1/*
   2 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
   3 * Copyright (c) 2020, Intel Corporation. All rights reserved.
   4 *
   5 * This software is available to you under a choice of one of two
   6 * licenses.  You may choose to be licensed under the terms of the GNU
   7 * General Public License (GPL) Version 2, available from the file
   8 * COPYING in the main directory of this source tree, or the
   9 * OpenIB.org BSD license below:
  10 *
  11 *     Redistribution and use in source and binary forms, with or
  12 *     without modification, are permitted provided that the following
  13 *     conditions are met:
  14 *
  15 *      - Redistributions of source code must retain the above
  16 *        copyright notice, this list of conditions and the following
  17 *        disclaimer.
  18 *
  19 *      - Redistributions in binary form must reproduce the above
  20 *        copyright notice, this list of conditions and the following
  21 *        disclaimer in the documentation and/or other materials
  22 *        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 OF
  26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31 * SOFTWARE.
  32 */
  33
  34
  35#include <linux/kref.h>
  36#include <linux/random.h>
  37#include <linux/debugfs.h>
  38#include <linux/export.h>
  39#include <linux/delay.h>
  40#include <linux/dma-buf.h>
  41#include <linux/dma-resv.h>
  42#include <rdma/ib_umem_odp.h>
  43#include "dm.h"
  44#include "mlx5_ib.h"
  45#include "umr.h"
  46
  47enum {
  48	MAX_PENDING_REG_MR = 8,
  49};
  50
  51#define MLX5_UMR_ALIGN 2048
  52
  53static void
  54create_mkey_callback(int status, struct mlx5_async_work *context);
  55static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem,
  56				     u64 iova, int access_flags,
  57				     unsigned int page_size, bool populate);
  58
  59static void set_mkc_access_pd_addr_fields(void *mkc, int acc, u64 start_addr,
  60					  struct ib_pd *pd)
  61{
  62	struct mlx5_ib_dev *dev = to_mdev(pd->device);
  63
  64	MLX5_SET(mkc, mkc, a, !!(acc & IB_ACCESS_REMOTE_ATOMIC));
  65	MLX5_SET(mkc, mkc, rw, !!(acc & IB_ACCESS_REMOTE_WRITE));
  66	MLX5_SET(mkc, mkc, rr, !!(acc & IB_ACCESS_REMOTE_READ));
  67	MLX5_SET(mkc, mkc, lw, !!(acc & IB_ACCESS_LOCAL_WRITE));
  68	MLX5_SET(mkc, mkc, lr, 1);
  69
  70	if (acc & IB_ACCESS_RELAXED_ORDERING) {
  71		if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write))
  72			MLX5_SET(mkc, mkc, relaxed_ordering_write, 1);
  73
  74		if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read) ||
  75		    (MLX5_CAP_GEN(dev->mdev,
  76				  relaxed_ordering_read_pci_enabled) &&
  77		     pcie_relaxed_ordering_enabled(dev->mdev->pdev)))
  78			MLX5_SET(mkc, mkc, relaxed_ordering_read, 1);
  79	}
  80
  81	MLX5_SET(mkc, mkc, pd, to_mpd(pd)->pdn);
  82	MLX5_SET(mkc, mkc, qpn, 0xffffff);
  83	MLX5_SET64(mkc, mkc, start_addr, start_addr);
  84}
  85
  86static void assign_mkey_variant(struct mlx5_ib_dev *dev, u32 *mkey, u32 *in)
  87{
  88	u8 key = atomic_inc_return(&dev->mkey_var);
  89	void *mkc;
  90
  91	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
  92	MLX5_SET(mkc, mkc, mkey_7_0, key);
  93	*mkey = key;
  94}
  95
  96static int mlx5_ib_create_mkey(struct mlx5_ib_dev *dev,
  97			       struct mlx5_ib_mkey *mkey, u32 *in, int inlen)
  98{
  99	int ret;
 100
 101	assign_mkey_variant(dev, &mkey->key, in);
 102	ret = mlx5_core_create_mkey(dev->mdev, &mkey->key, in, inlen);
 103	if (!ret)
 104		init_waitqueue_head(&mkey->wait);
 105
 106	return ret;
 107}
 108
 109static int mlx5_ib_create_mkey_cb(struct mlx5r_async_create_mkey *async_create)
 110{
 111	struct mlx5_ib_dev *dev = async_create->ent->dev;
 112	size_t inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
 113	size_t outlen = MLX5_ST_SZ_BYTES(create_mkey_out);
 114
 115	MLX5_SET(create_mkey_in, async_create->in, opcode,
 116		 MLX5_CMD_OP_CREATE_MKEY);
 117	assign_mkey_variant(dev, &async_create->mkey, async_create->in);
 118	return mlx5_cmd_exec_cb(&dev->async_ctx, async_create->in, inlen,
 119				async_create->out, outlen, create_mkey_callback,
 120				&async_create->cb_work);
 121}
 122
 123static int mkey_cache_max_order(struct mlx5_ib_dev *dev);
 124static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent);
 125
 126static int destroy_mkey(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
 127{
 128	WARN_ON(xa_load(&dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key)));
 129
 130	return mlx5_core_destroy_mkey(dev->mdev, mr->mmkey.key);
 131}
 132
 133static void create_mkey_warn(struct mlx5_ib_dev *dev, int status, void *out)
 134{
 135	if (status == -ENXIO) /* core driver is not available */
 136		return;
 137
 138	mlx5_ib_warn(dev, "async reg mr failed. status %d\n", status);
 139	if (status != -EREMOTEIO) /* driver specific failure */
 140		return;
 141
 142	/* Failed in FW, print cmd out failure details */
 143	mlx5_cmd_out_err(dev->mdev, MLX5_CMD_OP_CREATE_MKEY, 0, out);
 144}
 145
 146static int push_mkey_locked(struct mlx5_cache_ent *ent, u32 mkey)
 147{
 148	unsigned long tmp = ent->mkeys_queue.ci % NUM_MKEYS_PER_PAGE;
 149	struct mlx5_mkeys_page *page;
 150
 151	lockdep_assert_held(&ent->mkeys_queue.lock);
 152	if (ent->mkeys_queue.ci >=
 153	    ent->mkeys_queue.num_pages * NUM_MKEYS_PER_PAGE) {
 154		page = kzalloc(sizeof(*page), GFP_ATOMIC);
 155		if (!page)
 156			return -ENOMEM;
 157		ent->mkeys_queue.num_pages++;
 158		list_add_tail(&page->list, &ent->mkeys_queue.pages_list);
 159	} else {
 160		page = list_last_entry(&ent->mkeys_queue.pages_list,
 161				       struct mlx5_mkeys_page, list);
 162	}
 163
 164	page->mkeys[tmp] = mkey;
 165	ent->mkeys_queue.ci++;
 166	return 0;
 167}
 168
 169static int pop_mkey_locked(struct mlx5_cache_ent *ent)
 170{
 171	unsigned long tmp = (ent->mkeys_queue.ci - 1) % NUM_MKEYS_PER_PAGE;
 172	struct mlx5_mkeys_page *last_page;
 173	u32 mkey;
 174
 175	lockdep_assert_held(&ent->mkeys_queue.lock);
 176	last_page = list_last_entry(&ent->mkeys_queue.pages_list,
 177				    struct mlx5_mkeys_page, list);
 178	mkey = last_page->mkeys[tmp];
 179	last_page->mkeys[tmp] = 0;
 180	ent->mkeys_queue.ci--;
 181	if (ent->mkeys_queue.num_pages > 1 && !tmp) {
 182		list_del(&last_page->list);
 183		ent->mkeys_queue.num_pages--;
 184		kfree(last_page);
 185	}
 186	return mkey;
 187}
 188
 189static void create_mkey_callback(int status, struct mlx5_async_work *context)
 190{
 191	struct mlx5r_async_create_mkey *mkey_out =
 192		container_of(context, struct mlx5r_async_create_mkey, cb_work);
 193	struct mlx5_cache_ent *ent = mkey_out->ent;
 194	struct mlx5_ib_dev *dev = ent->dev;
 195	unsigned long flags;
 196
 197	if (status) {
 198		create_mkey_warn(dev, status, mkey_out->out);
 199		kfree(mkey_out);
 200		spin_lock_irqsave(&ent->mkeys_queue.lock, flags);
 201		ent->pending--;
 202		WRITE_ONCE(dev->fill_delay, 1);
 203		spin_unlock_irqrestore(&ent->mkeys_queue.lock, flags);
 204		mod_timer(&dev->delay_timer, jiffies + HZ);
 205		return;
 206	}
 207
 208	mkey_out->mkey |= mlx5_idx_to_mkey(
 209		MLX5_GET(create_mkey_out, mkey_out->out, mkey_index));
 210	WRITE_ONCE(dev->cache.last_add, jiffies);
 211
 212	spin_lock_irqsave(&ent->mkeys_queue.lock, flags);
 213	push_mkey_locked(ent, mkey_out->mkey);
 214	/* If we are doing fill_to_high_water then keep going. */
 215	queue_adjust_cache_locked(ent);
 216	ent->pending--;
 217	spin_unlock_irqrestore(&ent->mkeys_queue.lock, flags);
 218	kfree(mkey_out);
 219}
 220
 221static int get_mkc_octo_size(unsigned int access_mode, unsigned int ndescs)
 222{
 223	int ret = 0;
 224
 225	switch (access_mode) {
 226	case MLX5_MKC_ACCESS_MODE_MTT:
 227		ret = DIV_ROUND_UP(ndescs, MLX5_IB_UMR_OCTOWORD /
 228						   sizeof(struct mlx5_mtt));
 229		break;
 230	case MLX5_MKC_ACCESS_MODE_KSM:
 231		ret = DIV_ROUND_UP(ndescs, MLX5_IB_UMR_OCTOWORD /
 232						   sizeof(struct mlx5_klm));
 233		break;
 234	default:
 235		WARN_ON(1);
 236	}
 237	return ret;
 238}
 239
 240static void set_cache_mkc(struct mlx5_cache_ent *ent, void *mkc)
 241{
 242	set_mkc_access_pd_addr_fields(mkc, ent->rb_key.access_flags, 0,
 243				      ent->dev->umrc.pd);
 244	MLX5_SET(mkc, mkc, free, 1);
 245	MLX5_SET(mkc, mkc, umr_en, 1);
 246	MLX5_SET(mkc, mkc, access_mode_1_0, ent->rb_key.access_mode & 0x3);
 247	MLX5_SET(mkc, mkc, access_mode_4_2,
 248		(ent->rb_key.access_mode >> 2) & 0x7);
 249
 250	MLX5_SET(mkc, mkc, translations_octword_size,
 251		 get_mkc_octo_size(ent->rb_key.access_mode,
 252				   ent->rb_key.ndescs));
 253	MLX5_SET(mkc, mkc, log_page_size, PAGE_SHIFT);
 254}
 255
 256/* Asynchronously schedule new MRs to be populated in the cache. */
 257static int add_keys(struct mlx5_cache_ent *ent, unsigned int num)
 258{
 259	struct mlx5r_async_create_mkey *async_create;
 260	void *mkc;
 261	int err = 0;
 262	int i;
 263
 264	for (i = 0; i < num; i++) {
 265		async_create = kzalloc(sizeof(struct mlx5r_async_create_mkey),
 266				       GFP_KERNEL);
 267		if (!async_create)
 268			return -ENOMEM;
 269		mkc = MLX5_ADDR_OF(create_mkey_in, async_create->in,
 270				   memory_key_mkey_entry);
 271		set_cache_mkc(ent, mkc);
 272		async_create->ent = ent;
 273
 274		spin_lock_irq(&ent->mkeys_queue.lock);
 275		if (ent->pending >= MAX_PENDING_REG_MR) {
 276			err = -EAGAIN;
 277			goto free_async_create;
 278		}
 279		ent->pending++;
 280		spin_unlock_irq(&ent->mkeys_queue.lock);
 281
 282		err = mlx5_ib_create_mkey_cb(async_create);
 283		if (err) {
 284			mlx5_ib_warn(ent->dev, "create mkey failed %d\n", err);
 285			goto err_create_mkey;
 286		}
 287	}
 288
 289	return 0;
 290
 291err_create_mkey:
 292	spin_lock_irq(&ent->mkeys_queue.lock);
 293	ent->pending--;
 294free_async_create:
 295	spin_unlock_irq(&ent->mkeys_queue.lock);
 296	kfree(async_create);
 297	return err;
 298}
 299
 300/* Synchronously create a MR in the cache */
 301static int create_cache_mkey(struct mlx5_cache_ent *ent, u32 *mkey)
 302{
 303	size_t inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
 304	void *mkc;
 305	u32 *in;
 306	int err;
 307
 308	in = kzalloc(inlen, GFP_KERNEL);
 309	if (!in)
 310		return -ENOMEM;
 311	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
 312	set_cache_mkc(ent, mkc);
 313
 314	err = mlx5_core_create_mkey(ent->dev->mdev, mkey, in, inlen);
 315	if (err)
 316		goto free_in;
 317
 318	WRITE_ONCE(ent->dev->cache.last_add, jiffies);
 319free_in:
 320	kfree(in);
 321	return err;
 322}
 323
 324static void remove_cache_mr_locked(struct mlx5_cache_ent *ent)
 325{
 326	u32 mkey;
 327
 328	lockdep_assert_held(&ent->mkeys_queue.lock);
 329	if (!ent->mkeys_queue.ci)
 330		return;
 331	mkey = pop_mkey_locked(ent);
 332	spin_unlock_irq(&ent->mkeys_queue.lock);
 333	mlx5_core_destroy_mkey(ent->dev->mdev, mkey);
 334	spin_lock_irq(&ent->mkeys_queue.lock);
 335}
 336
 337static int resize_available_mrs(struct mlx5_cache_ent *ent, unsigned int target,
 338				bool limit_fill)
 339	__acquires(&ent->mkeys_queue.lock) __releases(&ent->mkeys_queue.lock)
 340{
 341	int err;
 342
 343	lockdep_assert_held(&ent->mkeys_queue.lock);
 344
 345	while (true) {
 346		if (limit_fill)
 347			target = ent->limit * 2;
 348		if (target == ent->pending + ent->mkeys_queue.ci)
 349			return 0;
 350		if (target > ent->pending + ent->mkeys_queue.ci) {
 351			u32 todo = target - (ent->pending + ent->mkeys_queue.ci);
 352
 353			spin_unlock_irq(&ent->mkeys_queue.lock);
 354			err = add_keys(ent, todo);
 355			if (err == -EAGAIN)
 356				usleep_range(3000, 5000);
 357			spin_lock_irq(&ent->mkeys_queue.lock);
 358			if (err) {
 359				if (err != -EAGAIN)
 360					return err;
 361			} else
 362				return 0;
 363		} else {
 364			remove_cache_mr_locked(ent);
 365		}
 366	}
 367}
 368
 369static ssize_t size_write(struct file *filp, const char __user *buf,
 370			  size_t count, loff_t *pos)
 371{
 372	struct mlx5_cache_ent *ent = filp->private_data;
 373	u32 target;
 374	int err;
 375
 376	err = kstrtou32_from_user(buf, count, 0, &target);
 377	if (err)
 378		return err;
 379
 380	/*
 381	 * Target is the new value of total_mrs the user requests, however we
 382	 * cannot free MRs that are in use. Compute the target value for stored
 383	 * mkeys.
