On 16/01/2020 20:10, Dmitry Osipenko wrote: > 15.01.2020 12:00, Jon Hunter пишет: >> >> On 14/01/2020 20:33, Dmitry Osipenko wrote: >>> 14.01.2020 18:09, Jon Hunter пишет: >>>> >>>> On 12/01/2020 17:29, Dmitry Osipenko wrote: >>>>> I was doing some experiments with I2C and noticed that Tegra APB DMA >>>>> driver crashes sometime after I2C DMA transfer termination. The crash >>>>> happens because tegra_dma_terminate_all() bails out immediately if pending >>>>> list is empty, thus it doesn't release the half-completed descriptors >>>>> which are getting re-used before ISR tasklet kicks-in. >>>> >>>> Can you elaborate a bit more on how these are getting re-used? What is >>>> the sequence of events which results in the panic? I believe that this >>>> was also reported in the past [0] and so I don't doubt there is an issue >>>> here, but would like to completely understand this. >>>> >>>> Thanks! >>>> Jon >>>> >>>> [0] https://lore.kernel.org/patchwork/patch/675349/ >>>> >>> >>> In my case it happens in the touchscreen driver during of the >>> touchscreen's interrupt handling (in a threaded IRQ handler) + CPU is >>> under load and there is other interrupts activity. So what happens here >>> is that the TS driver issues one I2C transfer, which fails with >>> (apparently bogus) timeout (because DMA descriptor is completed and >>> removed from the pending list, but tasklet not executed yet), and then >>> TS immediately issues another I2C transfer that re-uses the >>> yet-incompleted descriptor. That's my understanding. >> >> OK, but what is the exact sequence that it allowing it to re-use the >> incompleted descriptor? > > TDMA driver DMA Client > > 1. > dmaengine_prep() > > 2. > tegra_dma_desc_get() > dma_desc = kzalloc() > ... > tegra_dma_prep_slave_sg() > INIT_LIST_HEAD(&dma_desc->tx_list); > INIT_LIST_HEAD(&dma_desc->cb_node); > list_add_tail(sgreq->node, > dma_desc->tx_list) > > 3. > dma_async_issue_pending() > > 4. > tegra_dma_tx_submit() > list_splice_tail_init(dma_desc->tx_list, > tdc->pending_sg_req) > > 5. > tegra_dma_isr() > ... > handle_once_dma_done() > ... > sgreq = list_first_entry(tdc->pending_sg_req) > list_del(sgreq->node); > ... > list_add_tail(dma_desc->cb_node, > tdc->cb_desc); > list_add_tail(dma_desc->node, > tdc->free_dma_desc); Isn't this the problem here, that we have placed this on the free list before we are actually done? It seems to me that there could still be a potential race condition between the ISR and the tasklet running. Jon -- nvpublic
