Hi Brian, On Mon, Oct 10, 2016 at 4:47 PM, Brian Norris <briannorris@xxxxxxxxxxxx> wrote: > Hi Dmitry, > > On Mon, Oct 10, 2016 at 04:43:06PM -0700, Dmitry Torokhov wrote: >> On Thu, Oct 6, 2016 at 6:03 AM, Amitkumar Karwar <akarwar@xxxxxxxxxxx> wrote: >> >> From: linux-wireless-owner@xxxxxxxxxxxxxxx [mailto:linux-wireless- >> >> owner@xxxxxxxxxxxxxxx] On Behalf Of Brian Norris >> >> Sent: Wednesday, October 05, 2016 10:00 PM >> >> To: Amitkumar Karwar >> >> Cc: linux-wireless@xxxxxxxxxxxxxxx; Cathy Luo; Nishant Sarmukadam; >> >> rajatja@xxxxxxxxxx; Xinming Hu >> >> Subject: Re: [PATCH v2 1/2] mwifiex: reset card->adapter during device >> >> unregister >> >> >> >> On Wed, Oct 05, 2016 at 02:04:53PM +0000, Amitkumar Karwar wrote: >> >> > > From: Brian Norris [mailto:briannorris@xxxxxxxxxxxx] >> >> > > Sent: Wednesday, October 05, 2016 3:28 AM >> >> > > To: Amitkumar Karwar >> >> > > Cc: linux-wireless@xxxxxxxxxxxxxxx; Cathy Luo; Nishant Sarmukadam; >> >> > > rajatja@xxxxxxxxxx; briannorris@xxxxxxxxxx; Xinming Hu >> >> > > Subject: Re: [PATCH v2 1/2] mwifiex: reset card->adapter during >> >> > > device unregister >> >> > > >> >> > > On Tue, Oct 04, 2016 at 10:38:24PM +0530, Amitkumar Karwar wrote: >> >> >> >> > > > --- a/drivers/net/wireless/marvell/mwifiex/pcie.c >> >> > > > +++ b/drivers/net/wireless/marvell/mwifiex/pcie.c >> >> > > > @@ -3042,6 +3042,7 @@ static void mwifiex_unregister_dev(struct >> >> > > mwifiex_adapter *adapter) >> >> > > > pci_disable_msi(pdev); >> >> > > > } >> >> > > > } >> >> > > > + card->adapter = NULL; >> >> > > >> >> > > I think you have a similar problem here as in patch 2; there is no >> >> > > locking to protect fields in struct pcie_service_card or struct >> >> > > sdio_mmc_card below. That problem kind of already exists, except >> >> > > that you only write the value of card->adapter once at registration >> >> > > time, so it's not actually unsafe. But now that you're introducing a >> >> > > second write, you have a problem. >> >> > > >> >> > > Brian >> >> > > >> >> > >> >> > We have a global "add_remove_card_sem" semaphore in our code for >> >> > synchronizing initialization and teardown threads. Ideally "init + >> >> > teardown/reboot" should not have a race issue with this logic >> >> > >> >> > Later there was a loophole introduced in this after using async >> >> > firmware download API. During initialization, firmware downloads >> >> > asynchronously in a separate thread where might have released the >> >> > semaphore. I am working on a patch to fix this. >> >> > >> >> > So "card->adapter" doesn't need to have locking. Even if we have two >> >> > write operations, those two threads can't run simultaneously due to >> >> > above mentioned logic. >> >> >> >> What about writes racing with reads? You have lots of unsynchronized >> >> cases that read this, although most of them should be halted by now >> >> (e.g., cmd processing). I was looking at suspend() in particular, which >> >> I thought you were looking at in this patch series. >> > >> > Please note that "card->adapter" is used only in pcie.c/sdio.c/usb.c files >> > >> > Writes won't have race with reads. >> > >> > 1) write 1 --- "card->adapter = adapter;" in mwifiex_register_dev() >> > This place is at the beginning of initialization. >> > mwifiex_pcie_probe() -> mwifiex_add_card() -> adapter->if_ops.register_dev() >> > There is no chance that "card->adapter" is read anywhere at this point. FW is not yet downloaded >> > >> > 2) write 2 ---- "card->adapter = NULL;" in mwifiex_unregister_dev() >> > This place the end of teardown phase. >> > Interrupts are disabled and all cleanup is done. We have "card->adapter" NULL checks at entry point of suspend/remove/resume, if they get called after this. >> >> How exactly are you preventing ->suspend() from being called *while* >> you are running mwifiex_unregister_dev()? I.e. user slamming the lid >> of a laptop and throwing it in their backpack? > > As I already commented, isn't this primarily [*] called from the driver > remove() callback? remove() doesn't race with suspend(), does it? Nope, there is request_firmware_nowait() path that is asynchronous with device registration/unregistration and power management. Maybe the right way is to move the driver to async probing and switch that request_firmware_nowait() into plain request_firmware(). I haven't looked that closely. The wording "we may end up accessing invalid memory in some corner cases" which is a synonym for "the code is racy" caught my eye, thus the response on the patch. Thanks. -- Dmitry
