Re: [RFC PATCH 00/19] hugetlb support for KVM guest_mem
On Fri, Jun 16, 2023 at 11:28 AM Mike Kravetz wrote: > > On 06/06/23 19:03, Ackerley Tng wrote: > > Hello, > > > > This patchset builds upon a soon-to-be-published WIP patchset that Sean > > published at https://github.com/sean-jc/linux/tree/x86/kvm_gmem_solo, > > mentioned > > at [1]. > > > > The tree can be found at: > > https://github.com/googleprodkernel/linux-cc/tree/gmem-hugetlb-rfc-v1 > > > > In this patchset, hugetlb support for KVM's guest_mem (aka gmem) is > > introduced, > > allowing VM private memory (for confidential computing) to be backed by > > hugetlb > > pages. > > > > guest_mem provides userspace with a handle, with which userspace can > > allocate > > and deallocate memory for confidential VMs without mapping the memory into > > userspace. > > Hello Ackerley, > > I am not sure if you are aware or, have been following the hugetlb HGM > discussion in this thread: > https://lore.kernel.org/linux-mm/20230306191944.GA15773@monkey/ > > There we are trying to decide if HGM should be added to hugetlb, or if > perhaps a new filesystem/driver/allocator should be created. The concern > is added complexity to hugetlb as well as core mm special casing. Note > that HGM is addressing issues faced by existing hugetlb users. > > Your proposal here suggests modifying hugetlb so that it can be used in > a new way (use case) by KVM's guest_mem. As such it really seems like > something that should be done in a separate filesystem/driver/allocator. > You will likely not get much support for modifying hugetlb. > > -- > Mike Kravetz > IIUC mm/hugetlb.c implements memory manager for Hugetlb pages and fd/hugetlbfs/inode.c implements the filesystem logic for hugetlbfs. This series implements a new filesystem with limited operations parallel to hugetlbfs filesystem but tries to reuse hugetlb memory manager. The effort here is to not add any new feature to hugetlb memory manager but clean it up so that it can be used by a new filesystem. guest_mem warrants a new filesystem since it supports limited operations on the underlying files but there is no additional restriction on underlying memory management. Though one could argue that memory management for guest_mem files can be a very simple one that goes inline with limited operations on the files. If this series were to go a separate way of implementing a new memory manager, one immediate requirement that might spring up, would be to convert memory from hugetlb managed memory to be managed by this newly introduced memory manager and vice a versa at runtime since there could be a mix of VMs on the same platform using guest_mem and hugetlb. Maybe this can be satisfied by having a separate global pool for reservation that's consumed by both, which would need more changes in my understanding. Using guest_mem for all the VMs by default would be a future work contingent on all existing usecases/requirements being satisfied. Regards, Vishal
Re: [RFC PATCH 00/19] hugetlb support for KVM guest_mem
On 06/06/23 19:03, Ackerley Tng wrote: > Hello, > > This patchset builds upon a soon-to-be-published WIP patchset that Sean > published at https://github.com/sean-jc/linux/tree/x86/kvm_gmem_solo, > mentioned > at [1]. > > The tree can be found at: > https://github.com/googleprodkernel/linux-cc/tree/gmem-hugetlb-rfc-v1 > > In this patchset, hugetlb support for KVM's guest_mem (aka gmem) is > introduced, > allowing VM private memory (for confidential computing) to be backed by > hugetlb > pages. > > guest_mem provides userspace with a handle, with which userspace can allocate > and deallocate memory for confidential VMs without mapping the memory into > userspace. Hello Ackerley, I am not sure if you are aware or, have been following the hugetlb HGM discussion in this thread: https://lore.kernel.org/linux-mm/20230306191944.GA15773@monkey/ There we are trying to decide if HGM should be added to hugetlb, or if perhaps a new filesystem/driver/allocator should be created. The concern is added complexity to hugetlb as well as core mm special casing. Note that HGM is addressing issues faced by existing hugetlb users. Your proposal here suggests modifying hugetlb so that it can be used in a new way (use case) by KVM's guest_mem. As such it really seems