Re: [PATCH bpf-next v5 1/7] mm, bpf: Introduce try_alloc_pages() for opportunistic page allocation

[Sebastian Andrzej Siewior <bigeasy@linutronix.de>]

On 2025-01-14 18:17:40 [-0800], Alexei Starovoitov wrote:
> From: Alexei Starovoitov <ast@kernel.org>
> 
> Tracing BPF programs execute from tracepoints and kprobes where
> running context is unknown, but they need to request additional
> memory. The prior workarounds were using pre-allocated memory and
> BPF specific freelists to satisfy such allocation requests.
> Instead, introduce gfpflags_allow_spinning() condition that signals
> to the allocator that running context is unknown.
> Then rely on percpu free list of pages to allocate a page.
> try_alloc_pages() -> get_page_from_freelist() -> rmqueue() ->
> rmqueue_pcplist() will spin_trylock to grab the page from percpu
> free list. If it fails (due to re-entrancy or list being empty)
> then rmqueue_bulk()/rmqueue_buddy() will attempt to
> spin_trylock zone->lock and grab the page from there.
> spin_trylock() is not safe in RT when in NMI or in hard IRQ.
> Bailout early in such case.
> 
> The support for gfpflags_allow_spinning() mode for free_page and memcg
> comes in the next patches.
> 
> This is a first step towards supporting BPF requirements in SLUB
> and getting rid of bpf_mem_alloc.
> That goal was discussed at LSFMM: https://lwn.net/Articles/974138/
> 
> Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>

could you…

> Signed-off-by: Alexei Starovoitov <ast@kernel.org>
…
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 1cb4b8c8886d..74c2a7af1a77 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -7023,3 +7032,86 @@ static bool __free_unaccepted(struct page *page)
>  }
>  
>  #endif /* CONFIG_UNACCEPTED_MEMORY */
> +
> +/**
> + * try_alloc_pages_noprof - opportunistic reentrant allocation from any context
> + * @nid - node to allocate from
> + * @order - allocation order size
> + *
> + * Allocates pages of a given order from the given node. This is safe to
> + * call from any context (from atomic, NMI, and also reentrant
> + * allocator -> tracepoint -> try_alloc_pages_noprof).
> + * Allocation is best effort and to be expected to fail easily so nobody should
> + * rely on the success. Failures are not reported via warn_alloc().

Could you maybe add a pointer like "See AlwaysFailRestrictions below."
or something similar to make the user aware of the comment below where
certain always-fail restrictions are mentioned. Such as PREEMPT_RT + NMI
or deferred_pages_enabled().
It might not be easy to be aware of this.

I'm curious how this turns out in the long run :)

> + *
> + * Return: allocated page or NULL on failure.
> + */
> +struct page *try_alloc_pages_noprof(int nid, unsigned int order)
> +{
> +	/*
> +	 * Do not specify __GFP_DIRECT_RECLAIM, since direct claim is not allowed.
> +	 * Do not specify __GFP_KSWAPD_RECLAIM either, since wake up of kswapd
> +	 * is not safe in arbitrary context.
> +	 *
> +	 * These two are the conditions for gfpflags_allow_spinning() being true.
> +	 *
> +	 * Specify __GFP_NOWARN since failing try_alloc_pages() is not a reason
> +	 * to warn. Also warn would trigger printk() which is unsafe from
> +	 * various contexts. We cannot use printk_deferred_enter() to mitigate,
> +	 * since the running context is unknown.
> +	 *
> +	 * Specify __GFP_ZERO to make sure that call to kmsan_alloc_page() below
> +	 * is safe in any context. Also zeroing the page is mandatory for
> +	 * BPF use cases.
> +	 *
> +	 * Though __GFP_NOMEMALLOC is not checked in the code path below,
> +	 * specify it here to highlight that try_alloc_pages()
> +	 * doesn't want to deplete reserves.
> +	 */
> +	gfp_t alloc_gfp = __GFP_NOWARN | __GFP_ZERO | __GFP_NOMEMALLOC;
> +	unsigned int alloc_flags = ALLOC_TRYLOCK;
> +	struct alloc_context ac = { };
> +	struct page *page;
> +
> +	/*
> +	 * In RT spin_trylock() may call raw_spin_lock() which is unsafe in NMI.
PREEMPT_RT please. s/may/will

> +	 * If spin_trylock() is called from hard IRQ the current task may be
> +	 * waiting for one rt_spin_lock, but rt_spin_trylock() will mark the
> +	 * task as the owner of another rt_spin_lock which will confuse PI
> +	 * logic, so return immediately if called form hard IRQ or NMI.
> +	 *
> +	 * Note, irqs_disabled() case is ok. This function can be called
> +	 * from raw_spin_lock_irqsave region.
> +	 */
> +	if (IS_ENABLED(CONFIG_PREEMPT_RT) && (in_nmi() || in_hardirq()))
> +		return NULL;
> +	if (!pcp_allowed_order(order))
> +		return NULL;
…

Sebastian