Re: [PATCH 1/2] clk: Fix double free due to devm_clk_register()

From: Sylwester Nawrocki
Date: Wed May 14 2014 - 07:27:10 EST


On 19/04/14 01:29, Stephen Boyd wrote:
> Now that clk_unregister() frees the struct clk we're
> unregistering we'll free memory twice: first we'll call kfree()
> in __clk_release() with an address kmalloc doesn't know about and
> second we'll call kfree() in the devres layer. Remove the
> allocation of struct clk in devm_clk_register() and let
> clk_release() handle it. This fixes slab errors like:
>
> =============================================================================
> BUG kmalloc-128 (Not tainted): Invalid object pointer 0xed08e8d0
> -----------------------------------------------------------------------------
>
> Disabling lock debugging due to kernel taint
> INFO: Slab 0xeec503f8 objects=25 used=15 fp=0xed08ea00 flags=0x4081
> CPU: 2 PID: 73 Comm: rmmod Tainted: G B 3.14.0-11032-g526e9c764381 #34
> [<c0014be0>] (unwind_backtrace) from [<c0012240>] (show_stack+0x10/0x14)
> [<c0012240>] (show_stack) from [<c04b74dc>] (dump_stack+0x70/0xbc)
> [<c04b74dc>] (dump_stack) from [<c00f6778>] (slab_err+0x74/0x84)
> [<c00f6778>] (slab_err) from [<c04b6278>] (free_debug_processing+0x2cc/0x31c)
> [<c04b6278>] (free_debug_processing) from [<c04b6300>] (__slab_free+0x38/0x41c)
> [<c04b6300>] (__slab_free) from [<c03931bc>] (clk_unregister+0xd4/0x140)
> [<c03931bc>] (clk_unregister) from [<c02fb774>] (release_nodes+0x164/0x1d8)
> [<c02fb774>] (release_nodes) from [<c02f8698>] (__device_release_driver+0x60/0xb0)
> [<c02f8698>] (__device_release_driver) from [<c02f9080>] (driver_detach+0xb4/0xb8)
> [<c02f9080>] (driver_detach) from [<c02f8480>] (bus_remove_driver+0x5c/0xc4)
> [<c02f8480>] (bus_remove_driver) from [<c008c9b8>] (SyS_delete_module+0x148/0x1d8)
> [<c008c9b8>] (SyS_delete_module) from [<c000ef80>] (ret_fast_syscall+0x0/0x48)
> FIX kmalloc-128: Object at 0xed08e8d0 not freed
>
> Fixes: fcb0ee6a3d33 (clk: Implement clk_unregister)
> Cc: Jiada Wang <jiada_wang@xxxxxxxxxx>
> Cc: Sylwester Nawrocki <s.nawrocki@xxxxxxxxxxx>
> Cc: Kyungmin Park <kyungmin.park@xxxxxxxxxxx>
> Signed-off-by: Stephen Boyd <sboyd@xxxxxxxxxxxxxx>

Thank you for correcting this, and my apologies for introducing this bug.

Acked-by: Sylwester Nawrocki <s.nawrocki@xxxxxxxxxxx>

> ---
> drivers/clk/clk.c | 71 +++++++++++++++++++++++--------------------------------
> 1 file changed, 30 insertions(+), 41 deletions(-)
>
> diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c
> index dff0373f53c1..f71093bf83ab 100644
> --- a/drivers/clk/clk.c
> +++ b/drivers/clk/clk.c
> @@ -1984,9 +1984,28 @@ struct clk *__clk_register(struct device *dev, struct clk_hw *hw)
> }
> EXPORT_SYMBOL_GPL(__clk_register);
>
> -static int _clk_register(struct device *dev, struct clk_hw *hw, struct clk *clk)
> +/**
> + * clk_register - allocate a new clock, register it and return an opaque cookie
> + * @dev: device that is registering this clock
> + * @hw: link to hardware-specific clock data
> + *
> + * clk_register is the primary interface for populating the clock tree with new
> + * clock nodes. It returns a pointer to the newly allocated struct clk which
> + * cannot be dereferenced by driver code but may be used in conjuction with the
> + * rest of the clock API. In the event of an error clk_register will return an
> + * error code; drivers must test for an error code after calling clk_register.
> + */
> +struct clk *clk_register(struct device *dev, struct clk_hw *hw)
> {
> int i, ret;
> + struct clk *clk;
> +
> + clk = kzalloc(sizeof(*clk), GFP_KERNEL);
> + if (!clk) {
> + pr_err("%s: could not allocate clk\n", __func__);
> + ret = -ENOMEM;
> + goto fail_out;
> + }
>
> clk->name = kstrdup(hw->init->name, GFP_KERNEL);
> if (!clk->name) {
> @@ -2026,7 +2045,7 @@ static int _clk_register(struct device *dev, struct clk_hw *hw, struct clk *clk)
>
> ret = __clk_init(dev, clk);
> if (!ret)
> - return 0;
> + return clk;
>
> fail_parent_names_copy:
> while (--i >= 0)
> @@ -2035,36 +2054,6 @@ fail_parent_names_copy:
> fail_parent_names:
> kfree(clk->name);
> fail_name:
> - return ret;
> -}
> -
> -/**
> - * clk_register - allocate a new clock, register it and return an opaque cookie
> - * @dev: device that is registering this clock
> - * @hw: link to hardware-specific clock data
> - *
> - * clk_register is the primary interface for populating the clock tree with new
> - * clock nodes. It returns a pointer to the newly allocated struct clk which
> - * cannot be dereferenced by driver code but may be used in conjuction with the
> - * rest of the clock API. In the event of an error clk_register will return an
> - * error code; drivers must test for an error code after calling clk_register.
> - */
> -struct clk *clk_register(struct device *dev, struct clk_hw *hw)
> -{
> - int ret;
> - struct clk *clk;
> -
> - clk = kzalloc(sizeof(*clk), GFP_KERNEL);
> - if (!clk) {
> - pr_err("%s: could not allocate clk\n", __func__);
> - ret = -ENOMEM;
> - goto fail_out;
> - }
> -
> - ret = _clk_register(dev, hw, clk);
> - if (!ret)
> - return clk;
> -
> kfree(clk);
> fail_out:
> return ERR_PTR(ret);
> @@ -2173,7 +2162,7 @@ EXPORT_SYMBOL_GPL(clk_unregister);
>
> static void devm_clk_release(struct device *dev, void *res)
> {
> - clk_unregister(res);
> + clk_unregister(*(struct clk **)res);
> }
>
> /**
> @@ -2188,18 +2177,18 @@ static void devm_clk_release(struct device *dev, void *res)
> struct clk *devm_clk_register(struct device *dev, struct clk_hw *hw)
> {
> struct clk *clk;
> - int ret;
> + struct clk **clkp;
>
> - clk = devres_alloc(devm_clk_release, sizeof(*clk), GFP_KERNEL);
> - if (!clk)
> + clkp = devres_alloc(devm_clk_release, sizeof(*clkp), GFP_KERNEL);
> + if (!clkp)
> return ERR_PTR(-ENOMEM);
>
> - ret = _clk_register(dev, hw, clk);
> - if (!ret) {
> - devres_add(dev, clk);
> + clk = clk_register(dev, hw);
> + if (!IS_ERR(clk)) {
> + *clkp = clk;
> + devres_add(dev, clkp);
> } else {
> - devres_free(clk);
> - clk = ERR_PTR(ret);
> + devres_free(clkp);
> }
>
> return clk;
>


--
Sylwester Nawrocki
Samsung R&D Institute Poland
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/