before, where the vcpu thread
> >>>> should pull in this information. On the other hand there is now the direct
> >>>> mechanism, this will directly push the information to the guest.
> >>>>
> >>>> Still the vcpu thread should call check_completion to cleanup leftovers,
> >>>> that leaves most of the common code untouched.
> >>>>
> >>>> Signed-off-by: Dominik Dingel <[email protected]>
> >>>
> >>> Acked-by: Christian Borntraeger <[email protected]>
> >>> for the "why". We want to use the existing architectured interface.
> >>
> >> Shouldn't this be a runtime option?
> >>
> > Why? What is the advantage of using sync delivery when HW can do it
> > async?
>
> What's the advantage of having an option at all then? Who selects it?
>
x86 is stupid and cannot deliver the even asynchronously. Platform that
can do it select the option.

--
Gleb.

2013-07-10 10:51:25

[permalink] [raw]

Subject: Re: [PATCH 3/4] PF: Provide additional direct page notification

On 10.07.2013, at 12:49, Christian Borntraeger wrote:

> On 10/07/13 12:39, Alexander Graf wrote:
>>
>> On 09.07.2013, at 18:01, Christian Borntraeger wrote:
>>
>>> On 09/07/13 15:56, Dominik Dingel wrote:
>>>> By setting a Kconfig option, the architecture can control when
>>>> guest notifications will be presented by the apf backend.
>>>> So there is the default batch mechanism, working as before, where the vcpu thread
>>>> should pull in this information. On the other hand there is now the direct
>>>> mechanism, this will directly push the information to the guest.
>>>>
>>>> Still the vcpu thread should call check_completion to cleanup leftovers,
>>>> that leaves most of the common code untouched.
>>>>
>>>> Signed-off-by: Dominik Dingel <[email protected]>
>>>
>>> Acked-by: Christian Borntraeger <[email protected]>
>>> for the "why". We want to use the existing architectured interface.
>>
>> Shouldn't this be a runtime option?
>
> This is an a) or b) depending on the architecture. So making this a kconfig
> option is the most sane approach no?

I guess I'm just missing the patch that actually selects it. Last thing I remember you can have a kernel configured for s390x that runs on any 64bit capable system out there. What would you select? If that kernel runs on newer hardware, it would be able to do async pf, no?

There's a good chance I simply miss a critical component here :).

Alex

2013-07-10 10:52:09