Re: Unmapping of UIO logical memory causing a trace

[Ankit Jindal]

Hi Greg,

On Wed, Aug 12, 2015 at 10:41 PM, Greg Kroah-Hartman
<gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> On Wed, Aug 12, 2015 at 01:25:53PM +0530, Ankit Jindal wrote:
>> Hi Greg,
>>
>> On Wed, Aug 12, 2015 at 3:15 AM, Greg Kroah-Hartman
>> <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
>> > On Tue, Aug 11, 2015 at 04:39:08PM +0530, Ankit Jindal wrote:
>> >> Hi,
>> >>
>> >> We have observed an issue where kmalloc of a small sized memory causes
>> >> an occasional trace when unmapping the mmaped memory via UIO framework
>> >> This trace is coming when kernel sees a negative value in
>> >> page->_mapcount. Trace is pasted at the end of the mail.
>> >>
>> >> After debugging this issue further, we realized following sequence
>> >> occurs when kmalloc is used to allocate small memory using slub
>> >> allocator:
>> >> 1. Frozen bit (msb) of the page from which memory has been allocated
>> >> is set (which is an union with _mapcount).
>> >> 2. If there are free objects in the the same page then this frozen bit
>> >> remains set even after kernel boots completely.
>> >> 3. When user space calls unmap of this memory, vma_unmap_single()
>> >> treats the _mapcount as a negative (as frozen bit is set), causing a
>> >> trace.
>> >>
>> >> We are not sure whether exposing kernel memory of size
>> >> less than PAGE_SIZE via UIO is a valid use case ? In case this is an invalid
>> >> use case then shouldn't the UIO framework restrict mapping of non
>> >> PAGE_SIZE aligned memory and size not in order of PAGE_SIZE.
>> >
>> > We've had a few discussions about this in the past, and one proposed
>> > patch which had to be reverted because it broke some working systems, so
>> > it's a messy thing.
>> >
>> > What UIO driver are you using that causes this behavior?
>>
>> We have observed this during the development of new UIO driver for our
>> soc. In our driver, we need to parse non probable properties of device
>> and provide these details to our user application.
>
> What exactly do you mean by this?

There are certain the device specific properties (coming from device
tree) which we are exposing using UIO_LOGICAL_MEM to the user space.

>
>> For this we do a kmalloc of device info(approx size 80 bytes) and pass
>> this address to user space via UIO mem logical.
>
> For such tiny sizes, why not just use a normal sysfs file?

We did not think of sysfs as an option, since UIO_MEM_LOGICAL seems to
address our use case.

>
> Do you have a pointer to your driver so that I can see exactly what it
> is doing here?

Below is the code snippet of my driver.

struct reserved_qgroup {
        uint32_t qstart;
        uint32_t qend;
        uint32_t valid;
};

struct my_dev_info {
        uint32_t num_queues;
        uint32_t version;
        struct reserved_qgroup resvd_qgrps[MAX_RESVD_QGROUPS];
};

static int my_dev_probe(struct platform_device *pdev) {

struct property *prop;
struct my_dev_info *dev_info;

dev_info = devm_kzalloc(&pdev->dev, sizeof(*dev_info), GFP_KERNEL);

if (!dev_info)
      return -ENOMEM;

prop = of_find_property(pdev->dev.of_node, "reserved-qgroups", NULL);

if (prop) {
        const __be32 *reserved_qid = NULL;
        unsigned int i;
        uint32_t qid_start, qid_end;

        for (i = 0; i < MAX_RESVD_QGROUPS; i++) {
                reserved_qid = of_prop_next_u32(prop, reserved_qid, &qid_start);
                if (!reserved_qid)
                        break;
                reserved_qid = of_prop_next_u32(prop, reserved_qid, &qid_end);
                if (!reserved_qid)
                        break;

                dev_info->resvd_qgrps[i].qstart = qid_start;
                dev_info->resvd_qgrps[i].qend = qid_end;
                dev_info->resvd_qgrps[i].valid  = 1;
        }
}

ret = of_property_read_u32(pdev->dev.of_node, "num-queues",
                dev_info->num_queues);
if (ret < 0) {
        dev_err(&pdev->dev, "No num-queues resource specified\n");
        goto out_err;
}

info->mem[3].name = "dev-info";
info->mem[3].addr = (phys_addr_t) dev_info;
info->mem[3].size = sizeof(struct dev_info);
info->mem[3].memtype = UIO_MEM_LOGICAL;

......
}

Thanks,
Ankit
--
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/