| Message ID | 20180626055926.27703-3-aik@ozlabs.ru (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | KVM: PPC: Check if IOMMU page is contained in the pinned physical page | expand |
On Tue, Jun 26, 2018 at 03:59:26PM +1000, Alexey Kardashevskiy wrote: > We already have a check in drivers/vfio/vfio_iommu_spapr_tce.c that > an IOMMU page is contained in the physical page so the PCI hardware won't > get access to unassigned host memory. > > However we do not have this check in KVM fastpath (H_PUT_TCE accelerated > code) so the user space can pin memory backed with 64k pages and create > a hardware TCE table with a bigger page size. We were lucky so far and > did not hit this yet as the very first time the mapping happens > we do not have tbl::it_userspace allocated yet and fall back to > the userspace which in turn calls VFIO IOMMU driver and that fails > because of the check in vfio_iommu_spapr_tce.c which is really > sustainable solution. > > This stores the smallest preregistered page size in the preregistered > region descriptor and changes the mm_iommu_xxx API to check this against > the IOMMU page size. > > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> > --- > Changes: > v2: > * explicitly check for compound pages before calling compound_order() > > --- > The bug is: run QEMU _without_ hugepages (no -mempath) and tell it to > advertise 16MB pages to the guest; a typical pseries guest will use 16MB > for IOMMU pages without checking the mmu pagesize and this will fail > at https://git.qemu.org/?p=qemu.git;a=blob;f=hw/vfio/common.c;h=fb396cf00ac40eb35967a04c9cc798ca896eed57;hb=refs/heads/master#l256 > > With the change, mapping will fail in KVM and the guest will print: > > mlx5_core 0000:00:00.0: ibm,create-pe-dma-window(2027) 0 8000000 20000000 18 1f returned 0 (liobn = 0x80000001 starting addr = 8000000 0) > mlx5_core 0000:00:00.0: created tce table LIOBN 0x80000001 for /pci@800000020000000/ethernet@0 > mlx5_core 0000:00:00.0: failed to map direct window for > /pci@800000020000000/ethernet@0: -1 [snip] > @@ -124,7 +125,7 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > struct mm_iommu_table_group_mem_t **pmem) > { > struct mm_iommu_table_group_mem_t *mem; > - long i, j, ret = 0, locked_entries = 0; > + long i, j, ret = 0, locked_entries = 0, pageshift; > struct page *page = NULL; > > mutex_lock(&mem_list_mutex); > @@ -166,6 +167,8 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > goto unlock_exit; > } > > + mem->pageshift = 30; /* start from 1G pages - the biggest we have */ What about 16G pages on an HPT system? > for (i = 0; i < entries; ++i) { > if (1 != get_user_pages_fast(ua + (i << PAGE_SHIFT), > 1/* pages */, 1/* iswrite */, &page)) { > @@ -199,6 +202,11 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > } > } > populate: > + pageshift = PAGE_SHIFT; > + if (PageCompound(page)) > + pageshift += compound_order(compound_head(page)); > + mem->pageshift = min_t(unsigned int, mem->pageshift, pageshift); Why not make mem->pageshift and pageshift local the same type to avoid the min_t() ? > + > mem->hpas[i] = page_to_pfn(page) << PAGE_SHIFT; > } > > @@ -349,7 +357,7 @@ struct mm_iommu_table_group_mem_t *mm_iommu_find(struct mm_struct *mm, > EXPORT_SYMBOL_GPL(mm_iommu_find); > > long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem, > - unsigned long ua, unsigned long *hpa) > + unsigned long ua, unsigned int pageshift, unsigned long *hpa) > { > const long entry = (ua - mem->ua) >> PAGE_SHIFT; > u64 *va = &mem->hpas[entry]; > @@ -357,6 +365,9 @@ long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem, > if (entry >= mem->entries) > return -EFAULT; > > + if (pageshift > mem->pageshift) > + return -EFAULT; > + > *hpa = *va | (ua & ~PAGE_MASK); > > return 0; > @@ -364,7 +375,7 @@ long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem, > EXPORT_SYMBOL_GPL(mm_iommu_ua_to_hpa); > > long mm_iommu_ua_to_hpa_rm(struct mm_iommu_table_group_mem_t *mem, > - unsigned long ua, unsigned long *hpa) > + unsigned long ua, unsigned int pageshift, unsigned long *hpa) > { > const long entry = (ua - mem->ua) >> PAGE_SHIFT; > void *va = &mem->hpas[entry]; > @@ -373,6 +384,9 @@ long mm_iommu_ua_to_hpa_rm(struct mm_iommu_table_group_mem_t *mem, > if (entry >= mem->entries) > return -EFAULT; > > + if (pageshift > mem->pageshift) > + return -EFAULT; > + > pa = (void *) vmalloc_to_phys(va); > if (!pa) > return -EFAULT; > diff --git a/drivers/vfio/vfio_iommu_spapr_tce.c b/drivers/vfio/vfio_iommu_spapr_tce.c > index 2da5f05..7cd63b0 100644 > --- a/drivers/vfio/vfio_iommu_spapr_tce.c > +++ b/drivers/vfio/vfio_iommu_spapr_tce.c > @@ -467,7 +467,7 @@ static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container, > if (!mem) > return -EINVAL; > > - ret = mm_iommu_ua_to_hpa(mem, tce, phpa); > + ret = mm_iommu_ua_to_hpa(mem, tce, shift, phpa); > if (ret) > return -EINVAL; >
On Fri, 29 Jun 2018 14:12:41 +1000 David Gibson <david@gibson.dropbear.id.au> wrote: > On Tue, Jun 26, 2018 at 03:59:26PM +1000, Alexey Kardashevskiy wrote: > > We already have a check in drivers/vfio/vfio_iommu_spapr_tce.c that > > an IOMMU page is contained in the physical page so the PCI hardware won't > > get access to unassigned host memory. > > > > However we do not have this check in KVM fastpath (H_PUT_TCE accelerated > > code) so the user space can pin memory backed with 64k pages and create > > a hardware TCE table with a bigger page size. We were lucky so far and > > did not hit this yet as the very first time the mapping happens > > we do not have tbl::it_userspace allocated yet and fall back to > > the userspace which in turn calls VFIO IOMMU driver and that fails > > because of the check in vfio_iommu_spapr_tce.c which is really > > sustainable solution. > > > > This stores the smallest preregistered page size in the preregistered > > region descriptor and changes the mm_iommu_xxx API to check this against > > the IOMMU page size. > > > > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> > > --- > > Changes: > > v2: > > * explicitly check for compound pages before calling compound_order() > > > > --- > > The bug is: run QEMU _without_ hugepages (no -mempath) and tell it to > > advertise 16MB pages to the guest; a typical pseries guest will use 16MB > > for IOMMU pages without checking the mmu pagesize and this will fail > > at https://git.qemu.org/?p=qemu.git;a=blob;f=hw/vfio/common.c;h=fb396cf00ac40eb35967a04c9cc798ca896eed57;hb=refs/heads/master#l256 > > > > With the change, mapping will fail in KVM and the guest will print: > > > > mlx5_core 0000:00:00.0: ibm,create-pe-dma-window(2027) 0 8000000 20000000 18 1f returned 0 (liobn = 0x80000001 starting addr = 8000000 0) > > mlx5_core 0000:00:00.0: created tce table LIOBN 0x80000001 for /pci@800000020000000/ethernet@0 > > mlx5_core 0000:00:00.0: failed to map direct window for > > /pci@800000020000000/ethernet@0: -1 > > [snip] > > @@ -124,7 +125,7 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > struct mm_iommu_table_group_mem_t **pmem) > > { > > struct mm_iommu_table_group_mem_t *mem; > > - long i, j, ret = 0, locked_entries = 0; > > + long i, j, ret = 0, locked_entries = 0, pageshift; > > struct page *page = NULL; > > > > mutex_lock(&mem_list_mutex); > > @@ -166,6 +167,8 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > goto unlock_exit; > > } > > > > + mem->pageshift = 30; /* start from 1G pages - the biggest we have */ > > What about 16G pages on an HPT system? Below in the loop mem->pageshift will reduce to the biggest actual size which will be 16mb/64k/4k. Or remain 1GB if no memory is actually pinned, no loss there. > > > for (i = 0; i < entries; ++i) { > > if (1 != get_user_pages_fast(ua + (i << PAGE_SHIFT), > > 1/* pages */, 1/* iswrite */, &page)) { > > @@ -199,6 +202,11 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > } > > } > > populate: > > + pageshift = PAGE_SHIFT; > > + if (PageCompound(page)) > > + pageshift += compound_order(compound_head(page)); > > + mem->pageshift = min_t(unsigned int, mem->pageshift, pageshift); > > Why not make mem->pageshift and pageshift local the same type to avoid > the min_t() ? I was under impression min() is deprecated (misinterpret checkpatch.pl may be) and therefore did not pay attention to it. I can fix this and repost if there is no other question. > > > + > > mem->hpas[i] = page_to_pfn(page) << PAGE_SHIFT; > > } > > > > @@ -349,7 +357,7 @@ struct mm_iommu_table_group_mem_t *mm_iommu_find(struct mm_struct *mm, > > EXPORT_SYMBOL_GPL(mm_iommu_find); > > > > long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem, > > - unsigned long ua, unsigned long *hpa) > > + unsigned long ua, unsigned int pageshift, unsigned long *hpa) > > { > > const long entry = (ua - mem->ua) >> PAGE_SHIFT; > > u64 *va = &mem->hpas[entry]; > > @@ -357,6 +365,9 @@ long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem, > > if (entry >= mem->entries) > > return -EFAULT; > > > > + if (pageshift > mem->pageshift) > > + return -EFAULT; > > + > > *hpa = *va | (ua & ~PAGE_MASK); > > > > return 0; > > @@ -364,7 +375,7 @@ long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem, > > EXPORT_SYMBOL_GPL(mm_iommu_ua_to_hpa); > > > > long mm_iommu_ua_to_hpa_rm(struct mm_iommu_table_group_mem_t *mem, > > - unsigned long ua, unsigned long *hpa) > > + unsigned long ua, unsigned int pageshift, unsigned long *hpa) > > { > > const long entry = (ua - mem->ua) >> PAGE_SHIFT; > > void *va = &mem->hpas[entry]; > > @@ -373,6 +384,9 @@ long mm_iommu_ua_to_hpa_rm(struct mm_iommu_table_group_mem_t *mem, > > if (entry >= mem->entries) > > return -EFAULT; > > > > + if (pageshift > mem->pageshift) > > + return -EFAULT; > > + > > pa = (void *) vmalloc_to_phys(va); > > if (!pa) > > return -EFAULT; > > diff --git a/drivers/vfio/vfio_iommu_spapr_tce.c b/drivers/vfio/vfio_iommu_spapr_tce.c > > index 2da5f05..7cd63b0 100644 > > --- a/drivers/vfio/vfio_iommu_spapr_tce.c > > +++ b/drivers/vfio/vfio_iommu_spapr_tce.c > > @@ -467,7 +467,7 @@ static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container, > > if (!mem) > > return -EINVAL; > > > > - ret = mm_iommu_ua_to_hpa(mem, tce, phpa); > > + ret = mm_iommu_ua_to_hpa(mem, tce, shift, phpa); > > if (ret) > > return -EINVAL; > > -- Alexey
On Fri, Jun 29, 2018 at 02:51:21PM +1000, Alexey Kardashevskiy wrote: > On Fri, 29 Jun 2018 14:12:41 +1000 > David Gibson <david@gibson.dropbear.id.au> wrote: > > > On Tue, Jun 26, 2018 at 03:59:26PM +1000, Alexey Kardashevskiy wrote: > > > We already have a check in drivers/vfio/vfio_iommu_spapr_tce.c that > > > an IOMMU page is contained in the physical page so the PCI hardware won't > > > get access to unassigned host memory. > > > > > > However we do not have this check in KVM fastpath (H_PUT_TCE accelerated > > > code) so the user space can pin memory backed with 64k pages and create > > > a hardware TCE table with a bigger page size. We were lucky so far and > > > did not hit this yet as the very first time the mapping happens > > > we do not have tbl::it_userspace allocated yet and fall back to > > > the userspace which in turn calls VFIO IOMMU driver and that fails > > > because of the check in vfio_iommu_spapr_tce.c which is really > > > sustainable solution. > > > > > > This stores the smallest preregistered page size in the preregistered > > > region descriptor and changes the mm_iommu_xxx API to check this against > > > the IOMMU page size. > > > > > > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> > > > --- > > > Changes: > > > v2: > > > * explicitly check for compound pages before calling compound_order() > > > > > > --- > > > The bug is: run QEMU _without_ hugepages (no -mempath) and tell it to > > > advertise 16MB pages to the guest; a typical pseries guest will use 16MB > > > for IOMMU pages without checking the mmu pagesize and this will fail > > > at https://git.qemu.org/?p=qemu.git;a=blob;f=hw/vfio/common.c;h=fb396cf00ac40eb35967a04c9cc798ca896eed57;hb=refs/heads/master#l256 > > > > > > With the change, mapping will fail in KVM and the guest will print: > > > > > > mlx5_core 0000:00:00.0: ibm,create-pe-dma-window(2027) 0 8000000 20000000 18 1f returned 0 (liobn = 0x80000001 starting addr = 8000000 0) > > > mlx5_core 0000:00:00.0: created tce table LIOBN 0x80000001 for /pci@800000020000000/ethernet@0 > > > mlx5_core 0000:00:00.0: failed to map direct window for > > > /pci@800000020000000/ethernet@0: -1 > > > > [snip] > > > @@ -124,7 +125,7 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > struct mm_iommu_table_group_mem_t **pmem) > > > { > > > struct mm_iommu_table_group_mem_t *mem; > > > - long i, j, ret = 0, locked_entries = 0; > > > + long i, j, ret = 0, locked_entries = 0, pageshift; > > > struct page *page = NULL; > > > > > > mutex_lock(&mem_list_mutex); > > > @@ -166,6 +167,8 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > goto unlock_exit; > > > } > > > > > > + mem->pageshift = 30; /* start from 1G pages - the biggest we have */ > > > > What about 16G pages on an HPT system? > > > Below in the loop mem->pageshift will reduce to the biggest actual size > which will be 16mb/64k/4k. Or remain 1GB if no memory is actually > pinned, no loss there. Are you saying that 16G IOMMU pages aren't supported? Or that there's some reason a guest can never use them? > > > for (i = 0; i < entries; ++i) { > > > if (1 != get_user_pages_fast(ua + (i << PAGE_SHIFT), > > > 1/* pages */, 1/* iswrite */, &page)) { > > > @@ -199,6 +202,11 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > } > > > } > > > populate: > > > + pageshift = PAGE_SHIFT; > > > + if (PageCompound(page)) > > > + pageshift += compound_order(compound_head(page)); > > > + mem->pageshift = min_t(unsigned int, mem->pageshift, pageshift); > > > > Why not make mem->pageshift and pageshift local the same type to avoid > > the min_t() ? > > I was under impression min() is deprecated (misinterpret checkpatch.pl > may be) and therefore did not pay attention to it. I can fix this and > repost if there is no other question. Hm, it's possible.
