• Jialing Fu's avatar
    mmc: core: fix race condition in mmc_wait_data_done · 71f8a4b8
    Jialing Fu authored
    The following panic is captured in ker3.14, but the issue still exists
    in latest kernel.
    ---------------------------------------------------------------------
    [   20.738217] c0 3136 (Compiler) Unable to handle kernel NULL pointer dereference
    at virtual address 00000578
    ......
    [   20.738499] c0 3136 (Compiler) PC is at _raw_spin_lock_irqsave+0x24/0x60
    [   20.738527] c0 3136 (Compiler) LR is at _raw_spin_lock_irqsave+0x20/0x60
    [   20.740134] c0 3136 (Compiler) Call trace:
    [   20.740165] c0 3136 (Compiler) [<ffffffc0008ee900>] _raw_spin_lock_irqsave+0x24/0x60
    [   20.740200] c0 3136 (Compiler) [<ffffffc0000dd024>] __wake_up+0x1c/0x54
    [   20.740230] c0 3136 (Compiler) [<ffffffc000639414>] mmc_wait_data_done+0x28/0x34
    [   20.740262] c0 3136 (Compiler) [<ffffffc0006391a0>] mmc_request_done+0xa4/0x220
    [   20.740314] c0 3136 (Compiler) [<ffffffc000656894>] sdhci_tasklet_finish+0xac/0x264
    [   20.740352] c0 3136 (Compiler) [<ffffffc0000a2b58>] tasklet_action+0xa0/0x158
    [   20.740382] c0 3136 (Compiler) [<ffffffc0000a2078>] __do_softirq+0x10c/0x2e4
    [   20.740411] c0 3136 (Compiler) [<ffffffc0000a24bc>] irq_exit+0x8c/0xc0
    [   20.740439] c0 3136 (Compiler) [<ffffffc00008489c>] handle_IRQ+0x48/0xac
    [   20.740469] c0 3136 (Compiler) [<ffffffc000081428>] gic_handle_irq+0x38/0x7c
    ----------------------------------------------------------------------
    Because in SMP, "mrq" has race condition between below two paths:
    path1: CPU0: <tasklet context>
      static void mmc_wait_data_done(struct mmc_request *mrq)
      {
         mrq->host->context_info.is_done_rcv = true;
         //
         // If CPU0 has just finished "is_done_rcv = true" in path1, and at
         // this moment, IRQ or ICache line missing happens in CPU0.
         // What happens in CPU1 (path2)?
         //
         // If the mmcqd thread in CPU1(path2) hasn't entered to sleep mode:
         // path2 would have chance to break from wait_event_interruptible
         // in mmc_wait_for_data_req_done and continue to run for next
         // mmc_request (mmc_blk_rw_rq_prep).
         //
         // Within mmc_blk_rq_prep, mrq is cleared to 0.
         // If below line still gets host from "mrq" as the result of
         // compiler, the panic happens as we traced.
         wake_up_interruptible(&mrq->host->context_info.wait);
      }
    
    path2: CPU1: <The mmcqd thread runs mmc_queue_thread>
      static int mmc_wait_for_data_req_done(...
      {
         ...
         while (1) {
               wait_event_interruptible(context_info->wait,
                       (context_info->is_done_rcv ||
                        context_info->is_new_req));
         	   static void mmc_blk_rw_rq_prep(...
               {
               ...
               memset(brq, 0, sizeof(struct mmc_blk_request));
    
    This issue happens very coincidentally; however adding mdelay(1) in
    mmc_wait_data_done as below could duplicate it easily.
    
       static void mmc_wait_data_done(struct mmc_request *mrq)
       {
         mrq->host->context_info.is_done_rcv = true;
    +    mdelay(1);
         wake_up_interruptible(&mrq->host->context_info.wait);
        }
    
    At runtime, IRQ or ICache line missing may just happen at the same place
    of the mdelay(1).
    
    This patch gets the mmc_context_info at the beginning of function, it can
    avoid this race condition.
    Signed-off-by: default avatarJialing Fu <jlfu@marvell.com>
    Tested-by: default avatarShawn Lin <shawn.lin@rock-chips.com>
    Fixes: 2220eedf ("mmc: fix async request mechanism ....")
    Signed-off-by: default avatarShawn Lin <shawn.lin@rock-chips.com>
    Signed-off-by: default avatarUlf Hansson <ulf.hansson@linaro.org>
    71f8a4b8
core.c 69.2 KB