703 lines · c
1/*2 * Copyright (c) 2004 Topspin Communications. All rights reserved.3 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.4 *5 * This software is available to you under a choice of one of two6 * licenses. You may choose to be licensed under the terms of the GNU7 * General Public License (GPL) Version 2, available from the file8 * COPYING in the main directory of this source tree, or the9 * OpenIB.org BSD license below:10 *11 * Redistribution and use in source and binary forms, with or12 * without modification, are permitted provided that the following13 * conditions are met:14 *15 * - Redistributions of source code must retain the above16 * copyright notice, this list of conditions and the following17 * disclaimer.18 *19 * - Redistributions in binary form must reproduce the above20 * copyright notice, this list of conditions and the following21 * disclaimer in the documentation and/or other materials22 * provided with the distribution.23 *24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE31 * SOFTWARE.32 */33 34#include <linux/slab.h>35#include <linux/errno.h>36 37#include "mthca_dev.h"38#include "mthca_cmd.h"39#include "mthca_memfree.h"40 41struct mthca_mtt {42 struct mthca_buddy *buddy;43 int order;44 u32 first_seg;45};46 47/*48 * Must be packed because mtt_seg is 64 bits but only aligned to 32 bits.49 */50struct mthca_mpt_entry {51 __be32 flags;52 __be32 page_size;53 __be32 key;54 __be32 pd;55 __be64 start;56 __be64 length;57 __be32 lkey;58 __be32 window_count;59 __be32 window_count_limit;60 __be64 mtt_seg;61 __be32 mtt_sz; /* Arbel only */62 u32 reserved[2];63} __packed;64 65#define MTHCA_MPT_FLAG_SW_OWNS (0xfUL << 28)66#define MTHCA_MPT_FLAG_MIO (1 << 17)67#define MTHCA_MPT_FLAG_BIND_ENABLE (1 << 15)68#define MTHCA_MPT_FLAG_PHYSICAL (1 << 9)69#define MTHCA_MPT_FLAG_REGION (1 << 8)70 71#define MTHCA_MTT_FLAG_PRESENT 172 73#define MTHCA_MPT_STATUS_SW 0xF074#define MTHCA_MPT_STATUS_HW 0x0075 76#define SINAI_FMR_KEY_INC 0x100000077 78/*79 * Buddy allocator for MTT segments (currently not very efficient80 * since it doesn't keep a free list and just searches linearly81 * through the bitmaps)82 */83 84static u32 mthca_buddy_alloc(struct mthca_buddy *buddy, int order)85{86 int o;87 int m;88 u32 seg;89 90 spin_lock(&buddy->lock);91 92 for (o = order; o <= buddy->max_order; ++o)93 if (buddy->num_free[o]) {94 m = 1 << (buddy->max_order - o);95 seg = find_first_bit(buddy->bits[o], m);96 if (seg < m)97 goto found;98 }99 100 spin_unlock(&buddy->lock);101 return -1;102 103 found:104 __clear_bit(seg, buddy->bits[o]);105 --buddy->num_free[o];106 107 while (o > order) {108 --o;109 seg <<= 1;110 __set_bit(seg ^ 1, buddy->bits[o]);111 ++buddy->num_free[o];112 }113 114 spin_unlock(&buddy->lock);115 116 seg <<= order;117 118 return seg;119}120 121static void mthca_buddy_free(struct mthca_buddy *buddy, u32 seg, int order)122{123 seg >>= order;124 125 spin_lock(&buddy->lock);126 127 while (test_bit(seg ^ 1, buddy->bits[order])) {128 __clear_bit(seg ^ 1, buddy->bits[order]);129 --buddy->num_free[order];130 seg >>= 1;131 ++order;132 }133 134 __set_bit(seg, buddy->bits[order]);135 ++buddy->num_free[order];136 137 spin_unlock(&buddy->lock);138}139 140static int mthca_buddy_init(struct mthca_buddy *buddy, int max_order)141{142 int i;143 144 buddy->max_order = max_order;145 spin_lock_init(&buddy->lock);146 147 buddy->bits = kcalloc(buddy->max_order + 1, sizeof(long *),148 GFP_KERNEL);149 buddy->num_free = kcalloc((buddy->max_order + 1), sizeof *buddy->num_free,150 GFP_KERNEL);151 if (!buddy->bits || !buddy->num_free)152 goto err_out;153 154 for (i = 0; i <= buddy->max_order; ++i) {155 buddy->bits[i] = bitmap_zalloc(1 << (buddy->max_order - i),156 GFP_KERNEL);157 if (!buddy->bits[i])158 goto err_out_free;159 }160 161 __set_bit(0, buddy->bits[buddy->max_order]);162 buddy->num_free[buddy->max_order] = 1;163 164 return 0;165 166err_out_free:167 for (i = 0; i <= buddy->max_order; ++i)168 bitmap_free(buddy->bits[i]);169 170err_out:171 kfree(buddy->bits);172 kfree(buddy->num_free);173 174 return -ENOMEM;175}176 177static void mthca_buddy_cleanup(struct mthca_buddy *buddy)178{179 int i;180 181 for (i = 0; i <= buddy->max_order; ++i)182 bitmap_free(buddy->bits[i]);183 184 kfree(buddy->bits);185 kfree(buddy->num_free);186}187 188static u32 mthca_alloc_mtt_range(struct mthca_dev *dev, int order,189 struct mthca_buddy *buddy)190{191 u32 seg = mthca_buddy_alloc(buddy, order);192 193 if (seg == -1)194 return -1;195 196 if (mthca_is_memfree(dev))197 if (mthca_table_get_range(dev, dev->mr_table.mtt_table, seg,198 seg + (1 << order) - 1)) {199 mthca_buddy_free(buddy, seg, order);200 seg = -1;201 }202 203 return seg;204}205 206static struct mthca_mtt *__mthca_alloc_mtt(struct mthca_dev *dev, int size,207 struct mthca_buddy *buddy)208{209 struct mthca_mtt *mtt;210 int i;211 212 if (size <= 0)213 return ERR_PTR(-EINVAL);214 215 mtt = kmalloc(sizeof *mtt, GFP_KERNEL);216 if (!mtt)217 return ERR_PTR(-ENOMEM);218 219 mtt->buddy = buddy;220 mtt->order = 0;221 for (i = dev->limits.mtt_seg_size / 8; i < size; i <<= 1)222 ++mtt->order;223 224 mtt->first_seg = mthca_alloc_mtt_range(dev, mtt->order, buddy);225 if (mtt->first_seg == -1) {226 kfree(mtt);227 return ERR_PTR(-ENOMEM);228 }229 230 return mtt;231}232 233struct mthca_mtt *mthca_alloc_mtt(struct mthca_dev *dev, int size)234{235 return __mthca_alloc_mtt(dev, size, &dev->mr_table.mtt_buddy);236}237 238void mthca_free_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt)239{240 if (!mtt)241 return;242 243 mthca_buddy_free(mtt->buddy, mtt->first_seg, mtt->order);244 245 mthca_table_put_range(dev, dev->mr_table.mtt_table,246 mtt->first_seg,247 mtt->first_seg + (1 << mtt->order) - 1);248 249 kfree(mtt);250}251 252static int __mthca_write_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt,253 int start_index, u64 *buffer_list, int list_len)254{255 struct mthca_mailbox *mailbox;256 __be64 *mtt_entry;257 int err = 0;258 int i;259 260 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);261 if (IS_ERR(mailbox))262 return PTR_ERR(mailbox);263 mtt_entry = mailbox->buf;264 265 while (list_len > 0) {266 mtt_entry[0] = cpu_to_be64(dev->mr_table.mtt_base +267 mtt->first_seg * dev->limits.mtt_seg_size +268 start_index * 8);269 mtt_entry[1] = 0;270 for (i = 0; i < list_len && i < MTHCA_MAILBOX_SIZE / 8 - 2; ++i)271 mtt_entry[i + 2] = cpu_to_be64(buffer_list[i] |272 MTHCA_MTT_FLAG_PRESENT);273 274 /*275 * If we have an odd number of entries to write, add276 * one more dummy entry for firmware efficiency.277 */278 if (i & 1)279 mtt_entry[i + 2] = 0;280 281 err = mthca_WRITE_MTT(dev, mailbox, (i + 1) & ~1);282 if (err) {283 mthca_warn(dev, "WRITE_MTT failed (%d)\n", err);284 goto out;285 }286 287 list_len -= i;288 start_index += i;289 buffer_list += i;290 }291 292out:293 mthca_free_mailbox(dev, mailbox);294 return err;295}296 297int mthca_write_mtt_size(struct mthca_dev *dev)298{299 if (dev->mr_table.fmr_mtt_buddy != &dev->mr_table.mtt_buddy ||300 !(dev->mthca_flags & MTHCA_FLAG_FMR))301 /*302 * Be friendly to WRITE_MTT command303 * and leave two empty slots for the304 * index and reserved fields of the305 * mailbox.306 */307 return PAGE_SIZE / sizeof (u64) - 2;308 309 /* For Arbel, all MTTs must fit in the same page. */310 return mthca_is_memfree(dev) ? (PAGE_SIZE / sizeof (u64)) : 0x7ffffff;311}312 313static void mthca_tavor_write_mtt_seg(struct mthca_dev *dev,314 struct mthca_mtt *mtt, int start_index,315 u64 *buffer_list, int list_len)316{317 u64 __iomem *mtts;318 int i;319 320 mtts = dev->mr_table.tavor_fmr.mtt_base + mtt->first_seg * dev->limits.mtt_seg_size +321 start_index * sizeof (u64);322 for (i = 0; i < list_len; ++i)323 mthca_write64_raw(cpu_to_be64(buffer_list[i] | MTHCA_MTT_FLAG_PRESENT),324 mtts + i);325}326 327static void mthca_arbel_write_mtt_seg(struct mthca_dev *dev,328 struct mthca_mtt *mtt, int start_index,329 u64 *buffer_list, int list_len)330{331 __be64 *mtts;332 dma_addr_t dma_handle;333 int i;334 int s = start_index * sizeof (u64);335 336 /* For Arbel, all MTTs must fit in the same page. */337 BUG_ON(s / PAGE_SIZE != (s + list_len * sizeof(u64) - 1) / PAGE_SIZE);338 /* Require full segments */339 BUG_ON(s % dev->limits.mtt_seg_size);340 341 mtts = mthca_table_find(dev->mr_table.mtt_table, mtt->first_seg +342 s / dev->limits.mtt_seg_size, &dma_handle);343 344 BUG_ON(!mtts);345 346 dma_sync_single_for_cpu(&dev->pdev->dev, dma_handle,347 list_len * sizeof (u64), DMA_TO_DEVICE);348 349 for (i = 0; i < list_len; ++i)350 mtts[i] = cpu_to_be64(buffer_list[i] | MTHCA_MTT_FLAG_PRESENT);351 352 dma_sync_single_for_device(&dev->pdev->dev, dma_handle,353 list_len * sizeof (u64), DMA_TO_DEVICE);354}355 356int mthca_write_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt,357 int start_index, u64 *buffer_list, int list_len)358{359 int size = mthca_write_mtt_size(dev);360 int chunk;361 362 if (dev->mr_table.fmr_mtt_buddy != &dev->mr_table.mtt_buddy ||363 !