Re: [PATCH v9 bpf-next 1/9] x86/Kconfig: select HAVE_ARCH_HUGE_VMALLOC with HAVE_ARCH_HUGE_VMAP

From: "Edgecombe, Rick P" <rick.p.edgecombe@intel.com>
To: "songliubraving@fb.com" <songliubraving@fb.com>
Cc: "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"daniel@iogearbox.net" <daniel@iogearbox.net>,
	"peterz@infradead.org" <peterz@infradead.org>,
	"ast@kernel.org" <ast@kernel.org>,
	"bpf@vger.kernel.org" <bpf@vger.kernel.org>,
	"Kernel-team@fb.com" <Kernel-team@fb.com>,
	"song@kernel.org" <song@kernel.org>, "hch@lst.de" <hch@lst.de>,
	"pmenzel@molgen.mpg.de" <pmenzel@molgen.mpg.de>,
	"andrii@kernel.org" <andrii@kernel.org>,
	"x86@kernel.org" <x86@kernel.org>,
	"iii@linux.ibm.com" <iii@linux.ibm.com>,
	"netdev@vger.kernel.org" <netdev@vger.kernel.org>,
	"urezki@gmail.com" <urezki@gmail.com>,
	"npiggin@gmail.com" <npiggin@gmail.com>
Subject: Re: [PATCH v9 bpf-next 1/9] x86/Kconfig: select HAVE_ARCH_HUGE_VMALLOC with HAVE_ARCH_HUGE_VMAP
Date: Tue, 29 Mar 2022 18:39:49 +0000	[thread overview]
Message-ID: <3ecfbf80feff3487cbb26b492375cef5a5fe8ac4.camel@intel.com> (raw)
In-Reply-To: <6080EC28-E3FE-4B00-B94A-ED7EBA1F55ED@fb.com>

CC some vmalloc folks. What happened was vmalloc huge pages was turned
on for x86 with the rest of the kernel unprepared and problems have
popped up. We are discussing a possible new config and vmap flag such
that for some arch's, huge pages would only be used for certain
allocations such as BPF's.

Thread:

https://lore.kernel.org/lkml/6080EC28-E3FE-4B00-B94A-ED7EBA1F55ED@fb.com/

On Tue, 2022-03-29 at 08:23 +0000, Song Liu wrote:
> > On Mar 28, 2022, at 5:18 PM, Edgecombe, Rick P <
> > rick.p.edgecombe@intel.com> wrote:
> > 
> > On Mon, 2022-03-28 at 23:27 +0000, Song Liu wrote:
> > > I like this direction. But I am afraid this is not enough. Using
> > > VM_NO_HUGE_VMAP in module_alloc() will make sure we don't
> > > allocate 
> > > huge pages for modules. But other users of
> > > __vmalloc_node_range(), 
> > > such as vzalloc in Paul's report, may still hit the issue. 
> > > 
> > > Maybe we need another flag VM_FORCE_HUGE_VMAP that bypasses 
> > > vmap_allow_huge check. Something like the diff below.
> > > 
> > > Would this work?
> > 
> > Yea, that looks like a safer direction. It's too bad we can't have
> > automatic large pages, but it doesn't seem ready to just turn on
> > for
> > the whole x86 kernel.
> > 
> > I'm not sure about this implementation though. It would let large
> > pages
> > get enabled without HAVE_ARCH_HUGE_VMALLOC and also despite the
> > disable
> > kernel parameter.
> > 
> > Apparently some architectures can handle large pages automatically
> > and
> > it has benefits for them, so maybe vmalloc should support both
> > behaviors based on config. Like there should a
> > ARCH_HUGE_VMALLOC_REQUIRE_FLAG config. If configured it requires
> > VM_HUGE_VMAP (or some name). I don't think FORCE fits, because the
> > current logic would not always give huge pages.
> > 
> > But yea, seems risky to leave it on generally, even if you could
> > fix
> > Paul's specific issue.
> > 
> 
> 
> How about something like the following? (I still need to fix
> something, but
> would like some feedbacks on the API).
> 
> Thanks,
> Song
> 
> 
> diff --git a/arch/Kconfig b/arch/Kconfig
> index 84bc1de02720..defef50ff527 100644
> --- a/arch/Kconfig
> +++ b/arch/Kconfig
> @@ -858,6 +858,15 @@ config HAVE_ARCH_HUGE_VMALLOC
>  	depends on HAVE_ARCH_HUGE_VMAP
>  	bool
>  
> +#
> +# HAVE_ARCH_HUGE_VMALLOC_FLAG allows users of __vmalloc_node_range
> to allocate
> +# huge page without HAVE_ARCH_HUGE_VMALLOC. To allocate huge pages,,
> the user
> +# need to call __vmalloc_node_range with VM_PREFER_HUGE_VMAP.
> +#
> +config HAVE_ARCH_HUGE_VMALLOC_FLAG
> +	depends on HAVE_ARCH_HUGE_VMAP
> +	bool
> +
>  config ARCH_WANT_HUGE_PMD_SHARE
>  	bool
>  
> diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
> index 7340d9f01b62..e64f00415575 100644
> --- a/arch/x86/Kconfig
> +++ b/arch/x86/Kconfig
> @@ -161,7 +161,7 @@ config X86
>  	select HAVE_ALIGNED_STRUCT_PAGE		if SLUB
>  	select HAVE_ARCH_AUDITSYSCALL
>  	select HAVE_ARCH_HUGE_VMAP		if X86_64 || X86_PAE
> -	select HAVE_ARCH_HUGE_VMALLOC		if X86_64
> +	select HAVE_ARCH_HUGE_VMALLOC_FLAG	if X86_64
>  	select HAVE_ARCH_JUMP_LABEL
>  	select HAVE_ARCH_JUMP_LABEL_RELATIVE
>  	select HAVE_ARCH_KASAN			if X86_64
> diff --git a/include/linux/vmalloc.h b/include/linux/vmalloc.h
> index 3b1df7da402d..98f8a93fcfd4 100644
> --- a/include/linux/vmalloc.h
> +++ b/include/linux/vmalloc.h
> @@ -35,6 +35,11 @@ struct notifier_block;		/* in
> notifier.h */
>  #define VM_DEFER_KMEMLEAK	0
>  #endif
>  
> +#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC_FLAG
> +#define VM_PREFER_HUGE_VMAP	0x00001000	/* prefer PMD_SIZE
> mapping (bypass vmap_allow_huge check) */

