On Fri, Jan 04 2019 at 14:02 -0700, Evan Green wrote:
On Thu, Jan 3, 2019 at 9:47 AM Stephen Boyd <[email protected]> wrote:Using the batch API from the interconnect driver sometimes leads to a KASAN error due to an access to freed memory. This is easier to trigger with threadirqs on the kernel commandline. BUG: KASAN: use-after-free in rpmh_tx_done+0x114/0x12c Read of size 1 at addr fffffff51414ad84 by task irq/110-apps_rs/57 CPU: 0 PID: 57 Comm: irq/110-apps_rs Tainted: G W 4.19.10 #72 Call trace: dump_backtrace+0x0/0x2f8 show_stack+0x20/0x2c __dump_stack+0x20/0x28 dump_stack+0xcc/0x10c print_address_description+0x74/0x240 kasan_report+0x250/0x26c __asan_report_load1_noabort+0x20/0x2c rpmh_tx_done+0x114/0x12c tcs_tx_done+0x450/0x768 irq_forced_thread_fn+0x58/0x9c irq_thread+0x120/0x1dc kthread+0x248/0x260 ret_from_fork+0x10/0x18 Allocated by task 385: kasan_kmalloc+0xac/0x148 __kmalloc+0x170/0x1e4 rpmh_write_batch+0x174/0x540 qcom_icc_set+0x8dc/0x9ac icc_set+0x288/0x2e8 a6xx_gmu_stop+0x320/0x3c0 a6xx_pm_suspend+0x108/0x124 adreno_suspend+0x50/0x60 pm_generic_runtime_suspend+0x60/0x78 __rpm_callback+0x214/0x32c rpm_callback+0x54/0x184 rpm_suspend+0x3f8/0xa90 pm_runtime_work+0xb4/0x178 process_one_work+0x544/0xbc0 worker_thread+0x514/0x7d0 kthread+0x248/0x260 ret_from_fork+0x10/0x18 Freed by task 385: __kasan_slab_free+0x12c/0x1e0 kasan_slab_free+0x10/0x1c kfree+0x134/0x588 rpmh_write_batch+0x49c/0x540 qcom_icc_set+0x8dc/0x9ac icc_set+0x288/0x2e8 a6xx_gmu_stop+0x320/0x3c0 a6xx_pm_suspend+0x108/0x124 adreno_suspend+0x50/0x60 cr50_spi spi5.0: SPI transfer timed out pm_generic_runtime_suspend+0x60/0x78 __rpm_callback+0x214/0x32c rpm_callback+0x54/0x184 rpm_suspend+0x3f8/0xa90 pm_runtime_work+0xb4/0x178 process_one_work+0x544/0xbc0 worker_thread+0x514/0x7d0 kthread+0x248/0x260 ret_from_fork+0x10/0x18 The buggy address belongs to the object at fffffff51414ac80 which belongs to the cache kmalloc-512 of size 512 The buggy address is located 260 bytes inside of 512-byte region [fffffff51414ac80, fffffff51414ae80) The buggy address belongs to the page: page:ffffffbfd4505200 count:1 mapcount:0 mapping:fffffff51e00c680 index:0x0 compound_mapcount: 0 flags: 0x4000000000008100(slab|head) raw: 4000000000008100 ffffffbfd4529008 ffffffbfd44f9208 fffffff51e00c680 raw: 0000000000000000 0000000000200020 00000001ffffffff 0000000000000000 page dumped because: kasan: bad access detected Memory state around the buggy address: fffffff51414ac80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fffffff51414ad00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb >fffffff51414ad80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ^ fffffff51414ae00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fffffff51414ae80: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc The batch API sets the same completion for each rpmh message that's sent and then loops through all the messages and waits for that single completion declared on the stack to be completed before returning from the function and freeing the message structures. Unfortunately, some messages may still be in process and 'stuck' in the TCS. At some later point, the tcs_tx_done() interrupt will run and try to process messages that have already been freed at the end of rpmh_write_batch(). This will in turn access the 'needs_free' member of the rpmh_request structure and cause KASAN to complain. Let's fix this by allocating a chunk of completions for each message and waiting for all of them to be completed before returning from the batch API. Alternatively, we could wait for the last message in the batch, but that may be a more complicated change because it looks like tcs_tx_done() just iterates through the indices of the queue and completes each message instead of tracking the last inserted message and completing that first. Cc: Lina Iyer <[email protected]> Cc: "Raju P.L.S.S.S.N" <[email protected]> Cc: Matthias Kaehlcke <[email protected]> Cc: Evan Green <[email protected]> Fixes: c8790cb6da58 ("drivers: qcom: rpmh: add support for batch RPMH request") Signed-off-by: Stephen Boyd <[email protected]> --- drivers/soc/qcom/rpmh.c | 25 +++++++++++++++++-------- 1 file changed, 17 insertions(+), 8 deletions(-) diff --git a/drivers/soc/qcom/rpmh.c b/drivers/soc/qcom/rpmh.c index c7beb6841289..3b3e8b0b2d95 100644 --- a/drivers/soc/qcom/rpmh.c +++ b/drivers/soc/qcom/rpmh.c @@ -348,11 +348,12 @@ int rpmh_write_batch(const struct device *dev, enum rpmh_state state, { struct batch_cache_req *req; struct rpmh_request *rpm_msgs; - DECLARE_COMPLETION_ONSTACK(compl); + struct completion *compls; struct rpmh_ctrlr *ctrlr = get_rpmh_ctrlr(dev); unsigned long time_left; int count = 0; int ret, i, j; + void *ptr; if (!cmd || !n) return -EINVAL; @@ -362,10 +363,15 @@ int rpmh_write_batch(const struct device *dev, enum rpmh_state state, if (!count) return -EINVAL; - req = kzalloc(sizeof(*req) + count * sizeof(req->rpm_msgs[0]), + ptr = kzalloc(sizeof(*req) + + count * (sizeof(req->rpm_msgs[0]) + sizeof(*compls)), GFP_ATOMIC); - if (!req) + if (!ptr) return -ENOMEM; + + req = ptr; + compls = ptr + sizeof(*req) + count * sizeof(*rpm_msgs); + req->count = count; rpm_msgs = req->rpm_msgs; @@ -380,7 +386,10 @@ int rpmh_write_batch(const struct device *dev, enum rpmh_state state, } for (i = 0; i < count; i++) { - rpm_msgs[i].completion = &compl; + struct completion *compl = &compls[i]; + + init_completion(compl); + rpm_msgs[i].completion = compl; ret = rpmh_rsc_send_data(ctrlr_to_drv(ctrlr), &rpm_msgs[i].msg); if (ret) { pr_err("Error(%d) sending RPMH message addr=%#x\n",It's a little weird that we call rpmh_tx_done on a bunch of transfers we never submitted, just so the completion will get signaled so we can wait on it in the next loop. We could just do count = i; break; here instead.
It seems like it was carried over from my earlier submissions, where I was reference counting the number of completions for a batch. I beleive, with what we are doing here, we don't need to call tx_done with this approach.
@@ -393,12 +402,12 @@ int rpmh_write_batch(const struct device *dev, enum rpmh_state state, time_left = RPMH_TIMEOUT_MS; for (i = 0; i < count; i++) { - time_left = wait_for_completion_timeout(&compl, time_left); + time_left = wait_for_completion_timeout(&compls[i], time_left);So we give RPMH_TIMEOUT_MS for all the completions to finish. I wonder if it would be better to have that as RPMH_TIMEOUT_MS per completion.
That would work too. RPMH_TIMEOUT_MS is a large number that it should not be a problem with this approach either. Thanks, Lina