like something that should be done in a separate filesystem/driver/allocator. You will likely not get much support for modifying hugetlb. -- Mike Kravetz > Why use hugetlb instead of introducing a new allocator, like gmem does for 4K > and transparent hugepages? > > + hugetlb provides the following useful functionality, which would otherwise > have to be reimplemented: > + Allocation of hugetlb pages at boot time, including > + Parsing of kernel boot parameters to configure hugetlb > + Tracking of usage in hstate > + gmem will share the same system-wide pool of hugetlb pages, so users > don't have to have separate pools for hugetlb and gmem > + Page accounting with subpools > + hugetlb pages are tracked in subpools, which gmem uses to reserve > pages from the global hstate > + Memory charging > + hugetlb provides code that charges memory to cgroups > + Reporting: hugetlb usage and availability are available at > /proc/meminfo, > etc > > The first 11 patches in this patchset is a series of refactoring to decouple > hugetlb and hugetlbfs. > > The central thread binding the refactoring is that some functions (like > inode_resv_map(), inode_subpool(), inode_hstate(), etc) rely on a hugetlbfs > concept, that the resv_map, subpool, hstate, are in a specific field in a > hugetlb inode. > > Refactoring to parametrize functions by hstate, subpool, resv_map will allow > hugetlb to be used by gmem and in other places where these data structures > aren't necessarily stored in the same positions in the inode. > > The refactoring proposed here is just the minimum required to get a > proof-of-concept working with gmem. I would like to get opinions on this > approach before doing further refactoring. (See TODOs) > > TODOs: > > + hugetlb/hugetlbfs refactoring > + remove_inode_hugepages() no longer needs to be exposed, it is hugetlbfs > specific and used only in inode.c > + remove_mapping_hugepages(), remove_inode_single_folio(), > hugetlb_unreserve_pages() shouldn't need to take inode as a parameter > + Updating inode->i_blocks can be refactored to a separate function > and > called from hugetlbfs and gmem > + alloc_hugetlb_folio_from_subpool() shouldn't need to be parametrized by > vma > + hugetlb_reserve_pages() should be refactored to be symmetric with > hugetlb_unreserve_pages() > + It should be parametrized by resv_map > + alloc_hugetlb_folio_from_subpool() could perhaps use > hugetlb_reserve_pages()? > + gmem > + Figure out if resv_map should be used by gmem at all > + Probably needs more refactoring to decouple resv_map from hugetlb > functions > > Questions for the community: > > 1. In this patchset, every gmem file backed with hugetlb is given a new >subpool. Is that desirable? > + In hugetlbfs, a subpool always belongs to a mount, and hugetlbfs has one > mount per hugetlb size (2M, 1G, etc) > + memfd_create(MFD_HUGETLB) effectively returns a full hugetlbfs file, so > it > (rightfully) uses the hugetlbfs kernel mounts and their subpools > + I gave each file a subpool mostly to speed up implementation and still > be > able to reserve hugetlb pages from the global hstate based on the gmem > file size. > + gmem, unlike hugetlbfs, isn't meant to be a full filesystem, so > + Should there be multiple mounts, one for each hugetlb size? > + Will the mounts be initialized on boot or on first gmem file > creation? > + Or is one subpool per gmem file fine? > 2. Should resv_map be used for gmem at all,
Re: [RFC PATCH 00/19] hugetlb support for KVM guest_mem
On Tue, Jun 06, 2023 at 07:03:45PM +,
Ackerley Tng wrote:
> Hello,
>
> This patchset builds upon a soon-to-be-published WIP patchset that Sean
> published at https://github.com/sean-jc/linux/tree/x86/kvm_gmem_solo,
> mentioned
> at [1].
>
> The tree can be found at:
> https://github.com/googleprodkernel/linux-cc/tree/gmem-hugetlb-rfc-v1
>
> In this patchset, hugetlb support for KVM's guest_mem (aka gmem) is
> introduced,
> allowing VM private memory (for confidential computing) to be backed by
> hugetlb
> pages.
>
> guest_mem provides userspace with a handle, with which userspace can allocate
> and deallocate memory for confidential VMs without mapping the memory into
> userspace.
>
> Why use hugetlb instead of introducing a new allocator, like gmem does for 4K
> and transparent hugepages?
