2 * Setup routines for AGP 3.5 compliant bridges.
5 #include <linux/list.h>
7 #include <linux/agp_backend.h>
8 #include <linux/module.h>
9 #include <linux/slab.h>
13 /* Generic AGP 3.5 enabling routines */
16 struct list_head list;
22 static void agp_3_5_dev_list_insert(struct list_head *head, struct list_head *new)
24 struct agp_3_5_dev *cur, *n = list_entry(new, struct agp_3_5_dev, list);
25 struct list_head *pos;
27 list_for_each(pos, head) {
28 cur = list_entry(pos, struct agp_3_5_dev, list);
29 if(cur->maxbw > n->maxbw)
32 list_add_tail(new, pos);
35 static void agp_3_5_dev_list_sort(struct agp_3_5_dev *list, unsigned int ndevs)
37 struct agp_3_5_dev *cur;
39 struct list_head *pos, *tmp, *head = &list->list, *start = head->next;
44 for (pos=start; pos!=head; ) {
45 cur = list_entry(pos, struct agp_3_5_dev, list);
48 pci_read_config_dword(dev, cur->capndx+AGPNISTAT, &nistat);
49 cur->maxbw = (nistat >> 16) & 0xff;
53 agp_3_5_dev_list_insert(head, tmp);
58 * Initialize all isochronous transfer parameters for an AGP 3.0
59 * node (i.e. a host bridge in combination with the adapters
63 static int agp_3_5_isochronous_node_enable(struct agp_bridge_data *bridge,
64 struct agp_3_5_dev *dev_list, unsigned int ndevs)
67 * Convenience structure to make the calculations clearer
68 * here. The field names come straight from the AGP 3.0 spec.
76 struct agp_3_5_dev *dev;
79 struct pci_dev *td = bridge->dev, *dev;
80 struct list_head *head = &dev_list->list, *pos;
81 struct agp_3_5_dev *cur;
82 struct isoch_data *master, target;
83 unsigned int cdev = 0;
84 u32 mnistat, tnistat, tstatus, mcmd;
87 u32 tot_bw = 0, tot_n = 0, tot_rq = 0, y_max, rq_isoch, rq_async;
88 u32 step, rem, rem_isoch, rem_async;
92 * We'll work with an array of isoch_data's (one for each
93 * device in dev_list) throughout this function.
95 if ((master = kmalloc(ndevs * sizeof(*master), GFP_KERNEL)) == NULL) {
101 * Sort the device list by maxbw. We need to do this because the
102 * spec suggests that the devices with the smallest requirements
103 * have their resources allocated first, with all remaining resources
104 * falling to the device with the largest requirement.
106 * We don't exactly do this, we divide target resources by ndevs
107 * and split them amongst the AGP 3.0 devices. The remainder of such
108 * division operations are dropped on the last device, sort of like
109 * the spec mentions it should be done.
111 * We can't do this sort when we initially construct the dev_list
112 * because we don't know until this function whether isochronous
113 * transfers are enabled and consequently whether maxbw will mean
116 agp_3_5_dev_list_sort(dev_list, ndevs);
118 pci_read_config_dword(td, bridge->capndx+AGPNISTAT, &tnistat);
119 pci_read_config_dword(td, bridge->capndx+AGPSTAT, &tstatus);
121 /* Extract power-on defaults from the target */
122 target.maxbw = (tnistat >> 16) & 0xff;
123 target.n = (tnistat >> 8) & 0xff;
124 target.y = (tnistat >> 6) & 0x3;
125 target.l = (tnistat >> 3) & 0x7;
126 target.rq = (tstatus >> 24) & 0xff;
131 * Extract power-on defaults for each device in dev_list. Along
132 * the way, calculate the total isochronous bandwidth required
133 * by these devices and the largest requested payload size.
135 list_for_each(pos, head) {
136 cur = list_entry(pos, struct agp_3_5_dev, list);
139 mcapndx = cur->capndx;
141 pci_read_config_dword(dev, cur->capndx+AGPNISTAT, &mnistat);
143 master[cdev].maxbw = (mnistat >> 16) & 0xff;
144 master[cdev].n = (mnistat >> 8) & 0xff;
145 master[cdev].y = (mnistat >> 6) & 0x3;
146 master[cdev].dev = cur;
148 tot_bw += master[cdev].maxbw;
149 y_max = max(y_max, master[cdev].y);
154 /* Check if this configuration has any chance of working */
155 if (tot_bw > target.maxbw) {
156 printk(KERN_ERR PFX "isochronous bandwidth required "
157 "by AGP 3.0 devices exceeds that which is supported by "
158 "the AGP 3.0 bridge!\n");
166 * Write the calculated payload size into the target's NICMD
167 * register. Doing this directly effects the ISOCH_N value
168 * in the target's NISTAT register, so we need to do this now
169 * to get an accurate value for ISOCH_N later.
