Re: [PATCH] staging/android: refactor SYNC_IOC_FILE_INFO
From: Emil Velikov
Date: Mon Feb 29 2016 - 03:34:25 EST
Hi Gustavo,
On 27 February 2016 at 15:25, Gustavo Padovan
<gustavo.padovan@xxxxxxxxxxxxxxx> wrote:
> Hi Emil,
>
> 2016-02-27 Emil Velikov <emil.l.velikov@xxxxxxxxx>:
>
>> Hi Gustavo,
>>
>> On 26 February 2016 at 21:00, Gustavo Padovan <gustavo@xxxxxxxxxxx> wrote:
>> > From: Gustavo Padovan <gustavo.padovan@xxxxxxxxxxxxxxx>
>> >
>> > Change SYNC_IOC_FILE_INFO behaviour to avoid future API breaks and
>> > optimize buffer allocation. In the new approach the ioctl needs to be called
>> > twice to retrieve the array of fence_infos pointed by info->sync_fence_info.
>> >
>> I might have misunderstood things but I no one says you "need" to call
>> it twice - you can just request a "random" amount of fences_info. Upon
>> return (if num_fences was non zero) it will report how many fence_info
>> were retrieved.
>
> Right, I don't see any problem doing it in one request, I just didn't
> think about that in the new proposal. I'll update the code and commit
> message accordinly.
>
>>
>> > The first call should pass num_fences = 0, the kernel will then fill
>> > info->num_fences. Userspace receives back the number of fences and
>> > allocates a buffer size num_fences * sizeof(struct sync_fence_info) on
>> > info->sync_fence_info.
>> >
>> > It then call the ioctl again passing num_fences received in info->num_fences.
>> "calls"
>>
>> > The kernel checks if info->num_fences > 0 and if yes it fill
>> > info->sync_fence_info with an array containing all fence_infos.
>> >
>> The above sentence sounds a bit strange. I believe you meant to say
>> something like "Then the kernel fills the fence_infos array with data.
>> One should read back the actual number from info->num_fences." ?
>>
>> > info->len now represents the length of the buffer sync_fence_info points
>> > to.
>> Now that I think about it, I'm wondering if there'll be a case where
>> len != info->num_fences * sizeof(struct sync_file_info). If that's not
>> possible one could just drop len and nicely simplify things.
>>
>> > Also, info->sync_fence_info was converted to __u64 pointer.
>> >
>> ... pointer to prevent 32bit compatibility issues.
>>
>> > An example userspace code would be:
>> >
>> > struct sync_file_info *info;
>> > int err, size, num_fences;
>> >
>> > info = malloc(sizeof(*info));
>> >
>> > memset(info, 0, sizeof(*info));
>> >
>> > err = ioctl(fd, SYNC_IOC_FILE_INFO, info);
>> > num_fences = info->num_fences;
>> >
>> > if (num_fences) {
>> > memset(info, 0, sizeof(*info));
>> > size = sizeof(struct sync_fence_info) * num_fences;
>> > info->len = size;
>> > info->num_fences = num_fences;
>> > info->sync_fence_info = (uint64_t) calloc(num_fences,
>> > sizeof(struct sync_fence_info));
>> >
>> > err = ioctl(fd, SYNC_IOC_FILE_INFO, info);
>> > }
>> >
>> > v2: fix fence_info memory leak
>> >
>> > Signed-off-by: Gustavo Padovan <gustavo.padovan@xxxxxxxxxxxxxxx>
>> > ---
>> > drivers/staging/android/sync.c | 52 +++++++++++++++++++++++++++++--------
>> > drivers/staging/android/uapi/sync.h | 9 +++----
>> > 2 files changed, 45 insertions(+), 16 deletions(-)
>> >
>> > diff --git a/drivers/staging/android/sync.c b/drivers/staging/android/sync.c
>> > index dc5f382..2379f23 100644
>> > --- a/drivers/staging/android/sync.c
>> > +++ b/drivers/staging/android/sync.c
>> > @@ -502,21 +502,22 @@ static int sync_fill_fence_info(struct fence *fence, void *data, int size)
>> > static long sync_file_ioctl_fence_info(struct sync_file *sync_file,
>> > unsigned long arg)
>> > {
>> > - struct sync_file_info *info;
>> > + struct sync_file_info in, *info;
>> > + struct sync_fence_info *fence_info = NULL;
>> > __u32 size;
>> > __u32 len = 0;
>> > int ret, i;
>> >
>> > - if (copy_from_user(&size, (void __user *)arg, sizeof(size)))
>> > + if (copy_from_user(&in, (void __user *)arg, sizeof(*info)))
>> s/*info/in/
>>
>> > return -EFAULT;
>> >
>> > - if (size < sizeof(struct sync_file_info))
>> > - return -EINVAL;
>> > + if (in.status || strcmp(in.name, "\0"))
>> Afaict these two are outputs, so we should be checking them ?
