Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Subject: Re: [PATCH v3 1/2] tpm: fix reference counting for struct tpm_chip
To:     Jason Gunthorpe <jgg@ziepe.ca>,
        Lino Sanfilippo <l.sanfilippo@kunbus.com>
Cc:     peterhuewe@gmx.de, jarkko@kernel.org, stefanb@linux.vnet.ibm.com,
        James.Bottomley@hansenpartnership.com, stable@vger.kernel.org,
        linux-integrity@vger.kernel.org, linux-kernel@vger.kernel.org
References: <1612482643-11796-1-git-send-email-LinoSanfilippo@gmx.de>
 <1612482643-11796-2-git-send-email-LinoSanfilippo@gmx.de>
 <20210205130511.GI4718@ziepe.ca>
 <3b821bf9-0f54-3473-d934-61c0c29f8957@kunbus.com>
 <20210205151511.GM4718@ziepe.ca>
 <f6e5dd7d-30df-26d9-c712-677c127a8026@kunbus.com>
 <20210205155808.GO4718@ziepe.ca>
From:   Lino Sanfilippo <LinoSanfilippo@gmx.de>
Message-ID: <db7c90c3-d86a-65c9-81a2-be1527b47e11@gmx.de>
Date:   Fri, 5 Feb 2021 22:50:02 +0100
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101
 Thunderbird/68.10.0
MIME-Version: 1.0
In-Reply-To: <20210205155808.GO4718@ziepe.ca>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On 05.02.21 at 16:58, Jason Gunthorpe wrote:
eference in the first place).
>
> No, they are all chained together because they are all in the same
> struct:
>
> struct tpm_chip {
> 	struct device dev;
> 	struct device devs;
> 	struct cdev cdev;
> 	struct cdev cdevs;
>
> dev holds the refcount on memory, when it goes 0 the whole thing is
> kfreed.
>
> The rule is dev's refcount can't go to zero while any other refcount
> is !=3D 0.
>
> For instance devs holds a get on dev that is put back only when devs
> goes to 0:
>
> static void tpm_devs_release(struct device *dev)
> {
> 	struct tpm_chip *chip =3D container_of(dev, struct tpm_chip, devs);
>
> 	/* release the master device reference */
> 	put_device(&chip->dev);
> }
>
> Both cdev elements do something similar inside the cdev layer.

Well this chaining is exactly what does not work nowadays and what the pat=
ch is supposed
to fix: currently we dont ever take the extra ref (not even in TPM 2 case,=
 note that
TPM_CHIP_FLAG_TMP2 is never set), so

-	if (chip->flags & TPM_CHIP_FLAG_TPM2)
-		get_device(&chip->dev);
+	get_device(&chip->dev);


and tpm_devs_release() is never called, since there is nothing that ever p=
uts devs, so


+	rc =3D devm_add_action_or_reset(pdev,
+				      (void (*)(void *)) put_device,
+				      &chip->devs);


The race with only get_device()/putdevice() in tpm_common_open()/tpm_commo=
n_release() is:

1. tpm chip is allocated with dev refcount =3D 1, devs refcount =3D 1
2. /dev/tpmrm is opened but before we get the ref to dev in tpm_common() a=
nother thread
rmmmods the chip driver:
3. the chip is unregistered, dev is put with refcount =3D 0 and the whole =
chip struct is freed
3. Now open() proceeds, tries to grab the extra ref chip->dev from a chip =
that has already
been deallocated and the system crashes.

As I already wrote, that approach was my first thought, too, but since the=
 result crashed due to the
race condition, I chose the approach in patch 1.

Regards,
Lino

> The net result is during any open() the tpm_chip is guarenteed to have
> a positive refcount.
>