 384	 */
 385	spin_lock_irq(&ent->mkeys_queue.lock);
 386	if (target < ent->in_use) {
 387		err = -EINVAL;
 388		goto err_unlock;
 389	}
 390	target = target - ent->in_use;
 391	if (target < ent->limit || target > ent->limit*2) {
 392		err = -EINVAL;
 393		goto err_unlock;
 394	}
 395	err = resize_available_mrs(ent, target, false);
 396	if (err)
 397		goto err_unlock;
 398	spin_unlock_irq(&ent->mkeys_queue.lock);
 399
 400	return count;
 401
 402err_unlock:
 403	spin_unlock_irq(&ent->mkeys_queue.lock);
 404	return err;
 405}
 406
 407static ssize_t size_read(struct file *filp, char __user *buf, size_t count,
 408			 loff_t *pos)
 409{
 410	struct mlx5_cache_ent *ent = filp->private_data;
 411	char lbuf[20];
 412	int err;
 413
 414	err = snprintf(lbuf, sizeof(lbuf), "%ld\n",
 415		       ent->mkeys_queue.ci + ent->in_use);
 416	if (err < 0)
 417		return err;
 418
 419	return simple_read_from_buffer(buf, count, pos, lbuf, err);
 420}
 421
 422static const struct file_operations size_fops = {
 423	.owner	= THIS_MODULE,
 424	.open	= simple_open,
 425	.write	= size_write,
 426	.read	= size_read,
 427};
 428
 429static ssize_t limit_write(struct file *filp, const char __user *buf,
 430			   size_t count, loff_t *pos)
 431{
 432	struct mlx5_cache_ent *ent = filp->private_data;
 433	u32 var;
 434	int err;
 435
 436	err = kstrtou32_from_user(buf, count, 0, &var);
 437	if (err)
 438		return err;
 439
 440	/*
 441	 * Upon set we immediately fill the cache to high water mark implied by
 442	 * the limit.
 443	 */
 444	spin_lock_irq(&ent->mkeys_queue.lock);
 445	ent->limit = var;
 446	err = resize_available_mrs(ent, 0, true);
 447	spin_unlock_irq(&ent->mkeys_queue.lock);
 448	if (err)
 449		return err;
 450	return count;
 451}
 452
 453static ssize_t limit_read(struct file *filp, char __user *buf, size_t count,
 454			  loff_t *pos)
 455{
 456	struct mlx5_cache_ent *ent = filp->private_data;
 457	char lbuf[20];
 458	int err;
 459
 460	err = snprintf(lbuf, sizeof(lbuf), "%d\n", ent->limit);
 461	if (err < 0)
 462		return err;
 463
 464	return simple_read_from_buffer(buf, count, pos, lbuf, err);
 465}
 466
 467static const struct file_operations limit_fops = {
 468	.owner	= THIS_MODULE,
 469	.open	= simple_open,
 470	.write	= limit_write,
 471	.read	= limit_read,
 472};
 473
 474static bool someone_adding(struct mlx5_mkey_cache *cache)
 475{
 476	struct mlx5_cache_ent *ent;
 477	struct rb_node *node;
 478	bool ret;
 479
 480	mutex_lock(&cache->rb_lock);
 481	for (node = rb_first(&cache->rb_root); node; node = rb_next(node)) {
 482		ent = rb_entry(node, struct mlx5_cache_ent, node);
 483		spin_lock_irq(&ent->mkeys_queue.lock);
 484		ret = ent->mkeys_queue.ci < ent->limit;
 485		spin_unlock_irq(&ent->mkeys_queue.lock);
 486		if (ret) {
 487			mutex_unlock(&cache->rb_lock);
 488			return true;
 489		}
 490	}
 491	mutex_unlock(&cache->rb_lock);
 492	return false;
 493}
 494
 495/*
 496 * Check if the bucket is outside the high/low water mark and schedule an async
 497 * update. The cache refill has hysteresis, once the low water mark is hit it is
 498 * refilled up to the high mark.
 499 */
 500static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent)
 501{
 502	lockdep_assert_held(&ent->mkeys_queue.lock);
 503
 504	if (ent->disabled || READ_ONCE(ent->dev->fill_delay) || ent->is_tmp)
 505		return;
 506	if (ent->mkeys_queue.ci < ent->limit) {
 507		ent->fill_to_high_water = true;
 508		mod_delayed_work(ent->dev->cache.wq, &ent->dwork, 0);
 509	} else if (ent->fill_to_high_water &&
 510		   ent->mkeys_queue.ci + ent->pending < 2 * ent->limit) {
 511		/*
 512		 * Once we start populating due to hitting a low water mark
 513		 * continue until we pass the high water mark.
 514		 */
 515		mod_delayed_work(ent->dev->cache.wq, &ent->dwork, 0);
 516	} else if (ent->mkeys_queue.ci == 2 * ent->limit) {
 517		ent->fill_to_high_water = false;
 518	} else if (ent->mkeys_queue.ci > 2 * ent->limit) {
 519		/* Queue deletion of excess entries */
 520		ent->fill_to_high_water = false;
 521		if (ent->pending)
 522			queue_delayed_work(ent->dev->cache.wq, &ent->dwork,
 523					   msecs_to_jiffies(1000));
 524		else
 525			mod_delayed_work(ent->dev->cache.wq, &ent->dwork, 0);
 526	}
 527}
 528
 529static void __cache_work_func(struct mlx5_cache_ent *ent)
 530{
 531	struct mlx5_ib_dev *dev = ent->dev;
 532	struct mlx5_mkey_cache *cache = &dev->cache;
 533	int err;
 534
 535	spin_lock_irq(&ent->mkeys_queue.lock);
 536	if (ent->disabled)
 537		goto out;
 538
 539	if (ent->fill_to_high_water &&
 540	    ent->mkeys_queue.ci + ent->pending < 2 * ent->limit &&
 541	    !READ_ONCE(dev->fill_delay)) {
 542		spin_unlock_irq(&ent->mkeys_queue.lock);
 543		err = add_keys(ent, 1);
 544		spin_lock_irq(&ent->mkeys_queue.lock);
 545		if (ent->disabled)
 546			goto out;
 547		if (err) {
 548			/*
 549			 * EAGAIN only happens if there are pending MRs, so we
 550			 * will be rescheduled when storing them. The only
 551			 * failure path here is ENOMEM.
 552			 */
 553			if (err != -EAGAIN) {
 554				mlx5_ib_warn(
 555					dev,
 556					"add keys command failed, err %d\n",
 557					err);
 558				queue_delayed_work(cache->wq, &ent->dwork,
 559						   msecs_to_jiffies(1000));
 560			}
 561		}
 562	} else if (ent->mkeys_queue.ci > 2 * ent->limit) {
 563		bool need_delay;
 564
 565		/*
 566		 * The remove_cache_mr() logic is performed as garbage
 567		 * collection task. Such task is intended to be run when no
 568		 * other active processes are running.
 569		 *
 570		 * The need_resched() will return TRUE if there are user tasks
 571		 * to be activated in near future.
 572		 *
 573		 * In such case, we don't execute remove_cache_mr() and postpone
 574		 * the garbage collection work to try to run in next cycle, in
 575		 * order to free CPU resources to other tasks.
 576		 */
 577		spin_unlock_irq(&ent->mkeys_queue.lock);
 578		need_delay = need_resched() || someone_adding(cache) ||
 579			     !time_after(jiffies,
 580					 READ_ONCE(cache->last_add) + 300 * HZ);
 581		spin_lock_irq(&ent->mkeys_queue.lock);
 582		if (ent->disabled)
 583			goto out;
 584		if (need_delay) {
 585			queue_delayed_work(cache->wq, &ent->dwork, 300 * HZ);
 586			goto out;
 587		}
 588		remove_cache_mr_locked(ent);
 589		queue_adjust_cache_locked(ent);
 590	}
 591out:
 592	spin_unlock_irq(&ent->mkeys_queue.lock);
 593}
 594
 595static void delayed_cache_work_func(struct work_struct *work)
 596{
 597	struct mlx5_cache_ent *ent;
 598
 599	ent = container_of(work, struct mlx5_cache_ent, dwork.work);
 600	__cache_work_func(ent);
 601}
 602
 603static int cache_ent_key_cmp(struct mlx5r_cache_rb_key key1,
 604			     struct mlx5r_cache_rb_key key2)
 605{
 606	int res;
 607
 608	res = key1.ats - key2.ats;
 609	if (res)
 610		return res;
 611
 612	res = key1.access_mode - key2.access_mode;
 613	if (res)
 614		return res;
 615
 616	res = key1.access_flags - key2.access_flags;
 617	if (res)
 618		return res;
 619
 620	/*
 621	 * keep ndescs the last in the compare table since the find function
 622	 * searches for an exact match on all properties and only closest
 623	 * match in size.
 624	 */
 625	return key1.ndescs - key2.ndescs;
 626}
 627
 628static int mlx5_cache_ent_insert(struct mlx5_mkey_cache *cache,
 629				 struct mlx5_cache_ent *ent)
 630{
 631	struct rb_node **new = &cache->rb_root.rb_node, *parent = NULL;
 632	struct mlx5_cache_ent *cur;
 633	int cmp;
 634
 635	/* Figure out where to put new node */
 636	while (*new) {
 637		cur = rb_entry(*new, struct mlx5_cache_ent, node);
 638		parent = *new;
 639		cmp = cache_ent_key_cmp(cur->rb_key, ent->rb_key);
 640		if (cmp > 0)
 641			new = &((*new)->rb_left);
 642		if (cmp < 0)
 643			new = &((*new)->rb_right);
 644		if (cmp == 0) {
 645			mutex_unlock(&cache->rb_lock);
 646			return -EEXIST;
 647		}
 648	}
 649
 650	/* Add new node and rebalance tree. */
 651	rb_link_node(&ent->node, parent, new);
 652	rb_insert_color(&ent->node, &cache->rb_root);
 653
 654	return 0;
 655}
 656
 657static struct mlx5_cache_ent *
 658mkey_cache_ent_from_rb_key(struct mlx5_ib_dev *dev,
 659			   struct mlx5r_cache_rb_key rb_key)
 660{
 661	struct rb_node *node = dev->cache.rb_root.rb_node;
 662	struct mlx5_cache_ent *cur, *smallest = NULL;
 663	int cmp;
 664
 665	/*
 666	 * Find the smallest ent with order >= requested_order.