>
> Hmm. Maybe not.
>
>>
>> > + return -EFAULT;
>> >
>> As originally, input validation should return -EINVAL on error.
>>
>>
>> > - if (size > 4096)
>> > - size = 4096;
>> > + if (in.num_fences && !in.sync_fence_info)
>> > + return -EFAULT;
>> >
>> Ditto.
>>
>> > - info = kzalloc(size, GFP_KERNEL);
>> > + info = kzalloc(sizeof(*info), GFP_KERNEL);
>> > if (!info)
>> > return -ENOMEM;
>> >
>> > @@ -525,14 +526,33 @@ static long sync_file_ioctl_fence_info(struct sync_file *sync_file,
>> > if (info->status >= 0)
>> > info->status = !info->status;
>> >
>> > - info->num_fences = sync_file->num_fences;
>> > + /*
>> > + * Passing num_fences = 0 means that userspace want to know how
>> > + * many fences are in the sync_file to be able to allocate a buffer to
>> > + * fit all sync_fence_infos and call the ioctl again with the buffer
>> > + * assigned to info->sync_fence_info. The second call pass the
>> > + * num_fences value received in the first call.
>> > + */
>> > + if (!in.num_fences)
>> > + goto no_fences;
>> > +
>> We should clamp in.num_fences to min2(in.num_fences,
>> sync_file->num_fences) and use it over sync_file->num_fences though
>> the rest of the function. Or just bail out when the two are not the
>> same.
>>
>> Depends on what the planned semantics are. Fwiw I'm leaning towards the former.
>
> If num_fences received is smaller than the actual num_fences I think we
> should fails, otherwise we should just fill the buffer with all
> fence_infos...
>
Fair enough. Just make sure that this is clearly explained to the use
- manpages, other ?
>>
>> > + size = sync_file->num_fences * sizeof(*fence_info);
>> > + if (in.len != size) {
>> > + ret = -EFAULT;
>> EINVAL or just drop len from the struct.
>
> ...so this check now would be in.len < size.
>
>>
>> > + goto out;
>> > + }
>> >
>> > - len = sizeof(struct sync_file_info);
>> > + fence_info = kzalloc(size, GFP_KERNEL);
>> > + if (!fence_info) {
>> > + ret = -ENOMEM;
>> > + goto out;
>> > + }
>> >
>> > for (i = 0; i < sync_file->num_fences; ++i) {
>> > struct fence *fence = sync_file->cbs[i].fence;
>> >
>> > - ret = sync_fill_fence_info(fence, (u8 *)info + len, size - len);
>> > + ret = sync_fill_fence_info(fence, (u8 *)fence_info + len,
>> A few comments about sync_fill_fence_info()
>> - Internal function so make the second argument of the correct type -
>> struct sync_fence_info *
>> - Drop the third argument size, as that one is always sizeof(sync_fence_info).
>> - Remove the size checking in the same function and make its return type void
>>
>> Then one can simplify this loop even further :-)
>
> Sounds good to me.
>
>>
>> > + size - len);
>> >
>> > if (ret < 0)
>> > goto out;
>> > @@ -540,14 +560,24 @@ static long sync_file_ioctl_fence_info(struct sync_file *sync_file,
>> > len += ret;
>> > }
>> >
>> > + if (copy_to_user((void __user *)in.sync_fence_info, fence_info, size)) {
>> > + ret = -EFAULT;
>> > + goto out;
>> > + }
>> > +
>> > info->len = len;
>> > + info->sync_fence_info = (__u64) in.sync_fence_info;
>> Why the cast ?
>>
>> > +
>> > +no_fences:
>> > + info->num_fences = sync_file->num_fences;
>> >
>> > - if (copy_to_user((void __user *)arg, info, len))
>> > + if (copy_to_user((void __user *)arg, info, sizeof(*info)))
>> Don't know if we should be returning (copying) any other information
>> but info->num_fences in case of "no_fences". In case it's not clear -
>> I'm thinking about the data we already have in in info->name and
>> info->status.
>
> Userspace might want to know all info about the sync_file but
> sync_fence_info.
>
IMHO, that does sound like a rather strange use of the API. Whichever
route one goes for, it would be nice to have the behaviour clearly
documented. Otherwise things will end up quite confusing.
Documentation can be done in separate patch/series.
Thanks
Emil