Skip to contentSkip to conversation

net/mlx4_core: Implement the master-slave communication channel

DISCUSSION Dan Carpenter 3 participants 3 messages Nov 26, 2013 – Nov 27, 2013 Kernel janitors

lore
Share & tools
lore
#0 lore raw

Hello Yevgeny Petrilin,

The patch e8f081aacdbf: "net/mlx4_core: Implement the master-slave communication channel" from Dec 13, 2011, leads to the following static checker warning:
"drivers/net/ethernet/mellanox/mlx4/cmd.c:1747 mlx4_master_do_cmd()

	 warn: buffer overflow 'priv->mfunc.master.gen_eqe_mutex' 80 <= 127"

drivers/net/ethernet/mellanox/mlx4/cmd.c
1741 switch (cmd) {
1742 case MLX4_COMM_CMD_VHCR0:
1743 if (slave_state[slave].last_cmd != MLX4_COMM_CMD_RESET)
1744 goto reset_slave;

  1745                  slave_state[slave].vhcr_dma = ((u64) param) << 48;
  1746                  priv->mfunc.master.slave_state[slave].cookie = 0;
  1747                  mutex_init(&priv->mfunc.master.gen_eqe_mutex[slave]);
                                                       ^^^^^^^^^^^^^

There are 80 mutexes in this array. Smatch thinks "slave" can go up to 127 as explained below.

  1748                  break;
  1749          case MLX4_COMM_CMD_VHCR1:
  1750                  if (slave_state[slave].last_cmd != MLX4_COMM_CMD_VHCR0)
  1751                          goto reset_slave;
  1752                  slave_state[slave].vhcr_dma |= ((u64) param) << 32;
  1753                  break;

[ snip ]

  1841          bit_vec = master->comm_arm_bit_vector;
  1842          for (i = 0; i < COMM_CHANNEL_BIT_ARRAY_SIZE; i++) {

i is between 0 and 3.

  1843                  vec = be32_to_cpu(bit_vec[i]);
  1844                  for (j = 0; j < 32; j++) {
  1845                          if (!(vec & (1 << j)))
  1846                                  continue;
  1847                          ++reported;
  1848                          slave = (i * 32) + j;

slave can be 3 * 32 + 31 = 127.

  1849                          comm_cmd = swab32(readl(
  1850                                            &mfunc->comm[slave].slave_write));
  1851                          slt = swab32(readl(&mfunc->comm[slave].slave_read))
  1852                                       >> 31;

regards,
dan carpenter

#1 lore raw
Quoted 12 lines

The patch e8f081aacdbf: "net/mlx4_core: Implement the master-slave communication channel" from Dec 13, 2011, leads to the following static checker warning:
"drivers/net/ethernet/mellanox/mlx4/cmd.c:1747 mlx4_master_do_cmd()

	 warn: buffer overflow 'priv->mfunc.master.gen_eqe_mutex' 80 <> 127"

drivers/net/ethernet/mellanox/mlx4/cmd.c
1741 switch (cmd) {
1742 case MLX4_COMM_CMD_VHCR0:
1743 if (slave_state[slave].last_cmd != MLX4_COMM_CMD_RESET)
1744 goto reset_slave;

Quoted 7 lines
  1745                  slave_state[slave].vhcr_dma = ((u64) param) << 48;
  1746                  priv->mfunc.master.slave_state[slave].cookie = 0;
  1747                  mutex_init(&priv->mfunc.master.gen_eqe_mutex[slave]);
                                                       ^^^^^^^^^^^^^

There are 80 mutexes in this array. Smatch thinks "slave" can go up to 127 as explained below.

Thanks Dan, good catch.
We'll send a fix for this.

Yevgeny

#2 lore raw
In reply to Dan Carpenter

This is correct, sort of. But I agree that the code should be modified to avoid the problem under all conditions.

It is true that the scan is over all the comm-channel-fired bits. However, if a VF does not send something on the comm channel, the bit will not be set. That is why we have not seen any problems at all here. If we want to be strict about it, shorten the scan time, and protect against any FW glitches, we could do the following:

	for (i = 0; i < COMM_CHANNEL_BIT_ARRAY_SIZE; i++) {
		if ((i * 32) >= MLX4_NUM_MFUNC)
			break;
		vec = be32_to_cpu(bit_vec[i]);
		for (j = 0; j < 32; j++) {
			slave = (i * 32) + j;
			if (slave >= MLX4_NUM_MFUNC)
				break;
			if (!(vec & (1 << j)))
				continue;
			++reported;
			comm_cmd = swab32(readl(
					  &mfunc->comm[slave].slave_write));

Or maybe better yet:

	Use " > dev->num_vfs" instead of ">= MLX4_NUM_MFUNC"

This will take care of the potential overrun (which has never happened in practice).

I'll prepare a patch (within the next couple of days -- I'm swamped right now).

-Jack

-----Original Message-----
From: Yevgeny Petrilin
Sent: Wednesday, November 27, 2013 10:30 AM
To: Dan Carpenter; Or Gerlitz
Cc: [email protected]; Jack Morgenstein
Subject: RE: net/mlx4_core: Implement the master-slave communication channel

Quoted 12 lines

The patch e8f081aacdbf: "net/mlx4_core: Implement the master-slave communication channel" from Dec 13, 2011, leads to the following static checker warning:
"drivers/net/ethernet/mellanox/mlx4/cmd.c:1747 mlx4_master_do_cmd()

	 warn: buffer overflow 'priv->mfunc.master.gen_eqe_mutex' 80 <= 127"

drivers/net/ethernet/mellanox/mlx4/cmd.c
1741 switch (cmd) {
1742 case MLX4_COMM_CMD_VHCR0:
1743 if (slave_state[slave].last_cmd != MLX4_COMM_CMD_RESET)
1744 goto reset_slave;

Quoted 8 lines
  1745                  slave_state[slave].vhcr_dma = ((u64) param) << 48;
  1746                  priv->mfunc.master.slave_state[slave].cookie = 0;
  1747                  mutex_init(&priv->mfunc.master.gen_eqe_mutex[slave]);
                                                       ^^^^^^^^^^^^^ 

There are 80 mutexes in this array. Smatch thinks "slave" can go up to
127 as explained below.

Thanks Dan, good catch.
We'll send a fix for this.

Yevgeny