double-check me?
Subsystems (2)
Maybe you guys can spot something I'm missing here. alloc_netdev is supposed to guarantee that dev->priv is aligned by 32 bytes:
struct net_device *alloc_netdev(int sizeof_priv, const char *mask,
void (*setup)(struct net_device *))
{
struct net_device *dev;
int alloc_size;
/* ensure 32-byte alignment of the private area */
alloc_size = sizeof (*dev) + sizeof_priv + 31;
dev = (struct net_device *) kmalloc (alloc_size, GFP_KERNEL);
if (dev == NULL)
{
printk(KERN_ERR "alloc_dev: Unable to allocate device
memory.\n"
);
return NULL;
}
memset(dev, 0, alloc_size);
if (sizeof_priv)
dev->priv = (void *) (((long)(dev + 1) + 31) & ~31);Now... shouldn't that last line of code be "dev + 1 + sizeof(*dev)" ?
It seems to work 2.[456] for a long time, so I am doubting myself... surely it would have caused memory corruption or something by now if I have really found a bug.
Jeff
Quoted 9 lines · Jeff Garzik
Maybe you guys can spot something I'm missing here.
alloc_netdevis supposed to guarantee thatdev->privis aligned by 32 bytes:struct net_device *dev; if (sizeof_priv) dev->priv = (void *) (((long)(dev + 1) + 31) & ~31);
Quoted 5 lines
Now... shouldn't that last line of code be "dev + 1 +
sizeof(*dev)" ?It seems to work 2.[456] for a long time, so I am doubting myself... surely it would have caused memory corruption or something by now if I have really found a bug.
Looks ok... not very readable, though.
dev + 1 = ((u8*)dev) + sizeof(dev) = pointer to end of net_device struct.
(X + 31) & ~31 makes sure X is 32-bytes aligned (0-31 bytes are added). Hope the alloc_size has enough space for this.
31 should be better 0x1F I think.
I also like sizeof(struct net_device) more than sizeof(*dev).
Why do you want "+ 1" if you add sizeof(*dev)?
Signature · 2 lines
Krzysztof Halasa Network Administrator
Quoted 5 lines · [email protected] said:
if (sizeof_priv) dev->priv = (void *) (((long)(dev + 1) + 31) & ~31);Now... shouldn't that last line of code be "dev + 1 +
sizeof(*dev)" ?
are you missing that the "dev + 1" pointer arithmetic is already adding sizeof(*dev) to dev, rather than just 1 byte? "(dev + 1)" is a pointer to the private area after the actual struct net_device, and "((long)(dev + 1) + 31)" adds 31 bytes of padding. The final "& ~31" chops off any excess padding from the 31 that was added and actually aligns the pointer.
Seems right to me, but I'm not used to playing alignment tricks.
Jason
Quoted 14 lines · Krzysztof Halasa
Quoted 11 lines · Jeff Garzik
Maybe you guys can spot something I'm missing here.
alloc_netdevis supposed to guarantee thatdev->privis aligned by 32 bytes:struct net_device *dev; if (sizeof_priv) dev->priv = (void *) (((long)(dev + 1) + 31) & ~31);Now... shouldn't that last line of code be "dev + 1 +
sizeof(*dev)" ?
Quoted 8 lines
Quoted 3 lines
It seems to work 2.[456] for a long time, so I am doubting myself... surely it would have caused memory corruption or something by now if I have really found a bug.
Looks ok... not very readable, though.
dev + 1 = ((u8*)dev) + sizeof(dev) = pointer to end of
net_devicestruct.
C pointer arithmatic. Ok, duh, thanks :)
Jeff, too used to void pointer arith
Quoted 2 lines · [email protected] said:
C pointer arithmatic. Ok, duh, thanks :)
Jeff, too used to void pointer arith
Don't get _too_ used to it. arithmetic on (void *) is a gcc extension. it's undefined in ansi C.
Jason
Quoted 6 lines · Jason Lunz
Quoted 2 lines · [email protected] said:
C pointer arithmatic. Ok, duh, thanks :)
Jeff, too used to void pointer arithDon't get _too_ used to it. arithmetic on (void *) is a gcc extension. it's undefined in ansi C.
Yes, I know.
It's used extensively in the kernel, however.
Jeff