 667	 */
 668	while (node) {
 669		cur = rb_entry(node, struct mlx5_cache_ent, node);
 670		cmp = cache_ent_key_cmp(cur->rb_key, rb_key);
 671		if (cmp > 0) {
 672			smallest = cur;
 673			node = node->rb_left;
 674		}
 675		if (cmp < 0)
 676			node = node->rb_right;
 677		if (cmp == 0)
 678			return cur;
 679	}
 680
 681	return (smallest &&
 682		smallest->rb_key.access_mode == rb_key.access_mode &&
 683		smallest->rb_key.access_flags == rb_key.access_flags &&
 684		smallest->rb_key.ats == rb_key.ats) ?
 685		       smallest :
 686		       NULL;
 687}
 688
 689static struct mlx5_ib_mr *_mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
 690					struct mlx5_cache_ent *ent,
 691					int access_flags)
 692{
 693	struct mlx5_ib_mr *mr;
 694	int err;
 695
 696	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 697	if (!mr)
 698		return ERR_PTR(-ENOMEM);
 699
 700	spin_lock_irq(&ent->mkeys_queue.lock);
 701	ent->in_use++;
 702
 703	if (!ent->mkeys_queue.ci) {
 704		queue_adjust_cache_locked(ent);
 705		ent->miss++;
 706		spin_unlock_irq(&ent->mkeys_queue.lock);
 707		err = create_cache_mkey(ent, &mr->mmkey.key);
 708		if (err) {
 709			spin_lock_irq(&ent->mkeys_queue.lock);
 710			ent->in_use--;
 711			spin_unlock_irq(&ent->mkeys_queue.lock);
 712			kfree(mr);
 713			return ERR_PTR(err);
 714		}
 715	} else {
 716		mr->mmkey.key = pop_mkey_locked(ent);
 717		queue_adjust_cache_locked(ent);
 718		spin_unlock_irq(&ent->mkeys_queue.lock);
 719	}
 720	mr->mmkey.cache_ent = ent;
 721	mr->mmkey.type = MLX5_MKEY_MR;
 722	init_waitqueue_head(&mr->mmkey.wait);
 723	return mr;
 724}
 725
 726static int get_unchangeable_access_flags(struct mlx5_ib_dev *dev,
 727					 int access_flags)
 728{
 729	int ret = 0;
 730
 731	if ((access_flags & IB_ACCESS_REMOTE_ATOMIC) &&
 732	    MLX5_CAP_GEN(dev->mdev, atomic) &&
 733	    MLX5_CAP_GEN(dev->mdev, umr_modify_atomic_disabled))
 734		ret |= IB_ACCESS_REMOTE_ATOMIC;
 735
 736	if ((access_flags & IB_ACCESS_RELAXED_ORDERING) &&
 737	    MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write) &&
 738	    !MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write_umr))
 739		ret |= IB_ACCESS_RELAXED_ORDERING;
 740
 741	if ((access_flags & IB_ACCESS_RELAXED_ORDERING) &&
 742	    (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read) ||
 743	     MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read_pci_enabled)) &&
 744	    !MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read_umr))
 745		ret |= IB_ACCESS_RELAXED_ORDERING;
 746
 747	return ret;
 748}
 749
 750struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
 751				       int access_flags, int access_mode,
 752				       int ndescs)
 753{
 754	struct mlx5r_cache_rb_key rb_key = {
 755		.ndescs = ndescs,
 756		.access_mode = access_mode,
 757		.access_flags = get_unchangeable_access_flags(dev, access_flags)
 758	};
 759	struct mlx5_cache_ent *ent = mkey_cache_ent_from_rb_key(dev, rb_key);
 760
 761	if (!ent)
 762		return ERR_PTR(-EOPNOTSUPP);
 763
 764	return _mlx5_mr_cache_alloc(dev, ent, access_flags);
 765}
 766
 767static void clean_keys(struct mlx5_ib_dev *dev, struct mlx5_cache_ent *ent)
 768{
 769	u32 mkey;
 770
 771	cancel_delayed_work(&ent->dwork);
 772	spin_lock_irq(&ent->mkeys_queue.lock);
 773	while (ent->mkeys_queue.ci) {
 774		mkey = pop_mkey_locked(ent);
 775		spin_unlock_irq(&ent->mkeys_queue.lock);
 776		mlx5_core_destroy_mkey(dev->mdev, mkey);
 777		spin_lock_irq(&ent->mkeys_queue.lock);
 778	}
 779	spin_unlock_irq(&ent->mkeys_queue.lock);
 780}
 781
 782static void mlx5_mkey_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
 783{
 784	if (!mlx5_debugfs_root || dev->is_rep)
 785		return;
 786
 787	debugfs_remove_recursive(dev->cache.fs_root);
 788	dev->cache.fs_root = NULL;
 789}
 790
 791static void mlx5_mkey_cache_debugfs_add_ent(struct mlx5_ib_dev *dev,
 792					    struct mlx5_cache_ent *ent)
 793{
 794	int order = order_base_2(ent->rb_key.ndescs);
 795	struct dentry *dir;
 796
 797	if (!mlx5_debugfs_root || dev->is_rep)
 798		return;
 799
 800	if (ent->rb_key.access_mode == MLX5_MKC_ACCESS_MODE_KSM)
 801		order = MLX5_IMR_KSM_CACHE_ENTRY + 2;
 802
 803	sprintf(ent->name, "%d", order);
 804	dir = debugfs_create_dir(ent->name, dev->cache.fs_root);
 805	debugfs_create_file("size", 0600, dir, ent, &size_fops);
 806	debugfs_create_file("limit", 0600, dir, ent, &limit_fops);
 807	debugfs_create_ulong("cur", 0400, dir, &ent->mkeys_queue.ci);
 808	debugfs_create_u32("miss", 0600, dir, &ent->miss);
 809}
 810
 811static void mlx5_mkey_cache_debugfs_init(struct mlx5_ib_dev *dev)
 812{
 813	struct dentry *dbg_root = mlx5_debugfs_get_dev_root(dev->mdev);
 814	struct mlx5_mkey_cache *cache = &dev->cache;
 815
 816	if (!mlx5_debugfs_root || dev->is_rep)
 817		return;
 818
 819	cache->fs_root = debugfs_create_dir("mr_cache", dbg_root);
 820}
 821
 822static void delay_time_func(struct timer_list *t)
 823{
 824	struct mlx5_ib_dev *dev = from_timer(dev, t, delay_timer);
 825
 826	WRITE_ONCE(dev->fill_delay, 0);
 827}
 828
 829static int mlx5r_mkeys_init(struct mlx5_cache_ent *ent)
 830{
 831	struct mlx5_mkeys_page *page;
 832
 833	page = kzalloc(sizeof(*page), GFP_KERNEL);
 834	if (!page)
 835		return -ENOMEM;
 836	INIT_LIST_HEAD(&ent->mkeys_queue.pages_list);
 837	spin_lock_init(&ent->mkeys_queue.lock);
 838	list_add_tail(&page->list, &ent->mkeys_queue.pages_list);
 839	ent->mkeys_queue.num_pages++;
 840	return 0;
 841}
 842
 843static void mlx5r_mkeys_uninit(struct mlx5_cache_ent *ent)
 844{
 845	struct mlx5_mkeys_page *page;
 846
 847	WARN_ON(ent->mkeys_queue.ci || ent->mkeys_queue.num_pages > 1);
 848	page = list_last_entry(&ent->mkeys_queue.pages_list,
 849			       struct mlx5_mkeys_page, list);
 850	list_del(&page->list);
 851	kfree(page);
 852}
 853
 854struct mlx5_cache_ent *
 855mlx5r_cache_create_ent_locked(struct mlx5_ib_dev *dev,
 856			      struct mlx5r_cache_rb_key rb_key,
 857			      bool persistent_entry)
 858{
 859	struct mlx5_cache_ent *ent;
 860	int order;
 861	int ret;
 862
 863	ent = kzalloc(sizeof(*ent), GFP_KERNEL);
 864	if (!ent)
 865		return ERR_PTR(-ENOMEM);
 866
 867	ret = mlx5r_mkeys_init(ent);
 868	if (ret)
 869		goto mkeys_err;
 870	ent->rb_key = rb_key;
 871	ent->dev = dev;
 872	ent->is_tmp = !persistent_entry;
 873
 874	INIT_DELAYED_WORK(&ent->dwork, delayed_cache_work_func);
 875
 876	ret = mlx5_cache_ent_insert(&dev->cache, ent);
 877	if (ret)
 878		goto ent_insert_err;
 879
 880	if (persistent_entry) {
 881		if (rb_key.access_mode == MLX5_MKC_ACCESS_MODE_KSM)
 882			order = MLX5_IMR_KSM_CACHE_ENTRY;
 883		else
 884			order = order_base_2(rb_key.ndescs) - 2;
 885
 886		if ((dev->mdev->profile.mask & MLX5_PROF_MASK_MR_CACHE) &&
 887		    !dev->is_rep && mlx5_core_is_pf(dev->mdev) &&
 888		    mlx5r_umr_can_load_pas(dev, 0))
 889			ent->limit = dev->mdev->profile.mr_cache[order].limit;
 890		else
 891			ent->limit = 0;
 892
 893		mlx5_mkey_cache_debugfs_add_ent(dev, ent);
 894	} else {
 895		mod_delayed_work(ent->dev->cache.wq,
 896				 &ent->dev->cache.remove_ent_dwork,
 897				 msecs_to_jiffies(30 * 1000));
 898	}
 899
 900	return ent;
 901ent_insert_err:
 902	mlx5r_mkeys_uninit(ent);
 903mkeys_err:
 904	kfree(ent);
 905	return ERR_PTR(ret);
 906}
 907
 908static void remove_ent_work_func(struct work_struct *work)
 909{
 910	struct mlx5_mkey_cache *cache;
 911	struct mlx5_cache_ent *ent;
 912	struct rb_node *cur;
 913
 914	cache = container_of(work, struct mlx5_mkey_cache,
 915			     remove_ent_dwork.work);
 916	mutex_lock(&cache->rb_lock);
 917	cur = rb_last(&cache->rb_root);
 918	while (cur) {
 919		ent = rb_entry(cur, struct mlx5_cache_ent, node);
 920		cur = rb_prev(cur);
 921		mutex_unlock(&cache->rb_lock);
 922
 923		spin_lock_irq(&ent->mkeys_queue.lock);
 924		if (!ent->is_tmp) {
 925			spin_unlock_irq(&ent->mkeys_queue.lock);
 926			mutex_lock(&cache->rb_lock);
 927			continue;
 928		}
 929		spin_unlock_irq(&ent->mkeys_queue.lock);
 930
 931		clean_keys(ent->dev, ent);
 932		mutex_lock(&cache->rb_lock);
 933	}
 934	mutex_unlock(&cache->rb_lock);
 935}
 936
 937int mlx5_mkey_cache_init(struct mlx5_ib_dev *dev)
 938{
 939	struct mlx5_mkey_cache *cache = &dev->cache;
 940	struct rb_root *root = &dev->cache.rb_root;
 941	struct mlx5r_cache_rb_key rb_key = {
 942		.access_mode = MLX5_MKC_ACCESS_MODE_MTT,
 943	};
 944	struct mlx5_cache_ent *ent;
 945	struct rb_node *node;
 946	int ret;
 947	int i;
 948
 949	mutex_init(&dev->slow_path_mutex);
 950	mutex_init(&dev->cache.rb_lock);
 951	dev->cache.rb_root = RB_ROOT;
 952	INIT_DELAYED_WORK(&dev->cache.remove_ent_dwork, remove_ent_work_func);
 953	cache->wq = alloc_ordered_workqueue("mkey_cache", WQ_MEM_RECLAIM);
 954	if (!cache->wq) {
 955		mlx5_ib_warn(dev, "failed to create work queue\n");
 956		return -ENOMEM;
 957	}
 958
 959	mlx5_cmd_init_async_ctx(dev->mdev, &dev->async_ctx);
 960	timer_setup(&dev->delay_timer, delay_time_func, 0);
 961	mlx5_mkey_cache_debugfs_init(dev);
 962	mutex_lock(&cache->rb_lock);
 963	for (i = 0; i <= mkey_cache_max_order(dev); i++) {
 964		rb_key.ndescs = 1 << (i + 2);
 965		ent = mlx5r_cache_create_ent_locked(dev, rb_key, true);
 966		if (IS_ERR(ent)) {
 967			ret = PTR_ERR(ent);
 968			goto err;
 969		}
 970	}
 971
 972	ret = mlx5_odp_init_mkey_cache(dev);
 973	if (ret)
 974		goto err;
 975
 976	mutex_unlock(&cache->rb_lock);
 977	for (node = rb_first(root); node; node = rb_next(node)) {
 978		ent = rb_entry(node, struct mlx5_cache_ent, node);
 979		spin_lock_irq(&ent->mkeys_queue.lock);
 980		queue_adjust_cache_locked(ent);
 981		spin_unlock_irq(&ent->mkeys_queue.lock);
 982	}
 983
 984	return 0;
 985
 986err:
 987	mutex_unlock(&cache->rb_lock);
 988	mlx5_mkey_cache_debugfs_cleanup(dev);
 989	mlx5_ib_warn(dev, "failed to create mkey cache entry\n");
 990	return ret;
 991}
 992
 993void mlx5_mkey_cache_cleanup(struct mlx5_ib_dev *dev)
 994{
 995	struct rb_root *root = &dev->cache.rb_root;
 996	struct mlx5_cache_ent *ent;
 997	struct rb_node *node;
 998
 999	if (!dev->cache.wq)
1000		return;
1001
1002	mutex_lock(&dev->cache.rb_lock);
1003	cancel_delayed_work(&dev->cache.remove_ent_dwork);
1004	for (node = rb_first(root); node; node = rb_next(node)) {
1005		ent = rb_entry(node, struct mlx5_cache_ent, node);
1006		spin_lock_irq(&ent->mkeys_queue.lock);
1007		ent->disabled = true;
1008		spin_unlock_irq(&ent->mkeys_queue.lock);
1009		cancel_delayed_work(&ent->dwork);
1010	}
1011	mutex_unlock(&dev->cache.rb_lock);
1012
1013	/*
1014	 * After all entries are disabled and will not reschedule on WQ,
1015	 * flush it and all async commands.
