Re: [PATCH v2] hid: hid-led: fix issue with transfer buffer not being dma capable

[Heiner Kallweit <hkallweit1@gmail.com>]

Am 07.10.2016 um 18:22 schrieb Benjamin Tissoires:
> On Oct 03 2016 or thereabouts, Heiner Kallweit wrote:
>> The hid-led driver works fine under 4.8.0, however with the next
>> kernel from today I get this:
>>
>> ------------[ cut here ]------------
>> WARNING: CPU: 0 PID: 2578 at drivers/usb/core/hcd.c:1584 usb_hcd_map_urb_for_dma+0x373/0x550 [usbcore]
>> transfer buffer not dma capable
>> Modules linked in: hid_led(+) usbhid vfat fat ir_sony_decoder iwlmvm led_class mac80211 snd_hda_codec_realtek snd_hda_codec_generic x86_pkg_temp_thermal iwlwifi crc32c_intel snd_hda_codec_hdmi i2c_i801 i2c_smbus snd_hda_intel cfg80211 snd_hda_codec snd_hda_core snd_pcm r8169 snd_timer mei_me mii snd mei ir_lirc_codec lirc_dev nuvoton_cir rc_core btusb btintel bluetooth rfkill usb_storage efivarfs ipv6 ehci_pci ehci_hcd xhci_pci xhci_hcd usbcore usb_common ext4 jbd2 mbcache ahci libahci libata
>> CPU: 0 PID: 2578 Comm: systemd-udevd Not tainted 4.8.0-rc8-next-20161003 #1
>> Hardware name: ZOTAC ZBOX-CI321NANO/ZBOX-CI321NANO, BIOS B246P105 06/01/2015
>>  ffffc90003dbb7e0 ffffffff81280425 ffffc90003dbb830 0000000000000000
>>  ffffc90003dbb820 ffffffff8105b086 0000063003dbb800 ffff88006f374480
>>  0000000000000000 0000000000000000 0000000000000001 ffff880079544000
>> Call Trace:
>>  [<ffffffff81280425>] dump_stack+0x68/0x93
>>  [<ffffffff8105b086>] __warn+0xc6/0xe0
>>  [<ffffffff8105b0ea>] warn_slowpath_fmt+0x4a/0x50
>>  [<ffffffffa0143a43>] usb_hcd_map_urb_for_dma+0x373/0x550 [usbcore]
>>  [<ffffffffa01441b6>] usb_hcd_submit_urb+0x316/0x9c0 [usbcore]
>>  [<ffffffff810bce80>] ? rcu_read_lock_sched_held+0x40/0x80
>>  [<ffffffff810e0043>] ? module_assert_mutex_or_preempt+0x13/0x50
>>  [<ffffffff810e0c07>] ? __module_address+0x27/0xf0
>>  [<ffffffffa01456e4>] usb_submit_urb+0x2c4/0x520 [usbcore]
>>  [<ffffffffa0145fea>] usb_start_wait_urb+0x5a/0xe0 [usbcore]
>>  [<ffffffffa014612c>] usb_control_msg+0xbc/0xf0 [usbcore]
>>  [<ffffffff810e0c07>] ? __module_address+0x27/0xf0
>>  [<ffffffffa079a724>] usbhid_raw_request+0xa4/0x180 [usbhid]
>>  [<ffffffffa07a93b1>] hidled_recv+0x71/0xe0 [hid_led]
>>  [<ffffffffa07a947d>] thingm_init+0x2d/0x50 [hid_led]
>>  [<ffffffffa07a969b>] hidled_probe+0xcb/0x24a [hid_led]
>>  [<ffffffff814d96f2>] hid_device_probe+0xd2/0x150
>>  [<ffffffff8146023d>] driver_probe_device+0x1fd/0x2c0
>>  [<ffffffff8146039a>] __driver_attach+0x9a/0xa0
>>  [<ffffffff81460300>] ? driver_probe_device+0x2c0/0x2c0
>>  [<ffffffff8145e25d>] bus_for_each_dev+0x5d/0x90
>>  [<ffffffff8145fa79>] driver_attach+0x19/0x20
>>  [<ffffffff8145f5ff>] bus_add_driver+0x11f/0x220
>>  [<ffffffffa07ac000>] ? 0xffffffffa07ac000
>>  [<ffffffff8146086b>] driver_register+0x5b/0xd0
>>  [<ffffffffa07ac000>] ? 0xffffffffa07ac000
>>  [<ffffffff814d83d1>] __hid_register_driver+0x61/0xa0
>>  [<ffffffffa07ac01e>] hidled_driver_init+0x1e/0x20 [hid_led]
>>  [<ffffffff81000408>] do_one_initcall+0x38/0x150
>>  [<ffffffff810bce80>] ? rcu_read_lock_sched_held+0x40/0x80
>>  [<ffffffff81194ca0>] ? kmem_cache_alloc_trace+0x1d0/0x230
>>  [<ffffffff811342f9>] do_init_module+0x5a/0x1cb
>>  [<ffffffff810e3862>] load_module+0x1e42/0x2530
>>  [<ffffffff810e0990>] ? __symbol_put+0x50/0x50
>>  [<ffffffff810dfc50>] ? show_coresize+0x30/0x30
>>  [<ffffffff811ad650>] ? kernel_read_file+0x100/0x190
>>  [<ffffffff811ad794>] ? kernel_read_file_from_fd+0x44/0x70
>>  [<ffffffff810e415a>] SYSC_finit_module+0xba/0xc0
>>  [<ffffffff810e4179>] SyS_finit_module+0x9/0x10
>>  [<ffffffff815e082a>] entry_SYSCALL_64_fastpath+0x18/0xad
>> ---[ end trace c9e6ea27003ecf9e ]---
>>
>> Fix this by using a kmalloc'ed buffer when calling hid_hw_raw_request.
>>
>> Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
>> ---
>> v2:
>> - Based on a comment from Alan Stern allocate the buffer to be provided to
>>   hid_hw_raw_request separately (and not as part of struct hidled_device).
>>   Alternative would have been to allocate the buffer dynamically in each
>>   function calling hidled_send/_recv. However this would mean more overhead
>>   IMHO, and we'd need an error path in callers to free the buffer in case
>>   of an error.
>>   In addition we have better control that a proper buffer is used in case
>>   the driver is extended by somebody for supporting another LED device.
>> ---
> 
> Looks like the receive function is only called from .probe(), so this
> should be safe.
> However, for the send function, is there a chance there can be a
> concurrent access of the buffer? (like 2 userspace processes writing a
> different values at the same time).
> 
Buffer ldev->buf is accessed with mutex ldev->lock held in hidled_send.
So the locking exists already.

