[PATCH 8/8] KVM: PPC: Add hugepage support for IOMMU in-kernel handling

[Alexander Graf]

On 10.07.2013, at 01:29, Alexey Kardashevskiy wrote:

> On 07/10/2013 03:32 AM, Alexander Graf wrote:
>> On 07/06/2013 05:07 PM, Alexey Kardashevskiy wrote:
>>> This adds special support for huge pages (16MB).  The reference
>>> counting cannot be easily done for such pages in real mode (when
>>> MMU is off) so we added a list of huge pages.  It is populated in
>>> virtual mode and get_page is called just once per a huge page.
>>> Real mode handlers check if the requested page is huge and in the list,
>>> then no reference counting is done, otherwise an exit to virtual mode
>>> happens.  The list is released at KVM exit.  At the moment the fastest
>>> card available for tests uses up to 9 huge pages so walking through this
>>> list is not very expensive.  However this can change and we may want
>>> to optimize this.
>>> 
>>> Signed-off-by: Paul Mackerras<paulus at samba.org>
>>> Signed-off-by: Alexey Kardashevskiy<aik at ozlabs.ru>
>>> 
>>> ---
>>> 
>>> Changes:
>>> 2013/06/27:
>>> * list of huge pages replaces with hashtable for better performance
>> 
>> So the only thing your patch description really talks about is not true
>> anymore?
>> 
>>> * spinlock removed from real mode and only protects insertion of new
>>> huge [ages descriptors into the hashtable
>>> 
>>> 2013/06/05:
>>> * fixed compile error when CONFIG_IOMMU_API=n
>>> 
>>> 2013/05/20:
>>> * the real mode handler now searches for a huge page by gpa (used to be pte)
>>> * the virtual mode handler prints warning if it is called twice for the same
>>> huge page as the real mode handler is expected to fail just once - when a
>>> huge
>>> page is not in the list yet.
>>> * the huge page is refcounted twice - when added to the hugepage list and
>>> when used in the virtual mode hcall handler (can be optimized but it will
>>> make the patch less nice).
>>> 
>>> Signed-off-by: Alexey Kardashevskiy<aik at ozlabs.ru>
>>> ---
>>>  arch/powerpc/include/asm/kvm_host.h |  25 +++++++++
>>>  arch/powerpc/kernel/iommu.c         |   6 ++-
>>>  arch/powerpc/kvm/book3s_64_vio.c    | 104
>>> +++++++++++++++++++++++++++++++++---
>>>  arch/powerpc/kvm/book3s_64_vio_hv.c |  21 ++++++--
>>>  4 files changed, 146 insertions(+), 10 deletions(-)
>>> 
>>> diff --git a/arch/powerpc/include/asm/kvm_host.h
>>> b/arch/powerpc/include/asm/kvm_host.h
>>> index 53e61b2..a7508cf 100644
>>> --- a/arch/powerpc/include/asm/kvm_host.h
>>> +++ b/arch/powerpc/include/asm/kvm_host.h
>>> @@ -30,6 +30,7 @@
>>>  #include<linux/kvm_para.h>
>>>  #include<linux/list.h>
>>>  #include<linux/atomic.h>
>>> +#include<linux/hashtable.h>
>>>  #include<asm/kvm_asm.h>
>>>  #include<asm/processor.h>
>>>  #include<asm/page.h>
>>> @@ -182,10 +183,34 @@ struct kvmppc_spapr_tce_table {
>>>      u32 window_size;
>>>      struct iommu_group *grp;        /* used for IOMMU groups */
>>>      struct vfio_group *vfio_grp;        /* used for IOMMU groups */
>>> +    DECLARE_HASHTABLE(hash_tab, ilog2(64));    /* used for IOMMU groups */
>>> +    spinlock_t hugepages_write_lock;    /* used for IOMMU groups */
>>>      struct { struct { unsigned long put, indir, stuff; } rm, vm; } stat;
>>>      struct page *pages[0];
>>>  };
>>> 
>>> +/*
>>> + * The KVM guest can be backed with 16MB pages.
>>> + * In this case, we cannot do page counting from the real mode
>>> + * as the compound pages are used - they are linked in a list
>>> + * with pointers as virtual addresses which are inaccessible
>>> + * in real mode.
>>> + *
>>> + * The code below keeps a 16MB pages list and uses page struct
>>> + * in real mode if it is already locked in RAM and inserted into
>>> + * the list or switches to the virtual mode where it can be
>>> + * handled in a usual manner.
>>> + */
>>> +#define KVMPPC_SPAPR_HUGEPAGE_HASH(gpa)    hash_32(gpa>>  24, 32)
>>> +
>>> +struct kvmppc_spapr_iommu_hugepage {
>>> +    struct hlist_node hash_node;
>>> +    unsigned long gpa;    /* Guest physical address */
>>> +    unsigned long hpa;    /* Host physical address */
>>> +    struct page *page;    /* page struct of the very first subpage */
>>> +    unsigned long size;    /* Huge page size (always 16MB at the moment) */
>>> +};
>>> +
>>>  struct kvmppc_linear_info {
>>>      void        *base_virt;
>>>      unsigned long     base_pfn;
>>> diff --git a/arch/powerpc/kernel/iommu.c b/arch/powerpc/kernel/iommu.c
>>> index 51678ec..e0b6eca 100644
>>> --- a/arch/powerpc/kernel/iommu.c
>>> +++ b/arch/powerpc/kernel/iommu.c
>>> @@ -999,7 +999,8 @@ int iommu_free_tces(struct iommu_table *tbl, unsigned
>>> long entry,
>>>              if (!pg) {
>>>                  ret = -EAGAIN;
>>>              } else if (PageCompound(pg)) {
>>> -                ret = -EAGAIN;
>>> +                /* Hugepages will be released at KVM exit */
>>> +                ret = 0;
>>>              } else {
>>>                  if (oldtce&  TCE_PCI_WRITE)
>>>                      SetPageDirty(pg);
>>> @@ -1009,6 +1010,9 @@ int iommu_free_tces(struct iommu_table *tbl,
>>> unsigned long entry,
>>>              struct page *pg = pfn_to_page(oldtce>>  PAGE_SHIFT);
>>>              if (!pg) {
>>>                  ret = -EAGAIN;
>>> +            } else if (PageCompound(pg)) {
>>> +                /* Hugepages will be released at KVM exit */
>>> +                ret = 0;
>>>              } else {
>>>                  if (oldtce&  TCE_PCI_WRITE)
>>>                      SetPageDirty(pg);
>>> diff --git a/arch/powerpc/kvm/book3s_64_vio.c
>>> b/arch/powerpc/kvm/book3s_64_vio.c
>>> index 2b51f4a..c037219 100644
>>> --- a/arch/powerpc/kvm/book3s_64_vio.c
>>> +++ b/arch/powerpc/kvm/book3s_64_vio.c
>>> @@ -46,6 +46,40 @@
>>> 
>>>  #define ERROR_ADDR      ((void *)~(unsigned long)0x0)
>>> 
>>> +#ifdef CONFIG_IOMMU_API
>> 
>> Can't you just make CONFIG_IOMMU_API mandatory in Kconfig?
> 
> 
> Sure I can. I can do anything. Why should I?