1016	 */
1017	flush_workqueue(dev->cache.wq);
1018
1019	mlx5_mkey_cache_debugfs_cleanup(dev);
1020	mlx5_cmd_cleanup_async_ctx(&dev->async_ctx);
1021
1022	/* At this point all entries are disabled and have no concurrent work. */
1023	mutex_lock(&dev->cache.rb_lock);
1024	node = rb_first(root);
1025	while (node) {
1026		ent = rb_entry(node, struct mlx5_cache_ent, node);
1027		node = rb_next(node);
1028		clean_keys(dev, ent);
1029		rb_erase(&ent->node, root);
1030		mlx5r_mkeys_uninit(ent);
1031		kfree(ent);
1032	}
1033	mutex_unlock(&dev->cache.rb_lock);
1034
1035	destroy_workqueue(dev->cache.wq);
1036	del_timer_sync(&dev->delay_timer);
1037}
1038
1039struct ib_mr *mlx5_ib_get_dma_mr(struct ib_pd *pd, int acc)
1040{
1041	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1042	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1043	struct mlx5_ib_mr *mr;
1044	void *mkc;
1045	u32 *in;
1046	int err;
1047
1048	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1049	if (!mr)
1050		return ERR_PTR(-ENOMEM);
1051
1052	in = kzalloc(inlen, GFP_KERNEL);
1053	if (!in) {
1054		err = -ENOMEM;
1055		goto err_free;
1056	}
1057
1058	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1059
1060	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_PA);
1061	MLX5_SET(mkc, mkc, length64, 1);
1062	set_mkc_access_pd_addr_fields(mkc, acc | IB_ACCESS_RELAXED_ORDERING, 0,
1063				      pd);
1064
1065	err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen);
1066	if (err)
1067		goto err_in;
1068
1069	kfree(in);
1070	mr->mmkey.type = MLX5_MKEY_MR;
1071	mr->ibmr.lkey = mr->mmkey.key;
1072	mr->ibmr.rkey = mr->mmkey.key;
1073	mr->umem = NULL;
1074
1075	return &mr->ibmr;
1076
1077err_in:
1078	kfree(in);
1079
1080err_free:
1081	kfree(mr);
1082
1083	return ERR_PTR(err);
1084}
1085
1086static int get_octo_len(u64 addr, u64 len, int page_shift)
1087{
1088	u64 page_size = 1ULL << page_shift;
1089	u64 offset;
1090	int npages;
1091
1092	offset = addr & (page_size - 1);
1093	npages = ALIGN(len + offset, page_size) >> page_shift;
1094	return (npages + 1) / 2;
1095}
1096
1097static int mkey_cache_max_order(struct mlx5_ib_dev *dev)
1098{
1099	if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
1100		return MKEY_CACHE_LAST_STD_ENTRY;
1101	return MLX5_MAX_UMR_SHIFT;
1102}
1103
1104static void set_mr_fields(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr,
1105			  u64 length, int access_flags, u64 iova)
1106{
1107	mr->ibmr.lkey = mr->mmkey.key;
1108	mr->ibmr.rkey = mr->mmkey.key;
1109	mr->ibmr.length = length;
1110	mr->ibmr.device = &dev->ib_dev;
1111	mr->ibmr.iova = iova;
1112	mr->access_flags = access_flags;
1113}
1114
1115static unsigned int mlx5_umem_dmabuf_default_pgsz(struct ib_umem *umem,
1116						  u64 iova)
1117{
1118	/*
1119	 * The alignment of iova has already been checked upon entering
1120	 * UVERBS_METHOD_REG_DMABUF_MR
1121	 */
1122	umem->iova = iova;
1123	return PAGE_SIZE;
1124}
1125
1126static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd,
1127					     struct ib_umem *umem, u64 iova,
1128					     int access_flags)
1129{
1130	struct mlx5r_cache_rb_key rb_key = {
1131		.access_mode = MLX5_MKC_ACCESS_MODE_MTT,
1132	};
1133	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1134	struct mlx5_cache_ent *ent;
1135	struct mlx5_ib_mr *mr;
1136	unsigned int page_size;
1137
1138	if (umem->is_dmabuf)
1139		page_size = mlx5_umem_dmabuf_default_pgsz(umem, iova);
1140	else
1141		page_size = mlx5_umem_find_best_pgsz(umem, mkc, log_page_size,
1142						     0, iova);
1143	if (WARN_ON(!page_size))
1144		return ERR_PTR(-EINVAL);
1145
1146	rb_key.ndescs = ib_umem_num_dma_blocks(umem, page_size);
1147	rb_key.ats = mlx5_umem_needs_ats(dev, umem, access_flags);
1148	rb_key.access_flags = get_unchangeable_access_flags(dev, access_flags);
1149	ent = mkey_cache_ent_from_rb_key(dev, rb_key);
1150	/*
1151	 * If the MR can't come from the cache then synchronously create an uncached
1152	 * one.
1153	 */
1154	if (!ent) {
1155		mutex_lock(&dev->slow_path_mutex);
1156		mr = reg_create(pd, umem, iova, access_flags, page_size, false);
1157		mutex_unlock(&dev->slow_path_mutex);
1158		if (IS_ERR(mr))
1159			return mr;
1160		mr->mmkey.rb_key = rb_key;
 
1161		return mr;
1162	}
1163
1164	mr = _mlx5_mr_cache_alloc(dev, ent, access_flags);
1165	if (IS_ERR(mr))
1166		return mr;
1167
1168	mr->ibmr.pd = pd;
1169	mr->umem = umem;
1170	mr->page_shift = order_base_2(page_size);
 
1171	set_mr_fields(dev, mr, umem->length, access_flags, iova);
1172
1173	return mr;
1174}
1175
1176/*
1177 * If ibmr is NULL it will be allocated by reg_create.
1178 * Else, the given ibmr will be used.
1179 */
1180static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem,
1181				     u64 iova, int access_flags,
1182				     unsigned int page_size, bool populate)
1183{
1184	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1185	struct mlx5_ib_mr *mr;
1186	__be64 *pas;
1187	void *mkc;
1188	int inlen;
1189	u32 *in;
1190	int err;
1191	bool pg_cap = !!(MLX5_CAP_GEN(dev->mdev, pg));
1192
1193	if (!page_size)
1194		return ERR_PTR(-EINVAL);
1195	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1196	if (!mr)
1197		return ERR_PTR(-ENOMEM);
1198
1199	mr->ibmr.pd = pd;
1200	mr->access_flags = access_flags;
1201	mr->page_shift = order_base_2(page_size);
1202
1203	inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1204	if (populate)
1205		inlen += sizeof(*pas) *
1206			 roundup(ib_umem_num_dma_blocks(umem, page_size), 2);
1207	in = kvzalloc(inlen, GFP_KERNEL);
1208	if (!in) {
1209		err = -ENOMEM;
1210		goto err_1;
1211	}
1212	pas = (__be64 *)MLX5_ADDR_OF(create_mkey_in, in, klm_pas_mtt);
1213	if (populate) {
1214		if (WARN_ON(access_flags & IB_ACCESS_ON_DEMAND)) {
1215			err = -EINVAL;
1216			goto err_2;
1217		}
1218		mlx5_ib_populate_pas(umem, 1UL << mr->page_shift, pas,
1219				     pg_cap ? MLX5_IB_MTT_PRESENT : 0);
1220	}
1221
1222	/* The pg_access bit allows setting the access flags
1223	 * in the page list submitted with the command.
1224	 */
1225	MLX5_SET(create_mkey_in, in, pg_access, !!(pg_cap));
1226
1227	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1228	set_mkc_access_pd_addr_fields(mkc, access_flags, iova,
1229				      populate ? pd : dev->umrc.pd);
1230	MLX5_SET(mkc, mkc, free, !populate);
1231	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_MTT);
1232	MLX5_SET(mkc, mkc, umr_en, 1);
1233
1234	MLX5_SET64(mkc, mkc, len, umem->length);
1235	MLX5_SET(mkc, mkc, bsf_octword_size, 0);
1236	MLX5_SET(mkc, mkc, translations_octword_size,
1237		 get_octo_len(iova, umem->length, mr->page_shift));
1238	MLX5_SET(mkc, mkc, log_page_size, mr->page_shift);
1239	if (mlx5_umem_needs_ats(dev, umem, access_flags))
1240		MLX5_SET(mkc, mkc, ma_translation_mode, 1);
1241	if (populate) {
1242		MLX5_SET(create_mkey_in, in, translations_octword_actual_size,
1243			 get_octo_len(iova, umem->length, mr->page_shift));
1244	}
1245
1246	err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen);
1247	if (err) {
1248		mlx5_ib_warn(dev, "create mkey failed\n");
1249		goto err_2;
1250	}
1251	mr->mmkey.type = MLX5_MKEY_MR;
1252	mr->mmkey.ndescs = get_octo_len(iova, umem->length, mr->page_shift);
1253	mr->umem = umem;
1254	set_mr_fields(dev, mr, umem->length, access_flags, iova);
1255	kvfree(in);
1256
1257	mlx5_ib_dbg(dev, "mkey = 0x%x\n", mr->mmkey.key);
1258
1259	return mr;
1260
1261err_2:
1262	kvfree(in);
1263err_1:
1264	kfree(mr);
1265	return ERR_PTR(err);
1266}
1267
1268static struct ib_mr *mlx5_ib_get_dm_mr(struct ib_pd *pd, u64 start_addr,
1269				       u64 length, int acc, int mode)
1270{
1271	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1272	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1273	struct mlx5_ib_mr *mr;
1274	void *mkc;
1275	u32 *in;
1276	int err;
1277
1278	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1279	if (!mr)
1280		return ERR_PTR(-ENOMEM);
1281
1282	in = kzalloc(inlen, GFP_KERNEL);
1283	if (!in) {
1284		err = -ENOMEM;
1285		goto err_free;
1286	}
1287
1288	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1289
1290	MLX5_SET(mkc, mkc, access_mode_1_0, mode & 0x3);
1291	MLX5_SET(mkc, mkc, access_mode_4_2, (mode >> 2) & 0x7);
1292	MLX5_SET64(mkc, mkc, len, length);
1293	set_mkc_access_pd_addr_fields(mkc, acc, start_addr, pd);
1294
1295	err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen);
1296	if (err)
1297		goto err_in;
1298
1299	kfree(in);
1300
1301	set_mr_fields(dev, mr, length, acc, start_addr);
1302
1303	return &mr->ibmr;
1304
1305err_in:
1306	kfree(in);
1307
1308err_free:
1309	kfree(mr);
1310
1311	return ERR_PTR(err);
1312}
1313
1314int mlx5_ib_advise_mr(struct ib_pd *pd,
1315		      enum ib_uverbs_advise_mr_advice advice,
1316		      u32 flags,
1317		      struct ib_sge *sg_list,
1318		      u32 num_sge,
1319		      struct uverbs_attr_bundle *attrs)
1320{
1321	if (advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH &&
1322	    advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
1323	    advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT)
1324		return -EOPNOTSUPP;
1325
1326	return mlx5_ib_advise_mr_prefetch(pd, advice, flags,
1327					 sg_list, num_sge);
1328}
1329
1330struct ib_mr *mlx5_ib_reg_dm_mr(struct ib_pd *pd, struct ib_dm *dm,
1331				struct ib_dm_mr_attr *attr,
1332				struct uverbs_attr_bundle *attrs)
1333{
1334	struct mlx5_ib_dm *mdm = to_mdm(dm);
1335	struct mlx5_core_dev *dev = to_mdev(dm->device)->mdev;
1336	u64 start_addr = mdm->dev_addr + attr->offset;
1337	int mode;
1338
1339	switch (mdm->type) {
1340	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
1341		if (attr->access_flags & ~MLX5_IB_DM_MEMIC_ALLOWED_ACCESS)
1342			return ERR_PTR(-EINVAL);
1343
1344		mode = MLX5_MKC_ACCESS_MODE_MEMIC;
1345		start_addr -= pci_resource_start(dev->pdev, 0);
1346		break;
1347	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
1348	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
1349	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_PATTERN_SW_ICM:
1350	case MLX5_IB_UAPI_DM_TYPE_ENCAP_SW_ICM:
1351		if (attr->access_flags & ~MLX5_IB_DM_SW_ICM_ALLOWED_ACCESS)
1352			return ERR_PTR(-EINVAL);
1353
1354		mode = MLX5_MKC_ACCESS_MODE_SW_ICM;
1355		break;
1356	default:
1357		return ERR_PTR(-EINVAL);
1358	}
1359
1360	return mlx5_ib_get_dm_mr(pd, start_addr, attr->length,
1361				 attr->access_flags, mode);
1362}
1363
1364static struct ib_mr *create_real_mr(struct ib_pd *pd, struct ib_umem *umem,
1365				    u64 iova, int access_flags)
1366{
1367	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1368	struct mlx5_ib_mr *mr = NULL;
1369	bool xlt_with_umr;
1370	int err;
1371
1372	xlt_with_umr = mlx5r_umr_can_load_pas(dev, umem->length);
1373	if (xlt_with_umr) {
1374		mr = alloc_cacheable_mr(pd, umem, iova, access_flags);
1375	} else {
1376		unsigned int page_size = mlx5_umem_find_best_pgsz(
1377			umem, mkc, log_page_size, 0, iova);
1378
1379		mutex_lock(&dev->slow_path_mutex);
1380		mr = reg_create(pd, umem, iova, access_flags, page_size, true);
1381		mutex_unlock(&dev->slow_path_mutex);
1382	}
1383	if (IS_ERR(mr)) {
1384		ib_umem_release(umem);
1385		return ERR_CAST(mr);
1386	}
1387
1388	mlx5_ib_dbg(dev, "mkey 0x%x\n", mr->mmkey.key);
1389
1390	atomic_add(ib_umem_num_pages(umem), &dev->mdev->priv.reg_pages);
1391
1392	if (xlt_with_umr) {
1393		/*
1394		 * If the MR was created with reg_create then it will be
1395		 * configured properly but left disabled. It is safe to go ahead
1396		 * and configure it again via UMR while enabling it.