171 pci_read_config_word(td, bridge->capndx+AGPNICMD, &tnicmd);
172 tnicmd &= ~(0x3 << 6);
173 tnicmd |= target.y << 6;
174 pci_write_config_word(td, bridge->capndx+AGPNICMD, tnicmd);
176 /* Reread the target's ISOCH_N */
177 pci_read_config_dword(td, bridge->capndx+AGPNISTAT, &tnistat);
178 target.n = (tnistat >> 8) & 0xff;
180 /* Calculate the minimum ISOCH_N needed by each master */
181 for (cdev=0; cdev<ndevs; cdev++) {
182 master[cdev].y = target.y;
183 master[cdev].n = master[cdev].maxbw / (master[cdev].y + 1);
185 tot_n += master[cdev].n;
188 /* Exit if the minimal ISOCH_N allocation among the masters is more
189 * than the target can handle. */
190 if (tot_n > target.n) {
191 printk(KERN_ERR PFX "number of isochronous "
192 "transactions per period required by AGP 3.0 devices "
193 "exceeds that which is supported by the AGP 3.0 "
199 /* Calculate left over ISOCH_N capability in the target. We'll give
200 * this to the hungriest device (as per the spec) */
201 rem = target.n - tot_n;
204 * Calculate the minimum isochronous RQ depth needed by each master.
205 * Along the way, distribute the extra ISOCH_N capability calculated
208 for (cdev=0; cdev<ndevs; cdev++) {
210 * This is a little subtle. If ISOCH_Y > 64B, then ISOCH_Y
211 * byte isochronous writes will be broken into 64B pieces.
212 * This means we need to budget more RQ depth to account for
213 * these kind of writes (each isochronous write is actually
214 * many writes on the AGP bus).
216 master[cdev].rq = master[cdev].n;
217 if(master[cdev].y > 0x1)
218 master[cdev].rq *= (1 << (master[cdev].y - 1));
220 tot_rq += master[cdev].rq;
222 master[ndevs-1].n += rem;
224 /* Figure the number of isochronous and asynchronous RQ slots the
225 * target is providing. */
226 rq_isoch = (target.y > 0x1) ? target.n * (1 << (target.y - 1)) : target.n;
227 rq_async = target.rq - rq_isoch;
229 /* Exit if the minimal RQ needs of the masters exceeds what the target
231 if (tot_rq > rq_isoch) {
232 printk(KERN_ERR PFX "number of request queue slots "
233 "required by the isochronous bandwidth requested by "
234 "AGP 3.0 devices exceeds the number provided by the "
235 "AGP 3.0 bridge!\n");
240 /* Calculate asynchronous RQ capability in the target (per master) as
241 * well as the total number of leftover isochronous RQ slots. */
242 step = rq_async / ndevs;
243 rem_async = step + (rq_async % ndevs);
244 rem_isoch = rq_isoch - tot_rq;
246 /* Distribute the extra RQ slots calculated above and write our
247 * isochronous settings out to the actual devices. */
248 for (cdev=0; cdev<ndevs; cdev++) {
249 cur = master[cdev].dev;
252 mcapndx = cur->capndx;
254 master[cdev].rq += (cdev == ndevs - 1)
255 ? (rem_async + rem_isoch) : step;
257 pci_read_config_word(dev, cur->capndx+AGPNICMD, &mnicmd);
258 pci_read_config_dword(dev, cur->capndx+AGPCMD, &mcmd);
260 mnicmd &= ~(0xff << 8);
261 mnicmd &= ~(0x3 << 6);
262 mcmd &= ~(0xff << 24);
264 mnicmd |= master[cdev].n << 8;
265 mnicmd |= master[cdev].y << 6;
266 mcmd |= master[cdev].rq << 24;
268 pci_write_config_dword(dev, cur->capndx+AGPCMD, mcmd);
269 pci_write_config_word(dev, cur->capndx+AGPNICMD, mnicmd);
280 * This function basically allocates request queue slots among the
281 * AGP 3.0 systems in nonisochronous nodes. The algorithm is
282 * pretty stupid, divide the total number of RQ slots provided by the
283 * target by ndevs. Distribute this many slots to each AGP 3.0 device,
284 * giving any left over slots to the last device in dev_list.