On Fri, 29 Jun 2018 14:57:02 +1000 David Gibson <david@gibson.dropbear.id.au> wrote: > On Fri, Jun 29, 2018 at 02:51:21PM +1000, Alexey Kardashevskiy wrote: > > On Fri, 29 Jun 2018 14:12:41 +1000 > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > On Tue, Jun 26, 2018 at 03:59:26PM +1000, Alexey Kardashevskiy wrote: > > > > We already have a check in drivers/vfio/vfio_iommu_spapr_tce.c that > > > > an IOMMU page is contained in the physical page so the PCI hardware won't > > > > get access to unassigned host memory. > > > > > > > > However we do not have this check in KVM fastpath (H_PUT_TCE accelerated > > > > code) so the user space can pin memory backed with 64k pages and create > > > > a hardware TCE table with a bigger page size. We were lucky so far and > > > > did not hit this yet as the very first time the mapping happens > > > > we do not have tbl::it_userspace allocated yet and fall back to > > > > the userspace which in turn calls VFIO IOMMU driver and that fails > > > > because of the check in vfio_iommu_spapr_tce.c which is really > > > > sustainable solution. > > > > > > > > This stores the smallest preregistered page size in the preregistered > > > > region descriptor and changes the mm_iommu_xxx API to check this against > > > > the IOMMU page size. > > > > > > > > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> > > > > --- > > > > Changes: > > > > v2: > > > > * explicitly check for compound pages before calling compound_order() > > > > > > > > --- > > > > The bug is: run QEMU _without_ hugepages (no -mempath) and tell it to > > > > advertise 16MB pages to the guest; a typical pseries guest will use 16MB > > > > for IOMMU pages without checking the mmu pagesize and this will fail > > > > at https://git.qemu.org/?p=qemu.git;a=blob;f=hw/vfio/common.c;h=fb396cf00ac40eb35967a04c9cc798ca896eed57;hb=refs/heads/master#l256 > > > > > > > > With the change, mapping will fail in KVM and the guest will print: > > > > > > > > mlx5_core 0000:00:00.0: ibm,create-pe-dma-window(2027) 0 8000000 20000000 18 1f returned 0 (liobn = 0x80000001 starting addr = 8000000 0) > > > > mlx5_core 0000:00:00.0: created tce table LIOBN 0x80000001 for /pci@800000020000000/ethernet@0 > > > > mlx5_core 0000:00:00.0: failed to map direct window for > > > > /pci@800000020000000/ethernet@0: -1 > > > > > > [snip] > > > > @@ -124,7 +125,7 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > struct mm_iommu_table_group_mem_t **pmem) > > > > { > > > > struct mm_iommu_table_group_mem_t *mem; > > > > - long i, j, ret = 0, locked_entries = 0; > > > > + long i, j, ret = 0, locked_entries = 0, pageshift; > > > > struct page *page = NULL; > > > > > > > > mutex_lock(&mem_list_mutex); > > > > @@ -166,6 +167,8 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > goto unlock_exit; > > > > } > > > > > > > > + mem->pageshift = 30; /* start from 1G pages - the biggest we have */ > > > > > > What about 16G pages on an HPT system? > > > > > > Below in the loop mem->pageshift will reduce to the biggest actual size > > which will be 16mb/64k/4k. Or remain 1GB if no memory is actually > > pinned, no loss there. > > Are you saying that 16G IOMMU pages aren't supported? Or that there's > some reason a guest can never use them? ah, 16_G_, not _M_. My bad. I just never tried such huge pages, I will lift the limit up to 64 then, easier this way. > > > > > for (i = 0; i < entries; ++i) { > > > > if (1 != get_user_pages_fast(ua + (i << PAGE_SHIFT), > > > > 1/* pages */, 1/* iswrite */, &page)) { > > > > @@ -199,6 +202,11 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > } > > > > } > > > > populate: > > > > + pageshift = PAGE_SHIFT; > > > > + if (PageCompound(page)) > > > > + pageshift += compound_order(compound_head(page)); > > > > + mem->pageshift = min_t(unsigned int, mem->pageshift, pageshift); > > > > > > Why not make mem->pageshift and pageshift local the same type to avoid > > > the min_t() ? > > > > I was under impression min() is deprecated (misinterpret checkpatch.pl > > may be) and therefore did not pay attention to it. I can fix this and > > repost if there is no other question. > > Hm, it's possible. Nah, tried min(), compiles fine. -- Alexey
On Fri, 29 Jun 2018 15:18:20 +1000 Alexey Kardashevskiy <aik@ozlabs.ru> wrote: > On Fri, 29 Jun 2018 14:57:02 +1000 > David Gibson <david@gibson.dropbear.id.au> wrote: > > > On Fri, Jun 29, 2018 at 02:51:21PM +1000, Alexey Kardashevskiy wrote: > > > On Fri, 29 Jun 2018 14:12:41 +1000 > > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > > > On Tue, Jun 26, 2018 at 03:59:26PM +1000, Alexey Kardashevskiy wrote: > > > > > We already have a check in drivers/vfio/vfio_iommu_spapr_tce.c that > > > > > an IOMMU page is contained in the physical page so the PCI hardware won't > > > > > get access to unassigned host memory. > > > > > > > > > > However we do not have this check in KVM fastpath (H_PUT_TCE accelerated > > > > > code) so the user space can pin memory backed with 64k pages and create > > > > > a hardware TCE table with a bigger page size. We were lucky so far and > > > > > did not hit this yet as the very first time the mapping happens > > > > > we do not have tbl::it_userspace allocated yet and fall back to > > > > > the userspace which in turn calls VFIO IOMMU driver and that fails > > > > > because of the check in vfio_iommu_spapr_tce.c which is really > > > > > sustainable solution. > > > > > > > > > > This stores the smallest preregistered page size in the preregistered > > > > > region descriptor and changes the mm_iommu_xxx API to check this against > > > > > the IOMMU page size. > > > > > > > > > > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> > > > > > --- > > > > > Changes: > > > > > v2: > > > > > * explicitly check for compound pages before calling compound_order() > > > > > > > > > > --- > > > > > The bug is: run QEMU _without_ hugepages (no -mempath) and tell it to > > > > > advertise 16MB pages to the guest; a typical pseries guest will use 16MB > > > > > for IOMMU pages without checking the mmu pagesize and this will fail > > > > > at https://git.qemu.org/?p=qemu.git;a=blob;f=hw/vfio/common.c;h=fb396cf00ac40eb35967a04c9cc798ca896eed57;hb=refs/heads/master#l256 > > > > > > > > > > With the change, mapping will fail in KVM and the guest will print: > > > > > > > > > > mlx5_core 0000:00:00.0: ibm,create-pe-dma-window(2027) 0 8000000 20000000 18 1f returned 0 (liobn = 0x80000001 starting addr = 8000000 0) > > > > > mlx5_core 0000:00:00.0: created tce table LIOBN 0x80000001 for /pci@800000020000000/ethernet@0 > > > > > mlx5_core 0000:00:00.0: failed to map direct window for > > > > > /pci@800000020000000/ethernet@0: -1 > > > > > > > > [snip] > > > > > @@ -124,7 +125,7 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > struct mm_iommu_table_group_mem_t **pmem) > > > > > { > > > > > struct mm_iommu_table_group_mem_t *mem; > > > > > - long i, j, ret = 0, locked_entries = 0; > > > > > + long i, j, ret = 0, locked_entries = 0, pageshift; > > > > > struct page *page = NULL; > > > > > > > > > > mutex_lock(&mem_list_mutex); > > > > > @@ -166,6 +167,8 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > goto unlock_exit; > > > > > } > > > > > > > > > > + mem->pageshift = 30; /* start from 1G pages - the biggest we have */ > > > > > > > > What about 16G pages on an HPT system? > > > > > > > > > Below in the loop mem->pageshift will reduce to the biggest actual size > > > which will be 16mb/64k/4k. Or remain 1GB if no memory is actually > > > pinned, no loss there. > > > > Are you saying that 16G IOMMU pages aren't supported? Or that there's > > some reason a guest can never