1397		 */
1398		err = mlx5r_umr_update_mr_pas(mr, MLX5_IB_UPD_XLT_ENABLE);
1399		if (err) {
1400			mlx5_ib_dereg_mr(&mr->ibmr, NULL);
1401			return ERR_PTR(err);
1402		}
1403	}
1404	return &mr->ibmr;
1405}
1406
1407static struct ib_mr *create_user_odp_mr(struct ib_pd *pd, u64 start, u64 length,
1408					u64 iova, int access_flags,
1409					struct ib_udata *udata)
1410{
1411	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1412	struct ib_umem_odp *odp;
1413	struct mlx5_ib_mr *mr;
1414	int err;
1415
1416	if (!IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING))
1417		return ERR_PTR(-EOPNOTSUPP);
1418
1419	err = mlx5r_odp_create_eq(dev, &dev->odp_pf_eq);
1420	if (err)
1421		return ERR_PTR(err);
1422	if (!start && length == U64_MAX) {
1423		if (iova != 0)
1424			return ERR_PTR(-EINVAL);
1425		if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
1426			return ERR_PTR(-EINVAL);
1427
1428		mr = mlx5_ib_alloc_implicit_mr(to_mpd(pd), access_flags);
1429		if (IS_ERR(mr))
1430			return ERR_CAST(mr);
1431		return &mr->ibmr;
1432	}
1433
1434	/* ODP requires xlt update via umr to work. */
1435	if (!mlx5r_umr_can_load_pas(dev, length))
1436		return ERR_PTR(-EINVAL);
1437
1438	odp = ib_umem_odp_get(&dev->ib_dev, start, length, access_flags,
1439			      &mlx5_mn_ops);
1440	if (IS_ERR(odp))
1441		return ERR_CAST(odp);
1442
1443	mr = alloc_cacheable_mr(pd, &odp->umem, iova, access_flags);
1444	if (IS_ERR(mr)) {
1445		ib_umem_release(&odp->umem);
1446		return ERR_CAST(mr);
1447	}
1448	xa_init(&mr->implicit_children);
1449
1450	odp->private = mr;
1451	err = mlx5r_store_odp_mkey(dev, &mr->mmkey);
1452	if (err)
1453		goto err_dereg_mr;
1454
1455	err = mlx5_ib_init_odp_mr(mr);
1456	if (err)
1457		goto err_dereg_mr;
1458	return &mr->ibmr;
1459
1460err_dereg_mr:
1461	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
1462	return ERR_PTR(err);
1463}
1464
1465struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
1466				  u64 iova, int access_flags,
1467				  struct ib_udata *udata)
1468{
1469	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1470	struct ib_umem *umem;
1471
1472	if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM))
1473		return ERR_PTR(-EOPNOTSUPP);
1474
1475	mlx5_ib_dbg(dev, "start 0x%llx, iova 0x%llx, length 0x%llx, access_flags 0x%x\n",
1476		    start, iova, length, access_flags);
1477
1478	if (access_flags & IB_ACCESS_ON_DEMAND)
1479		return create_user_odp_mr(pd, start, length, iova, access_flags,
1480					  udata);
1481	umem = ib_umem_get(&dev->ib_dev, start, length, access_flags);
1482	if (IS_ERR(umem))
1483		return ERR_CAST(umem);
1484	return create_real_mr(pd, umem, iova, access_flags);
1485}
1486
1487static void mlx5_ib_dmabuf_invalidate_cb(struct dma_buf_attachment *attach)
1488{
1489	struct ib_umem_dmabuf *umem_dmabuf = attach->importer_priv;
1490	struct mlx5_ib_mr *mr = umem_dmabuf->private;
1491
1492	dma_resv_assert_held(umem_dmabuf->attach->dmabuf->resv);
1493
1494	if (!umem_dmabuf->sgt)
1495		return;
1496
1497	mlx5r_umr_update_mr_pas(mr, MLX5_IB_UPD_XLT_ZAP);
1498	ib_umem_dmabuf_unmap_pages(umem_dmabuf);
1499}
1500
1501static struct dma_buf_attach_ops mlx5_ib_dmabuf_attach_ops = {
1502	.allow_peer2peer = 1,
1503	.move_notify = mlx5_ib_dmabuf_invalidate_cb,
1504};
1505
1506struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 offset,
1507					 u64 length, u64 virt_addr,
1508					 int fd, int access_flags,
1509					 struct ib_udata *udata)
1510{
1511	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1512	struct mlx5_ib_mr *mr = NULL;
1513	struct ib_umem_dmabuf *umem_dmabuf;
1514	int err;
1515
1516	if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM) ||
1517	    !IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING))
1518		return ERR_PTR(-EOPNOTSUPP);
1519
1520	mlx5_ib_dbg(dev,
1521		    "offset 0x%llx, virt_addr 0x%llx, length 0x%llx, fd %d, access_flags 0x%x\n",
1522		    offset, virt_addr, length, fd, access_flags);
1523
1524	/* dmabuf requires xlt update via umr to work. */
1525	if (!mlx5r_umr_can_load_pas(dev, length))
1526		return ERR_PTR(-EINVAL);
1527
1528	umem_dmabuf = ib_umem_dmabuf_get(&dev->ib_dev, offset, length, fd,
1529					 access_flags,
1530					 &mlx5_ib_dmabuf_attach_ops);
1531	if (IS_ERR(umem_dmabuf)) {
1532		mlx5_ib_dbg(dev, "umem_dmabuf get failed (%ld)\n",
1533			    PTR_ERR(umem_dmabuf));
1534		return ERR_CAST(umem_dmabuf);
1535	}
1536
1537	mr = alloc_cacheable_mr(pd, &umem_dmabuf->umem, virt_addr,
1538				access_flags);
1539	if (IS_ERR(mr)) {
1540		ib_umem_release(&umem_dmabuf->umem);
1541		return ERR_CAST(mr);
1542	}
1543
1544	mlx5_ib_dbg(dev, "mkey 0x%x\n", mr->mmkey.key);
1545
1546	atomic_add(ib_umem_num_pages(mr->umem), &dev->mdev->priv.reg_pages);
1547	umem_dmabuf->private = mr;
1548	err = mlx5r_store_odp_mkey(dev, &mr->mmkey);
1549	if (err)
1550		goto err_dereg_mr;
1551
1552	err = mlx5_ib_init_dmabuf_mr(mr);
1553	if (err)
1554		goto err_dereg_mr;
1555	return &mr->ibmr;
1556
1557err_dereg_mr:
1558	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
1559	return ERR_PTR(err);
1560}
1561
1562/*
1563 * True if the change in access flags can be done via UMR, only some access
1564 * flags can be updated.
1565 */
1566static bool can_use_umr_rereg_access(struct mlx5_ib_dev *dev,
1567				     unsigned int current_access_flags,
1568				     unsigned int target_access_flags)
1569{
1570	unsigned int diffs = current_access_flags ^ target_access_flags;
1571
1572	if (diffs & ~(IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE |
1573		      IB_ACCESS_REMOTE_READ | IB_ACCESS_RELAXED_ORDERING))
 
1574		return false;
1575	return mlx5r_umr_can_reconfig(dev, current_access_flags,
1576				      target_access_flags);
1577}
1578
1579static bool can_use_umr_rereg_pas(struct mlx5_ib_mr *mr,
1580				  struct ib_umem *new_umem,
1581				  int new_access_flags, u64 iova,
1582				  unsigned long *page_size)
1583{
1584	struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device);
1585
1586	/* We only track the allocated sizes of MRs from the cache */
1587	if (!mr->mmkey.cache_ent)
1588		return false;
1589	if (!mlx5r_umr_can_load_pas(dev, new_umem->length))
1590		return false;
1591
1592	*page_size =
1593		mlx5_umem_find_best_pgsz(new_umem, mkc, log_page_size, 0, iova);
1594	if (WARN_ON(!*page_size))
1595		return false;
1596	return (mr->mmkey.cache_ent->rb_key.ndescs) >=
1597	       ib_umem_num_dma_blocks(new_umem, *page_size);
1598}
1599
1600static int umr_rereg_pas(struct mlx5_ib_mr *mr, struct ib_pd *pd,
1601			 int access_flags, int flags, struct ib_umem *new_umem,
1602			 u64 iova, unsigned long page_size)
1603{
1604	struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device);
1605	int upd_flags = MLX5_IB_UPD_XLT_ADDR | MLX5_IB_UPD_XLT_ENABLE;
1606	struct ib_umem *old_umem = mr->umem;
1607	int err;
1608
1609	/*
1610	 * To keep everything simple the MR is revoked before we start to mess
1611	 * with it. This ensure the change is atomic relative to any use of the
1612	 * MR.
1613	 */
1614	err = mlx5r_umr_revoke_mr(mr);
1615	if (err)
1616		return err;
1617
1618	if (flags & IB_MR_REREG_PD) {
1619		mr->ibmr.pd = pd;
1620		upd_flags |= MLX5_IB_UPD_XLT_PD;
1621	}
1622	if (flags & IB_MR_REREG_ACCESS) {
1623		mr->access_flags = access_flags;
1624		upd_flags |= MLX5_IB_UPD_XLT_ACCESS;
1625	}
1626
1627	mr->ibmr.iova = iova;
1628	mr->ibmr.length = new_umem->length;
1629	mr->page_shift = order_base_2(page_size);
1630	mr->umem = new_umem;
1631	err = mlx5r_umr_update_mr_pas(mr, upd_flags);
1632	if (err) {
1633		/*
1634		 * The MR is revoked at this point so there is no issue to free
1635		 * new_umem.
