On 18.01.22 15:29, Andrew Cooper wrote:
> On 16/01/2022 08:33, Juergen Gross wrote:
>> diff --git a/lib/xs.c b/lib/xs.c
>> index 4af0f960..c12341aa 100644
>> --- a/lib/xs.c
>> +++ b/lib/xs.c
>> @@ -18,23 +18,56 @@ static inline int _xs_fileno(struct xs_handle *h) {
>>       return (intptr_t) h;
>>   }
>>   
>> +static int xs_close_fd(struct file *file)
>> +{
>> +    struct xenbus_event *event, *next;
>> +
>> +    for (event = file->dev; event; event = next)
>> +    {
>> +        next = event->next;
>> +        free(event);
>> +    }
>> +
>> +    return 0;
>> +}
>> +
>> +static bool xs_can_read(struct file *file)
>> +{
>> +    return file && file->dev;
> 
> Just 'return file->dev;' ?

Yes.

> 
>> @@ -169,18 +202,20 @@ char **xs_directory(struct xs_handle *h, xs_transaction_t t,
>>   
>>   bool xs_watch(struct xs_handle *h, const char *path, const char *token)
>>   {
>> -    int fd = _xs_fileno(h);
>> +    struct file *file = get_file_from_fd(_xs_fileno(h));
>> +
>>       printk("xs_watch(%s, %s)\n", path, token);
>>       return xs_bool(xenbus_watch_path_token(XBT_NULL, path, token,
>> -                   (xenbus_event_queue *)&files[fd].dev));
>> +                   (xenbus_event_queue *)&file->dev));
> 
> This is utterly mad.  In particular, close() looks to be very racy with
> new watches arriving.

In practice it should be no problem, though (closing the file in one
thread while the other one is adding a watch would be rather strange).

Additionally close() for xenbus in Mini-OS is called only when stopping
the domain today.

> However, can the indentation at least be fixed here as the line is
> changing.  That's a parameter to xenbus_watch_path_token(), not xs_bool().

Yes, that should be done.

> 
>> diff --git a/xenbus/xenbus.c b/xenbus/xenbus.c
>> index b687678f..785389fb 100644
>> --- a/xenbus/xenbus.c
>> +++ b/xenbus/xenbus.c
>> @@ -393,6 +393,7 @@ static int allocate_xenbus_id(void)
>>   void init_xenbus(void)
>>   {
>>       int err;
>> +
> 
> Spurious hunk?

Yes.


Juergen