Re: [PATCH v2 2/2] mm: prevent gup_fast from racing with COW during fork

[Peter Xu <peterx@redhat.com>]

Hi, Jason,

I think majorly the patch looks good to me, but I have a few pure questions
majorly not directly related to the patch itself, but around the contexts.
Since I _feel_ like there'll be a new version to update the comments below,
maybe I can still ask aloud... Please bare with me. :)

On Fri, Oct 30, 2020 at 11:46:21AM -0300, Jason Gunthorpe wrote:
> Slow GUP is safe against this race because copy_page_range() is only
> called while holding the exclusive side of the mmap_lock on the src
> mm_struct.

Pure question: I understand that this patch requires this, but... Could anyone
remind me why read lock of mmap_sem is not enough for fork() before this one?

> 
> Fixes: f3c64eda3e50 ("mm: avoid early COW write protect games during fork()")
> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
> Link: https://lore.kernel.org/r/CAHk-=wi=iCnYCARbPGjkVJu9eyYeZ13N64tZYLdOB8CP5Q_PLw@mail.gmail.com
> Reviewed-by: John Hubbard <jhubbard@nvidia.com>
> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
> ---
>  arch/x86/kernel/tboot.c    |  1 +
>  drivers/firmware/efi/efi.c |  1 +
>  include/linux/mm_types.h   |  7 +++++++
>  kernel/fork.c              |  1 +
>  mm/gup.c                   | 19 +++++++++++++++++++
>  mm/init-mm.c               |  1 +
>  mm/memory.c                | 10 +++++++++-
>  7 files changed, 39 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/x86/kernel/tboot.c b/arch/x86/kernel/tboot.c
> index 992fb1415c0f1f..6a2f542d9588a4 100644
> --- a/arch/x86/kernel/tboot.c
> +++ b/arch/x86/kernel/tboot.c
> @@ -93,6 +93,7 @@ static struct mm_struct tboot_mm = {
>  	.pgd            = swapper_pg_dir,
>  	.mm_users       = ATOMIC_INIT(2),
>  	.mm_count       = ATOMIC_INIT(1),
> +	.write_protect_seq = SEQCNT_ZERO(tboot_mm.write_protect_seq),
>  	MMAP_LOCK_INITIALIZER(init_mm)
>  	.page_table_lock =  __SPIN_LOCK_UNLOCKED(init_mm.page_table_lock),
>  	.mmlist         = LIST_HEAD_INIT(init_mm.mmlist),
> diff --git a/drivers/firmware/efi/efi.c b/drivers/firmware/efi/efi.c
> index 5e5480a0a32d7d..2520f6e05f4d44 100644
> --- a/drivers/firmware/efi/efi.c
> +++ b/drivers/firmware/efi/efi.c
> @@ -57,6 +57,7 @@ struct mm_struct efi_mm = {
>  	.mm_rb			= RB_ROOT,
>  	.mm_users		= ATOMIC_INIT(2),
>  	.mm_count		= ATOMIC_INIT(1),
> +	.write_protect_seq      = SEQCNT_ZERO(efi_mm.write_protect_seq),
>  	MMAP_LOCK_INITIALIZER(efi_mm)
>  	.page_table_lock	= __SPIN_LOCK_UNLOCKED(efi_mm.page_table_lock),
>  	.mmlist			= LIST_HEAD_INIT(efi_mm.mmlist),

Another pure question: I'm just curious how you find all the statically
definied mm_structs, and to make sure all of them are covered (just in case
un-initialized seqcount could fail strangely).

Actually I'm thinking whether we should have one place to keep all the init
vars for all the statically definied mm_structs, so we don't need to find them
everytime, but only change that one place.

> diff --git a/mm/memory.c b/mm/memory.c
> index c48f8df6e50268..294c2c3c4fe00d 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -1171,6 +1171,12 @@ copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
>  		mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
>  					0, src_vma, src_mm, addr, end);
>  		mmu_notifier_invalidate_range_start(&range);
> +		/*
> +		 * The read side doesn't spin, it goes to the mmap_lock, so the
> +		 * raw version is used to avoid disabling preemption here
> +		 */
> +		mmap_assert_write_locked(src_mm);
> +		raw_write_seqcount_t_begin(&src_mm->write_protect_seq);

Would raw_write_seqcount_begin() be better here?

My understanding is that we used raw_write_seqcount_t_begin() because we're
with spin lock so assuming we disabled preemption already.  However I'm
thinking whether raw_write_seqcount_begin() would be even better to guarantee
that.  I have no idea of how the rt kernel merging topic, but if rt kernel
merged into mainline then IIUC preemption is allowed here (since pgtable spin
lock should be rt_spin_lock, not raw spin locks).

An even further pure question on __seqcount_preemptible() (feel free to ignore
this question!): I saw that __seqcount_preemptible() seems to have been
constantly defined as "return false".  Not sure what happened there..

Thanks,

-- 
Peter Xu