> If so, then you'll need to add a locking mechanism (can't recall if the
> LED API provide one for us or not), or just alloc/free the buffers
> directly.
> 
> If no, the patch is:
> Reviewed-by: Benjamin Tissoires <benjamin.tissoires@redhat.com>
> 
> Cheers,
> Benjamin
> 
>>  drivers/hid/hid-led.c | 23 +++++++++++++++++++----
>>  1 file changed, 19 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/hid/hid-led.c b/drivers/hid/hid-led.c
>> index d8d55f3..555c88e 100644
>> --- a/drivers/hid/hid-led.c
>> +++ b/drivers/hid/hid-led.c
>> @@ -100,6 +100,7 @@ struct hidled_device {
>>  	const struct hidled_config *config;
>>  	struct hid_device       *hdev;
>>  	struct hidled_rgb	*rgb;
>> +	u8			*buf;
>>  	struct mutex		lock;
>>  };
>>  
>> @@ -118,13 +119,19 @@ static int hidled_send(struct hidled_device *ldev, __u8 *buf)
>>  
>>  	mutex_lock(&ldev->lock);
>>  
>> +	/*
>> +	 * buffer provided to hid_hw_raw_request must not be on the stack
>> +	 * and must not be part of a data structure
>> +	 */
>> +	memcpy(ldev->buf, buf, ldev->config->report_size);
>> +
>>  	if (ldev->config->report_type == RAW_REQUEST)
>> -		ret = hid_hw_raw_request(ldev->hdev, buf[0], buf,
>> +		ret = hid_hw_raw_request(ldev->hdev, buf[0], ldev->buf,
>>  					 ldev->config->report_size,
>>  					 HID_FEATURE_REPORT,
>>  					 HID_REQ_SET_REPORT);
>>  	else if (ldev->config->report_type == OUTPUT_REPORT)
>> -		ret = hid_hw_output_report(ldev->hdev, buf,
>> +		ret = hid_hw_output_report(ldev->hdev, ldev->buf,
>>  					   ldev->config->report_size);
>>  	else
>>  		ret = -EINVAL;
>> @@ -147,17 +154,21 @@ static int hidled_recv(struct hidled_device *ldev, __u8 *buf)
>>  
>>  	mutex_lock(&ldev->lock);
>>  
>> -	ret = hid_hw_raw_request(ldev->hdev, buf[0], buf,
>> +	memcpy(ldev->buf, buf, ldev->config->report_size);
>> +
>> +	ret = hid_hw_raw_request(ldev->hdev, buf[0], ldev->buf,
>>  				 ldev->config->report_size,
>>  				 HID_FEATURE_REPORT,
>>  				 HID_REQ_SET_REPORT);
>>  	if (ret < 0)
>>  		goto err;
>>  
>> -	ret = hid_hw_raw_request(ldev->hdev, buf[0], buf,
>> +	ret = hid_hw_raw_request(ldev->hdev, buf[0], ldev->buf,
>>  				 ldev->config->report_size,
>>  				 HID_FEATURE_REPORT,
>>  				 HID_REQ_GET_REPORT);
>> +
>> +	memcpy(buf, ldev->buf, ldev->config->report_size);
>>  err:
>>  	mutex_unlock(&ldev->lock);
>>  
>> @@ -447,6 +458,10 @@ static int hidled_probe(struct hid_device *hdev, const struct hid_device_id *id)
>>  	if (!ldev)
>>  		return -ENOMEM;
>>  
>> +	ldev->buf = devm_kmalloc(&hdev->dev, MAX_REPORT_SIZE, GFP_KERNEL);
>> +	if (!ldev->buf)
>> +		return -ENOMEM;
>> +
>>  	ret = hid_parse(hdev);
>>  	if (ret)
>>  		return ret;
>> -- 
>> 2.10.0
>>
>>
>