286 static void agp_3_5_nonisochronous_node_enable(struct agp_bridge_data *bridge,
287 struct agp_3_5_dev *dev_list, unsigned int ndevs)
289 struct agp_3_5_dev *cur;
290 struct list_head *head = &dev_list->list, *pos;
293 unsigned int cdev = 0;
295 pci_read_config_dword(bridge->dev, bridge->capndx+AGPSTAT, &tstatus);
297 trq = (tstatus >> 24) & 0xff;
300 rem = mrq + (trq % ndevs);
302 for (pos=head->next; cdev<ndevs; cdev++, pos=pos->next) {
303 cur = list_entry(pos, struct agp_3_5_dev, list);
305 pci_read_config_dword(cur->dev, cur->capndx+AGPCMD, &mcmd);
306 mcmd &= ~(0xff << 24);
307 mcmd |= ((cdev == ndevs - 1) ? rem : mrq) << 24;
308 pci_write_config_dword(cur->dev, cur->capndx+AGPCMD, mcmd);
313 * Fully configure and enable an AGP 3.0 host bridge and all the devices
316 int agp_3_5_enable(struct agp_bridge_data *bridge)
318 struct pci_dev *td = bridge->dev, *dev = NULL;
321 u32 tstatus, mstatus, ncapid;
324 struct agp_3_5_dev *dev_list, *cur;
325 struct list_head *head, *pos;
326 unsigned int ndevs = 0;
329 /* Extract some power-on defaults from the target */
330 pci_read_config_dword(td, bridge->capndx+AGPSTAT, &tstatus);
331 isoch = (tstatus >> 17) & 0x1;
332 if (isoch == 0) /* isoch xfers not available, bail out. */
335 arqsz = (tstatus >> 13) & 0x7;
338 * Allocate a head for our AGP 3.5 device list
339 * (multiple AGP v3 devices are allowed behind a single bridge).
341 if ((dev_list = kmalloc(sizeof(*dev_list), GFP_KERNEL)) == NULL) {
345 head = &dev_list->list;
346 INIT_LIST_HEAD(head);
348 /* Find all AGP devices, and add them to dev_list. */
349 for_each_pci_dev(dev) {
350 mcapndx = pci_find_capability(dev, PCI_CAP_ID_AGP);
354 switch ((dev->class >>8) & 0xff00) {
355 case 0x0600: /* Bridge */
356 /* Skip bridges. We should call this function for each one. */
359 case 0x0001: /* Unclassified device */
360 /* Don't know what this is, but log it for investigation. */
362 printk (KERN_INFO PFX "Wacky, found unclassified AGP device. %x:%x\n",
363 dev->vendor, dev->device);
367 case 0x0300: /* Display controller */
368 case 0x0400: /* Multimedia controller */
369 if((cur = kmalloc(sizeof(*cur), GFP_KERNEL)) == NULL) {
386 * Take an initial pass through the devices lying behind our host
387 * bridge. Make sure each one is actually an AGP 3.0 device, otherwise
388 * exit with an error message. Along the way store the AGP 3.0
389 * cap_ptr for each device
391 list_for_each(pos, head) {
392 cur = list_entry(pos, struct agp_3_5_dev, list);
395 pci_read_config_word(dev, PCI_STATUS, &mpstat);
396 if ((mpstat & PCI_STATUS_CAP_LIST) == 0)
399 pci_read_config_byte(dev, PCI_CAPABILITY_LIST, &mcapndx);
402 pci_read_config_dword(dev, mcapndx, &ncapid);
403 if ((ncapid & 0xff) != 2)
404 mcapndx = (ncapid >> 8) & 0xff;
406 while (((ncapid & 0xff) != 2) && (mcapndx != 0));
410 printk(KERN_ERR PFX "woah! Non-AGP device "
411 "found on the secondary bus of an AGP 3.5 bridge!\n");
416 mmajor = (ncapid >> AGP_MAJOR_VERSION_SHIFT) & 0xf;
418 printk(KERN_ERR PFX "woah! AGP 2.0 device "
419 "found on the secondary bus of an AGP 3.5 "
420 "bridge operating with AGP 3.0 electricals!\n");
425 cur->capndx = mcapndx;
427 pci_read_config_dword(dev, cur->capndx+AGPSTAT, &mstatus);
429 if (((mstatus >> 3) & 0x1) == 0) {
430 printk(KERN_ERR PFX "woah! AGP 3.x device "
431 "not operating in AGP 3.x mode found on the "
432 "secondary bus of an AGP 3.5 bridge operating "
433 "with AGP 3.0 electricals!\n");
440 * Call functions to divide target resources amongst the AGP 3.0
441 * masters. This process is dramatically different depending on
442 * whether isochronous transfers are supported.
445 ret = agp_3_5_isochronous_node_enable(bridge, dev_list, ndevs);
447 printk(KERN_INFO PFX "Something bad happened setting "
448 "up isochronous xfers. Falling back to "
449 "non-isochronous xfer mode.\n");
454 agp_3_5_nonisochronous_node_enable(bridge, dev_list, ndevs);
457 /* Be sure to free the dev_list */
458 for (pos=head->next; pos!=head; ) {
459 cur = list_entry(pos, struct agp_3_5_dev, list);