Re: [RFC PATCH v2 1/3] x86/mm/tlb: Change __flush_tlb_one_user interface

[Juergen Gross]

On 24.08.19 00:52, Nadav Amit wrote:
> __flush_tlb_one_user() currently flushes a single entry, and flushes it
> both in the kernel and user page-tables, when PTI is enabled.
>
> Change __flush_tlb_one_user() and related interfaces into
> __flush_tlb_range() that flushes a range and does not flush the user
> page-table.
>
> This refactoring is needed for the next patch, but regardless makes
> sense and has several advantages. First, only Xen-PV, which does not
> use PTI, implements the paravirtual interface of flush_tlb_one_user() so
> nothing is broken by separating the user and kernel page-table flushes,
> and the interface is more intuitive.
>
> Second, INVLPG can flush unrelated mappings, and it is also a
> serializing instruction. It is better to have a tight loop that flushes
> the entries.
>
> Third, currently __flush_tlb_one_kernel() also flushes the user
> page-tables, which is not needed. This allows to avoid this redundant
> flush.
>
> Cc: Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>
> Cc: Juergen Gross <jgross@xxxxxxxx>
> Cc: Stefano Stabellini <sstabellini@xxxxxxxxxx>
> Cc: xen-devel@xxxxxxxxxxxxxxxxxxxx
> Signed-off-by: Nadav Amit <namit@xxxxxxxxxx>
> ---
>   arch/x86/include/asm/paravirt.h       |  5 ++--
>   arch/x86/include/asm/paravirt_types.h |  3 ++-
>   arch/x86/include/asm/tlbflush.h       | 24 +++++------------
>   arch/x86/kernel/paravirt.c            |  7 ++---
>   arch/x86/mm/tlb.c                     | 39 ++++++++++++++++++++++-----
>   arch/x86/xen/mmu_pv.c                 | 21 +++++++++------
>   6 files changed, 62 insertions(+), 37 deletions(-)

...

> diff --git a/arch/x86/xen/mmu_pv.c b/arch/x86/xen/mmu_pv.c
> index 48f7c7eb4dbc..ed68657f5e77 100644
> --- a/arch/x86/xen/mmu_pv.c
> +++ b/arch/x86/xen/mmu_pv.c
> @@ -1325,22 +1325,27 @@ static noinline void xen_flush_tlb(void)
>   	preempt_enable();
>   }
>   
> -static void xen_flush_tlb_one_user(unsigned long addr)
> +static void xen_flush_tlb_range(unsigned long start, unsigned long end,
> +				u8 stride_shift)
>   {
>   	struct mmuext_op *op;
>   	struct multicall_space mcs;
> -
> -	trace_xen_mmu_flush_tlb_one_user(addr);
> +	unsigned long addr;
>   
>   	preempt_disable();
>   
>   	mcs = xen_mc_entry(sizeof(*op));
>   	op = mcs.args;
> -	op->cmd = MMUEXT_INVLPG_LOCAL;
> -	op->arg1.linear_addr = addr & PAGE_MASK;
> -	MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
>   
> -	xen_mc_issue(PARAVIRT_LAZY_MMU);
> +	for (addr = start; addr < end; addr += 1ul << stride_shift) {
> +		trace_xen_mmu_flush_tlb_one_user(addr);
> +
> +		op->cmd = MMUEXT_INVLPG_LOCAL;
> +		op->arg1.linear_addr = addr & PAGE_MASK;
> +		MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
> +
> +		xen_mc_issue(PARAVIRT_LAZY_MMU);
> +	}

For this kind of usage (a loop) you should:

- replace the call of xen_mc_entry() with xen_mc_batch()
- use xen_extend_mmuext_op() for each loop iteration
- call xen_mc_issue() after the loop

Additionally I'd like you to replace trace_xen_mmu_flush_tlb_one_user()
with trace_xen_mmu_flush_tlb_range() taking all three parameters and
keep it where it was (out of the loop).

The paravirt parts seem to be okay.


Juergen