Re: [Bug report] Memory leak in scmi_device_create

[Catalin Marinas <catalin.marinas@arm.com>]

On Thu, Mar 06, 2025 at 11:09:33AM +0000, Alice Ryhl wrote:
> On Wed, Mar 05, 2025 at 05:10:16PM +0000, Cristian Marussi wrote:
> > On Wed, Mar 05, 2025 at 11:59:58AM +0000, Alice Ryhl wrote:
> > > This was with a kernel running v6.13-rc3, but as far as I can tell, no
> > > relevant changes have landed since v6.13-rc3. My tree *does* include
> > > commit 295416091e44 ("firmware: arm_scmi: Fix slab-use-after-free in
> > > scmi_bus_notifier()"). I've only seen this kmemleak report once, so it's
> > > not happening consistently.
> > > 
> > > See below for the full kmemleak report.
> > > 
> > > Alice
> > > 
> > > $ sudo cat /sys/kernel/debug/kmemleak
> > > unreferenced object 0xffffff8106c86000 (size 2048):
> > >   comm "swapper/0", pid 1, jiffies 4294893094
> > >   hex dump (first 32 bytes):
> > >     02 00 00 00 10 00 00 00 c0 01 bc 03 81 ff ff ff  ................
> > >     60 67 ba 03 81 ff ff ff 18 60 c8 06 81 ff ff ff  `g.......`......
> > >   backtrace (crc feae9680):
> > >     [<00000000197aa008>] kmemleak_alloc+0x34/0xa0
> > >     [<0000000056fe02c9>] __kmalloc_cache_noprof+0x1e0/0x450
> > >     [<00000000a8b3dfe1>] __scmi_device_create+0xb4/0x2b4
> > >     [<000000008714917b>] scmi_device_create+0x40/0x194
> > >     [<000000001818f3cf>] scmi_chan_setup+0x144/0x3b8
> > >     [<00000000970bad38>] scmi_probe+0x584/0xa78
> > >     [<000000002600d2fd>] platform_probe+0xbc/0xf0
> > >     [<00000000f6f556b4>] really_probe+0x1b8/0x520
> > >     [<00000000eed93d59>] __driver_probe_device+0xe0/0x1d8
> > >     [<00000000d613b754>] driver_probe_device+0x6c/0x208
> > >     [<00000000187a9170>] __driver_attach+0x168/0x328
> > >     [<00000000e3ff1834>] bus_for_each_dev+0x14c/0x178
> > >     [<00000000984a3176>] driver_attach+0x34/0x44
> > >     [<00000000fc35bf2a>] bus_add_driver+0x1bc/0x358
> > >     [<00000000747fce19>] driver_register+0xc0/0x1a0
> > >     [<0000000081cb8754>] __platform_driver_register+0x40/0x50
> > > unreferenced object 0xffffff8103bc01c0 (size 32):
> > 
> > I could not reproduce on my setup, even though I run a system with
> > all the existent SCMI protocols (and related drivers) enabled (and
> > so a lot of device creations) and a downstream test driver that causes
> > even more SCMI devices to be created/destroyed at load/unload.
> > 
> > Coming down the path from scmi_chan_setup(), it seems something around
> > transport devices creation, but it is not obvious to me where the leak
> > could hide....
> > 
> > ...any particular setup on your side ? ...using LKMs, loading/unloading,
> > any usage pattern that could help me reproduce ?
> 
> I looked into this a bit more, and actually it does happen consistently.
> It's just that kmemleak doesn't report it until 10 minutes after
> booting, so I did not notice it.

You can force the scanning with:

  echo scan > /sys/kernel/debug/kmemleak

Just do it a couple of times after boot, no need to wait 10 min for the
default background scanning.

> user@rk3588-ci:~$ sudo cat /sys/kernel/debug/kmemleak
> unreferenced object 0xffffff81068c0000 (size 2048):
>   comm "swapper/0", pid 1, jiffies 4294893128
>   hex dump (first 32 bytes):
>     02 00 00 00 10 00 00 00 40 a3 7a 03 81 ff ff ff  ........@.z.....
>     60 c8 79 03 81 ff ff ff 18 00 8c 06 81 ff ff ff  `.y.............
>   backtrace (crc 60df30fb):
>     kmemleak_alloc+0x34/0xa0
>     __kmalloc_cache_noprof+0x1e0/0x450
>     __scmi_device_create+0xb4/0x2b4

Is this the kzalloc() for sizeof(*scmi_dev)? It's surprisingly large, I
thought it would go for the kmalloc-1k slab as struct device is below
this side, at least for my builds. Anyway...

>     scmi_device_create+0x40/0x194
>     scmi_chan_setup+0x144/0x3b8
>     scmi_probe+0x51c/0x9fc
>     platform_probe+0xbc/0xf0
>     really_probe+0x1b8/0x520
>     __driver_probe_device+0xe0/0x1d8
>     driver_probe_device+0x6c/0x208
>     __driver_attach+0x168/0x328
>     bus_for_each_dev+0x14c/0x178
>     driver_attach+0x34/0x44
>     bus_add_driver+0x1bc/0x358
>     driver_register+0xc0/0x1a0
>     __platform_driver_register+0x40/0x50
> unreferenced object 0xffffff81037aa340 (size 32):
>   comm "swapper/0", pid 1, jiffies 4294893128
>   hex dump (first 32 bytes):
>     5f 5f 73 63 6d 69 5f 74 72 61 6e 73 70 6f 72 74  __scmi_transport
>     5f 64 65 76 69 63 65 5f 72 78 5f 31 30 00 ff ff  _device_rx_10...
>   backtrace (crc 8dab7ca7):
>     kmemleak_alloc+0x34/0xa0
>     __kmalloc_node_track_caller_noprof+0x234/0x528
>     kstrdup+0x48/0x80
>     kstrdup_const+0x30/0x3c

These are referenced from the main structure above, so they'd be
reported as leaks as well.

This loop in scmi_device_create() looks strange:

	list_for_each_entry(rdev, phead, node) {
		struct scmi_device *sdev;

		sdev = __scmi_device_create(np, parent,
					    rdev->id_table->protocol_id,
					    rdev->id_table->name);
		/* Report errors and carry on... */
		if (sdev)
			scmi_dev = sdev;
		else
			pr_err("(%s) Failed to create device for protocol 0x%x (%s)\n",
			       of_node_full_name(parent->of_node),
			       rdev->id_table->protocol_id,
			       rdev->id_table->name);
	}

We can override scmi_dev a few times in the loop and lose the previous
sdev allocations. Is this intended?

-- 
Catalin