2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
6 * Copyright (c) 2000-2006 Silicon Graphics, Inc. All Rights Reserved.
9 #include <linux/config.h>
10 #include <linux/module.h>
11 #include <asm/sn/nodepda.h>
12 #include <asm/sn/addrs.h>
13 #include <asm/sn/arch.h>
14 #include <asm/sn/sn_cpuid.h>
15 #include <asm/sn/pda.h>
16 #include <asm/sn/shubio.h>
17 #include <asm/nodedata.h>
18 #include <asm/delay.h>
20 #include <linux/bootmem.h>
21 #include <linux/string.h>
22 #include <linux/sched.h>
24 #include <asm/sn/bte.h>
27 #define L1_CACHE_MASK (L1_CACHE_BYTES - 1)
30 /* two interfaces on two btes */
31 #define MAX_INTERFACES_TO_TRY 4
32 #define MAX_NODES_TO_TRY 2
34 static struct bteinfo_s *bte_if_on_node(nasid_t nasid, int interface)
36 nodepda_t *tmp_nodepda;
38 if (nasid_to_cnodeid(nasid) == -1)
39 return (struct bteinfo_s *)NULL;
41 tmp_nodepda = NODEPDA(nasid_to_cnodeid(nasid));
42 return &tmp_nodepda->bte_if[interface];
46 static inline void bte_start_transfer(struct bteinfo_s *bte, u64 len, u64 mode)
49 BTE_CTRL_STORE(bte, (IBLS_BUSY | ((len) | (mode) << 24)));
51 BTE_LNSTAT_STORE(bte, len);
52 BTE_CTRL_STORE(bte, mode);
56 /************************************************************************
57 * Block Transfer Engine copy related functions.
59 ***********************************************************************/
62 * bte_copy(src, dest, len, mode, notification)
64 * Use the block transfer engine to move kernel memory from src to dest
65 * using the assigned mode.
68 * src - physical address of the transfer source.
69 * dest - physical address of the transfer destination.
70 * len - number of bytes to transfer from source to dest.
71 * mode - hardware defined. See reference information
72 * for IBCT0/1 in the SHUB Programmers Reference
73 * notification - kernel virtual address of the notification cache
74 * line. If NULL, the default is used and
75 * the bte_copy is synchronous.
77 * NOTE: This function requires src, dest, and len to
78 * be cacheline aligned.
80 bte_result_t bte_copy(u64 src, u64 dest, u64 len, u64 mode, void *notification)
85 struct bteinfo_s *bte;
86 bte_result_t bte_status;
87 unsigned long irq_flags;
88 unsigned long itc_end = 0;
89 int nasid_to_try[MAX_NODES_TO_TRY];
90 int my_nasid = cpuid_to_nasid(raw_smp_processor_id());
91 int bte_if_index, nasid_index;
92 int bte_first, btes_per_node = BTES_PER_NODE;
94 BTE_PRINTK(("bte_copy(0x%lx, 0x%lx, 0x%lx, 0x%lx, 0x%p)\n",
95 src, dest, len, mode, notification));
101 BUG_ON((len & L1_CACHE_MASK) ||
102 (src & L1_CACHE_MASK) || (dest & L1_CACHE_MASK));
103 BUG_ON(!(len < ((BTE_LEN_MASK + 1) << L1_CACHE_SHIFT)));
106 * Start with interface corresponding to cpu number
108 bte_first = raw_smp_processor_id() % btes_per_node;
110 if (mode & BTE_USE_DEST) {
111 /* try remote then local */
112 nasid_to_try[0] = NASID_GET(dest);
113 if (mode & BTE_USE_ANY) {
114 nasid_to_try[1] = my_nasid;
116 nasid_to_try[1] = (int)NULL;
119 /* try local then remote */
120 nasid_to_try[0] = my_nasid;
121 if (mode & BTE_USE_ANY) {
122 nasid_to_try[1] = NASID_GET(dest);
124 nasid_to_try[1] = (int)NULL;
130 local_irq_save(irq_flags);
132 bte_if_index = bte_first;
135 /* Attempt to lock one of the BTE interfaces. */
136 while (nasid_index < MAX_NODES_TO_TRY) {
137 bte = bte_if_on_node(nasid_to_try[nasid_index],bte_if_index);
144 if (spin_trylock(&bte->spinlock)) {
145 if (!(*bte->most_rcnt_na & BTE_WORD_AVAILABLE) ||
146 (BTE_LNSTAT_LOAD(bte) & BTE_ACTIVE)) {
147 /* Got the lock but BTE still busy */
148 spin_unlock(&bte->spinlock);
150 /* we got the lock and it's not busy */
155 bte_if_index = (bte_if_index + 1) % btes_per_node; /* Next interface */
156 if (bte_if_index == bte_first) {
158 * We've tried all interfaces on this node
170 local_irq_restore(irq_flags);
172 if (!(mode & BTE_WACQUIRE)) {
173 return BTEFAIL_NOTAVAIL;
177 if (notification == NULL) {
178 /* User does not want to be notified. */
