linux-kernel - Re: mm/percpu.c: use smarter memory allocation for struct pcpu_alloc

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <nycvar.YSQ.7.76.1711201512300.16045@knanqh.ubzr>
Date:   Mon, 20 Nov 2017 15:21:32 -0500 (EST)
From:   Nicolas Pitre <nicolas.pitre@...aro.org>
To:     Guenter Roeck <linux@...ck-us.net>
cc:     Tejun Heo <tj@...nel.org>, Christoph Lameter <cl@...ux.com>,
        linux-mm@...ck.org, linux-kernel@...r.kernel.org,
        Mikael Starvik <starvik@...s.com>,
        Jesper Nilsson <jesper.nilsson@...s.com>,
        linux-cris-kernel@...s.com
Subject: Re: mm/percpu.c: use smarter memory allocation for struct pcpu_alloc_info
 (crisv32 hang)

On Mon, 20 Nov 2017, Guenter Roeck wrote:

> On Mon, Nov 20, 2017 at 01:18:38PM -0500, Nicolas Pitre wrote:
> > On Sun, 19 Nov 2017, Guenter Roeck wrote:
> > 
> > > On 11/19/2017 08:08 PM, Nicolas Pitre wrote:
> > > > On Sun, 19 Nov 2017, Guenter Roeck wrote:
> > > > > On 11/19/2017 12:36 PM, Nicolas Pitre wrote:
> > > > > > On Sat, 18 Nov 2017, Guenter Roeck wrote:
> > > > > > > On Tue, Oct 03, 2017 at 06:29:49PM -0400, Nicolas Pitre wrote:
> > > > > > > > @@ -2295,6 +2295,7 @@ void __init setup_per_cpu_areas(void)
> > > > > > > >      	if (pcpu_setup_first_chunk(ai, fc) < 0)
> > > > > > > >    		panic("Failed to initialize percpu areas.");
> > > > > > > > +	pcpu_free_alloc_info(ai);
> > > > > > > 
> > > > > > > This is the culprit. Everything works fine if I remove this line.
> > > > > > 
> > > > > > Without this line, the memory at the ai pointer is leaked. Maybe this is
> > > > > > modifying the memory allocation pattern and that triggers a bug later on
> > > > > > in your case.
> > > > > > 
> > > > > > At that point the console driver is not yet initialized and any error
> > > > > > message won't be printed. You should enable the early console mechanism
> > > > > > in your kernel (see arch/cris/arch-v32/kernel/debugport.c) and see what
> > > > > > that might tell you.
> > > > > > 
> > > > > 
> > > > > The problem is that BUG() on crisv32 does not yield useful output.
> > > > > Anyway, here is the culprit.
> > > > > 
> > > > > diff --git a/mm/bootmem.c b/mm/bootmem.c
> > > > > index 6aef64254203..2bcc8901450c 100644
> > > > > --- a/mm/bootmem.c
> > > > > +++ b/mm/bootmem.c
> > > > > @@ -382,7 +382,8 @@ static int __init mark_bootmem(unsigned long start,
> > > > > unsigned long end,
> > > > >                          return 0;
> > > > >                  pos = bdata->node_low_pfn;
> > > > >          }
> > > > > -       BUG();
> > > > > +       WARN(1, "mark_bootmem(): memory range 0x%lx-0x%lx not found\n",
> > > > > start,
> > > > > end);
> > > > > +       return -ENOMEM;
> > > > >   }
> > > > > 
> > > > >   /**
> > > > > diff --git a/mm/percpu.c b/mm/percpu.c
> > > > > index 79e3549cab0f..c75622d844f1 100644
> > > > > --- a/mm/percpu.c
> > > > > +++ b/mm/percpu.c
> > > > > @@ -1881,6 +1881,7 @@ struct pcpu_alloc_info * __init
> > > > > pcpu_alloc_alloc_info(int nr_groups,
> > > > >    */
> > > > >   void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
> > > > >   {
> > > > > +       printk("pcpu_free_alloc_info(%p (0x%lx))\n", ai, __pa(ai));
> > > > >          memblock_free_early(__pa(ai), ai->__ai_size);
> > > > 
> > > > The problem here is that there is two possibilities for
> > > > memblock_free_early(). From include/linux/bootmem.h:
> > > > 
> > > > #if defined(CONFIG_HAVE_MEMBLOCK) && defined(CONFIG_NO_BOOTMEM)
> > > > 
> > > > static inline void __init memblock_free_early(
> > > >                                          phys_addr_t base, phys_addr_t size)
> > > > {
> > > >          __memblock_free_early(base, size);
> > > > }
> > > > 
> > > > #else
> > > > 
> > > > static inline void __init memblock_free_early(
> > > >                                          phys_addr_t base, phys_addr_t size)
> > > > {
> > > >          free_bootmem(base, size);
> > > > }
> > > > 
> > > > #endif
> > > > 
> > > > It looks like most architectures use the memblock variant, including all
> > > > the ones I have access to.
> > > > 
> > > > > results in:
> > > > > 
> > > > > pcpu_free_alloc_info(c0534000 (0x40534000))
> > > > > ------------[ cut here ]------------
> > > > > WARNING: CPU: 0 PID: 0 at mm/bootmem.c:385 mark_bootmem+0x9a/0xaa
> > > > > mark_bootmem(): memory range 0x2029a-0x2029b not found
> > > > 
> > > > Well... PFN_UP(0x40534000) should give 0x40534. How you might end up
> > > > with 0x2029a in mark_bootmem(), let alone not exit on the first "if (max
> > > > == end) return 0;" within the loop is rather weird.
> > > > 
> > > pcpu_free_alloc_info: ai=c0536000, __pa(ai)=0x40536000,
> > > PFN_UP(__pa(ai))=0x2029b, PFN_UP(ai)=0x6029b
> > > 
> > > bootmem range is 0x60000..0x61000. It doesn't get to "if (max == end)"
> > > because "pos (=0x2029b) < bdata->node_min_pfn (=0x60000)".
> > 
> > OK. the 0x2029b is the result of PAGE_SIZE being 8192 in your case.
> > However the bootmem allocator deals with physical addresses not virtual 
> > ones. So it shouldn't give you a 0x60000..0x61000 range.
> > 
> > Would be interesting to see what result you get on line 860 of 
> > mm/bootmem.c.
> > 
> Nothing; __alloc_bootmem_low_node() is not called.
> 
> Call chain is:
>   pcpu_alloc_alloc_info
>     memblock_virt_alloc_nopanic
>       __alloc_bootmem_nopanic
>         ___alloc_bootmem_nopanic

But from there it should continue with: 

	alloc_bootmem_core() -->
	  alloc_bootmem_bdata() -->
	    [...]
	    region = phys_to_virt(PFN_PHYS(bdata->node_min_pfn) + start_off);

That's line 585, not 860 as I mentioned. Sorry for the confusion.


Nicolas