use them? > > > ah, 16_G_, not _M_. My bad. I just never tried such huge pages, I will > lift the limit up to 64 then, easier this way. Ah, no, rather this as the upper limit: mem->pageshift = ilog2(entries) + PAGE_SHIFT; @entries here is a number of system pages being pinned in that function. > > > > > > > > for (i = 0; i < entries; ++i) { > > > > > if (1 != get_user_pages_fast(ua + (i << PAGE_SHIFT), > > > > > 1/* pages */, 1/* iswrite */, &page)) { > > > > > @@ -199,6 +202,11 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > } > > > > > } > > > > > populate: > > > > > + pageshift = PAGE_SHIFT; > > > > > + if (PageCompound(page)) > > > > > + pageshift += compound_order(compound_head(page)); > > > > > + mem->pageshift = min_t(unsigned int, mem->pageshift, pageshift); > > > > > > > > Why not make mem->pageshift and pageshift local the same type to avoid > > > > the min_t() ? > > > > > > I was under impression min() is deprecated (misinterpret checkpatch.pl > > > may be) and therefore did not pay attention to it. I can fix this and > > > repost if there is no other question. > > > > Hm, it's possible. > > Nah, tried min(), compiles fine. -- Alexey
On Fri, Jun 29, 2018 at 05:07:47PM +1000, Alexey Kardashevskiy wrote: > On Fri, 29 Jun 2018 15:18:20 +1000 > Alexey Kardashevskiy <aik@ozlabs.ru> wrote: > > > On Fri, 29 Jun 2018 14:57:02 +1000 > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > On Fri, Jun 29, 2018 at 02:51:21PM +1000, Alexey Kardashevskiy wrote: > > > > On Fri, 29 Jun 2018 14:12:41 +1000 > > > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > > > > > On Tue, Jun 26, 2018 at 03:59:26PM +1000, Alexey Kardashevskiy wrote: > > > > > > We already have a check in drivers/vfio/vfio_iommu_spapr_tce.c that > > > > > > an IOMMU page is contained in the physical page so the PCI hardware won't > > > > > > get access to unassigned host memory. > > > > > > > > > > > > However we do not have this check in KVM fastpath (H_PUT_TCE accelerated > > > > > > code) so the user space can pin memory backed with 64k pages and create > > > > > > a hardware TCE table with a bigger page size. We were lucky so far and > > > > > > did not hit this yet as the very first time the mapping happens > > > > > > we do not have tbl::it_userspace allocated yet and fall back to > > > > > > the userspace which in turn calls VFIO IOMMU driver and that fails > > > > > > because of the check in vfio_iommu_spapr_tce.c which is really > > > > > > sustainable solution. > > > > > > > > > > > > This stores the smallest preregistered page size in the preregistered > > > > > > region descriptor and changes the mm_iommu_xxx API to check this against > > > > > > the IOMMU page size. > > > > > > > > > > > > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> > > > > > > --- > > > > > > Changes: > > > > > > v2: > > > > > > * explicitly check for compound pages before calling compound_order() > > > > > > > > > > > > --- > > > > > > The bug is: run QEMU _without_ hugepages (no -mempath) and tell it to > > > > > > advertise 16MB pages to the guest; a typical pseries guest will use 16MB > > > > > > for IOMMU pages without checking the mmu pagesize and this will fail > > > > > > at https://git.qemu.org/?p=qemu.git;a=blob;f=hw/vfio/common.c;h=fb396cf00ac40eb35967a04c9cc798ca896eed57;hb=refs/heads/master#l256 > > > > > > > > > > > > With the change, mapping will fail in KVM and the guest will print: > > > > > > > > > > > > mlx5_core 0000:00:00.0: ibm,create-pe-dma-window(2027) 0 8000000 20000000 18 1f returned 0 (liobn = 0x80000001 starting addr = 8000000 0) > > > > > > mlx5_core 0000:00:00.0: created tce table LIOBN 0x80000001 for /pci@800000020000000/ethernet@0 > > > > > > mlx5_core 0000:00:00.0: failed to map direct window for > > > > > > /pci@800000020000000/ethernet@0: -1 > > > > > > > > > > [snip] > > > > > > @@ -124,7 +125,7 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > struct mm_iommu_table_group_mem_t **pmem) > > > > > > { > > > > > > struct mm_iommu_table_group_mem_t *mem; > > > > > > - long i, j, ret = 0, locked_entries = 0; > > > > > > + long i, j, ret = 0, locked_entries = 0, pageshift; > > > > > > struct page *page = NULL; > > > > > > > > > > > > mutex_lock(&mem_list_mutex); > > > > > > @@ -166,6 +167,8 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > goto unlock_exit; > > > > > > } > > > > > > > > > > > > + mem->pageshift = 30; /* start from 1G pages - the biggest we have */ > > > > > > > > > > What about 16G pages on an HPT system? > > > > > > > > > > > > Below in the loop mem->pageshift will reduce to the biggest actual size > > > > which will be 16mb/64k/4k. Or remain 1GB if no memory is actually > > > > pinned, no loss there. > > > > > > Are you saying that 16G IOMMU pages aren't supported? Or that there's > > > some reason a guest can never use them? > > > > > > ah, 16_G_, not _M_. My bad. I just never tried such huge pages, I will > > lift the limit up to 64 then, easier this way. > > > Ah, no, rather this as the upper limit: > > mem->pageshift = ilog2(entries) + PAGE_SHIFT; I can't make sense of this comment in context. I see how you're computing the minimum page size in the reserved region. My question is about what the "maximum minimum" is - the starting value from which you calculate. Currently it's 1G, but I can't immediately see a reason that 16G is impossible here. > @entries here is a number of system pages being pinned in that > function. > > > > > > > > > > > > > > for (i = 0; i < entries; ++i) { > > > > > > if (1 != get_user_pages_fast(ua + (i << PAGE_SHIFT), > > > > > > 1/* pages */, 1/* iswrite */, &page)) { > > > > > > @@ -199,6 +202,11 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > } > > > > > > } > > > > > > populate: > > > > > > + pageshift = PAGE_SHIFT; > > > > > > + if (PageCompound(page)) > > > > > > + pageshift += compound_order(compound_head(page)); > > > > > > + mem->pageshift = min_t(unsigned int, mem->pageshift, pageshift); > > > > > > > > > > Why not make mem->pageshift and pageshift local the same type to avoid > > > > > the min_t() ? > > > > > > > > I was under impression min() is deprecated (misinterpret checkpatch.pl > > > > may be) and therefore did not pay attention to it. I can fix this and > > > > repost if there is no other question. > > > > > > Hm, it's possible. > > > > Nah, tried min(), compiles fine. > > >
On Mon, 2 Jul 2018 14:08:52 +1000 David Gibson <david@gibson.dropbear.id.au> wrote: > On Fri, Jun 29, 2018 at 05:07:47PM +1000, Alexey Kardashevskiy wrote: > > On Fri, 29 Jun 2018 15:18:20 +1000 > > Alexey Kardashevskiy <aik@ozlabs.ru> wrote: > > > > > On Fri, 29 Jun 2018 14:57:02 +1000 > > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > > > On Fri, Jun 29, 2018 at 02:51:21PM +1000, Alexey Kardashevskiy wrote: > > > > > On Fri, 29 Jun 2018 14:12:41 +1000 > > > > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > > > > > > > On Tue, Jun 26, 2018 at 03:59:26PM +1000, Alexey Kardashevskiy wrote: > > > > > > > We already have a check in drivers/vfio/vfio_iommu_spapr_tce.c that > > > > > > > an IOMMU page is contained in the physical page so the PCI hardware won't > > > > > > > get access to unassigned host memory. > > > > > > > > > > > > > > However we do not have this check in KVM fastpath (H_PUT_TCE accelerated > > > > > > > code) so the user space can pin memory backed with 64k pages and create > > > > > > > a hardware TCE table with a bigger page size. We were lucky so far and > > > > > > > did not hit this yet as the very first time the mapping happens > > > > > > > we do not have tbl::it_userspace allocated yet and fall back to > > > > > > > the userspace