Re: [PATCH -next V2 1/2] sata_fsl: fix UAF in sata_fsl_port_stop when rmmod sata_fsl

From: Sergei Shtylyov
Date: Mon Nov 22 2021 - 13:58:29 EST


On 22.11.2021 5:03, libaokun (A) wrote:

When the `rmmod sata_fsl.ko` command is executed in the PPC64 GNU/Linux,
a bug is reported:
  BUG: Unable to handle kernel data access on read at 0x80000800805b502c
  Oops: Kernel access of bad area, sig: 11 [#1]
  NIP [c0000000000388a4] .ioread32+0x4/0x20
  LR [80000000000c6034] .sata_fsl_port_stop+0x44/0xe0 [sata_fsl]
  Call Trace:
   .free_irq+0x1c/0x4e0 (unreliable)
   .ata_host_stop+0x74/0xd0 [libata]
   .fsl_sata_driver_exit+0x18/0xa20 [sata_fsl]

The triggering of the BUG is shown in the following stack:

       drv->remove(dev) --> platform_drv_remove/platform_remove
         drv->remove(dev) --> sata_fsl_remove
           iounmap(host_priv->hcr_base);            <---- unmap
           kfree(host_priv); <---- free
           dr->node.release(dev, dr->data) --> ata_host_stop
             ap->ops->port_stop(ap) --> sata_fsl_port_stop
                 ioread32(hcr_base + HCONTROL) <---- UAF

The iounmap(host_priv->hcr_base) and kfree(host_priv) commands should


OK! I'm going to modify this in V3.

not be executed in drv->remove. These commands should be executed in
host_stop after port_stop. Therefore, we move these commands to the
new function sata_fsl_host_stop and bind the new function to host_stop
by referring to achi.

  You mean AHCI? I don't see where you reference ahci (or achi)...

Yes, it's AHCI, I'm sorry for a spelling error here..

ahci_platform_ops in drivers/ata/libahci_platform.c

You should have (at least) written "the AHCI platform driver"...

diff --git a/drivers/ata/sata_fsl.c b/drivers/ata/sata_fsl.c
index e5838b23c9e0..30759fd1c3a2 100644
--- a/drivers/ata/sata_fsl.c
+++ b/drivers/ata/sata_fsl.c
@@ -1430,12 +1430,25 @@ static struct ata_port_operations sata_fsl_ops = {
      .pmp_detach = sata_fsl_pmp_detach,
  +static void sata_fsl_host_stop(struct ata_host *host)
+    struct sata_fsl_host_priv *host_priv = host->private_data;
+    iounmap(host_priv->hcr_base);
+    kfree(host_priv);
+static struct ata_port_operations sata_fsl_platform_ops = {
+    .inherits       = &sata_fsl_ops,
+    .host_stop      = sata_fsl_host_stop,

  Why not just add it to the initializer for sata_fsl_ops?

This is the AHCI of the reference.

Most ATA drivers add host_stop to to the  initializer for xxx_platform_ops,

Most? Even if so, I guess they add it this way because they're in the separate modules with the ops they inherit -- in this case it's not so.

such as ahci_platform_ops, ahci_brcm_platform_ops, and ahci_imx_ops.

Note that these are all AHCI drivers, not just (more general) ATA.

It feels like this separates the port operation from the host operation,

Why separate them? The 'struct ata_port_operations' embraces many different aspects of ATA, the arguments do not always include a 'struct *ata_port' (I don't quite like that part in libata).

making the hierarchy of the code clearer.

Clear as mud. In your case, there's no separate modules in play, so blindly parroting what the AHCI platform drivers do gives you nothing but memory waste... :-(


MBR, Sergei

Thank you very much for your advice.

You're welcome. :-)

If there's nothing else to modify, I'll send a patch V3.

Please use a single structure, it's already large enough to have 2 of them in the same module for no good reason.

With Best Regards,
Baokun Li

MBR, Sergei