On Thu, 13 Sep 2018, Jia-Ju Bai wrote:

hid_alloc_report_buf() has to be called with GFP_ATOMIC in 
__hid_request(), because there are the following callchains 
leading to __hid_request() being an atomic context:

picolcd_send_and_wait (acquire a spinlock)
  hid_hw_request
    __hid_request
      hid_alloc_report_buf(GFP_KERNEL)

picolcd_reset (acquire a spinlock)
  hid_hw_request
    __hid_request
      hid_alloc_report_buf(GFP_KERNEL)

lg4ff_play (acquire a spinlock)
  hid_hw_request
    __hid_request
      hid_alloc_report_buf(GFP_KERNEL)

lg4ff_set_autocenter_ffex (acquire a spinlock)
  hid_hw_request
    __hid_request
      hid_alloc_report_buf(GFP_KERNEL)

Hm, so it's always drivers calling out into core in atomic context. So 
either we take this, and put our bets on being able to allocate the buffer 
without sleeping, or actually fix the few drivers (it's just lg4ff and 
picolcd at the end of the day) not to do that, and explicitly anotate 
__hid_request() with might_sleep().

Hmm?

Thanks,

-- 
Jiri Kosina
SUSE Labs