(dev->mthca_flags & MTHCA_FLAG_FMR))364 return __mthca_write_mtt(dev, mtt, start_index, buffer_list, list_len);365 366 while (list_len > 0) {367 chunk = min(size, list_len);368 if (mthca_is_memfree(dev))369 mthca_arbel_write_mtt_seg(dev, mtt, start_index,370 buffer_list, chunk);371 else372 mthca_tavor_write_mtt_seg(dev, mtt, start_index,373 buffer_list, chunk);374 375 list_len -= chunk;376 start_index += chunk;377 buffer_list += chunk;378 }379 380 return 0;381}382 383static inline u32 tavor_hw_index_to_key(u32 ind)384{385 return ind;386}387 388static inline u32 tavor_key_to_hw_index(u32 key)389{390 return key;391}392 393static inline u32 arbel_hw_index_to_key(u32 ind)394{395 return (ind >> 24) | (ind << 8);396}397 398static inline u32 arbel_key_to_hw_index(u32 key)399{400 return (key << 24) | (key >> 8);401}402 403static inline u32 hw_index_to_key(struct mthca_dev *dev, u32 ind)404{405 if (mthca_is_memfree(dev))406 return arbel_hw_index_to_key(ind);407 else408 return tavor_hw_index_to_key(ind);409}410 411static inline u32 key_to_hw_index(struct mthca_dev *dev, u32 key)412{413 if (mthca_is_memfree(dev))414 return arbel_key_to_hw_index(key);415 else416 return tavor_key_to_hw_index(key);417}418 419static inline u32 adjust_key(struct mthca_dev *dev, u32 key)420{421 if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT)422 return ((key << 20) & 0x800000) | (key & 0x7fffff);423 else424 return key;425}426 427int mthca_mr_alloc(struct mthca_dev *dev, u32 pd, int buffer_size_shift,428 u64 iova, u64 total_size, u32 access, struct mthca_mr *mr)429{430 struct mthca_mailbox *mailbox;431 struct mthca_mpt_entry *mpt_entry;432 u32 key;433 int i;434 int err;435 436 WARN_ON(buffer_size_shift >= 32);437 438 key = mthca_alloc(&dev->mr_table.mpt_alloc);439 if (key == -1)440 return -ENOMEM;441 key = adjust_key(dev, key);442 mr->ibmr.rkey = mr->ibmr.lkey = hw_index_to_key(dev, key);443 444 if (mthca_is_memfree(dev)) {445 err = mthca_table_get(dev, dev->mr_table.mpt_table, key);446 if (err)447 goto err_out_mpt_free;448 }449 450 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);451 if (IS_ERR(mailbox)) {452 err = PTR_ERR(mailbox);453 goto err_out_table;454 }455 mpt_entry = mailbox->buf;456 457 mpt_entry->flags = cpu_to_be32(MTHCA_MPT_FLAG_SW_OWNS |458 MTHCA_MPT_FLAG_MIO |459 MTHCA_MPT_FLAG_REGION |460 access);461 if (!mr->mtt)462 mpt_entry->flags |= cpu_to_be32(MTHCA_MPT_FLAG_PHYSICAL);463 464 mpt_entry->page_size = cpu_to_be32(buffer_size_shift - 12);465 mpt_entry->key = cpu_to_be32(key);466 mpt_entry->pd = cpu_to_be32(pd);467 mpt_entry->start = cpu_to_be64(iova);468 mpt_entry->length = cpu_to_be64(total_size);469 470 memset_startat(mpt_entry, 0, lkey);471 472 if (mr->mtt)473 mpt_entry->mtt_seg =474 cpu_to_be64(dev->mr_table.mtt_base +475 mr->mtt->first_seg * dev->limits.mtt_seg_size);476 477 if (0) {478 mthca_dbg(dev, "Dumping