in agp/isoch.c [64:274]
static int agp_3_5_isochronous_node_enable(struct agp_bridge_data *bridge,
struct agp_3_5_dev *dev_list, unsigned int ndevs)
{
/*
* Convenience structure to make the calculations clearer
* here. The field names come straight from the AGP 3.0 spec.
*/
struct isoch_data {
u32 maxbw;
u32 n;
u32 y;
u32 l;
u32 rq;
struct agp_3_5_dev *dev;
};
struct pci_dev *td = bridge->dev, *dev;
struct list_head *head = &dev_list->list, *pos;
struct agp_3_5_dev *cur;
struct isoch_data *master, target;
unsigned int cdev = 0;
u32 mnistat, tnistat, tstatus, mcmd;
u16 tnicmd, mnicmd;
u32 tot_bw = 0, tot_n = 0, tot_rq = 0, y_max, rq_isoch, rq_async;
u32 step, rem, rem_isoch, rem_async;
int ret = 0;
/*
* We'll work with an array of isoch_data's (one for each
* device in dev_list) throughout this function.
*/
master = kmalloc_array(ndevs, sizeof(*master), GFP_KERNEL);
if (master == NULL) {
ret = -ENOMEM;
goto get_out;
}
/*
* Sort the device list by maxbw. We need to do this because the
* spec suggests that the devices with the smallest requirements
* have their resources allocated first, with all remaining resources
* falling to the device with the largest requirement.
*
* We don't exactly do this, we divide target resources by ndevs
* and split them amongst the AGP 3.0 devices. The remainder of such
* division operations are dropped on the last device, sort of like
* the spec mentions it should be done.
*
* We can't do this sort when we initially construct the dev_list
* because we don't know until this function whether isochronous
* transfers are enabled and consequently whether maxbw will mean
* anything.
*/
agp_3_5_dev_list_sort(dev_list, ndevs);
pci_read_config_dword(td, bridge->capndx+AGPNISTAT, &tnistat);
pci_read_config_dword(td, bridge->capndx+AGPSTAT, &tstatus);
/* Extract power-on defaults from the target */
target.maxbw = (tnistat >> 16) & 0xff;
target.n = (tnistat >> 8) & 0xff;
target.y = (tnistat >> 6) & 0x3;
target.l = (tnistat >> 3) & 0x7;
target.rq = (tstatus >> 24) & 0xff;
y_max = target.y;
/*
* Extract power-on defaults for each device in dev_list. Along
* the way, calculate the total isochronous bandwidth required
* by these devices and the largest requested payload size.
*/
list_for_each(pos, head) {
cur = list_entry(pos, struct agp_3_5_dev, list);
dev = cur->dev;
pci_read_config_dword(dev, cur->capndx+AGPNISTAT, &mnistat);
master[cdev].maxbw = (mnistat >> 16) & 0xff;
master[cdev].n = (mnistat >> 8) & 0xff;
master[cdev].y = (mnistat >> 6) & 0x3;
master[cdev].dev = cur;
tot_bw += master[cdev].maxbw;
y_max = max(y_max, master[cdev].y);
cdev++;
}
/* Check if this configuration has any chance of working */
if (tot_bw > target.maxbw) {
dev_err(&td->dev, "isochronous bandwidth required "
"by AGP 3.0 devices exceeds that which is supported by "
"the AGP 3.0 bridge!\n");
ret = -ENODEV;
goto free_and_exit;
}
target.y = y_max;
/*
* Write the calculated payload size into the target's NICMD
* register. Doing this directly effects the ISOCH_N value
* in the target's NISTAT register, so we need to do this now
* to get an accurate value for ISOCH_N later.
*/
pci_read_config_word(td, bridge->capndx+AGPNICMD, &tnicmd);
tnicmd &= ~(0x3 << 6);
tnicmd |= target.y << 6;
pci_write_config_word(td, bridge->capndx+AGPNICMD, tnicmd);
/* Reread the target's ISOCH_N */
pci_read_config_dword(td, bridge->capndx+AGPNISTAT, &tnistat);
target.n = (tnistat >> 8) & 0xff;
/* Calculate the minimum ISOCH_N needed by each master */
for (cdev=0; cdev<ndevs; cdev++) {
master[cdev].y = target.y;
master[cdev].n = master[cdev].maxbw / (master[cdev].y + 1);
tot_n += master[cdev].n;
}
/* Exit if the minimal ISOCH_N allocation among the masters is more
* than the target can handle. */
if (tot_n > target.n) {
dev_err(&td->dev, "number of isochronous "
"transactions per period required by AGP 3.0 devices "
"exceeds that which is supported by the AGP 3.0 "
"bridge!\n");
ret = -ENODEV;
goto free_and_exit;
}
/* Calculate left over ISOCH_N capability in the target. We'll give
* this to the hungriest device (as per the spec) */
rem = target.n - tot_n;
/*
* Calculate the minimum isochronous RQ depth needed by each master.
* Along the way, distribute the extra ISOCH_N capability calculated
* above.
*/
for (cdev=0; cdev<ndevs; cdev++) {
/*
* This is a little subtle. If ISOCH_Y > 64B, then ISOCH_Y
* byte isochronous writes will be broken into 64B pieces.
* This means we need to budget more RQ depth to account for
* these kind of writes (each isochronous write is actually
* many writes on the AGP bus).
*/
master[cdev].rq = master[cdev].n;
if (master[cdev].y > 0x1)
master[cdev].rq *= (1 << (master[cdev].y - 1));
tot_rq += master[cdev].rq;
}
master[ndevs-1].n += rem;
/* Figure the number of isochronous and asynchronous RQ slots the
* target is providing. */
rq_isoch = (target.y > 0x1) ? target.n * (1 << (target.y - 1)) : target.n;
rq_async = target.rq - rq_isoch;
/* Exit if the minimal RQ needs of the masters exceeds what the target
* can provide. */
if (tot_rq > rq_isoch) {
dev_err(&td->dev, "number of request queue slots "
"required by the isochronous bandwidth requested by "
"AGP 3.0 devices exceeds the number provided by the "
"AGP 3.0 bridge!\n");
ret = -ENODEV;
goto free_and_exit;
}
/* Calculate asynchronous RQ capability in the target (per master) as
* well as the total number of leftover isochronous RQ slots. */
step = rq_async / ndevs;
rem_async = step + (rq_async % ndevs);
rem_isoch = rq_isoch - tot_rq;
/* Distribute the extra RQ slots calculated above and write our
* isochronous settings out to the actual devices. */
for (cdev=0; cdev<ndevs; cdev++) {
cur = master[cdev].dev;
dev = cur->dev;
master[cdev].rq += (cdev == ndevs - 1)
? (rem_async + rem_isoch) : step;
pci_read_config_word(dev, cur->capndx+AGPNICMD, &mnicmd);
pci_read_config_dword(dev, cur->capndx+AGPCMD, &mcmd);
mnicmd &= ~(0xff << 8);
mnicmd &= ~(0x3 << 6);
mcmd &= ~(0xff << 24);
mnicmd |= master[cdev].n << 8;
mnicmd |= master[cdev].y << 6;
mcmd |= master[cdev].rq << 24;
pci_write_config_dword(dev, cur->capndx+AGPCMD, mcmd);
pci_write_config_word(dev, cur->capndx+AGPNICMD, mnicmd);
}
free_and_exit:
kfree(master);
get_out:
return ret;
}