which in turn calls VFIO IOMMU driver and that fails > > > > > > > because of the check in vfio_iommu_spapr_tce.c which is really > > > > > > > sustainable solution. > > > > > > > > > > > > > > This stores the smallest preregistered page size in the preregistered > > > > > > > region descriptor and changes the mm_iommu_xxx API to check this against > > > > > > > the IOMMU page size. > > > > > > > > > > > > > > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> > > > > > > > --- > > > > > > > Changes: > > > > > > > v2: > > > > > > > * explicitly check for compound pages before calling compound_order() > > > > > > > > > > > > > > --- > > > > > > > The bug is: run QEMU _without_ hugepages (no -mempath) and tell it to > > > > > > > advertise 16MB pages to the guest; a typical pseries guest will use 16MB > > > > > > > for IOMMU pages without checking the mmu pagesize and this will fail > > > > > > > at https://git.qemu.org/?p=qemu.git;a=blob;f=hw/vfio/common.c;h=fb396cf00ac40eb35967a04c9cc798ca896eed57;hb=refs/heads/master#l256 > > > > > > > > > > > > > > With the change, mapping will fail in KVM and the guest will print: > > > > > > > > > > > > > > mlx5_core 0000:00:00.0: ibm,create-pe-dma-window(2027) 0 8000000 20000000 18 1f returned 0 (liobn = 0x80000001 starting addr = 8000000 0) > > > > > > > mlx5_core 0000:00:00.0: created tce table LIOBN 0x80000001 for /pci@800000020000000/ethernet@0 > > > > > > > mlx5_core 0000:00:00.0: failed to map direct window for > > > > > > > /pci@800000020000000/ethernet@0: -1 > > > > > > > > > > > > [snip] > > > > > > > @@ -124,7 +125,7 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > > struct mm_iommu_table_group_mem_t **pmem) > > > > > > > { > > > > > > > struct mm_iommu_table_group_mem_t *mem; > > > > > > > - long i, j, ret = 0, locked_entries = 0; > > > > > > > + long i, j, ret = 0, locked_entries = 0, pageshift; > > > > > > > struct page *page = NULL; > > > > > > > > > > > > > > mutex_lock(&mem_list_mutex); > > > > > > > @@ -166,6 +167,8 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > > goto unlock_exit; > > > > > > > } > > > > > > > > > > > > > > + mem->pageshift = 30; /* start from 1G pages - the biggest we have */ > > > > > > > > > > > > What about 16G pages on an HPT system? > > > > > > > > > > > > > > > Below in the loop mem->pageshift will reduce to the biggest actual size > > > > > which will be 16mb/64k/4k. Or remain 1GB if no memory is actually > > > > > pinned, no loss there. > > > > > > > > Are you saying that 16G IOMMU pages aren't supported? Or that there's > > > > some reason a guest can never use them? > > > > > > > > > ah, 16_G_, not _M_. My bad. I just never tried such huge pages, I will > > > lift the limit up to 64 then, easier this way. > > > > > > Ah, no, rather this as the upper limit: > > > > mem->pageshift = ilog2(entries) + PAGE_SHIFT; > > I can't make sense of this comment in context. I see how you're > computing the minimum page size in the reserved region. > > My question is about what the "maximum minimum" is - the starting > value from which you calculate. Currently it's 1G, but I can't > immediately see a reason that 16G is impossible here. 16GB is impossible if the chunk we are preregistering here is smaller than that, for example, the entire guest ram is 4GB. If that is the case and we try mapping a 16GB IOMMU page, this should fail as I do not really know what happens to the memory between 4GB..16GB. imho if not that, than 1<<64 would make a good upper limit. > > > @entries here is a number of system pages being pinned in that > > function. > > > > > > > > > > > > > > > > > > > > for (i = 0; i < entries; ++i) { > > > > > > > if (1 != get_user_pages_fast(ua + (i << PAGE_SHIFT), > > > > > > > 1/* pages */, 1/* iswrite */, &page)) { > > > > > > > @@ -199,6 +202,11 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > > } > > > > > > > } > > > > > > > populate: > > > > > > > + pageshift = PAGE_SHIFT; > > > > > > > + if (PageCompound(page)) > > > > > > > + pageshift += compound_order(compound_head(page)); > > > > > > > + mem->pageshift = min_t(unsigned int, mem->pageshift, pageshift); > > > > > > > > > > > > Why not make mem->pageshift and pageshift local the same type to avoid > > > > > > the min_t() ? > > > > > > > > > > I was under impression min() is deprecated (misinterpret checkpatch.pl > > > > > may be) and therefore did not pay attention to it. I can fix this and > > > > > repost if there is no other question. > > > > > > > > Hm, it's possible. > > > > > > Nah, tried min(), compiles fine. > > > > > > > > > > -- > David Gibson | I'll have my music baroque, and my code > david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_ > | _way_ _around_! > http://www.ozlabs.org/~dgibson -- Alexey
On Mon, Jul 02, 2018 at 02:33:30PM +1000, Alexey Kardashevskiy wrote: > On Mon, 2 Jul 2018 14:08:52 +1000 > David Gibson <david@gibson.dropbear.id.au> wrote: > > > On Fri, Jun 29, 2018 at 05:07:47PM +1000, Alexey Kardashevskiy wrote: > > > On Fri, 29 Jun 2018 15:18:20 +1000 > > > Alexey Kardashevskiy <aik@ozlabs.ru> wrote: > > > > > > > On Fri, 29 Jun 2018 14:57:02 +1000 > > > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > > > > > On Fri, Jun 29, 2018 at 02:51:21PM +1000, Alexey Kardashevskiy wrote: > > > > > > On Fri, 29 Jun 2018 14:12:41 +1000 > > > > > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > > > > > > > > > On Tue, Jun 26, 2018 at 03:59:26PM +1000, Alexey Kardashevskiy wrote: > > > > > > > > We already have a check in drivers/vfio/vfio_iommu_spapr_tce.c that > > > > > > > > an IOMMU page is contained in the physical page so the PCI hardware won't > > > > > > > > get access to unassigned host memory. > > > > > > > > > > > > > > > > However we do not have this check in KVM fastpath (H_PUT_TCE accelerated > > > > > > > > code) so the user space can pin memory backed with 64k pages and create > > > > > > > > a hardware TCE table with a bigger page size. We were lucky so far and > > > > > > > > did not hit this yet as the very first time the mapping happens > > > > > > > > we do not have tbl::it_userspace allocated yet and fall back to > > > > > > > > the userspace which in turn calls VFIO IOMMU driver and that fails > > > > > > > > because of the check in vfio_iommu_spapr_tce.c which is really > > > > > > > > sustainable solution. > > > > > > > > > > > > > > > > This stores the smallest preregistered page size in the preregistered > > > > > > > > region descriptor and changes the mm_iommu_xxx API to check this against > > > > > > > > the IOMMU page size. > > > > > > > > > > > > > > > > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> > > > > > > > > --- > > > > > > > > Changes: > > > > > > > > v2: > > > > > > > > * explicitly check for compound pages before calling compound_order() > > > > > > > > > > > > > > > > --- > > > > > > > > The bug is: run QEMU _without_ hugepages (no -mempath) and tell it to > > > > > > > > advertise 16MB pages to the guest; a typical pseries guest will use 16MB > > > > > > > > for IOMMU pages without checking the mmu pagesize and this will fail > > > > > > > > at https://git.qemu.org/?p=qemu.git;a=blob;f=hw/vfio/common.c;h=fb396cf00ac40eb35967a04c9cc798ca896eed57;hb=refs/heads/master#l256 > > > > > > > > > > > > > > > > With the change, mapping will fail in KVM and the guest will print: > > > > > > > > > > > > > > > > mlx5_core 0000:00:00.0: ibm,create-pe-dma-window(2027) 0 8000000 20000000 18 1f returned 0 (liobn = 0x80000001 starting addr = 8000000 0) > > > > > > > > mlx5_core 0000:00:00.0: created tce table LIOBN 0x80000001 for /pci@800000020000000/ethernet@0 > > > > > > > > mlx5_core 