>
> + hugetlb provides the following useful functionality, which would otherwise
> have to be reimplemented:
> + Allocation of hugetlb pages at boot time, including
> + Parsing of kernel boot parameters to configure hugetlb
> + Tracking of usage in hstate
> + gmem will share the same system-wide pool of hugetlb pages, so users
> don't have to have separate pools for hugetlb and gmem
> + Page accounting with subpools
> + hugetlb pages are tracked in subpools, which gmem uses to reserve
> pages from the global hstate
> + Memory charging
> + hugetlb provides code that charges memory to cgroups
> + Reporting: hugetlb usage and availability are available at
> /proc/meminfo,
> etc
>
> The first 11 patches in this patchset is a series of refactoring to decouple
> hugetlb and hugetlbfs.
>
> The central thread binding the refactoring is that some functions (like
> inode_resv_map(), inode_subpool(), inode_hstate(), etc) rely on a hugetlbfs
> concept, that the resv_map, subpool, hstate, are in a specific field in a
> hugetlb inode.
>
> Refactoring to parametrize functions by hstate, subpool, resv_map will allow
> hugetlb to be used by gmem and in other places where these data structures
> aren't necessarily stored in the same positions in the inode.
>
> The refactoring proposed here is just the minimum required to get a
> proof-of-concept working with gmem. I would like to get opinions on this
> approach before doing further refactoring. (See TODOs)
>
> TODOs:
>
> + hugetlb/hugetlbfs refactoring
> + remove_inode_hugepages() no longer needs to be exposed, it is hugetlbfs
> specific and used only in inode.c
> + remove_mapping_hugepages(), remove_inode_single_folio(),
> hugetlb_unreserve_pages() shouldn't need to take inode as a parameter
> + Updating inode->i_blocks can be refactored to a separate function
> and
> called from hugetlbfs and gmem
> + alloc_hugetlb_folio_from_subpool() shouldn't need to be parametrized by
> vma
> + hugetlb_reserve_pages() should be refactored to be symmetric with
> hugetlb_unreserve_pages()
> + It should be parametrized by resv_map
> + alloc_hugetlb_folio_from_subpool() could perhaps use
> hugetlb_reserve_pages()?
> + gmem
> + Figure out if resv_map should be used by gmem at all
> + Probably needs more refactoring to decouple resv_map from hugetlb
> functions
Hi. If kvm gmem is compiled as kernel module, many symbols are failed to link.
You need to add EXPORT_SYMBOL{,_GPL} for exported symbols.
Or compile it to kernel instead of module?
Thanks,
> Questions for the community:
>
> 1. In this patchset, every gmem file backed with hugetlb is given a new
>subpool. Is that desirable?
> + In hugetlbfs, a subpool always belongs to a mount, and hugetlbfs has one
> mount per hugetlb size (2M, 1G, etc)
> + memfd_create(MFD_HUGETLB) effectively returns a full hugetlbfs file, so
> it
> (rightfully) uses the hugetlbfs kernel mounts and their subpools
> + I gave each file a subpool mostly to speed up implementation and still
> be
> able to reserve hugetlb pages from the global hstate based on the gmem
> file size.
> + gmem, unlike hugetlbfs, isn't meant to be a full filesystem, so
> + Should there be multiple mounts, one for each hugetlb size?
> + Will the mounts be initialized on boot or on first gmem file
> creation?
> + Or is one subpool per gmem file fine?
> 2. Should resv_map be used for gmem at all, since gmem doesn't allow userspace
>reservations?
