linuxnext-2019119 edac warns (was Re: edac KASAN warning in experimen…

…tal arm64 allmodconfig boot)

Hi John,

thanks for testing and reporting this. See inline.

On 21.11.19 12:34:22, John Garry wrote:
> On 14/10/2019 16:18, John Garry wrote:
> JFYI, I see an issue on linuxnext-2019119, as follows:
>
>    21.645388] io scheduler kyber registered
> [   21.734011] input: Power Button as
> /devices/LNXSYSTM:00/LNXSYBUS:00/PNP0C0C:00/input/input0
> [   21.743295] ACPI: Power Button [PWRB]
> [   21.809644] [Firmware Bug]: APEI: Invalid bit width + offset in GAR
> [0x94110034/64/0/3/0]
> [   21.821974] EDAC MC0: Giving out device to module ghes_edac.c controller
> ghes_edac: DEV ghes (INTERRUPT)
> [   21.831763] ------------[ cut here ]------------
> [   21.836374] refcount_t: increment on 0; use-after-free.
> [   21.841620] WARNING: CPU: 36 PID: 1 at lib/refcount.c:156
> refcount_inc_checked+0x44/0x50
> [   21.849697] Modules linked in:
> [   21.852745] CPU: 36 PID: 1 Comm: swapper/0 Not tainted
> 5.4.0-rc8-next-20191119-00003-g141a9fef5092-dirty torvalds#650
> [   21.862645] Hardware name: Huawei D06 /D06, BIOS Hisilicon D06 UEFI RC0 -
> V1.16.01 03/15/2019
> [   21.871157] pstate: 60c00009 (nZCv daif +PAN +UAO)
> [   21.875936] pc : refcount_inc_checked+0x44/0x50
> [   21.880455] lr : refcount_inc_checked+0x44/0x50

This is a warning from the refcount framework. It warns if we
increment from zero. This is reasonable as typically a kernel object
is created with a refcount of 1 and thrown away once the refcount is
zero. Afterwards the object is used-after-free.

For ghes the refcount is initialized with zero, and that is why we see
this message. However, we protect the refcount with the ghes_reg_mutex
and thus there is no use after free. The device is allocated and
registered if the refcount is zero. So this works fine.

Enclosed a fix that avoids the warning, please test.

But see below...