To get rid of #ifdef's. They make code hard to maintain.

> Do I have to do that to get
> this accepted? I do not understand this comment. It has already been
> discussed how to enable this option.
> 
> 
>>> +static void kvmppc_iommu_hugepages_init(struct kvmppc_spapr_tce_table *tt)
>>> +{
>>> +    spin_lock_init(&tt->hugepages_write_lock);
>>> +    hash_init(tt->hash_tab);
>>> +}
>>> +
>>> +static void kvmppc_iommu_hugepages_cleanup(struct kvmppc_spapr_tce_table
>>> *tt)
>>> +{
>>> +    int bkt;
>>> +    struct kvmppc_spapr_iommu_hugepage *hp;
>>> +    struct hlist_node *tmp;
>>> +
>>> +    spin_lock(&tt->hugepages_write_lock);
>>> +    hash_for_each_safe(tt->hash_tab, bkt, tmp, hp, hash_node) {
>>> +        pr_debug("Release HP liobn=%llx #%u gpa=%lx hpa=%lx size=%ld\n",
>>> +                tt->liobn, bkt, hp->gpa, hp->hpa, hp->size);
>> 
>> trace point
>> 
>>> +        hlist_del_rcu(&hp->hash_node);
>>> +
>>> +        put_page(hp->page);
>> 
>> Don't you have to mark them dirty?
> 
> 
> get_user_pages_fast is called with writing==1. Does not it do the same?

