Date: Fri, 5 Apr 2024 09:04:16 -0700
From: Alison Schofield <alison.schofield@...el.com>
To: Kwangjin Ko <kwangjin.ko@...com>
Cc: dave@...olabs.net, jonathan.cameron@...wei.com, dave.jiang@...el.com,
	vishal.l.verma@...el.com, ira.weiny@...el.com,
	dan.j.williams@...el.com, linux-cxl@...r.kernel.org,
	linux-kernel@...r.kernel.org, kernel_team@...ynix.com
Subject: Re: [PATCH v2 1/1] cxl/core: Fix initialization of mbox_cmd.size_out
 in get event

On Tue, Apr 02, 2024 at 05:14:03PM +0900, Kwangjin Ko wrote:
> Since mbox_cmd.size_out is overwritten with the actual output size in
> the function below, it needs to be initialized every time.
> 
> cxl_internal_send_cmd -> __cxl_pci_mbox_send_cmd
> 
> Problem scenario:
> 
> 1) The size_out variable is initially set to the size of the mailbox.
> 2) Read an event.
>    - size_out is set to 160 bytes(header 32B + one event 128B).
>    - Two event are created while reading.
> 3) Read the new *two* events.
>    - size_out is still set to 160 bytes.
>    - Although the value of out_len is 288 bytes, only 160 bytes are
>      copied from the mailbox register to the local variable.
>    - record_count is set to 2.
>    - Accessing records[1] will result in reading incorrect data.

Agree with the other comments on need to set .out_size when doing
cxl_internal_send_cmd() in a loop. Poison list retrieval can hit
this case if the MORE flag is set and a follow on read of the list
delivers more records than the previous read. ie. device gives one
record, sets the _MORE flag, then gives 5.

2 other things appeared to me while looking at this:

First, it seems that there is another cleanup wrt accessing records
with invalid data.  Still focusing on get_events and get_poison
since those loop through output data based on a device supplied
record count. The min_out check means the driver at least gets a
count of records to expect. That's good. The problem occurs::

if (mbox.size_out != struct_size(payload, records, 'record_count'))

The driver will log garbage trace events, and that could lead to
bad actions based on bad data. (like a needless scan of device based
on a false overflow flag). So, checking that size.out is the proper
multiple of record_count protects driver from bad device behavior.

I think that can be combined w the patch Dan is suggesting to
reset mbox.size_out on each loop.

Second thing is the pci-driver quiet handling of PAYLOAD LENGTH
values reported by the device. It seems like at a minimum the
pci-driver could emit an info or debug message when the device
is reporting payload lengths that exceed what the driver can
copy in. I'm referring to the mbox.size_out adjustment in
__cxl_pci_mbox_send_cmd(). Or, if it's not the pci-drivers job
to judge, pass that actual payload length value back in the
mbox structure (new field) so that the cxl-driver can use it.
The pci driver would still do it's "#8 Sanitize the copy" work,
but it would allow the cxl-driver to clearly see why it got the
size_out it got, and squawk about it if needed.

--Alison

> 
> Signed-off-by: Kwangjin Ko <kwangjin.ko@...com>
> ---
>  drivers/cxl/core/mbox.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
> index 9adda4795eb7..a38531a055c8 100644
> --- a/drivers/cxl/core/mbox.c
> +++ b/drivers/cxl/core/mbox.c
> @@ -958,13 +958,14 @@ static void cxl_mem_get_records_log(struct cxl_memdev_state *mds,
>  		.payload_in = &log_type,
>  		.size_in = sizeof(log_type),
>  		.payload_out = payload,
> -		.size_out = mds->payload_size,
>  		.min_out = struct_size(payload, records, 0),
>  	};
>  
>  	do {
>  		int rc, i;
>  
> +		mbox_cmd.size_out = mds->payload_size;
> +
>  		rc = cxl_internal_send_cmd(mds, &mbox_cmd);
>  		if (rc) {
>  			dev_err_ratelimited(dev,
> -- 
> 2.34.1
> 

Powered by blists - more mailing lists