1636		 */
1637		mr->umem = old_umem;
1638		return err;
1639	}
1640
1641	atomic_sub(ib_umem_num_pages(old_umem), &dev->mdev->priv.reg_pages);
1642	ib_umem_release(old_umem);
1643	atomic_add(ib_umem_num_pages(new_umem), &dev->mdev->priv.reg_pages);
1644	return 0;
1645}
1646
1647struct ib_mr *mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start,
1648				    u64 length, u64 iova, int new_access_flags,
1649				    struct ib_pd *new_pd,
1650				    struct ib_udata *udata)
1651{
1652	struct mlx5_ib_dev *dev = to_mdev(ib_mr->device);
1653	struct mlx5_ib_mr *mr = to_mmr(ib_mr);
1654	int err;
1655
1656	if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM))
1657		return ERR_PTR(-EOPNOTSUPP);
1658
1659	mlx5_ib_dbg(
1660		dev,
1661		"start 0x%llx, iova 0x%llx, length 0x%llx, access_flags 0x%x\n",
1662		start, iova, length, new_access_flags);
1663
1664	if (flags & ~(IB_MR_REREG_TRANS | IB_MR_REREG_PD | IB_MR_REREG_ACCESS))
1665		return ERR_PTR(-EOPNOTSUPP);
1666
1667	if (!(flags & IB_MR_REREG_ACCESS))
1668		new_access_flags = mr->access_flags;
1669	if (!(flags & IB_MR_REREG_PD))
1670		new_pd = ib_mr->pd;
1671
1672	if (!(flags & IB_MR_REREG_TRANS)) {
1673		struct ib_umem *umem;
1674
1675		/* Fast path for PD/access change */
1676		if (can_use_umr_rereg_access(dev, mr->access_flags,
1677					     new_access_flags)) {
1678			err = mlx5r_umr_rereg_pd_access(mr, new_pd,
1679							new_access_flags);
1680			if (err)
1681				return ERR_PTR(err);
1682			return NULL;
1683		}
1684		/* DM or ODP MR's don't have a normal umem so we can't re-use it */
1685		if (!mr->umem || is_odp_mr(mr) || is_dmabuf_mr(mr))
1686			goto recreate;
1687
1688		/*
1689		 * Only one active MR can refer to a umem at one time, revoke
1690		 * the old MR before assigning the umem to the new one.
1691		 */
1692		err = mlx5r_umr_revoke_mr(mr);
1693		if (err)
1694			return ERR_PTR(err);
1695		umem = mr->umem;
1696		mr->umem = NULL;
1697		atomic_sub(ib_umem_num_pages(umem), &dev->mdev->priv.reg_pages);
1698
1699		return create_real_mr(new_pd, umem, mr->ibmr.iova,
1700				      new_access_flags);
1701	}
1702
1703	/*
1704	 * DM doesn't have a PAS list so we can't re-use it, odp/dmabuf does
1705	 * but the logic around releasing the umem is different
1706	 */
1707	if (!mr->umem || is_odp_mr(mr) || is_dmabuf_mr(mr))
1708		goto recreate;
1709
1710	if (!(new_access_flags & IB_ACCESS_ON_DEMAND) &&
1711	    can_use_umr_rereg_access(dev, mr->access_flags, new_access_flags)) {
1712		struct ib_umem *new_umem;
1713		unsigned long page_size;
1714
1715		new_umem = ib_umem_get(&dev->ib_dev, start, length,
1716				       new_access_flags);
1717		if (IS_ERR(new_umem))
1718			return ERR_CAST(new_umem);
1719
1720		/* Fast path for PAS change */
1721		if (can_use_umr_rereg_pas(mr, new_umem, new_access_flags, iova,
1722					  &page_size)) {
1723			err = umr_rereg_pas(mr, new_pd, new_access_flags, flags,
1724					    new_umem, iova, page_size);
1725			if (err) {
1726				ib_umem_release(new_umem);
1727				return ERR_PTR(err);
1728			}
1729			return NULL;
1730		}
1731		return create_real_mr(new_pd, new_umem, iova, new_access_flags);
1732	}
1733
1734	/*
1735	 * Everything else has no state we can preserve, just create a new MR
1736	 * from scratch
1737	 */
1738recreate:
1739	return mlx5_ib_reg_user_mr(new_pd, start, length, iova,
1740				   new_access_flags, udata);
1741}
1742
1743static int
1744mlx5_alloc_priv_descs(struct ib_device *device,
1745		      struct mlx5_ib_mr *mr,
1746		      int ndescs,
1747		      int desc_size)
1748{
1749	struct mlx5_ib_dev *dev = to_mdev(device);
1750	struct device *ddev = &dev->mdev->pdev->dev;
1751	int size = ndescs * desc_size;
1752	int add_size;
1753	int ret;
1754
1755	add_size = max_t(int, MLX5_UMR_ALIGN - ARCH_KMALLOC_MINALIGN, 0);
1756	if (is_power_of_2(MLX5_UMR_ALIGN) && add_size) {
1757		int end = max_t(int, MLX5_UMR_ALIGN, roundup_pow_of_two(size));
1758
1759		add_size = min_t(int, end - size, add_size);
1760	}
1761
1762	mr->descs_alloc = kzalloc(size + add_size, GFP_KERNEL);
1763	if (!mr->descs_alloc)
1764		return -ENOMEM;
1765
1766	mr->descs = PTR_ALIGN(mr->descs_alloc, MLX5_UMR_ALIGN);
1767
1768	mr->desc_map = dma_map_single(ddev, mr->descs, size, DMA_TO_DEVICE);
1769	if (dma_mapping_error(ddev, mr->desc_map)) {
1770		ret = -ENOMEM;
1771		goto err;
1772	}
1773
1774	return 0;
1775err:
1776	kfree(mr->descs_alloc);
1777
1778	return ret;
1779}
1780
1781static void
1782mlx5_free_priv_descs(struct mlx5_ib_mr *mr)
1783{
1784	if (!mr->umem && mr->descs) {
1785		struct ib_device *device = mr->ibmr.device;
1786		int size = mr->max_descs * mr->desc_size;
1787		struct mlx5_ib_dev *dev = to_mdev(device);
1788
1789		dma_unmap_single(&dev->mdev->pdev->dev, mr->desc_map, size,
1790				 DMA_TO_DEVICE);
1791		kfree(mr->descs_alloc);
1792		mr->descs = NULL;
1793	}
1794}
1795
1796static int cache_ent_find_and_store(struct mlx5_ib_dev *dev,
1797				    struct mlx5_ib_mr *mr)
1798{
1799	struct mlx5_mkey_cache *cache = &dev->cache;
1800	struct mlx5_cache_ent *ent;
1801	int ret;
1802
1803	if (mr->mmkey.cache_ent) {
1804		spin_lock_irq(&mr->mmkey.cache_ent->mkeys_queue.lock);
1805		mr->mmkey.cache_ent->in_use--;
1806		goto end;
1807	}
1808
1809	mutex_lock(&cache->rb_lock);
1810	ent = mkey_cache_ent_from_rb_key(dev, mr->mmkey.rb_key);
1811	if (ent) {
1812		if (ent->rb_key.ndescs == mr->mmkey.rb_key.ndescs) {
1813			if (ent->disabled) {
1814				mutex_unlock(&cache->rb_lock);
1815				return -EOPNOTSUPP;
1816			}
1817			mr->mmkey.cache_ent = ent;
1818			spin_lock_irq(&mr->mmkey.cache_ent->mkeys_queue.lock);
1819			mutex_unlock(&cache->rb_lock);
1820			goto end;
1821		}
1822	}
1823
1824	ent = mlx5r_cache_create_ent_locked(dev, mr->mmkey.rb_key, false);
1825	mutex_unlock(&cache->rb_lock);
1826	if (IS_ERR(ent))
1827		return PTR_ERR(ent);
1828
1829	mr->mmkey.cache_ent = ent;
1830	spin_lock_irq(&mr->mmkey.cache_ent->mkeys_queue.lock);
1831
1832end:
1833	ret = push_mkey_locked(mr->mmkey.cache_ent, mr->mmkey.key);
1834	spin_unlock_irq(&mr->mmkey.cache_ent->mkeys_queue.lock);
1835	return ret;
1836}
1837
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1838int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1839{
1840	struct mlx5_ib_mr *mr = to_mmr(ibmr);
1841	struct mlx5_ib_dev *dev = to_mdev(ibmr->device);
1842	int rc;
1843
1844	/*
1845	 * Any async use of the mr must hold the refcount, once the refcount
1846	 * goes to zero no other thread, such as ODP page faults, prefetch, any
1847	 * UMR activity, etc can touch the mkey. Thus it is safe to destroy it.
1848	 */
1849	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) &&
1850	    refcount_read(&mr->mmkey.usecount) != 0 &&
1851	    xa_erase(&mr_to_mdev(mr)->odp_mkeys, mlx5_base_mkey(mr->mmkey.key)))
1852		mlx5r_deref_wait_odp_mkey(&mr->mmkey);
1853
1854	if (ibmr->type == IB_MR_TYPE_INTEGRITY) {
1855		xa_cmpxchg(&dev->sig_mrs, mlx5_base_mkey(mr->mmkey.key),
1856			   mr->sig, NULL, GFP_KERNEL);
1857
1858		if (mr->mtt_mr) {
1859			rc = mlx5_ib_dereg_mr(&mr->mtt_mr->ibmr, NULL);
1860			if (rc)
1861				return rc;
1862			mr->mtt_mr = NULL;
1863		}
1864		if (mr->klm_mr) {
1865			rc = mlx5_ib_dereg_mr(&mr->klm_mr->ibmr, NULL);
1866			if (rc)
1867				return rc;
1868			mr->klm_mr = NULL;
1869		}
1870
1871		if (mlx5_core_destroy_psv(dev->mdev,
1872					  mr->sig->psv_memory.psv_idx))
1873			mlx5_ib_warn(dev, "failed to destroy mem psv %d\n",
1874				     mr->sig->psv_memory.psv_idx);
1875		if (mlx5_core_destroy_psv(dev->mdev, mr->sig->psv_wire.psv_idx))
1876			mlx5_ib_warn(dev, "failed to destroy wire psv %d\n",
1877				     mr->sig->psv_wire.psv_idx);
1878		kfree(mr->sig);
1879		mr->sig = NULL;
1880	}
1881
1882	/* Stop DMA */
1883	if (mr->umem && mlx5r_umr_can_load_pas(dev, mr->umem->length))
1884		if (mlx5r_umr_revoke_mr(mr) ||
1885		    cache_ent_find_and_store(dev, mr))
1886			mr->mmkey.cache_ent = NULL;
1887
1888	if (!mr->mmkey.cache_ent) {
1889		rc = destroy_mkey(to_mdev(mr->ibmr.device), mr);
1890		if (rc)
1891			return rc;
1892	}
1893
1894	if (mr->umem) {
1895		bool is_odp = is_odp_mr(mr);
1896
1897		if (!is_odp)
1898			atomic_sub(ib_umem_num_pages(mr->umem),
1899				   &dev->mdev->priv.reg_pages);
1900		ib_umem_release(mr->umem);
1901		if (is_odp)
1902			mlx5_ib_free_odp_mr(mr);
1903	}
1904
1905	if (!mr->mmkey.cache_ent)
1906		mlx5_free_priv_descs(mr);
1907
1908	kfree(mr);
1909	return 0;
1910}
1911
1912static void mlx5_set_umr_free_mkey(struct ib_pd *pd, u32 *in, int ndescs,
1913				   int access_mode, int page_shift)
1914{
1915	void *mkc;
1916
1917	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1918
1919	/* This is only used from the kernel, so setting the PD is OK. */
1920	set_mkc_access_pd_addr_fields(mkc, IB_ACCESS_RELAXED_ORDERING, 0, pd);
1921	MLX5_SET(mkc, mkc, free, 1);
1922	MLX5_SET(mkc, mkc, translations_octword_size, ndescs);
1923	MLX5_SET(mkc, mkc, access_mode_1_0, access_mode & 0x3);
1924	MLX5_SET(mkc, mkc, access_mode_4_2, (access_mode >> 2) & 0x7);
1925	MLX5_SET(mkc, mkc, umr_en, 1);
1926	MLX5_SET(mkc, mkc, log_page_size, page_shift);
1927}
1928
1929static int _mlx5_alloc_mkey_descs(struct ib_pd *pd, struct mlx5_ib_mr *mr,
1930				  int ndescs, int desc_size, int page_shift,