MPT entry %08x:\n", mr->ibmr.lkey);479 for (i = 0; i < sizeof (struct mthca_mpt_entry) / 4; ++i) {480 if (i % 4 == 0)481 printk("[%02x] ", i * 4);482 printk(" %08x", be32_to_cpu(((__be32 *) mpt_entry)[i]));483 if ((i + 1) % 4 == 0)484 printk("\n");485 }486 }487 488 err = mthca_SW2HW_MPT(dev, mailbox,489 key & (dev->limits.num_mpts - 1));490 if (err) {491 mthca_warn(dev, "SW2HW_MPT failed (%d)\n", err);492 goto err_out_mailbox;493 }494 495 mthca_free_mailbox(dev, mailbox);496 return err;497 498err_out_mailbox:499 mthca_free_mailbox(dev, mailbox);500 501err_out_table:502 mthca_table_put(dev, dev->mr_table.mpt_table, key);503 504err_out_mpt_free:505 mthca_free(&dev->mr_table.mpt_alloc, key);506 return err;507}508 509int mthca_mr_alloc_notrans(struct mthca_dev *dev, u32 pd,510 u32 access, struct mthca_mr *mr)511{512 mr->mtt = NULL;513 return mthca_mr_alloc(dev, pd, 12, 0, ~0ULL, access, mr);514}515 516int mthca_mr_alloc_phys(struct mthca_dev *dev, u32 pd,517 u64 *buffer_list, int buffer_size_shift,518 int list_len, u64 iova, u64 total_size,519 u32 access, struct mthca_mr *mr)520{521 int err;522 523 mr->mtt = mthca_alloc_mtt(dev, list_len);524 if (IS_ERR(mr->mtt))525 return PTR_ERR(mr->mtt);526 527 err = mthca_write_mtt(dev, mr->mtt, 0, buffer_list, list_len);528 if (err) {529 mthca_free_mtt(dev, mr->mtt);530 return err;531 }532 533 err = mthca_mr_alloc(dev, pd, buffer_size_shift, iova,534 total_size, access, mr);535 if (err)536 mthca_free_mtt(dev, mr->mtt);537 538 return err;539}540 541/* Free mr */542static void mthca_free_region(struct mthca_dev *dev, u32 lkey)543{544 mthca_table_put(dev, dev->mr_table.mpt_table,545 key_to_hw_index(dev, lkey));546 547 mthca_free(&dev->mr_table.mpt_alloc, key_to_hw_index(dev, lkey));548}549 550void mthca_free_mr(struct mthca_dev *dev, struct mthca_mr *mr)551{552 int err;553 554 err = mthca_HW2SW_MPT(dev, NULL,555 key_to_hw_index(dev, mr->ibmr.lkey) &556 (dev->limits.num_mpts - 1));557 if (err)558 mthca_warn(dev, "HW2SW_MPT failed (%d)\n", err);559 560 mthca_free_region(dev, mr->ibmr.lkey);561 mthca_free_mtt(dev, mr->mtt);562}563 564int mthca_init_mr_table(struct mthca_dev *dev)565{566 phys_addr_t addr;567 int mpts, mtts, err, i;568 569 err = mthca_alloc_init(&dev->mr_table.mpt_alloc,570 dev->limits.num_mpts,571 ~0, dev->limits.reserved_mrws);572 if (err)573 return err;574 575 if (!mthca_is_memfree(dev) &&576 (dev->mthca_flags & MTHCA_FLAG_DDR_HIDDEN))577 dev->limits.fmr_reserved_mtts = 0;578 else579 dev->mthca_flags |= MTHCA_FLAG_FMR;580 581 if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT)582 mthca_dbg(dev, "Memory key throughput optimization activated.\n");583 584 err = mthca_buddy_init(&dev->mr_table.mtt_buddy,585 fls(dev->limits.num_mtt_segs - 1));586 587 if (err)588 goto err_mtt_buddy;589 590 dev->mr_table.tavor_fmr.mpt_base = NULL;591 dev->mr_table.tavor_fmr.mtt_base = NULL;592 593 if (dev->limits.fmr_reserved_mtts) {594 i = fls(dev->limits.fmr_reserved_mtts - 1);595 596 if (i >= 31) {597 mthca_warn(dev, "Unable to reserve 2^31 FMR MTTs.