It's not exactly obvious that you're calling it with writing == 1 :). Can you create a new local variable "is_write" in the calling function, set that to 1 before the call to get_user_pages_fast and pass it in instead of the 1? The compiler should easily optimize all of that away, but it makes the code by far easier to read.

> 
>> 
>>> +        kfree(hp);
>>> +    }
>>> +    spin_unlock(&tt->hugepages_write_lock);
>>> +}
>>> +#else
>>> +static void kvmppc_iommu_hugepages_init(struct kvmppc_spapr_tce_table *tt)
>>> +{
>>> +}
>>> +
>>> +static void kvmppc_iommu_hugepages_cleanup(struct kvmppc_spapr_tce_table
>>> *tt)
>>> +{
>>> +}
>>> +#endif /* CONFIG_IOMMU_API */
>>> +
>>>  static long kvmppc_stt_npages(unsigned long window_size)
>>>  {
>>>      return ALIGN((window_size>>  SPAPR_TCE_SHIFT)
>>> @@ -112,6 +146,7 @@ static void release_spapr_tce_table(struct
>>> kvmppc_spapr_tce_table *stt)
>>> 
>>>      mutex_lock(&kvm->lock);
>>>      list_del(&stt->list);
>>> +    kvmppc_iommu_hugepages_cleanup(stt);
>>> 
>>>  #ifdef CONFIG_IOMMU_API
>>>      if (stt->grp) {
>>> @@ -200,6 +235,7 @@ long kvm_vm_ioctl_create_spapr_tce(struct kvm *kvm,
>>>      kvm_get_kvm(kvm);
>>> 
>>>      mutex_lock(&kvm->lock);
>>> +    kvmppc_iommu_hugepages_init(stt);
>>>      list_add(&stt->list,&kvm->arch.spapr_tce_tables);
>>> 
>>>      mutex_unlock(&kvm->lock);
>>> @@ -283,6 +319,7 @@ long kvm_vm_ioctl_create_spapr_tce_iommu(struct kvm
>>> *kvm,
>>> 
>>>      kvm_get_kvm(kvm);
>>>      mutex_lock(&kvm->lock);
>>> +    kvmppc_iommu_hugepages_init(tt);
>>>      list_add(&tt->list,&kvm->arch.spapr_tce_tables);
>>>      mutex_unlock(&kvm->lock);
>>> 
>>> @@ -307,10 +344,17 @@ long kvm_vm_ioctl_create_spapr_tce_iommu(struct kvm
>>> *kvm,
>>> 
>>>  /* Converts guest physical address to host virtual address */
>>>  static void __user *kvmppc_vm_gpa_to_hva_and_get(struct kvm_vcpu *vcpu,
>>> +        struct kvmppc_spapr_tce_table *tt,
>>>          unsigned long gpa, struct page **pg, unsigned long *hpa)
>>>  {
>>>      unsigned long hva, gfn = gpa>>  PAGE_SHIFT;
>>>      struct kvm_memory_slot *memslot;
>>> +#ifdef CONFIG_IOMMU_API
>>> +    struct kvmppc_spapr_iommu_hugepage *hp;
>>> +    unsigned key = KVMPPC_SPAPR_HUGEPAGE_HASH(gpa);
>>> +    pte_t *ptep;
>>> +    unsigned int shift = 0;
>>> +#endif
>>> 
>>>      memslot = search_memslots(kvm_memslots(vcpu->kvm), gfn);
>>>      if (!memslot)
>>> @@ -325,6 +369,54 @@ static void __user
>>> *kvmppc_vm_gpa_to_hva_and_get(struct kvm_vcpu *vcpu,
>>>          *hpa = __pa((unsigned long) page_address(*pg)) +
>>>                  (hva&  ~PAGE_MASK);
>>> 
>>> +#ifdef CONFIG_IOMMU_API
>> 
>> This function is becoming incredibly large. Please split it up. Also please
>> document the code.
> 
> 
> Less than 100 lines is incredibly large? There are _many_ functions bigger
> than that. I do not really see the point in making a separate function
> which is going to be called only once.

Anything above 20 lines is too big usually, with very few exceptions. As mnemonic you can always imagine Linus sitting there with an 80x25 UNIX terminal, reading your code ;).


Alex