179 bte->most_rcnt_na = &bte->notify;
181 bte->most_rcnt_na = notification;
184 /* Calculate the number of cache lines to transfer. */
185 transfer_size = ((len >> L1_CACHE_SHIFT) & BTE_LEN_MASK);
187 /* Initialize the notification to a known value. */
188 *bte->most_rcnt_na = BTE_WORD_BUSY;
189 notif_phys_addr = (u64)bte->most_rcnt_na;
191 /* Set the source and destination registers */
192 BTE_PRINTKV(("IBSA = 0x%lx)\n", src));
193 BTE_SRC_STORE(bte, src);
194 BTE_PRINTKV(("IBDA = 0x%lx)\n", dest));
195 BTE_DEST_STORE(bte, dest);
197 /* Set the notification register */
198 BTE_PRINTKV(("IBNA = 0x%lx)\n", notif_phys_addr));
199 BTE_NOTIF_STORE(bte, notif_phys_addr);
201 /* Initiate the transfer */
202 BTE_PRINTK(("IBCT = 0x%lx)\n", BTE_VALID_MODE(mode)));
203 bte_start_transfer(bte, transfer_size, BTE_VALID_MODE(mode));
205 itc_end = ia64_get_itc() + (40000000 * local_cpu_data->cyc_per_usec);
207 spin_unlock_irqrestore(&bte->spinlock, irq_flags);
209 if (notification != NULL) {
213 while ((transfer_stat = *bte->most_rcnt_na) == BTE_WORD_BUSY) {
215 if (ia64_get_itc() > itc_end) {
216 BTE_PRINTK(("BTE timeout nasid 0x%x bte%d IBLS = 0x%lx na 0x%lx\n",
217 NASID_GET(bte->bte_base_addr), bte->bte_num,
218 BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na) );
219 bte->bte_error_count++;
220 bte->bh_error = IBLS_ERROR;
221 bte_error_handler((unsigned long)NODEPDA(bte->bte_cnode));
222 *bte->most_rcnt_na = BTE_WORD_AVAILABLE;
227 BTE_PRINTKV((" Delay Done. IBLS = 0x%lx, most_rcnt_na = 0x%lx\n",
228 BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na));
230 if (transfer_stat & IBLS_ERROR) {
231 bte_status = transfer_stat & ~IBLS_ERROR;
233 bte_status = BTE_SUCCESS;
235 *bte->most_rcnt_na = BTE_WORD_AVAILABLE;
237 BTE_PRINTK(("Returning status is 0x%lx and most_rcnt_na is 0x%lx\n",
238 BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na));
243 EXPORT_SYMBOL(bte_copy);
246 * bte_unaligned_copy(src, dest, len, mode)
248 * use the block transfer engine to move kernel
249 * memory from src to dest using the assigned mode.
252 * src - physical address of the transfer source.
253 * dest - physical address of the transfer destination.
254 * len - number of bytes to transfer from source to dest.
255 * mode - hardware defined. See reference information
256 * for IBCT0/1 in the SGI documentation.
258 * NOTE: If the source, dest, and len are all cache line aligned,
259 * then it would be _FAR_ preferrable to use bte_copy instead.
261 bte_result_t bte_unaligned_copy(u64 src, u64 dest, u64 len, u64 mode)
263 int destFirstCacheOffset;
266 u64 headBcopySrcOffset;
274 char *bteBlock, *bteBlock_unaligned;
280 /* temporary buffer used during unaligned transfers */
281 bteBlock_unaligned = kmalloc(len + 3 * L1_CACHE_BYTES,
282 GFP_KERNEL | GFP_DMA);
283 if (bteBlock_unaligned == NULL) {
284 return BTEFAIL_NOTAVAIL;
286 bteBlock = (char *)L1_CACHE_ALIGN((u64) bteBlock_unaligned);
288 headBcopySrcOffset = src & L1_CACHE_MASK;
289 destFirstCacheOffset = dest & L1_CACHE_MASK;
292 * At this point, the transfer is broken into
293 * (up to) three sections. The first section is
294 * from the start address to the first physical
295 * cache line, the second is from the first physical
296 * cache line to the last complete cache line,
297 * and the third is from the last cache line to the
298 * end of the buffer. The first and third sections
299 * are handled by bte copying into a temporary buffer
300 * and then bcopy'ing the necessary section into the
301 * final location. The middle section is handled with
302 * a standard bte copy.
304 * One nasty exception to the above rule is when the
305 * source and destination are not symetrically
306 * mis-aligned. If the source offset from the first
307 * cache line is different from the destination offset,
308 * we make the first section be the entire transfer
309 * and the bcopy the entire block into place.