0000:00:00.0: failed to map direct window for > > > > > > > > /pci@800000020000000/ethernet@0: -1 > > > > > > > > > > > > > > [snip] > > > > > > > > @@ -124,7 +125,7 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > > > struct mm_iommu_table_group_mem_t **pmem) > > > > > > > > { > > > > > > > > struct mm_iommu_table_group_mem_t *mem; > > > > > > > > - long i, j, ret = 0, locked_entries = 0; > > > > > > > > + long i, j, ret = 0, locked_entries = 0, pageshift; > > > > > > > > struct page *page = NULL; > > > > > > > > > > > > > > > > mutex_lock(&mem_list_mutex); > > > > > > > > @@ -166,6 +167,8 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > > > goto unlock_exit; > > > > > > > > } > > > > > > > > > > > > > > > > + mem->pageshift = 30; /* start from 1G pages - the biggest we have */ > > > > > > > > > > > > > > What about 16G pages on an HPT system? > > > > > > > > > > > > > > > > > > Below in the loop mem->pageshift will reduce to the biggest actual size > > > > > > which will be 16mb/64k/4k. Or remain 1GB if no memory is actually > > > > > > pinned, no loss there. > > > > > > > > > > Are you saying that 16G IOMMU pages aren't supported? Or that there's > > > > > some reason a guest can never use them? > > > > > > > > > > > > ah, 16_G_, not _M_. My bad. I just never tried such huge pages, I will > > > > lift the limit up to 64 then, easier this way. > > > > > > > > > Ah, no, rather this as the upper limit: > > > > > > mem->pageshift = ilog2(entries) + PAGE_SHIFT; > > > > I can't make sense of this comment in context. I see how you're > > computing the minimum page size in the reserved region. > > > > My question is about what the "maximum minimum" is - the starting > > value from which you calculate. Currently it's 1G, but I can't > > immediately see a reason that 16G is impossible here. > > > 16GB is impossible if the chunk we are preregistering here is smaller > than that, for example, the entire guest ram is 4GB. Of course. Just like it was for 1GiB if you had a 512MiB guest, for example. I'm talking about a case where you have a guest that's >=16GiB and you *have* allocated 16GiB hugepages to back it. > If that is the > case and we try mapping a 16GB IOMMU page, this should fail as I do not > really know what happens to the memory between 4GB..16GB. > > imho if not that, than 1<<64 would make a good upper limit. > > > > > > > > @entries here is a number of system pages being pinned in that > > > function. > > > > > > > > > > > > > > > > > > > > > > > > > > for (i = 0; i < entries; ++i) { > > > > > > > > if (1 != get_user_pages_fast(ua + (i << PAGE_SHIFT), > > > > > > > > 1/* pages */, 1/* iswrite */, &page)) { > > > > > > > > @@ -199,6 +202,11 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > > > } > > > > > > > > } > > > > > > > > populate: > > > > > > > > + pageshift = PAGE_SHIFT; > > > > > > > > + if (PageCompound(page)) > > > > > > > > + pageshift += compound_order(compound_head(page)); > > > > > > > > + mem->pageshift = min_t(unsigned int, mem->pageshift, pageshift); > > > > > > > > > > > > > > Why not make mem->pageshift and pageshift local the same type to avoid > > > > > > > the min_t() ? > > > > > > > > > > > > I was under impression min() is deprecated (misinterpret checkpatch.pl > > > > > > may be) and therefore did not pay attention to it. I can fix this and > > > > > > repost if there is no other question. > > > > > > > > > > Hm, it's possible. > > > > > > > > Nah, tried min(), compiles fine. > > > > > > > > > > > > > > > > > >
On Mon, 2 Jul 2018 14:52:43 +1000 David Gibson <david@gibson.dropbear.id.au> wrote: > On Mon, Jul 02, 2018 at 02:33:30PM +1000, Alexey Kardashevskiy wrote: > > On Mon, 2 Jul 2018 14:08:52 +1000 > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > On Fri, Jun 29, 2018 at 05:07:47PM +1000, Alexey Kardashevskiy wrote: > > > > On Fri, 29 Jun 2018 15:18:20 +1000 > > > > Alexey Kardashevskiy <aik@ozlabs.ru> wrote: > > > > > > > > > On Fri, 29 Jun 2018 14:57:02 +1000 > > > > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > > > > > > > On Fri, Jun 29, 2018 at 02:51:21PM +1000, Alexey Kardashevskiy wrote: > > > > > > > On Fri, 29 Jun 2018 14:12:41 +1000 > > > > > > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > > > > > > > > > > > On Tue, Jun 26, 2018 at 03:59:26PM +1000, Alexey Kardashevskiy wrote: > > > > > > > > > We already have a check in drivers/vfio/vfio_iommu_spapr_tce.c that > > > > > > > > > an IOMMU page is contained in the physical page so the PCI hardware won't > > > > > > > > > get access to unassigned host memory. > > > > > > > > > > > > > > > > > > However we do not have this check in KVM fastpath (H_PUT_TCE accelerated > > > > > > > > > code) so the user space can pin memory backed with 64k pages and create > > > > > > > > > a hardware TCE table with a bigger page size. We were lucky so far and > > > > > > > > > did not hit this yet as the very first time the mapping happens > > > > > > > > > we do not have tbl::it_userspace allocated yet and fall back to > > > > > > > > > the userspace which in turn calls VFIO IOMMU driver and that fails > > > > > > > > > because of the check in vfio_iommu_spapr_tce.c which is really > > > > > > > > > sustainable solution. > > > > > > > > > > > > > > > > > > This stores the smallest preregistered page size in the preregistered > > > > > > > > > region descriptor and changes the mm_iommu_xxx API to check this against > > > > > > > > > the IOMMU page size. > > > > > > > > > > > > > > > > > > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> > > > > > > > > > --- > > > > > > > > > Changes: > > > > > > > > > v2: > > > > > > > > > * explicitly check for compound pages before calling compound_order() > > > > > > > > > > > > > > > > > > --- > > > > > > > > > The bug is: run QEMU _without_ hugepages (no -mempath) and tell it to > > > > > > > > > advertise 16MB pages to the guest; a typical pseries guest will use 16MB > > > > > > > > > for IOMMU pages without checking the mmu pagesize and this will fail > > > > > > > > > at https://git.qemu.org/?p=qemu.git;a=blob;f=hw/vfio/common.c;h=fb396cf00ac40eb35967a04c9cc798ca896eed57;hb=refs/heads/master#l256 > > > > > > > > > > > > > > > > > > With the change, mapping will fail in KVM and the guest will print: > > > > > > > > > > > > > > > > > > mlx5_core 0000:00:00.0: ibm,create-pe-dma-window(2027) 0 8000000 20000000 18 1f returned 0 (liobn = 0x80000001 starting addr = 8000000 0) > > > > > > > > > mlx5_core 0000:00:00.0: created tce table LIOBN 0x80000001 for /pci@800000020000000/ethernet@0 > > > > > > > > > mlx5_core 0000:00:00.0: failed to map direct window for > > > > > > > > > /pci@800000020000000/ethernet@0: -1 > > > > > > > > > > > > > > > > [snip] > > > > > > > > > @@ -124,7 +125,7 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > > > > struct mm_iommu_table_group_mem_t **pmem) > > > > > > > > > { > > > > > > > > > struct mm_iommu_table_group_mem_t *mem; > > > > > > > > > - long i, j, ret = 0, locked_entries = 0; > > > > > > > > > + long i, j, ret = 0, locked_entries = 0, pageshift; > > > > > > > > > struct page *page = NULL; > > > > > > > > > > > > > > > > > > mutex_lock(&mem_list_mutex); > > > > > > > > > @@ -166,6 +167,8 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > > > > goto unlock_exit; > > > > > > > > > } > > > > > > > > > > > > > > > > > > + mem->pageshift = 30; /* start from 1G pages - the biggest we have */ > > > > > > > > > > > > > > > > What about 16G pages on an HPT system? > > > > > > > > > > > > > > > > > > > > > Below in the loop mem->pageshift will reduce to the biggest actual size > > > > > > > which will be 16mb/64k/4k. Or