>
> [1] https://lore.kernel.org/lkml/zem5zq8oo+xna...@google.com/
>
> ---
>
> Ackerley Tng (19):
> mm: hugetlb: Expose get_hstate_idx()
> mm: hugetlb: Move and expose hugetlbfs_zero_partial_page
> mm: hugetlb: Expose remove_inode_hugepages
> mm: hugetlb: Decouple hstate, subpool from inode
> mm: hugetlb: Allow alloc_hugetlb_folio() to be parametrized by subpool
> and hstate
> mm: hugetlb: Provide hugetlb_filemap_add_folio()
> mm: hugetlb: Refactor
[RFC PATCH 00/19] hugetlb support for KVM guest_mem
Hello, This patchset builds upon a soon-to-be-published WIP patchset that Sean published at https://github.com/sean-jc/linux/tree/x86/kvm_gmem_solo, mentioned at [1]. The tree can be found at: https://github.com/googleprodkernel/linux-cc/tree/gmem-hugetlb-rfc-v1 In this patchset, hugetlb support for KVM's guest_mem (aka gmem) is introduced, allowing VM private memory (for confidential computing) to be backed by hugetlb pages. guest_mem provides userspace with a handle, with which userspace can allocate and deallocate memory for confidential VMs without mapping the memory into userspace. Why use hugetlb instead of introducing a new allocator, like gmem does for 4K and transparent hugepages? + hugetlb provides the following useful functionality, which would otherwise have to be reimplemented: + Allocation of hugetlb pages at boot time, including + Parsing of kernel boot parameters to configure hugetlb + Tracking of usage in hstate + gmem will share the same system-wide pool of hugetlb pages, so users don't have to have separate pools for hugetlb and gmem + Page accounting with subpools + hugetlb pages are tracked in subpools, which gmem uses to reserve pages from the global hstate + Memory charging + hugetlb provides code that charges memory to cgroups + Reporting: hugetlb usage and availability are available at /proc/meminfo, etc The first 11 patches in this patchset is a series of refactoring to decouple hugetlb and hugetlbfs. The central thread binding the refactoring is that some functions (like inode_resv_map(), inode_subpool(), inode_hstate(), etc) rely on a hugetlbfs concept, that the resv_map, subpool, hstate, are in a specific field in a hugetlb inode. Refactoring to parametrize functions by hstate, subpool, resv_map will allow hugetlb to be used by gmem and in other places where these data structures aren't necessarily stored in the same positions in the inode. The refactoring proposed here is just the minimum required to get a proof-of-concept working with gmem. I would like to get opinions on this approach before doing further refactoring. (See TODOs) TODOs: + hugetlb/hugetlbfs refactoring + remove_inode_hugepages() no longer needs to be exposed, it is hugetlbfs specific and used only in inode.c + remove_mapping_hugepages(), remove_inode_single_folio(), hugetlb_unreserve_pages() shouldn't need to take inode as a parameter + Updating inode->i_blocks can be refactored to a separate function and called from hugetlbfs and gmem + alloc_hugetlb_folio_from_subpool() shouldn't need to be parametrized by vma + hugetlb_reserve_pages() should be refactored to be symmetric with hugetlb_unreserve_pages() + It should be parametrized by resv_map + alloc_hugetlb_folio_from_subpool() could perhaps use hugetlb_reserve_pages()? + gmem + Figure out if resv_map should be used by gmem at all + Probably needs more refactoring to decouple resv_map from hugetlb functions Questions for the community: 1. In this patchset, every gmem file backed with hugetlb is given a new subpool. Is that desirable? + In hugetlbfs, a subpool always belongs to a mount, and hugetlbfs has one mount per hugetlb size (2M, 1G, etc) + memfd_create(MFD_HUGETLB) effectively returns a full hugetlbfs file, so it (rightfully) uses the hugetlbfs kernel mounts and their subpools + I gave each file a subpool mostly to speed up implementation and still be able to reserve hugetlb pages from the global hstate based on the gmem file size. + gmem, unlike hugetlbfs, isn't meant to be a full filesystem, so + Should there be multiple mounts, one for each hugetlb size? + Will the mounts be initialized on boot or on first gmem file creation? + Or is one subpool per gmem file fine? 2. Should resv_map be used for gmem at all, since gmem doesn't allow userspace reservations? [1] https://lore.kernel.org/lkml/zem5zq8oo+xna...@google.com/ --- Ackerley Tng (19): mm: hugetlb: Expose get_hstate_idx() mm: hugetlb: Move and expose hugetlbfs_zero_partial_page mm: hugetlb: Expose remove_inode_hugepages mm: hugetlb: Decouple hstate, subpool from inode mm: hugetlb: Allow alloc_hugetlb_folio() to be parametrized by subpool and hstate mm: hugetlb: Provide hugetlb_filemap_add_folio() mm: hugetlb: Refactor vma_*_reservation functions mm: hugetlb: Refactor restore_reserve_on_error mm: hugetlb: Use restore_reserve_on_error directly in filesystems mm: hugetlb: Parametrize alloc_hugetlb_folio_from_subpool() by resv_map mm: hugetlb: Parametrize hugetlb functions by resv_map mm: truncate: Expose preparation steps for truncate_inode_pages_final KVM: guest_mem: Refactor kvm_gmem fd creation to be in layers KVM: guest_mem: Refactor cleanup to separate inode and file cleanup KVM:
