2022-07-09 20:47:10

by Tyler Hicks

[permalink] [raw]
Subject: [PATCH] net/9p: Initialize the iounit field during fid creation

Ensure that the fid's iounit field is set to zero when a new fid is
created. Certain 9P operations, such as OPEN and CREATE, allow the
server to reply with an iounit size which the client code assigns to the
fid struct shortly after the fid is created in p9_fid_create(). Other
operations that follow a call to p9_fid_create(), such as an XATTRWALK,
don't include an iounit value in the reply message from the server. In
the latter case, the iounit field remained uninitialized. Depending on
allocation patterns, the iounit value could have been something
reasonable that was carried over from previously freed fids or, in the
worst case, could have been arbitrary values from non-fid related usages
of the memory location.

The bug was detected in the Windows Subsystem for Linux 2 (WSL2) kernel
after the uninitialized iounit field resulted in the typical sequence of
two getxattr(2) syscalls, one to get the size of an xattr and another
after allocating a sufficiently sized buffer to fit the xattr value, to
hit an unexpected ERANGE error in the second call to getxattr(2). An
uninitialized iounit field would sometimes force rsize to be smaller
than the xattr value size in p9_client_read_once() and the 9P server in
WSL refused to chunk up the READ on the attr_fid and, instead, returned
ERANGE to the client. The virtfs server in QEMU seems happy to chunk up
the READ and this problem goes undetected there. However, there are
likely other non-xattr implications of this bug that could cause
inefficient communication between the client and server.

Cc: [email protected]
Signed-off-by: Tyler Hicks <[email protected]>
---

Note that I haven't had a chance to identify when this bug was
introduced so I don't yet have a proper Fixes tag. The history looked a
little tricky to me but I'll have another look in the coming days. We
started hitting this bug after trying to move from linux-5.10.y to
linux-5.15.y but I didn't see any obvious changes between those two
series. I'm not confident of this theory but perhaps the fid refcounting
changes impacted the fid allocation patterns enough to uncover the
latent bug?

net/9p/client.c | 1 +
1 file changed, 1 insertion(+)

diff --git a/net/9p/client.c b/net/9p/client.c
index 8bba0d9cf975..1dfceb9154f7 100644
--- a/net/9p/client.c
+++ b/net/9p/client.c
@@ -899,6 +899,7 @@ static struct p9_fid *p9_fid_create(struct p9_client *clnt)
fid->clnt = clnt;
fid->rdir = NULL;
fid->fid = 0;
+ fid->iounit = 0;
refcount_set(&fid->count, 1);

idr_preload(GFP_KERNEL);
--
2.25.1


2022-07-10 06:33:37

by Tyler Hicks

[permalink] [raw]
Subject: Re: [PATCH] net/9p: Initialize the iounit field during fid creation

On 2022-07-09 15:00:05, Tyler Hicks wrote:
> Ensure that the fid's iounit field is set to zero when a new fid is
> created. Certain 9P operations, such as OPEN and CREATE, allow the
> server to reply with an iounit size which the client code assigns to the
> fid struct shortly after the fid is created in p9_fid_create(). Other
> operations that follow a call to p9_fid_create(), such as an XATTRWALK,
> don't include an iounit value in the reply message from the server. In
> the latter case, the iounit field remained uninitialized. Depending on
> allocation patterns, the iounit value could have been something
> reasonable that was carried over from previously freed fids or, in the
> worst case, could have been arbitrary values from non-fid related usages
> of the memory location.
>
> The bug was detected in the Windows Subsystem for Linux 2 (WSL2) kernel
> after the uninitialized iounit field resulted in the typical sequence of
> two getxattr(2) syscalls, one to get the size of an xattr and another
> after allocating a sufficiently sized buffer to fit the xattr value, to
> hit an unexpected ERANGE error in the second call to getxattr(2). An
> uninitialized iounit field would sometimes force rsize to be smaller
> than the xattr value size in p9_client_read_once() and the 9P server in
> WSL refused to chunk up the READ on the attr_fid and, instead, returned
> ERANGE to the client. The virtfs server in QEMU seems happy to chunk up
> the READ and this problem goes undetected there. However, there are
> likely other non-xattr implications of this bug that could cause
> inefficient communication between the client and server.
>
> Cc: [email protected]
> Signed-off-by: Tyler Hicks <[email protected]>
> ---
>
> Note that I haven't had a chance to identify when this bug was
> introduced so I don't yet have a proper Fixes tag. The history looked a
> little tricky to me but I'll have another look in the coming days. We
> started hitting this bug after trying to move from linux-5.10.y to
> linux-5.15.y but I didn't see any obvious changes between those two
> series. I'm not confident of this theory but perhaps the fid refcounting
> changes impacted the fid allocation patterns enough to uncover the
> latent bug?

From reading the source, I believe that this first showed up in commit
ebf46264a004 ("fs/9p: Add support user. xattr") which landed in v2.6.36.
Before that commit, p9_client_read(), p9_client_write(), and
p9_client_readdir() were always passed a fid that came from a file's
private_data and went through the open/create functions that initialized
iounit. That commit was the first that passed a fid directly from
p9_fid_create() to p9_client_read().

Tyler

2022-07-10 07:32:09

by Tyler Hicks

[permalink] [raw]
Subject: Re: [PATCH] net/9p: Initialize the iounit field during fid creation

On 2022-07-10 01:26:13, Tyler Hicks wrote:
> On 2022-07-09 15:00:05, Tyler Hicks wrote:
> > Ensure that the fid's iounit field is set to zero when a new fid is
> > created. Certain 9P operations, such as OPEN and CREATE, allow the
> > server to reply with an iounit size which the client code assigns to the
> > fid struct shortly after the fid is created in p9_fid_create(). Other
> > operations that follow a call to p9_fid_create(), such as an XATTRWALK,
> > don't include an iounit value in the reply message from the server. In
> > the latter case, the iounit field remained uninitialized. Depending on
> > allocation patterns, the iounit value could have been something
> > reasonable that was carried over from previously freed fids or, in the
> > worst case, could have been arbitrary values from non-fid related usages
> > of the memory location.
> >
> > The bug was detected in the Windows Subsystem for Linux 2 (WSL2) kernel
> > after the uninitialized iounit field resulted in the typical sequence of
> > two getxattr(2) syscalls, one to get the size of an xattr and another
> > after allocating a sufficiently sized buffer to fit the xattr value, to
> > hit an unexpected ERANGE error in the second call to getxattr(2). An
> > uninitialized iounit field would sometimes force rsize to be smaller
> > than the xattr value size in p9_client_read_once() and the 9P server in
> > WSL refused to chunk up the READ on the attr_fid and, instead, returned
> > ERANGE to the client. The virtfs server in QEMU seems happy to chunk up
> > the READ and this problem goes undetected there. However, there are
> > likely other non-xattr implications of this bug that could cause
> > inefficient communication between the client and server.

^ I think this last sentence can be removed. I now believe that this
only affects xattr get/set operations because nothing else calling the
functions that honor iounit is getting the fid directly from a call to
p9_fid_create().

> >

Please add the following tag:

Fixes: ebf46264a004 ("fs/9p: Add support user. xattr")

I'm happy to do both of these things in a v2 if any changes/improvements
are requested. Thanks!

Tyler

> > Cc: [email protected]
> > Signed-off-by: Tyler Hicks <[email protected]>
> > ---
> >
> > Note that I haven't had a chance to identify when this bug was
> > introduced so I don't yet have a proper Fixes tag. The history looked a
> > little tricky to me but I'll have another look in the coming days. We
> > started hitting this bug after trying to move from linux-5.10.y to
> > linux-5.15.y but I didn't see any obvious changes between those two
> > series. I'm not confident of this theory but perhaps the fid refcounting
> > changes impacted the fid allocation patterns enough to uncover the
> > latent bug?
>
> From reading the source, I believe that this first showed up in commit
> ebf46264a004 ("fs/9p: Add support user. xattr") which landed in v2.6.36.
> Before that commit, p9_client_read(), p9_client_write(), and
> p9_client_readdir() were always passed a fid that came from a file's
> private_data and went through the open/create functions that initialized
> iounit. That commit was the first that passed a fid directly from
> p9_fid_create() to p9_client_read().
>
> Tyler

2022-07-10 07:54:38

by Christophe JAILLET

[permalink] [raw]
Subject: Re: [PATCH] net/9p: Initialize the iounit field during fid creation

Le 09/07/2022 à 22:00, Tyler Hicks a écrit :
> Ensure that the fid's iounit field is set to zero when a new fid is
> created. Certain 9P operations, such as OPEN and CREATE, allow the
> server to reply with an iounit size which the client code assigns to the
> fid struct shortly after the fid is created in p9_fid_create(). Other
> operations that follow a call to p9_fid_create(), such as an XATTRWALK,
> don't include an iounit value in the reply message from the server. In
> the latter case, the iounit field remained uninitialized. Depending on
> allocation patterns, the iounit value could have been something
> reasonable that was carried over from previously freed fids or, in the
> worst case, could have been arbitrary values from non-fid related usages
> of the memory location.
>
> The bug was detected in the Windows Subsystem for Linux 2 (WSL2) kernel
> after the uninitialized iounit field resulted in the typical sequence of
> two getxattr(2) syscalls, one to get the size of an xattr and another
> after allocating a sufficiently sized buffer to fit the xattr value, to
> hit an unexpected ERANGE error in the second call to getxattr(2). An
> uninitialized iounit field would sometimes force rsize to be smaller
> than the xattr value size in p9_client_read_once() and the 9P server in
> WSL refused to chunk up the READ on the attr_fid and, instead, returned
> ERANGE to the client. The virtfs server in QEMU seems happy to chunk up
> the READ and this problem goes undetected there. However, there are
> likely other non-xattr implications of this bug that could cause
> inefficient communication between the client and server.
>
> Cc: [email protected]
> Signed-off-by: Tyler Hicks <[email protected]>
> ---
>
> Note that I haven't had a chance to identify when this bug was
> introduced so I don't yet have a proper Fixes tag. The history looked a
> little tricky to me but I'll have another look in the coming days. We
> started hitting this bug after trying to move from linux-5.10.y to
> linux-5.15.y but I didn't see any obvious changes between those two
> series. I'm not confident of this theory but perhaps the fid refcounting
> changes impacted the fid allocation patterns enough to uncover the
> latent bug?
>
> net/9p/client.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/net/9p/client.c b/net/9p/client.c
> index 8bba0d9cf975..1dfceb9154f7 100644
> --- a/net/9p/client.c
> +++ b/net/9p/client.c
> @@ -899,6 +899,7 @@ static struct p9_fid *p9_fid_create(struct p9_client *clnt)
> fid->clnt = clnt;
> fid->rdir = NULL;
> fid->fid = 0;
> + fid->iounit = 0;
> refcount_set(&fid->count, 1);
>
> idr_preload(GFP_KERNEL);

Hi,
you could also kzalloc 'fid' and remove the memset, "= NULL" and "= 0".
This would be even more future proof and would save some LoC.

Just my 2c,

CJ

2022-07-10 13:45:09

by Dominique Martinet

[permalink] [raw]
Subject: Re: [PATCH] net/9p: Initialize the iounit field during fid creation

Tyler Hicks wrote on Sat, Jul 09, 2022 at 03:00:05PM -0500:
> Ensure that the fid's iounit field is set to zero when a new fid is
> created. Certain 9P operations, such as OPEN and CREATE, allow the
> server to reply with an iounit size which the client code assigns to the
> fid struct shortly after the fid is created in p9_fid_create(). Other
> operations that follow a call to p9_fid_create(), such as an XATTRWALK,
> don't include an iounit value in the reply message from the server. In
> the latter case, the iounit field remained uninitialized. Depending on
> allocation patterns, the iounit value could have been something
> reasonable that was carried over from previously freed fids or, in the
> worst case, could have been arbitrary values from non-fid related usages
> of the memory location.
>
> The bug was detected in the Windows Subsystem for Linux 2 (WSL2) kernel
> after the uninitialized iounit field resulted in the typical sequence of
> two getxattr(2) syscalls, one to get the size of an xattr and another
> after allocating a sufficiently sized buffer to fit the xattr value, to
> hit an unexpected ERANGE error in the second call to getxattr(2). An
> uninitialized iounit field would sometimes force rsize to be smaller
> than the xattr value size in p9_client_read_once() and the 9P server in
> WSL refused to chunk up the READ on the attr_fid and, instead, returned
> ERANGE to the client. The virtfs server in QEMU seems happy to chunk up
> the READ and this problem goes undetected there. However, there are
> likely other non-xattr implications of this bug that could cause
> inefficient communication between the client and server.
>
> Cc: [email protected]
> Signed-off-by: Tyler Hicks <[email protected]>

Thanks for the fix!

> ---
>
> Note that I haven't had a chance to identify when this bug was
> introduced so I don't yet have a proper Fixes tag. The history looked a
> little tricky to me but I'll have another look in the coming days. We
> started hitting this bug after trying to move from linux-5.10.y to
> linux-5.15.y but I didn't see any obvious changes between those two
> series. I'm not confident of this theory but perhaps the fid refcounting
> changes impacted the fid allocation patterns enough to uncover the
> latent bug?
>
> net/9p/client.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/net/9p/client.c b/net/9p/client.c
> index 8bba0d9cf975..1dfceb9154f7 100644
> --- a/net/9p/client.c
> +++ b/net/9p/client.c
> @@ -899,6 +899,7 @@ static struct p9_fid *p9_fid_create(struct p9_client *clnt)
> fid->clnt = clnt;
> fid->rdir = NULL;
> fid->fid = 0;
> + fid->iounit = 0;

ugh, this isn't the first we've missed so I'll be tempted to agree with
Christophe -- let's make that a kzalloc and only set non-zero fields.

--
Dominique

2022-07-10 14:23:46

by Tyler Hicks

[permalink] [raw]
Subject: Re: [PATCH] net/9p: Initialize the iounit field during fid creation

On 2022-07-10 22:21:33, Dominique Martinet wrote:
> Tyler Hicks wrote on Sat, Jul 09, 2022 at 03:00:05PM -0500:
> > Ensure that the fid's iounit field is set to zero when a new fid is
> > created. Certain 9P operations, such as OPEN and CREATE, allow the
> > server to reply with an iounit size which the client code assigns to the
> > fid struct shortly after the fid is created in p9_fid_create(). Other
> > operations that follow a call to p9_fid_create(), such as an XATTRWALK,
> > don't include an iounit value in the reply message from the server. In
> > the latter case, the iounit field remained uninitialized. Depending on
> > allocation patterns, the iounit value could have been something
> > reasonable that was carried over from previously freed fids or, in the
> > worst case, could have been arbitrary values from non-fid related usages
> > of the memory location.
> >
> > The bug was detected in the Windows Subsystem for Linux 2 (WSL2) kernel
> > after the uninitialized iounit field resulted in the typical sequence of
> > two getxattr(2) syscalls, one to get the size of an xattr and another
> > after allocating a sufficiently sized buffer to fit the xattr value, to
> > hit an unexpected ERANGE error in the second call to getxattr(2). An
> > uninitialized iounit field would sometimes force rsize to be smaller
> > than the xattr value size in p9_client_read_once() and the 9P server in
> > WSL refused to chunk up the READ on the attr_fid and, instead, returned
> > ERANGE to the client. The virtfs server in QEMU seems happy to chunk up
> > the READ and this problem goes undetected there. However, there are
> > likely other non-xattr implications of this bug that could cause
> > inefficient communication between the client and server.
> >
> > Cc: [email protected]
> > Signed-off-by: Tyler Hicks <[email protected]>
>
> Thanks for the fix!

No problem!

>
> > ---
> >
> > Note that I haven't had a chance to identify when this bug was
> > introduced so I don't yet have a proper Fixes tag. The history looked a
> > little tricky to me but I'll have another look in the coming days. We
> > started hitting this bug after trying to move from linux-5.10.y to
> > linux-5.15.y but I didn't see any obvious changes between those two
> > series. I'm not confident of this theory but perhaps the fid refcounting
> > changes impacted the fid allocation patterns enough to uncover the
> > latent bug?
> >
> > net/9p/client.c | 1 +
> > 1 file changed, 1 insertion(+)
> >
> > diff --git a/net/9p/client.c b/net/9p/client.c
> > index 8bba0d9cf975..1dfceb9154f7 100644
> > --- a/net/9p/client.c
> > +++ b/net/9p/client.c
> > @@ -899,6 +899,7 @@ static struct p9_fid *p9_fid_create(struct p9_client *clnt)
> > fid->clnt = clnt;
> > fid->rdir = NULL;
> > fid->fid = 0;
> > + fid->iounit = 0;
>
> ugh, this isn't the first we've missed so I'll be tempted to agree with
> Christophe -- let's make that a kzalloc and only set non-zero fields.

Agreed - This is the better approach. V2 will be sent out shortly.

Tyler

>
> --
> Dominique
>