remain 1GB if no memory is actually > > > > > > > pinned, no loss there. > > > > > > > > > > > > Are you saying that 16G IOMMU pages aren't supported? Or that there's > > > > > > some reason a guest can never use them? > > > > > > > > > > > > > > > ah, 16_G_, not _M_. My bad. I just never tried such huge pages, I will > > > > > lift the limit up to 64 then, easier this way. > > > > > > > > > > > > Ah, no, rather this as the upper limit: > > > > > > > > mem->pageshift = ilog2(entries) + PAGE_SHIFT; > > > > > > I can't make sense of this comment in context. I see how you're > > > computing the minimum page size in the reserved region. > > > > > > My question is about what the is - the starting > > > value from which you calculate. Currently it's 1G, but I can't > > > immediately see a reason that 16G is impossible here. > > > > > > 16GB is impossible if the chunk we are preregistering here is smaller > > than that, for example, the entire guest ram is 4GB. > > Of course. Just like it was for 1GiB if you had a 512MiB guest, for > example. I'm talking about a case where you have a guest that's > >=16GiB and you *have* allocated 16GiB hugepages to back it. Then, assuming we are preregistering entire RAM as a single chunk, the "maximum minimum" will be initialized as ">=16GiB" (but floor-aligned to power of two) before the pinning loop and then reduce to the actual page size, inside the loop. I feel like I am missing something in the question, what is that? -- Alexey
On Mon, Jul 02, 2018 at 04:32:27PM +1000, Alexey Kardashevskiy wrote: > On Mon, 2 Jul 2018 14:52:43 +1000 > David Gibson <david@gibson.dropbear.id.au> wrote: > > > On Mon, Jul 02, 2018 at 02:33:30PM +1000, Alexey Kardashevskiy wrote: > > > On Mon, 2 Jul 2018 14:08:52 +1000 > > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > > > On Fri, Jun 29, 2018 at 05:07:47PM +1000, Alexey Kardashevskiy wrote: > > > > > On Fri, 29 Jun 2018 15:18:20 +1000 > > > > > Alexey Kardashevskiy <aik@ozlabs.ru> wrote: > > > > > > > > > > > On Fri, 29 Jun 2018 14:57:02 +1000 > > > > > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > > > > > > > > > On Fri, Jun 29, 2018 at 02:51:21PM +1000, Alexey Kardashevskiy wrote: > > > > > > > > On Fri, 29 Jun 2018 14:12:41 +1000 > > > > > > > > David Gibson <david@gibson.dropbear.id.au> wrote: > > > > > > > > > > > > > > > > > On Tue, Jun 26, 2018 at 03:59:26PM +1000, Alexey Kardashevskiy wrote: > > > > > > > > > > We already have a check in drivers/vfio/vfio_iommu_spapr_tce.c that > > > > > > > > > > an IOMMU page is contained in the physical page so the PCI hardware won't > > > > > > > > > > get access to unassigned host memory. > > > > > > > > > > > > > > > > > > > > However we do not have this check in KVM fastpath (H_PUT_TCE accelerated > > > > > > > > > > code) so the user space can pin memory backed with 64k pages and create > > > > > > > > > > a hardware TCE table with a bigger page size. We were lucky so far and > > > > > > > > > > did not hit this yet as the very first time the mapping happens > > > > > > > > > > we do not have tbl::it_userspace allocated yet and fall back to > > > > > > > > > > the userspace which in turn calls VFIO IOMMU driver and that fails > > > > > > > > > > because of the check in vfio_iommu_spapr_tce.c which is really > > > > > > > > > > sustainable solution. > > > > > > > > > > > > > > > > > > > > This stores the smallest preregistered page size in the preregistered > > > > > > > > > > region descriptor and changes the mm_iommu_xxx API to check this against > > > > > > > > > > the IOMMU page size. > > > > > > > > > > > > > > > > > > > > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> > > > > > > > > > > --- > > > > > > > > > > Changes: > > > > > > > > > > v2: > > > > > > > > > > * explicitly check for compound pages before calling compound_order() > > > > > > > > > > > > > > > > > > > > --- > > > > > > > > > > The bug is: run QEMU _without_ hugepages (no -mempath) and tell it to > > > > > > > > > > advertise 16MB pages to the guest; a typical pseries guest will use 16MB > > > > > > > > > > for IOMMU pages without checking the mmu pagesize and this will fail > > > > > > > > > > at https://git.qemu.org/?p=qemu.git;a=blob;f=hw/vfio/common.c;h=fb396cf00ac40eb35967a04c9cc798ca896eed57;hb=refs/heads/master#l256 > > > > > > > > > > > > > > > > > > > > With the change, mapping will fail in KVM and the guest will print: > > > > > > > > > > > > > > > > > > > > mlx5_core 0000:00:00.0: ibm,create-pe-dma-window(2027) 0 8000000 20000000 18 1f returned 0 (liobn = 0x80000001 starting addr = 8000000 0) > > > > > > > > > > mlx5_core 0000:00:00.0: created tce table LIOBN 0x80000001 for /pci@800000020000000/ethernet@0 > > > > > > > > > > mlx5_core 0000:00:00.0: failed to map direct window for > > > > > > > > > > /pci@800000020000000/ethernet@0: -1 > > > > > > > > > > > > > > > > > > [snip] > > > > > > > > > > @@ -124,7 +125,7 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > > > > > struct mm_iommu_table_group_mem_t **pmem) > > > > > > > > > > { > > > > > > > > > > struct mm_iommu_table_group_mem_t *mem; > > > > > > > > > > - long i, j, ret = 0, locked_entries = 0; > > > > > > > > > > + long i, j, ret = 0, locked_entries = 0, pageshift; > > > > > > > > > > struct page *page = NULL; > > > > > > > > > > > > > > > > > > > > mutex_lock(&mem_list_mutex); > > > > > > > > > > @@ -166,6 +167,8 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, > > > > > > > > > > goto unlock_exit; > > > > > > > > > > } > > > > > > > > > > > > > > > > > > > > + mem->pageshift = 30; /* start from 1G pages - the biggest we have */ > > > > > > > > > > > > > > > > > > What about 16G pages on an HPT system? > > > > > > > > > > > > > > > > > > > > > > > > Below in the loop mem->pageshift will reduce to the biggest actual size > > > > > > > > which will be 16mb/64k/4k. Or remain 1GB if no memory is actually > > > > > > > > pinned, no loss there. > > > > > > > > > > > > > > Are you saying that 16G IOMMU pages aren't supported? Or that there's > > > > > > > some reason a guest can never use them? > > > > > > > > > > > > > > > > > > ah, 16_G_, not _M_. My bad. I just never tried such huge pages, I will > > > > > > lift the limit up to 64 then, easier this way. > > > > > > > > > > > > > > > Ah, no, rather this as the upper limit: > > > > > > > > > > mem->pageshift = ilog2(entries) + PAGE_SHIFT; > > > > > > > > I can't make sense of this comment in context. I see how you're > > > > computing the minimum page size in the reserved region. > > > > > > > > My question is about what the is - the starting > > > > value from which you calculate. Currently it's 1G, but I can't > > > > immediately see a reason that 16G is impossible here. > > > > > > > > > 16GB is impossible if the chunk we are preregistering here is smaller > > > than that, for example, the entire guest ram is 4GB. > > > > Of course. Just like it was for 1GiB if you had a 512MiB guest, for > > example. I'm talking about a case where you have a guest that's > > >=16GiB and you *have* allocated 16GiB hugepages to back it. > > > Then, assuming we are preregistering entire RAM as a single chunk, the > "maximum minimum" will be initialized as ">=16GiB" (but floor-aligned > to power of two) before the pinning loop and then reduce to the actual > page size, inside the loop. I feel like I am missing something in the > question, what is that? I wish I knew... By the "maximum minimum" I'm talking about the shift value you start with before you start looking at the actual pages. Currently that is exactly 30 (1 GiB). AFAICT that prevents using 16GiB pages, and I can't see a good reason for that.