> [   21.884972] sp : ffff00236ffbf8a0
> [   21.888274] x29: ffff00236ffbf8a0 x28: 0000000000000002
> [   21.893576] x27: ffff00236cd07900 x26: ffff002369063010
> [   21.898876] x25: 0000000000000000 x24: ffff00233c236824
> [   21.904177] x23: ffffa000137b9000 x22: ffffa00016fbb7c0
> [   21.909477] x21: ffffa00012dfd000 x20: 1fffe0046dff7f24
> [   21.914777] x19: ffff00233c236000 x18: 0000000000000000
> [   21.920077] x17: 0000000000000000 x16: 0000000000000000
> [   21.925377] x15: 0000000000007700 x14: 64655f7365686720
> [   21.930677] x13: 72656c6c6f72746e x12: 1ffff40002719618
> [   21.935977] x11: ffff940002719618 x10: dfffa00000000000
> [   21.941278] x9 : ffff940002719619 x8 : 0000000000000001
> [   21.946578] x7 : 0000000000000000 x6 : 0000000000000001
> [   21.951877] x5 : ffff940002719618 x4 : ffff00236ffb0010
> [   21.957178] x3 : ffffa000112415e4 x2 : ffff80046dff7ede
> [   21.962478] x1 : 5aff78756b1cf400 x0 : 0000000000000000
> [   21.967779] Call trace:
> [   21.970214]  refcount_inc_checked+0x44/0x50
> [   21.974389]  ghes_edac_register+0x258/0x388
> [   21.978562]  ghes_probe+0x28c/0x5f0
> [   21.982041]  platform_drv_probe+0x70/0xd8
> [   21.986039]  really_probe+0x174/0x468
> [   21.989690]  driver_probe_device+0x7c/0x148
> [   21.993862]  device_driver_attach+0x94/0xa0
> [   21.998033]  __driver_attach+0xa4/0x110
> [   22.001857]  bus_for_each_dev+0xe8/0x158
> [   22.005768]  driver_attach+0x30/0x40
> [   22.009331]  bus_add_driver+0x234/0x2f0
> [   22.013156]  driver_register+0xbc/0x1d0
> [   22.016981]  __platform_driver_register+0x7c/0x88
> [   22.021675]  ghes_init+0xbc/0x14c
> [   22.024979]  do_one_initcall+0xb4/0x254
> [   22.028805]  kernel_init_freeable+0x248/0x2f4
> [   22.033151]  kernel_init+0x10/0x118
> [   22.036628]  ret_from_fork+0x10/0x18
> [   22.040194] ---[ end trace 33655bb65a9835fe ]---
> [   22.046666] EDAC MC: bug in low-level driver: attempt to assign
> [   22.046666]     duplicate mc_idx 0 in add_mc_to_global_list()
> [   22.058311] ghes_edac: Can't register at EDAC core
> [   22.065402] EDAC MC: bug in low-level driver: attempt to assign
> [   22.065402]     duplicate mc_idx 0 in add_mc_to_global_list()
> [   22.077080] ghes_edac: Can't register at EDAC core
> [   22.084140] EDAC MC: bug in low-level driver: attempt to assign
> [   22.084140]     duplicate mc_idx 0 in add_mc_to_global_list()
> [   22.095789] ghes_edac: Can't register at EDAC core
> [   22.102873] EDAC MC: bug in low-level driver: attempt to assign
> [   22.102873]     duplicate mc_idx 0 in add_mc_to_global_list()
> [   22.115442] ghes_edac: Can't register at EDAC core
> [   22.122536] EDAC MC: bug in low-level driver: attempt to assign
> [   22.122536]     duplicate mc_idx 0 in add_mc_to_global_list()
> [   22.134344] ghes_edac: Can't register at EDAC core
> [   22.141441] EDAC MC: bug in low-level driver: attempt to assign
> [   22.141441]     duplicate mc_idx 0 in add_mc_to_global_list()
> [   22.153089] ghes_edac: Can't register at EDAC core
> [   22.160161] EDAC MC: bug in low-level driver: attempt to assign
> [   22.160161]     duplicate mc_idx 0 in add_mc_to_global_list()
> [   22.171810] ghes_edac: Can't register at EDAC core

What I am more concerned is this here. In total this implies 8 ghes
users that all try to register a (single-instance) ghes mc device. For
non-x86 only one instance is allowed (see ghes_edac_register(), idx =
0).

So on your platform, when parsing the HEST table
(hest_ghes_dev_register()), more than one "GHES" device is parsed,
allocated and registered. Mind sending me your HEST table
(/sys/firmware/acpi/tables/HEST), or explain what happens here? If
this is a valid use case, we need to change ghes_edac_register() to
support more than one instance.

Again, please try the patch below.

Thanks,

-Robert

From 6962f8af4a7c1051c9e87a5ac60571f70d2b6814 Mon Sep 17 00:00:00 2001
From: Robert Richter <[email protected]>
Date: Thu, 21 Nov 2019 15:01:28 +0100
Subject: [PATCH] EDAC/ghes: Do not warn on when increment refcount on 0

Signed-off-by: Robert Richter <[email protected]>

drivers/edac/ghes_edac.c

@@ -556,8 +556,8 @@ int ghes_edac_register(struct ghes *ghes, struct device *dev)
 	ghes_pvt = pvt;
 	spin_unlock_irqrestore(&ghes_lock, flags);
 
-	/* only increment on success */
-	refcount_inc(&ghes_refcount);
+	/* only set on success */
+	refcount_set(&ghes_refcount, 1);
 
 unlock:
 	mutex_unlock(&ghes_reg_mutex);