Re: [PATCH] libata: Ensure ata_port probe has completed before detach

[John Garry]

On 16/10/2019 20:09, Jens Axboe wrote:
> On 10/16/19 4:19 AM, John Garry wrote:
> > With CONFIG_DEBUG_TEST_DRIVER_REMOVE set, we may find the following WARN:
> >
> > [   23.452574] ------------[ cut here ]------------
> > [   23.457190] WARNING: CPU: 59 PID: 1 at drivers/ata/libata-core.c:6676 ata_host_detach+0x15c/0x168
> > [   23.466047] Modules linked in:
> > [   23.469092] CPU: 59 PID: 1 Comm: swapper/0 Not tainted 5.4.0-rc1-00010-g5b83fd27752b-dirty #296
> > [   23.477776] Hardware name: Huawei D06 /D06, BIOS Hisilicon D06 UEFI RC0 - V1.16.01 03/15/2019
> > [   23.486286] pstate: a0c00009 (NzCv daif +PAN +UAO)
> > [   23.491065] pc : ata_host_detach+0x15c/0x168
> > [   23.495322] lr : ata_host_detach+0x88/0x168
> > [   23.499491] sp : ffff800011cabb50
> > [   23.502792] x29: ffff800011cabb50 x28: 0000000000000007
> > [   23.508091] x27: ffff80001137f068 x26: ffff8000112c0c28
> > [   23.513390] x25: 0000000000003848 x24: ffff0023ea185300
> > [   23.518689] x23: 0000000000000001 x22: 00000000000014c0
> > [   23.523987] x21: 0000000000013740 x20: ffff0023bdc20000
> > [   23.529286] x19: 0000000000000000 x18: 0000000000000004
> > [   23.534584] x17: 0000000000000001 x16: 00000000000000f0
> > [   23.539883] x15: ffff0023eac13790 x14: ffff0023eb76c408
> > [   23.545181] x13: 0000000000000000 x12: ffff0023eac13790
> > [   23.550480] x11: ffff0023eb76c228 x10: 0000000000000000
> > [   23.555779] x9 : ffff0023eac13798 x8 : 0000000040000000
> > [   23.561077] x7 : 0000000000000002 x6 : 0000000000000001
> > [   23.566376] x5 : 0000000000000002 x4 : 0000000000000000
> > [   23.571674] x3 : ffff0023bf08a0bc x2 : 0000000000000000
> > [   23.576972] x1 : 3099674201f72700 x0 : 0000000000400284
> > [   23.582272] Call trace:
> > [   23.584706]  ata_host_detach+0x15c/0x168
> > [   23.588616]  ata_pci_remove_one+0x10/0x18
> > [   23.592615]  ahci_remove_one+0x20/0x40
> > [   23.596356]  pci_device_remove+0x3c/0xe0
> > [   23.600267]  really_probe+0xdc/0x3e0
> > [   23.603830]  driver_probe_device+0x58/0x100
> > [   23.608000]  device_driver_attach+0x6c/0x90
> > [   23.612169]  __driver_attach+0x84/0xc8
> > [   23.615908]  bus_for_each_dev+0x74/0xc8
> > [   23.619730]  driver_attach+0x20/0x28
> > [   23.623292]  bus_add_driver+0x148/0x1f0
> > [   23.627115]  driver_register+0x60/0x110
> > [   23.630938]  __pci_register_driver+0x40/0x48
> > [   23.635199]  ahci_pci_driver_init+0x20/0x28
> > [   23.639372]  do_one_initcall+0x5c/0x1b0
> > [   23.643199]  kernel_init_freeable+0x1a4/0x24c
> > [   23.647546]  kernel_init+0x10/0x108
> > [   23.651023]  ret_from_fork+0x10/0x18
> > [   23.654590] ---[ end trace 634a14b675b71c13 ]---
> >
> > With KASAN also enabled, we may also get many use-after-free reports.
> >
> > The issue is that when CONFIG_DEBUG_TEST_DRIVER_REMOVE is set, we may
> > attempt to detach the ata_port before it has been probed.
> >
> > This is because the ata_ports are async probed, meaning that there is no
> > guarantee that the ata_port has probed prior to detach. When the ata_port
> > does probe in this scenario, we get all sorts of issues as the detach may
> > have already happened.
> >
> > Fix by ensuring synchronisation with async_synchronize_full(). We could
> > alternatively use the cookie returned from the ata_port probe
> > async_schedule() call, but that means managing the cookie, so more
> > complicated.
> >
> > Signed-off-by: John Garry <john.garry@xxxxxxxxxx>
> > ---
> > Note: This has only been boot tested and manual driver remove/add.
> >            My system has no disk attached to the ahci host.
> >
> > diff --git a/drivers/ata/libata-core.c b/drivers/ata/libata-core.c
> > index 28c492be0a57..74c9b3032d46 100644
> > --- a/drivers/ata/libata-core.c
> > +++ b/drivers/ata/libata-core.c
> > @@ -6708,6 +6708,9 @@ void ata_host_detach(struct ata_host *host)
> >    {
> >    	int i;
> >    
> > +	/* Ensure ata_port probe has completed */
> > +	async_synchronize_full();
> > +
> >    	for (i = 0; i < host->n_ports; i++)
> >    		ata_port_detach(host->ports[i]);
> >    
>
> Nice debugging, and the fix looks appropriate to me. I don't think
> there's any point in trying to individually synchronize cookies.
> I'll let this simmer on the list for a day or two to let other folks
> take a look at it, before queuing it up.

Hi Jens,

FWIW, I did also now test this on qemu with an emulated disk and it was ok.

Anyway, I don't mind if prefer to queue this early for 5.6 so it can sit 
on next for longer.

Cheers,
John