Message ID | 1454366363-10564-1-git-send-email-joshua.henderson@microchip.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On Mon, Feb 01, 2016 at 03:39:23PM -0700, Joshua Henderson wrote: > From: Purna Chandra Mandal <purna.mandal@microchip.com> > There is a BUG in the way SPI_MASTER_MUST_RX/TX is implemented which can create Bug is a WORD like any other WORD... > (1) spi core assigns dummy_rx buffer to transfer.rx_buf member and > (2) passes it to lower layer for handling. and lower layer completed the > transfer/message in due time. > (3) spi core deletes the buffer if no other requests pending, but > 'transfer.rx_buf' continues to hold *stale* dummy buffer pointer. > (4) If spi client driver (like mmc_spi) reuses the same transfer structure and > don't touch .rx_buf to NULL > mmc_spi doesn't reset the ptr unless data transfer direction changes in future > transaction(s). spi core will skip assigning new dummy buffer and underlying > master driver will treat .rx_buf as legitimate ptr. This will result into memory > corruption due to usage of free'd ptr. It's not clear to me that this is the best fix, it's causing problems to free the transfer but we could also fix that by just not freeing the dummy data once we realize we need it unless the adaptor is freed. That should also be more efficient since it saves us having to allocate and free things.
Thanks Mark. On 02/02/2016 04:47 AM, Mark Brown wrote: > On Mon, Feb 01, 2016 at 03:39:23PM -0700, Joshua Henderson wrote: >> From: Purna Chandra Mandal <purna.mandal@microchip.com> >> There is a BUG in the way SPI_MASTER_MUST_RX/TX is implemented which can create > Bug is a WORD like any other WORD... ack. >> (1) spi core assigns dummy_rx buffer to transfer.rx_buf member and >> (2) passes it to lower layer for handling. and lower layer completed the >> transfer/message in due time. >> (3) spi core deletes the buffer if no other requests pending, but >> 'transfer.rx_buf' continues to hold *stale* dummy buffer pointer. >> (4) If spi client driver (like mmc_spi) reuses the same transfer structure and >> don't touch .rx_buf to NULL >> mmc_spi doesn't reset the ptr unless data transfer direction changes in future >> transaction(s). spi core will skip assigning new dummy buffer and underlying >> master driver will treat .rx_buf as legitimate ptr. This will result into memory >> corruption due to usage of free'd ptr. > It's not clear to me that this is the best fix, it's causing problems to > free the transfer but we could also fix that by just not freeing the > dummy data once we realize we need it unless the adaptor is freed. That > should also be more efficient since it saves us having to allocate and > free things. Idea is good, but not sufficient. Dummy buffers are _reallocated_ to accommodate larger size of transfer. In this if [originally NULL] .rx_buf/.tx_buf is not reset back to NULL after the transfer is over spi-core will find those .rx/tx_buf non-NULL in next _transfer() call and will pass the stale rx/tx_buf to spi controller driver which will definitely corrupt the memory and crash the system. Above all the whole design depends on trust that core will not play with any data-structure which will break object-oriented/layered approach. So better to undo the modification (when needed to facilitate some functionality) unless core wants those information to be passed back to caller/client for reporting success or error or else. -- To unsubscribe from this list: send the line "unsubscribe linux-spi" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, Feb 05, 2016 at 10:30:24AM +0530, Purna Chandra Mandal wrote: Please fix your mail client to word wrap within paragraphs at something substantially less than 80 columns. Doing this makes your messages much easier to read and reply to. > Idea is good, but not sufficient. > Dummy buffers are _reallocated_ to accommodate larger size of transfer. In this if > [originally NULL] .rx_buf/.tx_buf is not reset back to NULL after the transfer > is over spi-core will find those .rx/tx_buf non-NULL in next _transfer() call and > will pass the stale rx/tx_buf to spi controller driver which will definitely > corrupt the memory and crash the system. This needs to be clear to readers; a fairly obvious way of dealing with this would be to rellocate down to a page rather than freeing. > Above all the whole design depends on trust that core will not play with any data-structure > which will break object-oriented/layered approach. So better to undo the modification > (when needed to facilitate some functionality) unless core wants those information to be passed > back to caller/client for reporting success or error or else. That's really not the case, we already make a range of other modifications to complete partially filled transfers in order to simplify driver code.
Mark, On 02/08/2016 09:45 PM, Mark Brown wrote: > On Fri, Feb 05, 2016 at 10:30:24AM +0530, Purna Chandra Mandal wrote: > > Please fix your mail client to word wrap within paragraphs at something > substantially less than 80 columns. Doing this makes your messages much > easier to read and reply to. > >> Idea is good, but not sufficient. >> Dummy buffers are _reallocated_ to accommodate larger size of transfer. In this if >> [originally NULL] .rx_buf/.tx_buf is not reset back to NULL after the transfer >> is over spi-core will find those .rx/tx_buf non-NULL in next _transfer() call and >> will pass the stale rx/tx_buf to spi controller driver which will definitely >> corrupt the memory and crash the system. > This needs to be clear to readers; a fairly obvious way of dealing with > this would be to rellocate down to a page rather than freeing. Yea. But current krealloc() implementation allocates new memory if new size is more than the allocated space, (and frees the old). If we allocate PAGE_SIZE of dummy buffer (at first call irrespective of required size), re-use it and don't allow transfer size to grow more than PAGE_SIZE we will be fine provided all SPI clients agree to the size restriction. The moment we'll try to re-allocate new buffer we will reach the same point we wanted to avoid here. >> Above all the whole design depends on trust that core will not play with any data-structure >> which will break object-oriented/layered approach. So better to undo the modification >> (when needed to facilitate some functionality) unless core wants those information to be passed >> back to caller/client for reporting success or error or else. > That's really not the case, we already make a range of other > modifications to complete partially filled transfers in order to > simplify driver code. Purna -- To unsubscribe from this list: send the line "unsubscribe linux-spi" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/drivers/spi/spi.c b/drivers/spi/spi.c index 47eff80..deabd6f 100644 --- a/drivers/spi/spi.c +++ b/drivers/spi/spi.c @@ -819,6 +819,15 @@ static int __spi_unmap_msg(struct spi_master *master, struct spi_message *msg) struct spi_transfer *xfer; struct device *tx_dev, *rx_dev; + if (master->flags & (SPI_MASTER_MUST_RX | SPI_MASTER_MUST_TX)) { + list_for_each_entry(xfer, &msg->transfers, transfer_list) { + if (xfer->rx_buf == master->dummy_rx) + xfer->rx_buf = NULL; + if (xfer->tx_buf == master->dummy_tx) + xfer->tx_buf = NULL; + } + } + if (!master->cur_msg_mapped || !master->can_dma) return 0; @@ -1264,12 +1273,11 @@ void spi_finalize_current_message(struct spi_master *master) unsigned long flags; int ret; + spi_unmap_msg(master, master->cur_msg); spin_lock_irqsave(&master->queue_lock, flags); mesg = master->cur_msg; spin_unlock_irqrestore(&master->queue_lock, flags); - spi_unmap_msg(master, mesg); - if (master->cur_msg_prepared && master->unprepare_message) { ret = master->unprepare_message(master, mesg); if (ret) {