1931				  int access_mode, u32 *in, int inlen)
1932{
1933	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1934	int err;
1935
1936	mr->access_mode = access_mode;
1937	mr->desc_size = desc_size;
1938	mr->max_descs = ndescs;
1939
1940	err = mlx5_alloc_priv_descs(pd->device, mr, ndescs, desc_size);
1941	if (err)
1942		return err;
1943
1944	mlx5_set_umr_free_mkey(pd, in, ndescs, access_mode, page_shift);
1945
1946	err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen);
1947	if (err)
1948		goto err_free_descs;
1949
1950	mr->mmkey.type = MLX5_MKEY_MR;
1951	mr->ibmr.lkey = mr->mmkey.key;
1952	mr->ibmr.rkey = mr->mmkey.key;
1953
1954	return 0;
1955
1956err_free_descs:
1957	mlx5_free_priv_descs(mr);
1958	return err;
1959}
1960
1961static struct mlx5_ib_mr *mlx5_ib_alloc_pi_mr(struct ib_pd *pd,
1962				u32 max_num_sg, u32 max_num_meta_sg,
1963				int desc_size, int access_mode)
1964{
1965	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1966	int ndescs = ALIGN(max_num_sg + max_num_meta_sg, 4);
1967	int page_shift = 0;
1968	struct mlx5_ib_mr *mr;
1969	u32 *in;
1970	int err;
1971
1972	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1973	if (!mr)
1974		return ERR_PTR(-ENOMEM);
1975
1976	mr->ibmr.pd = pd;
1977	mr->ibmr.device = pd->device;
1978
1979	in = kzalloc(inlen, GFP_KERNEL);
1980	if (!in) {
1981		err = -ENOMEM;
1982		goto err_free;
1983	}
1984
1985	if (access_mode == MLX5_MKC_ACCESS_MODE_MTT)
1986		page_shift = PAGE_SHIFT;
1987
1988	err = _mlx5_alloc_mkey_descs(pd, mr, ndescs, desc_size, page_shift,
1989				     access_mode, in, inlen);
1990	if (err)
1991		goto err_free_in;
1992
1993	mr->umem = NULL;
1994	kfree(in);
1995
1996	return mr;
1997
1998err_free_in:
1999	kfree(in);
2000err_free:
2001	kfree(mr);
2002	return ERR_PTR(err);
2003}
2004
2005static int mlx5_alloc_mem_reg_descs(struct ib_pd *pd, struct mlx5_ib_mr *mr,
2006				    int ndescs, u32 *in, int inlen)
2007{
2008	return _mlx5_alloc_mkey_descs(pd, mr, ndescs, sizeof(struct mlx5_mtt),
2009				      PAGE_SHIFT, MLX5_MKC_ACCESS_MODE_MTT, in,
2010				      inlen);
2011}
2012
2013static int mlx5_alloc_sg_gaps_descs(struct ib_pd *pd, struct mlx5_ib_mr *mr,
2014				    int ndescs, u32 *in, int inlen)
2015{
2016	return _mlx5_alloc_mkey_descs(pd, mr, ndescs, sizeof(struct mlx5_klm),
2017				      0, MLX5_MKC_ACCESS_MODE_KLMS, in, inlen);
2018}
2019
2020static int mlx5_alloc_integrity_descs(struct ib_pd *pd, struct mlx5_ib_mr *mr,
2021				      int max_num_sg, int max_num_meta_sg,
2022				      u32 *in, int inlen)
2023{
2024	struct mlx5_ib_dev *dev = to_mdev(pd->device);
2025	u32 psv_index[2];
2026	void *mkc;
2027	int err;
2028
2029	mr->sig = kzalloc(sizeof(*mr->sig), GFP_KERNEL);
2030	if (!mr->sig)
2031		return -ENOMEM;
2032
2033	/* create mem & wire PSVs */
2034	err = mlx5_core_create_psv(dev->mdev, to_mpd(pd)->pdn, 2, psv_index);
2035	if (err)
2036		goto err_free_sig;
2037
2038	mr->sig->psv_memory.psv_idx = psv_index[0];
2039	mr->sig->psv_wire.psv_idx = psv_index[1];
2040
2041	mr->sig->sig_status_checked = true;
2042	mr->sig->sig_err_exists = false;
2043	/* Next UMR, Arm SIGERR */
2044	++mr->sig->sigerr_count;
2045	mr->klm_mr = mlx5_ib_alloc_pi_mr(pd, max_num_sg, max_num_meta_sg,
2046					 sizeof(struct mlx5_klm),
2047					 MLX5_MKC_ACCESS_MODE_KLMS);
2048	if (IS_ERR(mr->klm_mr)) {
2049		err = PTR_ERR(mr->klm_mr);
2050		goto err_destroy_psv;
2051	}
2052	mr->mtt_mr = mlx5_ib_alloc_pi_mr(pd, max_num_sg, max_num_meta_sg,
2053					 sizeof(struct mlx5_mtt),
2054					 MLX5_MKC_ACCESS_MODE_MTT);
2055	if (IS_ERR(mr->mtt_mr)) {
2056		err = PTR_ERR(mr->mtt_mr);
2057		goto err_free_klm_mr;
2058	}
2059
2060	/* Set bsf descriptors for mkey */
2061	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
2062	MLX5_SET(mkc, mkc, bsf_en, 1);
2063	MLX5_SET(mkc, mkc, bsf_octword_size, MLX5_MKEY_BSF_OCTO_SIZE);
2064
2065	err = _mlx5_alloc_mkey_descs(pd, mr, 4, sizeof(struct mlx5_klm), 0,
2066				     MLX5_MKC_ACCESS_MODE_KLMS, in, inlen);
2067	if (err)
2068		goto err_free_mtt_mr;
2069
2070	err = xa_err(xa_store(&dev->sig_mrs, mlx5_base_mkey(mr->mmkey.key),
2071			      mr->sig, GFP_KERNEL));
2072	if (err)
2073		goto err_free_descs;
2074	return 0;
2075
2076err_free_descs:
2077	destroy_mkey(dev, mr);
2078	mlx5_free_priv_descs(mr);
2079err_free_mtt_mr:
2080	mlx5_ib_dereg_mr(&mr->mtt_mr->ibmr, NULL);
2081	mr->mtt_mr = NULL;
2082err_free_klm_mr:
2083	mlx5_ib_dereg_mr(&mr->klm_mr->ibmr, NULL);
2084	mr->klm_mr = NULL;
2085err_destroy_psv:
2086	if (mlx5_core_destroy_psv(dev->mdev, mr->sig->psv_memory.psv_idx))
2087		mlx5_ib_warn(dev, "failed to destroy mem psv %d\n",
2088			     mr->sig->psv_memory.psv_idx);
2089	if (mlx5_core_destroy_psv(dev->mdev, mr->sig->psv_wire.psv_idx))
2090		mlx5_ib_warn(dev, "failed to destroy wire psv %d\n",
2091			     mr->sig->psv_wire.psv_idx);
2092err_free_sig:
2093	kfree(mr->sig);
2094
2095	return err;
2096}
2097
2098static struct ib_mr *__mlx5_ib_alloc_mr(struct ib_pd *pd,
2099					enum ib_mr_type mr_type, u32 max_num_sg,
2100					u32 max_num_meta_sg)
2101{
2102	struct mlx5_ib_dev *dev = to_mdev(pd->device);
2103	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
2104	int ndescs = ALIGN(max_num_sg, 4);
2105	struct mlx5_ib_mr *mr;
2106	u32 *in;
2107	int err;
2108
2109	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2110	if (!mr)
2111		return ERR_PTR(-ENOMEM);
2112
2113	in = kzalloc(inlen, GFP_KERNEL);
2114	if (!in) {
2115		err = -ENOMEM;
2116		goto err_free;
2117	}
2118
2119	mr->ibmr.device = pd->device;
2120	mr->umem = NULL;
2121
2122	switch (mr_type) {
2123	case IB_MR_TYPE_MEM_REG:
2124		err = mlx5_alloc_mem_reg_descs(pd, mr, ndescs, in, inlen);
2125		break;
2126	case IB_MR_TYPE_SG_GAPS:
2127		err = mlx5_alloc_sg_gaps_descs(pd, mr, ndescs, in, inlen);
2128		break;
2129	case IB_MR_TYPE_INTEGRITY:
2130		err = mlx5_alloc_integrity_descs(pd, mr, max_num_sg,
2131						 max_num_meta_sg, in, inlen);
2132		break;
2133	default:
2134		mlx5_ib_warn(dev, "Invalid mr type %d\n", mr_type);
2135		err = -EINVAL;
2136	}
2137
2138	if (err)
2139		goto err_free_in;
2140
2141	kfree(in);
2142
2143	return &mr->ibmr;
2144
2145err_free_in:
2146	kfree(in);
2147err_free:
2148	kfree(mr);
2149	return ERR_PTR(err);
2150}
2151
2152struct ib_mr *mlx5_ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2153			       u32 max_num_sg)
2154{
2155	return __mlx5_ib_alloc_mr(pd, mr_type, max_num_sg, 0);
2156}
2157
2158struct ib_mr *mlx5_ib_alloc_mr_integrity(struct ib_pd *pd,
2159					 u32 max_num_sg, u32 max_num_meta_sg)
2160{
2161	return __mlx5_ib_alloc_mr(pd, IB_MR_TYPE_INTEGRITY, max_num_sg,
2162				  max_num_meta_sg);
2163}
2164
2165int mlx5_ib_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
2166{
2167	struct mlx5_ib_dev *dev = to_mdev(ibmw->device);
2168	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
2169	struct mlx5_ib_mw *mw = to_mmw(ibmw);
2170	unsigned int ndescs;
2171	u32 *in = NULL;
2172	void *mkc;
2173	int err;
2174	struct mlx5_ib_alloc_mw req = {};
2175	struct {
2176		__u32	comp_mask;
2177		__u32	response_length;
2178	} resp = {};
2179
2180	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
2181	if (err)
2182		return err;
2183
2184	if (req.comp_mask || req.reserved1 || req.reserved2)
2185		return -EOPNOTSUPP;
2186
2187	if (udata->inlen > sizeof(req) &&
2188	    !ib_is_udata_cleared(udata, sizeof(req),
2189				 udata->inlen - sizeof(req)))
2190		return -EOPNOTSUPP;
2191
2192	ndescs = req.num_klms ? roundup(req.num_klms, 4) : roundup(1, 4);
2193
2194	in = kzalloc(inlen, GFP_KERNEL);
2195	if (!in)
2196		return -ENOMEM;
2197
2198	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
2199
2200	MLX5_SET(mkc, mkc, free, 1);
2201	MLX5_SET(mkc, mkc, translations_octword_size, ndescs);
2202	MLX5_SET(mkc, mkc, pd, to_mpd(ibmw->pd)->pdn);
2203	MLX5_SET(mkc, mkc, umr_en, 1);
2204	MLX5_SET(mkc, mkc, lr, 1);
2205	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_KLMS);
2206	MLX5_SET(mkc, mkc, en_rinval, !!((ibmw->type == IB_MW_TYPE_2)));
2207	MLX5_SET(mkc, mkc, qpn, 0xffffff);
2208
2209	err = mlx5_ib_create_mkey(dev, &mw->mmkey, in, inlen);
2210	if (err)
2211		goto free;
2212
2213	mw->mmkey.type = MLX5_MKEY_MW;
2214	ibmw->rkey = mw->mmkey.key;
2215	mw->mmkey.ndescs = ndescs;
2216
2217	resp.response_length =
2218		min(offsetofend(typeof(resp), response_length), udata->outlen);
2219	if (resp.response_length) {
2220		err = ib_copy_to_udata(udata, &resp, resp.response_length);
2221		if (err)
2222			goto free_mkey;
2223	}
2224
2225	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
2226		err = mlx5r_store_odp_mkey(dev, &mw->mmkey);
2227		if (err)
2228			goto free_mkey;
2229	}
2230
2231	kfree(in);
2232	return 0;
2233
2234free_mkey:
2235	mlx5_core_destroy_mkey(dev->mdev, mw->mmkey.key);
2236free:
2237	kfree(in);
2238	return err;
2239}
2240
2241int mlx5_ib_dealloc_mw(struct ib_mw *mw)
2242{
2243	struct mlx5_ib_dev *dev = to_mdev(mw->device);
2244	struct mlx5_ib_mw *mmw = to_mmw(mw);
2245
2246	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) &&
2247	    xa_erase(&dev->odp_mkeys, mlx5_base_mkey(mmw->mmkey.key)))
2248		/*
2249		 * pagefault_single_data_segment() may be accessing mmw
2250		 * if the user bound an ODP MR to this MW.
