Hi!
One of the places in the kernel where HZ is interfaced with userspace
is in the SG driver. With HZ different from USER_HZ this may result
in all kinds of unexpected timeouts when using SG.
This patch is an attempt to fix this by transforming USER_HZ based timing to
HZ based timing an v.v. A problem I know of is the fact that SG_GET_TIMEOUT
may return a different value than the value that was passed on with
SG_SET_TIMEOUT. Depending on its impact this may need to be solved by
adding a field "user_timeout" to Sg_fd which contains the USER_HZ based value.
Cheers,
Rolf
--- linux-2.5.34.orig/drivers/scsi/sg.c Sun Sep 8 09:00:32 2002
+++ linux-2.5.34/drivers/scsi/sg.c Fri Sep 13 21:30:24 2002
@@ -731,6 +731,17 @@
return 0;
}
+/*
+ * Suppose you want to calculate the formula muldiv(x,m,d)=int(x * m / d)
+ * Then when using 32 bit integers x * m may overflow during the calculation.
+ * Replacing muldiv(x) by muldiv(x)=((x % d) * m) / d + int(x / d) * m
+ * calculates the same, but prevents the overflow when both m and d
+ * are "small" numbers (like HZ and USER_HZ).
+ * Of course an overflow is inavoidable if the result of muldiv doesn't fit
+ * in 32 bits.
+ */
+#define MULDIV(X,MUL,DIV) ((((X % DIV) * MUL) / DIV) + ((X / DIV) * MUL))
+
static int
sg_ioctl(struct inode *inode, struct file *filp,
unsigned int cmd_in, unsigned long arg)
@@ -790,10 +801,15 @@
return result;
if (val < 0)
return -EIO;
- sfp->timeout = val;
+ if (val >= MULDIV (INT_MAX, USER_HZ, HZ))
+ val = MULDIV (INT_MAX, USER_HZ, HZ);
+ sfp->timeout = MULDIV (val, HZ, USER_HZ);
+
return 0;
case SG_GET_TIMEOUT: /* N.B. User receives timeout as return value */
- return sfp->timeout; /* strange ..., for backward compatibility */
+ val = sfp->timeout;
+ val = MULDIV (val, USER_HZ, HZ);
+ return val; /* strange ..., for backward compatibility */
case SG_SET_FORCE_LOW_DMA:
result = get_user(val, (int *) arg);
if (result)