diff --git a/arch/powerpc/include/asm/mmu_context.h b/arch/powerpc/include/asm/mmu_context.h index 896efa5..79d570c 100644 --- a/arch/powerpc/include/asm/mmu_context.h +++ b/arch/powerpc/include/asm/mmu_context.h @@ -35,9 +35,9 @@ extern struct mm_iommu_table_group_mem_t *mm_iommu_lookup_rm( extern struct mm_iommu_table_group_mem_t *mm_iommu_find(struct mm_struct *mm, unsigned long ua, unsigned long entries); extern long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem, - unsigned long ua, unsigned long *hpa); + unsigned long ua, unsigned int pageshift, unsigned long *hpa); extern long mm_iommu_ua_to_hpa_rm(struct mm_iommu_table_group_mem_t *mem, - unsigned long ua, unsigned long *hpa); + unsigned long ua, unsigned int pageshift, unsigned long *hpa); extern long mm_iommu_mapped_inc(struct mm_iommu_table_group_mem_t *mem); extern void mm_iommu_mapped_dec(struct mm_iommu_table_group_mem_t *mem); #endif diff --git a/arch/powerpc/kvm/book3s_64_vio.c b/arch/powerpc/kvm/book3s_64_vio.c index d066e37..8c456fa 100644 --- a/arch/powerpc/kvm/book3s_64_vio.c +++ b/arch/powerpc/kvm/book3s_64_vio.c @@ -449,7 +449,7 @@ long kvmppc_tce_iommu_do_map(struct kvm *kvm, struct iommu_table *tbl, /* This only handles v2 IOMMU type, v1 is handled via ioctl() */ return H_TOO_HARD; - if (WARN_ON_ONCE(mm_iommu_ua_to_hpa(mem, ua, &hpa))) + if (WARN_ON_ONCE(mm_iommu_ua_to_hpa(mem, ua, tbl->it_page_shift, &hpa))) return H_HARDWARE; if (mm_iommu_mapped_inc(mem)) diff --git a/arch/powerpc/kvm/book3s_64_vio_hv.c b/arch/powerpc/kvm/book3s_64_vio_hv.c index 925fc31..5b298f5 100644 --- a/arch/powerpc/kvm/book3s_64_vio_hv.c +++ b/arch/powerpc/kvm/book3s_64_vio_hv.c @@ -279,7 +279,8 @@ static long kvmppc_rm_tce_iommu_do_map(struct kvm *kvm, struct iommu_table *tbl, if (!mem) return H_TOO_HARD; - if (WARN_ON_ONCE_RM(mm_iommu_ua_to_hpa_rm(mem, ua, &hpa))) + if (WARN_ON_ONCE_RM(mm_iommu_ua_to_hpa_rm(mem, ua, tbl->it_page_shift, + &hpa))) return H_HARDWARE; pua = (void *) vmalloc_to_phys(pua); @@ -469,7 +470,8 @@ long kvmppc_rm_h_put_tce_indirect(struct kvm_vcpu *vcpu, mem = mm_iommu_lookup_rm(vcpu->kvm->mm, ua, IOMMU_PAGE_SIZE_4K); if (mem) - prereg = mm_iommu_ua_to_hpa_rm(mem, ua, &tces) == 0; + prereg = mm_iommu_ua_to_hpa_rm(mem, ua, + IOMMU_PAGE_SHIFT_4K, &tces) == 0; } if (!prereg) { diff --git a/arch/powerpc/mm/mmu_context_iommu.c b/arch/powerpc/mm/mmu_context_iommu.c index abb4364..dc0e8cd 100644 --- a/arch/powerpc/mm/mmu_context_iommu.c +++ b/arch/powerpc/mm/mmu_context_iommu.c @@ -27,6 +27,7 @@ struct mm_iommu_table_group_mem_t { struct rcu_head rcu; unsigned long used; atomic64_t mapped; + unsigned int pageshift; u64 ua; /* userspace address */ u64 entries; /* number of entries in hpas[] */ u64 *hpas; /* vmalloc'ed */ @@ -124,7 +125,7 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, struct mm_iommu_table_group_mem_t **pmem) { struct mm_iommu_table_group_mem_t *mem; - long i, j, ret = 0, locked_entries = 0; + long i, j, ret = 0, locked_entries = 0, pageshift; struct page *page = NULL; mutex_lock(&mem_list_mutex); @@ -166,6 +167,8 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, goto unlock_exit; } + mem->pageshift = 30; /* start from 1G pages - the biggest we have */ + for (i = 0; i < entries; ++i) { if (1 != get_user_pages_fast(ua + (i << PAGE_SHIFT), 1/* pages */, 1/* iswrite */, &page)) { @@ -199,6 +202,11 @@ long mm_iommu_get(struct mm_struct *mm, unsigned long ua, unsigned long entries, } } populate: + pageshift = PAGE_SHIFT; + if (PageCompound(page)) + pageshift += compound_order(compound_head(page)); + mem->pageshift = min_t(unsigned int, mem->pageshift, pageshift); + mem->hpas[i] = page_to_pfn(page) << PAGE_SHIFT; } @@ -349,7 +357,7 @@ struct mm_iommu_table_group_mem_t *mm_iommu_find(struct mm_struct *mm, EXPORT_SYMBOL_GPL(mm_iommu_find); long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem, - unsigned long ua, unsigned long *hpa) + unsigned long ua, unsigned int pageshift, unsigned long *hpa) { const long entry = (ua - mem->ua) >> PAGE_SHIFT; u64 *va = &mem->hpas[entry]; @@ -357,6 +365,9 @@ long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem, if (entry >= mem->entries) return -EFAULT; + if (pageshift > mem->pageshift) + return -EFAULT; + *hpa = *va | (ua & ~PAGE_MASK); return 0; @@ -364,7 +375,7 @@ long mm_iommu_ua_to_hpa(struct mm_iommu_table_group_mem_t *mem, EXPORT_SYMBOL_GPL(mm_iommu_ua_to_hpa); long mm_iommu_ua_to_hpa_rm(struct mm_iommu_table_group_mem_t *mem, - unsigned long ua, unsigned long *hpa) + unsigned long ua, unsigned int pageshift, unsigned long *hpa) { const long entry = (ua - mem->ua) >> PAGE_SHIFT; void *va = &mem->hpas[entry]; @@ -373,6 +384,9 @@ long mm_iommu_ua_to_hpa_rm(struct mm_iommu_table_group_mem_t *mem, if (entry >= mem->entries) return -EFAULT; + if (pageshift > mem->pageshift) + return -EFAULT; + pa = (void *) vmalloc_to_phys(va); if (!pa) return -EFAULT; diff --git a/drivers/vfio/vfio_iommu_spapr_tce.c b/drivers/vfio/vfio_iommu_spapr_tce.c index 2da5f05..7cd63b0 100644 --- a/drivers/vfio/vfio_iommu_spapr_tce.c +++ b/drivers/vfio/vfio_iommu_spapr_tce.c @@ -467,7 +467,7 @@ static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container, if (!mem) return -EINVAL; - ret = mm_iommu_ua_to_hpa(mem, tce, phpa); + ret = mm_iommu_ua_to_hpa(mem, tce, shift, phpa); if (ret) return -EINVAL;
We already have a check in drivers/vfio/vfio_iommu_spapr_tce.c that an IOMMU page is contained in the physical page so the PCI hardware won't get access to unassigned host memory. However we do not have this check in KVM fastpath (H_PUT_TCE accelerated code) so the user space can pin memory backed with 64k pages and create a hardware TCE table with a bigger page size. We were lucky so far and did not hit this yet as the very first time the mapping happens we do not have tbl::it_userspace allocated yet and fall back to the userspace which in turn calls VFIO IOMMU driver and that fails because of the check in vfio_iommu_spapr_tce.c which is really sustainable solution. This stores the smallest preregistered page size in the preregistered region descriptor and changes the mm_iommu_xxx API to check this against the IOMMU page size. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> --- Changes: v2: * explicitly check for compound pages before calling compound_order() --- The bug is: run QEMU _without_ hugepages (no -mempath) and tell it to advertise 16MB pages to the guest; a typical pseries guest will use 16MB for IOMMU pages without checking the mmu pagesize and this will fail at https://git.qemu.org/?p=qemu.git;a=blob;f=hw/vfio/common.c;h=fb396cf00ac40eb35967a04c9cc798ca896eed57;hb=refs/heads/master#l256 With the change, mapping will fail in KVM and the guest will print: mlx5_core 0000:00:00.0: ibm,create-pe-dma-window(2027) 0 8000000 20000000 18 1f returned 0 (liobn = 0x80000001 starting addr = 8000000 0) mlx5_core 0000:00:00.0: created tce table LIOBN 0x80000001 for /pci@800000020000000/ethernet@0 mlx5_core 0000:00:00.0: failed to map direct window for /pci@800000020000000/ethernet@0: -1 --- arch/powerpc/include/asm/mmu_context.h | 4 ++-- arch/powerpc/kvm/book3s_64_vio.c | 2 +- arch/powerpc/kvm/book3s_64_vio_hv.c | 6 ++++-- arch/powerpc/mm/mmu_context_iommu.c | 20 +++++++++++++++++--- drivers/vfio/vfio_iommu_spapr_tce.c | 2 +- 5 files changed, 25 insertions(+), 9 deletions(-)