[resend due to bad alias expansion resulting in some recipients
being bogus]
Some hardware and software systems maintain page tables outside the normal
Linux page tables, which reference userspace memory. This includes
Infiniband, other RDMA-capable devices, and kvm (with a pending patch).
Because these systems maintain external page tables (and external tlbs),
Linux cannot demand page this memory and it must be locked. For kvm at
least, this is a significant reduction in functionality.
This sample patch adds a new mechanism, pte notifiers, that allows drivers
to register an interest in a changes to ptes. Whenever Linux changes a
pte, it will call a notifier to allow the driver to adjust the external
page table and flush its tlb.
Note that only one notifier is implemented, ->clear(), but others should be
similar.
pte notifiers are different from paravirt_ops: they extend the normal
page tables rather than replace them; and they provide high-level information
such as the vma and the virtual address for the driver to use.
Signed-off-by: Avi Kivity <[email protected]>
diff --git a/include/linux/mm.h b/include/linux/mm.h
index 655094d..5d2bbee 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -14,6 +14,7 @@
#include <linux/debug_locks.h>
#include <linux/backing-dev.h>
#include <linux/mm_types.h>
+#include <linux/pte_notifier.h>
struct mempolicy;
struct anon_vma;
@@ -108,6 +109,9 @@ struct vm_area_struct {
#ifndef CONFIG_MMU
atomic_t vm_usage; /* refcount (VMAs shared if !MMU) */
#endif
+#ifdef CONFIG_PTE_NOTIFIERS
+ struct list_head pte_notifier_list;
+#endif
#ifdef CONFIG_NUMA
struct mempolicy *vm_policy; /* NUMA policy for the VMA */
#endif
diff --git a/include/linux/pte_notifier.h b/include/linux/pte_notifier.h
new file mode 100644
index 0000000..d28832b
--- /dev/null
+++ b/include/linux/pte_notifier.h
@@ -0,0 +1,52 @@
+#ifndef _LINUX_PTE_NOTIFIER_H
+#define _LINUX_PTE_NOTIFIER_H
+
+#include <linux/list.h>
+
+struct vm_area_struct;
+
+#ifdef CONFIG_PTE_NOTIFIERS
+
+struct pte_notifier;
+
+struct pte_notifier_ops {
+ void (*close)(struct pte_notifier *pn, struct vm_area_struct *vma);
+ void (*clear)(struct pte_notifier *pn, struct vm_area_struct *vma,
+ unsigned long address);
+};
+
+struct pte_notifier {
+ struct list_head link;
+ const struct pte_notifier_ops *ops;
+};
+
+
+void vma_init_pte_notifiers(struct vm_area_struct *vma);
+void vma_close_pte_notifiers(struct vm_area_struct *vma);
+void pte_notifier_register(struct pte_notifier *pn,
+ struct vm_area_struct *vma);
+void pte_notifier_unregister(struct pte_notifier *pn);
+
+#define pte_notifier_call(vma, function, args...) \
+ do { \
+ struct pte_notifier *__pn; \
+ \
+ list_for_each_entry(__pn, &vma->pte_notifier_list, link) \
+ __pn->ops->function(__pn, vma, args); \
+ } while (0)
+
+#else
+
+static inline void vma_init_pte_notifiers(struct vm_area_struct *vma) {}
+static inline void vma_close_pte_notifiers(struct vm_area_struct *vma) {}
+static inline void pte_notifier_register(struct pte_notifier *pn,
+ struct vm_area_struct *vma) {}
+static inline void pte_notifier_unregister(struct pte_notifier *pn) {}
+
+#define pte_notifier_call(vma, function, args...) \
+ do { } while (0)
+
+#endif
+
+
+#endif
diff --git a/mm/Kconfig b/mm/Kconfig
index e24d348..7b10151 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -176,3 +176,6 @@ config NR_QUICK
config VIRT_TO_BUS
def_bool y
depends on !ARCH_NO_VIRT_TO_BUS
+
+config PTE_NOTIFIERS
+ bool
diff --git a/mm/Makefile b/mm/Makefile
index 245e33a..59f6a03 100644
--- a/mm/Makefile
+++ b/mm/Makefile
@@ -29,4 +29,5 @@ obj-$(CONFIG_FS_XIP) += filemap_xip.o
obj-$(CONFIG_MIGRATION) += migrate.o
obj-$(CONFIG_SMP) += allocpercpu.o
obj-$(CONFIG_QUICKLIST) += quicklist.o
+obj-$(CONFIG_PTE_NOTIFIERS) += pte_notifiers.o
diff --git a/mm/mmap.c b/mm/mmap.c
index b653721..cc6c4fe 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -1134,6 +1134,7 @@ munmap_back:
vma->vm_page_prot = protection_map[vm_flags &
(VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
vma->vm_pgoff = pgoff;
+ vma_init_pte_notifiers(vma);
if (file) {
error = -EINVAL;
diff --git a/mm/pte_notifier.c b/mm/pte_notifier.c
new file mode 100644
index 0000000..0b9076c
--- /dev/null
+++ b/mm/pte_notifier.c
@@ -0,0 +1,32 @@
+
+#include <linux/pte_notifier.h>
+
+void vma_init_pte_notifiers(struct vm_area_struct *vma)
+{
+ INIT_LIST_HEAD(&vma->pte_notifier_list);
+}
+EXPORT_SYMBOL_GPL(vma_init_pte_notifiers);
+
+void vma_destroy_pte_notifiers(struct vm_area_struct *vma)
+{
+ struct pte_notifier *pn;
+ struct list_head *n;
+
+ list_for_each_entry_safe(pn, n, &vma->pte_notifier_list, link) {
+ pn->ops->close(__pn, vma);
+ __list_del(n);
+ }
+}
+
+void pte_notifier_register(struct pte_notifier *pn, struct vm_area_struct *vma)
+{
+ list_add(&pn->link, &vma->pte_notifier_list);
+}
+EXPORT_SYMBOL_GPL(pte_notifier_register);
+
+void pte_notifier_unregister(struct pte_notifier *pn)
+{
+ list_del(&pn->link);
+}
+EXPORT_SYMBOL_GPL(pte_notifier_unregister);
+
diff --git a/mm/rmap.c b/mm/rmap.c
index 41ac397..3f61d38 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -682,6 +682,7 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
}
/* Nuke the page table entry. */
+ pte_notifier_call(vma, clear, address);
flush_cache_page(vma, address, page_to_pfn(page));
pteval = ptep_clear_flush(vma, address, pte);
On Wed, 2007-09-05 at 22:32 +0300, Avi Kivity wrote:
> [resend due to bad alias expansion resulting in some recipients
> being bogus]
>
> Some hardware and software systems maintain page tables outside the normal
> Linux page tables, which reference userspace memory. This includes
> Infiniband, other RDMA-capable devices, and kvm (with a pending patch).
And lguest. I can't tell until I've actually implemented it, but I
think it will seriously reduce the need for page pinning which is why
only root can currently launch guests.
My concern is locking: this is called with the page lock held, and I
guess we have to bump the guest out if it's currently running.
(Oh, and this means lguest needs to do a reverse mapping somehow, but
I'll come up with something).
Cheers,
Rusty.
Rusty Russell wrote:
> On Wed, 2007-09-05 at 22:32 +0300, Avi Kivity wrote:
>
>> [resend due to bad alias expansion resulting in some recipients
>> being bogus]
>>
>> Some hardware and software systems maintain page tables outside the normal
>> Linux page tables, which reference userspace memory. This includes
>> Infiniband, other RDMA-capable devices, and kvm (with a pending patch).
>>
>
> And lguest. I can't tell until I've actually implemented it, but I
> think it will seriously reduce the need for page pinning which is why
> only root can currently launch guests.
>
>
Ah yes, lguest.
> My concern is locking: this is called with the page lock held, and I
> guess we have to bump the guest out if it's currently running.
>
This will complicate kvm's locking too. We usually take kvm->lock to do
mmu ops, but that is now a mutex.
--
Any sufficiently difficult bug is indistinguishable from a feature.
On Wed, 2007-09-05 at 22:32 +0300, Avi Kivity wrote:
> [resend due to bad alias expansion resulting in some recipients
> being bogus]
>
> Some hardware and software systems maintain page tables outside the normal
> Linux page tables, which reference userspace memory. This includes
> Infiniband, other RDMA-capable devices, and kvm (with a pending patch).
>
> Because these systems maintain external page tables (and external tlbs),
> Linux cannot demand page this memory and it must be locked. For kvm at
> least, this is a significant reduction in functionality.
>
> This sample patch adds a new mechanism, pte notifiers, that allows drivers
> to register an interest in a changes to ptes. Whenever Linux changes a
> pte, it will call a notifier to allow the driver to adjust the external
> page table and flush its tlb.
>
> Note that only one notifier is implemented, ->clear(), but others should be
> similar.
>
> pte notifiers are different from paravirt_ops: they extend the normal
> page tables rather than replace them; and they provide high-level
> information
> such as the vma and the virtual address for the driver to use.
Looks great. So for kvm, all guest pages will be vma mapped?
There are lock issues in kvm between kvm lock and page lock.
Will shadow page table be still stored in page->private? If yes, the
page->private must be cleaned before add_to_swap.
Thanks,
Shaohua
Shaohua Li wrote:
> On Wed, 2007-09-05 at 22:32 +0300, Avi Kivity wrote:
>
>> [resend due to bad alias expansion resulting in some recipients
>> being bogus]
>>
>> Some hardware and software systems maintain page tables outside the normal
>> Linux page tables, which reference userspace memory. This includes
>> Infiniband, other RDMA-capable devices, and kvm (with a pending patch).
>>
>> Because these systems maintain external page tables (and external tlbs),
>> Linux cannot demand page this memory and it must be locked. For kvm at
>> least, this is a significant reduction in functionality.
>>
>> This sample patch adds a new mechanism, pte notifiers, that allows drivers
>> to register an interest in a changes to ptes. Whenever Linux changes a
>> pte, it will call a notifier to allow the driver to adjust the external
>> page table and flush its tlb.
>>
>> Note that only one notifier is implemented, ->clear(), but others should be
>> similar.
>>
>> pte notifiers are different from paravirt_ops: they extend the normal
>> page tables rather than replace them; and they provide high-level
>> information
>> such as the vma and the virtual address for the driver to use.
>>
> Looks great. So for kvm, all guest pages will be vma mapped?
> There are lock issues in kvm between kvm lock and page lock.
>
Yes, locking will be a headache.
> Will shadow page table be still stored in page->private? If yes, the
> page->private must be cleaned before add_to_swap.
>
page->private can be in use by filesystems, so we will need to move rmap
somewhere else.
--
Any sufficiently difficult bug is indistinguishable from a feature.
Avi Kivity <avi-atKUWr5tajBWk0Htik3J/[email protected]> writes:
>
> pte notifiers are different from paravirt_ops: they extend the normal
> page tables rather than replace them; and they provide high-level information
> such as the vma and the virtual address for the driver to use.
Sounds like a locking horror to me. To do anything with page tables
you need locks. Both for the kernel page tables and for your new tables.
What happens when people add all
things of complicated operations in these notifiers? That will likely
happen and then everytime you change something in VM code they
will break. This has the potential to increase the cost of maintaining
VM code considerably, which would be a bad thing.
This is quite different from paravirt ops because low level pvops
can typically run lockless by just doing some kind of hypercall directly.
But that won't work for maintaining your custom page tables.
-Andi
[ugh, what happened to the cc-list?]
Andi Kleen wrote:
> Avi Kivity <avi-atKUWr5tajBWk0Htik3J/[email protected]> writes:
>
>> pte notifiers are different from paravirt_ops: they extend the normal
>> page tables rather than replace them; and they provide high-level information
>> such as the vma and the virtual address for the driver to use.
>>
>
> Sounds like a locking horror to me. To do anything with page tables
> you need locks. Both for the kernel page tables and for your new tables.
>
> What happens when people add all
> things of complicated operations in these notifiers? That will likely
> happen and then everytime you change something in VM code they
> will break. This has the potential to increase the cost of maintaining
> VM code considerably, which would be a bad thing.
>
> This is quite different from paravirt ops because low level pvops
> can typically run lockless by just doing some kind of hypercall directly.
> But that won't work for maintaining your custom page tables.
>
This is a real problem. I don't have a solution yet.
Obviously that needs to be addressed before something like this can go
in; but as it's been done for the quadrics driver, presumably it is doable.
--
Any sufficiently difficult bug is indistinguishable from a feature.
Avi Kivity <avi-atKUWr5tajBWk0Htik3J/[email protected]> writes:
>
> pte notifiers are different from paravirt_ops: they extend the normal
> page tables rather than replace them; and they provide high-level information
> such as the vma and the virtual address for the driver to use.
Sounds like a locking horror to me. To do anything with page tables
you need locks. Both for the kernel page tables and for your new tables.
What happens when people add all
things of complicated operations in these notifiers? That will likely
happen and then everytime you change something in VM code they
will break. This has the potential to increase the cost of maintaining
VM code considerably, which would be a bad thing.
This is quite different from paravirt ops because low level pvops
can typically run lockless by just doing some kind of hypercall directly.
But that won't work for maintaining your custom page tables.
-Andi
Andi Kleen wrote:
> Avi Kivity <avi-atKUWr5tajBWk0Htik3J/[email protected]> writes:
>
>> pte notifiers are different from paravirt_ops: they extend the normal
>> page tables rather than replace them; and they provide high-level information
>> such as the vma and the virtual address for the driver to use.
>>
>
> Sounds like a locking horror to me. To do anything with page tables
> you need locks. Both for the kernel page tables and for your new tables.
>
> What happens when people add all
> things of complicated operations in these notifiers? That will likely
> happen and then everytime you change something in VM code they
> will break. This has the potential to increase the cost of maintaining
> VM code considerably, which would be a bad thing.
>
> This is quite different from paravirt ops because low level pvops
> can typically run lockless by just doing some kind of hypercall directly.
> But that won't work for maintaining your custom page tables.
>
Okay, here's a possible fix: add ->lock() and ->unlock() callbacks, to
be called when mmap_sem is taken either for read or write. Also add a
->release() for when the mm goes away to avoid the need to care about
the entire data structure going away.
The notifier list would need to be kept sorted to avoid deadlocks.
--
Any sufficiently difficult bug is indistinguishable from a feature.