[PATCH] riscv: mm: Pre-allocate PGD entries vmalloc/modules area

[Palmer Dabbelt]

On Mon, 29 May 2023 11:00:23 PDT (-0700), bjorn at kernel.org wrote:
> From: Björn Töpel <bjorn at rivosinc.com>
>
> The RISC-V port requires that kernel PGD entries are to be
> synchronized between MMs. This is done via the vmalloc_fault()
> function, that simply copies the PGD entries from init_mm to the
> faulting one.
>
> Historically, faulting in PGD entries have been a source for both bugs
> [1], and poor performance.
>
> One way to get rid of vmalloc faults is by pre-allocating the PGD
> entries. Pre-allocating the entries potientially wastes 64 * 4K (65 on
> SV39). The pre-allocation function is pulled from Jörg Rödel's x86
> work, with the addition of 3-level page tables (PMD allocations).
>
> The pmd_alloc() function needs the ptlock cache to be initialized
> (when split page locks is enabled), so the pre-allocation is done in a
> RISC-V specific pgtable_cache_init() implementation.
>
> Pre-allocate the kernel PGD entries for the vmalloc/modules area, but
> only for 64b platforms.
>
> Link: https://lore.kernel.org/lkml/20200508144043.13893-1-joro@8bytes.org/ # [1]
> Signed-off-by: Björn Töpel <bjorn at rivosinc.com>
> ---
>  arch/riscv/mm/fault.c | 20 +++------------
>  arch/riscv/mm/init.c  | 58 +++++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 62 insertions(+), 16 deletions(-)
>
> diff --git a/arch/riscv/mm/fault.c b/arch/riscv/mm/fault.c
> index 8685f85a7474..6b0b5e517e12 100644
> --- a/arch/riscv/mm/fault.c
> +++ b/arch/riscv/mm/fault.c
> @@ -230,32 +230,20 @@ void handle_page_fault(struct pt_regs *regs)
>  		return;
>
>  	/*
> -	 * Fault-in kernel-space virtual memory on-demand.
> -	 * The 'reference' page table is init_mm.pgd.
> +	 * Fault-in kernel-space virtual memory on-demand, for 32-bit
> +	 * architectures.  The 'reference' page table is init_mm.pgd.

That wording seems a little odd to me: I think English allows for these 
"add something after the comma to change the meaning of a sentence" 
things, but they're kind of complicated.  Maybe it's easier to just flip 
the order?

That said, it's very early so maybe it's fine...

>  	 *
>  	 * NOTE! We MUST NOT take any locks for this case. We may
>  	 * be in an interrupt or a critical region, and should
>  	 * only copy the information from the master page table,
>  	 * nothing more.
>  	 */
> -	if (unlikely((addr >= VMALLOC_START) && (addr < VMALLOC_END))) {
> +	if (!IS_ENABLED(CONFIG_64BIT) &&
> +	    unlikely(addr >= VMALLOC_START && addr < VMALLOC_END)) {
>  		vmalloc_fault(regs, code, addr);
>  		return;
>  	}
>
> -#ifdef CONFIG_64BIT
> -	/*
> -	 * Modules in 64bit kernels lie in their own virtual region which is not
> -	 * in the vmalloc region, but dealing with page faults in this region
> -	 * or the vmalloc region amounts to doing the same thing: checking that
> -	 * the mapping exists in init_mm.pgd and updating user page table, so
> -	 * just use vmalloc_fault.
> -	 */
> -	if (unlikely(addr >= MODULES_VADDR && addr < MODULES_END)) {
> -		vmalloc_fault(regs, code, addr);
> -		return;
> -	}
> -#endif
>  	/* Enable interrupts if they were enabled in the parent context. */
>  	if (!regs_irqs_disabled(regs))
>  		local_irq_enable();
> diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
> index 747e5b1ef02d..38bd4dd95276 100644
> --- a/arch/riscv/mm/init.c
> +++ b/arch/riscv/mm/init.c
> @@ -1363,3 +1363,61 @@ int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
>  	return vmemmap_populate_basepages(start, end, node, NULL);
>  }
>  #endif
> +
> +#ifdef CONFIG_64BIT
> +/*
> + * Pre-allocates page-table pages for a specific area in the kernel
> + * page-table. Only the level which needs to be synchronized between
> + * all page-tables is allocated because the synchronization can be
> + * expensive.
> + */
> +static void __init preallocate_pgd_pages_range(unsigned long start, unsigned long end,
> +					       const char *area)
> +{
> +	unsigned long addr;
> +	const char *lvl;
> +
> +	for (addr = start; addr < end && addr >= start; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
> +		pgd_t *pgd = pgd_offset_k(addr);
> +		p4d_t *p4d;
> +		pud_t *pud;
> +		pmd_t *pmd;
> +
> +		lvl = "p4d";
> +		p4d = p4d_alloc(&init_mm, pgd, addr);
> +		if (!p4d)
> +			goto failed;
> +
> +		if (pgtable_l5_enabled)
> +			continue;
> +
> +		lvl = "pud";
> +		pud = pud_alloc(&init_mm, p4d, addr);
> +		if (!pud)
> +			goto failed;
> +
> +		if (pgtable_l4_enabled)
> +			continue;
> +
> +		lvl = "pmd";
> +		pmd = pmd_alloc(&init_mm, pud, addr);
> +		if (!pmd)
> +			goto failed;
> +	}
> +	return;
> +
> +failed:
> +	/*
> +	 * The pages have to be there now or they will be missing in
> +	 * process page-tables later.
> +	 */
> +	panic("Failed to pre-allocate %s pages for %s area\n", lvl, area);
> +}
> +
> +void __init pgtable_cache_init(void)
> +{
> +	preallocate_pgd_pages_range(VMALLOC_START, VMALLOC_END, "vmalloc");
> +	if (IS_ENABLED(CONFIG_MODULES))
> +		preallocate_pgd_pages_range(MODULES_VADDR, MODULES_END, "bpf/modules");
> +}
> +#endif
>
> base-commit: ac9a78681b921877518763ba0e89202254349d1b

Reviewed-by: Palmer Dabbelt <palmer at rivosinc.com>

aside from the build issue, which seems pretty straight-forward.  I'm 
going to drop this from patchwork.