MIME-Version: 1.0
References: <cover.1675111415.git.andreyknvl@google.com> <9fbb4d2bf9b2676a29b120980b5ffbda8e2304ee.1675111415.git.andreyknvl@google.com>
In-Reply-To: <9fbb4d2bf9b2676a29b120980b5ffbda8e2304ee.1675111415.git.andreyknvl@google.com>
From:   Alexander Potapenko <glider@google.com>
Date:   Tue, 31 Jan 2023 10:29:51 +0100
Message-ID: <CAG_fn=VO0iO4+EuwDR0bKP-4om9_Afir3fY6CExKGRNad+uPLA@mail.gmail.com>
Subject: Re: [PATCH 01/18] lib/stackdepot: fix setting next_slab_inited in init_stack_slab
To:     andrey.konovalov@linux.dev
Cc:     Marco Elver <elver@google.com>,
        Andrey Konovalov <andreyknvl@gmail.com>,
        Vlastimil Babka <vbabka@suse.cz>, kasan-dev@googlegroups.com,
        Evgenii Stepanov <eugenis@google.com>,
        Andrew Morton <akpm@linux-foundation.org>, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org,
        Andrey Konovalov <andreyknvl@google.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Mon, Jan 30, 2023 at 9:49 PM <andrey.konovalov@linux.dev> wrote:
>
> From: Andrey Konovalov <andreyknvl@google.com>
>
> In commit 305e519ce48e ("lib/stackdepot.c: fix global out-of-bounds in
> stack_slabs"), init_stack_slab was changed to only use preallocated
> memory for the next slab if the slab number limit is not reached.
> However, setting next_slab_inited was not moved together with updating
> stack_slabs.
>
> Set next_slab_inited only if the preallocated memory was used for the
> next slab.
>
> Fixes: 305e519ce48e ("lib/stackdepot.c: fix global out-of-bounds in stack_slabs")
> Signed-off-by: Andrey Konovalov <andreyknvl@google.com>

Wait, I think there's a problem here.

> diff --git a/lib/stackdepot.c b/lib/stackdepot.c
> index 79e894cf8406..0eed9bbcf23e 100644
> --- a/lib/stackdepot.c
> +++ b/lib/stackdepot.c
> @@ -105,12 +105,13 @@ static bool init_stack_slab(void **prealloc)
>                 if (depot_index + 1 < STACK_ALLOC_MAX_SLABS) {
If we get to this branch, but the condition is false, this means that:
 - next_slab_inited == 0
 - depot_index == STACK_ALLOC_MAX_SLABS+1
 - stack_slabs[depot_index] != NULL.

So stack_slabs[] is at full capacity, but upon leaving
init_stack_slab() we'll always keep next_slab_inited==0.

Now every time __stack_depot_save() is called for a known stack trace,
it will preallocate 1<<STACK_ALLOC_ORDER pages (because
next_slab_inited==0), then find the stack trace id in the hash, then
pass the preallocated pages to init_stack_slab(), which will not
change the value of next_slab_inited.
Then the preallocated pages will be freed, and next time
__stack_depot_save() is called they'll be allocated again.