\n");598 err = -EINVAL;599 goto err_fmr_mpt;600 }601 mpts = mtts = 1 << i;602 } else {603 mtts = dev->limits.num_mtt_segs;604 mpts = dev->limits.num_mpts;605 }606 607 if (!mthca_is_memfree(dev) &&608 (dev->mthca_flags & MTHCA_FLAG_FMR)) {609 610 addr = pci_resource_start(dev->pdev, 4) +611 ((pci_resource_len(dev->pdev, 4) - 1) &612 dev->mr_table.mpt_base);613 614 dev->mr_table.tavor_fmr.mpt_base =615 ioremap(addr, mpts * sizeof(struct mthca_mpt_entry));616 617 if (!dev->mr_table.tavor_fmr.mpt_base) {618 mthca_warn(dev, "MPT ioremap for FMR failed.\n");619 err = -ENOMEM;620 goto err_fmr_mpt;621 }622 623 addr = pci_resource_start(dev->pdev, 4) +624 ((pci_resource_len(dev->pdev, 4) - 1) &625 dev->mr_table.mtt_base);626 627 dev->mr_table.tavor_fmr.mtt_base =628 ioremap(addr, mtts * dev->limits.mtt_seg_size);629 if (!dev->mr_table.tavor_fmr.mtt_base) {630 mthca_warn(dev, "MTT ioremap for FMR failed.\n");631 err = -ENOMEM;632 goto err_fmr_mtt;633 }634 }635 636 if (dev->limits.fmr_reserved_mtts) {637 err = mthca_buddy_init(&dev->mr_table.tavor_fmr.mtt_buddy, fls(mtts - 1));638 if (err)639 goto err_fmr_mtt_buddy;640 641 /* Prevent regular MRs from using FMR keys */642 err = mthca_buddy_alloc(&dev->mr_table.mtt_buddy, fls(mtts - 1));643 if (err)644 goto err_reserve_fmr;645 646 dev->mr_table.fmr_mtt_buddy =647 &dev->mr_table.tavor_fmr.mtt_buddy;648 } else649 dev->mr_table.fmr_mtt_buddy = &dev->mr_table.mtt_buddy;650 651 /* FMR table is always the first, take reserved MTTs out of there */652 if (dev->limits.reserved_mtts) {653 i = fls(dev->limits.reserved_mtts - 1);654 655 if (mthca_alloc_mtt_range(dev, i,656 dev->mr_table.fmr_mtt_buddy) == -1) {657 mthca_warn(dev, "MTT table of order %d is too small.\n",658 dev->mr_table.fmr_mtt_buddy->max_order);659 err = -ENOMEM;660 goto err_reserve_mtts;661 }662 }663 664 return 0;665 666err_reserve_mtts:667err_reserve_fmr:668 if (dev->limits.fmr_reserved_mtts)669 mthca_buddy_cleanup(&dev->mr_table.tavor_fmr.mtt_buddy);670 671err_fmr_mtt_buddy:672 if (dev->mr_table.tavor_fmr.mtt_base)673 iounmap(dev->mr_table.tavor_fmr.mtt_base);674 675err_fmr_mtt:676 if (dev->mr_table.tavor_fmr.mpt_base)677 iounmap(dev->mr_table.tavor_fmr.mpt_base);678 679err_fmr_mpt:680 mthca_buddy_cleanup(&dev->mr_table.mtt_buddy);681 682err_mtt_buddy:683 mthca_alloc_cleanup(&dev->mr_table.mpt_alloc);684 685 return err;686}687 688void mthca_cleanup_mr_table(struct mthca_dev *dev)689{690 /* XXX check if any MRs are still allocated? */691 if (dev->limits.fmr_reserved_mtts)692 mthca_buddy_cleanup(&dev->mr_table.tavor_fmr.mtt_buddy);693 694 mthca_buddy_cleanup(&dev->mr_table.mtt_buddy);695 696 if (dev->mr_table.tavor_fmr.mtt_base)697 iounmap(dev->mr_table.tavor_fmr.mtt_base);698 if (dev->mr_table.tavor_fmr.mpt_base)699 iounmap(dev->mr_table.tavor_fmr.mpt_base);700 701 mthca_alloc_cleanup(&dev->mr_table.mpt_alloc);702}703