311 if (headBcopySrcOffset == destFirstCacheOffset) {
314 * Both the source and destination are the same
315 * distance from a cache line boundary so we can
316 * use the bte to transfer the bulk of the
319 headBteSource = src & ~L1_CACHE_MASK;
320 headBcopyDest = dest;
321 if (headBcopySrcOffset) {
325 headBcopySrcOffset) ? L1_CACHE_BYTES
326 - headBcopySrcOffset : len);
327 headBteLen = L1_CACHE_BYTES;
333 if (len > headBcopyLen) {
334 footBcopyLen = (len - headBcopyLen) & L1_CACHE_MASK;
335 footBteLen = L1_CACHE_BYTES;
337 footBteSource = src + len - footBcopyLen;
338 footBcopyDest = dest + len - footBcopyLen;
340 if (footBcopyDest == (headBcopyDest + headBcopyLen)) {
342 * We have two contigous bcopy
343 * blocks. Merge them.
345 headBcopyLen += footBcopyLen;
346 headBteLen += footBteLen;
347 } else if (footBcopyLen > 0) {
348 rv = bte_copy(footBteSource,
349 ia64_tpa((unsigned long)bteBlock),
350 footBteLen, mode, NULL);
351 if (rv != BTE_SUCCESS) {
352 kfree(bteBlock_unaligned);
356 memcpy(__va(footBcopyDest),
357 (char *)bteBlock, footBcopyLen);
364 if (len > (headBcopyLen + footBcopyLen)) {
365 /* now transfer the middle. */
366 rv = bte_copy((src + headBcopyLen),
369 (len - headBcopyLen -
370 footBcopyLen), mode, NULL);
371 if (rv != BTE_SUCCESS) {
372 kfree(bteBlock_unaligned);
380 * The transfer is not symetric, we will
381 * allocate a buffer large enough for all the
382 * data, bte_copy into that buffer and then
383 * bcopy to the destination.
386 /* Add the leader from source */
387 headBteLen = len + (src & L1_CACHE_MASK);
388 /* Add the trailing bytes from footer. */
389 headBteLen += L1_CACHE_BYTES - (headBteLen & L1_CACHE_MASK);
390 headBteSource = src & ~L1_CACHE_MASK;
391 headBcopySrcOffset = src & L1_CACHE_MASK;
392 headBcopyDest = dest;
396 if (headBcopyLen > 0) {
397 rv = bte_copy(headBteSource,
398 ia64_tpa((unsigned long)bteBlock), headBteLen,
400 if (rv != BTE_SUCCESS) {
401 kfree(bteBlock_unaligned);
405 memcpy(__va(headBcopyDest), ((char *)bteBlock +
406 headBcopySrcOffset), headBcopyLen);
408 kfree(bteBlock_unaligned);
412 EXPORT_SYMBOL(bte_unaligned_copy);
414 /************************************************************************
415 * Block Transfer Engine initialization functions.
417 ***********************************************************************/
420 * bte_init_node(nodepda, cnode)
422 * Initialize the nodepda structure with BTE base addresses and
425 void bte_init_node(nodepda_t * mynodepda, cnodeid_t cnode)
430 * Indicate that all the block transfer engines on this node
435 * Allocate one bte_recover_t structure per node. It holds
436 * the recovery lock for node. All the bte interface structures
437 * will point at this one bte_recover structure to get the lock.
439 spin_lock_init(&mynodepda->bte_recovery_lock);
440 init_timer(&mynodepda->bte_recovery_timer);
441 mynodepda->bte_recovery_timer.function = bte_error_handler;
442 mynodepda->bte_recovery_timer.data = (unsigned long)mynodepda;
444 for (i = 0; i < BTES_PER_NODE; i++) {
447 /* Which link status register should we use? */
449 REMOTE_HUB_ADDR(cnodeid_to_nasid(cnode), BTE_BASE_ADDR(i));
450 mynodepda->bte_if[i].bte_base_addr = base_addr;
451 mynodepda->bte_if[i].bte_source_addr = BTE_SOURCE_ADDR(base_addr);
452 mynodepda->bte_if[i].bte_destination_addr = BTE_DEST_ADDR(base_addr);
453 mynodepda->bte_if[i].bte_control_addr = BTE_CTRL_ADDR(base_addr);
454 mynodepda->bte_if[i].bte_notify_addr = BTE_NOTIF_ADDR(base_addr);
457 * Initialize the notification and spinlock
458 * so the first transfer can occur.
460 mynodepda->bte_if[i].most_rcnt_na =
461 &(mynodepda->bte_if[i].notify);
462 mynodepda->bte_if[i].notify = BTE_WORD_AVAILABLE;
463 spin_lock_init(&mynodepda->bte_if[i].spinlock);
465 mynodepda->bte_if[i].bte_cnode = cnode;
466 mynodepda->bte_if[i].bte_error_count = 0;
467 mynodepda->bte_if[i].bte_num = i;
468 mynodepda->bte_if[i].cleanup_active = 0;
469 mynodepda->bte_if[i].bh_error = 0;