2251		 */
2252		mlx5r_deref_wait_odp_mkey(&mmw->mmkey);
2253
2254	return mlx5_core_destroy_mkey(dev->mdev, mmw->mmkey.key);
2255}
2256
2257int mlx5_ib_check_mr_status(struct ib_mr *ibmr, u32 check_mask,
2258			    struct ib_mr_status *mr_status)
2259{
2260	struct mlx5_ib_mr *mmr = to_mmr(ibmr);
2261	int ret = 0;
2262
2263	if (check_mask & ~IB_MR_CHECK_SIG_STATUS) {
2264		pr_err("Invalid status check mask\n");
2265		ret = -EINVAL;
2266		goto done;
2267	}
2268
2269	mr_status->fail_status = 0;
2270	if (check_mask & IB_MR_CHECK_SIG_STATUS) {
2271		if (!mmr->sig) {
2272			ret = -EINVAL;
2273			pr_err("signature status check requested on a non-signature enabled MR\n");
2274			goto done;
2275		}
2276
2277		mmr->sig->sig_status_checked = true;
2278		if (!mmr->sig->sig_err_exists)
2279			goto done;
2280
2281		if (ibmr->lkey == mmr->sig->err_item.key)
2282			memcpy(&mr_status->sig_err, &mmr->sig->err_item,
2283			       sizeof(mr_status->sig_err));
2284		else {
2285			mr_status->sig_err.err_type = IB_SIG_BAD_GUARD;
2286			mr_status->sig_err.sig_err_offset = 0;
2287			mr_status->sig_err.key = mmr->sig->err_item.key;
2288		}
2289
2290		mmr->sig->sig_err_exists = false;
2291		mr_status->fail_status |= IB_MR_CHECK_SIG_STATUS;
2292	}
2293
2294done:
2295	return ret;
2296}
2297
2298static int
2299mlx5_ib_map_pa_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
2300			int data_sg_nents, unsigned int *data_sg_offset,
2301			struct scatterlist *meta_sg, int meta_sg_nents,
2302			unsigned int *meta_sg_offset)
2303{
2304	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2305	unsigned int sg_offset = 0;
2306	int n = 0;
2307
2308	mr->meta_length = 0;
2309	if (data_sg_nents == 1) {
2310		n++;
2311		mr->mmkey.ndescs = 1;
2312		if (data_sg_offset)
2313			sg_offset = *data_sg_offset;
2314		mr->data_length = sg_dma_len(data_sg) - sg_offset;
2315		mr->data_iova = sg_dma_address(data_sg) + sg_offset;
2316		if (meta_sg_nents == 1) {
2317			n++;
2318			mr->meta_ndescs = 1;
2319			if (meta_sg_offset)
2320				sg_offset = *meta_sg_offset;
2321			else
2322				sg_offset = 0;
2323			mr->meta_length = sg_dma_len(meta_sg) - sg_offset;
2324			mr->pi_iova = sg_dma_address(meta_sg) + sg_offset;
2325		}
2326		ibmr->length = mr->data_length + mr->meta_length;
2327	}
2328
2329	return n;
2330}
2331
2332static int
2333mlx5_ib_sg_to_klms(struct mlx5_ib_mr *mr,
2334		   struct scatterlist *sgl,
2335		   unsigned short sg_nents,
2336		   unsigned int *sg_offset_p,
2337		   struct scatterlist *meta_sgl,
2338		   unsigned short meta_sg_nents,
2339		   unsigned int *meta_sg_offset_p)
2340{
2341	struct scatterlist *sg = sgl;
2342	struct mlx5_klm *klms = mr->descs;
2343	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2344	u32 lkey = mr->ibmr.pd->local_dma_lkey;
2345	int i, j = 0;
2346
2347	mr->ibmr.iova = sg_dma_address(sg) + sg_offset;
2348	mr->ibmr.length = 0;
2349
2350	for_each_sg(sgl, sg, sg_nents, i) {
2351		if (unlikely(i >= mr->max_descs))
2352			break;
2353		klms[i].va = cpu_to_be64(sg_dma_address(sg) + sg_offset);
2354		klms[i].bcount = cpu_to_be32(sg_dma_len(sg) - sg_offset);
2355		klms[i].key = cpu_to_be32(lkey);
2356		mr->ibmr.length += sg_dma_len(sg) - sg_offset;
2357
2358		sg_offset = 0;
2359	}
2360
2361	if (sg_offset_p)
2362		*sg_offset_p = sg_offset;
2363
2364	mr->mmkey.ndescs = i;
2365	mr->data_length = mr->ibmr.length;
2366
2367	if (meta_sg_nents) {
2368		sg = meta_sgl;
2369		sg_offset = meta_sg_offset_p ? *meta_sg_offset_p : 0;
2370		for_each_sg(meta_sgl, sg, meta_sg_nents, j) {
2371			if (unlikely(i + j >= mr->max_descs))
2372				break;
2373			klms[i + j].va = cpu_to_be64(sg_dma_address(sg) +
2374						     sg_offset);
2375			klms[i + j].bcount = cpu_to_be32(sg_dma_len(sg) -
2376							 sg_offset);
2377			klms[i + j].key = cpu_to_be32(lkey);
2378			mr->ibmr.length += sg_dma_len(sg) - sg_offset;
2379
2380			sg_offset = 0;
2381		}
2382		if (meta_sg_offset_p)
2383			*meta_sg_offset_p = sg_offset;
2384
2385		mr->meta_ndescs = j;
2386		mr->meta_length = mr->ibmr.length - mr->data_length;
2387	}
2388
2389	return i + j;
2390}
2391
2392static int mlx5_set_page(struct ib_mr *ibmr, u64 addr)
2393{
2394	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2395	__be64 *descs;
2396
2397	if (unlikely(mr->mmkey.ndescs == mr->max_descs))
2398		return -ENOMEM;
2399
2400	descs = mr->descs;
2401	descs[mr->mmkey.ndescs++] = cpu_to_be64(addr | MLX5_EN_RD | MLX5_EN_WR);
2402
2403	return 0;
2404}
2405
2406static int mlx5_set_page_pi(struct ib_mr *ibmr, u64 addr)
2407{
2408	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2409	__be64 *descs;
2410
2411	if (unlikely(mr->mmkey.ndescs + mr->meta_ndescs == mr->max_descs))
2412		return -ENOMEM;
2413
2414	descs = mr->descs;
2415	descs[mr->mmkey.ndescs + mr->meta_ndescs++] =
2416		cpu_to_be64(addr | MLX5_EN_RD | MLX5_EN_WR);
2417
2418	return 0;
2419}
2420
2421static int
2422mlx5_ib_map_mtt_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
2423			 int data_sg_nents, unsigned int *data_sg_offset,
2424			 struct scatterlist *meta_sg, int meta_sg_nents,
2425			 unsigned int *meta_sg_offset)
2426{
2427	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2428	struct mlx5_ib_mr *pi_mr = mr->mtt_mr;
2429	int n;
2430
2431	pi_mr->mmkey.ndescs = 0;
2432	pi_mr->meta_ndescs = 0;
2433	pi_mr->meta_length = 0;
2434
2435	ib_dma_sync_single_for_cpu(ibmr->device, pi_mr->desc_map,
2436				   pi_mr->desc_size * pi_mr->max_descs,
2437				   DMA_TO_DEVICE);
2438
2439	pi_mr->ibmr.page_size = ibmr->page_size;
2440	n = ib_sg_to_pages(&pi_mr->ibmr, data_sg, data_sg_nents, data_sg_offset,
2441			   mlx5_set_page);
2442	if (n != data_sg_nents)
2443		return n;
2444
2445	pi_mr->data_iova = pi_mr->ibmr.iova;
2446	pi_mr->data_length = pi_mr->ibmr.length;
2447	pi_mr->ibmr.length = pi_mr->data_length;
2448	ibmr->length = pi_mr->data_length;
2449
2450	if (meta_sg_nents) {
2451		u64 page_mask = ~((u64)ibmr->page_size - 1);
2452		u64 iova = pi_mr->data_iova;
2453
2454		n += ib_sg_to_pages(&pi_mr->ibmr, meta_sg, meta_sg_nents,
2455				    meta_sg_offset, mlx5_set_page_pi);
2456
2457		pi_mr->meta_length = pi_mr->ibmr.length;
2458		/*
2459		 * PI address for the HW is the offset of the metadata address
2460		 * relative to the first data page address.
2461		 * It equals to first data page address + size of data pages +
2462		 * metadata offset at the first metadata page
2463		 */
2464		pi_mr->pi_iova = (iova & page_mask) +
2465				 pi_mr->mmkey.ndescs * ibmr->page_size +
2466				 (pi_mr->ibmr.iova & ~page_mask);
2467		/*
2468		 * In order to use one MTT MR for data and metadata, we register
2469		 * also the gaps between the end of the data and the start of
2470		 * the metadata (the sig MR will verify that the HW will access
2471		 * to right addresses). This mapping is safe because we use
2472		 * internal mkey for the registration.
2473		 */
2474		pi_mr->ibmr.length = pi_mr->pi_iova + pi_mr->meta_length - iova;
2475		pi_mr->ibmr.iova = iova;
2476		ibmr->length += pi_mr->meta_length;
2477	}
2478
2479	ib_dma_sync_single_for_device(ibmr->device, pi_mr->desc_map,
2480				      pi_mr->desc_size * pi_mr->max_descs,
2481				      DMA_TO_DEVICE);
2482
2483	return n;
2484}
2485
2486static int
2487mlx5_ib_map_klm_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
2488			 int data_sg_nents, unsigned int *data_sg_offset,
2489			 struct scatterlist *meta_sg, int meta_sg_nents,
2490			 unsigned int *meta_sg_offset)
2491{
2492	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2493	struct mlx5_ib_mr *pi_mr = mr->klm_mr;
2494	int n;
2495
2496	pi_mr->mmkey.ndescs = 0;
2497	pi_mr->meta_ndescs = 0;
2498	pi_mr->meta_length = 0;
2499
2500	ib_dma_sync_single_for_cpu(ibmr->device, pi_mr->desc_map,
2501				   pi_mr->desc_size * pi_mr->max_descs,
2502				   DMA_TO_DEVICE);
2503
2504	n = mlx5_ib_sg_to_klms(pi_mr, data_sg, data_sg_nents, data_sg_offset,
2505			       meta_sg, meta_sg_nents, meta_sg_offset);
2506
2507	ib_dma_sync_single_for_device(ibmr->device, pi_mr->desc_map,
2508				      pi_mr->desc_size * pi_mr->max_descs,
2509				      DMA_TO_DEVICE);
2510
2511	/* This is zero-based memory region */
2512	pi_mr->data_iova = 0;
2513	pi_mr->ibmr.iova = 0;
2514	pi_mr->pi_iova = pi_mr->data_length;
2515	ibmr->length = pi_mr->ibmr.length;
2516
2517	return n;
2518}
2519
2520int mlx5_ib_map_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
2521			 int data_sg_nents, unsigned int *data_sg_offset,
2522			 struct scatterlist *meta_sg, int meta_sg_nents,
2523			 unsigned int *meta_sg_offset)
2524{
2525	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2526	struct mlx5_ib_mr *pi_mr = NULL;
2527	int n;
2528
2529	WARN_ON(ibmr->type != IB_MR_TYPE_INTEGRITY);
2530
2531	mr->mmkey.ndescs = 0;
2532	mr->data_length = 0;
2533	mr->data_iova = 0;
2534	mr->meta_ndescs = 0;
2535	mr->pi_iova = 0;
2536	/*
2537	 * As a performance optimization, if possible, there is no need to
2538	 * perform UMR operation to register the data/metadata buffers.
2539	 * First try to map the sg lists to PA descriptors with local_dma_lkey.
2540	 * Fallback to UMR only in case of a failure.
2541	 */
2542	n = mlx5_ib_map_pa_mr_sg_pi(ibmr, data_sg, data_sg_nents,
2543				    data_sg_offset, meta_sg, meta_sg_nents,
2544				    meta_sg_offset);
2545	if (n == data_sg_nents + meta_sg_nents)
2546		goto out;
2547	/*
2548	 * As a performance optimization, if possible, there is no need to map
2549	 * the sg lists to KLM descriptors. First try to map the sg lists to MTT
2550	 * descriptors and fallback to KLM only in case of a failure.
2551	 * It's more efficient for the HW to work with MTT descriptors
2552	 * (especially in high load).
2553	 * Use KLM (indirect access) only if it's mandatory.
2554	 */
2555	pi_mr = mr->mtt_mr;
2556	n = mlx5_ib_map_mtt_mr_sg_pi(ibmr, data_sg, data_sg_nents,
2557				     data_sg_offset, meta_sg, meta_sg_nents,
2558				     meta_sg_offset);
2559	if (n == data_sg_nents + meta_sg_nents)
2560		goto out;
2561
2562	pi_mr = mr->klm_mr;
2563	n = mlx5_ib_map_klm_mr_sg_pi(ibmr, data_sg, data_sg_nents,
2564				     data_sg_offset, meta_sg, meta_sg_nents,
2565				     meta_sg_offset);
2566	if (unlikely(n != data_sg_nents + meta_sg_nents))
2567		return -ENOMEM;
2568
2569out:
2570	/* This is zero-based memory region */
2571	ibmr->iova = 0;
2572	mr->pi_mr = pi_mr;
2573	if (pi_mr)
2574		ibmr->sig_attrs->meta_length = pi_mr->meta_length;
2575	else
2576		ibmr->sig_attrs->meta_length = mr->meta_length;
2577
2578	return 0;
2579}
2580
2581int mlx5_ib_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
2582		      unsigned int *sg_offset)
2583{
2584	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2585	int n;
2586
2587	mr->mmkey.ndescs = 0;
2588
2589	ib_dma_sync_single_for_cpu(ibmr->device, mr->desc_map,
2590				   mr->desc_size * mr->max_descs,
2591				   DMA_TO_DEVICE);
2592
2593	if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
2594		n = mlx5_ib_sg_to_klms(mr, sg, sg_nents, sg_offset, NULL, 0,
2595				       NULL);
2596	else
2597		n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset,
2598				mlx5_set_page);
2599
2600	ib_dma_sync_single_for_device(ibmr->device, mr->desc_map,
2601				      mr->desc_size * mr->max_descs,
2602				      DMA_TO_DEVICE);
2603
2604	return n;
2605}