I don't think we should tie this to vmap_allow_huge by definition.
Also, what it does is try 2MB pages for allocations larger than 2MB.
For smaller allocations it doesn't try or "prefer" them.

> +#else
> +#define VM_PREFER_HUGE_VMAP	0
> +#endif
>  /* bits [20..32] reserved for arch specific ioremap internals */
>  
>  /*
> @@ -51,7 +56,7 @@ struct vm_struct {
>  	unsigned long		size;
>  	unsigned long		flags;
>  	struct page		**pages;
> -#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
> +#if (defined(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
> defined(CONFIG_HAVE_ARCH_HUGE_VMALLOC_FLAG))
>  	unsigned int		page_order;
>  #endif
>  	unsigned int		nr_pages;
> @@ -225,7 +230,7 @@ static inline bool is_vm_area_hugepages(const
> void *addr)
>  	 * prevents that. This only indicates the size of the physical
> page
>  	 * allocated in the vmalloc layer.
>  	 */
> -#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
> +#if (defined(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
> defined(CONFIG_HAVE_ARCH_HUGE_VMALLOC_FLAG))
>  	return find_vm_area(addr)->page_order > 0;
>  #else
>  	return false;
> diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
> index 13e9dbeeedf3..fc9dae095079 100644
> --- a/kernel/bpf/core.c
> +++ b/kernel/bpf/core.c
> @@ -851,13 +851,28 @@ static LIST_HEAD(pack_list);
>  #define BPF_HPAGE_MASK PAGE_MASK
>  #endif
>  
> +static void *bpf_prog_pack_vmalloc(unsigned long size)
> +{
> +#if defined(MODULES_VADDR)
> +	unsigned long start = MODULES_VADDR;
> +	unsigned long end = MODULES_END;
> +#else
> +	unsigned long start = VMALLOC_START;
> +	unsigned long end = VMALLOC_END;
> +#endif
> +
> +	return __vmalloc_node_range(size, PAGE_SIZE, start, end,
> GFP_KERNEL, PAGE_KERNEL,
> +				    VM_DEFER_KMEMLEAK |
> VM_PREFER_HUGE_VMAP,
> +				    NUMA_NO_NODE,
> __builtin_return_address(0));
> +}
> +
>  static size_t select_bpf_prog_pack_size(void)
>  {
>  	size_t size;
>  	void *ptr;
>  
>  	size = BPF_HPAGE_SIZE * num_online_nodes();
> -	ptr = module_alloc(size);
> +	ptr = bpf_prog_pack_vmalloc(size);
>  
>  	/* Test whether we can get huge pages. If not just use
> PAGE_SIZE
>  	 * packs.
> @@ -881,7 +896,7 @@ static struct bpf_prog_pack *alloc_new_pack(void)
>  		       GFP_KERNEL);
>  	if (!pack)
>  		return NULL;
> -	pack->ptr = module_alloc(bpf_prog_pack_size);
> +	pack->ptr = bpf_prog_pack_vmalloc(bpf_prog_pack_size);
>  	if (!pack->ptr) {
>  		kfree(pack);
>  		return NULL;
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index e163372d3967..9d3c1ab8bdf1 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -2261,7 +2261,7 @@ static inline unsigned int
> vm_area_page_order(struct vm_struct *vm)
>  
>  static inline void set_vm_area_page_order(struct vm_struct *vm,
> unsigned int order)
>  {
> -#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
> +#if (defined(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
> defined(CONFIG_HAVE_ARCH_HUGE_VMALLOC_FLAG))
>  	vm->page_order = order;
>  #else
>  	BUG_ON(order != 0);
> @@ -3106,7 +3106,8 @@ void *__vmalloc_node_range(unsigned long size,
> unsigned long align,
>  		return NULL;
>  	}
>  
> -	if (vmap_allow_huge && !(vm_flags & VM_NO_HUGE_VMAP)) {
> +	if ((vmap_allow_huge && !(vm_flags & VM_NO_HUGE_VMAP)) ||
> +	    (IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC_FLAG) &&
> (vm_flags & VM_PREFER_HUGE_VMAP))) {

vmap_allow_huge is how the kernel parameter that can disable vmalloc
huge pages works. I don't think we should separate it. Disabling
vmalloc huge pages should still disable this alternate mode.

>  		unsigned long size_per_node;
>  
>  		/*